content
stringlengths 1
15.9M
|
|---|
\section{Background}
\mylabel{s:background}
A \textbf{Legendrian knot} in the standard contact structure on $\ensuremath{\mathbb{R}}^3$ is a smooth knot $\ensuremath{L} : S^1 \to \ensuremath{\mathbb{R}}^3$ satisfying $\ensuremath{L}'(t) \in \ker(dz-y\,dx)$ for all $t \in S^1$. A smooth one-parameter family $\ensuremath{L}_t$, $0 \leq t \leq 1$, of Legendrian knots is a \textbf{Legendrian isotopy} between $\ensuremath{L}_0$ and $\ensuremath{L}_1$. The \textbf{front diagram} of $\ensuremath{L}$ is the projection of $\ensuremath{L}$ to the $xz$-plane. Every Legendrian knot is Legendrian isotopic, by an arbitrarily small Legendrian isotopy, to a Legendrian knot whose front diagram is embedded except at transverse self-intersections, called \textbf{crossings}, and semi-cubical cusps such that, in addition, all of these exceptional points have distinct $x$-coordinates. A Legendrian knot with such a front diagram is said to have a \textbf{$\sigma$-generic} front diagram; see, for example, the front diagram in Figure~\ref{f:knot}. In a neighborhood of an $x$ value that is not the $x$-coordinate of a crossing or cusp, the front diagram looks like a collection of non-intersecting line segments commonly called the \textbf{strands} of $\ensuremath{D}$ at $x$. Orient $\ensuremath{L}$. The \textbf{rotation number} $r(\ensuremath{L})$ is $(d - u)/2$ where $d$ (resp. $u$) is the number of cusps at which the orientation travels downward (resp. upward) with respect to the $z$-axis.
\begin{figure}[t]
\centering
\includegraphics[scale=.9]{./images/knot}
\caption{A $\sigma$-generic front diagram of a Legendrian knot with rotation number $0$.}
\label{f:knot}
\end{figure}
\subsection{Chekanov-Eliashberg Algebra}
Fix a Legendrian knot $\ensuremath{L}$ with $\sigma$-generic front diagram $\ensuremath{D}$ and rotation number $0$. A \textbf{Maslov potential} is a map $\ensuremath{\mu} : \ensuremath{L} \to \ensuremath{\mathbb{Z}}$ that is constant except at cusp points of $\ensuremath{L}$ where the Maslov potential of the lower strand of the cusp is 1 less than the upper strand. Let $A(\ensuremath{D})$ be the $\ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}}$ vector space generated by the labels $Q = \{q_1, \hdots, q_n\}$ assigned to the crossings and right cusps of $\ensuremath{D}$. A generator $q \in Q$ is assigned a \textbf{grading} $|q|$, also called a \textbf{degree}, so that $|q|$ is 1 if $q$ is a right cusp and, otherwise, $|q|$ is $\ensuremath{\mu}(T) - \ensuremath{\mu}(B)$ where $T$ and $B$ are the strands of $\ensuremath{D}$ crossing at $q$ and $T$ has smaller slope. The graded algebra $\ensuremath{\mathcal{A}}(\ensuremath{D})$ is the unital tensor algebra $TA(\ensuremath{D})$. The \textbf{Chekanov-Eliashberg algebra}, written $(\ensuremath{\mathcal{A}}(\ensuremath{D}), \ensuremath{\partial})$, is the algebra $\ensuremath{\mathcal{A}}(\ensuremath{D})$ along with a degree -1 differential $\ensuremath{\partial} : \ensuremath{\mathcal{A}}(\ensuremath{D}) \to \ensuremath{\mathcal{A}}(\ensuremath{D})$ that, in the case of the front diagram description from \cite{Ng2003}, is defined by counting certain admissible maps of the two-disk $D^2$ into the $xz$-plane. We refer the reader to \cite{Ng2003} for a careful definition of $\ensuremath{\partial}$ as we will need to investigate only a small subset of these maps.
An \textbf{augmentation} is a map $\ensuremath{\epsilon} : \ensuremath{\mathcal{A}}(\ensuremath{D}) \to \ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}}$ satisfying $\ensuremath{\epsilon} \circ \ensuremath{\partial} = 0$, $\ensuremath{\epsilon}(1) = 1$, and $\ensuremath{\epsilon}(q) = 1$ only if $|q|=0$. The set $\mbox{Aug}(\ensuremath{D})$ is the set of all augmentations of $(\ensuremath{\mathcal{A}}(\ensuremath{D}), \ensuremath{\partial})$. We say a crossing $q$ is \textbf{augmented} by $\ensuremath{\epsilon}$ if $\ensuremath{\epsilon}(q) = 1$. An augmentation can be thought of as a morphism between the differential graded algebra $(\ensuremath{\mathcal{A}}(\ensuremath{D}), \ensuremath{\partial})$ and the differential graded algebra $(\ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}}, \ensuremath{\partial}')$ whose only non-zero element is in degree $0$ and where $\ensuremath{\partial}' = 0$. From this perspective, there is a natural algebraic equivalence relation on $\mbox{Aug}(\ensuremath{D})$. Given $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ in $\mbox{Aug}(\ensuremath{D})$, a \textbf{chain homotopy} from $\ensuremath{\epsilon}$ to $\ensuremath{\epsilon}'$ is a degree $1$ linear map $H : (\ensuremath{\mathcal{A}}(\ensuremath{D}), \ensuremath{\partial}) \to (\ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}}, \ensuremath{\partial}')$ satisfying $\ensuremath{\epsilon} - \ensuremath{\epsilon}' = \ensuremath{\partial}' \circ H +H \circ \ensuremath{\partial} $ and $H(ab) = H(a) \ensuremath{\epsilon}'(b) + (-1)^{|a|}\ensuremath{\epsilon}(a) H(b)$ for all $a, b \in \ensuremath{\mathcal{A}}(\ensuremath{D})$. Since we are working over $\ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}}$ and $\ensuremath{\partial}'=0$, these conditions simplify to
\begin{equation}
\label{eq:chain-homotopy}
\ensuremath{\epsilon} - \ensuremath{\epsilon}' = H \circ \ensuremath{\partial} \mbox{ and } H(ab) = H(a)\ensuremath{\epsilon}'(b) + \ensuremath{\epsilon}(a)H(b).
\end{equation}
By Lemma 2.18 of \cite{Kalman2005}, a chain homotopy $H$ is determined by the values it takes on the degree $-1$ crossings of $\ensuremath{D}$.
We say augmentations $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ are \textbf{homotopic} and write $\ensuremath{\epsilon} \simeq \ensuremath{\epsilon}'$ if there exists a chain homotopy from $\ensuremath{\epsilon}$ to $\ensuremath{\epsilon}'$. As the notation implies and as is proven in \cite{YvesFelix1995}, chain homotopy provides an equivalence relation on the set $\mbox{Aug}(\ensuremath{D})$. We let $\mbox{Aug}^{ch}(\ensuremath{D})$ be $\mbox{Aug}(\ensuremath{D}) / \simeq$. By Proposition 4.5 of \cite{Henry2011}, the count of homotopy classes of augmentations is a Legendrian isotopy invariant.
Suppose $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ are augmentations in $\mbox{Aug}(\ensuremath{D})$ and there exists a chain homotopy $H$ from $\ensuremath{\epsilon}$ to $\ensuremath{\epsilon}'$. Suppose $q$ is a degree 0 crossing and $\langle \ensuremath{\partial} q, \prod_{i=1}^m q_{k_i} \rangle$ is $1$, where $\langle \ensuremath{\partial} q, \prod_{i=1}^m q_{k_i} \rangle$ is the coefficient of $\prod_{i=1}^m q_{k_i}$ in $\ensuremath{\partial} q$. Then, by Equation (\ref{eq:chain-homotopy}),
\begin{eqnarray*}
(\ensuremath{\epsilon} - \ensuremath{\epsilon}')(q) &=& H \circ \partial (q) \\ \nonumber
&=& H \left (\prod_{i=1}^m q_{k_i} + \hdots \right ) \\ \nonumber
&=& H \left (\prod_{i=1}^m q_{k_i} \right ) + H( \hdots) \\ \nonumber
&=& \sum_{j=1}^m \left [ \left ( \prod_{i=1}^{j-1} \ensuremath{\epsilon}(q_{k_i}) \right ) H(q_{k_j}) \left ( \prod_{i=j+1}^m \ensuremath{\epsilon}' (q_{k_i}) \right ) \right ] + H( \hdots). \\ \nonumber
\end{eqnarray*}
At most one term in the sum $$\sum_{j=1}^m \left [ \left ( \prod_{i=1}^{j-1} \ensuremath{\epsilon}(q_{k_i}) \right ) H(q_{k_j}) \left ( \prod_{i=j+1}^m \ensuremath{\epsilon}' (q_{k_i}) \right ) \right ]$$ may be non-zero, since $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ are non-zero only on generators of degree 0 and $H$ is non-zero only on generators of degree $-1$. Note that, for a fixed $j \in \{1, \hdots, m\}$, the term $$ \left ( \prod_{i=1}^{j-1} \ensuremath{\epsilon}(q_{k_i}) \right ) H(q_{k_j}) \left ( \prod_{i=j+1}^m \ensuremath{\epsilon}' (q_{k_i}) \right ) $$ is non-zero if and only if $H(q_{k_j})=1$ holds and for $1 \leq i \leq j-1$ (resp. $j+1 \leq i \leq m$), the crossing $q_{k_i}$ is augmented by $\ensuremath{\epsilon}$ (resp. $\ensuremath{\epsilon}'$).
\begin{remark}
The monomials $\prod_{i=1}^m q_{k_i}$ appearing in $\partial(q)$ correspond to certain mappings of the two-disk $D^2$ into the $xz$-plane that are immersions except for allowable exceptions along $\partial D^2$. Only monomials containing generators of degree $0$ or $-1$ are relevant for our purposes. Therefore, we present only the description of such disks in the following definitions. Note that this restriction allows us to rule out some additional behaviors of $\partial D^2$ near right cusps that appear in \cite{Ng2003} and lead to monomials that contain generators of degree $1$.
\end{remark}
Let $D^2$ be the disk of radius 1 centered at the origin in $\ensuremath{\mathbb{R}}^2$. Choose $m$ points from $\ensuremath{\partial} D^2 \setminus \{(1,0)\}$. Label the chosen points $\{b_1, \hdots, b_m\}$ counter-clockwise with $b_1$ the first point counter-clockwise from $(1,0)$.
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {(a)} [tl] at 28 78
\pinlabel {(b)} [tl] at 148 78
\pinlabel {(c)} [tl] at 271 44
\pinlabel {(d)} [tl] at 28 -2
\pinlabel {(e)} [tl] at 148 -2
\pinlabel {(f)} [tl] at 271 -2
\pinlabel {i} [tl] at 233 33
\pinlabel {j} [tl] at 233 6
\pinlabel {$\{x_0\} \times [i,j]$} [tl] at 268 20
\endlabellist
\centering
\includegraphics[scale=.9]{./images/admissible2}
\caption{The possible singularities of the disk in Definition~\ref{defn:admissible} and the half-disks in Definitions~\ref{defn:aug-half-disk} and \ref{defn:half-disk}. The crossings in (d) and (e) are called convex corners. Near a boundary point that maps to a right cusp the image of a disk overlaps itself as indicated in (c) by the darkly shaded region.}
\label{f:admissible}
\end{figure}
\begin{definition}
\label{defn:admissible}
In terms of the notation above, a \textbf{$(0,-1)$-admissible disk} is a map from $D^2$ into the $xz$-plane that maps $\ensuremath{\partial} D^2$ to the front diagram $\ensuremath{D}$ and is a smooth orientation preserving immersion when restricted to the interior of $D^2$ satisfying the following:
\begin{enumerate}
\item The mapping takes $(1,0)$ to a degree 0 crossing $q$ and the image of $f$ in a neighborhood of $(1,0)$ looks as in Figure~\ref{f:admissible} (a). We say the $(0,-1)$-admissible disk \textbf{originates at $q$};
\item For exactly one $1 \leq j \leq m$, $f(b_j)$ is a degree $-1$ crossing $q_{k_j}$ and the image of $f$ in a neighborhood of $b_j$ looks as in Figure~\ref{f:admissible} (d) or (e).
\item For all $i \neq j$, $f(b_i)$ is a degree 0 crossing $q_{k_i}$ and the image of $f$ in a neighborhood of $b_i$ looks as in Figure~\ref{f:admissible} (d) or (e).
\item Along $\partial D^2$ the mapping is smooth except at $\{b_1, \hdots, b_m\} \cup \{(1,0)\}$ as described in (1)-(3) and at points in $\ensuremath{\partial} D^2 \setminus (\{b_1, \hdots, b_m\} \cup \{(1,0)\})$ where the image of $f$ looks like either Figure~\ref{f:admissible} (b) or (c).
\end{enumerate}
We say the $(0,-1)$-admissible disk has \textbf{convex corners} at $q_{k_1}, \hdots, q_{k_m}$. The $(0, -1)$-admissible disk is assigned the monomial $\prod_{i=1}^m q_{k_i}$. We say a $(0,-1)$-admissible disk is an \textbf{$(\ensuremath{\epsilon}, \ensuremath{\epsilon}', H)$-admissible disk} if, for some $1 \leq j \leq m$, $H(q_{k_j})=1$ holds and for $1 \leq i \leq j-1$ (resp. $j+1 \leq i \leq m$), the crossing $q_{k_i}$ is augmented by $\ensuremath{\epsilon}$ (resp. $\ensuremath{\epsilon}'$); see Figure \ref{f:eHeDisk}.
Henceforth, we consider admissible disks up to orientation preserving reparametrization of the domain (fixing $\{b_1, \hdots, b_m\} \cup \{(1,0)\}$), and all counts of disks are up to this equivalence relation.
\end{definition}
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {$q$} [l] at 170 80
\pinlabel {$\ensuremath{\epsilon}$} [bl] at 152 133
\pinlabel {$\ensuremath{\epsilon}$} [b] at 109 161
\pinlabel {$\ensuremath{\epsilon}$} [b] at 62 161
\pinlabel {$H$} [b] at 15 133
\pinlabel {$\ensuremath{\epsilon}'$} [r] at -1 80
\pinlabel {$\ensuremath{\epsilon}'$} [tr] at 15 27
\pinlabel {$\ensuremath{\epsilon}'$} [t] at 58 -1
\pinlabel {$\ensuremath{\epsilon}'$} [t] at 107 -1
\pinlabel {$\ensuremath{\epsilon}'$} [tl] at 153 28
\endlabellist
\centering
\includegraphics[scale=.6]{./images/eHeDisk}
\caption{The domain of an $(\ensuremath{\epsilon}, \ensuremath{\epsilon}', H)$-admissible disk with labels indicating marked points mapped to crossings augmented by $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ and the marked point mapped to the crossing satisfying $H(q_{k_j})=1$.}
\label{f:eHeDisk}
\end{figure}
The restrictions on the types of non-smooth points of an $(0,-1)$-admissible disk imply that $q$ is the right-most point of the disk. From Section 2 of \cite{Ng2003}, when a single $q_{k_j}$ has degree $1$ while $q$ and all of the remaining $q_{k_i}$ have degree $0$, $\langle \ensuremath{\partial} q, \prod_{i=1}^m q_{k_i} \rangle = 1$ holds if and only if there are an odd number of $(0,-1)$-admissible disks originating at $q$ and with monomial $\prod_{i=1}^m q_{k_i}$. Proposition~\ref{prop:admissible-disks} follows directly from the discussion above.
\begin{proposition}
\label{prop:admissible-disks}
Suppose $\ensuremath{D}$ is a $\sigma$-generic front diagram of a Legendrian knot and $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ are augmentations in $\mbox{Aug}(\ensuremath{D})$. If $q$ is a degree $0$ crossing and $H$ is a chain homotopy from $\ensuremath{\epsilon}$ to $\ensuremath{\epsilon}'$, then $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ differ at $q$ if and only if there are an odd number of $(\ensuremath{\epsilon}, \ensuremath{\epsilon}', H)$-admissible disks originating at $q$.
\end{proposition}
\section{Introduction}
\mylabel{s:intro}
The symplectic techniques of holomorphic curves and generating families provide two effective classes of invariants of Legendrian knots in standard contact $\ensuremath{\mathbb{R}}^3$. The holomorphic curve approach, which in this low-dimensional setting takes on a combinatorial flavor, can be used to define a Differential Graded Algebra (DGA), known alternatively as the Legendrian contact homology DGA or the Chekanov-Eliashberg DGA and originally defined in \cite{Chekanov2002a} and \cite{Eliashberg2000}. Generating families of Legendrian submanifolds in $1$-jet spaces, including $\ensuremath{\mathbb{R}}^3$, have also been used to produce homological Legendrian invariants; see, for instance, \cite{Jordan2006}, \cite{SabloffT2013}, \cite{Traynor1997}, \cite{Traynor2001}. In addition to distinguishing Legendrian isotopy classes of knots, both the holomorphic and generating family invariants carry useful information about Lagrangian cobordisms, cf. \cite{EkholmHK}, \cite{SabloffT2013}.
For Legendrian knots in $\ensuremath{\mathbb{R}}^3$, several close connections have been discovered between holomorphic curve and generating family invariants, although many questions remain. For example, the existence of a linear at infinity generating family for a Legendrian knot is known to be equivalent to the existence of a certain DGA morphism, called an augmentation, from the Chekanov-Eliashberg DGA to its ground ring, (\cite{Chekanov2005}, \cite{Fuchs2003}, \cite{Fuchs2004}, \cite{Fuchs2008}, \cite{Sabloff2005}). However, it is unknown if this statement can be strengthened to a bijective correspondence between appropriate equivalence classes of generating families and augmentations. In this article, we approach this question using a discrete analog of a generating family called a Morse complex sequence, abbreviated MCS. MCSs have proven to be more tractable for explicit construction and computation; see, for example, \cite{Henry2011,Henry2013,Henry2014}.
A generating family for a Legendrian $\ensuremath{L} \subset \ensuremath{\mathbb{R}}^3$ is a one-parameter family of functions whose critical values coincide with the projection of $\ensuremath{L}$ to the $xz$-plane, also called the front diagram of $\ensuremath{L}$. A Morse complex sequence for $\ensuremath{L}$ is a collection of chain complexes and formal handleslide marks on the front diagram of $\ensuremath{L}$ that obey a set of restrictions identical to those satisfied by Morse complexes in a generic one-parameter family of functions. Thus, by considering MCSs rather than generating families, a family of functions is replaced by discrete, algebraic data. There is a natural equivalence relation on the set of MCSs on $\ensuremath{L}$ that reflects the generic bifurcations appearing in two-parameter families of Morse complexes.
The concept of a Morse complex sequence originally appeared in unpublished work of Petya Pushkar, and first appears in print in the work of the first author \cite{Henry2011} where MCSs are studied in connection with augmentations. In \cite{Henry2011}, a surjective map is defined from MCSs of $\ensuremath{L}$ to augmentations of the Chekanov-Eliashberg DGA of $\ensuremath{L}$. Moreover, equivalent MCSs are mapped to homotopic augmentations. In the present article, we complement the results of \cite{Henry2011} by showing in Lemma~\ref{l:equiv} that two MCSs mapped to homotopic augmentations must, in fact, be equivalent as MCSs. Combined with \cite{Henry2011} this gives the following.
\begin{theorem} \label{t:bijection}
For any Legendrian knot $\ensuremath{L} \subset \ensuremath{\mathbb{R}}^3$ with generic front diagram, there is a bijection between equivalence classes of Morse complex sequences for $\ensuremath{L}$ and homotopy classes of augmentations of the Chekanov-Eliashberg DGA of $\ensuremath{L}$.
\end{theorem}
As a consequence, the number of MCS equivalence classes is a Legendrian isotopy invariant; see Corollary \ref{c:MCS-invt}. The less immediate Corollary \ref{c:LCH} combines Theorem \ref{t:bijection} with previous work of the authors from \cite{Henry2013} to deduce that homotopic augmentations must have isomorphic linearized homology groups. The set of linearized homology groups is a Legendrian isotopy invariant. Corollary \ref{c:LCH} allows for a refinement of this invariant by considering multiplicities.
The remainder of the article is organized as follows. Section \ref{s:background} recalls background concerning augmentations and Morse complex sequences, while Section \ref{s:results} contains the proof of Theorem \ref{t:bijection} and its corollaries.
\subsection{Acknowledgments}
We gratefully acknowledge The Royal Academy of Belgium where conversations between the authors on this project began in August 2013 at a workshop on Legendrian submanifolds, holomorphic curves, and generating families. We also thank the workshop organizer Fr\'{e}d\'{e}ric Bourgeois. In addition, we thank the American Institute of Mathematics for supporting a SQuaRE research group on augmentations and related topics, and we thank our fellow SQuaRE participants Dmitry Fuchs, Paul Melvin, Josh Sabloff, and Lisa Traynor.
\section{The Main Result}
\label{s:results}
Suppose $\ensuremath{L}$ is a Legendrian knot with $\sigma$-generic front diagram $\ensuremath{D}$, rotation number 0, and Maslov potential $\ensuremath{\mu}$. Before proving Theorem~\ref{t:bijection}, we require two definitions and a technical lemma.
Let $D^2$ be the disk of radius 1 centered at the origin in $\ensuremath{\mathbb{R}}^2$. Choose $m+2$ points on $\ensuremath{\partial} D^2$. Label the chosen points $\{b_0, \hdots, b_{m+1}\}$ counter-clockwise. Let $\gamma$ be the arc of $\ensuremath{\partial} D^2$ with endpoints $b_{m+1}$ and $b_0$ and so that $b_1, \hdots, b_m$ are not in $\gamma$. Given $x_0 \in \ensuremath{\mathbb{R}}$ that is not the $x$-coordinate of any crossing or cusp of $\ensuremath{D}$, we let $\{x_0\} \times [i,j]$ denote the vertical line segment with $x$-coordinate $x_0$ and endpoints on strands $i$ and $j$ of $\ensuremath{D}$, where the strands of $\ensuremath{D}$ above $x=x_0$ are numbered $1, 2, \hdots$ from top to bottom and $i < j$.
\begin{definition}
\label{defn:half-disk}
Let $\ensuremath{\epsilon}$ and $\ensuremath{\epsilon}'$ be homotopic augmentations in $\mbox{Aug}(\ensuremath{D})$ and let $H$ be a chain homotopy from $\ensuremath{\epsilon}$ to $\ensuremath{\epsilon}'$. An \textbf{$(\ensuremath{\epsilon},\ensuremath{\epsilon}',H)$-half-disk} is a mapping of the two-disk $D^2$ into the $xz$-plane as in Definition \ref{defn:admissible} except for the following variations along the boundary:
\begin{enumerate}
\item The arc $\gamma$ maps to a vertical line $\{x_0\} \times [i,j]$ where $\ensuremath{\mu}(i)=\ensuremath{\mu}(j)$; see Figure~\ref{f:admissible} (f). We say the $(\ensuremath{\epsilon},\ensuremath{\epsilon}',H)$-half-disk \textbf{originates at $\{x_0\} \times [i,j]$};
\item For exactly one $1 \leq j \leq m$, $f(b_j)$ is a degree $-1$ crossing $q_{k_j}$, $H(q_{k_j})=1$ holds, and $f$ has a convex corner at $f(b_j)$; see Figure~\ref{f:admissible} (d) or (e);
\item If $1 \leq i < j$ (resp. $j < i \leq m$), $f(b_i)$ is a degree $0$ crossing augmented by $\ensuremath{\epsilon}$ (resp. $\ensuremath{\epsilon}'$) and $f$ has a convex corner at $f(b_i)$;
\item The restriction of $f$ to $\partial D^2$ is smooth except at $\{b_0, \hdots, b_{m+1}\}$ as described in (1) and (2) and at points in $\ensuremath{\partial} D^2 \setminus (\{b_0, \hdots, b_{m+1}\})$ where the image of $f$ looks like Figure~\ref{f:admissible} (b) or (c).
\end{enumerate}
The set $\mathcal{H}(x_0, [i, j])$ consists of all $(\ensuremath{\epsilon},\ensuremath{\epsilon}',H)$-half-disks originating at $\{x_0\}\times[i,j]$ up to reparametrization, and $\# \mathcal{H}(x_0, [i, j])$ is the mod 2 count of elements in $\mathcal{H}(x_0, [i, j])$.
\end{definition}
\begin{definition}
\label{defn:aug-half-disk}
Let $\ensuremath{\epsilon}$ be an augmentation in $\mbox{Aug}(\ensuremath{D})$. An \textbf{$\ensuremath{\epsilon}$-half-disk} is a mapping $f$ of the two-disk $D^2$ into the $xz$-plane as in Definition \ref{defn:half-disk} except that conditions (2) and (3) are replaced with the requirement that all convex corners are at crossings that are augmented by $\ensuremath{\epsilon}$.
The set $\mathcal{G}^{\ensuremath{\epsilon}}(x_0, [i, j])$ consists of all $\ensuremath{\epsilon}$-half-disks originating at $\{x_0\}\times[i,j]$ up to reparametrization, and $\# \mathcal{G}^{\ensuremath{\epsilon}}(x_0, [i, j])$ is the mod 2 count of elements in $\mathcal{G}^{\ensuremath{\epsilon}}(x_0, [i, j])$.
\end{definition}
As in Definition~\ref{defn:admissible}, the points in the vertical line $\{x_0\} \times [i,j]$ are the right-most points of either an $(\ensuremath{\epsilon}, \ensuremath{\epsilon}', H)$-half-disk or an $\ensuremath{\epsilon}$-half-disk.
By Corollary 6.21 of \cite{Henry2011}, the map $\Phi: \mbox{MCS}_A(\ensuremath{D}) \to \mbox{Aug}(\ensuremath{D})$ defined as follows is a bijection. Given $\ensuremath{\mathcal{C}} \in \mbox{MCS}_A(\ensuremath{D})$ and a generator $q$ of $\ensuremath{\mathcal{A}}(\ensuremath{D})$, $\Phi(\ensuremath{\mathcal{C}})(q) = 1$ holds if and only if $q$ is a marked crossing of $\ensuremath{\mathcal{C}}$. We let $\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}$ be the augmentation $\Phi(\ensuremath{\mathcal{C}})$.
Lemma~\ref{l:aug-half-disks} below generalizes Lemma 7.10 in \cite{Henry2013} and Lemma 5.4 in \cite{Henry2014} by removing the assumption that the front diagram $D$ is nearly plat.
Note that ``gradient paths'' from Lemma 7.10 in \cite{Henry2013} correspond to $\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}$-half-disks in our terminology, and that Lemma 5.4 of \cite{Henry2014} allows more general coefficients.
\begin{lemma}
\label{l:aug-half-disks}
Suppose $\ensuremath{D}$ is a $\sigma$-generic front diagram and $\ensuremath{\mathcal{C}} = ( \{C_m, d_m\}, \{x_m\}, H)$ is in $\mbox{MCS}_A(\ensuremath{D})$. Suppose $p \in \{1, \hdots, M\}$ and $x_p$ is to the immediate right of a crossing or cusp. Then, for all $i < j$,
\begin{equation} \label{eq:mathcalG}
\langle d_p e_i, e_j \rangle= \#\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [i, j]) \mbox{ holds.}
\end{equation}
\end{lemma}
\begin{proof}
We induct on $p$. The base case, $p=1$, follows since there is a unique disk in $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_1, [1, 2])$, as in Figure \ref{f:admissible} (b), while $\langle d_1 e_1, e_2 \rangle=1$ holds according to (4) of Definition \ref{defn:MCS}.
Assume now that $x_p$ sits to the immediate right of a crossing or cusp and that the result is known for smaller values of $p$. We complete the inductive step by considering cases.
{\bf Left cusp.} Suppose $x_p$ is to the right of a left cusp with the two strands that meet at the cusp labeled $k$ and $k+1$ at $x_p$. Define $\tau: \{1,\ldots, s_{p-1}\} \rightarrow \{1, \ldots, s_p\}$ by $\tau(i) =\left\{ \begin{array}{cr} i, & \mbox{if $i<k$} \\ i+2 & \mbox{if $i\geq k$.} \end{array} \right.$ (Note that $s_{p-1} = s_p -2$.) For any $1\leq i' < j' \leq s_{p-1}$ there is a bijection between $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [i', j'])$ and $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [\tau(i'), \tau(j')])$; see, for example, Figure~\ref{f:leftcuspcase}. Moreover, (5)(c) of Definition \ref{defn:MCS} together with the requirement that $\ensuremath{\mathcal{C}}$ has simple left cusps give
\[
\langle d_{p-1} e_{i'}, e_{j'} \rangle = \langle d_{p} e_{\tau(i')}, e_{\tau(j')} \rangle,
\]
so (\ref{eq:mathcalG}) follows when $i= \tau(i')$ and $j=\tau(j')$.
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {(a)} [tl] at 76 25
\pinlabel {(b)} [tl] at 280 25
\endlabellist
\centering
\includegraphics[scale=.9]{./images/leftcuspTheorem}
\caption{Possible extensions of an $\ensuremath{\epsilon}$-half-disk or $(\ensuremath{\epsilon}, \ensuremath{\epsilon}', H)$-half-disk past a left cusp.}
\label{f:leftcuspcase}
\end{figure}
It remains to consider those cases where $\{i,j\} \cap \{k,k+1\} \neq \emptyset$. Suppose that precisely one of $i$ or $j$ belongs to $\{k,k+1\}$. As $\ensuremath{\mathcal{C}}$ has simple left cusps, we have $\langle d_{p} e_{i}, e_{j} \rangle=0$. In addition, the restriction on the behavior of an $\ensuremath{\epsilon}$-half disk near a left cusp from Figure \ref{f:admissible} (b) gives that $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [i, j])= \emptyset$, so (\ref{eq:mathcalG}) holds. Finally, when $i=k$ and $j=k+1$, there is a unique $\ensuremath{\epsilon}$-half disk in $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [k, k+1])$. (This disk has no convex corners, so (2) of Definition \ref{defn:aug-half-disk} is vacuously satisfied.) Therefore, (\ref{eq:mathcalG}) follows in view of (4) from Definition \ref{defn:MCS}.
{\bf Crossing.} When $x_p$ sits immediately to the right of a crossing, the inductive step is achieved precisely as in Lemma 7.10 of \cite{Henry2013} or Lemma 5.4 of \cite{Henry2014} (the signs in the latter reference may be ignored). The arguments in these references apply regardless of whether or not the crossing is marked.
{\bf Right cusp.} Suppose a right cusp sits between $x_p$ and $x_{p-1}$ with the strands that meet at the cusp labeled $k$ and $k+1$ at $x_{p-1}$. Let $a_{i,j}$ be $\langle d_{p-1} e_i, e_j\rangle$. In the quotient of $(C_{p-1},d_{p-1})$ by the subcomplex spanned by $e_k$ and $d_{p-1} e_k$, we have
\[
0 = [d_{p-1}e_{k}] = [e_{k+1}] + \sum_{k+1 < j} a_{k,j} [e_j],
\]
so
\[
d_{p-1} [e_i] = \sum_{i < j} a_{i,j} [e_j] = \sum_{i< j< k} a_{i,j} [e_j] + \sum_{k+1 < j} (a_{i,j} + a_{i,k+1}\cdot a_{k,j}) [e_j].
\]
Using Definition \ref{defn:MCS}~(5)~(b), this gives the computation of the differential in $(C_p,d_p)$ as
\begin{equation} \label{eq:RCuspCompd}
\langle d_p e_{i}, e_{j}\rangle = \langle d_{p-1}e_{\pi(i)}, e_{\pi(j)} \rangle + \langle d_{p-1} e_{\pi(i)}, e_{k+1} \rangle\cdot \langle d_{p-1} e_{k}, e_{\pi(j)}\rangle,
\end{equation}
where $\pi: \{1,\ldots, s_{p}\} \rightarrow \{1, \ldots, s_{p-1}\}$ is defined by $\pi(i) =\left\{ \begin{array}{cr} i, & \mbox{if $i<k$} \\ i+2 & \mbox{if $i\geq k$.} \end{array} \right.$
We note that the second term on the right can be non-zero only if $i<k \leq j$; see Figure \ref{f:epsilonLemma}.
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {$x_{p-1}$} [t] at 32 -2
\pinlabel {$x_p$} [t] at 104 -2
\pinlabel {$a_{i,j}+a_{i,k}\cdot a_{i,k+1}$} [l] at 106 52
\pinlabel {$a_{i,k+1}$} [l] at 34 64
\pinlabel {$a_{i,j}$} [r] at 14 64
\pinlabel {$a_{k,j}$} [l] at 50 44
\pinlabel {$i$} [l] at 278 72
\pinlabel {$j$} [l] at 278 16
\endlabellist
\centering
\includegraphics[scale=1.2]{./images/epsilonLemma4}
\caption{(left) The relation between differentials at a right cusp. A dotted arrow at $x_l$ pointing from strand $i$ to strand $j$ indicates the matrix coefficient $\langle d_l e_i, e_j \rangle$. (right) The appearance of disks in $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [i, j])$ with a boundary point at the right cusp between $x_{p-1}$ and $x_p$.}
\label{f:epsilonLemma}
\end{figure}
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {(a)} [tl] at 66 100
\pinlabel {(b)} [tl] at 254 100
\pinlabel {(c)} [tl] at 65 24
\endlabellist
\centering
\includegraphics[scale=.9]{./images/rightcuspTheorem}
\caption{Possible extensions of an $\ensuremath{\epsilon}$-half-disk or $(\ensuremath{\epsilon}, \ensuremath{\epsilon}', H)$-half-disk past a right cusp.}
\label{f:cuspcase}
\end{figure}
To complete the proof, we combine Equation (\ref{eq:RCuspCompd}) with the observation that $\ensuremath{\epsilon}$-half disks satisfy a bijection
\begin{align*}
\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [i, j]) &\cong \mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [\pi(i), \pi(j)]) \\ \nonumber
&\cup \left(\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [\pi(i),k+1])\times \mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [k,\pi(j)]) \right)
\end{align*}
explained as follows: Those disks in $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [i, j])$ whose boundaries do not intersect the cusp point are in bijection with $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [\pi(i), \pi(j)])$; see Figure~\ref{f:cuspcase}. Disks in $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_p, [i, j])$ whose boundaries do intersect the cusp point appear between $x_{p-1}$ and $x_p$ as pictured in Figure \ref{f:epsilonLemma}. Removing the portion of the disk between $x_{p-1}$ and $x_p$ leaves a pair of initially overlapping disks from $\mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [\pi(i),k+1])\times \mathcal{G}^{\ensuremath{\epsilon}_{\ensuremath{\mathcal{C}}}}(x_{p-1}, [k,\pi(j)])$, and this correspondence is bijective.
\end{proof}
\subsection{Morse Complex Sequences}
We briefly sketch the connection between generating families and Morse complex sequences and refer the reader to \cite{Henry2013} for more details. A one-parameter family of smooth functions $f_x: \ensuremath{\mathbb{R}}^N \to \ensuremath{\mathbb{R}}$, parameterized by $x \in \ensuremath{\mathbb{R}}$, is a generating family for a Legendrian knot $\ensuremath{L}$ with front diagram $\ensuremath{D}$ if $$\ensuremath{D} = \left \{ (x, z): z = f_x(e) \mbox{ for some } e \in \ensuremath{\mathbb{R}}^n \mbox{ satisfying } \frac{\ensuremath{\partial} f_x}{\ensuremath{\partial} e}(e) = 0 \right \}.$$ With an appropriately chosen metric, a generic $x \in \ensuremath{\mathbb{R}}$ determines a Morse chain complex $(C_x, d_x)$ on $\ensuremath{\mathbb{R}}^N$ and, as $x$ varies, the evolution of the Morse complexes of $f_x$ are well-understood; a cusp of $\ensuremath{D}$ corresponds to the creation or elimination of a canceling pair of critical points and a crossing corresponds to two critical points exchanging critical values. As $x$ varies, it is also possible for a fiberwise gradient flowline to momentarily flow between two critical points of the same index. Such an occurrence is called a handleslide and it determines an explicit chain isomorphism between successive Morse complexes. In summary, a generating family and choice of metric determine a one-parameter family of Morse chain complexes and the relationship between successive chain complexes is determined by the crossings and cusps of $\ensuremath{D}$ and the handleslides. A Morse complex sequence on $\ensuremath{D}$ is a finite sequence of chain complexes $(C_m, d_m)$ and vertical marks on $\ensuremath{D}$ that are meant to correspond to the Morse chain complexes and handleslides of a generating family and choice of metric. In addition, varying the choice of metric motivates an equivalence relation on MCSs.
Fix a Legendrian knot $\ensuremath{L}$ with $\sigma$-generic front diagram $\ensuremath{D}$, rotation number $0$, and Maslov potential $\ensuremath{\mu}$. Theorem~\ref{t:bijection} proves that a certain surjective map in \cite{Henry2011} from equivalence classes of Morse complex sequences to $\mbox{Aug}^{ch}(\ensuremath{D})$ is, in fact, a bijection. The definition of an MCS given in \cite{Henry2011} defines an MCS independent of a fixed front diagram. It is then shown that an MCS determines a front diagram. An alternative definition of an MCS given in \cite{Henry2013} defines an MCS as an object assigned to a fixed front diagram. Both definitions determine the same set of objects on a fixed front diagram. We will use the definition of a Morse complex sequence given in \cite{Henry2013}.
A \textbf{handleslide} on $\ensuremath{D}$ is a vertical line segment disjoint from all crossings and cusps and with endpoints on strands of $\ensuremath{D}$ that have the same Maslov potential.
\begin{definition}
\label{defn:MCS}
A \textbf{Morse complex sequence} on a $\sigma$-generic front diagram $\ensuremath{D}$ is the triple $\ensuremath{\mathcal{C}} = ( \{(C_m, d_m)\}, \{x_m\}, H)$ satisfying:
\begin{enumerate}
\item $H$ is a set of handleslides on $\ensuremath{D}$.
\item The real values $x_1 < x_2 < \hdots < x_M$ are $x$-coordinates distinct from the $x$-coordinates of crossings and cusps of $\ensuremath{D}$ and handleslides of $H$. For each $1 \leq m < M$, the set $\{(x, z): x_m \leq x \leq x_{m+1}\}$ contains a single crossing, cusp or handleslide. The set $\{(x, z): -\infty < x \leq x_1\}$ contains the left-most left cusp and the set $\{(x,z): x_M \leq x < \infty\}$ contains the right-most right cusp.
\item For each $1 \leq m \leq M$, the points of intersection of the vertical line $\{x_m\} \times \ensuremath{\mathbb{R}}$ and $\ensuremath{D}$ are labeled $e_1, e_2, \hdots, e_{s_m}$ from top to bottom. The vector space $C_m$ is the $\ensuremath{\mathbb{Z}}$-graded $\ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}}$ vector space generated by $e_1, e_2, \hdots, e_{s_m}$, where the degree of each generator is the value of the Maslov potential on the corresponding strand of $\ensuremath{D}$, $|e_i| = \ensuremath{\mu}(e_i)$. The map $d_m : C_m \to C_m$ is a degree $-1$ differential that is triangular in the sense that $$d_m e_i = \sum_{i<j} c_{ij} e_j, \mbox{ } c_{ij} \in \ensuremath{\mathbb{Z}} / 2 \ensuremath{\mathbb{Z}};$$
\item The coefficients $\langle d_1 e_1, e_2 \rangle$ and $\langle d_M e_1, e_2 \rangle$ are both $1$. Suppose $1 \leq m < M$ and let $T$ be the tangle $\ensuremath{D} \cap \{ (x, z) : x_m \leq x \leq x_{m+1}\}$. If $T$ contains a left (resp. right) cusp between strands $k$ and $k+1$, then $\langle d_{m+1} e_k, e_{k+1} \rangle $ is $1$ (resp. $\langle d_m e_k, e_{k+1} \rangle$ is $1$). If $T$ contains a crossing between strands $k$ and $k+1$, then $\langle d_m e_k, e_{k+1}\rangle$ is $0$.
\item For $1 \leq m < M$, the crossing, cusp, or handleslide mark in the tangle $T = \ensuremath{D} \cap \{ (x,z) : x_m \leq x \leq x_{m+1} \}$ determines an algebraic relationship between the chain complexes $(C_m, d_m)$ and $(C_{m+1},d_{m+1})$ as follows:
\begin{enumerate}
\item \textbf{Crossing:} If the crossing is between strands $k$ and $k+1$, then the map $\phi : (C_m, d_m) \to (C_{m+1}, d_{m+1})$ defined by:
\begin{equation*}
\phi(e_i) = \left\{ \begin{array}{rl}
e_i &\mbox{ if $i \notin \{k, k+1\}$} \\
e_{k+1} &\mbox{ if $i = k$} \\
e_{k} &\mbox{ if $i = k+1$}
\end{array} \right.
\end{equation*}
is an isomorphism of chain complexes.
\item \textbf{Right cusp:} If the right cusp is between strands $k$ and $k+1$, then the linear map
\begin{equation*}
\phi ( e_i) =
\begin{cases}
[e_i] &\mbox{ if $i < k $} \\
[e_{i+2}] &\mbox{ if $i \geq k $}.
\end{cases}
\end{equation*}
is an isomorphism of chain complexes from $(C_{m+1},d_{m+1})$ to the quotient of $(C_m, d_m)$ by the acyclic subcomplex generated by $\{e_k, d_{m}e_k\}$.
\item \textbf{Left cusp:} The case of a left cusp is the same as the case of a right cusp, though the roles of $(C_{m}, d_m)$ and $(C_{m+1}, d_{m+1})$ are reversed.
\item \textbf{Handleslide:} If the handleslide mark has endpoints on strands $k$ and $l$ with $k < l$, then the map $h_{k,l}:(C_m, d_m) \to (C_{m+1}, d_{m+1})$ defined by
\begin{equation*}
h_{k,l} ( e_i) =
\begin{cases}
e_i &\mbox{ if $i \neq k $} \\
e_k + e_l &\mbox{ if $i=k$}.
\end{cases}
\end{equation*}
is an isomorphism of chain complexes.
\end{enumerate}
\end{enumerate}
\end{definition}
The set $\mbox{MCS}(\ensuremath{D})$ is the set of all Morse complex sequences on $\ensuremath{D}$.
\begin{remark}
Morse complex sequences may be defined over more general coefficient rings than $\ensuremath{\mathbb{Z}}/2$, cf. \cite{Henry2014}. We restrict attention to $\ensuremath{\mathbb{Z}}/2$ coefficients as this is also done in \cite{Henry2011}.
\end{remark}
\begin{definition}
\label{defn:simple}
An MCS $\ensuremath{\mathcal{C}} = ( \{(C_m, d_m)\}, \{x_m\}, H)$ in $\mbox{MCS}(\ensuremath{D})$ has \textbf{simple left cusps} if, for each tangle $T = \{ (x,z) : x_m \leq x \leq x_{m+1} \}$ containing a left cusp between strands $k$ and $k+1$, the chain complex $(C_{m+1}, d_{m+1})$ satisfies $\langle d_{m+1} e_k, e_i \rangle = \langle d_{m+1} e_{k+1}, e_i \rangle = 0$
for all $k+1 < i$ and $\langle d_{m+1} e_j, e_{k+1} \rangle = \langle d_{m+1} e_j, e_{k} \rangle = 0$ for all $j < k$.
\end{definition}
The subset $\mbox{MCS}_b(\ensuremath{D}) \subset \mbox{MCS}(\ensuremath{D})$ denotes the set of MCSs with simple left cusps. We use the letter $b$ to be consistent with the notation of \cite{Henry2011}, where a left cusp is also called a ``birth''. This language is meant to draw a connection to the creation of a canceling pair of critical points, often called a birth, in a one-parameter family of Morse functions on a manifold.
Given an MCS $\ensuremath{\mathcal{C}} = ( \{(C_m, d_m)\}, \{x_m\}, H)$ with simple left cusps, the chain complexes $\{(C_m, d_m)\}$ are uniquely determined by the crossings and cusps of $\ensuremath{D}$, the handleslides $H$, and requirements (5) (a)-(d) of Definition~\ref{defn:MCS}. Consequently, $\ensuremath{\mathcal{C}}$ may be represented visually by placing the handleslide marks $H$ on the front diagram $\ensuremath{D}$; see Figure~\ref{f:MCS-example}.
\begin{figure}[t]
\centering
\includegraphics[scale=.9]{./images/MCS-example}
\caption{An MCS with simple left cusps. This MCS is also in A-form.}
\label{f:MCS-example}
\end{figure}
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {(1)} [tl] at 69 403
\pinlabel {(2)} [tl] at 261 403
\pinlabel {(3)} [tl] at 69 324
\pinlabel {(4)} [tl] at 261 324
\pinlabel {(5)} [tl] at 69 251
\pinlabel {(6)} [tl] at 261 251
\pinlabel {(7)} [tl] at 69 172
\pinlabel {(8)} [tl] at 261 172
\pinlabel {(9)} [tl] at 70 101
\pinlabel {(10)} [tl] at 258 101
\pinlabel {(11)} [tl] at 65 20
\pinlabel {(12)} [tl] at 258 20
\endlabellist
\centering
\includegraphics[scale=.9]{./images/MCS-moves}
\caption{Handleslide modifications, called MCS moves, that result in an equivalent MCS.}
\label{f:moves}
\end{figure}
\begin{figure}[t]
\labellist
\small\hair 2pt
\pinlabel {$i$} [tr] at -1 89
\pinlabel {$k$} [tr] at -1 57
\pinlabel {$l$} [tr] at -1 40
\pinlabel {$j$} [tr] at -1 9
\endlabellist
\centering
\includegraphics[scale=1.2]{./images/MCS-explosion}
\caption{MCS move (13). On the left, a dotted arrow from strand $\alpha$ to strand $\beta$ indicates that $\langle d_m e_{\alpha}, e_{\beta} \rangle$ is $1$.}
\label{f:explosion}
\end{figure}
In \cite{Henry2011} an equivalence relation on the set $\mbox{MCS}(\ensuremath{D})$ is defined that is motivated by a corresponding equivalence for generating families. (See also \cite{Henry2013}.) Here we denote the set of equivalence classes of this relation by $\widehat{\mbox{MCS}}(\ensuremath{D}) = \mbox{MCS}(\ensuremath{D}) / \simeq$. We recall a version of this equivalence relation that applies to the more restricted set of MCSs with simple left cusps, $\mbox{MCS}_b(\ensuremath{D})$. We denote equivalence classes with respect to this relation by $\widehat{\mbox{MCS}}_b(\ensuremath{D})$. By Proposition 3.17 of \cite{Henry2011}, the map from $\widehat{\mbox{MCS}}_b(\ensuremath{D})$ to $\widehat{\mbox{MCS}}(\ensuremath{D})$ induced by the inclusion $\mbox{MCS}_b(\ensuremath{D}) \subset \mbox{MCS}(\ensuremath{D})$ is a bijection. Therefore, to prove Theorem~\ref{t:bijection}, we need only consider MCSs in $\mbox{MCS}_b(\ensuremath{D})$ and MCS classes in $\widehat{\mbox{MCS}}_b(\ensuremath{D})$.
The equivalence relation on $\mbox{MCS}_b(\ensuremath{D})$ is generated by the \textbf{MCS moves} pictured in Figures~\ref{f:moves} and \ref{f:explosion}. The numbering indicated will be used throughout this article.
Additional moves result from reflecting each of the two figures in (3),(7),(9),(10), and (12) of Figure~\ref{f:moves} about a horizontal axis and reflecting each of the two figures in (4), (9), (11), and (12) of Figure~\ref{f:moves} about a vertical axis. The handleslide modification that results from reflecting Figure~\ref{f:moves} (10) about a vertical axis is \emph{not} an MCS move for MCSs with simple left cusps. (The absence of this reflected move is the only difference between the definitions of the equivalence relations on $\mbox{MCS}_b(\ensuremath{D})$ and $\mbox{MCS}(\ensuremath{D})$ discussed in the previous paragraph.) MCS move (13) requires explanation. Suppose $\ensuremath{\mathcal{C}} = ( \{(C_m, d_m)\}, \{x_m\}, H)$ is an MCS on $\ensuremath{D}$ and suppose there exists $x_m$ and $1 \leq k < l \leq s_m$ so that $\ensuremath{\mu}(e_k) = \ensuremath{\mu}(e_{l})-1$, then MCS move (13) introduces the collection of handleslides $K$ defined as follows. The handleslides in $K$ are of two types. First, if $i < k$ and $\langle d_m e_i, e_k \rangle=1$ holds, then $K$ contains a handleslide with endpoints on $i$ and $l$. Second, if $l<j$ and $\langle d_m e_{l}, e_j \rangle=1$ holds, then $K$ contains a handleslide with endpoints on $k$ and $j$.
By Proposition 3.8 of \cite{Henry2011}, modifying the handleslide set of an MCS in $\mbox{MCS}_b(\ensuremath{D})$ as in one of the cases in Figures~\ref{f:moves} and \ref{f:explosion} results in another MCS in $\mbox{MCS}_b(\ensuremath{D})$. Therefore, the notion of equivalence in the following definition is well-defined. In addition, if an MCS move is applied to an MCS, then only those chain complexes near the location of the MCS move are affected. In other words, the MCS moves are local in the sense that they change both the handleslides and chain complexes of an MCS only in a local neighborhood.
\begin{definition}
\label{defn:equiv}
Two MCSs $\ensuremath{\mathcal{C}}$ and $\ensuremath{\mathcal{C}}'$ in $\mbox{MCS}_b(\ensuremath{D})$ are \textbf{equivalent}, written $\ensuremath{\mathcal{C}} \simeq \ensuremath{\mathcal{C}}'$, if there exists a sequence $\ensuremath{\mathcal{C}}_1, \ensuremath{\mathcal{C}}_2, \hdots, \ensuremath{\mathcal{C}}_s$ in $\mbox{MCS}_b(\ensuremath{D})$ so that $\ensuremath{\mathcal{C}} = \ensuremath{\mathcal{C}}_1$, $\ensuremath{\mathcal{C}}'=\ensuremath{\mathcal{C}}_s$, and, for all $1 \leq i < s$, the set of handleslide marks of $\ensuremath{\mathcal{C}}_i$ and $\ensuremath{\mathcal{C}}_{i+1}$ differ by exactly one MCS move. The set $\widehat{\mbox{MCS}}_b(\ensuremath{D})$ is the set $\mbox{MCS}_b(\ensuremath{D}) / \simeq$.
\end{definition}
MCSs of the following type have a standard form that makes their relationship with augmentations particularly simple to describe.
\begin{definition}
\label{defn:A-form}
An MCS $\ensuremath{\mathcal{C}}$ in $\mbox{MCS}_b(\ensuremath{D})$ is in \textbf{A-form} if there exists a set $R$ of degree $0$ crossings so that just to the left of each $q$ in $R$ there is a handleslide with endpoints on the strands crossing at $q$ and $\ensuremath{\mathcal{C}}$ has no other handleslides. A crossing $q$ in $R$ is said to be \textbf{marked}.
Figure~\ref{f:MCS-example} shows an MCS in A-form where $R$ is the four left-most crossings. The subset $\mbox{MCS}_A(\ensuremath{D}) \subset \mbox{MCS}_b(\ensuremath{D})$ consists of all A-form MCSs on $\ensuremath{D}$.
\end{definition}
|
\section{Introduction}
Positron Emission Tomography is currently one of the best suited medical examination
methods for tumor detection.
Nowadays commercial PET scanners are made of crystal scintillators arranged in
a ring surrounding the patient~\cite{Conti2009,Humm2003,Karp2008,Townsend2004}.
New generation of PET scanners for the image reconstruction not only utilizes information about the hit position
of gamma quanta in the detectors but also takes advantage of the measurement of the time differences (TOF)
between the interactions of annihilation quanta in the detectors~\cite{Conti2011}.
This improves image reconstruction by increasing the signal to background ratio~\cite{Conti2009,Karp2008,Moses2003}.
A typical TOF resolution of presently used TOF-PET detectors amounts to about 500~ps~(FHWM)~\cite{Conti2011},
and there is a continuous endeavor to improve it (see e.g. results for small size
crystals~\cite{ContiErikson2009,Schaart2010,Schaart2009,Moses2008,Kuhn2006,MoszynskiSzczesniak2011}).
The Jagiellonian-PET (J-PET) collaboration aims at the construction of the TOF-PET
scanner with a large field of view (up to about 1 m)
and a superior TOF resolution by application of fast plastic scintillators instead of
organic crystals. The detector will be built out from strips of plastic scintillators forming
a diagnostic chamber~\cite{PCT2010,JPET-Genewa,NovelDetectorSystems,StripPETconcept,TOFPETDetector}
as shown schematically in the left panel of Fig.~\ref{uklad}.
\begin{figure}[h!]
\includegraphics[width=0.45\textwidth]{beczka2.eps} \hspace{-0.55cm} \includegraphics[width=0.58\textwidth]{uklad.eps}\\
\caption{(Left) Schematic view of the single layer of the J-PET scanner.
(Right) General scheme of the experimental setup used to test performance
of a single detection module. Radioactive source is held in the lead collimator.
\textit{Abbreviations:} PM and SDA denote photomultiplier and Serial Data Analyzer (Lecroy SDA6000A), respectively.
}
\label{uklad}
\end{figure}
A single detector module of the J-PET detector consists of a plastic scintillator strip
read out by photomultipliers at both ends. Plastic scintillators are less efficient for the detection of gamma quanta
than crystals but they posses better timing properties and allow to build large acceptance detectors in a cost efficient way.
Therefore, a PET scanner based on plastic scintillators constitutes a promising solution in view of the
TOF resolution and construction of the scanner allowing for simultaneous imaging of the whole human body.
Development of a cost-effective whole body PET scanner is
a technological challenge and there are various non-standard techniques being tested such as detectors based on straw tubes drift chambers~\cite{Lacy2001,Shehad2005}
or large area resistive plate chambers~\cite{Belli2006,Blanco2009}.
The J-PET detector with plastic scintillators arranged axially as shown in Fig.~\ref{uklad}
possesses also another advantage. Its diagnostic chamber is free of any electronic devices and magnetic materials,
thus giving unique possibility for simultaneous imaging of PET and MRI as well as PET and CT
in a way different from so far developed
configurations~\cite{Crosetto2003,Gilbert2006,Judenhofer2008,Marsden2002,Pichler2008,Quick2011,Townsend2008}.
A similar axial arrangement with crystal scintillators of the length of 10~cm
is being developed by the AX-PET collaboration aiming at improvement of resolution and sensitivity~\cite{AX-PET}.
Detectors based on plastic scintillators are commonly used in nuclear and particle physics experiments,
however, due to negligible probability of photo-electric effect,
their potential for registration of low energy gamma quanta (in the range of fraction of MeV)
was so far not explored except for few publications concentrated
on the light propagation studies~e.g.~\cite{NIM2008Moszynski},
or callibration methods~\cite{NIM2008,NIM2011,NIMKudomi}.
In this paper we show that plastic scintillators can be used for building
large area detection systems with very good time, position and energy resolution for the registration of low energy gamma quanta.
In Section 2 a comprehensive description of experimental setup used for investigations is presented.
Next, sections 3, 4 and 5 include description of methods and results for the determination of the energy, time and position resolution, respectively.
\section{Experimental setup}
General scheme of the experimental setup used for tests of a single module is presented in the right panel of Fig.~\ref{uklad}.
A prototype module consists of a BC-420~\cite{SaintGobain} scintillator strip with dimensions of
5~mm~x~19~mm~x~300~mm and of two Hamamatsu photomultipliers R5320~\cite{Hamamatsu} connected optically
to the most distant ends of the scintillator strip via optical gel EJ-550.
The module is tested with annihilation quanta from the $^{68}$Ge source
placed inside a lead collimator which can be moved along the scintillator
by means of a dedicated mechanical construction. A collimated beam emerging through 1.5~mm wide and 20~cm
long slit is used for irradiating desired points across the strip. A coincident registration
of signals from the tested module and a reference detector allows for a suppression of signals from other than annihilation quanta
to the negligible level. The reference detector consists of a scintillator strip with
thickness of 4~mm connected via light guide to the photomultiplier. The reference detector is fixed
to the collimator by means of an aluminum arm allowing to keep the relative setting between the collimator and
the reference detector unchanged while moving the collimator along the tested scintillator strip.
In this way the same collimating properties are ensured at every position of irradiation.
Signals from photomultipliers are probed with 50~ps intervals by means of Serial Data Analyzer
(LeCroy SDA6000A).
Exemplary sampled signals from the middle of the scintillator are shown in Fig.~\ref{shapes}.
For the full J-PET detector a dedicated electronics~\cite{palka,korcyl}
and analysis framework~\cite{krzemien,czerwinski}
for data collecting and processing is being developed.
\begin{figure}[h!]
\centerline{\includegraphics[width=0.7\textwidth]{signal.eps}}
\caption
Example of signals acquired from left and right photomultiplier
when irradiating the center of the scintillator strip.
$t_L$ and $t_R$ denote times at which left and right signal, respectively,
cross the reference voltage indicated by dashed green horizontal line.
For better visibility the signals were separated from each other by 19~ns.
(For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)
\label{shapes}
}
\end{figure}
In the J-PET detector a hit position along the scintillator strip,
as well as an annihilation point along the line-of-response, will be reconstructed based on the measurement
of times of light signals arrival to photomultipliers. Therefore, the detector needs to be optimized
for the best timing properties.
Moreover, for building a device with a large field of view,
a weak light attenuation in the scintillator material is mandatory.
These requirements led us to a choice
of the BC-420~\cite{SaintGobain}
(equivalent of EJ-230~\cite{Eljen})
plastic scintillator
as the most suitable among the currently available ones.
The rise time and bulk attenuation length of light signals in this scintillator
amount to 0.5~ns and 110~cm~\cite{SaintGobain}, respectively.
For the light detection, the Hamamatsu
R5320
photomultipliers~\cite{Hamamatsu} were chosen
with the rise time and the transit time spread of 0.7~ns and 0.16~ns, respectively~\cite{Hamamatsu}.
The rise time of signals shown in Fig.~\ref{shapes} is equal to about 1~ns,
as expected from 0.5 ns rise time of a light pulse in the BC-420 scintillator
convoluted with 0.7 ns rise time of signals from the Hamamatsu R5320 photomultiplier.
The decay time for the BC-420 scintillator amounts to 1.5~ns~\cite{SaintGobain}.
In a good approximation~\cite{Moszynski} an observed signal is a convolution
of the Gaussian and exponential functions and of the single photoelectron response of the photomultiplier.
The single photoelectron signals were measured using a method described in Ref.~\cite{CalibrationPMT},
and values of a rise time of 0.7~ns and FWHM of 1.5~ns were observed
in agreement with the values given in catalog~\cite{Hamamatsu}.
Moreover, we have checked that the observed signals shown in Fig.~\ref{shapes}
are consistent with the expectation for the decay time of 1.5~ns.
The studies presented in this paper were conducted for the scintillator strip wrapped with the tyvek foil.
For further details about
properties of the used photomultipliers and scintillators in view of the construction
of the J-PET detector the interested reader is referred to~\cite{CalibrationPMT,lukasz}.
\section{Energy resolution}
Energy resolution depends predominantly on the number of photoelectrons released from photocathodes of both photomultipliers.
The larger this number the better is the energy resolution due to decrease of the statistical fluctuation of the number of signal carriers.
Therefore, for the consideration of the energy resolution it is natural to express the energy deposition
in terms of the number of photoelectrons and to use an arithmetic mean as a measure of the deposited energy:
\begin{equation}
E_{deposited} = \alpha \frac{(N_L + N_R)}{2},
\end{equation}
where $\alpha$, $N_L$ and $N_R$
denote an energy calibration factor, and
number of photoelectrons registered
at left and right side of the scintillator, respectively.
For scintillator detectors the Fano factor is equal to one and therefore,
in case of uncorrelated errors of $N_L$ and $N_R$ the fractional energy resolution would read:
\begin{equation}
\frac{\sigma(E_{deposited})}{E_{deposited}} = \frac{1}{\sqrt{N_L+N_R}} = \frac{\sqrt{\alpha/2}}{\sqrt{E_{deposited}}}.
\label{beta1}
\end{equation}
Consequently, the energy resolution as a function of the deposited energy may be approximately parametrized as:
\begin{equation}
\frac {\sigma (E_{deposited}) }{E_{deposited}} = \frac {\beta }{\sqrt{E_{deposited}}},
\label{beta}
\end{equation}
where $\beta$ is an effective coefficient which in general may differ from $\sqrt {\alpha/2}$.
\begin{figure}[h!]
\centerline{\includegraphics[width=0.6\textwidth]{138Aall.eps}}
\caption{Distribution of arithmetic mean of the number of photoelectrons produced at photocathodes
of left and right photomultipliers. The spectrum was obtained by irradiating the center of the scintillator strip with the collimated
beam of annihilation quanta.
As indicated in the legend, solid and dashed lines indicate experimental and simulated spectra, respectively.
More detailed explanation is given in the text. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)
\label{GandA}}
\end{figure}
The value of $\beta$ for the tested detector
was obtained by comparing experimental distribution of $(N_L+N_R)/2$ (Fig.~\ref{GandA})
with the simulated histogram of deposited energy where $\alpha$,
$\beta$, and normalization constant $A$ were treated as free parameters.
A fit was conducted with a Neyman $\chi^2$ statistics defined as follows:
\begin{equation}
\chi^2(\alpha,\beta,A) = \sum_{i}\frac{(A \cdot N_{sim}(i\cdot\alpha,\beta) - N_{exp}(i))^2}{N_{exp}(i)},
\end{equation}
where $i$ denotes the $i^{th}$ bin of the histogram $N_{exp}$.
The simulated distribution of energy deposition of the annihilation quanta $N_{sim}(E,\beta)$
was obtained based on the Klein-Nishina formula~\cite{Klein} convoluted with the detector resolution
parametrized by Eq. \ref{beta}. Due to the large number of generated events
the statistical uncertainties of simulated distributions are negligible.
The fit was performed in the range from 90 to 150 photoelectrons.
The lower range of the spectrum was not taken into account since it is enhanced
by events with the scattering very close to the scintillator surface and
by signals originating from gamma quanta scattered
in the collimator and in material surrounding the detector. The best fit was obtained for $\beta = 0.044$.
Dashed and dotted lines in Fig.~\ref{GandA} show simulated energy loss spectra for the ideal detector and the detector
with the fractional energy resolution of $\sigma (E)/E = 0.044 / \sqrt{E(MeV)}$, as obtained from the fit.
It is also worth to stress that the determined value of $\beta$
is fairly close to the result expected for the fully uncorrelated
errors of $N_L$ and $N_R$ which (compare Eq.~\ref{beta1} and~\ref{beta})
gives
$\beta = \sqrt{\alpha/2} \approx \sqrt{0.341/138/2} \approx 0.035$,
where $\alpha$
is estimated knowing that $E_{deposited}$~=~0.341~MeV at the Compton edge
and that the corresponding mean value of photoelectrons
from left and right photomultipliers amounts to about 138 (see Fig.~\ref{GandA}).
The number of photoelectrons for each measured signal was determined based on the known average charge of signals
induced by single photons determined using the method described in Ref.~\cite{CalibrationPMT}.
To calculate the number of photoelectrons the charge of each measured signal was divided by the average charge of the single photoelectron signal.
In order to measure a charge spectrum originating from the single photoelectron
we have inserted between the tested photomultiplier and the scintillator an aperture with a hole with a diameter of 0.6~mm.
The ratio of the area of the hole (0.28~mm$^2$) and the side of the scintillator (95~mm$^2$) was equal to about 340,
thus providing that for the typical event out of about 300 photons reaching the edge of the scintillator
only zero, one or very rarely two photoelectrons were released from the photocathode.
The experimental spectrum shown in Fig.~\ref{GandA} is suppressed in the range below 20 photoelectrons due to the triggering conditions.
The superimposed red dashed line indicates the distribution simulated based on the Klein-Nishina formula~\cite{Klein}
convoluted with the detector response with a resolution of $\sigma(E)/E$ as indicated in the figure. The dotted line denotes the theoretical distribution
of the energy of electrons scattered via the Compton effect
by the gamma quantum with an energy of 511 keV.
The observed number of photoelectrons is consistent with
rough estimations of about 149 photoelectrons at the Compton edge (0.341~MeV)
which can be derived
taking into account that (i) the light output of BC-420 scintillator
equals to about 10,000 photons per MeV~\cite{SaintGobain} for electrons,
(ii) the fraction of light which can be conducted via internal reflections to the edge
in the rectangular strip surrounded by air is equal to about $\sqrt{1 - (1/n)^2} - \frac{1}{2}$~=~0.27
(with refractive index n~=~1.58~\cite{SaintGobain}), (iii)
the quantum effciency of Hamamatsu R5320 photomultipliers
is equal to about 0.2~\cite{Hamamatsu}
at 400 nm,
and (iv) the bulk light attenuation length is equal to about 110~cm~\cite{SaintGobain}
where on the average the light travels about 23~cm
from the center to the edge ($e^{-0.21}$~=~0.81).
In case of the reconstruction of the tomographic image it was estimated that only signals with $E_{deposited} > 0.2$~MeV
will be used in order to decrease the noise caused by the scattering
of the annihilation quanta in the patient's body~\cite{TOFPETDetector}.
In the energy range from 0.2~MeV to 0.34~MeV the value of $\beta~=~0.044$
gives fractional energy resolution of $\sigma(E)/E$ ranging from about 10\% to 7.5\%, respectively.
In the discussed case energy deposition of 0.2~MeV corresponds to about 81 photoelectrons.
Yet, in the further analysis, for conservative estimation of the time and position resolution we have selected
signals with at least 75 photoelectrons.
\section{Time resolution}
Time resolution is determined based on the distribution of time differences measured at a fixed point of irradiation.
As an example, Fig.~\ref{twoOnOne} presents time difference distributions ($\Delta t = t_R - t_L$) measured by irradiating the scintillator
strip close to the left photomultiplier (x~=~1.2~cm, right maximum) and at the position close to the right photomultiplier
(x~=~28.8~cm, left maximum). The times of pulses on both sides $t_L$ and $t_R$ were determined calculating
the time when the signal crosses a given threshold voltage, as it is illustrated in Fig.~\ref{shapes}.
The left panel of Fig.~\ref{thrAndfrac} presents the resolution of the time difference measurement as a function of
the irradiation position.
\begin{figure}[b!]
\centerline{\includegraphics[width=0.49\textwidth]{both.eps}}
\caption{Distributions of time difference $\Delta t = (t_R - t_L)$
for two positions as indicated in the figure and discribed in the text.
\label{twoOnOne}
}
\end{figure}
As expected, due to the “time walk” effect the resolution determined when applying threshold at 250~mV is worse with respect
to the one obtained at 50~mV. A value of 50~mV was chosen as a 2.5~$\sigma$
of a typical
electronic
noise level
equal to 20~mV~($\sigma$).
Right panel of Fig.~\ref{thrAndfrac} shows results obtained when determining the time at the constant fraction of the amplitude.
From Fig.~\ref{thrAndfrac} one can infer that the time resolution
is fairly independent of the irradiation position if time is determined
for constant fraction of the amplitude as well as for low threshold
at a constant level (50~mV). It is also visible that in the case of
the larger threshold (250~mV) resolutions become significantly worse at the edges of the scintillator strip which again is due to the
“time walk” effect. The resolution of the time difference ($\Delta t = t_R - t_L$) measurement at the center of the scintillator was determined
to be $\sigma(\Delta t)~=~(153~\pm~2)$~ps both for constant level discrimination
at 50~mV and constant fraction threshold of 10\% of the amplitude.
\begin{figure}[t!]
\begin{tabular}{c c}
\includegraphics[width=0.49\textwidth]{thr.eps} & \includegraphics[width=0.49\textwidth]{frac.eps}\\
\end{tabular}
\caption{ Left: Time resolution ($\sigma(\Delta t)$) as a function of irradiated position for constant level discrimination
at thresholds of 50 mV (squares) and 250 mV (triangles).
Right: Time resolution ($\sigma(\Delta t)$) as a function of irradiated position for times calculated at constant fraction of the amplitude
for fractions of 10\% (squares) and 50\% (triangles).
\label{thrAndfrac}
}
\end{figure}
It is important to stress that results shown in Figs.~\ref{twoOnOne} and~\ref{thrAndfrac}
were obtained taking into account only signals with the number of photoelectrons larger than 75.
This corresponds to the resolution of about 77~ps ($\sigma$)
for the determination of the interaction moment ($t_{hit}$) which may be expressed
as the average of times measured at the left and right photomultipliers independently of the hit position:
\begin{equation}
\frac {(t_R + t_L)}{2} = \frac {\left(t_{hit} + \frac {L-x}{v_{eff}} + t_{hit} + \frac {x}{v_{eff}} \right) } {2} = t_{hit} + \frac{L}{2 v_{eff}},
\end{equation}
where $v_{eff}$ denotes the effective velocity of the light signal inside the scintillator,
and L and x were defined in Fig.~\ref{uklad}. The constant time delays of electronics
were omitted in the above equation for the sake of simplicity.
Thus the uncertainty of the measurement of $t_{hit}$ may be expressed as:
\begin{equation}
\sigma (t_{hit}) = \frac {\sqrt { \sigma (t_L) ^2 + \sigma (t_R) ^ 2 }} {2} = \frac{\sigma (\Delta t)}{2}.
\end{equation}
\section{Spatial resolution}
In the first approximation, a hit position along the scintillator strip may be determined based
on the time difference of light signals arrival to the left and right photomultipliers using
the following formula:
\begin{equation}
x = \frac{\Delta t \cdot v_{eff}}{2},
\end{equation}
which may be derived from the relation:
\begin{equation}
\Delta t = (t_R - t_L) = t_{hit} + \frac{L-x}{v_{eff}} - t_{hit} - \frac{x}{v_{eff}} = \frac {-2x}{v_{eff}} + C.
\end{equation}
Thus the spatial resolution reads:
\begin{equation}
\sigma(x) = \sigma(\Delta t) \frac{v_{eff}}{2}.
\label{spatial}
\end{equation}
The effective speed of light signals along a scintillator strip ($v_{eff}$)
is smaller than the speed of light in a scintillator medium because most of photons do not travel
to the photomultipliers directly but rather undergoes many internal reflections.
In order to determine the effective speed of light signals in the
tested scintillator the time difference $\Delta t$ was determined
as a function of the irradiation position x, and $v_{eff}$ was extracted by fitting a straight line to the experimental points.
The determined value of $v_{eff}$ is shown in Fig.~\ref{lineAndCompar}
as a function of the applied threshold.
\begin{figure}[h!]
\centerline{\includegraphics[width=0.49\textwidth]{compar.eps}}
\caption{
Effective speed of light inside scintillator strip as a function of the applied threshold.
Superimposed line represents result of the fit of the second order polynomial to the data.
\label{lineAndCompar}
}
\end{figure}
The change of $v_{eff}$ with threshold is due to the “walk effect”
and the variation of the average amplitude of signals as a function of the distance
between the interaction point and photomultipliers.
In order to suppress the bias of the determined $v_{eff}$ due to the value of the applied threshold,
the effective speed of light signals was determined by
fitting the second order polynomial to the data points and
extrapolating the fitted curve to the threshold of 0 mV, as shown in Fig.~\ref{lineAndCompar}.
The systematic uncertainty due to the extrapolation method was estimated as a difference in results between
the fit with second and first order polynomials, and it was found to be negligible.
The resulting effective speed of light is equal to $v_{eff}~=~(12.61~\pm~0.05_{stat}~\pm~0.01_{sys})$~$\frac{cm}{ns}$.
The determined velocity is in the range of values obtained so far for signals in the plastic
scintillator bars (see e.g.~\cite{charpak,kouznetsov,shikaze}).
For the estimation of the position resolution
we apply in equation~\ref{spatial} the value of $v_{eff}$~=~12.2~cm/ns
and the value of $\sigma(\Delta t)$~=~153~ps both determined for the threshold of 50~mV.
As a result a spatial resolution of $\sigma(x)~=~0.93~cm$ is established
for the determination of the interaction point of the annihilation quanta along the strip.
\section{Conclusions}
Properties of a single plastic scintillator module of the J-PET detector were investigated in view of the detection of annihilation
gamma quanta with energy of 511 keV. The module was built out of BC-420 scintillator strip with dimensions
of 5~mm x 19~mm x 300~mm which was read out at both sides by Hamamatsu R5320 photomultipliers.
The measurements were performed using a collimated beam of annihilation quanta
from the $^{68}$Ge isotope and Serial Data Analyzer
for sampling of photomultiplier’s signals with 50~ps intervals.
The determined energy resolution amounts to $\sigma(E)/E \approx 0.044 / \sqrt{E(MeV)}$.
For the energy deposition ranging from 0.18~MeV to 0.34~MeV
the established time resolution
is equal to about 80~ps~($\sigma$) and the hit position resolution
along the scintillator strip equals to 0.93~cm~($\sigma$).
The achieved results are promising and as a next step the test will be conducted with
a dedicated front-end electronics which will allow
to sample the signals in the domain of voltage (using multi-threshold sampling) with
the electronic time resolution below 20~ps~\cite{palka}.
\section{Acknowledgements}
We acknowledge technical and administrative support of T.~Gucwa-Ry\'s, A.~Heczko, M.~Kajetanowicz, G.~Konopka-Cupia\l,
J.~Majewski, W.~Migda\l, A.~Misiak, and the financial support from the Polish National Center for Development
and Research through grant INNOTECH-K1/IN1/64/159174/NCBR/12,
the Foundation for Polish Science through MPD programme,
the EU and MSHE Grant No. POIG.02.03.00-161 00-013/09,
Doctus - the Lesser Poland PhD Scholarship Fund,
and Marian Smoluchowski Krak\'ow Research Consortium "Matter-Energy-Future".
|
\section{Introduction}
There are many problems in computational physics that involve solving partial
differential equations (PDEs) in complex geometries. Examples include fluid
flows in complicated systems, vein networks in plant leaves, and tumours in
human bodies. Standard solution methods for PDEs in complex domains typically
involve triangulation and unstructured grids. This rules out coarse-scale
discretizations and thus efficient geometric multi-level solutions. Also, mesh
generation for three-dimensional complex geometries remains a challenge, in
particular if we allow the geometry to evolve with time.
In the past several years, there has been much effort put into the development
of numerical methods for solving partial differential equations in complex
domains. However, most of these methods typically require tools
not frequently available in standard finite element and finite difference
software packages. Examples of such approaches include the extended and
composite finite element methods (e.g.,
\cite{GR07,dolbow09,fries10,duddu11,he11,PRE11,byfut12,bernauer12}), immersed
interface methods (e.g., \cite{LL94,LiIto_2006,SethianShan_2008,li12,wan12}),
virtual node methods with embedded boundary conditions (e.g.,
\cite{Bedrossian10,Zhu12,Hellrung12}), matched interface and boundary methods
(e.g., \cite{zhou06,zhao09,zhao10,xia11,zhou12}), modified finite
volume/embedded boundary/cut-cell methods/ghost-fluid methods (e.g.,
\cite{GlimmMarchesinMcBryan_1981,JohansenColella_1998,FedkiwAslamMerrimanOsher_1999,GibouFedkiwChengKang_2002,GibouFedkiw_2005,JiLienYee_2006,MacklinLowengrub_2006,zhong07,MacklinLowengrub_2008,Colellaetal_2008,lui09,Uzgoren09,OSK09,cisternino12,Coco13,Papac10,Papac13,Helgadottir11,Theillard13}).
In another approach, known as the fictitious domain method (e.g.,
\cite{GlowinskiPanPeriaux_CMAME_1994,GlowinskiPanWellsZhou_JCP_1996,RamiereAngotBelliard_CMAME_2007,lohner07}),
the original system is either augmented with equations for Lagrange multipliers
to enforce the boundary conditions, or the penalty method is used to enforce
the boundary conditions weakly. See also \cite{Gibou13} for a review of
numerical methods for solving the Poisson equation, the diffusion equation and
the Stefan problem on irregular domains.
An alternate approach for simulating PDEs in complex domains, which does not
require any modification of standard finite element or finite difference
software, is the diffuse-domain method. In this method, the domain is
represented implicitly by a phase-field function, which is an approximation of
the characteristic function of the domain. The domain boundary is replaced by
a narrow diffuse interface layer such that the phase-field function rapidly
transitions from one inside the domain to zero in the exterior of the domain.
The boundary of the domain can thus be represented as an isosurface of the
phase-field function. The PDE is then reformulated on a larger, regular domain
with additional source terms that approximate the boundary conditions.
Although uniform grids can be used, local grid refinement near domain
boundaries improves efficiency and enables the use of smaller interface
thicknesses than are achievable using uniform grids. A related approach
involves the level-set method \cite{osher88,Sethian99,osher03a} to describe the
implicitly embedded surface and to obtain the appropriate surface operators
(e.g., \cite{greer06}).
The diffuse-domain method (DDM) was introduced by Kockelkoren et al.\
\cite{Kockelkoren03} to study diffusion inside a cell with zero Neumann
boundary conditions at the cell boundary (a similar approach was also used in
\cite{Bueno06b,Bueno06a} using spectral methods). The DDM was later used to
simulate electrical waves in the heart \cite{Fenton05} and membrane-bound
Turing patterns \cite{Levine05}. More recently, diffuse-interface methods have
been developed for solving PDEs on stationary \cite{ratz} and evolving
\cite{DD07,dziuk08a,dziuk08b,elliott09a,elliott09,DE12} surfaces.
Diffuse-domain methods for solving PDEs in complex evolving domains with
Dirichlet, Neumann and Robin boundary conditions were developed by Li et al.\
\cite{Li09} and by Teigen et al.\ \cite{Teigen09-b} who modelled bulk-surface
coupling. The DDM was also used by Aland et al.\ \cite{Aland10} to simulate
incompressible two-phase flows in complex domains in 2D and 3D, and by Teigen
et al.\ \cite{Teigen11} to study two-phase flows with soluble surfactants.
Li et al.\ \cite{Li09} showed that in the DDM there exist several
approximations to the physical boundary conditions that converge asymptotically
to the correct sharp-interface problem. Li et al.\ presented some numerical
convergence results for a few selected problems and observed that the choice of
boundary condition can significantly affect the accuracy of the DDM. However,
Li et al.\ did not perform a quantitative comparison between the different
boundary-condition approximations, nor did they estimate convergence rates.
Further, Li et al.\ did not address the source of the different levels of
accuracy they observed for the different boundary-condition approximations.
In the context of Dirichlet boundary conditions, Franz et al.\
\cite{Franz12} recently presented a rigorous error analysis of the DDM for
a reaction-diffusion equation and found that the method converges only with
first-order accuracy in the interface thickness parameter $\epsilon$, which
they confirmed numerically. Similar results were obtained numerically by
Reuter et al.\ \cite{Reuter12} who reformulated the DDM using an integral
equation solver. Reuter et al.\ demonstrated that their generalized DDM,
with appropriate choices of approximate surface delta functions, converges
with first-order accuracy to solutions of the Poisson equation with Dirichlet
boundary conditions.
Here, we focus on Neumann and Robin boundary conditions and we present
a matched asymptotic analysis of general diffuse-domain methods in a fixed
complex geometry, focusing on the Poisson equation for Robin boundary
conditions and a steady reaction-diffusion equation for Neumann boundary
conditions. However, our approach applies to transient problems and more
general equations in the same way as shown in \cite{Li09}. Our analysis
shows that for certain choices of the boundary condition approximations, the
DDM is second-order accurate in $\epsilon$, which in practice is proportional
to the smallest mesh size. However, for other choices the DDM is only
first-order accurate. This helps to explain why the choice of boundary
condition approximation is important for rapid global convergence and high
accuracy.
Further, inspired by the work of Karma and Rappel \cite{Karma98} and Almgren
\cite{Almgren99}, who incorporated second-order corrections in their phase
field models of crystal growth and by the work of Folch et al.\
\cite{Folch1999} who added second-order corrections in phase-field models of
advection, we also suggest correction terms that may be added to yield a more
accurate version of the diffuse-domain method. Simple modifications of
first-order boundary condition approximations are proposed to achieve
asymptotically second-order accurate schemes. Our analytic results are
confirmed numerically for selected test problems.
The outline of the paper is as follows. In \cref{sec:dda} we introduce and
present an analysis of general diffuse-domain methods. In
\cref{sec:discretization} the numerical methods are described, and in
\cref{sec:results} the test cases are introduced and numerical results are
presented and discussed. We finally give some concluding remarks in
\cref{sec:conclusion}.
\section{The diffuse-domain method}
\label{sec:dda}
The main idea of the DDM is to extend PDEs that are defined inside complex and
possibly time-dependent domains into larger, regular domains. As a model
problem, consider the Poisson equation in a domain $D$,
\[
\boldsymbol\Delta u = f,
\]
with Neumann or Robin boundary conditions. As shown in Li et al.\ \cite{Li09},
the results for the Poisson equation can be used directly to obtain
diffuse-domain methods for more general second-order partial differential
equations in evolving domains.
The DDM equation is defined in a larger, regular domain $\Omega\supset D$ as
\begin{equation}
\div(\phi\grad u) + \text{BC} = \phi f,
\label{ddm}
\end{equation}
see \cref{fig:ddadomain}. Here $\phi$ approximates the characteristic function
of $D$,
\[
\chi_D = \begin{cases}
1 & \text{if $x\in D$,} \\
0 & \text{if $x\notin D$,}
\end{cases}
\]
and BC is chosen to approximate the physical boundary condition, cf.\
\cite{Li09}. This typically involves diffuse-interface approximations of the
surface delta function. A standard approximation of the characteristic
function is the phase-field function,
\begin{equation}
\chi_D \simeq \phi(\vct x,t) = \frac{1}{2} \left( 1
- \tanh \left( \frac{3r(\vct x,t)}{\epsilon} \right) \right).
\label{eq:characteristic}
\end{equation}
Here $\epsilon$ is the interface thickness and $r(\vct x,t)$ is the
signed-distance function with respect to $\partial D$, which is taken to be
negative inside $D$.
\begin{figure}[tbp]
\centering
\begin{tikzpicture}[scale=0.8]
\draw[thick] (0,0) rectangle(9,5);
\draw[thick] (2,1)
.. controls (1,1) and (1,4) .. (3,3.5)
.. controls (5,3) and (4,5) .. (6,4)
node[above right] {$\partial D$}
.. controls (8,3) and (8,1.5) .. (7,1.5)
.. controls (3,1.5) and (3,1) .. (2,1);
\node at (2.3,2.3) {$D$};
\node at (7.6,0.8) {$\Omega$};
\node at (4.8,2.4) {$\chi_D=1$};
\node at (5.4,0.5) {$\chi_D=0$};
\end{tikzpicture}
\caption{A complex domain $D$ covered by a larger, regular domain $\Omega$.}
\label{fig:ddadomain}
\end{figure}
As Li et al.\ \cite{Li09} described, there are a number of different choices
for BC in \cref{ddm}. For example, in the Neumann case, where $\vct n\cdot\nabla
u = g$ on $\partial D$, one may take:
\[
\text{BC} =
\begin{cases}
\text{BC1} = g|\nabla\phi|,
\quad\text{or}\\
\text{BC2} = \epsilon g|\nabla\phi|^2.
\end{cases}
\]
In the Robin case, where $\vct n\cdot\nabla u = k(u-g)$ on $\partial D$, one may use
analogous approximations:
\begin{equation}
\text{BC} =
\begin{cases}
\text{BC1} = k(u - g) |\nabla\phi |,
\quad\text{or} \\
\text{BC2} = \epsilon k(u - g) |\nabla\phi|^2.
\end{cases}
\label{eq:bc_robin}
\end{equation}
Note that the terms $|\nabla\phi|$ and $\epsilon|\nabla\phi|^2$ approximate the
surface delta function. Following Li et al.\ \cite{Li09} we assume that $g$ is
extended constant in the normal direction off $\partial D$ and that $f$ is
smooth up to $\partial D$ and is extended into the exterior of $D$ constant in
the normal direction. We next perform an asymptotic analysis to estimate the
rate of convergence of the corresponding approximations.
\subsection{Asymptotic analysis}
\label{sec:asymptotic_analysis}
To show asymptotic convergence, we need to consider the expansions of the
diffuse-domain variables in powers of the interface thickness $\epsilon$ in
regions close to and far from the interface. These are called inner and outer
expansions, respectively. The two expansions are then matched in a region
where both are valid, see \cref{fig:regions}, which provides the boundary
conditions for the outer variables. We refer the reader to \cite{Caginalp88}
and \cite{Pego88} for more details and discussion of the general procedure.
\begin{figure}[b!p]
\centering
\begin{tikzpicture}
[
yscale=0.8,
interface/.style={thick},
inner/.style={fill=gray,dotted,fill opacity=0.2},
outer/.style={fill=gray,dashed,fill opacity=0.3},
labels/.style={above right, font=\small},
]
\begin{scope}[scale=0.8]
\draw[thick] (0,0) rectangle(9,5);
\draw[thick] (2,1)
.. controls (1,1) and (1,4) .. (3,3.5)
.. controls (5,3) and (4,5) .. (6,4)
.. controls (8,3) and (8,1.5) .. (7,1.5) node (g1) {}
.. controls (3,1.5) and (3,1) .. (2,1);
\node at (7.4,0.4) {$\Omega$};
\node at (2.0,2.3) {$D$};
\coordinate (a) at (2.4,1.6);
\coordinate (b) at (3.4,1.6);
\coordinate (c) at (2.4,0.6);
\coordinate (d) at (3.4,0.6);
\draw[very thin] (c) rectangle (b);
\end{scope}
\begin{scope}[xshift=2.5cm,yshift=4.4cm]
\draw[very thin] (a) -- (0, 3);
\draw[very thin] (b) -- (8, 3);
\draw[very thin] (c) -- (0,-3);
\draw[very thin] (d) -- (8,-3);
\fill[white] (0,-3) rectangle (8,3);
\draw[very thin, densely dotted] (a) -- (0, 3);
\draw[very thin, densely dotted] (b) -- (8, 3);
\draw[interface] (0,0)
.. controls (2, 0.5) and (3, 0.5) .. (4,0)
.. controls (5,-0.5) and (6,-1.0) .. (8,0);
\draw[inner] (0,2.0)
.. controls (2,2.5) and (3,2.5) .. (4,2.0)
.. controls (5,1.5) and (6,1.0) .. (8,2.0) -- (8,-2.0)
.. controls (6,-3.0) and (5,-2.5) .. (4,-2.0)
.. controls (3,-1.5) and (2,-1.5) .. (0,-2.0) -- cycle;
\draw[outer] (0,3.0) -- (8,3.0) -- (8,1.0)
.. controls (6,0.0) and (5,0.5) .. (4,1.0)
.. controls (3,1.5) and (2,1.5) .. (0,1.0) -- cycle;
\draw[outer] (0,-3.0) -- (8,-3.0) -- (8,-1.0)
.. controls (6,-2.0) and (5,-1.5) .. (4,-1.0)
.. controls (3,-0.5) and (2,-0.5) .. (0,-1.0) -- cycle;
\node[labels] at (0.1,2.25) {Outer region};
\node[labels] at (0.1,1.25) {Overlapping region};
\node[labels] at (0.1,0.25) {Inner region};
\draw[decorate,decoration=brace] (8.1, 3.0) --
node[right=0.5em] {$D$} (8.1, 0.1);
\draw[decorate,decoration=brace] (8.1,-0.1) --
node[right=0.5em] {$\Omega$} (8.1,-3.0);
\node[right=0.5em] at (8.1,0) {$\partial D$};
\end{scope}
\end{tikzpicture}
\caption{A sketch of the regions used for the matched asymptotic expansions.
The inner region is marked with a light grey colour and the outer region
with a slightly darker grey colour. The overlapping region is marked with
the darkest grey colour.}
\label{fig:regions}
\end{figure}
The outer expansion for the variable $u(\vct x;\epsilon)$ is simply
\begin{equation}
u(\vct x;\epsilon) = u^{(0)}(\vct x)
+ \epsilon u^{(1)}(\vct x)
+ \epsilon^2 u^{(2)}(\vct x) + \cdots.
\label{eq:outer}
\end{equation}
The outer expansion of an equation is then found by inserting the expanded
variables into the equation.
The inner expansion is found by introducing a local coordinate system near the
interface $\partial D$,
\[
\vct x(\vct s,z;\epsilon) = \vct X(\vct s;\epsilon)
+ \epsilon z\vct n(\vct s;\epsilon),
\]
where $\vct X(\vct s;\epsilon)$ is a parametrization of the interface, $\vct
n(\vct s;\epsilon)$ is the interface normal vector that points out of $D$, $z$
is the stretched variable
\[
z = \frac{r(\vct x)}{\epsilon},
\]
and $r$ is the signed distance from the point $\vct x$ to $\partial D$. In the
local coordinate system, the derivatives become
\begin{align*}
\grad &= \frac{1}{\epsilon}\vct n\partial_z
+ \frac{1}{1+\epsilon z\kappa}\boldsymbol\nabla_{\text s}, \\
\boldsymbol\Delta &= \frac{1}{\epsilon^2}\partial_{zz}
+ \frac{1}{\epsilon}\frac{\kappa}{1+\epsilon z\kappa}\partial_z
+ \frac{1}{1+\epsilon z\kappa}\boldsymbol\nabla_{\text s}
\cdot \left(\frac{1}{1+\epsilon z\kappa}\boldsymbol\nabla_{\text s}\right),
\end{align*}
where $\kappa\equiv\boldsymbol\nabla_{\text s}\cdot\vct n$ is the curvature of the interface. Note that
$\vct n = -\frac{\nabla\phi}{|\nabla\phi|}$. The inner variable $\hat u(z,\vct
s;\epsilon)$ is now given by
\[
\hat u(z,\vct s;\epsilon) \equiv u(\vct x;\epsilon)
= u(\vct X(\vct s;\epsilon) + \epsilon z\vct n(\vct s;\epsilon);\epsilon),
\]
and the inner expansion is
\begin{equation}
\hat u(z,\vct s;\epsilon) = \hat u^{(0)}(z,\vct s)
+ \epsilon \hat u^{(1)}(z,\vct s)
+ \epsilon^2\hat u^{(2)}(z,\vct s) + \cdots.
\label{eq:inner}
\end{equation}
To obtain the matching conditions, we assume that there is a region of overlap
where both the expansions are valid. In this region, the solutions have to
match. In particular, if we evaluate the outer expansion in the inner
coordinates, this must match the limits of the inner solutions away from the
interface, that is
\[
u(\vct X + \epsilon z\vct n;\epsilon) \simeq \hat u(z,\vct s;\epsilon).
\]
Insert the expansions into \cref{eq:outer,eq:inner} to get
\[
\begin{split}
u^{(0)}(\vct X + \epsilon z\vct n)
+ \epsilon u^{(1)}(\vct X + \epsilon z\vct n)
+ \epsilon^2u^{(2)}(\vct X + \epsilon z\vct n) + \cdots \\
\simeq \hat u^{(0)}(z,\vct s)
+ \epsilon \hat u^{(1)}(z,\vct s)
+ \epsilon^2\hat u^{(2)}(z,\vct s) + \cdots.
\end{split}
\]
The terms on the left-hand side can be expanded as a Taylor series,
\[
u^{(k)}(\vct X + \epsilon z\vct n) = u^{(k)}(\vct s)
+ \epsilon z\vct n\cdot\grad u^{(k)}(\vct s)
+ \frac{\epsilon^2z^2}{2} \vct n\cdot\grad\grad u^{(k)}(\vct s)\cdot\vct n
+ \cdots,
\]
where $k \in \mathbb N$ and $u^{(k)}(\vct s)\equiv u^{(k)}(\vct X(\vct
s;\epsilon))$. Now we end up with the matching equation
\[
\begin{split}
u^{(0)}(\vct s)
+& \epsilon\left(u^{(1)}(\vct s) + z\vct n\cdot\grad u^{(0)}(\vct s)\right) \\
+& \epsilon^2\left(
u^{(2)}(\vct s) + z\vct n\cdot\grad u^{(1)}(\vct s)
+ \frac{z^2}{2} \vct n\cdot\grad\grad u^{(0)}(\vct s)\cdot\vct n \right) \\
+& \cdots \simeq \hat u^{(0)}(z,\vct s) + \epsilon \hat u^{(1)}(z,\vct s)
+ \epsilon^2\hat u^{(2)}(z,\vct s) + \cdots,
\end{split}
\]
which must hold when the interface width is decreased, that is $\epsilon \to
0$. In the matching region it is required that $\epsilon z = \bigo 1$. Under
this condition, if we let $z \to \pm\infty$, we get the following asymptotic
matching conditions:
\[
\ensuremath{\lim_{z\to\pm\infty}} \hat u^{(0)}(z,\vct s) = u^{(0)}(\vct s),
\label{eq:match1}
\]
and as $z \to \pm\infty$,
\begin{align}
\label{eq:match2}
\hat u^{(1)}(z,\vct s) &= u^{(1)}(\vct s)
+ z\vct n\cdot\grad u^{(0)}(\vct s) + \smallo 1, \\
\nonumber
\hat u^{(2)}(z,\vct s) &= u^{(2)}(\vct s)
+ z\vct n\cdot\grad u^{(1)}(\vct s) \\ &\quad
\label{eq:match3}
+ \frac{z^2}{2} \vct n\cdot\grad\grad u^{(0)}(\vct s)\cdot\vct n
+ \smallo 1,
\end{align}
where the quantities on the right-hand side are the limits from the interior
($-$) and exterior ($+$) of $D$. Here $\smallo 1$ means that the expressions
approach equality when $z\to\pm\infty$. That is, $\smallo 1$ is defined such
that if some function $f(z) = \smallo 1$, then we have $\ensuremath{\lim_{z\to\pm\infty}} f(z) = 0$.
\subsection{Poisson equation with Robin boundary conditions}
Now we are ready to consider the Poisson equation with Robin boundary
conditions,
\begin{equation}
\begin{alignedat}{2}
\boldsymbol\Delta u &= f & \ensuremath{\textup{in}\ } D, \\
\vct n\cdot\grad u &= k(u - g) \qquad & \ensuremath{\textup{on}\ } \partial D,
\end{alignedat}
\label{eq:poiss_robin}
\end{equation}
where $k\le 0$. Consider a general DDM approximation,
\begin{equation}
\div\left(\phi\grad u\right) + \frac{1}{\epsilon^2}\psi = \phi f.
\label{eq:general_dda}
\end{equation}
where $\psi$ represents the BC approximation in the DDM. The scaling factor
$1/\epsilon^2$ is taken for later convenience. If we assume that $\psi$ is
local to the interface (e.g., vanishes to all orders in $\epsilon$ away from
$\partial D$) and that $f$ is independent of $\epsilon$ (e.g., is smooth in
a neighbourhood of $\partial D$ and is extended constant in the normal
direction out of $D$), which is the case for the approximations BC1 and BC2
given in \cref{eq:bc_robin}, then the outer solution to this equation
when $z\to -\infty$ satisfies
\begin{equation}
\begin{split}
\boldsymbol\Delta u^{(0)} &= f, \\
\boldsymbol\Delta u^{(1)} &= 0, \\
\boldsymbol\Delta u^{(k)} &= 0,\qquad k = 2,3,\dots.
\end{split}
\label{eq:dda_outer}
\end{equation}
Now, if $u^{(0)}$ satisfies \cref{eq:poiss_robin} and $u^{(1)} \ne 0$ then the
outer expansion $u\approx u^{(0)} + \epsilon u^{(1)} +~\dots$ and the DDM is
asymptotically first-order accurate. However, if $u^{(1)} = 0$, then $u\approx
u^{(0)} + \epsilon^2 u^{(2)} + \dots$ and the DDM is asymptotically
second-order accurate. Determining which of these is the case requires
matching the outer solutions to the solutions of the inner equations.
\subsubsection{Matching conditions}
Before we analyse the inner expansions, we develop a higher-order matching
condition based on \cref{eq:match2,eq:match3} that matches a Robin boundary
condition for $u^{(1)}$. First we take the derivative of \cref{eq:match3} with
respect to $z$ and subtract $k$ times \cref{eq:match2}, which gives
\[
\hat u^{(2)}_z - k\hat u^{(1)}
= - ku^{(1)} - kz\vct n\cdot\grad u^{(0)}
+ \vct n\cdot\grad u^{(1)} + z \vct n\cdot\grad\grad u^{(0)}\cdot\vct n.
\]
Move the terms that make up a Robin condition for $u^{(1)}$ to the left-hand
side, and move the rest to the right-hand side, that is
\begin{equation}
\vct n\cdot\grad u^{(1)} - ku^{(1)}
= \hat u^{(2)}_z
- k\hat u^{(1)}
+ kz\vct n\cdot\grad u^{(0)}
- z \vct n\cdot\grad\grad u^{(0)}\cdot\vct n.
\label{eq:robin_mc}
\end{equation}
The Laplacian can be decomposed into normal and tangential components as
\begin{equation}
\boldsymbol\Delta u = \vct n\cdot\grad\grad u\cdot\vct n + \kappa\vct n\cdot\grad u + \boldsymbol\Delta_{\text s} u,
\label{laplace decomp}
\end{equation}
which can be shown by writing the gradient vector as $\grad = \vct n\vct n\cdot\grad
+ \boldsymbol\nabla_{\text s}$. We can therefore write
\[
\vct n\cdot\grad\grad u^{(0)}\cdot\vct n
= f - \kappa\vct n\cdot\grad u^{(0)} - \boldsymbol\Delta_{\text s} u^{(0)}
= \hat f^{(0)} - \kappa k\left(u^{(0)} - \hat g\right) - \boldsymbol\Delta_{\text s} u^{(0)},
\]
where we have assumed that $u^{(0)}$ satisfies the system
\eqref{eq:poiss_robin}, as demonstrated below. If we insert this into the
matching condition \eqref{eq:robin_mc}, we get
\begin{equation}
\vct n\cdot\grad u^{(1)} - ku^{(1)}
= \hat u^{(2)}_z
- k\hat u^{(1)}
- z\left(\hat f^{(0)}
- \left(\kappa+k\right)k\left(u^{0)}-\hat g\right)
- \boldsymbol\Delta_{\text s} u^{(0)}
\right),
\label{eq:robin_mc2}
\end{equation}
as $z\to - \infty$.
\subsubsection{Inner expansions}
\label{sec:dda_robin_inner}
Now consider the inner expansion of \cref{eq:general_dda},
\[
\frac{1}{\epsilon^2}\left(\phi\hat u_z\right)_z
+ \frac 1 \epsilon\frac{\kappa}{1+\epsilon z\kappa}\phi\hat u_z
+ \frac{\phi}{1+\epsilon z\kappa}\boldsymbol\nabla_{\text s}
\cdot\left(\frac{1}{1+\epsilon z\kappa}\boldsymbol\nabla_{\text s}\hat u\right)
+ \frac{1}{\epsilon^2}\hat\psi
= \phi\hat f.
\]
Expand $\hat u$, $\hat f$, $\hat \psi$ and $\displaystyle{\frac{1}{1+\epsilon
z\kappa}}$ in powers of $\epsilon$, to get
\begin{multline*}
\frac{1}{\epsilon^2}\left(\phi\hat u_z^{(0)}\right)_z
+ \frac{1}{\epsilon}\left(\phi\hat u_z^{(1)}\right)_z
+ \left(\phi\hat u_z^{(2)}\right)_z
+ \frac 1 \epsilon\kappa\phi\hat u_z^{(0)}
+ \kappa\phi\hat u_z^{(1)}-z\kappa^2\hat u_z^{(0)} \\
+ \phi\boldsymbol\Delta_{\text s}\hat u^{(0)}
+ \frac{1}{\epsilon^2}\hat\psi^{(0)}
+ \frac{1}{\epsilon}\hat\psi^{(1)}
+ \hat\psi^{(2)}
= \phi\hat f^{(0)} + \bigo{\epsilon}.
\end{multline*}
and then collect the leading order terms. Note that because $f$ is smooth up
to $\partial D$ and extended constantly outside $D$ we have that $\hat f^{(0)}$
is independent of $z$. The lowest power of $\epsilon$ gives
\[
\left( \phi \hat u^{(0)}_z \right)_z = - \hat\psi^{(0)}.
\]
Suppose that $\hat\psi^{(0)} = 0$, which is the case as we show below for BC1
and BC2, then we obtain $\hat u_z^{(0)} = 0$. By the matching condition
\eqref{eq:match1}, this gives $\hat u^{(0)}(z,\vct s) = u^{(0)}(\vct s)$,
where $u^{(0)}(\vct s)$ is the limiting value of $u^{(0)}$.
The next order terms give
\begin{equation}
\left( \phi\hat u_z^{(1)} \right)_z = - \hat\psi^{(1)}.
\label{eq:first_bc}
\end{equation}
Integrating from $-\infty$ to $+\infty$ in $z$ and using the matching condition
\eqref{eq:match2}, we get
\[
\vct n\cdot\grad u^{(0)} = \eint{\hat\psi^{(1)}}.
\]
To obtain a Robin boundary condition for $u^{(0)}$, we need that
\[
\eint{\hat\psi^{(1)}} = k(u^{(0)} - g).
\]
Now consider the zeroth order terms,
\begin{equation}
\left( \phi\hat u_z^{(2)} \right)_z
= \phi\hat f^{(0)}
- \kappa\phi\hat u_z^{(1)}
- \phi\boldsymbol\Delta_{\text s} u^{(0)}
- \hat\psi^{(2)}.
\label{eq:ddarobin0}
\end{equation}
If we subtract
\[
\left(\phi k\hat u^{(1)} + z\phi\left(
\hat f^{(0)} - (\kappa+ k)k\left(u^{(0)}-\hat g\right) - \boldsymbol\Delta_{\text s} u^{(0)}
\right)\right)_z
\]
from both sides of \cref{eq:ddarobin0}, we get
\begin{multline*}
\left(\phi\hat u^{(2)}_z
- \phi k\hat u^{(1)}
- z\phi\left(\hat f^{(0)}
- (\kappa+k) k\left(u^{(0)}-\hat g\right)
- \boldsymbol\Delta_{\text s} u^{(0)}
\right) \right)_z \\
= - \hat\psi^{(2)} - k\phi_z\hat u^{(1)}
- z\phi_z\left(\hat f^{(0)}
- (\kappa +k)k\left(u^{(0)}-\hat g\right)
- \boldsymbol\Delta_{\text s} u^{(0)} \right) \\
- \phi(\kappa + k)\left(\hat u^{(1)}_z-k\left(u^{(0)}-g\right)\right),
\end{multline*}
where we have taken into account the cancellation of terms and used the fact
that $\hat f^{(0)}$ and $\hat g$ are independent of $z$. The latter
holds when $g$ is extended as a constant in the normal direction off
$\partial D$, e.g., $\hat g(z,s) = g(s)$ and is independent of $z$ and
$\epsilon$. Next, we integrate and use the matching condition
\eqref{eq:robin_mc2} on the left-hand side,
\begin{multline}
\vct n\cdot\grad u^{(1)} - k u^{(1)}
= \eint{\left(\hat\psi^{(2)} + k\phi_z\hat u^{(1)}
+ \phi(\kappa + k)\left(\hat u^{(1)}_z-k\left(u^{(0)}-g\right)\right)
\right)} \\
+ \left(\hat f^{(0)} - (\kappa + k)k\left(u^{(0)} - \hat g\right)
- \boldsymbol\Delta_{\text s} u^{(0)}\right) \eint{z\phi_z}.
\label{order 1 BC}
\end{multline}
If the right-hand side of \cref{order 1 BC} vanishes, then we obtain
$\vct n\cdot\grad u^{(1)} - k u^{(1)}=0$ from which we can conclude that the outer
solution $u^{(1)} = 0$ since $u^{(1)}$ is harmonic: $\boldsymbol\Delta u^{(1)} = 0$. Next
we analyse the boundary condition approximations BC1 and BC2.
\subsubsection{Analysis of BC1}
The BC1 approximation corresponds to
\[
\frac{1}{\epsilon^2}\psi = k(u-g)|\nabla\phi|.
\]
Since $\phi=1$ in the outer region (interior part of $D$), we conclude that
$\psi$ vanishes in the outer region. In the inner region, we have
\begin{equation}
\frac{1}{\epsilon^2}\psi
= -\frac{k}{\epsilon}\left(\hat u - g\right)\phi_z
= -\frac{k}{\epsilon}\left(\hat u^{(0)}- g\right)\phi_z
- k\hat u^{(1)}\phi_z + \bigo{\epsilon},
\label{expansion of BC1}
\end{equation}
where we have used that $\hat g(z,s) = g(s)$ and is independent of $z$ and
$\epsilon$. Since $\hat\psi^{(0)} = 0$, we conclude from our analysis in
\cref{sec:dda_robin_inner} that $\hat u^{(0)}_z = 0$ and hence $\hat u^{(0)}
= u^{(0)}$.
The next orders of \cref{expansion of BC1} give
\[
\hat\psi^{(1)} = -k\left(u^{(0)} - g\right)\phi_z,
\qquad\text{and}\quad
\hat\psi^{(2)} = -k\hat u^{(1)}\phi_z.
\]
A direct calculation then shows that
\[
\eint{\hat\psi^{(1)}} = k(u^{(0)} - g),
\]
as desired. Thus, the leading order outer solution $u^{(0)}$ satisfies the
problem (\ref{eq:poiss_robin}).
To continue, we must first consider \cref{eq:first_bc},
\[
\left(\phi\hat u_z^{(1)}\right)_z = k\left(u^{(0)} - g\right)\phi_z,
\]
from which we get
\[
\hat u^{(1)}(z,\vct s) = u^{(1)}(\vct s) + zk\left(u^{(0)} - g\right),
\]
where $u^{(1)}(\vct s)$ is the limiting value of the outer solution (e.g., see
\cref{eq:match2}). Combining this with \cref{order 1 BC}, we obtain
\begin{equation}
\vct n\cdot\grad u^{(1)} - k u^{(1)} = \left(\hat f^{(0)}
- (\kappa + k)k\left(u^{(0)} - g\right)
- \boldsymbol\Delta_{\text s}\hat u^{(0)}\right)\eint{z\phi_z}.
\label{order 1 BC 1}
\end{equation}
Further, it follows from the definition of the phase-field
function~\eqref{eq:characteristic} that $z\phi_z$ is an odd function.
Therefore the integral on the right-hand side of \cref{order 1 BC 1} is
equal to zero. Thus $\vct n\cdot\grad u^{(1)}-k u^{(1)} = 0$ and so by our
arguments below \cref{eq:dda_outer}, the DDM with BC1 is second-order accurate
in $\epsilon$.
\subsubsection{Analysis of BC2}
When the BC2 approximation is used, we obtain
\[
\frac{1}{\epsilon^2}\psi = \epsilon k(u-g)|\nabla\phi|^2.
\]
Accordingly, in the inner region, we obtain
\[
\hat\psi^{(0)} = 0,
\qquad
\hat\psi^{(1)} = k\left( u^{(0)} - g\right)\phi_z^2,
\qquad\text{and}\quad
\hat\psi^{(2)} = k \hat u^{(1)}\phi_z^2.
\]
Since $\eint{\phi_z^2} = 1$, we get
\[
\eint{\hat\psi^{(1)}} = k(u^{(0)} - \hat g),
\]
as desired. From \cref{eq:first_bc} we have
\[
\left(\phi\hat u_z^{(1)}\right)_z = -k\left(u^{(0)} - g\right)\phi_z^2.
\]
Using that $\phi_z = -6\phi(1-\phi)$, this gives
\begin{equation}
\hat u^{(1)}(z,\vct s) = C(\vct s) - k\left(u^{(0)} - g\right) F(\phi),
\label{u1 inner solve B2}
\end{equation}
where
\begin{align}
\label{u1 inner solve B2 a}
F(\phi) &= -\frac{1}{6}\log\left(1 - \phi\right) + \frac{\phi}{3}, \\
\nonumber
C(\vct s) &= u^{(1)}(\vct s) + \frac{k}{3}\left(u^{(0)} - g\right).
\end{align}
Combining \cref{u1 inner solve B2,u1 inner solve B2 a,order 1 BC} we get
\begin{multline}
\vct n\cdot\grad u^{(1)} - k u^{(1)}
= kC(\vct s)\eint{\left(\phi_z^2 + \phi_z\right)} \\
- k^2\left(u^{(0)} - \hat g\right)
\eint{F(\phi)\left(\phi_z^2 + \phi_z\right)} \\
+ \left(\kappa + k\right)k\left(u^{(0)} - \hat g\right)
\eint{\phi\left(3\phi - 2\phi^2 - 1\right)}.
\label{order 1 BC 2}
\end{multline}
Direct calculations show that
\[
\eint{\left(\phi_z^2 + \phi_z\right)}
= \eint{\left(3\phi^2 - 2\phi^3 - \phi\right)} = 0,
\]
and
\[
\eint{F(\phi)\left(\phi_z^2 + \phi_z\right)} = -\frac{1}{36}.
\]
Using these in \cref{order 1 BC 2}, we get
\begin{equation}
\vct n\cdot\grad u^{(1)} - k u^{(1)} = \frac{1}{36}k^2\left(u^{(0)} - g\right).
\label{order 1 BC 2 final}
\end{equation}
This shows that the DDM with BC2 is only first-order accurate because the
solution $u^{(1)}$ of
\[
\begin{alignedat}{2}
\boldsymbol\Delta u^{(1)} &= 0 & \ensuremath{\textup{in}\ } D, \\
\vct n\cdot\grad u^{(1)} - ku^{(1)}
&= \frac{1}{36}k^2\left(u^{(0)} - g\right) \qquad
& \ensuremath{\textup{on}\ } \partial D,
\end{alignedat}
\]
is in general not equal to $0$, e.g., $u^{(1)}\ne 0$.
\subsubsection{Analysis of a second-order modification of BC2}
\label{sec:dda_robin_bc2m}
In order to modify BC2 to achieve second-order accuracy, we introduce
$\tilde\psi$ such that
\[
\hat \psi^{(0)} = 0,
\qquad
\hat \psi^{(1)} = k\left( u^{(0)}- g\right)\phi_z^2,
\qquad\text{and}\quad
\hat \psi^{(2)} = k \hat u^{(1)}\phi_z^2 +\hat{\tilde\psi}^{(0)}.
\]
That is, $\tilde\psi$ perturbs only the higher order terms in the inner
expansion and is chosen to cancel the term on the right-hand side of
\cref{order 1 BC 2 final} in order to achieve $\vct n\cdot\grad u^{(1)} - k u^{(1)}
= 0$, which in turn implies that $u^{(1)} = 0$ and the new formulation is
second-order accurate. The correction $\tilde\psi$ does not affect the
$O(\epsilon^{-2})$ or $O(\epsilon^{-1})$ orders in the system. Thus, $\hat
u^{(0)}$ and $\hat u^{(1)}$ are unchanged from the previous subsection.
\Cref{order 1 BC} now becomes
\[
\vct n\cdot\grad u^{(1)} - k u^{(1)}
= \frac{1}{36}k^2\left(u^{(0)} - g\right)
+ \eint{\hat{\tilde\psi}^{(0)}},
\]
so we wish to determine $\hat{\tilde\psi}^{(0)}$ such that
\[
\eint{\hat{\tilde\psi}^{(0)}}
= -\frac{1}{36}k^2\left(u^{(0)} - g\right).
\]
Two simple ways of achieving this are to take
\[
\hat{\tilde\psi}^{(0)} = -\frac{1}{36}k^2\left(u^{(0)} - g\right)
\times
\begin{cases}
-\phi_z,
\quad\text{or} \\
\phi_z^2.
\end{cases}
\]
Putting everything together, we can obtain BC2M, a second-order version of BC2,
using
\[
\text{BC2M} =
\begin{cases}
\text{BC2M1} = \epsilon k(u - g) |\nabla\phi|
\left(|\nabla\phi| - \frac{k}{36}\right),
\quad\text{or} \\
\text{BC2M2} = \epsilon k(u - g) |\nabla\phi|^2
\left(1 - \epsilon\frac{k}{36}\right).
\end{cases}
\]
The resulting DDM is an elliptic system since $k<0$, as required
for the Robin boundary condition. In each
instance, this is guaranteed if the interface thickness $\epsilon$ is
sufficiently small.
\subsubsection{Other approaches to second-order BCs}
\label{sec:dda_robin_other}
Thus far, we have taken advantage of integration to achieve second-order
accuracy. Alternatively, one may try to add correction terms to directly
obtain second-order boundary conditions without relying on integration. For
example, from \cref{order 1 BC} to achieve second-order accuracy we may take
\begin{multline}
\hat\psi^{(2)}
= -k\phi_z\hat u^{(1)} - \phi(\kappa+k)
\left(\hat u^{(1)}_z - k\left(u^{(0)} - g\right)\right) \\
- \left(\hat f^{(0)}
- (\kappa + k)k\left(u^{(0)} - \hat g\right)
- \boldsymbol\Delta_{\text s}\hat u^{(0)}\right) z\phi_z,
\label{local correction}
\end{multline}
where $\hat u^{(1)}$ is a functional of $\hat\psi^{(1)}$. This provides
another prescription of how to obtain a second-order accurate boundary
condition, which could in principle lead to faster asymptotic convergence since
it directly cancels a term in the inner expansion of the asymptotic matching.
As an illustration, let us use BC1 as a starting point even though this
boundary condition is already second-order accurate. Through the prescription
in \cref{local correction} above, we derive another second-order accurate
boundary condition. To see this, write
\[
\hat\psi^{(0)} = 0,
\qquad
\hat \psi^{(1)} = -k\left(u^{(0)} - g\right)\phi_z,
\qquad\text{and}\quad
\hat \psi^{(2)} = -k\hat u^{(1)}\phi_z + \hat{\tilde\psi}^{(0)},
\]
then from \cref{local correction} we get
\[
\hat{\tilde\psi}^{(0)} = -\left(\hat f^{(0)}
- (\kappa +k)k\left(u^{(0)} - \hat g\right)
- \boldsymbol\Delta_{\text s}\hat u^{(0)}\right) z\phi_z.
\]
This can be achieved by taking
\[
\frac{1}{\epsilon^2}\psi = k(u-g) |\nabla\phi| \left(1 - r(k + \kappa)\right)
+ r|\nabla\phi|\left(f - \boldsymbol\Delta_{\text s} u\right).
\]
where $r$ is the signed distance to $\partial D$ as defined earlier. Note that
we can also achieve second-order accuracy by taking instead
\[
\hat\psi^{(2)} = -k\phi_z\hat u^{(1)}
- \phi(\kappa + k)\left(\hat u^{(1)}_z - k\left(u^{(0)} - g\right)\right)
- \left(\hat f^{(0)} - (\kappa + k)k\left(u^{(0)} - \hat g\right)\right)
z\phi_z,
\]
where we use the fact that the integral involving $\boldsymbol\Delta_{\text s} u^{(0)}$ vanishes in
\cref{order 1 BC}. We refer to these choices, which are by no means
exhaustive, as
\begin{equation}
\text{BC1M} =
\begin{cases}
\text{BC1M1} = k(u-g) |\nabla\phi| \left(1 - r(k + \kappa)\right)
+ r|\nabla\phi| \left(f - \boldsymbol\Delta_{\text s} u\right),
\quad\text{or} \\
\text{BC1M2} = k(u-g) |\nabla\phi| \left(1 - r(k + \kappa)\right)
+ r|\nabla\phi| f.
\end{cases}
\label{BC1M versions}
\end{equation}
We remark, however, that this prescription may not always lead to an optimal
numerical method. For example, when using BC1M2, the system is guaranteed to
be elliptic when $1 - r(k + \kappa) > 0$ for $|r|\approx \epsilon$, which
puts an effective restriction on the interface thickness $\epsilon$ depending
on the values of $k$ and $\kappa$. When BC1M1 is used, the situation is more
delicate since ellipticity cannot be guaranteed when $r>0$ due to the $\boldsymbol\Delta_{\text s}
u$ term. Recall that $r>0$ outside the original domain $D$ and so this issue
is associated with the extending the modified boundary condition outside $D$.
In future work, we plan to consider different extensions that automatically
guarantee ellipticity.
To summarize, we have shown that the DDM
\[
\div\left(\phi\grad u\right) + \text{BC} = \phi f
\]
is a second-order accurate approximation of the system \eqref{eq:poiss_robin}
when BC1, BC2M, and BC1M are used. When BC2 is used, the DDM is only
first-order accurate.
\subsection{Reaction-diffusion equation with Neumann boundary conditions}
Since the Poisson equation with Neumann boundary conditions does not have
a unique solution, we instead consider the steady reaction-diffusion equation
with Neumann boundary conditions,
\begin{equation}
\begin{alignedat}{2}
\boldsymbol\Delta u - u &= f & \ensuremath{\textup{in}\ } D, \\
\vct n\cdot\grad u &= g \qquad & \ensuremath{\textup{on}\ } \partial D.
\end{alignedat}
\label{eq:poiss_neumann}
\end{equation}
Again we consider a general DDM approximation,
\begin{equation}
\div\left(\phi\grad u\right) - \phi u + \frac{1}{\epsilon^2}\psi = \phi f.
\label{eq:reacdiff_general_dda}
\end{equation}
Under the same conditions on $\psi$ as in the previous section, the outer
solution now satisfies
\begin{align*}
\boldsymbol\Delta u^{(0)} - u^{(0)} &= f, \\
\boldsymbol\Delta u^{(k)} - u^{(k)} &= 0,\qquad k = 1,2,3,\dots
\end{align*}
As in the Robin case, if $u^{(0)}$ satisfies \cref{eq:poiss_neumann} and
$u^{(1)} \ne 0$ then the DDM is first-order accurate. However, if $u^{(1)}
= 0$, then the DDM is second-order accurate.
To construct the boundary condition for $u^{(1)}$, we follow the approach from
the Robin case and combine \cref{eq:match3,laplace decomp} to get
\[
\vct n\cdot\grad\grad u^{(0)}\cdot\vct n
= \hat f^{(0)} - \kappa \hat g - \boldsymbol\Delta_{\text s} u^{(0)} + u^{(0)},
\]
assuming that $u^{(0)}$ satisfies the system \eqref{eq:poiss_neumann} as
demonstrated below, and to get
\[
\vct n\cdot\grad u^{(1)} = \hat u_z^{(2)}
- z\left(\hat f^{(0)} - \kappa\hat g - \boldsymbol\Delta_{\text s} u^{(0)} + u^{(0)}\right),
\]
as $z\to -\infty$.
\subsubsection{Inner expansions}
The inner expansion of \cref{eq:reacdiff_general_dda} is analogous to the Robin
case derived in \cref{sec:dda_robin_inner}. As before, if $\hat\psi^{(0)} = 0$
then $\hat u_z^{(0)}
= 0$ and $\hat u^{(0)}(z,\vct s) = u^{(0)}(\vct s)$, the limiting value of the
outer solution. \Cref{eq:first_bc} still holds at the next order and so to get
the desired boundary condition for $u^{(0)}$, we need
\begin{equation}
\eint{\hat\psi^{(1)}} = g.
\label{integral constraint Neumann}
\end{equation}
Analogously to \cref{eq:ddarobin0} the next order equation is
\begin{equation}
\left(\phi\hat u^{(2)}_z\right)_z
+ \phi\kappa\hat u_z^{(1)}
+ \phi\boldsymbol\Delta_{\text s} \hat u^{(0)}
- \phi\hat u^{(0)}
+ \hat\psi^{(2)}
= \phi\hat f^{(0)}.
\label{eq:o1}
\end{equation}
Subtracting
\[
-\left(z\phi\left(\hat f^{(0)}
- \kappa\hat g - \boldsymbol\Delta_{\text s} u^{(0)} + u^{(0)}\right)\right)_z
\]
from \cref{eq:o1} we get
\[
\begin{split}
\bigg(\phi&\hat u_z^{(2)}
- z\phi\left(\hat f^{(0)} - \kappa\hat g
- \boldsymbol\Delta_{\text s} u^{(0)} + u^{(0)}\right)\bigg)_z \\
&= - \hat\psi^{(2)} - \phi\kappa\left(\hat u^{(1)}_z - \hat g\right)
- z\phi_z\left(\hat f^{(0)} - \kappa \hat g
- \boldsymbol\Delta_{\text s} u^{(0)} + u^{(0)}\right),
\end{split}
\]
where we have used $\hat u^{(0)}=u^{(0)}$ as justified below. Integrating, we
obtain
\begin{equation}
\vct n\cdot\grad u^{(1)} = \eint{\left(
\hat\psi^{(2)} + \phi\kappa\left(\hat u^{(1)}_z - \hat g\right)
+ z\phi_z\left(\hat f^{(0)}
- \kappa\hat g
- \boldsymbol\Delta_{\text s} u^{(0)} + u^{(0)}\right)\right)}.
\label{order 1 BC Neumann}
\end{equation}
As in the Robin case, if the right-hand side of \cref{order 1 BC Neumann}
vanishes then we may conclude that $u^{(1)} = 0$ since $u^{(1)}$ satisfies
$\boldsymbol\Delta u^{(1)} - u^{(1)} = 0$ with zero Neumann boundary conditions. We next
analyse the boundary conditions BC1 and BC2.
\subsubsection{Analysis of BC1}
When the BC1 approximation is used, we obtain
\[
\frac{1}{\epsilon^2}\psi = g|\nabla\phi|,
\]
and
\begin{equation}
\hat \psi^{(0)} = 0,
\qquad
\hat \psi^{(1)} = -g\phi_z,
\qquad\text{and}\quad
\hat \psi^{(2)} = 0.
\label{psi expansion for BC1 Neumann}
\end{equation}
Accordingly, we find that $\hat u^{(0)}(z,\vct s) = u^{(0)}(\vct s)$ and
\cref{integral constraint Neumann} holds. Thus $u^{(0)}$ satisfies the system
\eqref{eq:poiss_neumann} as claimed above.
At the next order, from \cref{eq:first_bc,psi expansion for BC1 Neumann} we
obtain
\begin{equation}
\hat u^{(1)}(z,\vct s) = u^{(1)}(\vct s) + z\hat g.
\label{hat u1 Neumann}
\end{equation}
Thus, combining \cref{hat u1 Neumann,order 1 BC Neumann} we get
\[
\vct n\cdot\grad u^{(1)} = \left(\hat f^{(0)}
- \kappa\hat g
- \boldsymbol\Delta_{\text s}\hat u^{(0)}+u^{(0)}\right)\eint{z\phi_z} = 0,
\]
from which we conclude that $u^{(1)} = 0$ and the DDM with BC1 is second-order
accurate.
\subsubsection{Analysis of BC2}
When the BC2 approximation is used, we obtain
\[
\frac{1}{\epsilon^2}\psi = \epsilon g|\nabla\phi|^2,
\]
and
\begin{equation}
\hat \psi^{(0)} = 0,
\qquad
\hat \psi^{(1)} = g\phi_z^2,
\qquad\text{and}\quad
\hat \psi^{(2)} = 0.
\label{psi expansion for BC2 Neumann}
\end{equation}
Analogously to the case when BC1 is used, $\hat u^{(0)}(z,\vct s)
= u^{(0)}(\vct s)$, \cref{integral constraint Neumann} holds, and $u^{(0)}$
satisfies the system \eqref{eq:poiss_neumann}. At the next order, from
\cref{eq:first_bc,psi expansion for BC2 Neumann} we obtain
\begin{equation}
\hat u^{(1)}(z,\vct s) = u^{(1)}(\vct s) + z\hat g\left(3\phi
- 2\phi^2\right).
\label{hat u1 Neumann BC2}
\end{equation}
Combining \cref{hat u1 Neumann BC2,order 1 BC Neumann} we get
\[
\vct n\cdot\grad u^{(1)}
= \kappa\hat g\eint{\left(3\phi^2 - 2\phi^3 - \phi\right)}
+ \left(\hat f^{(0)} - \kappa\hat g \boldsymbol\Delta_{\text s}\hat u^{(0)} + u^{(0)}\right)
\eint{z\phi_z} = 0,
\]
from which we conclude that $u^{(1)} = 0$ and the DDM with BC2 is second-order
accurate for the Neumann problem as well, which is different from the Robin
case.
\subsubsection{Other approaches to second-order BCs}
Analogously to the Robin case, to achieve second-order accuracy we may also
take
\[
\hat\psi^{(2)} + \phi\kappa\left(\hat u^{(1)}_z - \hat g\right)
+ z\phi_z\left(\hat f^{(0)} - \kappa\hat g \boldsymbol\Delta_{\text s} u^{(0)}+u^{(0)}\right) = 0.
\]
Following the same reasoning, alternative boundary conditions analogous to
those in \cref{BC1M versions} may be derived
\[
\text{BC1M} =
\begin{cases}
\text{BC1M1} = g|\nabla\phi| \left(1 - r(k + \kappa)\right)
+ r|\nabla\phi| \left(f - \boldsymbol\Delta_{\text s} u\right),
\quad\text{or} \\
\text{BC1M2} = g|\nabla\phi| \left(1 - r(k + \kappa)\right)
+ r|\nabla\phi| f.
\end{cases}
\]
Note that as in the Robin case, when BC1M1 is used ellipticity cannot be
guaranteed when $r>0$ due to the $\boldsymbol\Delta_{\text s} u$ term.
To summarize, we have shown that the DDM
\[
\div\left(\phi\grad u\right) - \phi u + \text{BC} = \phi f
\]
is a second-order accurate approximation of the system \eqref{eq:poiss_robin}
when BC1, BC2, and BC1M are used.
\section{Discretizations and numerical methods}
\label{sec:discretization}
The equations are discretized on a uniform grid with the second-order
central-difference scheme. The discrete system is solved using a multigrid
method, where a red-black Gauss-Seidel type iterative method is used to relax
the solutions (see \cite{Wise07}). The equations are solved in two-dimensions
in a domain $\Omega = [-2,2]^2$ for all the test cases. Periodic boundary
conditions are used on the domain boundaries $\partial\Omega$. The iterations
are considered to be converged when the residual of the current solution has
reached a tolerance of $10^{-9}$.
Since the phase-field function quickly tends to zero outside the physical
domain $D$, it must be regularized in order to prevent the equations from
becoming ill-posed. We therefore use the modified phase-field function
\[
\hat\phi = \tau + (1 - \tau)\phi,
\]
where the regularization parameter is set to $\tau = 10^{-6}$ unless otherwise
specified. In addition, one should note that the chosen boundary condition for
the computational domain, $\Omega$, should not interfere with the physical
domain. Thus one has to make sure that the distance from the computational
wall to the diffuse interface of $D$ is large enough not to affect the results.
As shown in \cref{sec:asymptotic_analysis}, the normal vector and the
curvature can be calculated from the phase-field function as
\[
\vct n = -\frac{\grad\phi}{|\grad\phi|},
\]
and
\[
\kappa = -\div\frac{\grad\phi}{|\grad\phi|}.
\]
The surface Laplacian can be found from the identity
\[
\boldsymbol\Delta_{\text s} \equiv
\left(I - \vct n\vct n\right) \div\left(I - \vct n\vct n\right)\grad,
\]
where
\[
(I - \vct n\vct n)\grad \equiv (\delta_{ij} - n_in_j)\partial x_i.
\]
In 2D we get
\begin{align*}
\nonumber \boldsymbol\Delta_{\text s} u
&= \left(n_1n_2(n_1n_2)_x + n_1n_2(n_1^2)_y - (1-n_1^2)(n_1^2)_x
- (1-n_2^2)(n_1n_2)_y\right)u_x \\
&\qquad + \left(n_1n_2(n_1n_2)_y + n_1n_2(n_2^2)_x - (1-n_2^2)(n_2^2)_y
- (1-n_1^2)(n_1n_2)_x\right)u_y \\
&\qquad + \left(\left(1 - n_1^2 \right)^2 + n_1^2n_2^2\right)u_{xx} \\
&\qquad + \left(\left(1 - n_2^2 \right)^2 + n_1^2n_2^2\right)u_{yy} \\
&\qquad - 2n_1n_2\left(\left(1 - n_1^2 \right)
+ \left(1 - n_2^2\right)\right)u_{xy}.
\end{align*}
Below, we verify the accuracy of our numerical implementation on several test
problems in which we manufacture a solution to the DDM with different choices
of boundary conditions through particular choices of $f$. As suggested by our
analysis, we find that when we include the surface Laplacian, we are unable to
solve the discrete system using the multigrid method even though the correction
term and the subsequent loss of ellipticity outside $D$ is confined to the
interfacial region. As also mentioned in \cref{sec:dda_robin_other}, future
work involves developing
alternative extensions of the boundary conditions outside $D$ that maintain
ellipticity. Nevertheless, as a proof of principle, we still consider the
effect of this term by using the surface Laplacian of the analytic solution in
the DDM equations.
\section{Results}
\label{sec:results}
We next investigate the performance of the DDM with different choices of
boundary conditions and compare the results with the exact solution of the
sharp-interface equations for the reaction-diffusion equation with Neumann
boundary conditions and the Poisson equation with Robin boundary conditions.
We consider four different cases with Neumann boundary conditions and three
different cases with Robin boundary conditions. For each case, we calculate
and compare the error between the calculated solution $u$ and an analytic
solution $u_{\text{an}}$ of the original PDE, which is extended from $D$ into
$\Omega$. The error is defined as
\[
E_\epsilon = \frac{\|\phi(u_{\text{an}} - u)\|}{\|\phi u_\text{an}\|},
\]
where $\|\cdot\|$ is a norm and $\phi$ is used to restrict the error to the
physical domain $D$. The convergence rate in $\epsilon$ as $\epsilon\to 0$ is
calculated as
\[
k = \log\left(\frac{E_{\epsilon_i}}{E_{\epsilon_{i-1}}}\right) /
\log\left(\frac{\epsilon_i}{\epsilon_{i-1}}\right),
\]
for a decreasing sequence $\epsilon_i$. In the following results we mainly use
the $L^2$ norm,
\[
\|\psi\|_2 = \frac{\sqrt{\sum_{i=1}^N \psi_i^2}}{N},
\]
where $\psi$ is an array with $N$ elements. For a couple of cases, we also
present the results with the $L^\infty$ norm,
\[
\|\psi\|_\infty = \max_{i=1}^N\,|\psi_i|.
\]
For a given $\epsilon$, the error $E_\epsilon$ in both $L^2$ and $L^\infty$ is
calculated by refining the grid spacing until a minimum of two leading digits
converge (e.g., stop changing under refinement) in the $L^2$ norm. In some
cases for the smallest values of $\epsilon$, the error has not yet converged
to two digits in the $L^\infty$ norm. However, due to memory limits on our
computers, we were not able to use more than $n = 8192$ cells in each
direction. This limited our ability to obtain grid convergence, particularly
when BC2 (see below) is used for very small values of $\epsilon$.
\subsection{Neumann boundary conditions}
Consider the steady reaction-diffusion equation with Neumann boundary
conditions,
\[
\begin{alignedat}{3}
\boldsymbol\Delta u - u &= f && \ensuremath{\textup{in}\ } D, \\
\vct n\cdot\grad u &= g \quad && \ensuremath{\textup{on}\ } \partial D.
\end{alignedat}
\]
In this section we solve the DDM systems
\[
\div\left(\phi\grad u\right) - \phi u + \text{BC} = \phi f,
\]
where BC refers to selected boundary condition approximations considered in the
previous section. In the case of BC1M1, as remarked above, the surface
Laplacian term is not solved, rather the surface Laplacian of the analytic
solution is used and is treated as a known source term.
\subsubsection{Case 1}
Consider the case where $D$ is a circle of radius $R=1$ centred at $(0,0)$,
and where the analytic solution to the reaction-diffusion equation in $D$ is
\[
u_{\text{an}}(x,y) = \frac{1}{4}\left( x^2 + y^2 \right).
\]
This corresponds to $f = 1 - (x^2 + y^2)/4$, $g = 1/2$, and $\boldsymbol\Delta_{\text s}
u_{\text{an}} = 0$. In this case, the curvature is $\kappa = 1$.
\subsubsection{Case 2}
Now consider the case where $D$ is the square $D=[-1,1]^2$. Again let the
analytic solution in $D$ be
\[
u_{\text{an}}(x,y) = \frac{1}{4}\left( x^2 + y^2 \right),
\]
so that $f = 1 - (x^2 + y^2)/4$, $g = 1/2$, and $\boldsymbol\Delta_{\text s} u_{\text{an}} = 1/2$.
In this case the curvature is zero almost everywhere.
To initialize the square domain $D$, the signed-distance function is defined as
\[
r(x,y) =
\begin{cases}
|x| - 1 & \text{if $|x|\geq |y|$,} \\
|y| - 1 & \text{else.}
\end{cases}
\]
The phase-field function is then calculated directly from the signed-distance
function in \cref{eq:characteristic}.
\subsubsection{Case 3}
Again let $D$ be the circle centred at $(0,0)$ with radius $R=1$, but now
consider the case where the analytic solution is
\[
u_{\text{an}}(x,y) = y\sqrt{x^2 + y^2},
\]
which corresponds to
\[
f = \frac{3y}{\sqrt{x^2+y^2}} - y\sqrt{x^2 + y^2},
\]
$g = 2y$, and
\[
\boldsymbol\Delta_{\text s} u_{\text{an}} = -\frac{y}{\sqrt{x^2 + y^2}}.
\]
Note that in the DDM equations, $g$ is extrapolated constantly in the normal
direction off of the boundary $\partial D$.
\subsubsection{Case 4}
For the final Neumann case we again let $D=[-1,1]^2$, and we consider the case
where the analytic solution is
\[
u_{\text{an}}(x,y) = e^r,
\]
where $r = \frac{x^2 + y^2}{4}$. This corresponds to
\[
f = r e^r.
\]
The boundary function $g$ and the surface Laplacian of the analytic function
along the boundary are
\begin{align*}
g &= \frac 1 2 e^{\frac{1 + \xi^2}{4}}, \\
\boldsymbol\Delta_{\text s} u_\text{an} &= \frac 1 4\left(\xi^2 + 2\right)e^{\frac{1 + \xi^2}{4}},
\end{align*}
where $\xi \equiv x$ along the bottom and top boundaries, and $\xi \equiv y$
along the left and right boundaries.
\subsubsection{Results}
\Cref{fig:neumann_case1-plot,fig:neumann_case2-plot,fig:neumann_case3-plot,%
fig:neumann_case4-plot} and \cref{tab:neumann_error_per_eps} show convergence
results in the $L^2$ norm where~$\epsilon$ is reduced for Cases 1 to 4 with
BC1, BC1M1, BC1M2 and BC2.
\Cref{fig:neumann_case4-plot-2,tab:neumann_error_per_eps_infty} shows the
results for Case 4 where the $L^\infty$ norm is used. Although the DDM is most
efficient when adaptive meshes are used, here we consider only uniform meshes
to more easily control the discretization errors in order to focus on the
errors in the DDM. As in all diffuse-interface methods, fine grids are
necessary to accurately solve the equations when $\epsilon$ is small. This is
especially apparent for the cases with BC2 where even the finest grid spacing,
$n=8192$ in each direction, becomes too coarse to obtain results for small
$\epsilon$ that have converged with respect to the grid refinements.
The results confirm the second-order accuracy of all the considered boundary
condition approximations. Note that while the difference between BC1 and BC1M1
tends to be small, BC1M1 consistently performs better than BC1. In turn, BC1
performs better than BC2. In case 2 there is a noticeable improvement of
BC1M1 over BC1. Case 3 is the first case that has a nonconstant boundary
condition, and the surface Laplacian of the analytic solution along the
boundary is also nonconstant. An unexpected result for Case 3 is that BC1M2
performs the best. One possible explanation to this is errors due to grid
anisotropy. Therefore we also consider a fourth case, which again has
a nonconstant boundary condition and nonconstant surface Laplacian of the
analytic solution. Since the domain in this case is a square, the effect of
grid anisotropy is lessened. Correspondingly BC1M1 performs the best.
The cases were also calculated with the $L^\infty$ norm, which gave similar
results, although at the smallest values of $\epsilon$, the orders of accuracy
of BC1M1 and BC1M2 may deteriorate in $L^\infty$, as seen in
\cref{fig:neumann_case4-plot-2,tab:neumann_error_per_eps_infty} for case
4. This could be due to the influence of higher order terms in the expansion,
or the amplification of error when $\epsilon$ is small due to the condition
number of the system, which should scale like $\epsilon^{-2}$. This is
currently under investigation.
The difference between BC1 and BC2 is noticeable, especially with regard to the
required amount of grid refinement that is needed to obtain a convergent
result. This provides practical limits for the use of BC2.
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC2},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 3.4e-01)
(4.000e-01, 9.9e-02)
(2.000e-01, 2.6e-02)
(1.000e-01, 6.4e-03)
(5.000e-02, 1.6e-03)
(2.500e-02, 4.2e-04)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 4.8e-01)
(4.000e-01, 1.1e-01)
(2.000e-01, 2.7e-02)
(1.000e-01, 6.5e-03)
(5.000e-02, 1.6e-03)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 3.1e-01)
(4.000e-01, 9.5e-02)
(2.000e-01, 2.6e-02)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Neumann problem with respect to $\epsilon$ for
Case 1, as labelled.}
\label{fig:neumann_case1-plot}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC1M2,
BC2},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 2.5e-01)
(4.000e-01, 7.3e-02)
(2.000e-01, 1.9e-02)
(1.000e-01, 5.2e-03)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 2.0e-01)
(4.000e-01, 2.6e-02)
(2.000e-01, 5.2e-03)
(1.000e-01, 1.2e-03)
};
\addplot[dotted,mark=triangle*,mark options=solid,black] plot coordinates {
(8.000e-01, 2.9e-01)
(4.000e-01, 7.8e-02)
(2.000e-01, 2.0e-02)
(1.000e-01, 5.1e-03)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 2.2e-01)
(4.000e-01, 7.0e-02)
(2.000e-01, 2.0e-02)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Neumann problem with respect to $\epsilon$ for
Case 2, as labelled.}
\label{fig:neumann_case2-plot}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC1M2,
BC2},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 1.3e-01)
(4.000e-01, 3.1e-02)
(2.000e-01, 7.5e-03)
(1.000e-01, 1.8e-03)
(5.000e-02, 4.5e-04)
(2.500e-02, 1.2e-04)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 8.7e-02)
(4.000e-01, 2.9e-02)
(2.000e-01, 7.1e-03)
(1.000e-01, 1.8e-03)
(5.000e-02, 4.3e-04)
(2.500e-02, 1.1e-04)
};
\addplot[dotted,mark=triangle*,mark options=solid,black] plot coordinates {
(8.000e-01, 3.2e-02)
(4.000e-01, 1.3e-02)
(2.000e-01, 3.4e-03)
(1.000e-01, 8.6e-04)
(5.000e-02, 2.1e-04)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 1.2e-01)
(4.000e-01, 2.8e-02)
(2.000e-01, 8.1e-03)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Neumann problem with respect to $\epsilon$ for
Case 3, as labelled.}
\label{fig:neumann_case3-plot}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC1M2,
BC2},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 1.7e-01)
(4.000e-01, 4.4e-02)
(2.000e-01, 1.1e-02)
(1.000e-01, 3.0e-03)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 1.4e-01)
(4.000e-01, 2.0e-02)
(2.000e-01, 4.6e-03)
(1.000e-01, 1.2e-03)
};
\addplot[dotted,mark=triangle*,mark options=solid,black] plot coordinates {
(8.000e-01, 3.4e-01)
(4.000e-01, 5.5e-02)
(2.000e-01, 1.2e-02)
(1.000e-01, 3.1e-03)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 1.7e-01)
(4.000e-01, 4.6e-02)
(2.000e-01, 1.2e-02)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Neumann problem with respect to $\epsilon$ for
Case 4, as labelled.}
\label{fig:neumann_case4-plot}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC1M2,
BC2},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 1.8e-01)
(4.000e-01, 5.2e-02)
(2.000e-01, 1.5e-02)
(1.000e-01, 4.3e-03)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 1.5e-01)
(4.000e-01, 2.2e-02)
(2.000e-01, 1.2e-02)
(1.000e-01, 5.8e-03)
(5.000e-02, 2.8e-03)
};
\addplot[dotted,mark=triangle*,mark options=solid,black] plot coordinates {
(8.000e-01, 3.5e-01)
(4.000e-01, 6.3e-02)
(2.000e-01, 1.6e-02)
(1.000e-01, 4.6e-03)
(5.000e-02, 2.4e-03)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 1.8e-01)
(4.000e-01, 5.1e-02)
(2.000e-01, 1.4e-02)
(1.000e-01, 4.7e-03)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^\infty$ errors for the Neumann problem with respect to $\epsilon$
for Case 4, as labelled.}
\label{fig:neumann_case4-plot-2}
\end{figure}
\begin{table}[tbp]
\scriptsize
\centering
\begin{tabular}{rlrlrlrlr}
\toprule
& BC1 & & BC1M1 & & BC1M2 & & BC2 \\
$\epsilon$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ \\
\midrule
\multicolumn{4}{l}{Case 1} \\
0.800 & 3.39\e 1& & 4.77\e 1& & & & 3.09\e 1& \\
0.400 & 9.94\e 2& 1.8 & 1.12\e 1& 2.1 & & & 9.52\e 2& 1.7 \\
0.200 & 2.57\e 2& 2.0 & 2.68\e 2& 2.1 & & & 2.57\e 2& 1.9 \\
0.100 & 6.43\e 3& 2.0 & 6.51\e 3& 2.0 & & & & \\
0.050 & 1.61\e 3& 2.0 & 1.59\e 3& 2.0 & & & & \\
0.025 & 4.15\e 4& 2.0 & 3.87\e 4& 2.0 & & & & \\
\midrule
\multicolumn{4}{l}{Case 2} \\
0.800 & 2.46\e 1& & 1.96\e 1& & 2.88\e 1& & 2.22\e 1& \\
0.400 & 7.30\e 2& 1.8 & 2.58\e 2& 2.9 & 7.81\e 2& 1.9 & 6.99\e 2& 1.7 \\
0.200 & 1.94\e 2& 1.9 & 5.21\e 3& 2.3 & 1.96\e 2& 2.0 & 1.95\e 2& 1.8 \\
0.100 & 5.16\e 3& 1.9 & 1.20\e 3& 2.1 & 5.10\e 3& 1.9 & & \\
\midrule
\multicolumn{4}{l}{Case 3} \\
0.800 & 1.27\e 1& & 8.74\e 2& & 3.16\e 2& & 1.18\e 1& \\
0.400 & 3.12\e 2& 2.0 & 2.85\e 2& 1.6 & 1.28\e 2& 1.3 & 2.82\e 2& 2.1 \\
0.200 & 7.48\e 3& 2.1 & 7.08\e 3& 2.0 & 3.40\e 3& 1.9 & 8.13\e 3& 1.8 \\
0.100 & 1.81\e 3& 2.0 & 1.75\e 3& 2.0 & 8.58\e 4& 2.0 & & \\
0.050 & 4.48\e 4& 2.0 & 4.32\e 4& 2.0 & 2.12\e 4& 2.0 & & \\
0.025 & 1.15\e 4& 2.0 & 1.06\e 4& 2.0 & & & & \\
\midrule
\multicolumn{4}{l}{Case 4} \\
0.800 & 1.71\e 1& & 1.38\e 1& & 3.39\e 1& & 1.74\e 1& \\
0.400 & 4.42\e 2& 2.0 & 2.04\e 2& 2.8 & 5.51\e 2& 2.6 & 4.61\e 2& 1.9 \\
0.200 & 1.14\e 2& 2.0 & 4.58\e 3& 2.2 & 1.24\e 2& 2.2 & 1.20\e 2& 1.9 \\
0.100 & 2.95\e 3& 1.9 & 1.18\e 3& 2.0 & 3.09\e 3& 2.0 & & \\
\bottomrule
\end{tabular}
\caption{The $L^2$ error for the Neumann problem as a function of $\epsilon$
for all cases. All results are calculated with $n=8192$ in each direction
on uniform grids. Except for the scheme BC1M2 in Case 1, which was not
simulated, blank results indicate that the solutions require even finer
grids to converge.}
\label{tab:neumann_error_per_eps}
\end{table}
\begin{table}[tbp]
\scriptsize
\centering
\begin{tabular}{rlrlrlrlr}
\toprule
& BC1 & & BC1M1 & & BC1M2 & & BC2 & \\
$\epsilon$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ \\
\midrule
0.800 & 1.80\e 1& & 1.46\e 1& & 3.54\e 1& & 1.76\e 1& \\
0.400 & 5.17\e 2& 1.8 & 2.24\e 2& 2.7 & 6.32\e 2& 2.5 & 5.08\e 2& 1.8 \\
0.200 & 1.48\e 2& 1.8 & 1.18\e 2& 0.9 & 1.59\e 2& 2.0 & 1.44\e 2& 1.8 \\
0.100 & 4.29\e 3& 1.8 & 5.77\e 3& 1.0 & 4.56\e 3& 1.8 & 4.72\e 3& 1.6 \\
0.050 & & & 2.79\e 3& 1.0 & 2.40\e 3& 0.9 & & \\
\bottomrule
\end{tabular}
\caption{The $L^\infty$ error for the Neumann problem as a function of
$\epsilon$ for Case 4. All results are calculated with $n=8192$ in each
direction on uniform grids. Blank results indicate that the solutions
require even finer grids to converge.}
\label{tab:neumann_error_per_eps_infty}
\end{table}
\clearpage
\subsection{Robin boundary conditions}
Now consider the Poisson equation with Robin boundary conditions,
\[
\begin{alignedat}{3}
\boldsymbol\Delta u &= f && \ensuremath{\textup{in}\ } D, \\
\vct n\cdot\grad u &= k(u-g) \quad && \ensuremath{\textup{on}\ } \partial D.
\end{alignedat}
\]
As in the previous section, we solve the DDM equation
\[
\div\left(\phi\grad u\right) + \text{BC} = \phi f,
\]
using BC1, BC2, BC1M and BC2M.
\subsubsection{Case 1}
Consider the case where $D$ is a circle of radius $R=1$ centred at $(0,0)$,
and where the analytic solution to the Poisson equation in $D$ is
\[
u_{\text{an}}(x,y) = \frac{1}{4}\left( x^2 + y^2 \right).
\]
This corresponds to $f = 1$,
\[
g = \frac{1}{2} \left(\frac{1}{2} - \frac{1}{k}\right),
\]
and $\boldsymbol\Delta_{\text s} u_{\text{an}} = 0$. We will consider the case when $k=-1$, thus
$g=3/4$.
\subsubsection{Case 2}
Again let $D$ be the circle at $(0,0)$ with radius $R=1$, but now consider the
case where the analytic solution is
\[
u_{\text{an}}(x,y) = y\left(x^2 + y^2\right),
\]
which corresponds to
\begin{align*}
f &= -8y, \\
g &= y\left( 1 - \frac 3 k \right),
\end{align*}
and
\[
\boldsymbol\Delta_{\text s} u_{\text{an}} = -y,
\]
Again let $k=-1$ so that $g=4y$. Similar to the Neumann case 3, $g$ is
extended constantly in the normal direction in the DDM equations.
\subsubsection{Case 3}
For the final Robin case we let $D=[-1,1]^2$, and we consider a case that
corresponds to the Neumann Case 4 where the analytic solution is
\[
u_{\text{an}}(x,y) = e^r,
\]
where $r = \frac{x^2 + y^2}{4}$. This corresponds to
\[
f = (r+1) e^r.
\]
The boundary function $g$ and the surface Laplacian of the analytic function
along the boundary are
\begin{align*}
g &= \frac 3 2 e^{\frac{1 + \xi^2}{4}}, \\
\boldsymbol\Delta_{\text s} u_\text{an} &= \frac 1 4\left(\xi^2 + 2\right)e^{\frac{1 + \xi^2}{4}},
\end{align*}
where $\xi \equiv x$ along the bottom and top boundaries, and $\xi \equiv y$
along the left and right boundaries.
\subsubsection{Results}
The convergence results calculated with the $L^2$ norm are presented in
\cref{fig:robin_cases-plot1,fig:robin_cases-plot2,fig:robin_cases-plot3} and
\cref{tab:robin_error_per_eps}.
\Cref{fig:robin_cases-plot2-2,tab:robin_error_per_eps_infty} shows the results
for Case 2 where the $L^\infty$ norm is used. Again the results indicate that
BC1M1 performs better than BC1, although both methods are second-order
accurate, as predicted by our analysis. The results also show that BC1 gives
better results than BC2, which is approximately first-order accurate as also
predicted by theory. Further, as in the Neumann case, BC2 is seen to require
very fine grids to converge. For small $\epsilon$, the requirement exceeds our
finest grid.
The modified BC2M1 and BC2M2 schemes are also tested. The results with BC2M2
are almost indistinguishable from the results with BC2M1, so only the latter
results are shown in the following figures. All results are listed in
\cref{tab:robin_error_per_eps} and \cref{tab:robin_error_per_eps_infty}. The
BC2M schemes are shown to perform better than the BC2 scheme, but they also
require very fine grids to converge. Further, the orders of accuracy of BC2M1
and BC2M2 seem to deteriorate somewhat at the smallest values of $\epsilon$.
As discussed in \cref{sec:dda_robin_bc2m}, this could be due to the influence
of higher order terms in the expansion, or the amplification of error and is
under study.
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC2,
BC2M1,},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 2.1e-01)
(4.000e-01, 4.4e-02)
(2.000e-01, 9.0e-03)
(1.000e-01, 2.0e-03)
(5.000e-02, 4.6e-04)
(2.500e-02, 1.2e-04)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 1.2e-01)
(4.000e-01, 2.7e-02)
(2.000e-01, 6.4e-03)
(1.000e-01, 1.6e-03)
(5.000e-02, 3.8e-04)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 1.7e-01)
(4.000e-01, 5.2e-02)
(2.000e-01, 2.2e-02)
(1.000e-01, 1.0e-02)
};
\addplot[dashed,mark=square*,mark options=solid,green] plot coordinates {
(8.000e-01, 1.14e-01)
(4.000e-01, 2.29e-02)
(2.000e-01, 6.87e-03)
(1.000e-01, 2.54e-03)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Robin problem with respect to $\epsilon$ for
Case 1, as labelled.}
\label{fig:robin_cases-plot1}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC2,
BC2M1,},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 4.21e-01)
(4.000e-01, 1.02e-01)
(2.000e-01, 2.19e-02)
(1.000e-01, 4.77e-03)
(5.000e-02, 1.09e-03)
(2.500e-02, 2.57e-04)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 1.27e-1)
(4.000e-01, 4.87e-2)
(2.000e-01, 1.30e-2)
(1.000e-01, 3.27e-3)
(5.000e-02, 8.15e-4)
(2.500e-02, 2.04e-4)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 3.84e-1)
(4.000e-01, 9.16e-2)
(2.000e-01, 2.51e-2)
(1.000e-01, 9.13e-3)
};
\addplot[dashed,mark=square*,mark options=solid,green] plot coordinates {
(8.000e-01, 3.48e-1)
(4.000e-01, 7.24e-2)
(2.000e-01, 1.56e-2)
(1.000e-01, 4.65e-3)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Robin problem with respect to $\epsilon$ for
Case 2, as labelled.}
\label{fig:robin_cases-plot2}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC2,
BC2M1,},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 3.80e-1)
(4.000e-01, 9.39e-2)
(2.000e-01, 2.14e-2)
(1.000e-01, 4.90e-3)
(0.500e-01, 1.15e-3)
(0.250e-01, 2.78e-4)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 1.36e-1)
(4.000e-01, 4.30e-2)
(2.000e-01, 1.06e-2)
(1.000e-01, 2.53e-3)
(5.000e-02, 6.09e-4)
(2.500e-02, 1.49e-4)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 3.24e-1)
(4.000e-01, 7.49e-2)
(2.000e-01, 1.91e-2)
(1.000e-01, 6.98e-3)
};
\addplot[dashed,mark=square*,mark options=solid,green] plot coordinates {
(8.000e-01, 3.04e-1)
(4.000e-01, 6.94e-2)
(2.000e-01, 1.72e-2)
(1.000e-01, 6.40e-3)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^\infty$ errors for the Robin problem with respect to $\epsilon$
for Case 2, as labelled.}
\label{fig:robin_cases-plot2-2}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={Interface width, $\epsilon$},
ylabel={$E_\epsilon$},
x dir=reverse,
width=0.8\textwidth,
legend entries={BC1,
BC1M1,
BC2,
BC2M1,},
legend cell align=left,
legend style={column sep=0.5em,draw=white},
]
\addplot[solid,mark=*,black] plot coordinates {
(8.000e-01, 7.9e-02)
(4.000e-01, 1.6e-02)
(2.000e-01, 3.7e-03)
(1.000e-01, 9.0e-04)
};
\addplot[dashed,mark=square*,mark options=solid,black] plot coordinates {
(8.000e-01, 3.6e-02)
(4.000e-01, 7.4e-03)
(2.000e-01, 1.7e-03)
(1.000e-01, 4.3e-04)
};
\addplot[solid,mark=*,green] plot coordinates {
(8.000e-01, 8.2e-01)
(4.000e-01, 1.9e-02)
(2.000e-01, 5.9e-03)
};
\addplot[dashed,mark=square*,mark options=solid,green] plot coordinates {
(8.000e-01, 6.75e-02)
(4.000e-01, 1.27e-02)
(2.000e-01, 2.78e-03)
};
\end{loglogaxis}
\end{tikzpicture}
\caption{$L^2$ errors for the Robin problem with respect to $\epsilon$ for
Case 3, as labelled.}
\label{fig:robin_cases-plot3}
\end{figure}
\begin{table}[tbp]
\tiny
\centering
\begin{tabular}{clrlrlrlrlr}
\toprule
& BC1 & & BC1M1 & & BC2 & & BC2M1 & & BC2M2\\
$\epsilon$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ \\
\midrule
\multicolumn{4}{l}{Case 1} \\
0.800&2.11\e 1& &1.20\e 1& &1.70\e 1& &1.14\e 1& &1.15\e 1\\
0.400&4.40\e 2&2.3 &2.72\e 2&2.1 &5.21\e 2&1.7&2.29\e 2&2.3&2.31\e 2 &2.3\\
0.200&8.99\e 3&2.3 &6.42\e 3&2.1 &2.18\e 2&1.3&6.87\e 3&1.7&6.92\e 3 &1.7\\
0.100&1.95\e 3&2.2 &1.57\e 3&2.0 &1.03\e 2&1.1&2.54\e 3&1.4&2.55\e 3 &1.4\\
0.050&4.57\e 4&2.1 &3.79\e 4&2.0\\
0.025&1.23\e 4&1.9\\
\midrule
\multicolumn{4}{l}{Case 2} \\
0.800&4.21\e 1& &1.27\e 1& &3.84\e 1& &3.48\e 1& &3.49\e 1& \\
0.400&1.02\e 1&2.0&4.87\e 2&1.4&9.16\e 2&2.1 &7.24\e 2&2.3&7.26\e 2&2.3\\
0.200&2.19\e 2&2.2&1.30\e 2&1.9&2.51\e 2&1.9 &1.56\e 2&2.2&1.57\e 2&2.2\\
0.100&4.77\e 3&2.2&3.27\e 3&2.0&9.13\e 3&1.5 &4.65\e 3&1.7&4.67\e 3&1.7\\
0.050&1.09\e 3&2.1&8.15\e 4&2.0\\
0.025&2.57\e 4&2.1&2.04\e 4&2.0\\
\midrule
\multicolumn{4}{l}{Case 3} \\
0.800&7.89\e 2& &3.60\e 2& &8.23\e 2& &6.75\e 2& &6.80\e 2& \\
0.400&1.64\e 2&2.3&7.38\e 3&2.3 &1.89\e 2&2.2&1.27\e 2&2.3&1.28\e 2&2.4\\
0.200&3.70\e 3&2.2&1.71\e 3&2.1 &5.90\e 3&1.7&2.78\e 3&2.2&2.81\e 3&2.2\\
0.100&9.04\e 4&2.0&4.28\e 4&2.0\\
\bottomrule
\end{tabular}
\caption{The $L^2$ error for the Robin problem as a function of $\epsilon$
for all cases. All results are calculated with $n=8192$ in each direction
on uniform grids, except for Case 3 with BC2, where the results are
calculated with $n=4096$ in each direction. Blank results indicate that
the solutions require even finer grids to converge.}
\label{tab:robin_error_per_eps}
\end{table}
\begin{table}[tbp]
\tiny
\centering
\begin{tabular}{clrlrlrlrlr}
\toprule
& BC1 & & BC1M1 & & BC2 & & BC2M1 & & BC2M2 \\
$\epsilon$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ & $E$ & $k$ \\
\midrule
\multicolumn{4}{l}{Case 2} \\
0.800&3.80\e 1& &1.36\e 1& &3.24\e 1& &3.04\e 1& &3.05\e 1& \\
0.400&9.39\e 2&2.0&4.30\e 2&1.7&7.49\e 2&2.1 &6.94\e 2&2.1&6.95\e 2&2.1\\
0.200&2.14\e 2&2.1&1.06\e 2&2.0&1.91\e 2&2.0 &1.72\e 2&2.0&1.73\e 2&2.0\\
0.100&4.90\e 3&2.1&2.53\e 3&2.1&6.98\e 3&1.5 &6.40\e 3&1.4&6.42\e 3&1.4\\
0.050&1.15\e 3&2.1&6.09\e 4&2.1\\
0.025&2.78\e 4&2.1&1.49\e 4&2.0\\
\bottomrule
\end{tabular}
\caption{The $L^\infty$ error for the Robin problem as a function of
$\epsilon$ for Case 2. All results are calculated with $n=8192$ in each
direction on uniform grids. Blank results indicate that the solutions
require even finer grids to converge.}
\label{tab:robin_error_per_eps_infty}
\end{table}
\Cref{fig:plot_robin_case1} shows a plot of the solutions of Case 1 with
$\epsilon=0.2$ at $y=0$. The plot shows the solutions with BC1 (black dashed),
BC1M1 (black doted), BC2 (blue dashed) and BC2M1 (blue dotted). We see that
the solutions with the modified schemes BC1M1 and BC2M1 perform better than the
corresponding schemes with BC1 and BC2.
\begin{figure}[tbp]
\centering
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{plot_full}
\caption{}
\label{fig:plot_full}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{plot_zoom}
\caption{}
\label{fig:plot_zoom}
\end{subfigure}
\caption{A plot of the solutions of Case 1 through $y=0$. The solutions
with BC2M1 and BC2M2 are indistinguishable, so only BC2M1 is shown. (a)
Shows the full slice, where the domain boundary is depicted as thin
vertical lines at $x=\pm1$. (b) A zoom-in that shows the solutions near
the left boundary.}
\label{fig:plot_robin_case1}
\end{figure}
\clearpage
\section{Conclusion}
\label{sec:conclusion}
We have performed a matched asymptotic analysis of the DDM for the
Poisson equation with Robin boundary conditions and for a steady
reaction-diffusion equation with Neumann boundary conditions. Our analysis
shows that for certain choices of the boundary condition approximations, the
DDM is second-order accurate in the interface thickness $\epsilon$. However,
for other choices the DDM is only first-order accurate. This is confirmed
numerically and helps to explain why the choice of boundary-condition
approximation is important for rapid global convergence and high accuracy.
This helps to explain why the choice of boundary-condition approximation is
important for rapid global convergence and high accuracy. In particular, the
boundary condition BC1, which arises from representing the surface delta
function as $|\nabla\phi|$, is seen to give rise to a second-order
approximation for both the Neumann and Robin boundary conditions and thus is
perhaps the most reliable choice. The boundary condition BC2, which arises
from approximating the surface delta function as $\epsilon|\nabla\phi|^2$
yields a second-order accurate approximation for the Neumann problem, but
only first-order accuracy for the Robin problem. In addition, BC2 requires
very fine meshes to converge.
Our analysis also suggests correction terms that may be added to yield a more
accurate diffuse-domain method. We have presented several techniques for
obtaining second-order boundary conditions and performed numerical
simulations that confirm the predicted accuracy, although the order of
accuracy may deteriorate at the smallest values of $\epsilon$ possibly due to
amplification errors associated with conditioning of the system or the
influence of higher order terms in the asymptotic expansion. This is
currently under study. Further, the correction terms do not improve the mesh
requirements for convergence.
A common feature of the correction terms is that the interface thickness must
be sufficiently small in order for the DDM to remain an elliptic equation.
In addition, one choice of boundary condition involves the use of the surface
Laplacian of the solution, which could in principle lead to faster asymptotic
convergence since it directly cancels terms in the inner expansion of the
asymptotic matching. However, the extension of this term outside the domain
of interest can cause the loss of ellipticity of the DDM. As such, this is
an intriguing but not a practical scheme. Nevertheless, as a proof of
principle, we still considered the effect of this term, however, by using the
surface Laplacian of the analytic solution in the DDM. We found that this
choice gave the smallest errors in nearly all the cases considered. By
incorporating different extensions of the boundary conditions in the exterior
of the domain that automatically guarantee ellipticity, we aim to make this
method practical. This is the subject of future investigations.
We plan to extend our analysis to the Dirichlet problem where the boundary
condition approximations considered by Li et al.\ \cite{Li09} seem only to
yield first-order accuracy \cite{Franz12,Reuter12}. Our asymptotic analysis
thus has the potential to identify correction terms that can be used to
generate second-order accurate diffuse-domain methods for the Dirichlet
problem.
\medskip
{\bf Acknowledgement.}
KYL acknowledges support from the Fulbright foundation for a Visiting
Researcher Grant to fund a stay at the University of California, Irvine. KYL
also acknowledges support from Statoil and GDF SUEZ, and the Research Council
of Norway (193062/S60) for the research project Enabling low emission LNG
systems. JL acknowledges support from the National Science Foundation,
Division of Mathematical Sciences, and the National Institute of Health through
grant P50GM76516 for a Center of Excellence in Systems Biology at the
University of California, Irvine. The authors gratefully thank Bernhard Müller
(NTNU) and Svend Tollak Munkejord (SINTEF Energy Research) for helpful
discussions and for feedback on the manuscript. The authors also wish to thank
the anonymous reviewers for comments that greatly improved the manuscript.
\medskip
\bibliographystyle{siam}
|
\section{Introduction}
In \cite{sa-th:laa} it is shown that a certain class of groups acting on compact ultrametric spaces, the so-called
dually contracting local similarity groups, are $l^2$--invisible. The latter means that group homology with group von Neumann
algebra coefficients vanishes in every dimension, i.e.
\[H_k\big(G,\mathcal{N}(G)\big)=0\]
for all $k\geq 0$ where $\mathcal{N}(G)$ denotes the group von Neumann algebra of $G$. If $G$ is of type $F_\infty$, i.e.~there is
a classifying space for $G$ with finitely many cells in each dimension, then this is equivalent to
\[H_k\big(G,l^2(G)\big)=0\]
for all $k\geq 0$ by \cite{lue:lta}*{Lemmas 6.98 on p.~286 and 12.3 on p.~438}.
In \cite{thu:ogatfp} the author proposed to study fundamental groups of categories naturally associated to operads. This class
of groups, called operad groups, contains a lot of Thompson-like groups already existent in the literature. Among these are the
above mentioned local similarity groups (see \cite{thu:ogatfp}*{Subsection 3.5}).
This article is mainly concerned with generalizing the results of \cite{sa-th:laa} to the setting of symmetric operad groups
which form a much larger class of groups. The proof in \cite{sa-th:laa} consisted of constructing a suitable simplicial complex
on which the group in question acts and then applying a spectral sequence associated to this action which computes the
homology of the group in terms of the homology of the stabilizer subgroups. The proof in the case of operad groups goes exactly
the same way. However, it is a priori unclear how to construct the simplicial complex. The reason is the following: A local
similarity group is defined as a representation, i.e.~as a group of homeomorphisms of a compact ultrametric space. This space is
used to construct the simplicial complex as a poset of partitions of this space. The case of operad groups is more abstract. A priori,
there is no canonical space comparable to these ultrametric spaces on which an operad group acts. However, these spaces, called
limit spaces, are conjectured to exist if the operad satisfies the calculus of fractions
(see \cite{thu:ogatfp}*{Subsection 3.3} for the latter notion). We don't use these limit spaces here. Instead, we will take
the conjectured correspondence between calculus of fractions operads and their limit spaces as a motivation to mimic the necessary
notions for the construction of the desired simplicial complex in terms of the operad itself.
As in \cite{sa-th:laa}, we briefly want to discuss the relationship between these results and Gromov's
\emph{Zero-in-the-spectrum conjecture} (see \cite{gro:lrm}). The algebraic version of this conjecture states that if $\Gamma=\pi_1(M)$
is the fundamental group of a closed aspherical Riemannian manifold, then there always exists a dimension $p\geq 0$ such
that $H_p(\Gamma,\mathcal{N}\Gamma)\neq 0$ or equivalently $H_p(\Gamma,l^2\Gamma)\neq 0$. Conjecturally, the fundamental groups of
closed aspherical manifolds are precisely the Poincar\'e duality groups $G$ of type $F$, i.e.~there is a compact classifying space
for $G$ and a natural number $n\geq 0$ such that
\[H^i(G,\mathbb{Z} G)=\begin{cases}0&\text{if }i\neq n\\\mathbb{Z}&\text{if }i=n\end{cases}\]
(see \cite{dav:pdg}). Dropping Poincar\'e duality and relaxing type $F$ to type $F_\infty$, we arrive at a more general question which
has been posed by L\"uck in \cite{lue:lta}*{Remark 12.4 on p.~440}: If $G$ is a group of type $F_\infty$, does there always exist a $p$
with $H_p(G,\mathcal{N} G)\neq 0$? In \cite{thu:ogatfp} we discuss conditions for operads which imply that the associated operad groups
are of type $F_\infty$. Combining this with the results in the present article, we obtain a large class of groups of type $F_\infty$
which are also $l^2$--invisible. This class contains the well-known symmetric Thompson group $V$ and consequently, L\"uck's question
has to be answered in the negative. Unfortunately, all these groups $G$ are neither of type $F$ nor satisfy Poincar\'e duality since,
as another corollary of our main theorem (Theorem \ref{20869}), we can show $H^k(G,\mathbb{Z} G)=0$ for all $k\geq 0$.
\subsection{Prerequisites}
The present article is based on Sections 2 and 3 of \cite{thu:ogatfp}.
\subsection{Notation and Conventions}
When $f\colon A\rightarrow B$ and $g\colon B\rightarrow C$ are two composable arrows, we write $f*g$ for the composition $A\rightarrow C$
instead of the usual notation $g\circ f$. Consequently, it is often better to plug in arguments from the left. When we do this,
we use the notation $x{\triangleright} f$ for the evaluation of $f$ at $x$. However, we won't entirely drop the usual notation $f(x)$ and use
both notations side by side. Objects of type $\operatorname{Aut}(X)$ will be made into a group by the definition $fg=f\cdot g:=f*g$.
Conversely, a group $G$ is considered as a groupoid with one object and arrows the elements in $G$ together with the composition
$f*g:=f\cdot g$.
\subsection{Acknowledgments}
I want to thank my PhD adviser Roman Sauer for the opportunity to pursue mathematics, for his guidance, encouragement and
support over the last few years.
I also gratefully acknowledge financial support by the DFG grants 1661/3-1 and 1661/3-2.
\section{Statement of the main theorem}
\begin{defi}
Let $\mathcal{M} G$ be a $\mathbb{Z} G$--module for every group $G$. We call $\mathcal{M}$
\begin{itemize}
\item {\it K\"unneth} if for every two groups $G_1,G_2$ and $n_1,n_2\in\mathbb{Z}$ with $n_i\geq -1$ the following is satisfied:
\[\left.\begin{array}{r}\forall_{k\leq n_1}\ H_k(G_1,\mathcal{M} G_1)=0\\\forall_{k\leq n_2}\ H_k(G_2,\mathcal{M} G_2)=0\end{array}
\right\}\ \Longrightarrow\ \forall_{k\leq n}\ H_k(G,\mathcal{M} G)=0\]
where $G:=G_1\times G_2$ and $n:=n_1+n_2+1$.
\item {\it inductive} if whenever $H$ and $G$ are groups with $H$ a subgroup of $G$ and $k\geq 0$, then we have that
\[H_k(H,\mathcal{M} H)=0\hspace{3mm}\text{implies}\hspace{3mm}H_k(H,\mathcal{M} G)=0\]
\end{itemize}
Let $\mathfrak{P}$ be a property of groups. Then we say that $\mathcal{M}$ is $\mathfrak{P}$--K\"unneth if the
property K\"unneth has to be satisfied only for $\mathfrak{P}$--groups $G_1,G_2$. We say that $\mathcal{M}$ is $\mathfrak{P}$--inductive
if the property inductive has to be satisfied only for $\mathfrak{P}$--subgroups $H$ of the arbitrary group $G$.
Furthermore, one can formulate these two properties also in the cohomological case.
\end{defi}
\begin{defi}
Let $\mathcal{O}$ be a planar or symmetric or braided operad and $X$ an object in $\mathcal{S}(\mathcal{O})$. We say that $X$ is
\begin{itemize}
\item {\it split} if there are objects $A_1,A_2,A_3$ and an arrow $A_1\otimes X\otimes A_2\otimes X\otimes A_3\rightarrow X$
in $\mathcal{S}(\mathcal{O})$.
\item {\it progressive} if for every arrow $Y\rightarrow X$ there are objects $A_1,A_2$ and an arrow
$A_1\otimes X\otimes A_2\rightarrow Y$ such that the coordinates of $X$ are connected to only one
operation in this arrow (see Figure \ref{57690}).
\end{itemize}
\end{defi}
\begin{figure}[!ht]
\begin{center}\begin{tikzpicture}[scale=0.65]
\draw (0,4.6) -- +(-1,0);
\draw (0,5.4) -- +(-1,0);
\draw (0,4.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw (0,3) -- +(-1,0);
\draw (0,2.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw (0,0.4) -- +(-1,0);
\draw (0,1) -- +(-1,0);
\draw (0,1.6) -- +(-1,0);
\draw (0,0.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw[dotted] (-0.8,0) -- +(0,6);
\draw[white, line width=4pt] (-1,1.6) .. controls (-2,1.6) and (-3,3.5) .. (-4,3.5);
\draw (-1,1.6) .. controls (-2,1.6) and (-3,3.5) .. (-4,3.5);
\draw[white, line width=4pt] (-1,1) .. controls (-2,1) and (-3,0.5) .. (-4,0.5);
\draw (-1,1) .. controls (-2,1) and (-3,0.5) .. (-4,0.5);
\draw[white, line width=4pt] (-1,4.6) .. controls (-2,4.6) and (-3,1.5) .. (-4,1.5);
\draw (-1,4.6) .. controls (-2,4.6) and (-3,1.5) .. (-4,1.5);
\draw[white, line width=4pt] (-1,3) .. controls (-2,3) and (-3,5.5) .. (-4,5.5);
\draw (-1,3) .. controls (-2,3) and (-3,5.5) .. (-4,5.5);
\draw[white, line width=4pt] (-1,0.4) .. controls (-2,0.4) and (-3,2.5) .. (-4,2.5);
\draw (-1,0.4) .. controls (-2,0.4) and (-3,2.5) .. (-4,2.5);
\draw[white, line width=4pt] (-1,5.4) .. controls (-2,5.4) and (-3,4.5) .. (-4,4.5);
\draw (-1,5.4) .. controls (-2,5.4) and (-3,4.5) .. (-4,4.5);
\draw (-4.8,5.6) rectangle (-4.2,4.4);
\draw (-4.8,3.6) rectangle (-4.2,2.4);
\draw (-4.8,1.6) rectangle (-4.2,0.4);
\draw (2,5.2) rectangle (2.6,0.8);
\node at (-4.5,5) {$A_1$};
\node at (-4.5,3) {$X$};
\node at (-4.5,1) {$A_2$};
\node at (2.3,3) {$Y$};
\end{tikzpicture}\end{center}
\caption{An arrow $A_1\otimes X\otimes A_2\rightarrow Y$ such that $X$ is only connected to one operation.}\label{57690}
\end{figure}
\begin{rem}\label{92437}
If $X$ is just a single color, then $X$ is split if and only if there is an operation with output color $X$ and and at
least two inputs of color $X$. If $\mathcal{O}$ is monochromatic and $X\neq I$ is an object of $\mathcal{S}(\mathcal{O})$, then $X$ is
split if and only if there is at least one operation in $\mathcal{O}$ with at least two inputs. So in the monochromatic case,
the condition split is in fact a property of $\mathcal{O}$.
\end{rem}
\begin{rem}\label{71138}
If $X$ is just a single color, then $X$ is progressive if and only if for every operation $\theta$ with output
color $X$ there is another operation $\phi$ with at least one input of color $X$ and at least one input of
$\theta$ has the same color as the output of $\phi$. Now assume that $\mathcal{O}$ is monochromatic. Then an object
$X\neq I$ in $\mathcal{S}(\mathcal{O})$ (which is just a natural number $X>0$, e.g.~$X=3$) is progressive if and only if there
is an operation in $\mathcal{O}$ with at least $X$ inputs (e.g.~$3$ inputs). Note that $X=1$ is always progressive
in the monochromatic case.
\end{rem}
\begin{thm}\label{20869}
Let $\mathcal{O}$ be a planar or symmetric operad which satisfies the calculus of fractions.
Let $\mathcal{M}$ be a coefficient system which is K\"unneth and inductive. Let $X$ be a split progressive object of
$\mathcal{S}(\mathcal{O})$. Set $\Gamma:=\pi_1(\mathcal{O},X)$. Then
\[H_0(\Gamma,\mathcal{M}\Gamma)=0\hspace{5mm}\Longrightarrow\hspace{5mm}\forall_{k\geq 0}\ H_k(\Gamma,\mathcal{M}\Gamma)=0\]
The same is true for cohomology.\\
More generally, let $\mathfrak{P}$ be a property of groups which is closed under taking products.
Then the statement is true also for coefficient systems $\mathcal{M}$ which are only $\mathfrak{P}$--K\"unneth and
$\mathfrak{P}$--inductive, provided that $\Gamma$ satisfies $\mathfrak{P}$.
\end{thm}
\begin{rem}\label{50091}
Let $X,Y$ be objects in $\mathcal{S}(\mathcal{O})$. Generalizing the notion of progressiveness, we say that $X$ is
$Y$--progressive if for every arrow $Z\rightarrow X$ there is an arrow $A_1\otimes Y\otimes A_2\rightarrow Z$
and the coordinates of $Y$ are connected to only one operation in this arrow (call this the \emph{link condition}).
In particular, there is an arrow $A_1\otimes Y\otimes A_2\rightarrow X$.
With this notion, we can formulate a slightly more general version of Theorem \ref{20869}: Let $\mathcal{O},\mathfrak{P},\mathcal{M}$ be
as in the theorem. Let $X$ be an object of $\mathcal{S}(\mathcal{O})$ and set $\Gamma=\pi_1(\mathcal{O},X)$. Assume there is a
split object $Y$ such that $X$ is $Y$--progressive, $\Upsilon:=\pi_1(\mathcal{O},Y)$ satisfies $\mathfrak{P}$ and
$H_0(\Upsilon,\mathcal{M}\Upsilon)=0$. Then $H_k(\Gamma,\mathcal{M}\Gamma)=0$ for each $k\geq 0$. The same is true for
cohomology.
\end{rem}
\section{Proof of the main theorem}\label{96238}
We start with two general lemmas concerning the calculus of fractions.
\begin{lem}\label{25010}
Let $\mathcal{C}$ be a category satisfying the calculus of fractions. Then two square fillings of a given span
can be combined to a common square filling. This means: Let $x,y$ be two arrows with the
same codomain and assume having two square fillings as in the diagram
\begin{displaymath}\xymatrix@-5pt{
\bullet\ar[dd]_x & & \bullet\ar[ll]_i\ar[dd]^h \\
& \bullet\ar[ul]_j\ar[dr]^g & \\
\bullet & & \bullet\ar[ll]^y
}\end{displaymath}
then we can complete this diagram to the commutative diagram
\begin{displaymath}\xymatrix@-5pt{
& & & \bullet\ar@{-->}@/_10pt/[dlll]_\alpha\ar@{-->}[dl]^\delta
\ar@{-->}@/_8pt/[ddll]_\epsilon\ar@{-->}@/^10pt/[dddl]^\beta \\
\bullet\ar[dd] & & \bullet\ar[ll]\ar[dd] & \\
& \bullet\ar[ul]\ar[dr] & & \\
\bullet & & \bullet\ar[ll] &
}\end{displaymath}
\end{lem}
\begin{proof}
Let $c,d$ be a square filling of $a:=ix=hy,b:=jx=gy$, i.e.~$ca=db$. Then $ch$ and $dg$ are two parallel
arrows which are coequalized by $y$, i.e.~$(ch)y=(dg)y$. By the equalization property we find an equalizing
arrow $k$ with $k(ch)=k(dg)$. By the same reasoning we find an arrow $l$ with $l(ci)=l(dj)$. Let $m,n$ be a
square filling of $l,k$, i.e.~$ml=nk=:p$. Then one can easily calculate that the arrows
\[\delta=pc\hspace{8mm}\alpha=pci\hspace{8mm}\beta=pch\hspace{8mm}\epsilon=pd\]
fill the diagram as required.
\end{proof}
\begin{lem}\label{69343}
Let $\mathcal{C}$ be a category satisfying the calculus of fractions. Let $\bar{x}$ and $\bar{y}$ be two arrows
$A\rightarrow C$. Assume that there are arrows $x,y\colon A\rightarrow B$ and $a\colon B\rightarrow C$ such that
$xa=\bar{x}$ and $ya=\bar{y}$.
\begin{displaymath}\xymatrix{
C & B\ar[l]^a & A\ar[r]_y\ar@/^10pt/[rr]^{\bar{y}}\ar[l]^x\ar@/_10pt/[ll]_{\bar{x}} & B\ar[r]_a & C
}\end{displaymath}
Then the span $C\xleftarrow{\bar{x}}A\xrightarrow{\bar{y}}C$ is null-homotopic if and only if
the span $B\xleftarrow{x}A\xrightarrow{y}B$ is null-homotopic.
\end{lem}
\begin{proof}
First note that a span like $B\xleftarrow{x}A\xrightarrow{y}B$ is null-homotopic if and only if the
parallel arrows $x$ and $y$ are homotopic. Since $\mathcal{C}$ satisfies the calculus of fractions, this is the
case if and only if there is an equalizing arrow, i.e.~an arrow $d\colon D\rightarrow A$ with $dx=dy$. Now if
$x$ and $y$ are homotopic then clearly also $\bar{x}$ and $\bar{y}$ are homotopic. On the other hand, assume that
$\bar{x}$ and $\bar{y}$ are homotopic and $d\colon D\rightarrow A$ equalizes $\bar{x}$ and $\bar{y}$. Then we
have
\[(dx)a=d(xa)=d\bar{x}=d\bar{y}=d(ya)=(dy)a\]
Then by the equalization property we find an arrow $e\colon E\rightarrow D$ with $e(dx)=e(dy)$. Consequently,
the arrow $ed$ equalizes $x$ and $y$ and thus, $x$ and $y$ are homotopic.
\end{proof}
We now turn to the proof of Theorem \ref{20869}. In the following, let $\mathcal{O}$ be a planar or symmetric
operad satisfying the calculus of fractions with set of colors $C$ and let $\mathcal{S}:=\mathcal{S}(\mathcal{O})$.
\subsection{Marked objects}
Let $c=(c_1,...,c_n)$ be an object of $\mathcal{S}$, i.e.~$c_1,...,c_n$ are colors in $C$.
First we define a marking on $c$ in the symmetric case. It assigns to each coordinate of $c$ a
symbol. A symbol can be assigned several times and not every coordinate has to be marked by a symbol. More precisely,
a marking of $c$ is a set $S$ of symbols together with a subset $I\subset\{1,...,n\}$ and a surjective function
$f\colon I\rightarrow S$. In the planar case, we additionally require the marking to be ordered. This means
that whenever $i{\triangleright} f=j{\triangleright} f$ for $i<j$ then also $i{\triangleright} f=k{\triangleright} f=j{\triangleright} f$ for $i<k<j$.
Let $m_1,m_2$ be two markings of $c$ with symbol sets $S_1,S_2$. We say $m_1\subset m_2$ if there
is a function $i\colon S_1\rightarrow S_2$ and every coordinate which is marked with $s_1\in S_1$ is
also marked with $s_1{\triangleright} i\in S_2$. We say $m_1$ and $m_2$ are equivalent if $m_1\subset m_2$ and $m_2\subset m_1$.
This means that there is a bijection $i\colon S_1\rightarrow S_2$ and a coordinate is marked with
$s_1\in S_1$ if and only if it is marked with $s_1{\triangleright} i\in S_2$. By slight abuse of notation, we identify equivalent
markings and write $m_1=m_2$ if they are equivalent. Then $\subset$ becomes a partial order on the set of markings
on $c$ (see the first paragraph of Subsection \ref{52176}).
\subsection{Marked arrows}
Let $\alpha\colon c\rightarrow d$ be an arrow in $\mathcal{S}$ with objects $c=(c_1,...,c_n)$ and
$d=(d_1,...,d_m)$. A marking on $\alpha$ is a marking on the domain $c$. A comarking on $\alpha$ is a marking
on the codomain $d$. A comarking on $\alpha$ induces a marking on $\alpha$: Let $(\sigma,X)$ be a representative of
$\alpha$ where $\sigma$ is either an identity or a colored permutation,
depending on whether $\mathcal{O}$ is planar or symmetric. Write $X=(X_1,...,X_m)$.
The comarking yields a marking on the operations $X_i$: Mark each input of $X_i$
with the same symbol. Now push the markings through $\sigma$ to obtain a marking on the domain $c$.
Figure \ref{01354} illustrates this procedure. If $m$ is the comarking, then we denote this pull-backed
marking by $\alpha^*(m)$. Observe that this pull-back is functorial, i.e.~we have
\[(\alpha\beta)^*(m)=\alpha^*(\beta^*(m))\]
Furthermore, we have
\[m_1\subset m_2\ \ \Longleftrightarrow\ \ \alpha^*(m_1)\subset\alpha^*(m_2)\]
\begin{figure}[!ht]
\begin{center}\begin{tikzpicture}[scale=0.65]
\draw (-1,5.4) to[out=left,in=right] (-4,5.4);
\draw[white, line width=4pt] (-2.8,2.8) to[out=left,in=right] (-4,3);
\draw (-2.8,2.8) to[out=left,in=right] (-4,3);
\draw[white, line width=4pt] (-2,4.6) to[out=left,in=right] (-4,1.6);
\draw (-2,4.6) to[out=left,in=right] (-4,1.6);
\draw[white, line width=4pt] (-1,1.6) to[out=left,in=right] (-4,1);
\draw (-1,1.6) to[out=left,in=right] (-4,1);
\draw[white, line width=4pt] (-1,1) to[out=left,in=right] (-4,4.6);
\draw (-1,1) to[out=left,in=right] (-4,4.6);
\draw[white, line width=4pt] (-1.6,3) to[out=left,in=right] (-2.8,2.8);
\draw (-1.6,3) to[out=left,in=right] (-2.8,2.8);
\draw (-1,0.4) to[out=left,in=right] (-4,0.4);
\draw (-2,4.6) -- (-1,4.6);
\draw (-1.6,3) -- (-1,3);
\draw (0,4.6) -- +(-1,0);
\draw (0,5.4) -- +(-1,0);
\draw (0,4.1) -- +(0,1.8) -- +(1.8,0.9) node[right]{$\bigstar$} -- +(0,0);
\draw (0,3) -- +(-1,0);
\draw (0,2.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw (0,0.4) -- +(-1,0);
\draw (0,1) -- +(-1,0);
\draw (0,1.6) -- +(-1,0);
\draw (0,0.1) -- +(0,1.8) -- +(1.8,0.9) node[right]{$\diamondsuit$} -- +(0,0);
\draw[dotted] (-0.8,0) -- +(0,6);
\draw (8,5.4) to[out=left,in=right] (5,5.4) node[left]{$\bigstar$};
\draw[white, line width=4pt] (6.2,2.8) to[out=left,in=right] (5,3);
\draw (6.2,2.8) to[out=left,in=right] (5,3);
\draw[white, line width=4pt] (7,4.6) to[out=left,in=right] (5,1.6);
\draw (7,4.6) to[out=left,in=right] (5,1.6) node[left]{$\bigstar$};
\draw[white, line width=4pt] (8,1.6) to[out=left,in=right] (5,1);
\draw (8,1.6) to[out=left,in=right] (5,1) node[left]{$\diamondsuit$};
\draw[white, line width=4pt] (8,1) to[out=left,in=right] (5,4.6);
\draw (8,1) to[out=left,in=right] (5,4.6) node[left]{$\diamondsuit$};
\draw[white, line width=4pt] (7.4,3) to[out=left,in=right] (6.2,2.8);
\draw (7.4,3) to[out=left,in=right] (6.2,2.8);
\draw (8,0.4) to[out=left,in=right] (5,0.4) node[left]{$\diamondsuit$};
\draw (7,4.6) -- (8,4.6);
\draw (7.4,3) -- (8,3);
\draw (9,4.6) -- +(-1,0);
\draw (9,5.4) -- +(-1,0);
\draw (9,4.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw (9,3) -- +(-1,0);
\draw (9,2.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw (9,0.4) -- +(-1,0);
\draw (9,1) -- +(-1,0);
\draw (9,1.6) -- +(-1,0);
\draw (9,0.1) -- +(0,1.8) -- +(1.8,0.9) -- +(0,0);
\draw[dotted] (8.2,0) -- +(0,6);
\end{tikzpicture}\end{center}
\caption{A comarking (left) and the pull-backed marking (right).}\label{01354}
\end{figure}
\vspace{2mm}
\begin{center}
\framebox[1.1\width]{Now fix an object $x$ in $\mathcal{S}$.}
\end{center}
\vspace{2mm}
Let $(\alpha_1,m_1)$ and $(\alpha_2,m_2)$ be two marked arrows with codomain $x$, i.e.~$\alpha_i\colon c_i\rightarrow x$ is
an arrow and $m_i$ is a marking on $c_i$. We write
\[(\alpha_1,m_1)\subset(\alpha_2,m_2)\]
if there is a square filling
\begin{displaymath}\xymatrix{
d\ar@{..>}[r]^{\beta_2}\ar@{..>}[d]_{\beta_1}&c_2\ar[d]^{\alpha_2}\\
c_1\ar[r]_{\alpha_1}&x
}\end{displaymath}
with $\beta_1^*(m_1)\subset\beta_2^*(m_2)$. Observe that then {\it every} square filling satisfies this:
Let $(\gamma_1,\gamma_2)$ be another square filling of $(\alpha_1,\alpha_2)$. Then choose a
common square filling $(\delta_1,\delta_2)$ as in Lemma \ref{25010}. It is not hard to see that the property
$\delta_1^*(m_1)\subset\delta_2^*(m_2)$ is inherited from the square filling $(\beta_1,\beta_2)$. On the other
hand, this forces the property onto the square filling $(\gamma_1,\gamma_2)$, i.e.~we have $\gamma_1^*(m_1)\subset
\gamma_2^*(m_2)$.
\begin{rem}\label{00176}
This observation implies also the following: Let $(\alpha_1,m_1)\subset(\alpha_2,m_2)$
and assume that $\alpha_1=\alpha_2$. Then we necessarily have $m_1\subset m_2$. Indeed, we can
choose $\beta_1=\mathrm{id}=\beta_2$ in the above square filling.
\end{rem}
\begin{prop}\label{53020}
The relation $\subset$ on the set of marked arrows is reflexive and transitive.
\end{prop}
\begin{proof}
Reflexivity is clear. For transitivity assume $(\alpha_1,m_1)\subset(\delta,m)$ and
$(\delta,m)\subset(\alpha_2,m_2)$. Choose two square fillings
\begin{displaymath}\xymatrix{
a_1\ar@{..>}[d]_{\beta_1}\ar@{..>}[r]^{\delta_1}&d\ar[d]^{\delta}&
a_2\ar@{..>}[l]_{\delta_2}\ar@{..>}[d]^{\beta_2}\\
c_1\ar[r]_{\alpha_1}&x&c_2\ar[l]^{\alpha_2}
}\end{displaymath}
with $\beta_1^*(m_1)\subset\delta_1^*(m)$ and $\delta_2^*(m)\subset\beta_2^*(m_2)$. Choose a square filling of
$(\delta_1,\delta_2)$
\begin{displaymath}\xymatrix{
&e\ar[dl]_{\gamma_1}\ar[d]^{\eta}\ar[dr]^{\gamma_2}&\\
a_1\ar@{..>}[d]_{\beta_1}\ar@{..>}[r]_{\delta_1}&d\ar[d]^{\delta}&
a_2\ar@{..>}[l]^{\delta_2}\ar@{..>}[d]^{\beta_2}\\
c_1\ar[r]_{\alpha_1}&x&c_2\ar[l]^{\alpha_2}
}\end{displaymath}
Now we have
\begin{eqnarray*}
(\gamma_1\beta_1)^*(m_1)&=&\gamma_1^*(\beta_1^*(m_1))\\
&\subset &\gamma_1^*(\delta_1^*(m))\\
&=&(\gamma_1\delta_1)^*(m)\\
&=&\eta^*(m)\\
&=&(\gamma_2\delta_2)^*(m)\\
&=&\gamma_2^*(\delta_2^*(m))\\
&\subset &\gamma_2^*(\beta_2^*(m_2))\\
&=&(\gamma_2\beta_2)^*(m_2)
\end{eqnarray*}
This proves $(\alpha_1,m_1)\subset(\alpha_2,m_2)$.
\end{proof}
\subsection{Balls and partitions}\label{52176}
A transitive and reflexive relation $\preccurlyeq$ on a set $Z$ is not a poset in general since $a\preccurlyeq b$ together with
$b\preccurlyeq a$ does not imply $a=b$ in general. We can repair this in the following way: Define $a,b\in Z$ to be equivalent if
$a\preccurlyeq b$ and $b\preccurlyeq a$. This is indeed an equivalence relation because $\preccurlyeq$ is assumed to be reflexive
and transitive. Now if $\mathfrak{a}$ and $\mathfrak{b}$ are two equivalence classes, we write $\mathfrak{a}\leq\mathfrak{b}$
if there are representatives $a$ and $b$ respectively with $a\preccurlyeq b$. One can easily show that then \emph{any} two
representatives satisfy this. Using this, it is not hard to see that $\leq$ is indeed a partial order on the set of equivalence
classes. In particular, we have $\mathfrak{a}=\mathfrak{b}$ if and only if $\mathfrak{a}\leq\mathfrak{b}$ and
$\mathfrak{b}\leq\mathfrak{a}$.
\vspace{2mm}
We want to apply this observation to the reflexive and transitive relation $\subset$ on the set of marked arrows.
We say that two marked arrows $(\alpha_1,m_1)$ and $(\alpha_2,m_2)$ with common codomain $x$ are equivalent if both
$(\alpha_1,m_1)\subset(\alpha_2,m_2)$ and $(\alpha_2,m_2)\subset(\alpha_1,m_1)$ hold. We remark that this
is equivalent to the existence of a square filling
\begin{displaymath}\xymatrix{
d\ar@{..>}[r]^{\beta_2}\ar@{..>}[d]_{\beta_1}&c_2\ar[d]^{\alpha_2}\\
c_1\ar[r]_{\alpha_1}&x
}\end{displaymath}
with $\beta_1^*(m_1)=\beta_2^*(m_2)$ and moreover that {\it every} square filling satisfies this.
\begin{itemize}
\item A {\it semi-partition} is an equivalence class of marked arrows.
\item A {\it partition} is a semi-partition with fully marked domain for some (and therefore for every)
representative of the semi-partition. Here, an object in $\mathcal{S}$ is fully marked if every coordinate
is marked.
\item A {\it multiball} is a semi-partition with an uni-marked domain for some (and therefore for every)
representative of the semi-partition. Here, an object in $\mathcal{S}$ is uni-marked if there is only one
symbol in the marking.
\item A {\it ball} is a semi-partition such that there is a single-marked representative. Here, an object in $\mathcal{S}$ is
single-marked if only one coordinate is marked.
\end{itemize}
Note that these definitions depend on the base point $x$. Following the remarks in the first paragraph, we write
$\mathcal{P}\subset\mathcal{Q}$ for two semi-partitions $\mathcal{P}$ and $\mathcal{Q}$ if there are representatives $p$ of
$\mathcal{P}$ and $q$ of $\mathcal{Q}$ satisfying $p\subset q$. Then for \emph{all} such representatives $p,q$
we have $p\subset q$. It follows that $\subset$ is a partial order on the set of semi-partitions. In particular,
we have $\mathcal{P}=\mathcal{Q}$ if and only if $\mathcal{P}\subset\mathcal{Q}$ and $\mathcal{Q}\subset\mathcal{P}$.
\vspace{2mm}
We now investigate the relationship between semi-partitions and multiballs. Let $\mathcal{P}$ be a semi-partition with
representative $(\alpha,m)$. Picking out a symbol $s$ of $m$ and removing all markings except those with the chosen
symbol $s$ gives a uni-marked arrow $(\alpha,m^s)$. The corresponding equivalence class is a multiball and is
independent of the chosen representative $(\alpha,m)$ in the following sense: If we choose another representative
$(\beta,n)$, then $(\alpha,m)\sim(\beta,n)$ and to the chosen symbol $s$ of $m$ corresponds a unique symbol
$r$ of $n$. Deleting all markings in $n$ except those with the symbol $r$ gives a uni-marked arrow
$(\beta,n^r)$ which is equivalent to $(\alpha,m^s)$. Multiballs arising in this way are called submultiballs of
$\mathcal{P}$ and we write $P\in\mathcal{P}$ for submultiballs. Note that Remark \ref{00176} implies that two submultiballs
$P_1,P_2$ coming from a representative of $\mathcal{P}$ by choosing two different symbols satisfy $P_1\not\subset P_2$ and
$P_2\not\subset P_1$, in particular $P_1\neq P_2$. It follows that there is a canonical bijection between the set
$\{P\in\mathcal{P}\}$ of submultiballs of $\mathcal{P}$ and the set of symbols of $\mathcal{P}$ (which is, by definition, the set of
symbols of the marking of any representative for $\mathcal{P}$). Moreover, any two submultiballs $P_1,P_2\in\mathcal{P}$ with
$P_1\neq P_2$ satisfy the stronger property $(P_1\not\subset P_2)\wedge(P_2\not\subset P_1)$. Equivalently, whenever
$P_1\subset P_2$ or $P_2\subset P_1$, then already $P_1=P_2$.
\begin{prop}\label{45613}
Let $\mathcal{P},\mathcal{Q}$ be semi-partitions, then
\[\mathcal{Q}\subset\mathcal{P}\ \ \Longleftrightarrow\ \ \forall_{Q\in\mathcal{Q}}\ \exists_{P\in\mathcal{P}}\ Q\subset P\]
In particular, $\mathcal{P}=\mathcal{Q}$ if and only if $\{Q\in\mathcal{Q}\}=\{P\in\mathcal{P}\}$.
\end{prop}
\begin{proof}
We first prove the last statement since it is a formal consequence of the previous statement and the remarks preceding
the proposition. First recall that $\mathcal{P}=\mathcal{Q}$ is equivalent to $\mathcal{P}\subset\mathcal{Q}$ and $\mathcal{Q}\subset\mathcal{P}$.
The first statement of the proposition says
that there is a function $i\colon \{Q\in\mathcal{Q}\}\rightarrow\{P\in\mathcal{P}\}$ with the property that $Q\subset Q{\triangleright} i$
for each $Q\in\mathcal{Q}$. Since we also have $\mathcal{P}\subset\mathcal{Q}$, there is another function
$j\colon \{P\in\mathcal{P}\}\rightarrow\{Q\in\mathcal{Q}\}$ with the property that $P\subset P{\triangleright} j$ for each $P\in\mathcal{P}$. We have
\[Q\subset Q{\triangleright} i\subset (Q{\triangleright} i){\triangleright} j=Q{\triangleright}(ij)\]
for all $Q\in\mathcal{Q}$. Since both the left and right side are submultiballs of $\mathcal{Q}$, the remarks preceding
the proposition imply $Q=Q{\triangleright}(ij)$ for all $Q\in\mathcal{Q}$. We then have
\[Q\subset Q{\triangleright} i\subset Q\]
and therefore also $Q=Q{\triangleright} i$ for all $Q\in\mathcal{Q}$. This shows $\{Q\in\mathcal{Q}\}\subset\{P\in\mathcal{P}\}$. With a similar
argument applied to $ji$, we also obtain $\{Q\in\mathcal{Q}\}\supset\{P\in\mathcal{P}\}$. So we have indeed
$\{Q\in\mathcal{Q}\}=\{P\in\mathcal{P}\}$. The converse implication also follows easily from the first
statement of this proposition.
Now let's turn to the first statement: Assume $\mathcal{Q}\subset\mathcal{P}$. By the square filling technique, we know
that we can choose a common arrow $\alpha\colon c\rightarrow x$ with markings $m_\mathcal{Q}\subset m_\mathcal{P}$ such that
$[\alpha,m_\mathcal{Q}]=\mathcal{Q}$ and $[\alpha,m_\mathcal{P}]=\mathcal{P}$. If $Q\in\mathcal{Q}$, then we find a symbol $s_\mathcal{Q}$ of the
marking $m_\mathcal{Q}$ which corresponds to $Q$. But since $m_\mathcal{Q}\subset m_\mathcal{P}$, there is a unique symbol $s_\mathcal{P}$
of the marking $m_\mathcal{P}$ such that the coordinates of $c$ marked by $s_\mathcal{Q}$ are also marked by $s_\mathcal{P}$.
The submultiball obtained from $(\alpha,m_\mathcal{P})$ corresponding to the symbol $s_\mathcal{P}$ is the one we are looking
for.
Conversely, assume that there is a function $i\colon \{Q\in\mathcal{Q}\}\rightarrow\{P\in\mathcal{P}\}$ such that
$Q\subset Q{\triangleright} i$ for every $Q\in\mathcal{Q}$. Using the square filling technique, we find a common arrow
$\alpha\colon c\rightarrow x$ with markings $m_\mathcal{Q},m_\mathcal{P}$ such that $[\alpha,m_\mathcal{Q}]=\mathcal{Q}$ and
$[\alpha,m_\mathcal{P}]=\mathcal{P}$. We want to show $m_\mathcal{Q}\subset m_\mathcal{P}$. Let $s$ be any symbol of
$m_\mathcal{Q}$. To this symbol corresponds exactly one submultiball $Q\in\mathcal{Q}$ such that $Q=[\alpha,m_\mathcal{Q}^s]$
where $m_\mathcal{Q}^s$ is the submarking of $m_\mathcal{Q}$ with all markings removed except those with the symbol $s$.
To the submultiball $Q{\triangleright} i\in\mathcal{P}$ corresponds exactly one symbol $r$ of $m_\mathcal{P}$ such that
$Q{\triangleright} i=[\alpha,m_\mathcal{P}^r]$. Since $Q\subset Q{\triangleright} i$ we have
$(\alpha,m_\mathcal{Q}^s)\subset(\alpha,m_\mathcal{P}^r)$ and therefore $m_\mathcal{Q}^s\subset m_\mathcal{P}^r$ by Remark
\ref{00176}. It follows $m_\mathcal{Q}\subset m_\mathcal{P}$ and thus $\mathcal{Q}\subset\mathcal{P}$.
\end{proof}
\subsection{The action on the set of semi-partitions}
Here we will define an action of $\Gamma=\pi_1(\mathcal{S},x)$ on the set of semi-partitions over $x$.
So let $\gamma\in\Gamma$
and $\mathcal{P}$ be a semi-partition over $x$. We will define another semi-partition $\gamma\cdot\mathcal{P}$ over $x$.
Recall that $\gamma$ is represented by a span $x\xleftarrow{\gamma_d}a\xrightarrow{\gamma_n}x$ (the $d$
refers to \emph{denominator} and the $n$ refers to \emph{nominator}) and that $\mathcal{P}$ is represented by a marked arrow
$(\alpha\colon c\rightarrow x,m)$. First choose a square filling of $(\gamma_n,\alpha)$
\begin{displaymath}\xymatrix{
x&&a\ar[ll]_{\gamma_d}\ar[rr]^{\gamma_n}&&x&&c\ar[ll]_\alpha\\
&&&&&&\\
&&&&b\ar@{-->}[uullll]^{\delta}\ar@{..>}[uull]_{\beta_1}\ar@{..>}[uurr]^{\beta_2}&&
}\end{displaymath}
and then define $\delta:=\beta_1\gamma_d\colon b\rightarrow x$. Endow this arrow with the marking $\mu:=\beta_2^*(m)$.
Finally, define $\gamma\cdot\mathcal{P}:=[\delta,\mu]$. We have to show that this is well-defined, i.e.~we have to show
that the resulting class is independent of
\begin{itemize}
\item[\bf{1.}] the square filling $(\beta_1,\beta_2)$
\item[\bf{2.}] the marked arrow $(\alpha,m)$ as a representative of $\mathcal{P}$
\item[\bf{3.}] the span $(\gamma_d,\gamma_n)$ as a representative of $\gamma$
\end{itemize}
We now prove these points one by one.
{\bf 1.} Assume we have two square fillings of $(\gamma_n,\alpha)$ as in the following diagram:
\begin{displaymath}\xymatrix{
&&&&b'\ar@{..>}[dll]_{\beta'_1}\ar@{..>}[drr]^{\beta'_2}&&\\
x&&a\ar[ll]_{\gamma_d}\ar[rr]^{\gamma_n}&&x&&c\ar[ll]_\alpha\\
&&&&b\ar@{..>}[ull]^{\beta_1}\ar@{..>}[urr]_{\beta_2}&&
}\end{displaymath}
Choose a common square filling as in Lemma \ref{25010}:
\begin{displaymath}\xymatrix{
&&&e\ar@{-->}[dr]^{\eta'}\ar@{-->}[dddr]^{\eta}\ar@{-->}@/_10pt/[ddl]_{\delta_1}
\ar@{-->}@/^20pt/[ddrrr]^{\delta_2}&&&\\
&&&&b'\ar@{..>}[dll]_{\beta'_1}\ar@{..>}[drr]^{\beta'_2}&&\\
x&&a\ar[ll]_{\gamma_d}\ar[rr]^{\gamma_n}&&x&&c\ar[ll]_\alpha\\
&&&&b\ar@{..>}[ull]^{\beta_1}\ar@{..>}[urr]_{\beta_2}&&
}\end{displaymath}
Now the marked arrow $(\beta_1\gamma_d,\beta_2^*(m))$ is equivalent to the marked arrow
$(\delta_1\gamma_d,\delta_2^*(m))$ via $\eta$. Analogously, the marked arrow $(\beta'_1\gamma_d,{\beta'_2}^*(m))$ is
equivalent to $(\delta_1\gamma_d,\delta_2^*(m))$ via $\eta'$ and therefore equivalent to
$(\beta_1\gamma_d,\beta_2^*(m))$, q.e.d.
{\bf 2.} Let $(\alpha',m')$ be another marked arrow equivalent to $(\alpha,m)$ and choose a square filling
$(\beta,\beta')$ such that $\beta^*(m)={\beta'}^*(m')=:\mu$ as in the following diagram:
\begin{displaymath}\xymatrix{
&&&&&&c\ar[dll]_{\alpha}&&\\
x&&a\ar[ll]_{\gamma_d}\ar[rr]^{\gamma_n}&&x&&&&e\ar@{..>}[ull]_{\beta}\ar@{..>}[dll]^{\beta'}
\ar@{..>}[llll]_{\delta}\\
&&&&&&c'\ar[ull]^{\alpha'}&&
}\end{displaymath}
First choose a square filling $(\eta_1,\eta_2)$ of $(\gamma_n,\alpha)$ and then a square filling $(\nu_1,\nu_2)$
of $(\eta_2,\beta)$. Analogously, choose a square filling $(\eta'_1,\eta'_2)$ of $(\gamma_n,\alpha')$ and then a
square filling $(\nu'_1,\nu'_2)$ of $(\eta'_2,\beta')$
\begin{displaymath}\xymatrix{
&&&&&&z\ar@{-->}@/^10pt/[ddrr]^{\nu_2}\ar@{-->}[dll]_{\nu_1}&&\\
&&&&y\ar@{-->}[dll]_{\eta_1}\ar@{-->}[rr]^{\eta_2}&&c\ar[dll]_{\alpha}&&\\
x&&a\ar[ll]_{\gamma_d}\ar[rr]^{\gamma_n}&&x&&&&e\ar@{..>}[ull]_{\beta}\ar@{..>}[dll]^{\beta'}
\ar@{..>}[llll]_{\delta}\\
&&&&y'\ar@{-->}[ull]^{\eta'_1}\ar@{-->}[rr]_{\eta'_2}&&c'\ar[ull]^{\alpha'}&&\\
&&&&&&z'\ar@{-->}@/_10pt/[uurr]_{\nu'_2}\ar@{-->}[ull]^{\nu'_1}&&
}\end{displaymath}
The marked arrow $(\eta_1\gamma_d,\eta_2^*(m))$ is equivalent to $\Lambda:=(\nu_1\eta_1\gamma_d,\nu_2^*(\mu))$
via $\nu_1$. On the other side, the marked arrow $(\eta'_1\gamma_d,{\eta'_2}^*(m'))$ is equivalent to
$\Lambda':=(\nu'_1\eta'_1\gamma_d,{\nu'_2}^*(\mu))$ via $\nu'_1$. The marked arrows $\Lambda$ and $\Lambda'$ are both
constructed from the same marked arrow $(\delta,\mu)$ and so are equivalent by {\bf 1}. Consequently,
$(\eta_1\gamma_d,\eta_2^*(m))$ and $(\eta'_1\gamma_d,{\eta'_2}^*(m'))$ are equivalent, q.e.d.
{\bf 3.} Let $(\gamma'_d,\gamma'_n)$ be another representing span of $\gamma$ homotopic to the span
$(\gamma_d,\gamma_n)$. Then recall that the two spans can be filled by a diagram as follows:
\begin{displaymath}\xymatrix{
&&a\ar[dll]_{\gamma_d}\ar[drr]^{\gamma_n}&&&&\\
x&&e\ar@{..>}[u]_{\eta}\ar@{..>}[d]^{\eta'}\ar@{..>}[rr]^{\delta_n}\ar@{..>}[ll]_{\delta_d}&&x&&c\ar[ll]_\alpha\\
&&a'\ar[ull]^{\gamma'_d}\ar[urr]_{\gamma'_n}&&&&
}\end{displaymath}
Now choose a square filling $(\nu_1,\nu_2)$ of $(\delta_n,\alpha)$ and note that $(\epsilon,\nu_2)$,
where $\epsilon:=\nu_1\eta$, gives a square filling of $(\gamma_n,\alpha)$.
\begin{displaymath}\xymatrix{
&&&&z\ar@{-->}[ddll]^/-8pt/{\nu_1}\ar@{-->}[ddrr]_/-8pt/{\nu_2}\ar[dll]_\epsilon&&\\
&&a\ar[dll]_{\gamma_d}\ar[drr]^{\gamma_n}&&&&\\
x&&e\ar@{..>}[u]_{\eta}\ar@{..>}[d]^{\eta'}\ar@{..>}[rr]^{\delta_n}\ar@{..>}[ll]_{\delta_d}&&x&&c\ar[ll]_\alpha\\
&&a'\ar[ull]^{\gamma'_d}\ar[urr]_{\gamma'_n}&&&&
}\end{displaymath}
The marked arrow $(\epsilon\gamma_d,\nu_2^*(m))$ is equivalent to $(\nu_1\delta_d,\nu_2^*(m))$. Similarly,
define $\epsilon'=\nu_1\eta'$ and note that $(\epsilon',\nu_2)$ gives a square filling of $(\gamma'_n,\alpha)$.
Again, the marked arrow $(\epsilon'\gamma'_d,\nu_2^*(m))$ is equivalent to $(\nu_1\delta_d,\nu_2^*(m))$. Therefore,
$(\epsilon\gamma_d,\nu_2^*(m))$ and $(\epsilon'\gamma'_d,\nu_2^*(m))$ are equivalent, q.e.d.
\vspace{2mm}
Now we want to show that this is indeed an action, i.e.~$1\cdot\mathcal{P}=\mathcal{P}$ and
$\gamma^1\cdot(\gamma^2\cdot\mathcal{P})=(\gamma^1\gamma^2)\cdot\mathcal{P}$. The first property is easy to see. The
second property is not entirely trivial but straightforward. We will be explicit for completeness.
Choose two representing spans $(\gamma^1_d,\gamma^1_n)$ and $(\gamma^2_d,\gamma^2_n)$ for $\gamma^1$ and
$\gamma^2$ respectively. Let $(\alpha,m)$ represent $\mathcal{P}$. To get a representing span for the composition
$\gamma_1\gamma_2$, choose a square filling $(\beta_1,\beta_2)$ of $(\gamma^1_n,\gamma^2_d)$ and take
the span $(\beta_1\gamma^1_d,\beta_2\gamma^2_n)$. This span acts on $(\alpha,m)$ as before and is sketched
diagrammatically as follows:
\begin{displaymath}\xymatrix{
&&&&&&&&z\ar[dllll]_{\eta}\ar[ddrr]^{\nu}&&\\
&&&&y\ar[dll]^{\beta_1}\ar[drr]_{\beta_2}\ar[dllll]_{\delta_1}\ar[drrrr]^{\delta_2}&&&&&&\\
x&&a^1\ar[ll]^{\gamma^1_d}\ar[rr]_{\gamma^1_n}&&x&&a^2\ar[ll]^{\gamma^2_d}\ar[rr]_{\gamma^2_n}&&x&&
c\ar[ll]^\alpha
}\end{displaymath}
So a representative of $(\gamma^1\gamma^2)\cdot\mathcal{P}$ is given by $(\eta\delta_1,\nu^*(m))$. Now a representative
for $\gamma^2\cdot\mathcal{P}$ is given by $(\eta\beta_2\gamma^2_d,\nu^*(m))$ because $(\eta\beta_2,\nu)$ is a square
filling for $(\gamma^2_n,\alpha)$. Since $(\eta\beta_1,\mathrm{id}_z)$ is a square filling for
$(\gamma^1_n,\eta\beta_2\gamma^2_d)$, we obtain that $(\eta\beta_1\gamma^1_d,\mathrm{id}_z^*(\nu^*(m)))$ is
a representative of $\gamma^1\cdot(\gamma^2\cdot\mathcal{P})$. But this last marked arrow is equal to
$(\eta\delta_1,\nu^*(m))$, q.e.d.
\begin{rem}\label{12397}
It is not hard to see that $\mathcal{P}\subset\mathcal{Q}$ implies $\gamma\cdot\mathcal{P}\subset\gamma\cdot\mathcal{Q}$.
\end{rem}
\begin{rem}\label{72125}
The submultiballs of $\gamma\cdot\mathcal{P}$ are the multiballs $\gamma\cdot P$ with $P\in\mathcal{P}$.
\end{rem}
\subsection{Pointwise stabilizers of partitions}
Let $\mathcal{P}$ be a partition over $x$. By the pointwise stabilizer of $\mathcal{P}$ we mean
the subgroup
\[\Lambda:=\{\gamma\in\pi_1(\mathcal{S},x)\mid\gamma\cdot P=P\text{ for all }P\in\mathcal{P}\}\]
Fix some representative $(\alpha,m)$ of $\mathcal{P}$. We can assume without loss of generality that the marking
$m$ on the domain $c$ of $\alpha$ is ordered. That means that if $f\colon I\rightarrow S$ is the marking function
of $m$ and whenever $i{\triangleright} f=j{\triangleright} f$ for $i<j$, then also $i{\triangleright} f=k{\triangleright} f=j{\triangleright} f$ for every $k$ with
$i<k<j$. This is true in the planar case by definition. In the symmteric case, we can choose a
colored permutation $\sigma\in\mathfrak{Sym}(C)$ with $\sigma^*(m)$ ordered and replace $(\alpha,m)$ by the equivalent
marked arrow $(\sigma\alpha,\sigma^*(m))$.
\begin{prop}\label{88520}
Each symbol of the marking $m$ determines a subword of the word $c=\operatorname{dom}(\alpha)$. Denote
these subwords by $c_1,...,c_k$ and order them such that $c=c_1\otimes...\otimes c_k$.
Then we have a well-defined isomorphism of groups
\[\Xi\colon \pi_1(\mathcal{S},c_1)\times...\times\pi_1(\mathcal{S},c_k)\rightarrow\Lambda\]
which is given by applying the tensor product of paths and then conjugating with the arrow $\alpha$. More
explicitly, it is given by sending representing spans $p_1,...,p_k$ to the homotopy class represented by the path
\begin{displaymath}\begin{array}{ccccccccc}
&&c_1&\xleftarrow{p_1^\prec}&a_1&\xrightarrow{p_1^\succ}&c_1&&\\
&&\otimes &\otimes &\otimes &\otimes &\otimes &&\\
x&\xleftarrow{\hspace{5mm}\alpha\hspace{5mm}}{}&\vdots &\vdots &\vdots &
\vdots &\vdots &\xrightarrow{\hspace{5mm}\alpha\hspace{5mm}}&x\\
&&\otimes &\otimes &\otimes &\otimes &\otimes &&\\
&&c_k&\xleftarrow[p_k^\prec]{}&a_k&\xrightarrow[p_k^\succ]{}&c_k&&
\end{array}\end{displaymath}
where $p_i^\prec$ is the arrow pointing to the left and $p_i^\succ$ the arrow pointing to the right in the
span $p_i$.
\end{prop}
\begin{proof}
It is not hard to see that the map is independent of the chosen
representing spans $p_i$ and that it is a group homomorphism. Injectivity follows from Lemma \ref{69343}
and Lemma \ref{56209} below. Before we prove surjectivity, we want to see that the image really lies in the
subgroup $\Lambda$. We can use the representative $(\alpha,m)$ to extract representatives of submultiballs
$P\in\mathcal{P}$. The subwords $c_i$ are in one to one correspondence with the submultiballs $P\in\mathcal{P}$.
A representative $(\alpha,m_i)$ of $P\in\mathcal{P}$ corresponding to $c_i$ is obtained from $(\alpha,m)$
by removing all markings except the markings on the subword $c_i$. The representing span of $\Xi(p_1,...,p_k)$
pictured above can be written as $(p^\prec\alpha,p^\succ\alpha)$ where
$p^\prec=p_1^\prec\otimes...\otimes p_k^\prec$ and $p^\succ=p_1^\succ\otimes...\otimes p_k^\succ$.
Letting this span act on $(\alpha,m_i)$, we can choose $(\mathrm{id},p^\succ)$
as a square filling and the resulting representative is $(p^\prec\alpha,{p^\succ}^*(m_i))$. But this is
equivalent to $(\alpha,m_i)$ because ${p^\prec}^*(m_i)={p^\succ}^*(m_i)$.
Now we prove surjectivity. Let $\gamma\in\Lambda$ which can be represented by a path of the form
\[x\xleftarrow{\hspace{5mm}\alpha\hspace{5mm}}c\xleftarrow{\hspace{5mm}z^\prec\hspace{5mm}}a
\xrightarrow{\hspace{5mm}z^\succ\hspace{5mm}}c\xrightarrow{\hspace{5mm}\alpha\hspace{5mm}}x\]
Observe the representatives $(\alpha,m_i)$ of the submultiballs $P\in\mathcal{P}$ from above. A representative
of $\gamma\cdot[\alpha,m_i]$ is given by $(z^\prec\alpha,{z^\succ}^*(m_i))$. So we have
$(\alpha,m_i)\sim(z^\prec\alpha,{z^\succ}^*(m_i))$. Of course, $(z^\prec,\mathrm{id})$ is a square
filling of $(\alpha,z^\prec\alpha)$ and thus
\[{z^\prec}^*(m_i)={z^\succ}^*(m_i)\]
Now assume for the moment that the operad $\mathcal{O}$ is planar. Then it follows easily from these equalities that
the span $(z^\prec,z^\succ)$ splits as a product according to the decomposition $c=c_1\otimes...\otimes c_k$,
i.e.~there are $z_i^\prec\colon a_i\rightarrow c_i$ and $z_i^\succ\colon a_i\rightarrow c_i$ with
$z^\prec=z_1^\prec\otimes...\otimes z_k^\prec$ and $z^\succ=z_1^\succ\otimes...\otimes z_k^\succ$.
By construction, the spans $(z_i^\prec,z_i^\succ)$ give a preimage of $\gamma$ under $\Xi$. If, on the
other hand, $\mathcal{O}$ is symmetric, then there is colored permutation $\sigma\in\mathfrak{Sym}(C)$
such that the span $(\sigma z^\prec,\sigma z^\succ)$, which is homotopic to $(z^\prec,z^\succ)$, splits as
above and we can finish the proof also in this case.
\end{proof}
\begin{lem}\label{56209}
Let $a\xleftarrow{q}b\xrightarrow{p}a$ be a span in $\mathcal{S}$ which is a tensor product of $k$ spans
$a_i\xleftarrow{q_i}b_i\xrightarrow{p_i}a_i$ for $i=1,...,k$, i.e.~$q=q_1\otimes....\otimes q_k$ and
$p=p_1\otimes...\otimes p_k$. Then the span $(q,p)$ is null-homotopic if and only if each $(q_i,p_i)$
is null-homotopic.
\end{lem}
\begin{proof}
It is clear that if each $(q_i,p_i)$ is null-homotopic, then $(q,p)$ is null-homotopic. So we prove the converse.
We can assume without loss of generality that $q_i\neq \mathrm{id}_I\neq p_i$ where $I$ is the monoidal unit in $\mathcal{S}$,
i.e.~the empty word. First observe that $p,q$ are parallel arrows and since $\mathcal{S}$ satisfies the calculus of fractions,
they are homotopic if and only if there is an arrow $r\colon c\rightarrow b$ with $rq=rp$. Now, by precomposing with
an arrow in $\mathfrak{Sym}(C)$ if necessary, we can assume that $r$ is an arrow in
$\mathcal{S}(\mathcal{O}_\mathrm{pl})$, i.e.~a tensor product of operations in $\mathcal{O}$. Observe that in $\mathcal{S}$
we have $\alpha_1\otimes...\otimes\alpha_l=\beta_1\otimes...\otimes\beta_m$ for arrows $\alpha_i\neq\mathrm{id}_I\neq\beta_i$ if
and only if $l=m$ and $\alpha_i=\beta_i$ for each $i=1,...,l$. Now it follows easily that $r$ gives arrows $r_1,...,r_k$
such that $r_iq_i=r_ip_i$ for each $i=1,...,k$. Thus, $q_i$ is homotopic to $p_i$ for each $i=1,...,k$.
\end{proof}
\subsection{The poset of partitions}
From now on, fix some base object $x$ which is split and progressive. More generally, in view of Remark \ref{50091}:
\vspace{2mm}
\begin{center}
\framebox[1.1\width]{Let $y$ be a split object such that $x$ is $y$--progressive.}
\end{center}
\vspace{2mm}
Furthermore, let $n\in\mathbb{N}$.
Two objects $a,b$ in $\mathcal{S}$ are called equivalent if they are isomorphic in $\pi_1(\mathcal{S})$, i.e.~there is a path
(equivalently, a span) between them in $\mathcal{S}$. Of course, $\pi_1(\mathcal{S},a)\cong\pi_1(\mathcal{S},b)$ in this case.
Let $c=(c_1,...,c_k)$ with $c_i\in C$ an object in $\mathcal{S}$ and $m$ be a uni-marking on $c$, i.e.~there is only
one symbol in $m$. Then $m$ determines another object $\mathfrak{c}(m)$ by deleting all $c_i$'s which are not marked
by $m$. If $\alpha\colon a\rightarrow c$ is an arrow, then $\mathfrak{c}(\alpha^*(m))\sim\mathfrak{c}(m)$ in the above
sense.
Let $B$ be a multiball. If $(\alpha,m)$ and $(\alpha',m')$ are representatives, then
$\mathfrak{c}(m)\sim\mathfrak{c}(m')$. Thus, each multiball $B$ gives an equivalence class
$\mathfrak{cc}(B)$ of objects.
We say that a partition $\mathcal{P}$ (over $x$) satisfies the $n$--condition with respect to $y$ if at least $n$
submultiballs $P\in\mathcal{P}$ satisfy $y\in\mathfrak{cc}(P)$. The $n$--condition with respect to $y$ is preserved by
the action of $\Gamma=\pi_1(\mathcal{S},x)$ on the partitions: If $\mathcal{P}$ satisfies the $n$--condition with respect
to $y$, then also $\gamma\cdot\mathcal{P}$ satisfies it.
We define a poset $(\mathbb{P},\leq)$: The objects of $\mathbb{P}$ are partitions over $x$ and $\mathcal{P}\leq\mathcal{Q}$
if and only if $\mathcal{P}\supset\mathcal{Q}$. The group $\Gamma=\pi_1(\mathcal{S},x)$ acts on this poset via the action on partitions.
Because of Remark \ref{12397}, the action indeed respects the relation $\leq$.
Since the $n$--condition with respect to $y$ is invariant under the action of $\Gamma$, we can define the invariant
subposet $(\mathbb{P}_n,\leq)$ to be the full subpost consisting of partitions satisfying the $n$--condition with
respect to $y$. Next, we want
to show that
\begin{itemize}
\item[{\bf 1.}] $\mathbb{P}_n\neq\emptyset$ and
\item[{\bf 2.}] $(\mathbb{P}_n,\leq)$ is filtered.
\end{itemize}
This implies that the poset $\mathbb{P}_n$ is \emph{contractible}.
{\bf 1.} Since $x$ is $y$--progressive, there is an arrow $a_1\otimes y\otimes a_2\rightarrow x$. Apply $y$'s split condition
$(n-1)$ times to find an arrow $z\rightarrow x$ where $z$ has a tensor product decomposition with at least $n$ factors equal to
$y$. Mark each of these factors with a different symbol and the rest with yet another symbol. This yields a partition
$\mathcal{P}\in\mathbb{P}_n$.
{\bf 2.} Let $\mathcal{P},\mathcal{Q}\in\mathbb{P}_n$. We have to find $\mathcal{R}\in\mathbb{P}_n$ with $\mathcal{P},\mathcal{Q}\leq\mathcal{R}$.
First we find one in $\mathbb{P}$. Let $(\alpha_\mathcal{P},m_\mathcal{P})$ and $(\alpha_\mathcal{Q},m_\mathcal{Q})$ be representatives of
$\mathcal{P}$ and $\mathcal{Q}$ respectively. Choose a square filling $(\beta_\mathcal{P},\beta_\mathcal{Q})$ of
$(\alpha_\mathcal{P},\alpha_\mathcal{Q})$ and set $\delta=\beta_\mathcal{P}\alpha_\mathcal{P}=\beta_\mathcal{Q}\alpha_\mathcal{Q}$. Now find a full marking
$\mu\subset\beta_\mathcal{P}^*(m_\mathcal{P}),\beta_\mathcal{Q}^*(m_\mathcal{Q})$, for example by marking each coordinate of
$\operatorname{dom}(\delta)$ with a different symbol. Then $\mathcal{R}=[\delta,\mu]$ is a common refinement of $\mathcal{P}$
and $\mathcal{Q}$. Now use that $x$ is $y$--progressive to find an arrow $\eta\colon z\rightarrow\operatorname{dom}(\delta)$
where $z$ has a tensor product decomposition with at least one factor equal to $y$. Then apply $y$'s split condition
$(n-1)$ times to obtain an arrow $\nu\colon w\rightarrow z$ where $w$ has a tensor product decomposition with at least
$n$ factors equal to $y$. Observe the marked arrow $(\nu\eta\delta,(\nu\eta)^*(\mu))$. The so-called link condition
in Remark \ref{50091} ensures that the $n$ factors of $w$ equal to $y$ are marked with the same symbol
in the marking $(\nu\eta)^*(\mu)$. Refine this marking such that these factors are marked with
new different symbols. This gives a representative of a partition satisfying the $n$--condition with respect to $y$,
refining $\mathcal{R}$ and thus refining both $\mathcal{P}$ and $\mathcal{Q}$.
\vspace{2mm}
A simplex $\sigma$ in the poset $\mathbb{P}_n$ is a finite ascending sequence of objects, written
$[\mathcal{P}_0<\mathcal{P}_1<...<\mathcal{P}_p]$. We now observe the stabilizer subgroup $\Gamma_\sigma$ of such a simplex.
By definition, an element $\gamma$ is in this stabilizer subgroup if and only if $\{\mathcal{P}_0,...,\mathcal{P}_p\}=
\{\gamma\cdot\mathcal{P}_0,...,\gamma\cdot\mathcal{P}_p\}$. But since the action of $\gamma$ respects $\leq$, this is
equivalent to $\gamma\cdot\mathcal{P}_i=\mathcal{P}_i$ for each $i=0,...,p$. So each $\gamma\in\Gamma_\sigma$ fixes
$\sigma$ vertex-wise. Observe the subgroup
\[\Lambda_\sigma:=\{\gamma\in\Gamma\mid\gamma\cdot P=P\text{ for all }P\in\mathcal{P}_p\}<\Gamma\]
By Proposition \ref{88520}, we know that $\Lambda_\sigma\cong\pi_1(\mathcal{S},c_1)\times...\times\pi_1(\mathcal{S},c_k)$
for appropriate objects $c_i$. Since $\mathcal{P}_p$ satisfies the $n$--condition with respect to $y$, at least $n$
of these objects are equivalent to $y$ and thus at least $n$ of the factors in the product decomposition of
$\Lambda_\sigma$ are isomorphic to $\Upsilon:=\pi_1(\mathcal{S},y)$. So we find a normal subgroup
$\Lambda'_\sigma\vartriangleleft\Lambda_\sigma$ with $\Lambda'_\sigma\cong\Upsilon^n$. Below, we will show that
$\Lambda_\sigma$ is a normal subgroup of $\Gamma_\sigma$. So we arrive at the following situation
\vspace{2mm}
\begin{center}
\framebox[1.1\width]{$\Upsilon^n\cong\Lambda'_\sigma\vartriangleleft\Lambda_\sigma\vartriangleleft\Gamma_\sigma$}
\end{center}
\vspace{2mm}
\begin{lem}
Let $\mathcal{R}_1,\mathcal{R}_2$ be semi-partitions and $\mathcal{P}$ be a partition with $\mathcal{P}\subset\mathcal{R}_1$.
Assume
\[\forall_{R_1\in\mathcal{R}_1}\ \exists_{R_2\in\mathcal{R}_2}\ \forall_{P\in\mathcal{P}}\ P\subset R_1
\Longrightarrow P\subset R_2\]
Then we have $\mathcal{R}_1\subset\mathcal{R}_2$.
\end{lem}
\begin{proof}
By applying the square filling technique twice, we find an arrow $\delta$ with three markings
$m_\mathcal{P},m_{\mathcal{R}_1},m_{\mathcal{R}_2}$ on its domain such that $(\delta,m_\mathcal{P})$ represents $\mathcal{P}$ and
$(\delta,m_{\mathcal{R}_i})$ represents $\mathcal{R}_i$. Since $\mathcal{P}\subset\mathcal{R}_1$ we have $(\delta,m_\mathcal{P})\subset
(\delta,m_{\mathcal{R}_1})$ and therefore $m_\mathcal{P}\subset m_{\mathcal{R}_1}$. Note that $\mathcal{P}$ is a partition and
therefore $m_\mathcal{P}$ is a full marking. Now the assumption of the statement implies
$m_{\mathcal{R}_1}\subset m_{\mathcal{R}_2}$ and thus $\mathcal{R}_1\subset\mathcal{R}_2$.
\end{proof}
We first show that $\Lambda_\sigma$ is contained in $\Gamma_\sigma$. So let $\gamma\in\Lambda_\sigma$,
i.e.~$\gamma\cdot P=P$ for all $P\in\mathcal{P}_p$. In particular, we have $\gamma\cdot\mathcal{P}_p=\mathcal{P}_p$ (Proposition \ref{45613}
and Remark \ref{72125}).
We have to show $\gamma\cdot\mathcal{P}_i=\mathcal{P}_i$ also for the other $i$'s. Write $\mathcal{P}:=\mathcal{P}_p$ and $\mathcal{R}:=\mathcal{P}_i$ for some
other $i$. Then we have $\mathcal{P}\subset\mathcal{R}$. We want to apply the above lemma to $\mathcal{R}_1=\mathcal{R}$ and
$\mathcal{R}_2=\gamma\cdot\mathcal{R}$ and deduce $\mathcal{R}\subset\gamma\cdot\mathcal{R}$. So let $R\in\mathcal{R}$ and observe
$\gamma\cdot R\in\gamma\cdot\mathcal{R}$. Let $P\in\mathcal{P}$ with $P\subset R$. Then $P=\gamma\cdot P\subset\gamma\cdot R$
and the assumption of the lemma is satisfied. Similarly, we get $\mathcal{R}\subset\gamma^{-1}\cdot\mathcal{R}$ and thus
$\gamma\cdot\mathcal{R}\subset\mathcal{R}$. This yields $\gamma\cdot\mathcal{R}=\mathcal{R}$, q.e.d.
Now we show that $\Lambda_\sigma$ is normal in $\Gamma_\sigma$. Let $\gamma\in\Gamma_\sigma$ and
$\alpha\in\Lambda_\sigma$. We have to show $\gamma^{-1}\alpha\gamma\in\Lambda_\sigma$,
i.e.~$\gamma^{-1}\alpha\gamma\cdot P=P$ for all $P\in\mathcal{P}_p=:\mathcal{P}$ or equivalently
$\alpha\cdot(\gamma\cdot P)=\gamma\cdot P$ for all $P\in\mathcal{P}$. Since $\gamma\cdot\mathcal{P}=\mathcal{P}$,
we have a bijection $f\colon \{P\in\mathcal{P}\}\rightarrow\{P\in\mathcal{P}\}$ such that $\gamma\cdot P=P{\triangleright} f$ for all
$P\in\mathcal{P}$ (Proposition \ref{45613}). Consequently, if $P\in\mathcal{P}$,
$\alpha\cdot(\gamma\cdot P)=\alpha\cdot(P{\triangleright} f)=P{\triangleright} f=\gamma\cdot P$, q.e.d.
\subsection{End of the proof}
Let $\mathfrak{P}$ be a property of groups which is closed under taking products and let $\mathcal{M}$ be a coefficient
system which is $\mathfrak{P}$--K\"unneth and $\mathfrak{P}$--inductive.
We will only give the proof for homology. Using analogous devices for cohomology, we obtain a proof of the
cohomological version of the statement.
Our main tool will be a spectral sequence explained in Brown's book \cite{bro:fpo}*{Chapter VII.7} (see also
\cite{sa-th:laa}*{Subsection 4.1}). If we plug in our $\Gamma$--complex $(\mathbb{P}_n,\leq)$ and the $\mathbb{Z}\Gamma$--module $\mathcal{M}\Gamma$,
we obtain a spectral sequence $E^k_{pq}$ with
\[E^1_{pq}=\bigoplus_{\sigma\in\Sigma_p}H_q(\Gamma_\sigma,\mathcal{M}\Gamma)\Rightarrow
H_{p+q}(\Gamma,\mathcal{M}\Gamma)\]
where $\Sigma_p$ is set of $p$--cells representing the $\Gamma$--orbits of $(\mathbb{P}_n,\leq)$. This uses that the poset $\mathbb{P}_n$
is contractible and that the cell stabilizers fix the cells pointwise.
We assumed that $\Upsilon$ satisfies $\mathfrak{P}$ and that $H_0(\Upsilon,\mathcal{M}\Upsilon)=0$. Applying the
$\mathfrak{P}$--K\"unneth property $(n-1)$ times, we obtain $H_k(\Upsilon^n,\mathcal{M}\Upsilon^n)=0$ for $k\leq n-1$.
So we have $H_k(\Lambda'_\sigma,\mathcal{M}\Lambda'_\sigma)=0$ for $k\leq n-1$. The property $\mathfrak{P}$--inductive
yields $H_k(\Lambda'_\sigma,\mathcal{M}\Gamma)=0$ for $k\leq n-1$. Since $\Lambda'_\sigma\vartriangleleft
\Lambda_\sigma\vartriangleleft\Gamma_\sigma$, we can apply the Hochschild--Serre spectral sequence twice to
obtain $H_k(\Gamma_\sigma,\mathcal{M}\Gamma)=0$ for $k\leq n-1$. The above spectral sequence now yields
$H_k(\Gamma,\mathcal{M}\Gamma)=0$ for $k\leq n-1$. Since $n$ was arbitrary, the result follows.
\section{Non-amenability and infiniteness}
In this section we use the techniques from Section \ref{96238} to prove non-amenability and infiniteness of some operad groups.
Note that semi-partitions and the action on the set of semi-partitions can also be defined in the braided case.
\begin{lem}\label{67656}
If $\mathcal{O}$ satisfies the calculus of fractions, then the action of the colored permutations in
$\operatorname{Aut}_{\mathfrak{Sym}(C)}(X)$ or the colored braids in
$\operatorname{Aut}_{\mathfrak{Braid}(C)}(X)$ on the set of arrows $\operatorname{Hom}_{\mathcal{S}(\mathcal{O})}(X,Y)$
is free. In particular, in the operad $\mathcal{O}$, the action of the symmetric groups or the braid
groups on the sets of operations is free.
\end{lem}
\begin{proof}
Let $[\alpha,\Theta]$ be an element in $\operatorname{Hom}_{\mathcal{S}(\mathcal{O})}(X,Y)$ and $\sigma\in\operatorname{Aut}_{\mathfrak{Sym}(C)}(X)$
or $\sigma\in\operatorname{Aut}_{\mathfrak{Braid}(C)}(X)$. We have to show that $[\sigma,\mathrm{id}]*[\alpha,\Theta]=[\alpha,\Theta]$
implies that $\sigma$ is trivial. From this equality and the equalization property of $\mathcal{S}(\mathcal{O})$, we obtain an arrow
$z:=[\delta,\Psi]$ with $z*[\sigma,\mathrm{id}]=z$. We can assume without loss of generality that $\delta=\mathrm{id}$.
We then have $z*[\sigma,\mathrm{id}]=[\bar{\sigma},\bar{\Psi}]$ with $\bar{\sigma}=\Psi{\curvearrowright}\sigma$ and
$\bar{\Psi}=\Psi{\curvearrowleft}\sigma$. Consequently, the pairs $(\bar{\sigma},\bar{\Psi})$ and $(\mathrm{id},\Psi)$
are equivalent in $\mathfrak{Sym}(C)\times\mathcal{S}(\mathcal{O}_\mathrm{pl})$ or $\mathfrak{Braid}(C)\times\mathcal{S}(\mathcal{O}_\mathrm{pl})$.
This is only possible if $\sigma$ is trivial.
\end{proof}
Let $\mathcal{O}$ be a symmetric or braided operad. Let $\alpha$ be an arrow in $\mathcal{S}(\mathcal{O})$.
For any colored permutation $\sigma\in\mathfrak{Sym}(C)$ or colored braid $\sigma\in\mathfrak{Braid}(C)$ with suitable domain and
codomain, we can form the group element $\gamma$ represented by the span $(\alpha,[\sigma,\mathrm{id}]*\alpha)$. Recall that the first
arrow always denotes the denominator, i.e.~points to the left.
\begin{lem}\label{22286}
Assume $\mathcal{O}$ satisfies the calculus of fractions. Then
\[\sigma\neq 1\ \ \Longrightarrow\ \ \gamma\neq 1\]
\end{lem}
\begin{proof}
First consider the symmetric case. Observe the semi-partition $\mathcal{R}$ represented by the marked arrow
$(\alpha,m)$ where $m$ is a marking on the domain of $\alpha$ with
only one marked coordinate and this coordinate is non-trivially permuted by $\sigma^{-1}$.
It is easy to see that $\gamma\cdot\mathcal{R}$ is represented by $(\alpha,[\sigma,\mathrm{id}]^*(m))$.
The marking $m':=[\sigma,\mathrm{id}]^*(m)$ is different from $m$ because $\sigma^{-1}$ maps the only
marked coordinate of $m$ to a different coordinate by assumption.
From Remark \ref{00176} it follows that the marked
arrow $(\alpha,m)$ cannot be equivalent to $(\alpha,m')$ and thus $\gamma\cdot\mathcal{R}\neq\mathcal{R}$. Consequently
$\gamma\neq 1$.
Now if $\mathcal{O}$ is braided we can apply the above argument verbatim if we require that the braid $\sigma$
has a non-trivial permutation part. But there are of course non-trivial braids which are trivial as
permutations (so-called pure braids).
Assume that $\sigma$ is such a pure braid and that $\gamma=1$. Note that the latter means that
the parallel arrows $\sigma\alpha:=[\sigma,\mathrm{id}]*\alpha$ and $\alpha$ are homotopic.
Since $\mathcal{S}(\mathcal{O})$ satisfies the calculus of fractions, this means that there is an arrow $\delta$ with
$\delta*\sigma\alpha=\delta*\alpha$. We can assume without loss of generality that
$\delta\in\mathcal{S}(\mathcal{O}_{\mathrm{pl}})$, i.e.~that $\delta=[\mathrm{id},\Theta]$. Composing in
$\mathcal{S}(\mathcal{O})$, we get $\delta*[\sigma,\mathrm{id}]=[\bar{\sigma},\bar{\Theta}]$ where
$\bar{\sigma}=\Theta{\curvearrowright}\sigma$ and $\bar{\Theta}=\Theta{\curvearrowleft}\sigma$.
Using that $\sigma$ is pure we immediately see $\bar{\Theta}=\Theta$. So we have
\[[\bar{\sigma},\mathrm{id}]*\big([\mathrm{id},\Theta]*\alpha\big)=\big(\delta*[\sigma,\mathrm{id}]\big)*\alpha=[\mathrm{id},\Theta]*\alpha\]
Lemma \ref{67656} now gives $\bar{\sigma}=1$ and thus $\sigma=1$.
\end{proof}
Denoting the element $\gamma$ suggestively by $\xleftarrow{\alpha}\sigma\xrightarrow{\alpha}$, the lemma implies
that two such group elements $\xleftarrow{\alpha}\sigma\xrightarrow{\alpha}$ and
$\xleftarrow{\alpha}\sigma'\xrightarrow{\alpha}$ are equal if and only if $\sigma=\sigma'$. We will use this now
to give a proof of the following proposition.
\begin{prop}\label{22947}
Let $\mathcal{O}$ be a symmetric operad satisfying the calculus of fractions and let $X$ be a split
object of $\mathcal{S}(\mathcal{O})$. Then $\Gamma=\pi_1(\mathcal{O},X)$ contains a non-abelian free subgroup and
is therefore non-amenable.
\end{prop}
\begin{proof}
Using the split condition on $X$, we will explicitly construct two non-trivial elements
$\gamma_1,\gamma_2\in\Gamma$ of order $2$ and $3$ respectively. Then we will define two disjoint subsets
$A_1,A_2$ of the set of semi-partitions over $X$ such that $\gamma_1\cdot A_2\subset A_1$ and
$\gamma_2^n\cdot A_1\subset A_2$ for $n=1,2$. The Ping--Pong Lemma then shows that the subgroup
$\langle\gamma_1,\gamma_2\rangle$ generated by the two elements $\gamma_1$ and $\gamma_2$ is isomorphic
to the free product $\langle\gamma_1\rangle*\langle\gamma_2\rangle$. So we have found a subgroup
which is isomorphic to $\mathbb{Z}_2*\mathbb{Z}_3$. Since $\mathbb{Z}_2*\mathbb{Z}_3$ contains a free non-abelian subgroup, the
proof of the proposition is then complete.
We now give the constructions. Because $X$ is split, there is an arrow
\[\varpi\colon A_1\otimes X\otimes A_2\otimes X\otimes A_3\rightarrow X\]
For better readability, we assume that $X$ is a single color and $A_1=A_2=A_3=I$. The construction
goes the same way in the general case (with obvious modifications). So we assume that $\varpi$ is just
an operation with two inputs of color $X$ and an output of color $X$. Define $\gamma_1$ to be the
following element
\begin{center}\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (7,0) -- +(-2,2) -- +(-2,-2) -- +(0,0);
\draw (2,1.2) to[out=right,in=left] (5,-1.2);
\draw (2,-1.2) to[out=right,in=left] (5,1.2);
\node at (1.2,0) {$\varpi$};
\node at (5.8,0) {$\varpi$};
\end{tikzpicture}\end{center}
and $\gamma_2$ to be the following element
\begin{center}\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1) -- +(1,1) -- +(1,-1) -- +(0,0);
\draw (2,-1.2) to (3,-1.2);
\draw (3,1.6) to[out=right,in=left] (6,0.4);
\draw (3,0.4) to[out=right,in=left] (6,-1.2);
\draw (9,0) -- +(-2,2) -- +(-2,-2) -- +(0,0);
\draw (7,1) -- +(-1,1) -- +(-1,-1) -- +(0,0);
\draw (7,-1.2) to (6,-1.2);
\draw (3,-1.2) to[out=right,in=left] (6,1.6);
\node at (1.2,0) {$\varpi$};
\node at (7.8,0) {$\varpi$};
\node at (2.6,1) {$\varpi$};
\node at (6.4,1) {$\varpi$};
\end{tikzpicture}\end{center}
Lemma \ref{22286} implies that $\gamma_1$ is of order $2$ and $\gamma_2$ is of order $3$. Let $B_1$
be the ball represented by the marked arrow
\begin{center}\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1.2) to (3,1.2);
\draw (2,-1.2) to (3,-1.2) node[right]{$\bigstar$};
\node at (1.2,0) {$\varpi$};
\end{tikzpicture}\end{center}
and let $B_2$ be the ball represented by the marked arrow
\begin{center}\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1.2) to (3,1.2) node[right]{$\bigstar$};
\draw (2,-1.2) to (3,-1.2);
\node at (1.2,0) {$\varpi$};
\end{tikzpicture}\end{center}
Composing the operation $\varpi$ several times, one gets operations that look like binary trees. Call them
$\varpi$--tree operations. Now define $A_1$ to be the set of all balls $B\subset B_1$ which are represented
by $\varpi$--tree operations. Similarly, define $A_2$ to be the set of all balls $B\subset B_2$ which are
represented by $\varpi$--tree operations. For example, the following marked arrows represent balls in $A_1$
\begin{center}
\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1.2) to (3,1.2);
\draw (2,-1.2) to (3,-1.2) node[right]{$\bigstar$};
\node at (1.2,0) {$\varpi$};
\end{tikzpicture}
\hspace{3mm}
\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,-1) -- +(1,1) -- +(1,-1) -- +(0,0);
\draw (2,1.2) to (4,1.2);
\draw (3,-1.6) to (4,-1.6);
\draw (3,-0.4) to (4,-0.4) node[right]{$\bigstar$};
\node at (1.2,0) {$\varpi$};
\node at (2.6,-1) {$\varpi$};
\end{tikzpicture}
\hspace{3mm}
\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,-1) -- +(1,1) -- +(1,-1) -- +(0,0);
\draw (2,1.2) to (4,1.2);
\draw (3,-1.6) to (4,-1.6) node[right]{$\bigstar$};
\draw (3,-0.4) to (4,-0.4);
\node at (1.2,0) {$\varpi$};
\node at (2.6,-1) {$\varpi$};
\end{tikzpicture}
\end{center}
and the following marked arrows represent balls in $A_2$
\begin{center}
\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1.2) to (3,1.2) node[right]{$\bigstar$};
\draw (2,-1.2) to (3,-1.2);
\node at (1.2,0) {$\varpi$};
\end{tikzpicture}
\hspace{3mm}
\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1) -- +(1,1) -- +(1,-1) -- +(0,0);
\draw (2,-1.2) to (4,-1.2);
\draw (3,1.6) to (4,1.6) node[right]{$\bigstar$};
\draw (3,0.4) to (4,0.4);
\node at (1.2,0) {$\varpi$};
\node at (2.6,1) {$\varpi$};
\end{tikzpicture}
\hspace{3mm}
\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1) -- +(1,1) -- +(1,-1) -- +(0,0);
\draw (2,-1.2) to (4,-1.2);
\draw (3,1.6) to (4,1.6);
\draw (3,0.4) to (4,0.4) node[right]{$\bigstar$};
\node at (1.2,0) {$\varpi$};
\node at (2.6,1) {$\varpi$};
\end{tikzpicture}
\end{center}
It is straightforward to check $\gamma_1\cdot A_2\subset A_1$ and $\gamma_2\cdot A_1\subset A_2$ and
$\gamma_2^2\cdot A_1\subset A_2$, so the proof is completed.
\end{proof}
Next we give sufficient conditions for infiniteness of operad groups.
\begin{prop}\label{96213}
Let $\mathcal{O}$ be a planar, symmetric or braided operad satisfying the calculus of fractions and let
$X$ be a split object of $\mathcal{S}(\mathcal{O})$. Then $\Gamma:=\pi_1(\mathcal{O},X)$ contains an infinite
cyclic subgroup and is therefore infinite.
\end{prop}
\begin{proof}
Because $X$ is split, there is an arrow
\[\varpi\colon A_1\otimes X\otimes A_2\otimes X\otimes A_3\rightarrow X\]
For better readability, we assume that $X$ is a single color and $A_1=A_2=A_3=I$. The construction
goes the same way in the general case (with obvious modifications). So we assume that $\varpi$ is just
an operation with two inputs of color $X$ and an output of color $X$. Define $\gamma$ to be the
following element
\begin{center}\begin{tikzpicture}[scale=0.45]
\draw (0,0) -- +(2,2) -- +(2,-2) -- +(0,0);
\draw (2,1) -- +(1,1) -- +(1,-1) -- +(0,0);
\draw (9,0) -- +(-2,2) -- +(-2,-2) -- +(0,0);
\draw (7,-1) -- +(-1,1) -- +(-1,-1) -- +(0,0);
\draw (3,1.6) to[out=right,in=left] (6,1) to (7,1);
\draw (3,0.4) to[out=right,in=left] (6,-0.4);
\draw (2,-1) to (3,-1) to[out=right,in=left] (6,-1.6);
\node at (1.2,0) {$\varpi$};
\node at (7.8,0) {$\varpi$};
\node at (6.4,-1) {$\varpi$};
\node at (2.6,1) {$\varpi$};
\end{tikzpicture}\end{center}
Formally, $\gamma$ is represented by the span
$\big((\varpi,\mathrm{id})*\varpi,(\mathrm{id},\varpi)*\varpi\big)$. We claim that $\gamma$ has
infinite order. The element $\gamma^n$ is represented by the span (for better readability, we use the same symbol
$\mathrm{id}$ for different identities)
\[\big((\varpi,\mathrm{id})*...*(\varpi,\mathrm{id})*\varpi,
(\mathrm{id},\varpi)*...*(\mathrm{id},\varpi)*\varpi\big)\]
By Lemma \ref{69343}, this span is null-homotopic if and only of the span (remove the last $\varpi$ in both arrows)
\[\big((\varpi,\mathrm{id})*...*(\varpi,\mathrm{id}),
(\mathrm{id},\varpi)*...*(\mathrm{id},\varpi)\big)=:(\varpi_1,\varpi_2)\]
is null-homotopic. This is true if and only if there is an arrow $r$ with $r\varpi_1=r\varpi_2$. The arrow
$r$ can be chosen to lie in $\mathcal{S}(\mathcal{O}_{\mathrm{pl}})$. But note that $\varpi_1$ splits as
\[\varpi_1=\big((\varpi,\mathrm{id})*...*(\varpi,\mathrm{id})*\varpi\big)\otimes\mathrm{id}_X\]
and $\varpi_2$ splits as
\[\varpi_2=\mathrm{id}_X\otimes\big((\mathrm{id},\varpi)*...*(\mathrm{id},\varpi)*\varpi\big)\]
It can easily be seen that such an arrow $r$ cannot exist because otherwise operations with a different number
of inputs must be equal. Consequently, each $\gamma^n$ is non-trivial and therefore $\gamma$ has infinite order.
\end{proof}
\section{Applications}
Observe the $1$--dimensional planar cube cutting operads and the $d$--dimensional symmetric cube cutting operads from
\cite{thu:ogatfp}*{Subsection 3.5}. They all satisfy the (cancellative) calculus of fractions.
Furthermore, they are monochromatic and possess operations of arbitrarily high degree. From Remarks \ref{92437} and
\ref{71138} it follows that all objects (except the uninteresting unit object) are split and progressive. So Theorem
\ref{20869} is applicable to these operads. Furthermore, the corresponding operad groups are all infinite by Proposition
\ref{96213} and non-amenable in the symmetric case by Proposition \ref{22947}.
\vspace{2mm}
Observe now the local similarity operads. Let $\mathrm{Sim}_X$ be a finite similarity structure on the compact ultrametric space $X$.
When choosing a ball in each $\mathrm{Sim}_X$--equivalence of balls, we obtain a symmetric operad with transformations $\mathcal{O}$.
The colors of $\mathcal{O}$ are the chosen balls. We choose $X$ for the $\mathrm{Sim}_X$--equivalence class $[X]$. We already know
that $\mathcal{O}$ satisfies the (cancellative) calculus of fractions. In \cite{sa-th:laa}*{Definition 3.1} we called $\mathrm{Sim}_X$ dually
contracting if there are two disjoint proper subballs $B_1,B_2$ of $X$ together with similarities $X\rightarrow B_i$ in $\mathrm{Sim}_X$.
This easily implies that $X$ is split.
\begin{lem}
The color $X$ is progressive provided $\mathrm{Sim}_X$ is dually contracting.
\end{lem}
\begin{proof}
Let $\theta=(f_1,...,f_l)$ be an operation with output $X$. This means that the $f_i\colon B_i\rightarrow X$ are
$\mathrm{Sim}_X$--embeddings (i.e.~$f_i$ yields a similarity in $\mathrm{Sim}_X$ when the codomain is restricted to
the image) such that the images of the $f_i$ are pairwise disjoint and their union is $X$. So the images
$\operatorname{im}(f_i)$ form a partition $\mathcal{P}$ of $X$ into balls. If we apply \cite{sa-th:laa}*{Lemma 3.7} to this
partition, we find a $j$ and a small ball $B$ which is $\mathrm{Sim}_X$--equivalent to $X$ and such that
$B\subset\operatorname{im}(f_j)$. Using this, we can construct an operation $\psi=(g_1,...,g_k)$ with codomain $B_j$ such
that $g_1\colon X\rightarrow B_j$. From Remark \ref{71138} it now follows that $X$ is progressive.
\end{proof}
Consequently, Theorem \ref{20869} is applicable to dually contracting local similarity operads. Furthermore, the corresponding operad
groups based at $X$ are all infinite by Proposition \ref{96213} and non-amenable by Proposition \ref{22947}.
\subsection{$L^2$--homology}\label{72248}
For a group $G$, let $l^2G$ be the Hilbert space with Hilbert base $G$. Thus, elements in $l^2G$ are formal sums
$\sum_{g\in G}\lambda_g g$ with $\lambda_g\in\mathbb{C}$ such that $\sum_{g\in G}|\lambda_g|^2<\infty$. Left multiplication with
elements in $G$ induces an isometric $G$--action on $l^2G$. Denote the set of $G$--equivariant linear bounded operators
$l^2G\rightarrow l^2G$ by $\mathcal{B}^G(l^2G)$, a subalgebra of the algebra of all bounded linear operators $\mathcal{B}(l^2G)$.
\emph{Right} multiplication with an element $\gamma\in G$ induces a $G$--equivariant linear bounded operator
$\gamma{\triangleright}\rho\colon l^2G\rightarrow l^2G$. This induces a homomorphism $\rho\colon\mathbb{C} G\rightarrow\mathcal{B}(l^2G)$ from the group ring
into the algebra of bounded linear operators, i.e.~$1{\triangleright}\rho=\mathrm{id}$ and $(\gamma_1\gamma_2){\triangleright}\rho=
(\gamma_1{\triangleright}\rho)*(\gamma_1{\triangleright}\rho)$. The closure of the image of this map with respect to the weak or strong
operator topology is called the von Neumann algebra $\mathcal{N} G$ associated to $G$. It is equal to the subalgebra of all
$G$--equivariant bounded linear operators $\mathcal{B}^G(l^2G)\subset\mathcal{B}(l^2G)$ \cite{lue:lta}*{Example 9.7}.
We will cite some known results about this von Neumann algebra in order to deduce a corollary for $l^2$--homology.
\begin{itemize}
\item ({\it $\mathcal{N}$ is inductive}) Let $H$ be a subgroup of $G$ and $A\in\mathcal{B}^H(l^2H)$. Then
$\mathbb{C} G\otimes_{\mathbb{C} H}l^2H\subset l^2G$ is a dense $G$--invariant subspace and
\[\mathrm{id}_{\mathbb{C} G}\otimes_{\mathbb{C} H}A\colon \mathbb{C} G\otimes_{\mathbb{C} H}l^2H\rightarrow \mathbb{C} G\otimes_{\mathbb{C} H}l^2H\]
is a $G$--equivariant linear map which is bounded with respect to the norm coming from $l^2G$. Consequently,
it can be extended to an element in $\mathcal{B}^G(l^2G)$. We obtain a map $\mathcal{N} H\rightarrow\mathcal{N} G$ which is an
injective ring homomorphism. So if $H<G$, then $\mathcal{N} H$ is a subring of $\mathcal{N} G$. Even more is true:
It is a faithfully flat ring extension \cite{lue:lta}*{Theorem 6.29}. From this, it
follows easily that the coefficient system $\mathcal{N}$ is inductive.
\item ({\it $\mathcal{N}$ is K\"unneth}) If $H_1,H_2$ are two subgroups of $G$ which commute in $G$,
i.e.~$h_1h_2=h_2h_1$ for all $h_1\in H_1$ and $h_2\in H_2$, then $\mathcal{N} H_1$ and $\mathcal{N} H_2$ commute in
$\mathcal{N} G$. In particular, $\mathcal{N} H_1\otimes_\mathbb{C}\mathcal{N} H_2$ is a subring of $\mathcal{N} G$. This implies, using a
standard homological algebraic argument \cite{lue:lta}*{Lemma 12.11(3)}, that $\mathcal{N}$ is K\"unneth.
\item ({\it $H_0$ and amenability}) Going back to a result of Kesten, the $0$'th group homology of a group $G$
with coefficients in the von Neumann algebra $\mathcal{N} G$ vanishes if and only if $G$ is non-amenable
\cite{lue:lta}*{Lemma 6.36}. So we have
\[H_0(G,\mathcal{N} G)=0\hspace{5mm}\Longleftrightarrow\hspace{5mm}G\text{ non-amenable}\]
\item ({\it Relationship with $l^2$--homology}) From \cite{lue:lta}*{Lemma 6.97} or \cite{lue:lta}*{Theorem 6.24(3)}
we get for groups $G$ of type $F_{\infty}$ and every $k\geq 0$
\[H_k(G,\mathcal{N} G)=0\hspace{5mm}\Longleftrightarrow\hspace{5mm}H_k(G,l^2G)=0\]
\end{itemize}
Applying Theorem \ref{20869} to these observations, we get the following corollary.
\begin{cor}
Let $\mathcal{O}$ be a planar or symmetric operad which satisfies the calculus of fractions.
Let $X$ be a split progressive object of $\mathcal{S}(\mathcal{O})$. Set $\Gamma:=\pi_1(\mathcal{O},X)$ and assume
that $\Gamma$ is non-amenable. Then
\[H_k(\Gamma,\mathcal{N}\Gamma)=0\]
for all $k\geq 0$. If $\Gamma$ is also of type $F_\infty$ (e.g.~if the conditions in \cite{thu:ogatfp}*{Theorem 4.3}
are satisfied), we also have
\[H_k(\Gamma,l^2\Gamma)=0\]
for all $k\geq 0$.
\end{cor}
From Proposition \ref{22947}, we get the following corollary.
\begin{cor}
Let $\mathcal{O}$ be a symmetric operad which satisfies the calculus of fractions.
Let $X$ be a split progressive object of $\mathcal{S}(\mathcal{O})$. Set $\Gamma:=\pi_1(\mathcal{O},X)$. Then
\[H_k(\Gamma,\mathcal{N}\Gamma)=0\]
for all $k\geq 0$. If $\Gamma$ is also of type $F_\infty$, we have
\[H_k(\Gamma,l^2\Gamma)=0\]
for all $k\geq 0$.
\end{cor}
From the remarks at the beginning of this section and from \cite{thu:ogatfp}*{Subsection 4.6}, we get the following corollary.
\begin{cor}\label{57450}
Let $\mathcal{O}$ be a symmetric cube cutting operad or a local similarity operad coming from a dually contracting finite similarity structure
$\mathrm{Sim}_X$. In the first case, let $A$ be any object in $\mathcal{S}(\mathcal{O})$ different from the monoidal
unit $I$. In the second case, let $A$ be the object $X$. Set $\Gamma=\pi_1(\mathcal{O},A)$. Then
\[H_k(\Gamma,\mathcal{N}\Gamma)=0\]
for all $k\geq0$. Assume furthermore that $\mathrm{Sim}_X$ is rich in ball contractions \cite{fa-hu:fpo}*{Definition 5.12},
in other words, the associated operad $\mathcal{O}$ is color-tame in the sense of \cite{thu:ogatfp}*{Definition 4.2}. Then we also have
\[H_k(\Gamma,l^2\Gamma)=0\]
for all $k\geq 0$.
\end{cor}
In particular, we obtain that the Higman--Thompson groups $V_{n,r}$ and the higher-dimensional Thompson groups $nV$ (see \cite{bri:hdt})
are $l^2$-invisible. This answers a question posed by L\"uck \cite{lue:lta}*{Remark 12.4}: The \emph{Zero-in-the-spectrum conjecture} by Gromov
says that whenever $M$ is an aspherical closed Riemannian manifold, then there is always a dimension $p$ such that
zero is contained in the spectrum of the minimal closure of the Laplacian acting on smooth $p$--forms on the
universal covering of $M$:
\[\exists_{p\geq 0}\ \ 0\in\operatorname{spec}\big(\operatorname{cl}(\Delta_p)\colon D\subset L^2\Omega^p(\widetilde{M})
\rightarrow L^2\Omega^p(\widetilde{M})\big)\]
By \cite{lue:lta}*{Lemma 12.3}, this is equivalent to
\[\exists_{p\geq 0}\ \ H_p(G,\mathcal{N} G)\neq 0\]
for $G=\pi_1(M)$. Dropping Poincar\'e duality from the assumptions, we arrive at the following question: If $G$ is a
group of type $F$ (i.e.~there exists a compact classifying space for $G$), then is there a $p$ with
$H_p(G,\mathcal{N} G)\neq 0$? Relaxing the assumption on the finiteness property, we arrive at the following question: If $G$ is a
group of type $F_\infty$, then is there a $p$ with $H_p(G,\mathcal{N} G)\neq 0$? Corollary \ref{57450}
gives explicit counterexamples to this question.
\subsection{Cohomology with coefficients in the group ring}
We want to apply the cohomological version of Theorem \ref{20869} to $\mathcal{M} G:=\mathbb{Z} G$. To
this end, we want to show that $\mathbb{Z} G$ is $FP_\infty$--K\"unneth and $FP_\infty$--inductive (in cohomology).
The first property follows from \cite{sa-th:laa}*{Proposition 4.3}. The second property follows from the
observation that $\mathbb{Z} G$ is a free $\mathbb{Z} H$--module if $H<G$ and that group cohomology of groups of type
$FP_\infty$ commutes with direct limits in the coefficients \cite{bro:cog}*{Theorem VIII.4.8}.
From Theorem \ref{20869}, Proposition \ref{96213} and $H^0(G,\mathbb{Z} G)=(\mathbb{Z} G)^G=0$ for infinite $G$, we obtain:
\begin{cor}
Let $\mathcal{O}$ be a planar or symmetric operad which satisfies the calculus of fractions.
Let $X$ be a split progressive object of $\mathcal{S}(\mathcal{O})$. Set $\Gamma:=\pi_1(\mathcal{O},X)$ and assume
that $\Gamma$ is of type $FP_\infty$ (e.g.~if the conditions in Theorem \cite{thu:ogatfp}*{Theorem 4.3} are satisfied). Then
\[H^k(\Gamma,\mathbb{Z}\Gamma)=0\]
for all $k\geq 0$.
\end{cor}
Recall that type $F_\infty$ implies type $FP_\infty$ and note that, in this case, $H^k(\Gamma,\mathbb{Z}\Gamma)=0$ for all $k\geq 0$ implies that $\Gamma$ has
infinite cohomological dimension \cite{bro:cog}*{Propositions VIII.6.1 and VIII.6.7}.
Unfortunately, this tells us that none of these groups can be of type $F$ and consequently, we cannot find such a group
which is also $l^2$-invisible.
From the remarks at the beginning of this section and from \cite{thu:ogatfp}*{Subsection 4.6}, we get the following corollary.
\begin{cor}
Let $\mathcal{O}$ be a planar or symmetric cube cutting operad or a local similarity operad coming from a dually contracting
finite similarity structure $\mathrm{Sim}_X$ which is also rich in ball contractions. In the first two cases, let $A$ be any object in
$\mathcal{S}(\mathcal{O})$ different from the monoidal unit $I$. In the last case, let $A$ be the object $X$. Set $\Gamma=\pi_1(\mathcal{O},A)$. Then
\[H^k(\Gamma,\mathbb{Z}\Gamma)=0\]
for all $k\geq 0$.
\end{cor}
In particular, we obtain $H^k(F,\mathbb{Z} F)=0$ and $H^k(V,\mathbb{Z} V)=0$ for all $k\geq 0$. This has been shown before in \cite{br-ge:ait}*{Theorem 7.2}
and in \cite{bro:fpo}*{Theorem 4.21}.
\begin{bibdiv}
\begin{biblist}
\bib{bri:hdt}{article}{
author={Brin, Matthew G.},
title={Higher dimensional Thompson groups},
journal={Geom. Dedicata},
volume={108},
date={2004},
pages={163--192}
}
\bib{bro:cog}{book}{
author={Brown, Kenneth S.},
title={Cohomology of groups},
series={Graduate Texts in Mathematics},
volume={87},
publisher={Springer-Verlag},
place={New York},
date={1994}
}
\bib{bro:fpo}{article}{
author={Brown, Kenneth S.},
title={Finiteness properties of groups},
booktitle={Proceedings of the Northwestern conference on cohomology of
groups (Evanston, Ill., 1985)},
journal={J. Pure Appl. Algebra},
volume={44},
date={1987},
number={1-3},
pages={45--75}
}
\bib{br-ge:ait}{article}{
author={Brown, Kenneth S.},
author={Geoghegan, Ross},
title={An infinite-dimensional torsion-free ${\rm FP}_{\infty }$ group},
journal={Invent. Math.},
volume={77},
date={1984},
number={2},
pages={367--381}
}
\bib{dav:pdg}{article}{
author={Davis, Michael W.},
title={Poincar\'e duality groups},
conference={
title={Surveys on surgery theory, Vol. 1},
},
book={
series={Ann. of Math. Stud.},
volume={145},
publisher={Princeton Univ. Press, Princeton, NJ},
},
date={2000},
pages={167--193}
}
\bib{fa-hu:fpo}{article}{
author={Farley, Daniel S.},
author={Hughes, Bruce},
title={Finiteness properties of some groups of local similarities},
journal={Proc. Edinb. Math. Soc. (2)},
volume={58},
date={2015},
number={2},
pages={379--402}
}
\bib{gro:lrm}{article}{
author={Gromov, M.},
title={Large Riemannian manifolds},
conference={
title={Curvature and topology of Riemannian manifolds},
address={Katata},
date={1985},
},
book={
series={Lecture Notes in Math.},
volume={1201},
publisher={Springer},
place={Berlin},
},
date={1986},
pages={108--121}
}
\bib{lue:lta}{book}{
author={L{\"u}ck, Wolfgang},
title={$L^2$-invariants: theory and applications to geometry and
$K$-theory},
series={Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematics},
volume={44},
publisher={Springer-Verlag},
place={Berlin},
date={2002}
}
\bib{sa-th:laa}{article}{
author={Sauer, Roman},
author={Thumann, Werner},
title={$l^2$-invisibility and a class of local similarity groups},
journal={Compos. Math.},
volume={150},
date={2014},
number={10},
pages={1742--1754}
}
\bib{thu:ogatfp}{article}{
author={Thumann, Werner},
title={Operad groups and their finiteness properties},
journal={{\tt arXiv:1409.1085v3}},
date={2016}
}
\end{biblist}
\end{bibdiv}
\end{document} |
\section{Introduction}
The solar wind is the supersonic outflow of fully ionized gas from the solar corona streaming through
the magnetic field lines.
It is well established that the polar coronal hole is the source of fast
solar wind (Hassler {\it et al.,} 1999; Wilhelm {\it et al.,} 2000; Tu {\it et al.,}~2005)
while the slow solar wind originates from the boundary of active regions
and along the streamers in the equatorial corona (Habbal {\it et al.,} 1997; Sakao, 2007).
Hassler {\it et al.} (1999) have found that the plasma can be supplied
from the chromospheric heights in the network boundaries to the
solar wind in polar coronal holes. More precise estimate of the formation of nascent
wind in the coronal funnels between 5\,--\,20 Mm in coronal holes has been carried out by Tu {\it et al.} (2005).
The bases of the polar coronal holes are mostly the sources of
fast solar wind. It has also been suggested that plasma outflows observed at the edges of active regions are the
source of the slow solar wind. Active-region arches that can extend outward as the rays in the outer corona
may channel it (Slemzin {\it et al.,} 2013). In addition to the large-scale origin, it has also been found that the
small-scale outflows at the coronal hole boundaries (CHBs) can serve as the source of the slow solar wind
(Subramanian, Madjarska, and Doyle, 2010). Recently, Yang {\it et al.} (2013) have reported a numerical model to describe the
process of magnetic reconnection between moving magnetic features (MMFs) and the pre-existing ambient magnetic field
that drives anemone jet with inverted y-shape base and associated plasma blobs. They have found that an increase in
the thermal pressure at the base of the jet is also driven by the reconnection, which induces a train of slow-mode
shocks propagating upward resulting in plasma upflows. Their findings contribute to the formation of
jets and small-scale flows in the quiet-Sun corona where MMFs undergo into the
low atmospheric reconnection.
Two outstanding issues, however, remain unsettled: i)
what are the drivers of these winds in the outer corona? and ii) what are the source regions
and what drivers enable the mass supply to the lower solar atmosphere (Chromosphere-TR, and inner corona)?.
There exist several
studies advocating the role of Alfv\'en waves as a possible answer to the first question. The ion-cyclotron waves
at kinetic scales are visualized as one of the possible candidates to provide momentum and heat the outer
coronal winds both in theory and observations ({\it e.g.} Ofman and Davila, 1995;
Tu and Marsch, 1997; Suzuki and Inutsuka, 2005; Srivastava and Dwivedi, 2006, Jian {\it et al.,} 2009, and references cited therein).
The second question is crucial at present.
The consensus of solar-wind research, however, is unclear as to the origin of the
mass supply to the supersonic wind. Tian {\it et al.} (2010) have found upflows in the open field-lines of
coronal hole starting in the solar transition region and interpreted this as an
evidence of the
fast solar wind in the polar coronal holes.
As far as the slow solar wind is concerned, the outflows at the boundaries of active regions
can contribute at larger spatio--temporal scales to the mass supply as an expansion of the loops
lying over these active regions (Harra {\it et al.,} 2008). It has been shown recently that collimated
jet eruptions can also contribute to the formation of the solar wind (Madjarska, 2013). The magnetic-field topology of
structures such as jets ({\it e.g.} spray surges) may
not contribute to the solar wind (Uddin {\it et al.,} 2012). The contributions from the confined ejecta
in the solar-wind formation depend mainly on the local magnetic-field topology and
plasma conditions.
The quest continues on whether the mass supply to the slow solar wind comes from the lower
atmosphere expanding along the curved coronal fields. The question is what are the potential physical
drivers? It has been found that the open field lines at the boundary of active regions
reconnect periodically with closed field lines to guide the plasma motion in the form of
solar wind (Harra {\it et al.,} 2008). Similar examples are reported at the boundary of the
coronal holes at small spatial scales as the source of slow solar wind (Subramanian, Madjarska, and Doyle, 2010).
Therefore, an alternative option may be magnetic reconnection as a potential mechanism in the formation of slow solar wind.
Apart from magnetic reconnection, the wave-heating scenario
can shed light on the slow solar-wind source regions. Schmidt and Ofman (2011)
have found that the energy stored in the slow magnetoacoustic waves propagating towards the higher atmosphere within expanding loops. This may be a potential candidate for the acceleration
and formation of the slow solar wind. Wave activity at the bases
of the fast (polar coronal holes) and slow (equatorial corona) solar wind can be important to power the energized plasma at greater heights up to
the corona where it can be triggered supersonically in interplanetary space
({\it e.g.} Harrison {\it et al.,} 2002; Dwivedi and Srivastava 2006, 2008; De Pontieu {\it et al.,} 2007; McIntosh {\it et al.,} 2011, McIntosh 2012,
and references cited therein).
Thus, there seems to be compelling evidence for the role of magnetic reconnection and
wave phenomena in the solar wind source region.
In the present article, we report the evidence of the outflowing magnetic arches acting as coronal funnels
at the eastern boundary of an AR 10940 loop system observed on 5 February 2007. These coronal funnels seem to open up
in the higher atmosphere to transport the outflowing plasma. Their footpoints are rooted
in the boundary of the active region. They are the most likely heated regions that result in the activation of the outflowing plasma.
We present a 2D MHD simulation of the open and expanding funnel-type
model atmosphere in which a rarefied and hot region is implemented near the footpoint that
exhibits plasma perturbations similar to the observations.
In Section~2 we summarize the observational results. We present the numerical model in Section~3.
Discussion and conclusions are given in the last section.
\begin{figure*
\begin{center}
\mbox{
\hspace{-2.0cm}
\includegraphics[width=10.00cm,angle=0]{Obs/FOV-INT.ps}
\hspace{-2.5cm}
\includegraphics[width=10.00cm,angle=0]{Obs/Dop_Vel_FOV}}
\caption{\small Intensity (left) and Doppler velocity (right) maps of Active Region 10940 observed on 5 February 2007.
The core loops and the open arches at its eastern boundary are clearly evident.
}
\label{fig:obs-int}
\end{center}
\end{figure*}
\begin{figure*
\begin{center}
\includegraphics[width=8.00cm,angle=0]{Obs/pfss_roi.ps}
\caption{\small A view of the magnetic-field polarities of AR 10940
and sorrounding regions as observed by {\it Solar} {\it and} {\it Heliospheric} {\it Observatory} {\it (SOHO)}/{\it Michelson} {\it Doppler} {\it Imager} (MDI), and its
Potential Field Source Surface Extrapolation (PFSS).
The black arrow indicates the core loop system in the AR while the
south--east part of its boundary shows the large-scale open-field lines
extended into the upper corona. Some of these lines also connect to
the coronal hole lying in the North of AR.
}
\label{fig:obs-int}
\end{center}
\end{figure*}
\section{Observational Results}
\begin{figure*
\begin{center}
\hspace{-4.0cm}
\includegraphics[width=13.00cm,angle=0]{Obs/Dop_Vel_Path.ps}
\mbox{
\hspace{-4.0cm}
\includegraphics[width=8.00cm,angle=0]{Obs/Velocity_Path1.ps}
\hspace{-0.5cm}
\includegraphics[width=8.00cm,angle=0]{Obs/Velocity_Path2.ps}}
\mbox{
\hspace{-4.0cm}
\includegraphics[width=8.00cm,angle=0]{Obs/Velocity_Path3.ps}
\hspace{-0.5cm}
\includegraphics[width=8.00cm,angle=0]{Obs/Velocity_Path4.ps}
}
\caption{\small Top: the enlarged velocity map of the eastern
boundary of AR 10940 showing the open and expanding funnels
in which the plasma is outflowing. Middle and Bottom: the variation of L.O.S.
Doppler velocities along funnels 1 to 4.
}
\label{fig:obs-vel}
\end{center}
\end{figure*}
The active region AR 10940 was observed in the one-arcsecond-slit scan of {\it EUV Imaging Spectrometer} (EIS: Culhane {\it et al.}, 2006)
onboard the {\it Hinode} spacecraft on 05 February 2007. The EIS is an imaging spectrometer of which 40- and 266-arcsecond slots are used for the image analyses
using the light curves and emissions per pixel. The one- and two-arcsecond slits are utilized for spectral and Doppler analyses
using spectral-line profiles. EIS observes in two modes : i) scan; ii) sit-n-stare.
The EIS observes high-resolution spectra in two wavelength intervals :
170\,--\,211 \AA\ and 246\,--\,292 \AA~using respectively its short-wavelength (SW) and long-wavelength (LW) CCDs.
The spectral resolution of the EIS is 0.0223 \AA~per pixel. The analysed observations well taken on 05 February 2007
and the data-set contains spectra of
various lines formed at chromospheric, transition region (TR), and coronal temperatures.
The scanning observation started at 12:14:12 UT and ended at 13:31:21 UT on 05 February 2007.
The scanning steps were without any off-set in the region containing a coronal active region and
its eastern boundary with open and expanding magnetic arches where we are interested in the present investigation. This provides us an opportunity
to understand the plasma activity along the open-field regions at the eastern boundary of the active region in between
the diffused loop systems that reach up to a higher height
in the corona (Figure 1). We refer to these structures as $"$coronal funnels$"$. Such flow regions are heated at their base exhibiting outflows.
To understand the approximate magnetic-field geometry associated with AR 10940 and its sourrounding
region, we perform a Potential Field Source Surface (PFSS) extrapolation. Fan lines
over the {\it Solar and Heliospheric Observaory (SOHO)/Michelson Doppler Imager} (MDI) observations are shown in Figure 2. It is clear that the core loops of the active region
are bipolar and connect the central opposite magnetic polarities (white loops shown by the black arrow). Another
set of closed-field lines connect to the east-side weak negative polarity and the central positive polarity. At the eastern
boundary, there are weak patches of positive and negative polarities from where the open-field lines extend up to higher corona. These open-field lines form the large-scale coronal funnels at the eastern boundary of the active region. The lower parts
of the loops at the eastern boundary can also serve as funnels up to a certain height in the corona. The white lines represent
closed magnetic fields, while magenta and green lines represent the open fields that reach at the source surface, having opposite polarities. The funnels and the south-east boundaries of the AR are associated with the large-scale open field lines (magenta). They may also be the lower parts of the closed-loop system (white lines).
In order to obtain the
velocity structures in such coronal funnels, we select the strongest
EIS line, Fe \textsc{xii} 195.12 \AA\ in our study. We aim for the understanding of the
impulsively generated plasma outflows in such funnels and associated physical
processes. The slit step started the scanning of the polar coronal
hole with ($X_{\mathrm{cen}},Y_{\mathrm{cen}})$$\approx$(799.087 arcsecond, -19.185333 arcsecond).
The observational windows acquired on the CCDs are 128 pixels high
along the slit, while 111 pixels wide in the horizontal direction where spectra also disperse
with the spectral resolution of 0.0222 \AA per pixel.
\begin{figure*
\begin{center}
\mbox{
\hspace{-4.0cm}
\includegraphics[scale=0.45, angle=0]{figs/fig1a.eps}
\hspace{-1.0cm}
\includegraphics[scale=0.45, angle=0]{figs/rho_e.eps}
}
\hspace{-4.0cm}
\includegraphics[scale=0.45, angle=0]{figs/fig1b.eps}
\caption{\small
Equilibrium profiles of temperature (top-left panel), mass density (top-right panel),
and plasma-$\beta$ (bottom panel) along the vertical line [$y$] for $x=0$.
}
\label{fig:initial_profile}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[scale=0.5, angle=0]{figs/blocks.eps}
\vspace{-1.0cm}
\caption{\small Numerical blocks with their boundaries (solid lines).
}
\label{fig:blocks}
\end{center}
\end{figure}
We apply the standard EIS data-reduction procedures and calibration files/routines
to the data obtained from
the EUV-telescope, which is the raw (zeroth-level) data. The subroutines are found
in the {\sf sswidl} software tree under the IDL environment ({\sf www.darts.isas.jaxa.jp/pub/solar/ssw/hinode/eis/}).
These standard subroutines reduce the dark-current
subtraction, cosmic-ray removal, flat-field correction, hot pixels, warm pixels, and
bad/missing pixels. The data are saved in the level-1 data file, and the
associated errors are saved in the error file.
We choose the clean and strong line Fe \textsc{xii} 195.12 \AA~to examine the spatial variations
of the intensity, Doppler velocity in the observations scanned over AR 10940 on 5 February 2007.
We co-align the Fe \textsc{xii} 195.12 \AA~map with respect to the long-wavelength CCD observations
of He \textsc{ii} 256.86 \AA~by considering it as a reference image and by estimating
its offset. The orbital and slit-tilt are also corrected in the data using
the standard method described in the EIS-software notes.
We perform a double Gaussian fit for the removal of the weak blending
of Fe \textsc{xii} 195.18 \AA~line as per the procedure described by Young {\it et al.} (2009),
which is also outlined in the EIS Software Note 17.
We constrain the weakly blended line 195.18 \AA~to have the same width as the line at 195.12 \AA~and
to fix an offset of +0.06 \AA~relative to it. It is be noted that
we perform the fitting on 2$\times$2 pixel$^{2}$ binned data to enhance
the signal-to-noise ratio and also to be obtain a reasonable fit.
Figure~1 displays the intensity (left) and Doppler-velocity (right) maps
of the observed AR and open arches at its eastern boundary.
The intensity map shows that the AR is made up of
the core diffused loop systems lying at the lower heights
in the perpendicular plane to the line-of-sight. In actuality, however, they
may be tilted. These low-lying core loops exhibit high emission and downflows.
At the eastern boundary, we have identified the open and expanding arches
that exhibit coronal funnel behaviour to transport mass and energy
into the higher corona.
These expanding funnels may be the part of more quiescent and large-scale
loop systems opening higher up in the corona.
It is seen in the enlarged Doppler map (see Figure 3) that four such regions are
at least identified as blue-shifted and outflowing regions. As already noted, these could be the legs
of higher and large-scale loops, which form the building blocks
of the coronal funnels at relatively lower heights of the corona. They exhibit the slow and sub-sonic plasma outflows
mostly spreading along their field lines with a maximum speed
of about 10\,--\,15 km~s$^{-1}$.
Figure~3 shows the selection of Path 1\,--4\, (top-panel). The first two paths are drawn over
the large-scale open and magnetic-field arches (lines) lying in a QS\,-\,CH
region in the North East of this AR (not shown here). These two regions serve
as the open and expanding funnel regions. Paths 3 and 4 are drawn to the
expanding and blue-shifted regions that can also serve as coronal funnels. However,
they are actually the lower parts of some core-loop systems. On this account, we have
selected two coronal funnels (1 and 2), while the others (3 and 4) serve as funnels but are probably
associated at larger heights with the
curved-loop magnetic field.
The middle and bottom panels (left to right) respectively show the variation of projected
L.O.S. Doppler velocity along Paths 1\,--4\, with height in these funnels.
For funnels 1 and 2, the footpoints exhibit stronger outflows with a
maximum speed of 15 km~s$^{-1}$. At heights beyond 10 Mm,
the outflow speed weakens in these funnels. This signifies the start of the
outflows due to heating near the footpoints of these funnels; the outflows weaken
with the height as we move away from the heating source.
The other two funnels (3 and 4) are likely the legs of the diffused core loops,
exhibiting the outflows at certain lower heights with a maximum
speed of 10 km~s$^{-1}$. This indicates different locations of the
heating. It is also noted that the outflows
diminish at lower heights compared to the corressponding ones in Funnels 1 and 2 in the form of open arches. This
is because of their association with the curved loops at greater heights where
plasma is trapped and flows downward in order to maintain a new equilibrium.
It is also noted that the flow structure is more gentle in the open Funnels 1 and 2. However,
it is greater at the footpoint and decreases with height. In the Funnels 3 and 4, which are
the lower parts of curved loops, the generation of the flow is rather impulsive. The outflow starts
at a certain height above the loop's footpoint. It increases up to a certain distance and decreases thereafter.
This shows that impulsive heating is at work near the loop footpoint, which causes enhanced
upflows up to a certain distance. The downflowing plasma from the upper part of the loop
may counteract with the upflows. The red-shifted apex of the core loop system is clearly evident
in Figures 1 and 3. Funnels 3 and 4 are the lower parts of this loop system.
In the next section, we outline the details of the 2D numerical
simulation of such observed expanding coronal funnels and their plasma dynamics. We solve a set of ideal MHD equations
in the appropriate VAL-III C initial temperature
conditions and model atmosphere using the FLASH code.
\section{Numerical Model of the Hot Plasma Outflows in Coronal Funnels}
\subsection{Model Equations}
Our model system acquires a
gravitationally stratified solar atmosphere
which can be described by
the ideal two-dimensional (2D)
MHD equations:
\begin{equation}
\label{eq:MHD_rho}
\hspace{-3.3cm}
{{\partial \varrho}\over {\partial t}}+\nabla \cdot (\varrho\textbf{\textsl{V}})=0\, ,
\end{equation}
\begin{equation}
\label{eq:MHD_V}
\varrho{{\partial \textbf{\textsl{V}}}\over {\partial t}}+ \varrho\left (\textbf{\textsl{V}}\cdot \nabla\right )\textbf{\textsl{V}} =
-\nabla p+ \frac{1}{\mu}(\nabla\times\textbf{\textsl{B}})\times{\textbf{\textsl{B}}} +\varrho{\textbf{\textsl{g}}}\, ,
\end{equation}
\begin{equation}
\label{eq:MHD_p}
\hspace{-1.3cm}
{\partial p\over \partial t} + \nabla\cdot (p\textbf{\textsl{V}}) = (1-\gamma)p \nabla \cdot\textbf{\textsl{V}}\, ,
\end{equation}
\begin{equation}
\label{eq:MHD_B}
\hspace{1.8cm}
{{\partial\textbf{\textsl{B}}}\over {\partial t}}= \nabla \times (\textbf{\textsl{V}}\times\textbf{\textsl{B}})\, ,
\nabla\cdot\textbf{\textsl{B}} = 0\, .
\end{equation}
Here ${\varrho}$, $\textbf{\textsl{V}}$, $\textbf{\textsl{B}}$, $p = \frac{k_{\rm B}}{m} \varrho T$, $T$, $\gamma=5/3$,
$\textbf{\textsl{g}}=(0,-g)$ with its value $g=274$ m s$^{-2}$, $m$, and $k_{\rm B}$ are
the mass density, flow velocity, magnetic field, gas pressure, temperature,
adiabatic index, gravitational acceleration, mean particle mass, and Boltzmann's constant respectively.
It should be noted that we do not consider radiative cooling and
thermal conduction in our present model for the sake of simplicity. We simulate only the dynamics of the
plasma outflows to compare them with that of the observed coronal funnels.
\begin{figure*
\begin{center}
\vspace{-0.25cm}
\mbox{
\includegraphics[scale=0.75, angle=0]{figs/V/V1.eps}
\includegraphics[scale=0.75, angle=0]{figs/T/T1.eps}
\includegraphics[scale=0.75, angle=0]{figs/D/D1.eps}
}
\vspace{-0.50cm}
\mbox{
\includegraphics[scale=0.75, angle=0]{figs/V/V2.eps}
\includegraphics[scale=0.75, angle=0]{figs/T/T2.eps}
\includegraphics[scale=0.75, angle=0]{figs/D/D2.eps}
}
\vspace{-0.50cm}
\mbox{
\includegraphics[scale=0.75, angle=0]{figs/V/V3.eps}
\includegraphics[scale=0.75, angle=0]{figs/T/T3.eps}
\includegraphics[scale=0.75, angle=0]{figs/D/D3.eps}
}
\vspace{-0.50cm}
\mbox{
\includegraphics[scale=0.75, angle=0]{figs/V/V4.eps}
\includegraphics[scale=0.75, angle=0]{figs/T/T4.eps}
\includegraphics[scale=0.75, angle=0]{figs/D/D4.eps}
}
\vspace{0.50cm}
\caption{\small Numerical results (top to bottom rows): the velocity vectors superposed over the total velocity (left); the temperature (middle) and density (right) maps, for {\it t}\,=\,50, 100, 200, 300 seconds. White lines represent the diverging magnetic-field lines. The complete movies are given as an electronic
supplementary material for these numerical results.
}
\label{fig:num-V}
\end{center}
\end{figure*}
\subsection {Initial Conditions}
We assume that the solar atmosphere is at rest $\left[\textbf{\textsl{V}}_{\rm e}={\bf 0}\right]$ in equilibrium with a current-free
magnetic field,
$
\left[\nabla \times {\textbf{\textsl{B}}}_{\rm e}={\bf 0}\right].
$
As a result, the magnetic field is force-free
\begin{equation}\label{eq:B_e}
(\nabla\times{\textbf{\textsl{B}}}_{\rm e})\times{\textbf{\textsl{B}}}_{\rm e} = {\bf 0}\ .
\end{equation}
The divergence-free constraint is satisfied automatically if
the magnetic field is specified by the magnetic-flux function, $\left[A(x,y)\right]$,
as
\begin{equation}\label{eq:B_e1}
{\textbf{\textsl{B}}}_{\rm e}=\nabla \times (A\hat {\bf z})\, .
\end{equation}
Here the subscript $'e'$ corresponds to equilibrium quantities.
We set a curved magnetic field by choosing
\begin{equation}
A(x,y) = S\frac{x-a}{(x-a)^2+(y-b)^2},
\end{equation}
where
$S$
is
the strength of magnetic pole and $(a,b)=(0,-10)$ Mm is its position.
For such a choice of $(a,b)$, the magnetic-field vectors are weakly curved
to represent the expanding coronal funnels.
As a result of Equation~(\ref{eq:B_e})
the pressure gradient is balanced by the gravity force,
\begin{equation}
\label{eq:p}
-\nabla p_{\rm e} + \varrho_{\rm e} {\bf g} = {\bf 0}\, .
\end{equation}
With the ideal-gas law and the $y$-component of
Equation~(\ref{eq:p}), we
arrive at
\begin{equation}
\label{eq:pres}
p_{\rm e}(y)=p_{\rm 0}~{\rm exp}\left[ -\int_{y_{\rm r}}^{y}\frac{\bf \mathrm{dy^{'}}}{\Lambda (y^{'})} \right]\, ,\hspace{3mm}
\label{eq:eq_rho}
\varrho_{\rm e} (y)=\frac{p_{\rm e}(y)}{g \Lambda(y)}\, ,
\end{equation}
where
\begin{equation}
\Lambda(y) = \frac{k_{\rm B} T_{\rm e}(y)}{mg}
\end{equation}
is the pressure scale height, and $p_{\rm 0}$ denotes the gas
pressure at the reference level that we choose in the solar corona at $y_{\rm r}=10$ Mm.
We take
an equilibrium temperature profile $\left[T_{\rm e}(y)\right]$ (see top-left panel in Figure~4) for the solar atmosphere
derived from the VAL-C atmospheric model of Vernazza, Avrett, Loeser (1981) that is smoothly extended to the solar corona.
The transition region
is located at $y\approx 2.7$ Mm.
There is an extended solar corona above the transition region,
and with the temperature minimum level located at $y\approx0.9$ Mm below the solar chromosphere.
Having specified $T_{\rm e}(y)$ (see Figure~4, left-top panel) with Equation~(\ref{eq:pres}),
we obtain the corresponding gas-pressure and mass-density profiles.
\begin{figure*
\begin{center}
\mbox{
\includegraphics[scale=1.25, angle=0]{figs/Vel-T50.eps}
\includegraphics[scale=1.25, angle=0]{figs/Vel-T100.eps}
}
\mbox{
\includegraphics[scale=1.25, angle=0]{figs/Vel-T200.eps}
\includegraphics[scale=1.25, angle=0]{figs/Vel-T300.eps}
}
\caption{\small
Velocity profiles plotted w.r.t the height along the model funnel at
50, 100, 200, 300 seconds.}
\label{fig:num-V}
\end{center}
\end{figure*}
\subsection{Numerical Scheme and Computational Grid}
Equations (\ref{eq:MHD_rho})\,--\,(\ref{eq:MHD_B}) are solved numerically using the FLASH code
(Lee and Deane, 2009; Lee, 2013). This code uses a second-order unsplit Godunov solver (Godunov, 1959)
with various slope
limiters and Riemann solvers, as well as adaptive mesh refinement (AMR) (Murawski and Lee, 2011, 2012).
We set the simulation box
of $(-7,7)\, {\rm Mm} \times (2.6,51.6)\, {\rm Mm}$ along the $x$- and $y$-directions
(Figure~5). The lower boundary, where we apply the heating pulse, is set at
$x_{\rm 0}=0$ Mm, $y_{\rm 0}=2.6$ Mm. We set and hold fixed all of the plasma quantities at all boundaries of the simulation region to their equilibrium values
which are given by Equations~(\ref{eq:B_e1}) and (\ref{eq:eq_rho}).
As the magnetic field is curved and plasma is stratified gravitationally, open boundaries would not be a perfect choice with regard to the fixed boundary conditions. We have verified this experimentally. As the FLASH code uses a third-order accurate Godunov-type method, a characteristic method is already built into this procedure. However, the Riemann problem at the boundaries corresponds to open boundaries, resulting in numerically induced reflections from these boundaries. On the other hand, the fixed boundaries lead to negligibly small numerical reflections. Therefore, we adopt this approach in the numerical simulations.
In addition, we modify the equilibrium mass density and gas pressure at the bottom boundary as
\begin{equation}
\varrho(x,y,t) = \varrho_{\rm e}(y)
\left[
1 + A_{\rm \varrho} f(x,y) g(t)
\right]
\, ,
\label{eq:perturb}
\end{equation}
\begin{equation}
p(x,y,t) = p_{\rm e}(y)
\left[
1 + A_{\rm p} f(x,y) g(t)
\right]
\, ,
\label{eq:perturb}
\end{equation}
where
\begin{equation}
f(x,y) =
\exp\left[
-\frac{(x-x_{\rm 0})^2} {w_{\rm x}^2}
-\frac{(y-y_{\rm 0})^2} {w_{\rm y}^2}
\right]
\end{equation}
and
\begin{equation}
g(t) = -
\left[
\exp\left(
-\frac{t}{\tau}
\right) - 1
\right]
\, .
\end{equation}
Here $A_{\rm \varrho}$ and $A_{\rm p}$ are the amplitudes of the perturbations, $(x_{\rm 0},y_{\rm 0})$
are their initial position and
$w_{\rm x}$, $w_{\rm y}$ denote its widths along the $x$- and $y$-directions, respectively.
The symbol $\tau$ denotes the growth time of these perturbations.
We set and hold fixed $A_{\rm \varrho}=7$,
$A_{\rm p}=2$,
$x_{\rm 0}=0$ Mm, $y_{\rm 0}=2.6$ Mm,
$w_{\rm x}=3$ Mm, $w_{\rm y}=0.5$ Mm, and $\tau=10$~seconds.
These fixed boundary conditions perform much better than transparent boundaries,
leading to only negligibly small numerical
reflections of the wave signals from these boundaries.
In our modeling, we use an AMR grid with a minimum (maximum) level of
refinement set to $3$ ($8$) (see Figure~5). The refinement strategy is based on
controlling numerical errors in
mass density, which results in an excellent resolution of steep spatial profiles and
greatly reduces numerical diffusion at these locations.
A standard procedure to check the magnitude of a numerically induced flow is to run the code for the equilibrium alone,
without implementing any perturbation.
We have verified that the numerically induced flow is of the order of $2$ km s$^{-1}$ in the solar corona,
and the transition region is not affected by the spatial resolution which is about $20$ km around
the transition region.
This resolution is much smaller than the width of the transition region ($\approx 200$ km)
and the pressure scale height near the bottom, which is about $0.5-1$ Mm.
We have also performed grid convergence studies by increasing the spatial resolution by a factor of two at the transition
region. As the numerical results have been found essentially similar results for the finer and coarser grid, we have limited
our analysis to the latter.
\subsection{Results of the Numerical Simulation and Comparison with Observations}
The results of the numerical simulations and their comparison with the observed plasma outflows are summarized as follows:
Figure~6 (first column) displays the velocity vectors plotted over the total velocity maps for $t=50$,
$100$, $150$, $200$, $250$, $300$, $350$, $400$, $450$ seconds.
The diverging magnetic-field lines of the model funnel is also over-plotted
in these snapshots. It is clear from the $t=50$ seconds snapshot that
the implemented heating at the footpoint, just
below the transition region, results in the alteration of the ambient plasma pressure, and the plasma starts
flowing upward with a typical average speed of 40 km s$^{-1}$ near the footpoint at this time. The outflow velocity
is maximum near the heated region at the footpoint of the model funnel. This is higher than the observed line-of-sight outflow velocities (10\,--\,16 km s$^{-1}$) in various coronal funnels above their footpoints (see Figure~3).
The simulated outflow velocities greatly depend on the initial conditions of
the model funnel and magnitude of the heating pulse.
The observed trend of the flows in Funnel 1 and 2 (see Figure 3) and the same derived from the model match with each other (see Figures 6 and 7). At higher altitudes above the heating location in these funnels, the outflow weakens
which qualitatively match with the results obtained from our model (see Figures 6 and 7). Our model better fits the plasma flow conditions in the open coronal funnels ({\it e.g.} funnels 1 and 2) where the gentle flows start and expand due to their footpoint heating.
The flow becomes steady at each height after 300 seconds in the model funnel when the
heated plasma reach the height of 10 Mm. This also indicates that
the heating pulse is at work for a certain duration, and after some time the generated flows reach
a new equilibrium. The physical scenario is in agreement with the open coronal Funnels 1 and 2
(see Figures 3 and 7). Comparison of these two scenarios supports the outflow of the plasma due to heating. The heating causes
thermal flows guided in the magnetic-field lines of the open coronal funnels while it subsides away from the source.
The physical behaviour of the velocity field and its spatial distribution as observed by Hinode/EIS along each funnel
are consistent with the velocity field at a particular temporal span of the numerical simulation when
the outflowing plasma rise to a maximum height of $\approx$10 Mm.
For the funnels
that are the lower parts of the curved loop system (Funnels 3 and 4; Figure 3), the energy release might occur where blue-shift is enhanced at a certain height above the footpoint impulsively. Therefore, the locations in those observed funnels are identical to the energy release site of the modeled funnel where outflows start as a result of heating. On the contrary, it seems that impulsive heating in Funnels 3 and 4 causes the increment in the outflows up to a certain height and thereafter is balanced by
the downflowing and trapped plasma from the loop apex.
Figures~6, second and third columns, respectively display the temperature and density maps for $t=50$,
100, 200, 300 seconds. It is clear that during the heating, the
plasma maintained at inner coronal/TR temperatures (sub-MK and 1.0 MK; the hot plasma envelops the cool one) and with somewhat higher density, starts flowing from the footpoint
of the model funnel towards greater heights. The plasma is denser
near the footpoint. It flows streamline along the open funnels towards higher
heights. Thermal perturbations created the slow and subsonic flows
of the plasma; therefore, it reaches up to only the lower heights.
This is a similar situation that of the transition region within the funnel being pushed upward
due to the evolution of the thermal perturbation underneath.
Sub-mega-Kelvin and denser plasma from lower solar atmosphere move up and is
enveloped by the hot coronal plasma.
The denser plasma maintained at
the TR and inner coronal temperatures is visible up to a height of 10 Mm in the
model funnel. This is consistent with the
observations. Only the lower parts of the coronal arches are associated with enhanced fluxes and thus
with higher densities, while their higher parts are less intense and denser regions (see Figure 1, left panel).
\section{Discussion and Conclusions}
We have presented observations of outflowing coronal arches (coronal Funnels 1 and 2) lying at the
eastern boundary of the AR 10940 observed on 5 February 2007, 12:15\,--\,13:31 UT.
The scanning observations show that these arches open up in the nearby
quiet-Sun corona and exhibit plasma outflows maintained at coronal/TR temperature of around 1.0 MK.
They serve as the expanding coronal funnels from which the plasma moves to the
higher corona and serves as a source to the slow solar wind (Harra {\it et al.,} 2008).
The plasma outflows may be generated in such open field regions
because of the low atmospheric reconnections between the open- and closed-
field lines (Subramanian {\it et al.,} 2010). Episodic heating mechanisms
are now well observed and interpreted as the drivers of
outflowing plasma in the curved coronal loops (Klimchuk 2006; Del Zanna 2008; Brooks and Warren 2009).
The steady heating may generate the hot-plasma upflows in the corona (Tripathi {\it et al.,} 2012).
There has so far been little effort to model the plasma outflows in the corona and that too in an entirely different context
of the large-scale evolution of coronal magnetic field.
Murray {\it et al.} (2010) have modeled in 3D the origin and driver of coronal outflows,
and found that outflows are the result of the expansion of an active region during its development.
Harra {\it et al.} (2012) have modeled the AR coronal outflows as a consequence of compression
during the creation and annihilation of the magnetic-field lines.
However, our objective of the present investigation is to implement and test our 2D numerical
simulation model of the localized expanding coronal funnels with
{\it Hinode}/EIS observations of the outflowing open-field arches ({\it i.e.} funnels) at the
boundary of AR 10940. We solve the set of ideal MHD equations
in the appropriate VAL-III C initial temperature
conditions and model atmosphere using the FLASH code.
The key engredient of our model is the implementation of realistic ambient solar atmosphere, {\it e.g.}
realistic temperature, presence of the TR, implementation of expanding coronal fields, stratified
atmosphere in the initial equilibrium, which affect significantly such plasma dynamics.
We have implemented a rarefied and hotter region at the footpoint of the model funnel that
triggers the evolution of the slow and sub-sonic plasma perturbations propagating
outward in the form of plasma flows similar to the observed
dynamics.
We implemented the localized heating
below the transition region at the footpoint of the funnel.
The plasma is considered rarefied in the horizontal direction, mimicking
the structured, open, and expanding magnetic funnels. The heated plasma
evolves and exhibits the plasma perturbations as has been observed as
outlined in the {\it Hinode}/EIS observations. The outflows start at the base
of the funnel that further weakens with the height, which is suggestive of the
plasma dynamics due to heating near the footpoint. A similar physical
scenario is observed in the selected outflowing magnetic arches mimicking
the coronal funnels 1 and 2 at the eastern boundary of AR 10940.
We conclude that the implemented episodic heating
can excite plasma outflows in the solar active regions. These
slow, subsonic plasma outflows may not be launched at higher altitudes in the corona.
We have examined the presence of hot and denser plasma up to 10 Mm height
as triggered by thermal perturbations in our model. Observations also
show that the plasma only rises significantly in the lower parts of these
funnels up to inner coronal heights with significant intensity and velocity distributions. However, even if such flows reach up to
the inner coronal heights of 10 Mm in these funnels, they may also contribute
to the mass supply to the slow solar wind. Therefore, our model and observations invoke
the dynamics of the plasma in the localized coronal funnels, which
may be important candidates to transport mass and energy into the inner corona.
However, more observational studies should be performed
out new spectroscopic data ({\it e.g.} {\it Interface} {\it Region} {\it Imaging} {\it Spectrograph} (IRIS)) to compare with our proposed 2D model, specifically
under the physical conditions of different types of localized fluxtubes in the solar atmosphere,
which can serve as plasma outflowing regions due to episodic heating.
\begin{acks}
We thank the referees for their valuable suggestions
which improved the manuscript
considerably.
We acknowledge the use of {\it Hinode}/EIS observations.
AKS acknowledges the support of the International Exchange Scheme (Royal Society) between Indian and UK.
The software used in this work was in part developed by the DOE-supported ASCI/Alliance Center
for Astrophysical Thermonuclear Flashes at the University of Chicago.
\end{acks}
|
\section{introduction}
Students learn in the early days of their studies that the electromagnetic field can be described by a single quantity, the vector potential $A_{\mu}$, instead of the separate electric and magnetic fields \cite{jackson}. However, two distinct vector potentials $A_{\mu}$ and $A_{\mu}'=A_{\mu}+\partial_{\mu}\Lambda(x)$ (where $\Lambda(x)$ is a scalar field, the gauge field, that satisfies the Helmholtz equation) generate the same electric and magnetic fields. In order to remove this ambiguity, one should first \emph{fix the gauge}, namely determine uniquely the gauge function $\Lambda(x)$. Once the gauge has been chosen, the correspondence between the vector potential and the electromagnetic fields is made unique. The electromagnetic field, however, is not the only gauge field in Nature. In fact, all the fundamental forces, such as the weak force (responsible for radioactive decay and nuclear fusion of subatomic particles) and the strong force (that ensures the stability of ordinary matter) are both described by gauge fields \cite{weinberg}.
Since the quantum theory of a gauge field is often described by means of the path integral method \cite{das}, a generalization of the concept of gauge fixing within this framework is of paramount importance for the development of the theory itself. The Faddeev-Popov method exactly complies with this needing. From a rigorous point of view, this method consists in using a gauge fixing condition to reduce the number of allowed orbits of the mathematical configurations that represent a given physical system to a smaller set, where all the orbits are related by a smaller gauge group symmetry \cite{weinberg}. In simpler words, the Faddeev-Popov method consists in applying a constraint to the considered field, that automatically implements the gauge fixing condition, thus determining the field unambiguously \cite{brown}.
Usually, this method is object of doubts and of difficult understanding for the students (either graduate or undergraduate) that encounter it for the first time, and the development of this method reported in standard quantum field theory books \cite{brown,ryder,peskin} is not helpful for having a better understanding of it. In fact, most of the times it appears more like a magic trick rather than motivated by some physical argument.
The Legendre transform, on the other hand, is a very simple mathematical instrument that a student knows very well, since it finds application in a wide variety of physical problems, like the transformation between the Hamiltonian and the Lagrangian function of a system in classical mechanics \cite{goldstein}, the analysis of equilibrium states in statistical physics \cite{landau} or the transformation between the thermodynamical functions (entropy, enthalpy, Gibbs free energy, etc.) \cite{thermo}. This mathematical tool has the advantage, with respect to the Faddeev-Popov method, to be of immediate understanding, because it involves a very well known procedure that any student has seen at least one time during his studies. The possibility of illustrating the results of the Faddeev-Popov ``magic" method in terms of a Legendre transform would be of great didactical insight, since it will allow a better understanding of the physical foundations of a not so easy to understand common method in quantum field theory.
It is then the aim of this paper to unravel this connection, by using a simple example in which the Faddeev-Popov method is used to implement the Legendre transform of a given function $f(x)$. This is possible because these two methods, apparently very different and far away from each other, share in reality the same essence, namely they both consist in applying constraint to certain physical systems: the Faddeev-Popov method fixes the field gauge by imposing the gauge condition as a constraint to the field, and the Legendre transform implements the transformation between a function $f(x)$ and its transformed pair $g(p)$ by imposing a certain constraint to the function $F(x,p)=xp-f(x)$ \cite{convex}.
This work is organized as follows: in Sect. II we review very briefly the Faddeev-Popov method, in the simplest case of the electromagnetic field. This allows us to write Eq. \eqref{due}, that constitutes central point for our work. In Sect. III, a brief recall on the definition of Legendre transform is given. Finally, in Sect. IV we derive the Legendre transform formula by making use of the Faddeev-Popov rule. Conclusions are then drawn in Sect. V.
\section{The Faddeev-Popov method in a nutshell}
One of the possible ways to quantize a field is the so-called path integral quantization. In this scheme, the transition probability for a physical system to evolve from an initial configuration $\phi(x_i)$ to a final configuration $\phi(x_f)$ is obtained by summing all the probability amplitudes corresponding to all possible paths in space-time that the system will take to reach the final point $x_f$ starting from the initial point $x_i$ \cite{feynman}. Then, all the transition amplitudes (\emph{n}-point functions) can be obtained from a generating functional (the path integral) of the form
\begin{equation}\label{uno}
Z[J]=\int\mathcal{D}A_{\mu} e^{i\int d^4x\;\left(\mathcal{L}+J^{\mu}A_{\mu}\right)},
\end{equation}
where $\mathcal{L}\equiv\mathcal{L}[A_{\mu},\partial_{\nu}A_{\mu},x]$ is the Lagrangian density, $A_{\mu}$ represent the field variables and $J^{\mu}$ is an external current. For the free electromagnetic field the Lagrangian density is given by $\mathcal{L}=-(1/4)F_{\mu\nu}F^{\mu\nu}$, and since there are no field sources, $J^{\mu}$ has to be set to zero at the end of the calculation \cite{noteJ}.
Following Sect. 7.2 of Ref. \cite{ryder}, we can introduce heuristically the Faddeev-Popov method by making the following observations. First of all, the integral measure $\mathcal{D}A_{\mu}=\Pi_xdA_{\mu}(x)$ in Eq. \eqref{uno} accounts for every possible $A_{\mu}$, thus including those that are connected via a gauge transformation. We can make this fact explicit by introducing the following notation:
\begin{equation}
A_{\mu}\;\rightarrow \bar{A}_{\mu}^{\Lambda}.
\end{equation}
This means that now we are considering $A_{\mu}$ as a representative of the class of vector potentials that can be obtained from a given $\bar{A}_{\mu}$ by performing a gauge transformation of the type
\begin{equation}\label{gaugeA}
A_{\mu}=\bar{A}_{\mu}+\partial_{\mu}\Lambda(x).
\end{equation}
With this trick, we can rewrite the generating functional $Z[J]$ as follows, by separating the contributions of $\bar{A}_{\mu}$ and $\Lambda(x)$:
\begin{equation}\label{nonReg}
Z[J]=\int\mathcal{D}\bar{A}_{\mu}e^{i\int d^4x\;\left(\mathcal{L}+J^{\mu}\bar{A}_{\mu}\right)}\int\mathcal{D}\Lambda.
\end{equation}
The presence of the second integral $\int\mathcal{D}\Lambda$ is the one that causes $Z[J]$ to diverge, since there are infinitely many gauge fields $\Lambda(x)$ that satisfy Eq. \eqref{gaugeA}. The regularization of the integral \eqref{nonReg} is the scope of the Faddeev-Popov method.
In order to fully understand the essence of this method, let us consider the following resolution of the identity \cite{ryder}:
\begin{equation}\label{identity}
I=\mathrm{det}\Bigg(\frac{\delta G[A^{\Lambda}]}{\delta\Lambda}\Bigg)\Bigg|_{\Lambda=0}\int\mathcal{D}\Lambda\delta(G[A^{\Lambda}]),
\end{equation}
where the derivative in round brackets has to be intended in the sense of functional derivative \cite{brown}. The Dirac delta function inside the integral is nothing else but the gauge condition that one needs to impose to the Lagrangian of the free electromagnetic field \cite{comment} in order to fix the gauge, namely
\begin{equation}\label{gaugeCond}
G[A^{\Lambda}]=\partial^{\mu}A_{\mu}+\frac{1}{e}\partial^{\mu}\partial_{\mu}\Lambda(x),
\end{equation}
where $e$ is the electron charge. Eq. \eqref{identity} is simply a generalization to the domain of functional analysis of the well known expression \cite{matBook}
\begin{equation}
1=\left[\left(\frac{\partial g(x)}{\partial x}\right)\Bigg|_{x=x_0}\right]^{-1}\int\;dx\;\delta\left(g(x)\right),
\end{equation}
where $x_0$ is the point at which $g(x)$ vanishes. Inserting the expansion of the unity as given by Eq. \eqref{identity} into Eq. \eqref{uno} (and dropping the bar symbol from $\bar{A}_{\mu}$) gives:
\begin{equation}\label{due}
Z[J]=\int\mathcal{D}\Lambda\int\mathcal{D}A_{\mu}\Delta[A]\delta(G[A^{\Lambda}])e^{i\int\;d^4x\left(\mathcal{L}+J^{\mu}A_{\mu}\right)},
\end{equation}
where
\begin{equation}\label{deltaD}
\Delta[A]=\mathrm{det}\Bigg(\frac{\delta G[A^{\Lambda}]}{\delta\Lambda}\Bigg)\Bigg|_{\Lambda=0}.
\end{equation}
This result should be compared with Eq. \eqref{nonReg}. In this case, in fact, the generating functional $Z[J]$ is regularized, thanks to the presence of the Dirac delta function, whose role is to uniquely determine the vector potential $A_{\mu}$ by suitably fixing the gauge \eqref{gaugeCond} that is given as the argument of the Dirac delta function itself. This is the essence of the Faddeev-Popov method: by inserting in the generating functional \eqref{uno} the resolution of the identity as given by Eq. \eqref{identity}, one is now able to introduce a constraint (through a Dirac delta function) that automatically fixes the gauge, thus eliminating the arbitrariness of the vector potential $A_{\mu}$ and removing the problem of overcounting the field configurations.
Let us now put this result in a more appealing form, that will be helpful for the results of the next section. To this aim, let us rewrite Eq. \eqref{due} in the following way:
\begin{equation}\label{FPalt}
Z[J]=\int\mathcal{D}\Lambda\;\mathcal{F}[A_{\mu},\Lambda;J]\Delta[A_{\mu}]\delta(G[A^{\Lambda}]),
\end{equation}
where
\begin{equation}
\mathcal{F}[A_{\mu},\Lambda;J]=\int\mathcal{D}A_{\mu}e^{i\int\;d^4x\left(\mathcal{L}+J^{\mu}A_{\mu}\right)},
\end{equation}
and the dependence of $\mathcal{F}$ on $\Lambda$ is implicitly contained in $A_{\mu}$ through Eq. \eqref{gaugeA}. Although at a first glance Eq. \eqref{FPalt} may appear cumbersome and of little use, its form has a great advantage: it may be directly used to calculate the Legendre transform of an
arbitrary function $f(x)$, as we will discuss in the next section. Moreover, written in this form, the physical idea behind the Faddeev-Popov method appear more clear: given a functional $\mathcal{F}[A_{\mu},\Lambda;J]$ that depends on the particular choice of the gauge function $\Lambda(x)$, the Faddeev-Popov method consists in inserting the gauge fixing condition $G[A^{\Lambda}]$ in the form of an integral over the gauge configurations $\Lambda$. The presence of the term $\Delta[A_{\mu},0]$ simply accounts for the correct normalization term to insert in order not to modify the physics of the original system.
\section{The Legendre transform}
Before calculating the Legendre transform by means of the Faddeev-Popov method, let us first briefly recall the definition of Legendre transform and clarify its use with a simple example.
Following the standard textbook definition \cite{convex}, let us assume to have a function $f(x)$, and let us define, starting from $f(x)$, the two functions $F(x,p)=px-f(x)$ and $G(x,p)=\partial F(x,p)/\partial x$. It is then possible to define the Legendre transform $g(p)$ of $f(x)$ as:
\begin{equation}\label{leg1}
g(p)=F(x,p)|_{G(x,p)=0},
\end{equation}
where the constraint $G(x,p)=0$ is intended to be solved with respect to $x$. The reader should also keep in mind that, in order for the Legendre transform to make sense, the initial function $f(x)$ has to be convex, i.e., the function must have positive second derivative. It is also worth to be noted, that the definition of Legendre transform given here is only one of its possible form. Frequently, the Legendre transform is also defined as a constrained maximization of the function $f(x)$ \cite{convex}. For the purposes of this paper, however, this definition suites more the needings.
As an example of calculation of the Legendre transform, let us take the simple function $f(x)=x^2/2$. We then have that $F(x,p)=px-x^2/2$, $G(x,p)=p-x$, and the function $G(x,p)$ vanishes for $x_0=p$. Substituting this into Eq. \eqref{leg1} we have the following:
\begin{eqnarray}
g(p)&=&F(x,p)|_{G(x,p)=0}=\left(xp-\frac{x^2}{2}\right)\Bigg|_{x=x_0=p}\nonumber\\
&=&p^2-\frac{p^2}{2}=\frac{p^2}{2}.
\end{eqnarray}
\section{Implementing the Legendre Transform with the Faddeev-Popov Method}
We now want to show that another way of obtaining the Legendre transform is by applying the Faddeev-Popov trick to $F(x,p)$ as given by Eq. \eqref{FPalt}. To do this, we firstly rewrite Eq. \eqref{FPalt} for the case of ordinary functions (instead of functionals) as follows:
\begin{equation}\label{leg2}
\mathcal{I}=\int\;dx\;F(x,p)\Delta(p,x_0)\delta\Big(G(x,p)\Big),
\end{equation}
where $x_0$ is the point in which $G(x,p)=0$ with respect to $x$ and $\Delta(p,x_0)$ is a function to be yet determined. For the purpose of this note, we will assume that the function $G(x,p)$ vanishes only for a single given $x_0$ in the considered interval. We can expand the Dirac delta function in the integral with the usual expansion formula \cite{matBook}
\begin{equation}\label{dirac_fx}
\delta(w(x))=\frac{\delta(x-x_0)}{|w^{'} (x_0)|},
\end{equation}
where $x_0$ is the zero of the function $w(x)$ and $w'(x_{0})$ is the first derivative of $w(x)$ evaluated in $x=x_{0}$. Substituting this expression in eq. \eqref{leg2} brings to
\begin{equation}\label{FPx}
\mathcal{I}=\int\;dx\;F(x,p)\Delta(p,x_0)\frac{\delta(x-x_0)}{|G'(x_0,p)|},
\end{equation}
and it appear clear that we can choose $\Delta(x_0,p)$ in such a way that it compensates the extra term in the the denominator, namely
\begin{equation}
\Delta(p,x_0)=\left|\frac{\partial G(x,p)}{\partial x}\right|_{x=x_0}=|G'(x_0,p)|.
\end{equation}
Note that this definition is fully equivalent and consistent with the one given by Eq. \eqref{deltaD}. This brings to the following result:
\begin{equation}
\mathcal{I}=\int\;dx\;F(x,p)\delta(x-x_0)=F(x_0,p).
\end{equation}
Now, since $x_0$ is the value at which the function $G(x,p)=0$, the last term in the equality can be rewritten as $F(x,p)|_{G(x,p)=0}$, thus giving
\begin{eqnarray}\label{legendre}
\mathcal{I}&=&\int\;dx\;F(x,p)\Delta(x_0,p)\delta\Big(G(x,p)\Big)dx\nonumber\\
&=&F(x_0,p)=F(x,p)|_{G(x,p)=0}\equiv g(p)
\end{eqnarray}
This is exactly the Legendre transform $g(p)$ of the initial function $f(x)$. This result is worth a bit of discussion. To start with, let us consider again Eq. \eqref{leg1}. There, the Legendre transform is implemented by applying the constraint $G(x,p)=0$ to the function $F(x,p)$. In that case, the constraint is applied ``ad hoc" in order to produce the function $g(p)$. In Eq. \eqref{FPx}, instead, the same constraint $G(x,p)=0$ is applied to the function $F(x,p)$ in a more natural way by means of the Faddeev-Popov trick, by making the constraint the argument of the Dirac delta function. The Legendre transformation is now implemented by selecting (among all the possible values of $x$) only those values $x_0$ that make the argument of the Dirac delta in Eq. \eqref{FPx} vanish. The constraint, in this case, is therefore automatically applied, and the result is the same as in the case of the standard definition of Legendre transform as given by Eq. \eqref{leg1}.
We now repeat the example of Sect. III by using the Faddeev-popov method given by Eq. \eqref{leg2} to calculate the Legendre transform of the function $f(x)=x^2/2$. As before, $F(x,p)=xp-x^2/2$ and $G(x,p)=p-x$, that vanish for $x_0=p$. Therefore we have that $\Delta(p,x_0)=|G'(p,p)|=1$ and
\begin{equation}
\delta\Big(G(x,p)\Big)=\delta(x-p).
\end{equation}
Substituting these results into Eq. \eqref{leg2} then gives:
\begin{equation}
\mathcal{I}=\int\;dx\;F(x,p)\delta(x-p)=F(p,p)=\frac{p^2}{2}\equiv g(p).
\end{equation}
This trivial, but instructive, example proves that our definition of Legendre transform via the integral in Eq. \eqref{legendre} is a valid and fully compatible definition of Legendre transform. Moreover, this allows us to give a more intuitive and clear explanation on how to apply the Faddeev-Popov method and what is its physical meaning, by providing an example of application in a more familiar context for the students, than the framework of quantum field theory where this method finds its natural application.
As a second, more physical example, let us consider the problem of calculating the Hamiltonian of an harmonic oscillator given its Lagrangian
\begin{equation}
\mathcal{L}_{ho}(q,\dot{q})=\frac{1}{2}m\dot{q}^2-\frac{1}{2}m\omega^2q^2,
\end{equation}
where $\dot{q}=dq/dt$, $m$ is the oscillator mass and $\omega$ is the characteristic oscillator frequency. As it is well known, the Hamiltonian is calculated by firstly introducting the canonically conjugate momentum $p=\partial\mathcal{L}_{ho}/\partial\dot{q}=m\dot{q}$ and then by performing a Legendre transform \cite{goldstein}:
\begin{equation}\label{hoLeg}
\mathcal{H}_{ho}(q,p)=p\dot{q}-\mathcal{L}_{ho}(q,\dot{q})=\frac{p^2}{2m}+\frac{1}{2}m\omega^2q^2.
\end{equation}
We now want to reproduce the same result by applying the Faddeev-Popov method. First of all, it is worth noticing that in this case the Legendre transform is made with respect to the variable $\dot{q}$, and that $q$ acts simply as a parameter in Eq. \eqref{hoLeg}. So said, we can build the analogue of the function $F(x,p)$ as
\begin{equation}
F(\dot{q},p;q)=p\dot{q}-\mathcal{L}_{ho}(q,\dot{q}).
\end{equation}
Notice that we used the semi-colon in the argument of $F(\dot{q},p;q)$ to indicate that here $q$ plays the role of a parameter, as the application of the Faddeev-Popov method leaves $q$ unchanged. The constraint $G(x,p)$ is now given by
\begin{equation}
G(\dot{q},p;q)=\frac{\partial F(\dot{q},p;q)}{\partial\dot{q}}=p-\frac{\partial\mathcal{L}_{ho}(q,\dot{q})}{\partial\dot{q}}=p-m\dot{q},
\end{equation}
and therefore $\Delta(\dot{q}_0,p;q)=|G'(\dot{q}_0,p;q)|=m$, where $\dot{q}_0=p/m$ is the point at which $G(\dot{q},p;q)$ vanishes with respect to $\dot{q}$. Notice that $G(\dot{q},p;q)=0$ corresponds to the usual definition of the canonical momentum in terms of derivative of the Lagrangian with respect to $\dot{q}$. This is the actual constraint that implements the Legendre transform.
In this case, Eq. \eqref{leg2} becomes
\begin{eqnarray}
\mathcal{I}&=&\int\;d\dot{q}\Big[p\dot{q}-\mathcal{L}_{ho}(q,\dot{q})\Big]\;m\;\delta\Big(p-m\dot{q}\Big)\nonumber\\
&=&\int\;d\dot{q}\Big[p\dot{q}-\mathcal{L}_{ho}(q,\dot{q})\Big]\;\delta\Big(\dot{q}-\frac{p}{m}\Big)\nonumber\\
&=&p\left(\frac{p}{m}\right)-\mathcal{L}_{ho}\left(q,\frac{p}{m}\right)\nonumber\\
&=&\frac{p^2}{2m}+\frac{1}{2}m\omega^2q^2\equiv\mathcal{H}_{ho}.
\end{eqnarray}
\section{Conclusions}
In conclusion, we have discussed how the Faddeev-Popov method can be used to easily implement the Legendre transform of an arbitrary function $f(x)$. As an example, we considered the simple case of $f(x)=x^2/2$ as a direct verification of the validity of our correspondence. We have also presented, as a more physically meaningful example, how to use the Faddeev-Popov method to calculate the Hamiltonian of a classical harmonic oscillator starting from the knowledge of its Lagrangian function. This provides a novel point of view on the method itself, that only makes use of concepts familiar to the students such as the Legendre transform, without invoking any complicated theoretical framework as quantum field theory or gauge theory.
|
\section{Planar slab waveguide.}
To gain insight for the design of optical lattices with dielectric structures we consider the simplest structure that supports the existence of `Guided Modes' (GMs) for both polarizations of light: a dielectric slab of width $W$ and refractive index $n(=\sqrt{\varepsilon})$. For illustration we focus on the symmetric transverse electric (TE-like) GM, whose electric profile inside and outside the dielectric is given by;
\begin{align}
\label{eqTE}
& \mathbf{E}_{\mathrm{GM},\mathrm{in}}(\mathbf{r})=E_{\mathrm{in}} \cos(k_z z) e^{i k_\| x}\mathbf{y} \,, \nonumber \\
& \mathbf{E}_{\mathrm{GM},\mathrm{out}}(\mathbf{r})=E_{\mathrm{out}} e^{-\beta z} e^{i k_\| x} \mathbf{y}\,,
\end{align}
\noindent where $\beta$ and $k_z$ are related by the following trascendental equations \cite{liao91a}:
\begin{align}
\label{eqTEmodes}
& \beta = k_z\tan( k_z W/2)\,,\nonumber \\
& k_z^2+\beta^2=k_0^2(n^2-1)\, ,
\end{align}
\noindent where $k_0=\omega/c=2\pi/\lambda_0$. The in-plane momentum, $k_\|$, determines the effective wavelength of the GM and can be obtained from the solution ($\beta,k_z$) of the previous equations as $k_\|^2=k_0^2 n^2-k_z^2$. Contrary to conventional optical lattices $k_\|$ depends not only on its wavelength ($\lambda_0$) but also on the slab properties $(n,W)$. By sending counter-propagating GMs in both $x,y$ directions a standing-wave is generated as for conventional optical lattices with intensity periodicity $d_{\mathrm{slab}}=\pi/k_\|$. From Eqs.~\ref{eqTE}-\ref{eqTEmodes}, it can be shown that the upper bound of $k_\|$ yields to a minimum periodicity of: $d_{\mathrm{min},\mathrm{slab}}= \frac{\lambda_0}{2 n}$. This distance is a factor $n$ smaller than the one that can be achieved by interfering lasers in free-space. The electromagnetic profile of this standing wave has associated an optical potential with a periodicity in the $x,y$ plane, $V_{\mathrm{GM}}(\mathbf{r}+\
\mathbf{R})=V_{\mathrm{GM}}(\mathbf{r})$, where $\mathbf{R}$ is any vector of the Bravais lattice of the potential. However, in the $z$ direction the potential decays exponentially, $e^{-2\beta z}$, due to the evanescent character of the GMs outside the dielectric.
To stabilize the trap in the vertical $z$ direction normal to the slab, Side Illumination (SI) on both sides of the structure can be used. By sending two lasers with the same frequency but whose amplitudes are related by a phase factor $e^{i\phi}$, an interference pattern is generated, whose minima position can be controlled by chirping the phase $\phi$ between them. It is convenient to work in the condition where the slab is transparent to incident light, i.e., when $\lambda_{\mathrm{SI}}=2 n W$. The far off-resonance trapping (FORT) potential associated to SI can be approximated by:
\begin{equation}
V_{\mathrm{SI}}(\mathbf{r})=V_{\mathrm{SI}}(z)\approx-\frac{\Omega_{\mathrm{SI}}^2}{\delta_{\mathrm{SI}}}\sin^2\Big[k_{\mathrm{SI}} \big(z-z_t(\phi)\big)\Big] \,,
\end{equation}
where $\Omega_{\mathrm{SI}}=\vec{\mu}\cdot \mathbf{E}_{\mathrm{SI}} /\hbar$ is the Rabi frequency, $\vec{\mu}$ the atomic dipole moment, $ \mathbf{E}_{\mathrm{SI}}$ the electric field associated to SI, and $\delta_{\mathrm{SI}}$ the detuning between the SI and atomic frequencies. The position of the trapping minimum, $z_t$, is controlled through the phase, $\phi$, between up/down lasers.
The trapping minimum, $z_t$, must be close to the surface to interact with $V_{\mathrm{GM}}$, therefore, Casimir-Polder (CP) potentials must be taken into account \cite{buhmann07a}. In order to estimate their effect over the optically induced trap and under which conditions we can neglect them, we consider the CP potential above a semi-infinite dielectric interface, given by \cite{buhmann07a}:
\begin{align}
\label{vwdpot1}
V_{\mathrm{CP},\mathrm{plane}}(z)&=-\frac{1}{16}\frac{n^2-1}{
n^2+1}\frac{\Gamma_{\mathrm{a}}}{(k_{\mathrm{a}}z)^3}\,,
\end{align}
where $k_\mathrm{a}$ and $\Gamma_{\mathrm{a}}$ are the atomic momemtum and vacuum decay rate. The stability condition for the trap in the $z$-direction implies $V''_{\mathrm{SI}}(z_t)+V''_{\mathrm{CP},\mathrm{plane}}(z_t)>0$, which can always be satisfied by tuning the intensity of the laser such that
\begin{equation}
\label{intlas}
\Omega_{\mathrm{SI}}^2>\frac{3}{2}\frac{\delta_{\mathrm{SI}}V_{\mathrm{CP},\mathrm{plane}}(z_t) }{(k_{\mathrm{SI}} z_t)^2} =\frac{3}{32}\frac{n^2-1}{
n^2+1}\frac{\delta_{\mathrm{SI}}\Gamma_{\mathrm{a}}}{(k_{\mathrm{a}} z_t)^3(k_{\mathrm{SI}} z_t)^2}\,.
\end{equation}
Thus, combining $V_{\mathrm{SI}}(z)+V_{\mathrm{GM}}(\mathbf{r})$, a predominantly FORT optical potential is generated along the $z$ direction with subwavelength lattice constant, $d_{\mathrm{slab}}\gtrsim \frac{\lambda_0}{2 n}$ in the $x,y$ plane. We recognize that the estimate of intensity of Eq. \ref{intlas} using Eq. \ref{vwdpot1} for the CP potential is only a first approximation to the problem, and a full numerical calculation of CP potential has to be considered for more accurate estimations.
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.99\linewidth]{fig1supv4v3.pdf}
\caption{(a) General scheme of the system: dielectric slab of thickness $W$ and refractive index $n$ with periodic lattice of holes of radius $R$ and periodicity $d$. Panels (b-c): Proof of principle example for GaP slab with $d=125$ nm, $R=0.4d$ and $W=2d$. We use an incoherent superposition of two orthogonal TM GM modes for horizontal trapping in the $x,y$ plane with $\lambda_{\mathrm{TM}}=770$ nm and $\Omega_{\mathrm{GM}}/2\pi= 50$ GHz. For vertical trapping along $z$ we use SI FORT with $\lambda_{\mathrm{SI}}=760$ nm and $\Omega_{\mathrm{SI}}/2\pi= 50$ GHz. (b) [and (c)] Vertical [and horizontal] cut of the total potential (solid black) and its different contributions: CP (dashed red), SI (dotted blue) and TM (dashed-dotted purple) at $x=y=0$ [and $x=0$, $z=z_t=65$ nm], respectively.}
\label{fig1SM}
\end{figure*}
\section{Photonic Crystal Waveguides (PCWs) for subwavelength trapping.}
Let us now consider the situation where there is a periodic modulation of the refractive index of the structure, $n(\mathbf{r})$, e.g., by assuming that the dielectric slab has a periodic square lattice of circular holes, with radius $R$ and periodicity $d$ [see Fig.~\ref{fig1SM}(a)]-- the so-called Photonic Crystals Waveguide (PCW) slab. The presence of the holes decreases intensity requirements, as the CP potential above the holes is corrected with respect to $V_{\mathrm{plane},\mathrm{CP}}(z)$ \cite{eberlein11a} in the electrostatic regime by a factor, $f(z,R)$:
\begin{align}
\label{vwdpot2}
V_{\mathrm{hole},\mathrm{CP}}(\mathbf{r})&=V_{\mathrm{plane},\mathrm{CP}}(z)\times f(z,R)=V_{\mathrm{plane},\mathrm{CP}}(z)\times \Big[\frac{1}{2}+\frac{1}{\pi}\arctan\big(\frac{
z^2-R^2}{2 R z}\big) + \frac{2 z R (z^2-R^2)}{\pi(R^2+z^2)^2}\Big]\,,
\end{align}
which is always smaller than $1$ above the hole. PCW slabs also support GMs whose properties depend on geometrical/material parameters. The wavelength of the GM is $\lambda_{\mathrm{GM}}=\lambda_0/n'$, where $n'$ is the effective refractive index of the medium that takes into account geometrical effects. Typically $n' \lesssim n$, thus, the periodic lattice generated by two counter-propagating GMs [see Fig.~\ref{fig1SM}(a)] show periodicities lower bounded approximately by: $d_{\mathrm{min},\mathrm{GM}}\approx \frac{\lambda_0}{2 n}$. SI is used to stabilize the trap in the $z$-direction. At large enough distances $z_t$, the periodic modulation of the CP and SI potential in the $x,y$ plane can be neglected.
In Fig.~\ref{fig1SM}, we show a proof of principle example for a GaP slab with $d=125$ nm, $R=0.4d$ and $W=2d$ which corresponds to the line (i) in Fig. 1(b) of the main manuscript. Curves for the various trapping potentials are calculated numerically without simplifying assumptions as in Eq. (\ref{vwdpot2}). We use an incoherent superposition of two orthogonal TM GM modes for a horizontal trap with $\lambda_{\mathrm{TM}}=770$ nm and $\Omega_{\mathrm{GM}}/2\pi= 50$ GHz. For vertical trapping we use SI FORT with $\lambda_{\mathrm{SI}}=760$ nm and $\Omega_{\mathrm{SI}}/2\pi=50$ GHz. Line cuts of the total trap potential $V_{\mathrm{tot}}$ (solid black) are given in Fig.~\ref{fig1SM}(b,c), with the contributions from the TM FORT (dot-dashed purple), SI FORT (dotted blue) and CP (dashed red) potentials shown. The trap depth $V_d$ and frequencies $\omega_{\mathrm{t}}$ for the particular example in Fig.~\ref{fig1SM} are $\{V_{d,\mathrm{xy}},V_{d,\mathrm{z}}\}/2\pi \approx \{2.1,624\}$~MHz and $\{\omega_{t,\mathrm{xy}},
\omega_{t,\mathrm{z}}\}/2\pi\
\approx \{0.52,18.7\}$~MHz.
\subsection{Contrast loss of GM for deeply subwavelength scales.}
The scheme of the two previous sections use GMs for $x,y$ confinement and SI to trap in the vertical direction. In principle, it is possible to extend this method for deeper subwavelength scales, $d\ll \lambda_0/(2 n)$; however, several complications arise in this limit. For example, as the lattice constant $d$ decreases, the GM with energies around the atomic transition are closer to the light line. Consequently, the contrast of the GM intensity in the $x,y$ plane for a given trap distance $z_t$, that can be characterized through the function $C(z_t)=(\mathrm{max}\{|E_{\mathrm{GM}}(x,y,z_t)|^2\}-\mathrm{min}\{|E_{\mathrm{GM}}(x,y,z_t)|^2\})/(\mathrm{max}\{|E_{\mathrm{GM}}(x,y,z_t)|^2\}+\mathrm{min}\{|E_{\mathrm{GM}}(x,y,z_t)|^2\})$, also decreases. This contrast loss results in the necessity of larger laser intensities to guarantee the trapping condition, $2\omega_t\lesssim V_d$. The FORT trap depth and frequency scale as $V_d =C(z)|\Omega_{\mathrm{GM}}|^2/|\delta|$ and $\omega_t =\sqrt{h V_d /(2 m d^2)}$,
respectively, where $\Omega_{\mathrm{GM}}$ is the Rabi frequency associated to the GMs. Theoretically, the trapping condition can always be satisfied by compensating the contrast loss with the Rabi frequency strength. Concretely, if we denote $\Omega_{\mathrm{conf}}$ to be the value that guarantees the equality [$2\omega_t= V_d$], $\Omega_{\mathrm{conf}}$ is found to be
\begin{align}
\label{key}
\Omega_{\mathrm{conf}}\simeq\sqrt{\frac{2 h c |\Delta\lambda^{-1}|}{m d^2 C}}\,,
\end{align}
%
where $\Delta\lambda^{-1}=1/\lambda_{\mathrm{a}}-1/\lambda_{\mathrm{GM}}$.
In Fig.~\ref{fig3sup}, we show an example of these scalings for a GaP slab with $W=2d$, $R=0.4 d$, considering a fixed GM wavelength of $\lambda_{\mathrm{GM}}=760$ nm. From the study of the horizontal confinement provided by the GMs, we see that: i) The smaller the lattice parameter $d$, the smaller is the contrast of the GM modes, implying a considerable increase of the intensities required to reach the trapping condition in the $x,y$ plane (see Fig.~\ref{fig3sup}(a)); ii) For a fixed lattice constant parameter $d$, it is possible to decrease the intensity requirements by choosing a smaller vertical trapping position, $z_t$, as shown in panels (a-b) of Fig.~\ref{fig3sup}. However, this strategy requires increasing SI intensities to compensate the CP potential (which is not included in Eq. \ref{key} nor in Fig. \ref{fig3sup}). Summing up, obtaining deep subwavelength scales using GM (and SI) for horizontal (vertical) trapping can be done at the expense of significant increases in the laser intensities with
the
concomitant increase in the scattering and heating rates.
\begin{figure*}[!hb]
\centering
\includegraphics[width=0.7\linewidth]{fig3supv4v3.pdf}
\caption{(a) $\Omega_{\mathrm{conf}}$ for a GaP PCW with $W=2d$ and $R=0.4d$, as defined in Eq. \ref{key}, as a function of lattice parameter $d$ for different vertical distances $z_t$ from the surface, as depicted in the legend. (b) $\Omega_{\mathrm{conf}}$ for a GaP PCW with $W=2d$ and $R=0.4d$, as defined in Eq. \ref{key}, as a function of vertical distance from the surface, $z_t$, for different lattice parameters $d$ as depicted in the legend. }
\label{fig3sup}
\end{figure*}
\section{Scalings in ``vacuum lattices''.}
In the main manuscript we show how to engineer a new class of optical lattices taking advantage of the CP modulation in the XY-direction. In these ``vacuum lattices'' the XY confinement is mainly provided by CP, whereas the vertical confinement comes from SI. In Fig.~\ref{fig4sup} (a-c) we show how the CP potential varies in the vertical direction for different places of the unit cell and different post radii. Here, the unit cell is centered at the position of the vertical axis of the dielectric post. $x=y=0$. As expected the CP potential is larger in absolute value in the regions close to the dielectric, $y=0$, than along the border of the unit cell $y=d/2$. Moreover, the case with the biggest post radius, $R=10$ nm, also shows deeper potentials in the XY plane. The dependence of the CP for the XY confinement as a function of the trapping distance $z_t$ is summarized in panel (d) of Fig.~\ref{fig4sup}. The advantage of the CP potential compared to $V_{\mathrm{GM}}$ is that it does not suffer from the
contrast loss for smaller
distances $\{d,z_t \}
\ll \lambda_0$.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.99\linewidth]{fig2supv4v4.pdf}
\caption{Panel (a,b,c): Vertical cut of the CP potential $V_{\mathrm{CP}}(0,y,z)$ for three different horizontal positions in the unit cell: $y=0$ (solid black), $y=d/4$ (dashed red) and $y=d/2$ (dotted blue) for a structure with $d=50$~nm and $R=10$~nm (a), $R=7.5$~nm (b) and $R=5$~nm (c). Panel (d): Horizontal cut of the CP potential for a periodic lattice of GaP posts of $R=10$~nm and $d=50$ nm for different vertical distances, $z_t$, ranging from $45$ to $30$~nm. }
\label{fig4sup}
\vspace{-4mm}
\end{figure}
\section{Atom-atom interactions $\Gamma_{ij}$ mediated by Guided modes.}
When atoms are trapped close to the PCW structures, they interact with the GMs in the structure. This interaction can be described by the following hamiltonian~\cite{john90a,john94a,john95a}:
\begin{equation}
\label{hint1}
H_{int}=\sum_{i}\sum_{\mathbf{k}}\big( g_\mathbf{k} e^{i\mathbf{k}\cdot
\mathbf{r}_i}a_{\mathbf{k}}\ud{\sigma_i}+\mathrm{h.c.}\big)\,,
\end{equation}
where we neglected the multi-band effects (as we work with $\Delta_{\mathrm{a}c}=\omega_{\mathrm{a}}-\omega_c$ much smaller than the bandgap width) and focused on a single polarization. The coupling constant can then be written as:
\begin{equation}
\label{coup}
g_{\mathbf{k}}=\sqrt{\eta\frac{\omega(\mathbf{k})}{2 \varepsilon_0 \hbar L^2}} \vec{\mu}\cdot \mathbf{u}_{\mathbf{k}}\,,
\end{equation}
where $\vec{\mu}$ is the atomic dipole moment, $L^2$ is the quantization area, $\omega(\mathbf{k})$ [$\mathbf{u}_{\mathbf{k}}$] is the energy dispersion [spatial mode dependence] of the field $a_{\mathbf{k}}$, and $\eta$ a factor that takes into account polarization effects for the particular atomic level structure addressed. The hamiltonian of Eq.~\ref{hint1} induces an effective interaction between the atoms. If the coupling between atoms and dielectric modes can be treated under the Born-Markov approximation, it is then possible to obtain a master equation that describes effectively the dynamics of the atoms by tracing out the photonic degrees of freedom. The effective equation is then given by \cite{breuerbook02a}:
\begin{equation}
\label{efmaster2}
\frac{d\rho}{dt}=\sum_{i,j} \Gamma_{ij}\big( \sigma_i\rho\ud{\sigma_j}
-\ud{\sigma_j}\sigma_i\rho\big)+\mathrm{h.c} \,,
\end{equation}
\noindent where $\Gamma_{ij}$ is the collective coupling defined as follows:
\begin{align}
\label{coupling}
\Gamma_{ij}= L^2\lim_{s\rightarrow 0} \int_{\mathrm{BZ}}\frac{d^2 \mathbf{k}}{(2\pi)^2}\frac{|g_{\mathbf{k}}|^2}{s+i(\omega_{\mathrm{a}}-\omega(\mathbf{k}))}e^{i \mathbf{k}\cdot \mathbf{r}_{ij} }\,,
\end{align}
where $L^2$ is the quantization area and the $\mathbf{k}$-integration is over the entire Brillouin zone.
We can separate in Eq.~\ref{efmaster2} the contributions from the real and imaginary parts of $\Gamma_{ij}$. Denoting: $\Gamma_{ij}=\gamma_{ij}/2+i J_{ij}$, where both $\gamma_{ij}$ and $J_{ij}$ are real numbers, Eq.~\ref{efmaster2} can be rewritten as follows:
\begin{equation}
\label{efmaster22}
\frac{d\rho}{dt}=\sum_{i,j} \frac{\gamma_{ij}}{2}\big( 2\sigma_i\rho\ud{\sigma_j}
-\ud{\sigma_j}\sigma_i\rho-\rho\ud{\sigma_j}\sigma_i\big)+i\sum_{ij} J_{ij}[\rho,\ud{\sigma_j}\sigma_i]\,,
\end{equation}
which is separated into the coherent (incoherent) contribution coming from $J_{ij}$ ($\gamma_{ij}$), respectively.
For the structure considered in Figs.~3-4 in Ref.~\cite{gonzaleztudelaMain}, the main contributions of the integration over the Brillouin zone in Eq. \ref{coupling} are given by the regions within the four semi-circles depicted in Fig.~\ref{fig6sup}(a) for the four $X$ points of the Brillouin zone. Separating the different contributions, we arrive at:
\begin{align}
\label{coupling3}
\Gamma_{ij}=\Big[\Gamma_{ij,x} \cos(\mathbf{k}_{c,x}\cdot \mathbf{r}_{ij})+\Gamma_{ij,y} \cos(\mathbf{k}_{c,y}\cdot \mathbf{r}_{ij}) \Big] \,,
\end{align}
where $\Gamma_{ij,x (y)}$:
\begin{align}
\label{coupling4}
\Gamma_{ij,x (y)}=L^2\lim_{s\rightarrow 0} \int_{\mathrm{BZ}}\frac{d^2 \mathbf{k}}{(2\pi)^2}\frac{|g_{\mathbf{k}_{c,{x,y}}-\mathbf{k}}|^2}{s+i(\omega_{\mathrm{a}}-\omega(\mathbf{k}_{c,x(y)}-\mathbf{k}))}e^{i \mathbf{k}\cdot \mathbf{r}_{ij} } \,.
\end{align}
The atoms are placed at positions $\mathrm{r}_{ij}=(n,m)d$, with $n,m\in \mathbb{Z}$ such that the cosine terms only give phases: $(-1)^n,(-1)^m$. These phases could also be compensated by an appropriate configuration of the in-plane momenta of the driving lasers in our two-photon Raman coupling scheme (see Fig.~\ref{fig6sup}(d)). As we are interested in calculating the magnitude of $\Gamma_{ij}$, for simplicity, we drop these phases and also assume that the coupling is symmetric for the X and Y directions ($\Gamma_{ij,x}=\Gamma_{ij,y}$). Then, the absolute value of coupling $\Gamma_{ij}$ can be estimated by calculating twice the integral of the circumference area around the X point:
\begin{align}
\label{coupling5}
\Gamma_{ij}=2 L^2\lim_{s\rightarrow 0} \int_{BZ}\frac{d^2 \mathbf{k}}{(2\pi)^2}\frac{|g_{\mathbf{k}_{c}-\mathbf{k}}|^2}{s+i(\omega_{\mathrm{a}}-\omega(\mathbf{k}_{c}-\mathbf{k}))} e^{i \mathbf{k}\cdot \mathbf{r}_{ij} } \,.
\end{align}
\begin{figure}[!tb]
\centering
\includegraphics[width=0.89\linewidth]{fig6supv4v4.pdf}
\caption{(a) Brillouin zone highlighting in blue the four most relevant regions contributing to $\Gamma_{ij}$. (b) Inverse of effective mode length, $d/L_m(\mathbf{k})$, and (c) energy dispersion for the square hole lattice structure of Figs.~ 3-4 of Ref. \cite{gonzaleztudelaMain}. (d) Parabolic approximation of the band structure close to a band edge. (e) Simplified scheme in order to couple to a single polarization of light. (f) Coherent coupling, $J_{ij}/\Gamma_{\mathrm{a}}$, for $r_{ij}=d$ using the exact numerical integration and the analytical expression obtained using an isotropic approximation and an ``averaged'' $A_{\mathrm{fit}}\sim 1.8\times 10^{12} $ $\mu$m$^2/s$. The ``averaged'' value of $A$ is chosen to fit the numerical integration data (in solid black).}
\label{fig6sup}
\end{figure}
\subsection{Using ``effective mass'' for $\omega(\mathbf{k})$ and isotropic approximation.}
Up to this point we have not made assumptions about $g_{\mathbf{k}}$ or $\omega(\mathbf{k})$. In order to obtain analytical expressions that give us some intuition on the magnitude and scaling of $\Gamma_{ij}$, we use the effective mass approximation: $\omega(\mathbf{k})\approx \omega_{c}+A (\mathbf{k}_c-\mathbf{k})^2$ [see Fig.~3 in Ref.~\cite{gonzaleztudelaMain}], and assume that the coupling is isotropic around the X-points. Using these assumptions, we integrate the angular dependence in the integral of Eq.~\ref{coupling5} as follows:
\begin{align}
\label{coupling9}
\Gamma^{\mathrm{iso}}_{ij}= \lim_{s\rightarrow 0} \frac{L^2 |g_{\mathbf{k}_{c}}|^2}{\pi}\int_0^{k_c}dk \frac{k J_0(k |r_{ij}| )}{s+i(\Delta_{\mathrm{a}c}-A k^2)}\,,
\end{align}
where we have taken $|g_{\mathbf{k}_{c}-\mathbf{k}}|^2\rightarrow |g_{\mathbf{k}_{c}}|^2$ out of the integral for a simple model. For $A>0$ there are two different situations to consider, namely, $\Delta_{\mathrm{a}c}\lessgtr 0$. When the atomic frequency lies in the bandgap, $\Delta_{\mathrm{a}c}<0$, the limit has no singularity and the $\Gamma^{\mathrm{iso}}_{ij}=-i J^{\mathrm{iso}}_{ij}$ is purely imaginary, with:
\begin{align}
\label{coupling9}
J^{\mathrm{iso}}_{ij}=\frac{L^2 |g_{\mathbf{k}_{c}}|^2}{\pi}\int_0^{k_c}dk \frac{k J_0(k |r_{ij}| )}{\Delta_{\mathrm{a}c}-A k^2}\,.
\end{align}
Defining $\xi=\sqrt{|A/\Delta_{\mathrm{a}c}|}$ and performing the change of variables $q=k\xi$, we arrive to:
\begin{align}
\label{coupling99}
J^{\mathrm{iso}}_{ij}= \frac{|g_{\mathbf{k}_{c}}|^2 L^2}{\pi A}\int_0^{k_c\xi}dq q \frac{1}{1+q^2} J_0(q |r_{ij}|/\xi )\simeq \frac{ |g_{\mathbf{k}_{c}}|^2 L^2}{\pi A}K_0(|r_{ij}|/\xi)\equiv\Gamma_{2d} K_0(|r_{ij}|/\xi) \,,
\end{align}
where we have used $k_c\xi\gg 1$. $K_0(x)$ is a modified Bessel function. Defining the free-space decay rate as $\Gamma_{\mathrm{a}}=\frac{|\mu|^2 k_{\mathrm{a}}^3}{3\pi \varepsilon_0 \hbar}$, we write the coupling $|g_{\mathbf{k}}|^2$ as follows:
\begin{equation}
\label{couplomega}
|g_{\mathbf{k}}|^2=\Gamma_{\mathrm{a}}\frac{c \sigma \omega(\mathbf{k})}{8 L^2 L_m(\mathrm{k},\mathbf{r}_{\mathrm{a}}) \omega_{\mathrm{a}}} \,,
\end{equation}
where we have defined $\sigma=\frac{3 \eta}{2 \pi} \lambda^2_{\mathrm{a}}$ as the effective cross-section. Substituting the $|g_{\mathbf{k}_{c}}|^2$ into Eq.~\ref{coupling99}, we arrive to:
\begin{align}
\label{coupling10}
\Gamma_{2d}= \Gamma_\mathrm{a}\frac{c \sigma }{4\pi A L_m (\mathbf{k}_c,\mathbf{r}_\mathrm{a})}\,,
\end{align}
where we have introduced the effective mode length, $L_{m}(\mathbf{k},\mathbf{r}_\mathrm{a})$ defined as:
\begin{equation}
\label{effectivelength}
L_m (\mathbf{k},\mathbf{r}_\mathrm{a})=\frac{\int d^3\mathbf{r}\epsilon(\mathbf{r}) |\bar{E}_{\mathbf{k},m}(\mathbf{r})|^2}{d^2\epsilon(\mathbf{r}_{\mathrm{a}}) |\bar{E}_{\mathbf{k},m} (\mathbf{r}_{\mathrm{a}})|^2}\,,
\end{equation}
%
that takes into account the geometrical distribution of the field density, $|\bar{E}_{\mathbf{k},m} (\mathbf{r}_{\mathrm{a}})|^2$, at the atomic position $\mathbf{r}_{\mathrm{a}}$.
In the opposite limit, i.e., the dissipative regime, when $\Delta_{\mathrm{a}c}>0$ the limit $s\rightarrow 0$ has a singularity, that can be calculated using:
\begin{equation}
\label{limit}
\lim_{s\rightarrow 0} \frac{1}{s+i\Delta_k }=\pi \delta(\Delta_k)-i
\mathcal{P}\frac{1}{\Delta_k}\,.
\end{equation}
Therefore in this case $\Gamma_{ij}$ has both real and imaginary components and is given as follows:
\begin{align}
\label{couplingdelta}
\Gamma^{\mathrm{iso}}_{ij}\approx \Gamma_{2d}\frac{\pi}{2} H^{(1)}_0 \big[|r_{ij}|/\xi\big]\,,
\end{align}
where $H^{(1)}_0(x)=J_0(x)+iY_0(x)$ is the Hankel function of the first kind.
\subsection{Collective coherent coupling $J_{ij}$ using exact $\omega(\mathbf{k})$ and $g_{\mathbf{k}}$.}
The effective mass and isotropic approximation are convenient to obtain analytical formulas to understand scaling with parameters such as $A$, $\Delta_{\mathrm{a}c}$ or $r_{ij}$. For the structure of Figs. (3-4) of Ref.~\cite{gonzaleztudelaMain}, both the exact dispersion relation $\omega(\mathbf{k})$, and its effective length at the center of the hole, $L_m(\mathbf{k})$, are anisotropic [see Panels (b-c) of Fig.~\ref{fig6sup}]. For example, the curvature (effective length) along the $X-M$ direction is flatter (steeper) than in the $X-\Gamma$ direction. Therefore, it is not straightforward to determine an appropiate value of $A$ to input into the isotropic formula of Eq.~\ref{coupling10}. Moreover, we should also take into account the details of the implementation of the simplified $\Lambda$-scheme depicted in Fig.~\ref{fig6sup}(e).
To estimate up to which point the isotropic formula is giving the right scalings and order of magnitude for $\Gamma_{ij}$, we numerically integrate Eq.~\ref{coupling5} with the exact $\omega(\mathbf{k})$ and $g_{\mathbf{k}}$ extracted from band structure calculation with the parameter of Figs. (3-4) of Ref.~\cite{gonzaleztudelaMain}. We focus on the situation when the atomic transition lies in the bandgap, $\Gamma_{ij}=i J_{ij}$. To optimize the number of points of numerical integration, we integrate over one quadrant around the X point (that we denote by $BZ_4$ as schematically depicted in Panel (a)) and multiply by 4 to consider the contribution of the whole area; that is, we take:
\begin{align}
\label{coupling15}
J_{ij}=8 L^2\lim_{s\rightarrow 0} \int_{BZ_4}\frac{d^2 \mathbf{k}}{(2\pi)^2}\frac{|g_{\mathbf{k}_{c}-\mathbf{k}}|^2}{s+i(\omega_{\mathrm{a}}-\omega(\mathbf{k}_{c}-\mathbf{k}))} \cos(\mathbf{k}\cdot \mathbf{r}_{ij})\,.
\end{align}
By choosing an atomic position such that $|\mathbf{r}_{ij}|=d$, we obtain $J_{ij}/\Gamma_a \sim 30$ for detunings $\Delta_{\mathrm{a}c}/2\pi\sim 30$ MHz. The scaling with $\Delta_{\mathrm{a}c}$ is logarithmic as predicted by $K_0(r_{ij}/\xi)$ using the isotropic approximation [see Fig.~\ref{fig6sup}(g)]. The best fit of $A$ (such that the isotropic formula and the numerical integration match) is $A_{\mathrm{fit}}\sim 1.8\times 10^{12} $ $\mu$m$^2$/s (using $\eta=1/2$), which is closer to the curvature along the $X-M$ direction, $A_{X-M}\sim 1.5\times 10^{12} $ $\mu$m$^2$/s, than along the $X-\Gamma$ direction, $A_{X-\Gamma}\sim 1.1\times 10^{13} $ $\mu$m$^2$/s. This is because the flatter direction is also more weighted by $1/L_m$ than the one in $X-\Gamma$.
\subsection{Ratio between coherent and incoherent processes.}
In the previous Section, we estimated the rate of the coherent processes in our system ($J_{ij}$). However, there will be several factors that limit the coherence of our system. The absence of a complete $3$-D bandgap or the presence of other polarization modes yields a decay into other radiative channels, that we estimated using FDTD calculations (not shown) to be $\Gamma'\approx 0.4 \Gamma_{\mathrm{a}}$ for our structure in Figs. 3 and 4 in Ref. \cite{gonzaleztudelaMain}.
Moreover, the fabrication process of these structures results in imperfections of the photonic crystal that yields scattering of the guided photons to the non-guided ones at a rate $\kappa$, that is characterized by the so-called quality factor $Q=\omega_c/\kappa$. For square lattices of $15 \times 15$ holes $Q\sim 10^3-10^4$ has been reported~\cite{kwon03a,cho05a}. Further improvements in the material and fabrication quality yield higher $Q\sim 10^6-10^7$ as recently reported~\cite{srinivasan02a,taguchi11a,sekoguchi14a}.
It can be shown~\cite{douglas13a} that the error rate introduced by both $\kappa$ and $\Gamma'$ within the exchange time of an excitation between two spins ($1/J_{ij}$) through the off-resonant atom-induced cavity is given by $\kappa_{\mathrm{eff}}=\kappa J_{ij}/\Delta_{\mathrm{ac}}+\Gamma'$. In order to quantify the ratio between coherent and incoherent processes, we define:
\begin{equation}
\label{coop}
\mathcal{N}=\frac{J_{ij}}{\kappa_{\mathrm{eff}}}=\frac{J_{ij}}{\kappa J_{ij}/\Delta_{\mathrm{ac}}+\Gamma'}\,,
\end{equation}
which intuitively represents the number of cycles of the coherent exchange before a non-desired transition occurs (e.g., a decay event into a continuum of unguided modes). In Fig.~\ref{fig7sup}, we plot $\mathcal{N}$ for state of the art qualify factors $Q= 10^6,10^7$, as well as a projection to $Q=10^8$. We also consider improvements in design and fabrication of the PCW which might lead to reduced band curvature $A$ and radiation loss $\Gamma'$. We find that there is an optimal detuning that maximizes $\mathcal{N}$, i.e., when $\Delta_{\mathrm{ac}}\simeq\kappa J_{ij}/\Gamma'$. In Fig.~\ref{fig7sup}(a) we plot $\mathcal{N}$ for the structure considered in the main manuscript, using $\Gamma'=0.4\Gamma_{\mathrm{a}}$, and $Q=10^6$ (black), $10^7$ (blue) and $10^8$ (green). Using $Q=10^7$, the maximum $\mathcal{N}\sim 35$ is obtained for a detuning $\Delta_{\beta}\sim 10$~GHz, which yields $J_{ij}\sim 16 \Gamma_{\mathrm{a}}$ and $\xi_{\beta}\sim 16 d$.
In Fig.~\ref{fig7sup}(b), we explore the effect of a reduced curvature $A=A_{\mathrm{fit}}/10$ using $J_{ij}^\mathrm{iso}$ defined in the previous section, by keeping $\Gamma'=0.4\Gamma_{\mathrm{a}}$, obtaining approximately a two-fold enhancement of $\mathcal{N}$. Finally, we explore the effect of reducing $\Gamma'$ to $0.1\Gamma_{\mathrm{a}}$ in panel (c), obtaining approximately a five-fold enhancement of $\mathcal{N}$, being able to reach $\mathcal{N}\sim O(10^3)$ for $Q\sim 10^9$.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.99\linewidth]{cooperativitySupMatv2.pdf}
\caption{(a) Number of coherent cycles $\mathcal{N}$ as a function of detuning $\Delta_{\mathrm{ac}}$ for the structure of the main manuscript, $\Gamma'=0.4\Gamma_{\mathrm{a}}$ and $Q=10^6$ (black), $Q=10^7$ (blue) and $Q=10^8$ (green). Dashed red line correspond to the normalized effective interaction length $\xi/d$ (b) Same as (a), but using $J_{ij}^\mathrm{iso}$ with a reduced curvature, $A=A_{\mathrm{fit}}/10$, with respect to the structure of the main manuscript. (c) Same as (b), but with reduced $\Gamma'=0.1\Gamma_{\mathrm{a}}$.}
\label{fig7sup}
\end{figure}
\subsection{Engineering with a general $\Lambda$ scheme. }
Here we derive an effective hamiltonian describing the interaction between an effective spin level formed by two metastable atomic states and the guided modes of the structure by using a $\Lambda$-level as shown in Fig.~\ref{fig6sup}(e). This implementation has several advantages: i) it allows to engineer more complex hamiltonians, e.g., XXZ spin hamiltonians; ii) the Raman process effectively narrows the natural linewidth of the excited state keeping the cooperativity of the process constant \cite{douglas13a}; iii) it requires only a single polarization bandgap, relaxing the requirements in lattice geometry and index contrast. By going to a rotating frame of the driving fields, the complete hamiltonian, with the notation of Fig.~\ref{fig6sup}(e), is given by:
\begin{align}
\label{hamlaserLi}
H=H_0+H_L+H_I=(\omega_{g1}+\omega_{L,1}) \ket{g_1}\bra{g_1} +(\omega_{g2}+\omega_{L,2}) \ket{g_2}\bra{g_2}+\omega_{e} \ket{e}\bra{e}+\sum_k \omega_k \ud{a_k} a_k +\nonumber \\ \frac{\Omega_1}{2}\big(\ket{g_1}\bra{e} +\mathrm{h.c.}\big)+\frac{\Omega_2}{2}\big(\ket{g_2}\bra{e} +\mathrm{h.c.}\big)+\sum_k g_k \big(\ud{a_k} \ket{g_1}\bra{e} e^{-i \omega_{L,1} t}+\mathrm{h.c.}\big)\,,
\end{align}
which has three different contributions: the free energy, $H_0$, the coherent driving of the lasers, $H_L$, and the interaction with the modes of the structure, $H_I$. We define the following detunings: $\Delta_i=\omega_e-(\omega_{g_i}+\omega_{L,i})$ and apply the following transformation: $ U=e^S=e^{\frac{\Omega_1}{2 \Delta_1 }(\ket{e}\bra{g_1}-\ket{g_1}\bra{e})+\frac{\Omega_2}{2 \Delta_2 }(\ket{e}\bra{g_2}-\ket{g_2}\bra{e})}$. When $|\Delta_i|\gg \Omega_{i}$, the excited state is only virtually populated and the result of the transformation, $H\rightarrow e^S H e^{-S}$, yields in lower order of $\Omega_i/\Delta_i$ to a hamiltonian with different terms: $H=H_0+H_1+H_2+H_3+H_4$. The first one is the free energy hamiltonian, $H_0$, as in the first line of Eq.~\ref{hamlaserLi}. The second and third terms, $H_1, H_2$, are the Stark-shifts and coherent driving of the ground-state levels induced by the off-resonant driving:
\begin{align}
H_1=-\frac{\Omega_1^2}{4\Delta_1} \ket{g_1}\bra{g_1} -\frac{\Omega_2^2}{4\Delta_2} \ket{g_2}\bra{g_2}\,, \\
H_2=-\frac{\Omega_1 \Omega_2}{4}\Big(\frac{1}{\Delta_1}+\frac{1}{\Delta_2}\Big) \big(\ket{g_1}\bra{g_2} +\ket{g_2}\bra{g_1}\big)\,.
\end{align}
$H_1$ can be considered as a renormalization of the free energies where $H_2$ describes Raman processes. Finally, defining an effective qubit with $\{\ket{g_1},\ket{g_2}\}$, such that $\sigma=\ket{g_1}\bra{g_2}$, $\ud{\sigma}=\ket{g_2}\bra{g_1}$ and $\sigma_z=\ket{g_1}\bra{g_1}+1/2$, the transformation of the $H_I$ yields to \cite{footnote1}:
\begin{align}
\label{hamlaserLi4}
H_3=-\sum_k g_k \frac{\Omega_2 }{2\Delta_2} \big(\ud{a_k} \sigma e^{i(\omega_k-\omega_{g_2}+\omega_{g_1} -\omega_{L,2} )t} +\mathrm{h.c.}\big)\,, \\
H_4=-\sum_k g_k \frac{\Omega_1 }{2\Delta_1} \big(\ud{a_k} \sigma^z e^{i (\omega_k-\omega_{L,1} ) t} +\mathrm{h.c.}\big)\,.
\end{align}
\noindent which are the most relevant terms for our discussion, as they contain the interaction with the GMs of the structure ($a_k$). By extending these effective light-matter hamiltonians to many atoms, and adiabatically eliminating the GMs in the conditions where the atomic transitions lie in the bandgap leads to the following XXZ spin hamiltonian $H_{\mathrm{spin}}$ from Eqs. SM29-SM30:
\begin{equation}
\label{spin}
H_{\mathrm{spin}}=\sum_{i,j} [J^{z}_{ij} \sigma_i^z\sigma_j^z +J^{xy}_{ij}\ud{\sigma_i}\sigma_j]\,
\end{equation}
where $J^{\beta}_{ij}=h_{\beta} \Gamma_{2d} K_0(r_{ij}/\xi_\beta)$, where we have used the effective mass and isotropic approximation (see previous Section). The factor $h_\beta=\big(\Omega_l/(2\Delta_{\beta})\big)^2$ with $l=1\,(2)$ for $\beta=z\, (xy)$ can be controlled independently for each $\beta$ through the laser intensities, $\Omega_l$, and detunings, $\Delta_l$. The effective length scale, $\xi_\beta=\sqrt{A/\Delta_{\beta}}$, depends on both the curvature of the band, $A$, and effective detuning, $\{\Delta_{xy}, \Delta_z\} \lessgtr 0$, where $\Delta_{xy}=\omega_{g,2}-\omega_{g,1}+\omega_{L,2}-\omega_c$ and $\Delta_z=\omega_{L,1}-\omega_c $. Therefore, both the length scale, through $\xi_\beta$, and the strength of the interactions, through $h_\beta$, can be tuned independently for each $J^{\beta}_{ij}$-component.
\section{GaP: Material Properties and Fabrication. }
Our analysis in Ref. \cite{gonzaleztudelaMain} considers Gallium Phosphide (GaP) for the various examples presented. We have made this choice in order to present the most favorable case for sub-wavelength optical traps in nanophotonic dielectric waveguides. GaP is a high index ($n \simeq 3.25$), low loss III-IV semiconductor with an indirect band gap at $550$~nm. For other dielectrics transparent in the frequency range of electronic transitions for alkali atoms (e.g., SiN, SiC, TiO$_2$), the lattice constant $d_{\mathrm{min}}=\lambda_0/2n$ would be larger than for GaP (e.g., $n_{SiN}=2.0$, $n_{SiC}=2.6$, and $n_{TiO_2}=\{2.5.2.8\}$ for ordinary and extraordinary polarizations, respectively), even if these materials may have other more favorable properties (e.g., lower absorption) and more advanced processing capabilities. In fact, our schemes for vacuum trapping (Fig. 2) and photon-mediated atomic interactions (Figs. 3, 4) do not rely on such a high-index material as GaP.
That said, here we gather relevant information related to material properties and state-of-art fabrication for GaP. Beginning with material losses for bulk GaP, we note that Fig. 2 in Ref. \cite{dean67a} considers the intrinsic optical absorption of GaP and finds an absorption coefficient $\alpha < 0.1/$cm for wavelengths larger than approximately $550$~nm, where the intensity attenuation is given by $I(z)=I_0e^{-\alpha z}$. These authors further explore low-level interband absorption due to Arsenic dopants \cite{dean69a}, emphasizing the need for high purity materials for our application.
More recently, nanophotonic structures have been fabricated in GaP and quality factors measured over a range of wavelengths in the visible and near infrared. Examples include a hybrid ring resonator of $900$~nm diameter with quality factor $Q \simeq 6800$ for $\lambda_0 \simeq 637$~nm \cite{barclay11a} and a photonic crystal cavity with lattice constant $d \simeq 200$~nm and measured $Q \simeq 12000$ for $\lambda_0$ around $800$~nm \cite{rivoire09a}. The associated absorption coefficients are $\alpha \simeq 24/$~cm and $\alpha \simeq 11/$cm, respectively, more than $10^2$ times larger than the limit from Ref. \cite{dean67a}.
The highest reported quality factor is $Q\simeq 2.8\times 10^5$ at $\lambda_0 \simeq 1.55$~$\mu$m for a GaP micro disk of diameter approximately $6$~$\mu$m \cite{mitchell14a}. The corresponding absorption coefficient $\alpha \simeq 0.24/$cm, which approaches the ultra-low loss regime of Ref. \cite{dean67a}.
These loss values for fabricated GaP structures represent steady progress of improved performance, including for a photonic crystal \cite{rivoire09a} comparable to those in our manuscript \cite{gonzaleztudelaMain}. For example, the structure in Fig. 3 \cite{gonzaleztudelaMain} has lattice constant $d=316$nm. A lattice of $60\times60$ unit cells would have linear dimension $L=60 d \simeq 20$~$\mu$m and could accommodate a few thousand atoms. For $\alpha \simeq 11/$cm as for the photonic crystal in Ref. \cite{rivoire09a}, the reduction in intensity for a propagating guided mode across the lattice would be $\alpha L \sim 2 \times 10^{-2}$ (i.e., $2\%$ loss) far from a band edge, with further reduction by perhaps $>10 \times$ suggested by the limits in Refs. \cite{dean67a,mitchell14a}. However, the loss would be further increased near a band edge, especially for the flat bands considered in our manuscript. We are currently investigating this issue by way of FDTD calculations for our proposed structures.
Apart from reducing absorption for guided modes used for atom-atom interactions, minimizing absorption in the PCW structures of our manuscript is critical since the intensities employed for atom trapping will be high. The largest Rabi frequency is $\Omega_{\mathrm{SI}}/2\pi = 130$GHz for the scheme in Fig. 2 \cite{gonzaleztudelaMain}, corresponding to intensity $I \sim 0.1$~W/$\mu$m$^2$. This high value is mitigated by the fact that the planar PCW is located near a node in the standing wave formed by the two counter propagating side illumination beams. Nevertheless, thermal management and optical nonlinearities of the dielectric will be important issues.
To gain some perspective on these issues, we refer to the generation of frequency combs with microscopic ring resonators \cite{kippenberg11a}. Ref. \cite{levy10a} reports a Silicon Nitride (SiN) ring resonator of cross sectional area $\sim 3 \mu$m$^2$ pumped with $P_{in} \simeq 300$~mW leading to circulating power within the ring resonator $P_c \simeq 300$W and intracavity intensity $I \sim 10^{2}$~W/$\mu$m$^2$ without optical damage. The high intensity leads to nonlinear interactions to generate a frequency comb, from which the authors determine the value of the nonlinear index of refraction for SiN to be $n_2 \simeq 2.5\times10^{-15}$~cm$^2$/W (i.e., about $10\times$ larger than for silica).
For our structures with $I \sim 0.1$~W/$\mu$m$^2$ as in Fig. 2a for the vacuum lattice, the nonlinear phase shift across the thickness of the substrate $W \simeq 120$~nm would be $\delta \phi \lesssim 10^{-6}$~radians for a single side illumination beam. Here, we use the nonlinear index quoted in Ref. \cite{liu10a}, namely $n_2 \simeq 7\times10^{-14}$~cm$^2$/W determined for GaP illuminated with femtosecond laser pulses at $1040$nm.
Certainly, we recognize that these estimates provide only a basis for some optimism and motivation for further investigations of the feasibility of the structures described in our manuscript \cite{gonzaleztudelaMain}. The realization of our proposals for nanophotonic atomic lattices requires overcoming significant challenges in material characterization and device fabrication not by a literature survey but rather by a dedicated research program, including materials other than GaP.
|
\section{Introduction}
The study of decays is an important subject of atomic, nuclear and particle
physics \cite{ghirardi}. Some subatomic particles possess a lifetime which is
so short that they can be seen only through their decay products, and hence
one usually calls them resonances. This is indeed the case for the recently
discovered Higgs particle, see \emph{e.g.} Refs. \cite{djouadi,ellis} and references
therein. In the realm of the strong interactions also many hadrons were
discovered via their decay processes \cite{amslerrev}; in addition to that,
decays turn out to be crucially important for the understanding of their quantum
numbers and inner structure.
The main problem concerning the fundamental theory of quarks and gluons
(Quantum Chromodynamics or {\em QCD}) is the fact that this theory is
non-perturbative in the low-energy regime. Hence one relies on other
approaches, as for instance effective models based on symmetries
\cite{geffen,meissner}, where the physical degrees of freedom are not quarks
and gluons, but composite particles, namely hadrons. Decays of hadrons have often
been evaluated within such models in the lowest order approximation -- in
other words at tree-level, see \emph{e.g.} Refs. \cite{ko,tq,elsm,dick} and
references therein. In particular, in the recent work of Ref. \cite{dick}
decays of various mesons up to $1.5$ GeV were computed in a chirally and
dilatation invariant framework and were found to be in agreement with the
corresponding experimental values as provided by the PDG \cite{pdg}.
A two-body tree-level decay is the easiest nontrivial process in quantum field
theory. It is depicted in Fig. \ref{fig:fig1}a: The unstable bosonic
particle $S$ decays into two identical particles, denoted as
$\phi$. The decay amplitude is simply a constant in the case of scalar
particles and non-derivative interactions. When derivatives and/or particles
with nonzero total angular momentum $J$ are considered, a dependence of the
momenta appears in the tree-level amplitude(s).
The next step in the context of effective models has been the study of
(hadronic) loops, see for instance Refs.
\cite{dullemond,coito,tornqvist,pennington,achasov,giacosaSpectral,duecan,e38,salam,Caprini,garcia}
and references therein. The leading contribution to the self-energy is shown
by the diagram in Fig. \ref{fig:fig1}b. Both the mass and the width of the decaying particle are influenced by the quantum fluctuations due
to the coupling to hadronic intermediate states. The optical theorem assures
that the imaginary part of the one-loop diagram from Fig. \ref{fig:fig1}b coincides with the tree-level decay formula. The
unstable particle is described by a spectral function (\emph{i.e.}, an energy
distribution), which is given by the imaginary part of the one-loop resummed
propagator. Alternatively, the properties of the unstable particle can also be
described as a complex pole in the appropriate unphysical Riemann sheet, a
procedure first proposed by Peierls \cite{peierls} a long time ago. The
quantum theoretical treatment of unstable particles became an object of much
interest, see \emph{e.g.} Refs. \cite{hoehler,levypoles,aramaki,landshoff}. The general
outcome of such studies is -- disregarding problems like mixing -- that when
the particle is narrow-shaped, quantum fluctuations have a small influence on
its properties (\emph{i.e.}, mass and width) but are non-negligible for broad
resonances. It turns out to be the ratio `width over mass' that is decisive
here: as long as this number is smaller than $\sim0.3$ the loop contributions
have a small impact \cite{giacosaSpectral}.
\begin{figure}[h]
\centering
\hspace{-0.4cm}\includegraphics[scale=0.75]{fig1-eps-converted-to.pdf}
\caption{a) Decay process $S\rightarrow\phi\phi$ at leading order (tree-level). b) Self-energy at
leading order.}
\label{fig:fig1}
\end{figure}
There is, however, another open issue: what is the role of the next-to-leading
order (NLO) diagram for hadronic decays? We depict this kind of
triangle-shaped diagram in Fig. \ref{fig:fig2}; it is proportional
to the third power of the coupling constant. In the context of hadronic decays
in effective field theories/models it is usually not taken into account.
Nevertheless, one should stress that the coupling constant in hadronic models
is in general not a small number, thus there is a priori no guarantee that the
NLO diagram is smaller than the tree-level one.
\begin{figure}[ptb]
\centering
\hspace{-1.0cm}\includegraphics[scale=0.75]{fig2.pdf}\caption{Triangle-shaped NLO diagram for a two-body decay.}
\label{fig:fig2}
\end{figure}
The aim of this work is to close this gap. To this end, we evaluate the role
of the triangle diagram from Fig. \ref{fig:fig2} in the case of a
simple scalar field theory without derivative interactions. We do this in
plain perturbation theory (\emph{i.e.}, without resummation), meaning that the
virtual $S$-particle exchanged in Fig. \ref{fig:fig2} is described by
its free propagator. After discussing the analytic properties of the triangle
diagram, we adopt our results to some decays of well-known scalar resonances:
$f_{0}(500)$, $f_{0}(980)$, $f_{0}(1370)$ and $f_{0}(1500)$ \cite{pdg}. All
these resonances decay predominately in two pions and are therefore a good
test for our purpose. For completeness, we also look at the pion-pion and kaon-kaon decay channels of $f_{0}(1710)$ as well as the kaon-kaon decay of the scalar-isovector state $a_{0}(1450)$ \cite{pdg}. Yet, our investigation is quite general and applies also to decays involving derivatives and particles with spin.
The main result of our work is that the triangle contribution is indeed
\emph{negligible} and one can consequently justify a posteriori all previous
studies in which those types of contributions (and, in turn, higher-order
contributions as well) were not taken into account. Since in the field of
hadron physics there are usually other (and even larger) sources of
uncertainties due to various (and sometimes subtle) approximations and
simplifications, the restriction to the leading order tree-level diagram from
Fig. \ref{fig:fig1}a and to the (resummed) one-loop quantum
corrections from Fig. \ref{fig:fig1}b are reasonable and usually sufficient.
The paper is organized as follows: we present the model and some
analytic aspects of the triangle diagram in Section \ref{seq:model}, while the numerical
results are shown in Section \ref{seq:results}. The last Section \ref{seq:conc} contains the conclusions.
\section{The model and the triangle diagram}
\label{seq:model}
We introduce a model with the scalar fields $S$ (with mass $m_{S}$) and $\phi$ (with mass $m_{\phi}$) described by the Lagrangian
\begin{equation}
\mathcal{L}=\frac{1}{2}(\partial_{\mu}S)^{2}-\frac{m_{S}^{2}}{2}S^{2}+\frac{1}{2}(\partial_{\mu}\phi)^{2}-\frac{m_{\phi}^{2}}{2}\phi^{2}+gS\phi^{2} \ .
\end{equation}
The interaction term induces the decay process $S\rightarrow\phi\phi$. The
parameter $g$ is the coupling constant (with dimension of energy). For
previous studies and details of the model see Refs.
\cite{achasov,giacosaSpectral,duecan,e38,veltman,thomasthesis}. The decay
width in perturbation theory can be expressed as
\begin{equation}
\Gamma_{S\rightarrow\phi\phi}=\frac{1}{2}\frac{\sqrt{\frac{m_{S}^{2}}%
{4}-m_{\phi}^{2}}}{8\pi m_{S}^{2}}|\hspace{0.02cm}\mathbf{-}\hspace
{0.05cm}i\mathcal{M}|^{2}\text{ } \ , \label{gammafull}
\end{equation}
where the decay amplitude $-i\mathcal{M}$ in perturbation theory is written as
a sum
\begin{equation}
-i\mathcal{M}=-i\mathcal{M}_{1}-i\mathcal{M}_{3}+\dots \ .
\end{equation}
The term $-i\mathcal{M}_{2n+1}$ represents the contribution of order
$g^{2n+1}$ to the amplitude. It is quite remarkable that an exact solution to
this problem has not yet been found (there are, however, quantum mechanical
models for which this is possible, see Refs.
\cite{ghirardi,lee,kofman,facchiprl,giacosapra,duecan,shimizu}).
\bigskip
1) \emph{Leading order}: The leading order with $n=0$ is given by the
tree-level amplitude of Fig. \ref{fig:fig1}a, for which one obtains
\begin{equation}
-i\mathcal{M}_{1}=2ig \ .
\label{eq:tree_amplitude_phiphi}
\end{equation}
Thus, the tree-level decay width simply reads
\begin{equation}
\Gamma_{S\rightarrow\phi\phi}^{\text{tree}}=\frac{1}{2}\frac{\sqrt{\frac
{m_{S}^{2}}{4}-m_{\phi}^{2}}}{8\pi m_{S}^{2}}\hspace{0.03cm}(2g)^{2} \ .
\label{eq:tree_width_phiphi}
\end{equation}
2) \emph{The one-loop diagram}: The loop diagram shown in Fig.
\ref{fig:fig1}b actually does \emph{not} enter directly into the
expression of Eq. (\ref{gammafull}), because the latter is valid in plain
perturbation theory without resummation. The role of the one-loop resummation
has been widely studied and, although not directly relevant for our
calculation, we recall the main features in view of its general importance for
the problem that we are studying and for future works \cite{dullemond,coito,tornqvist,pennington,achasov,giacosaSpectral,duecan,e38,thomasthesis}%
. By denoting $\Pi(p^{2})$ as the loop contribution, the propagator of the
field $S$ changes upon resummation to
\begin{equation}
\Delta_{S}(p^{2})=\frac{i}{p^{2}-m_{S}^{2}+i\epsilon}\ \ \rightarrow \
\ G_{S}(p^{2})=\frac{i}{p^{2}-m_{S}^{2}-\Pi(p^{2})+i\epsilon} \ .
\end{equation}
The self-energy $\Pi(p^{2})$ is linked to the tree-level decay width via the
optical theorem:
\begin{equation}
-\operatorname{Im}\Pi(p^{2})=\sqrt{p^{2}}\ \Gamma_{S\rightarrow\phi\phi
}^{\text{tree}}\hspace{0.03cm}(m_{S}^{2}\rightarrow p^{2})=\sqrt{p^{2}}\ \frac{g^{2}\sqrt{\frac{p^{2}}{4}-m_{\phi}^{2}}}{4\pi p^{2}} \ .
\end{equation}
The properties of the unstable particle $S$ (\emph{i.e.}, its mass and decay width)
are often identified with the complex pole of the full propagator $G_{S}%
(p^{2})$ in the second Riemann sheet, $p^{2}=\big(m_{S}^{\text{pole}}%
-i\hspace{0.03cm}\Gamma_{S\rightarrow\phi\phi}^{\text{pole}}/2\big)^{2}$. The
(shifted) mass $m_{S}^{\text{pole}}$ is given by the real part of the pole and
the modified decay width $\Gamma_{S\rightarrow\phi\phi}^{\text{pole}}$ by the
negative imaginary part multiplied by two. For a small coupling constant $g$ the
quantities $\Gamma_{S\rightarrow\phi\phi}^{\text{pole}}$ and $\Gamma
_{S\rightarrow\phi\phi}^{\text{tree}}$ are close to each other, but then
deviate when increasing $g$ \cite{thomasthesis}.
At this point it should be stressed that hadronic theories do not need to
undergo a renormalization process as for theories of elementary particles,
since they are only valid in a limited energy regime. (For instance, if we
restrict our attention to mesons made of $u$-, $d$-, and $s$-quarks, the
associated `cutoff' has a value of about $\sim1.5$ GeV.) For this reason, a
finite energy cutoff is used when evaluating the loop diagram from Fig.
\ref{fig:fig1}b. Different forms for the consequently needed cutoff
function can be applied, but the general outcome shows just a soft dependence on
the precise choice as long as convergence is guaranteed \cite{giacosaSpectral}.
In an ideal scattering experiment of the type $\phi\phi\rightarrow\phi\phi,$
the unstable state $S$ manifests itself as an enhanced peak for a center of
mass energy close to $m_{S}.$ More precisely, the spectral function
$d_{S}(E)=-\frac{2E}{\pi}\operatorname{Im}G_{S}(E^{2})$ plays an important
role: as was argued long time ago by Matthews and Salam \cite{salam}, it can
be interpreted as a `mass distribution' of the unstable particle $S$, which
can be well described by a Breit--Wigner function for narrow resonances.
Namely, even if there are low-energy threshold(s) and high-energy distortions,
as long as the ratio $\Gamma_{S\rightarrow\phi\phi}^{\text{tree}}/m_{S}$ is a
small number, the role of hadronic loop contributions is small
\cite{giacosaSpectral,duecan}. However, when this ratio becomes large one
observes big deviations from a typical Breit--Wigner peak and in some cases a
very peculiar phenomenon takes place, often called pole-doubling. This means
that new poles can emerge in the unphysical Riemann sheet(s)
\cite{tornqvist,pennington,dullemond}. Although there is only one unstable
state in the Lagrangian, quantum fluctuations might be able to generate two
(or more) resonance poles. Such a
mechanism of dynamical generation is present in the ongoing debate among
hadron physicists on where some of the (known) resonances in the hadron
spectrum arise from.\footnote{For a general discussion concerning dynamical
generation see Ref. \cite{giacosaDynamical} and references therein.} A
particular interesting field of study is that of charmed mesons, see Refs.
\cite{coito,brambilla}.
\newpage
3) \emph{The triangle diagram}: We now turn our attention to the main subject of this work: the NLO (non-resummed) perturbation theory. To this end, we evaluate the
triangle diagram corresponding to the amplitude $-i\mathcal{M}_{3}$ as
depicted in Fig. \ref{fig:fig2}. Its analytic expression takes the
form
\begin{align}
-i\mathcal{M}_{3}(p & \rightarrow kk^{\prime})\label{eq:3amplitude}\\
& =\ 8ig^{3}\int\frac{\text{d}^{4}q}{(2\pi)^{4}}\ \frac{1}{q^{2}-m_{\phi}^{2}+i\epsilon}\hspace{0.03cm}\frac{1}{(q-p)^{2}-m_{\phi}^{2}+i\epsilon}\hspace{0.03cm}\frac{1}{(q-k)^{2}-m_{S}^{2}+i\epsilon} \ . \nonumber
\end{align}
Note, the propagator of $S$ is taken as the free propagator $\Delta_{S}%
(p^{2}).$ The factor of 8 arises due to identical particles in the final state. Solving the integration over $q^{0}$ by the residue theorem and
after introducing spherical coordinates the amplitude can be re-expressed
as
\begin{equation}
\mathcal{M}_{3}(p\rightarrow kk^{\prime})=\frac{2g^{3}}{\pi^{2}}\int
_{0}^{\infty}\mbox{d}u\;u^{2}\int_{-1}^{1}\mbox{d}\chi\;\big\{P_{12}(u,\chi)+P_{3}(u,\chi)\big\} \ , \label{eq:m3second}
\end{equation}
where
\begin{align}
P_{12}(u,\chi)\ = & \ \ \frac{1}{8\sqrt{u^{2}+m_{\phi}^{2}}\big(m_{\phi}%
^{2}+w^{2}(1-\chi^{2})\big)}\\
& \ \times\ \frac{u^{2}-2w^{2}+uw\chi}{(u-p_{1}-i\epsilon)(u-p_{2}%
+i\epsilon)(u-p_{3}(\chi)-i\epsilon\xi)(u-p_{4}(\chi)+i\epsilon\xi
)} \ ,\nonumber\\[10pt]
\xi\ = & \ \hspace{0.08cm}\sgn(-\triangle_{2}-2uw\chi) \ ,\\[7pt]
P_{3}(u,\chi)\ = & \ \ \frac{1}{8u\sqrt{u^{2}+m_{S}^{2}}\big(w^{2}(\chi
^{2}-1)-m^{2}\big)\big(u-p_{5}(u,\chi)-i\epsilon\xi^{\prime}\big)} \ ,\\[7pt]
\xi^{\prime}\ = & \ \hspace{0.08cm}\sgn\left(\frac{m_{S}^{2}}{u}%
-2w\chi\right) \ ,\\[7pt]
\triangle_{1}\ = & \ \ m_{S}^{2}-m_{\phi}^{2} \ ,\\[5pt]
\triangle_{2}\ = & \ \ 2m_{\phi}^{2}-m_{S}^{2} \ .
\end{align}
Here we introduced $|\mathbf{q}|=u$, $|\mathbf{k}|=w$ and $\chi=\cos\theta$,
where $\theta$ is the angle between $\mathbf{q}$ and $\mathbf{k}$. We applied
the shift $\mathbf{q}\rightarrow\mathbf{q}-\mathbf{k}$ for $P_{3}$ before the
transformation to spherical coordinates. The integrand in Eq.
(\ref{eq:m3second}) has five poles:
\begin{align}
p_{1}\ = & \ \ w\ ,\label{eq:Pole4Pre}\\
p_{2}\ = & \ -\hspace{-0.07cm}w\ ,\\
p_{3}(\chi)\ = & \ \ \frac{-w\chi\triangle_{2}-\sqrt{u^{2}\chi^{2}%
\triangle_{2}^{2}-4w^{2}\triangle_{1}(-w^{2}-m_{\phi}^{2}+w^{2}\chi^{2})}%
}{2(-w^{2}-m_{\phi}^{2}+w^{2}\chi^{2})}\nonumber\\[7pt]
\ = & \ \ \frac{-2w\chi\triangle_{2}-wm_{S}\sqrt{4w^{2}+m_{\phi}^{2}%
(3+\chi^{2})}}{2(-w^{2}-m_{\phi}^{2}+w^{2}\chi^{2})}\ ,\\[7pt]
p_{4}(\chi)\ = & \ \ \frac{-w\chi\triangle_{2}+\sqrt{w^{2}\chi^{2}%
\triangle_{2}^{2}-4w^{2}\triangle_{1}(-w^{2}-m_{\phi}^{2}+w^{2}\chi^{2})}%
}{2(-w^{2}-m_{\phi}^{2}+w^{2}\chi^{2})}\nonumber\\[7pt]
\ = & \ \ \frac{-2w\chi\triangle_{2}+wm_{S}\sqrt{4w^{2}+m_{\phi}^{2}%
(3+\chi^{2})}}{2(-w^{2}-m_{\phi}^{2}+w^{2}\chi^{2})}\ ,\\[5pt]
p_{5}(\chi)\ = & \ \ \frac{m_{S}^{2}w\chi}{w^{2}(\chi^{2}-1)-m^{2}}\ .
\end{align}
One finds that only $p_{1},$ $p_{3}$ and $p_{5}$ are positive and located on
the path of integration (where, for the latter, this in fact depends on the
value of $\chi$), and thus contribute to the imaginary part of the integral in Eq. (\ref{eq:m3second}). The contribution from $p_{1}$ is
easy to calculate analytically, whereas the one from $p_{3}$ yields a rather
complicated result and is therefore computed numerically. Notice that just the
term $P_{12}$ contains those two singularities and that the contribution of
$p_{5}$ vanishes.
Before moving to our final results, two comments are in order:
\begin{itemize}
\item The NLO diagram in Fig. \ref{fig:fig2} is convergent and
well-defined also for an infinite cutoff. However, as we explained above, a
finite value of the cutoff naturally comes into a hadronic theory because of
the non-elementary nature of the fields
\cite{closecutoff,giacosaSpectral,duecan} and the finite size of the
corresponding particles \cite{tornqvist,pennington}, respectively. One should
in principle evaluate this diagram by including such a cutoff even if
convergence is ensured in the limit of $\Lambda\rightarrow\infty$. It turns
out that for what concerns the triangle diagram the influence of the cutoff
parameter is small -- a value of about $\sim1$ GeV or taking the
infinity limit generates only small changes.
\item Triangle-shaped diagrams were indeed studied in hadron physics, but in a
rather different framework. For instance, one has studied the processes
$\pi^{0}\rightarrow\gamma\gamma$ and $f_{0}\rightarrow\gamma\gamma$
\cite{trianglegammagamma}, where $f_{0}$ represents a generic scalar state
(see Section \ref{seq:results}). These decays occur via a triangle-loop of quarks
and represent the leading order contributions (there is no tree-level diagram
for those processes). For a case in which mesonic loops contribute to $\gamma\gamma$
emission see Ref. \cite{hanhart}. The mentioned investigations have similar technical aspects to our
present interest, yet they could not give an answer to the question of the
role of the triangle diagram as a next-to-leading order contribution.
\end{itemize}
\section{Results}
\label{seq:results}
\subsection{General case}
We first present numerical results of our calculations without referring to any
particular mesonic state. To this end, we fix the energy units in the
following way:
\begin{equation}
\lbrack g]=[m_{\phi}]=[\mathcal{M}]=[\Gamma]=[\Lambda]=1\ m_{S} \ ,
\end{equation}
\emph{i.e.}, all dimensionful quantities are expressed in terms of the mass of the
unstable state $S.$
In the upper panel of Fig. \ref{fig:fig3} the ratio
$|\mathcal{M}_{3}|/|\mathcal{M}_{1}|$ of the decay amplitudes is shown for
different masses of the particle $\phi$ in dependence of the coupling constant
$g$. In this way the role of the triangle diagram is visualized. As expected,
the larger the coupling, the larger $|\mathcal{M}_{3}|/|\mathcal{M}_{1}|$ and
for equal $g$ the ratio is larger for smaller masses $m_{\phi}=m$. We denote
$g=g_{\ast}$ as the value of the coupling for which the amplitudes are equal,
$|\mathcal{M}_{3}|$ $=|\mathcal{M}_{1}|$, implying that then the triangle
diagram is exactly as large as the tree-level one. The value $g_{\ast}$ represents an upper limit for the validity of a tree-level calculation in particular and for a perturbative expansion in general (see Tab. \ref{tab:results} for a list of numerical values).
In Fig. \ref{fig:fig3}, lower panel, we also show the ratio
$\Gamma_{\text{NLO}}/\Gamma_{\text{LO}}$ for different values of $m_{\phi}$ as
function of $g,$ where the lowest order is the tree-level decay width
$\Gamma_{\text{LO}}=\Gamma_{S\rightarrow\phi\phi}^{\text{tree}}$, meaning $-i\mathcal{M}=-i\mathcal{M}_{1}=2ig$ in Eq. (\ref{gammafull}), and the
next-to-leading order is $\Gamma_{\text{NLO}}$, meaning $-i\mathcal{M}=-i\mathcal{M}_{1}-i\mathcal{M}_{3}$. We also define the two
specific couplings $g^{\prime}$ and $g^{\prime\prime}$, where $g^{\prime}$
corresponds to the ratio $\Gamma_{\text{NLO}}/\Gamma_{\text{LO}}=1.33$ (the decay
width at NLO is $33\%$ larger than the tree-level width, \emph{i.e.}, this marks a `soft' limit for the validity of the tree-level calculation) and $g^{\prime\prime}$ corresponds to $\Gamma_{\text{NLO}%
}/\Gamma_{\text{LO}}=2$ (a `hard' limit). The values of $g^{\prime}$ and
$g^{\prime\prime}$ are reported in Tab. \ref{tab:results}.
\begin{table}
\begin{center}
\begin{tabular}
[c]{|l|c|c|c|}\hline
$m_{\phi}$ & $g_{\ast}$ & $g^{\prime}$ & $g^{\prime\prime}$\\
\hline
\hline
$10^{-13}$ & $0.4581$ & $0.3459$ & $0.4571$%
\\\hline
$0.1$ & $1.6407$ & $1.1766$ & $1.5894$\\\hline
$0.2$ & $1.9365$ & $1.3327$ & $1.8316$\\\hline
$0.3$ & $2.2155$ & $1.4279$ & $2.0164$\\\hline
$0.4$ & $2.6122$ & $1.5672$ & $2.2767$\\\hline
$0.49$ & $3.4742$ & $1.4796$ & $2.3908$\\\hline
\end{tabular}
\end{center}
\caption{Specific coupling constants for different mass configurations: The case $|\mathcal{M}_{3}|$
$=|\mathcal{M}_{1}|$ is obtained for $g=g_{\ast}$, the case $g=g^{\prime}$ for
$\Gamma_{\text{NLO}}/\Gamma_{\text{LO}}=1.33$ and the case $g=g^{\prime\prime}$ for
$\Gamma_{\text{NLO}}/\Gamma_{\text{LO}}=2$. All quantities are in units of $m_{S}$.}
\label{tab:results}
\end{table}
The following comments are in order:
\vspace{0.3cm}
(i) Although the limit $m_{\phi}\rightarrow0$ is subtle because it contains
infrared divergences, even for the very small value $m_{\phi}=m=10^{-13}$ the
NLO becomes dominant only for $g\gtrsim0.45$. This case is however unrealistic
for hadronic physics in which $m_{S}\sim1$ GeV and $m_{\phi}\ge m_{\pi}$.
\vspace{0.3cm}
(ii) For the ratio $m_{\phi}/m_{S}\gtrsim0.1$, which is usually fulfilled for
decays of hadrons, one has $g_{\ast}\sim g^{\prime\prime}\gtrsim1.6$ and
$g^{\prime}\gtrsim1.2$, see Tab. \ref{tab:results}. These values of coupling constants
correspond to very large decay widths. In physical cases the value of $g$ is
usually safely smaller, showing that the NLO is subdominant, see next Subsection for explicit examples.
\begin{figure}
\centering
\hspace{1.0cm}\includegraphics[scale=0.6]{fig3_mod.pdf}
\caption{Upper panel: Ratio $|\mathcal{M}_{3}|/|\mathcal{M}_{1}|$ of the decay
amplitudes for the mass configurations $m_{\phi}=m=10^{-13},\hspace
{0.05cm}0.1,\hspace{0.05cm}0.2,\hspace{0.05cm}0.3,\hspace{0.05cm}%
0.4,\hspace{0.05cm}0.49$ in dependence of the coupling constant $g$. Lower
panel: Ratio $\Gamma_{\text{NLO}}/\Gamma_{\text{LO}}$ of the decay widths for
the same mass configurations in dependence of the coupling constant $g$.}
\label{fig:fig3}
\end{figure}
(iii) The fact that $g_{\ast}\sim g^{\prime\prime}$ is possible because the
NLO amplitude, $-i\mathcal{M}_{3}$, has a dominant imaginary part. As a
consequence
\begin{align}
\Gamma_{\text{NLO}} & = \vspace{0.25cm} \frac{1}{2}\frac{\sqrt
{\frac{m_{S}^{2}}{4}-m_{\phi}^{2}}}{8\pi m_{S}^{2}}|\hspace{0.02cm}%
\mathbf{-}\hspace{0.05cm}i\mathcal{M}_{1}\mathbf{-}\hspace{0.05cm}%
i\mathcal{M}_{3}|^{2}\simeq\vspace{0.25cm}\frac{1}{2}\frac{\sqrt{\frac{m_{S}^{2}}{4}-m_{\phi
}^{2}}}{8\pi m_{S}^{2}}\left(|\hspace{0.02cm}\mathbf{-}\hspace
{0.05cm}i\mathcal{M}_{1}|^{2}+|\hspace{0.02cm}\mathbf{-}\hspace{0.05cm}%
i\mathcal{M}_{3}|^{2}\right) \nonumber \\
& =\vspace{0.25cm}\Gamma_{\text{LO}}+\frac{1}{2}\frac{\sqrt
{\frac{m_{S}^{2}}{4}-m_{\phi}^{2}}}{8\pi m_{S}^{2}}|\hspace{0.02cm}%
\mathbf{-}\hspace{0.05cm}i\mathcal{M}_{3}|^{2} \ ,
\end{align}
which shows that $\Gamma_{\text{NLO}}$ is always larger
than $\Gamma_{\text{LO}}$ (interference effects that
involve only the real part of $\mathbf{-}\hspace{0.05cm}i\mathcal{M}_{3}$ are
small) and that, when the NLO equals the LO, the width doubles.
\vspace{0.3cm}
(iv) Although in our study one has a decay into particles with equal masses,
nothing substantial would change for the decay into two particles with different masses.
The numerical evaluation would be more involved because other poles could contribute.
\subsection{Specific examples}
We now turn to the concrete examples of the well-known scalar resonances
$f_{0}(500)$, $f_{0}(980)$, $f_{0}(1370)$ and $f_{0}(1500)$ and calculate
their decays into pions \cite{pdg} (for a
discussion of the internal structure of these states in terms of quarks and
gluons see also \emph{e.g.} Refs. \cite{ko,tq,elsm,dick,giacosaSpectral,dullemond,Caprini,lohse,Black}). We present the results in Fig. \ref{fig:fig4}, plotting the decay
widths as function of $g$. The experimental value
of the width (which is $\Gamma_{f_{0}\rightarrow\pi\pi}/3$ since only the
$\pi^{0}\pi^{0}$-channel is considered) is marked on the $y$-axis and the error bars are
indicated by a gray band. In the case of the two resonances below $1$ GeV the
data was taken from the dispersive analysis of Ref. \cite{garcia}, while
for the other two resonances the values from the PDG \cite{pdg} were used. Notice that the PDG value $(350\pm150)$ MeV refers to the full decay width of
$f_{0}(1370).$ We are thus making the simplifying assumption that the $\pi\pi
$-decay mode is dominant (see for instance the recent study in Ref. \cite{stanislaus} and
references therein). This choice represents an upper limit: for smaller $\pi\pi$
branching ratios, the effect of the triangle diagram will be smaller.
The calculation for the resonance $f_{0}(500)$ was done twice: once with a cutoff
$\Lambda\sim10^{5}/2$ GeV (\emph{i.e.}, the practical limit $\Lambda\rightarrow\infty$) and
once with a physical cutoff of $\Lambda\sim1$ GeV in order to demonstrate how the
outcome is influenced by a cutoff value which is typical in hadron physics. As can be seen in Fig. \ref{fig:fig4}, the NLO correction to the decay width is only
important for the resonance $f_{0}(500)$, for which (i) the mass of the decay
products is comparable to $m_{f_{0}(500)}$ \emph{and} (ii) $\Gamma_{f_{0}(500)}^{\text{tree}}\sim
m_{f_{0}(500)}$: in such a case $\Gamma_{\text{NLO}}\sim2\hspace{0.03cm}\Gamma
_{\text{LO}}$, thus the value of the coupling $g$ corresponds roughly to $g^{\prime\prime}$. However, it should be stressed that the NLO process does not include all the other $\pi\pi$-scattering contributions. In the low-energy regime in which $f_{0}(500)$ appears, those contributions interfere and, in virtue of chiral symmetry (see \emph{e.g.} Ref. \cite{garcia}), the full NLO result is expected to be smaller than what our simple model suggests.
For $f_{0}(980)$, $f_{0}(1370)$ and $f_{0}(1500)$ the NLO differs in a small amount with respect to the full result. (Since $f_{0}(1370)$ is quite broad, the NLO correction gives a small but non-negligible contribution to the width. This is so because the ratio $\Gamma_{f_{0}(1370)}^{\text{tree}}/m_{f_{0}(1370)}$ is already in the range of $\sim0.1$. However, the large error bars do not allow to distinguish the NL and the NLO decay widths.) We also observe that a finite cutoff parameter $\Lambda
\sim1$ GeV would affect the decays of $f_{0}(980)$,
$f_{0}(1370)$ and $f_{0}(1500)$ only marginally, since the outcome curve would lie somewhere in between $\Gamma_{\text{LO}}$ and $\Gamma_{\text{NLO}}$.
Furthermore, we show in Fig. \ref{fig:fig5} analogous plots for $f_{0}(1710)$ going into pions and kaons, and for $a_{0}(1450)$ going into kaons. The experimental value
of the width (which is $\Gamma_{f_{0}\rightarrow\pi\pi}/3$ and $\Gamma_{f_{0}/a_{0}\rightarrow KK}/2$, respectively) is marked on the $y$-axis and the error bars are again
indicated by a gray band. Here, data was taken from the PDG only \cite{pdg}. By looking at the plots it becomes pretty clear that the NLO correction has only a small influence. Notice that there is a subtle difference in the calculation of the kaon-kaon decay due to the fact that the kaons are distinguishable particles: the factor of $2$ in the LO expression of the amplitude in Eq. (\ref{eq:tree_amplitude_phiphi}), the factor of 8 in the NLO expression in Eq. (\ref{eq:3amplitude}), as well as the symmetry factor $1/2$ in Eq. (\ref{eq:tree_width_phiphi}) are replaced by the unity.
As a next step, we check if the coupling constants obtained in Fig. \ref{fig:fig4} and Fig. \ref{fig:fig5} are compatible with the ones chiral approaches deliver. In Ref. \cite{dick} the resonances $f_{0}(1370)$, $f_{0}(1500)$ and $a_{0}(1450)$ were studied. The various terms in the amplitude for a given channel (\emph{i.e.}, terms with and without derivatives) can be summarized in a unique effective coupling constant for that channel. One gets $\sim 1.7$ GeV for $f_{0}(1370)$ into pions, $\sim 0.7$ GeV for $f_{0}(1500)$ into pions, and $2.28$ GeV for $a_{0}(1450)$ into kaons. Also, the values of the effective couplings of $f_{0}(1710)$ were determined in Ref. \cite{stanislaus} to be $0.64$ GeV for pions and $1.88$ GeV for kaons.
All those values are well compatible with the values shown in Fig. \ref{fig:fig4} and Fig. \ref{fig:fig5}. Similar comments can be made concerning the resonances below $1$ GeV by regarding at the works done in Refs. \cite{giacosaSpectral,e38}. These considerations confirm also that the ranges of couplings studied in Fig. \ref{fig:fig3} were realistic.
\begin{figure}
\hspace{-3.0cm}\includegraphics[scale=0.9]{fig4_mod.pdf}
\caption{Decay widths of $f_{0}(500)$ (upper left), $f_{0}(980)$ (upper
right), $f_{0}(1370)$ (lower left) and $f_{0}(1500)$ (lower right). The
`known' value (for the two resonances below $1$ GeV taken from Ref.
\cite{garcia}, for the other two from the PDG \cite{pdg}) of the width is
marked on the $y$-axis and error bars are indicated by a gray band. Note that here $\Gamma=\Gamma_{f_{0}\rightarrow\pi^{0}\pi^{0}}=\Gamma_{f_{0}\rightarrow\pi\pi}/3$. The vertical lines correspond to the determination of the coupling constant at LO and NLO, respectively.}
\label{fig:fig4}
\end{figure}
\newpage
\begin{figure}
\hspace{-3.0cm}\includegraphics[scale=0.9]{fig5_mod.pdf}
\caption{Decay widths of $f_{0}(1710)$ going into pions (upper left) and into kaons (upper
right), and $a_{0}(1450)$ into kaons (lower middle). The
`known' value (taken from the PDG \cite{pdg}) of the width is
marked on the $y$-axis and error bars are indicated by a gray band. Note that here $\Gamma=\Gamma_{f_{0}\rightarrow\pi^{0}\pi^{0}}=\Gamma_{f_{0}\rightarrow\pi\pi}/3$ and $\Gamma=\Gamma_{f_{0}/a_{0}\rightarrow K^{+}K^{-}}=\Gamma_{f_{0}/a_{0}\rightarrow K^{0}\bar{K}^{0}}=\Gamma_{f_{0}/a_{0}\rightarrow KK}/2$. The vertical lines correspond to the determination of the coupling constant at LO and NLO, respectively.}
\label{fig:fig5}
\end{figure}
\clearpage
\section{Conclusions}
\label{seq:conc}
In this work we investigated the importance of the (usually neglected)
triangle-shaped NLO contribution to two-body hadronic decays of the form $S\rightarrow\phi\phi$, see Fig. \ref{fig:fig2}. To this end, the NLO diagram was calculated analytically and numerically in the framework of a quantum field theory involving scalar fields without derivatives (avoiding unnecessary complications due to spin and/or derivatives).
We studied different cases where the mass $m_{\phi}$ of the decay products
varied from nearly zero up to almost $m_{S}/2$, see Fig. \ref{fig:fig3}. The (finite)
contribution from the triangle diagram to the decay width turned out to be
negligible if the mass of the decaying particle is sufficiently large, a
condition which is usually met in hadron physics. We then moved to physical
decays into pions of the well-known scalar resonances $f_{0}(500)$,
$f_{0}(980)$, $f_{0}(1370)$, $f_{0}(1500)$ and $f_{0}(1710)$, as well as to the kaonic channels of $f_{0}(1710)$ and of $a_{0}(1450)$. The outcome was, too, very
clear, see Fig. \ref{fig:fig4} and Fig. \ref{fig:fig5}: the NLO correction gave only a relevant contribution for
$f_{0}(500)$, obviously because $\Gamma_{f_{0}(500)}^{\text{tree}}\sim m_{f_{0}(500)}$. The
resonance $f_{0}(500)$ is however an `extreme example': in all other cases the
correction is much smaller.
\begin{figure}[h]
\centering
\hspace{0.3cm}
\includegraphics[scale=0.71]{fig6.pdf}
\caption{NLO correction to the self-energy of the state $S$.}
\label{fig:fig6}
\end{figure}
Within the perturbative framework we have studied in this work, the unstable
state $S$ exchanged in the triangle diagram (Fig. \ref{fig:fig2}) has been considered as stable (\emph{i.e.}, the free propagator was used). Indeed, when considering the fact
that $S$ has a finite width, the contribution from the triangle diagram would be
even smaller. However, the correct way of going beyond the present study is the following:
Besides the one-loop diagram in Fig. \ref{fig:fig1}b, one should perform the resummation of the self-energy of the unstable state $S$ also by incorporating the NLO correction depicted in Fig. \ref{fig:fig6}. Such a study is certainly
nontrivial because the full propagator of $S$ enters here -- one is left with a typical problem of the Bethe--Salpeter type,
see \emph{e.g.} Ref. \cite{alkofer} and references therein. Quite interestingly, the results of
our work show that the modifications coming from such a computation are likely
to be in most cases negligible. Thus, the tree-level results or at most a
description using the (resummed) one-loop propagator of an unstable state, give(s) a
good description of unstable hadronic states.
Other future studies are possible by considering different forms of the
Lagrangian, including derivative interactions, particles with higher spin,
fermionic fields (\emph{i.e.}, baryons), three-body decays and unstable states which
decay in more than one channel.
\section*{Acknowledgements}
The authors thank G. Pagliara, D. H. Rischke and J. Wambach for useful discussions. T. W.
acknowledges financial support from HGS-HIRe, F\&E GSI/GU and HIC for FAIR Frankfurt.
\bigskip
\bigskip
{\small \bigskip}
|
\section{Introduction and statement of the main results}
As is known, in the study of the qualitative theory of real planar differential systems, one of the important
open problems is the determination of limit cycles. The second part of the famous Hilbert's 16th problem, proposed in 1900, asks for an upper bound on Hilbert number $H(n)$ and position of limit cycles for all planar polynomial differential systems of degree $n$, but it is still open even for $n=2$. It is so difficult that the weak form of this problem has been introduced. A classical way to obtain limit cycles is that perturbing the periodic orbits of a center. Let us consider the planar polynomial vector fields $X_{\varepsilon}=X_{0}+\varepsilon Y$, where $0<\varepsilon\ll1$ and $X_{0}=(-H_{y}/R,H_{x}/R)$ is a polynomial vector field having a continuum of periodic orbits. $H$ is a first integral of $X_{0}$ and $R$ is an integrating factor. If $R=1$, we call $X_{0}$ a Hamiltonian vector field, otherwise, we say that it's a non-Hamiltonian integrable vector field. In order to study the periodic orbits of $X_{\varepsilon}$ that remain among all the periodic orbits of $X_{0}$, it is necessary to study the number of zeros of an (generalized) Abelian integral, also known as the first order Melnikov function, i.e.,
\begin{equation}
M(h)=\displaystyle\oint_{H=h}R\left(Y_{1}(x,y)\mathrm{d}y-Y_{2}(x,y)\mathrm{d}x\right),
\end{equation}
where $\{H=h,h\in(h_{0},h_{1})\}$ are periodic orbits of $X_{0}$, and $Y_{i}(x,y),i=1,2$, are the two components of $Y$.
To the best of our knowledge, many authors have investigated the limit cycles for the quadratic Hamiltonian systems and non-Hamiltonian integrable systems under polynomial perturbations (e.g., \cite{CLY,CLP,GGJ,HI,I,LLLZ,LLLZ1,L,Z} and references therein). However, the studies on cubic and higher degree systems are relatively few (e.g., \cite{AZ,BL1,LLLZ,LLLZ2,LLM,YH,ZZ}).
In this paper, we concern with the number of zeros of Abelian integrals for a class of cubic non-Hamiltonian integrable systems of Lotka-Volterra type with a rational first integral of degree 2, the integrating factor of which consists of $x$ and $y$. In general, it is difficult to study the bifurcation of limit cycles for a non-Hamiltonian integrable systems with integrating factors including both the variables $x$ and $y$, thus the technique and the results are few (e.g., \cite{AZ,BL1,LLLZ2}).
The authors \cite{CL} first studied the planar cubic polynomial
vector fields of Lotka-Volterra type with a rational first integral
of degree 2 and got 28 non-topologically equivalent phase portraits,
among which there are only 2 cases having at least one center in the finite plane,
as shown in Figure 1 (see Figure 1 of \cite{CL}).
\begin{figure}[h]
\centering \epsfig{file=fig.eps,totalheight=50mm,width=95mm,angle=0}
\setlength{\abovecaptionskip}{0pt
\renewcommand{\figurename}{Figure}
\caption{All the non-topologically equivalent phase portraits with at least one center
of a planar cubic polynomial vector field of Lotka-Volterra type with a rational first
integral of degree 2.} \centering
\label{Fig.1}
\end{figure}
The system with a center in \cite{CL} is a differential system of the form
\begin{equation}\begin{array}{ll}\label{eq11}
\dot{x}=x(1+bx+x^2-y^2),\\[2ex]
\dot{y}=y(-1-cy+x^2-y^2),
\end{array}\end{equation}
which has a first integral
\begin{equation*}\vspace{.05in}
H=\dfrac{1+bx+cy+x^2+y^2}{xy}
\end{equation*}
with integrating factor $x^{-2}y^{-2}$.
When $0\leq b<2$, $0\leq
c<2$ and $b\neq c$, the phase portrait of system \eqref{eq11} corresponds to
$X_{2.9}$ in Figure 1; when $0<b=c<2$, the phase portrait of system \eqref{eq11}
is $X_{2.10}$ in Figure 1. For convenience, we will also denote the two subclasses of system \eqref{eq11}
by $X_{2.9}$ and $X_{2.10}$, respectively.
System $X_{2.9}$ has three finite singular points: a
hyperbolic saddle at $O (0,0)$ and two centers at $C_{\pm} (\delta(-b\delta\pm
c\gamma)/2(\delta^2-\gamma^2),\gamma(-c\gamma\pm
b\delta)/2(\gamma^2-\delta^2))$, where
$\gamma=\sqrt{4-b^2}, \delta=\sqrt{4-c^2}$. $C_{+}$ is located in the third
quadrant while $C_{-}$ is located in the fourth (resp. second) quadrant if $b<c$ (resp. $b>c$). If $b>c$, taking a change
$(x,y,t,b,c)\rightarrow (y,x,-t,c,b)$, system $X_{2.9}$ is
reduced to that in the case $b<c$. Thus without loss of generality, it suffices to
consider the latter case. $H=h_{\pm}=(-bc\pm\gamma\delta)/2$ correspond to
$C_{\pm}$. There are two families of periodic orbits
$\Gamma_{h}:H=h$, $h\in(-2,h_{-})\cup(h_{+},2)$,
which surround the centers $C_{-}$ and $C_{+}$, respectively.
System $X_{2.10}$ has two finite singular points. $O (0,0)$ is a
hyperbolic saddle and $C (-1/b,-1/b)$ is a center, which is located in the
third quadrant. There is a family of periodic orbits
$\Gamma_{h}:H=h$, $h\in(2-b^2,2)$, which surrounds the center $C$. $H=2-b^2$
is the value of the center.
In what follows we are going to study the polynomial perturbations of these two subclasses of system \eqref{eq11}:
\begin{equation}\begin{array}{ll}\label{eq12}
\dot{x}=x(1+bx+x^2-y^2)+\varepsilon f(x,y),\\[2ex]
\dot{y}=y(-1-cy+x^2-y^2)+\varepsilon g(x,y),
\end{array}\end{equation}
where $f(x,y)$ and $g(x,y)$ are polynomials in the variables $x$ and
$y$ with $\mbox{max}\{\deg f$, $\deg g\}=n$.
Consider \eqref{eq12} with $0<\varepsilon\ll1$. Let $H_{X_{2.9}}(n)$ and $H_{X_{2.10}}(n)$ be the maximum number of zeros (taking into account their
multiplicity) of the Abelian integrals associated with the two systems:
\[
\begin{array}{ll} \vspace{.1in}
X_{2.9}, & \hbox{i.e., $0\leq b<c<2$,}\\
X_{2.10},& \hbox{i.e., $0<b=c<2$, }
\end{array}
\]
respectively. Note that system $X_{2.9}$ has two period annuli and here $H_{X_{2.9}}(n)$ denotes the maximum number of zeros of Abelian integral $M(h)$ on
the interval $(-2,h-)$ or $(h_+,2)$.
The main results of this paper are:
\begin{theorem}\label{th1}
$H_{X_{2.9}}(n)\leq1$ if $n\leq 2$ and $H_{X_{2.9}}(n)\leq2n-3$ if $n\geq 3$.
\end{theorem}
\begin{theorem}\label{th2}
$H_{X_{2.10}}(n)\leq1$ if $n\leq 2$ and $H_{X_{2.10}}(n)\leq[(3n-3)/2]$ if $n\geq 3$.
\end{theorem}
\begin{remark}
By Proposition 1 of \cite{LLLZ2}, it is easy to
verify that the systems $X_{2.9}$ and $X_{2.10}$ can be transformed into $S^{*}$:
\begin{equation*}\begin{array}{ll}
\dot{x}=-y+\beta x^2-2\alpha xy-\beta y^2+x^2y,\\[2ex]
\dot{y}=x+\alpha x^2+2\beta xy-\alpha y^2+xy^2,
\end{array}\end{equation*}
in Table I of \cite{LLLZ2}. Thus the centers of these two systems are isochronous.
Moreover, it follows from the first integral that system $X_{2.10}$ is
reversible (with respect to the straight line $y=x$) while system $X_{2.9}$ is not. In fact, by the affine
transformation
\[(x,y,t)\rightarrow\left(y-x,\dfrac{b(x+y)+2}{\sqrt{4-b^2}},-\dfrac{\sqrt{4-b^2}}{b}t\right),
\]
system $X_{2.10}$ is transformed into $S^{*}$ with
$\alpha=0$, $\beta=-\sqrt{4-b^2}/2$, which belongs to the reversible case (A) of \cite{LLLZ2}.
As a special one, our result
is better than that of the general case (see Theorem 2 of \cite{LLLZ2}).
Similarly, by the affine
transformation
\[(x,y,t)\rightarrow\left(\dfrac{x+y-\xi-\eta}{\sqrt{1-(\xi+\eta)^2}},\dfrac{y-x+\xi-\eta}{\sqrt{(\xi-\eta)^2-1}},-\sqrt{4\xi^2\eta^2-(\xi^2+\eta^2-1)^2}t\right),
\]
system $X_{2.9}$ is reduced to $S^{*}$ with
$\alpha=(\xi+\eta)\sqrt{1-(\xi+\eta)^2}/(4\xi\eta)$,
$\beta=(\xi-\eta)\sqrt{(\xi-\eta)^2-1}/(4\xi\eta)$, where
$\xi=-\delta\zeta$, $\eta=\gamma\zeta$,
$\zeta=(b\delta+c\gamma)/2(\delta^2-\gamma^2)$. Since $\alpha\beta\neq
0$, system $X_{2.9}$ is not reversible. The problem that how many limit cycles bifurcate from the period annuli
of system $X_{2.9}$ has not yet been studied.
\end{remark}
Different from the classical methods which are used to study the number of
limit cycles that bifurcate from the periodic orbits of a center,
for example Poincar\'e-Melnikov integral method, Picard-Fuchs equation method \cite{HI,ZZ}, inverse integrating
factor method \cite{GLV}, averaging method \cite{BL1} and so on, this paper takes
advantage of some symmetric properties of the first integral and
integrating factor to give the number of zeros of the Abelian integrals by direct
computation. Compared with the method using Picard-Fuchs
equation, our method gives the better upper bound. The computational approach has similarities with the technique used in \cite{LLLZ1,LLLZ2},
but we do not use Green's theorem which makes the work
harder than computing the usual single Abelian integral directly in this paper.
Recently, Llibre {\it et al.} also study limit cycles of cubic polynomial differential systems with rational first integrals of degree 2 under polynomial perturbations of degree $3$ in \cite{LLM} using averaging method. They give six families of cubic polynomial differential systems, denote by $P_k$ for $k=1,2,\ldots,6$, where $P_3$ and $P_5$ are equivalent to ours. However, they only give the exact upper bound for $P_1,P_2,P_4$ and $P_6$, and give an example of class $P_5$ that has at most 3 limit cycles.
We investigate the sharp upper bound for the number of zeros of the Abelian integrals
with respect to the two systems for $n=3$. The result is the following one.
\begin{theorem}\label{th3}
$H_{X_{2.9}}(3)=3$ and $H_{X_{2.10}}(3)=3$.
\end{theorem}
As is introduced above, there are two families of periodic orbits of system $X_{2.9}$.
It is natural to consider the simultaneous bifurcation and distribution of
limit cycles that emerge from both period annuli (e.g., \cite{CGP,GGJ}).
The configuration of limit cycles $(u,v)$, $u\geq0, v\geq0$, is considered to be achievable
if, for $\varepsilon$ small enough, exactly $u$ (resp. $v$) limit cycles bifurcate from the periodic orbits surrounding $C_{+}$ (resp. $C_{-}$).
Though, up to first order in $\varepsilon$, three limit cycles can emerge from each period annulus of system $X_{2.9}$ under cubic polynomial perturbations, it does not means that the configuration (3,3) can be achievable (see for instance \cite{CLP,L}). In the present paper, we give a positive answer, i.e., (3,3) is impossible.
\begin{theorem}\label{th4}
Under cubic polynomial perturbations, all configurations $(u,v)$ of limit cycles bifurcated from the two period annuli of system
$X_{2.9}$ can be realized, where $0\leq u, v\leq 3, u+v\leq5$.
\end{theorem}
The paper is organized as follows: in sections 2 and 3, we study the number of zeros
of the Abelian integrals for systems $X_{2.9}$ and $X_{2.10}$, respectively. Upper bounds for $H_{X_{2.9}}(n)$ and $H_{X_{2.10}}(n)$ are obtained. Section 4 focuses on the analysis of the least upper bound for the number of zeros of the Abelian integrals associated to the two systems for $n=3$. \emph{Chebyshev criterion} is used to determine $H_{X_{2.9}}(3)$ and $H_{X_{2.10}}(3)$. Finally, we give the simultaneous bifurcation and distribution of limit cycles bifurcated from the two period annuli of system
$X_{2.9}$ under cubic polynomial perturbations in section 5.
\section{Zeros of the Abelian integral for system $X_{2.9}$}
In this section, we will study $H_{X_{2.9}}(n)$.
Let $\Gamma_{h}$ be the closed component of the algebraic
curve
\[
H(x,y,b,c)=\dfrac{1+bx+cy+x^2+y^2}{xy}=h
\]
and $\Gamma_{h}^{'}$ be the closed component
of the algebraic curve
\[
H(x,y,c,b)=\dfrac{1+cx+by+x^2+y^2}{xy}=h.
\]
The Abelian integral associated to system $X_{2.9}$ is defined as
\begin{equation}\begin{split}\label{Mh:29}
M(h)&=\displaystyle\oint_{\Gamma_{h}}x^{-2}y^{-2}(f\mathrm{d}y-g\mathrm{d}x)\\
&=\displaystyle\sum_{i+j=0}^{n}\oint_{\Gamma_{h}}(a_{ij}x^{i-2}y^{j-2}\mathrm{d}y-b_{ij}x^{i-2}y^{j-2}\mathrm{d}x), \quad h\in(-2,h_{-})\cup(h_{+},2),
\end{split}\end{equation}
where $\Gamma_{h}$ has the positive (resp. negative) orientation if it surrounds $C_{+}$ (resp. $C_{-}$).
Denote
\begin{equation}\label{Iij:29}
I_{ij}(h)=\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{j-2}\mathrm{d}x,\quad \bar{I}_{ij}(h)=\displaystyle\oint_{\Gamma_{h}^{'}}x^{i-2}y^{j-2}\mathrm{d}x.
\end{equation}
\begin{proposition}\label{Prop:29}
$M(h)$ has the following expression:
\begin{equation}
M(h)=\sum_{i+j=0}^{n}\left(-a_{ji}\bar{I}_{ij}(h)-b_{ij}I_{ij}(h)\right), \quad h\in(-2,h_{-})\cup(h_{+},2).
\end{equation}
\end{proposition}
\begin{proof}
Suppose the level curve $\Gamma_{h}$ is a periodic orbit that
surrounds the center $C_{+}$ of system $X_{2.9}$, then
\begin{equation*}\begin{split}
&\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{j-2}\mathrm{d}y\\
=&\displaystyle\int_{y_{1}}^{y_{2}}\left(\dfrac{hy-b}{2}+\sqrt{\Psi(y)
}\right)^{i-2}y^{j-2}\mathrm{d}y+\displaystyle\int_{y_{2}}^{y_{1}}\left(\dfrac{hy-b}{2}-\sqrt{\Psi(y)
}\right)^{i-2}y^{j-2}\mathrm{d}y,
\end{split}\end{equation*}
where $\Psi(y)=(h^{2}/4-1)y^{2}-(bh/2+c)y+b^{2}/4-1$ and $y_{1},
y_{2}$ denote the two roots of the equation $\Psi(y)=0$ with $y_{1}<y_{2}$.
Set $y=x$ in the right-hand-side of the last equality, then
\begin{equation}\begin{split}\label{1}
&\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{j-2}\mathrm{d}y\\[-3pt]
=&\displaystyle\int_{y_{1}}^{y_{2}}\left(\dfrac{hx-b}{2}+\sqrt{\Psi(x)
}\right)^{i-2}x^{j-2}\mathrm{d}x+\displaystyle\int_{y_{2}}^{y_{1}}\left(\dfrac{hx-b}{2}-\sqrt{\Psi(x)
}\right)^{i-2}x^{j-2}\mathrm{d}x\\[-3pt]
=&-\displaystyle\oint_{\Gamma_{h}^{'}}y^{i-2}x^{j-2}\mathrm{d}x=-\bar{I}_{ji}(h),
\end{split}\end{equation}
where $\Gamma_{h}^{'}$ has the same orientation as $\Gamma_{h}$. The second
equality holds since $y_{1}, y_{2}$ are also the roots of the equation $\Psi(x)=0$.
The proposition follows from \eqref{Mh:29}, \eqref{Iij:29} and \eqref{1}.
\end{proof}
By Proposition 6, owing to the fact that the form and algorithm of the two parts
of $M(h)$ are similar, we only need to calculate
\begin{equation}\label{M29}
\sum_{i+j=0}^{n}b_{ij}I_{ij}(h)=M_{1}(h)+M_{2}(h)+M_{3}(h), \quad h\in(-2,h_{-})\cup(h_{+},2),
\end{equation}
where
\begin{equation*}\begin{split}
M_{1}(h)&=\displaystyle\sum_{j\geq2,\atop i+j\leq n}b_{ij}I_{ij}(h)=\sum_{j\geq2,\atop i+j\leq n}b_{ij}\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{j-2}\mathrm{d}x,\\[-3pt]
M_{2}(h)&=\displaystyle\sum^{n-1}_{i=0}b_{i1}I_{i1}(h)=\sum_{i=0}^{n-1}b_{i1}\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{-1}\mathrm{d}x,\\[-3pt]
M_{3}(h)&=\displaystyle\sum^{n}_{i=0}b_{i0}I_{i0}(h)=\sum_{i=0}^{n}b_{i0}\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{-2}\mathrm{d}x.
\end{split}\end{equation*}
Let $\Delta=(h^{2}/4-1)x^{2}-(ch/2+b)x+c^{2}/4-1$ and $x_{1}, x_{2}$
are the two roots of the equation $\Delta=0$ with $x_{1}<
x_{2}$. Then
\begin{equation}\label{Viete1}
x_{1}+x_{2}=-\dfrac{2(2b+ch)}{4-h^{2}},\ \ \
x_{1}x_{2}=\dfrac{4-c^{2}}{4-h^{2}}.
\end{equation}
We have that
\begin{equation}\label{sign}
I_{ij}(h)=\displaystyle\pm\int_{x_{1}}^{x_{2}}x^{i-2}\left[\left(\dfrac{hx-c}{2}-\sqrt{\Delta
}\right)^{j-2}-\left(\dfrac{hx-c}{2}+\sqrt{\Delta }\right)^{j-2}\right]\mathrm{d}x.
\end{equation}
\begin{remark}
The appearance of ``$\pm$'' in \eqref{sign} comes from the fact that $\Gamma_{h}$ has the positive orientation when $h\in(h_+,2)$ and the negative orientation when $h\in(-2,h_-)$. However, it is easy to verify that the sign ``$\pm$'' does not affect the maximum number of zeros of the Abelian integral defined on each interval. Thus we will drop it in the computation of $M_{i}(h),i=1,2,3$ for $h\in(-2,h_{-})\cup(h_{+},2)$. The results obtained differ from the original functions at most by a negative sign.
\end{remark}
First, we get that
\begin{equation}\begin{split}\label{M1}
M_{1}(h)&=\displaystyle\sum_{j\geq2, \atop i+j\leq n} b_{ij}\int_{x_{1}}^{x_{2}}x^{i-2}\left[\left(\dfrac{hx-c}{2}-\sqrt{\Delta
}\right)^{j-2}-\left(\dfrac{hx-c}{2}+\sqrt{\Delta }\right)^{j-2}\right]\mathrm{d}x\\[-10pt]
&=\displaystyle\sum_{j\geq3, \atop i+j\leq n} b_{ij}\int_{x_{1}}^{x_{2}}x^{i-2}\left(\sum_{k=0}^{j-3}\tilde{m}_{j,k}(h)x^{k}\sqrt{\Delta}\right)\mathrm{d}x\\[-4pt]
&=\displaystyle\sum_{k=0}^{n-3}m_{k}(h)J_{k}(h),
\end{split}\end{equation}
where $\tilde{m}_{j,k}(h),m_{k}(h)$ are polynomials of $h$ with
$\deg \tilde{m}_{j,k}(h)\leq{k},\deg m_{k}(h)\leq{k}$, and
\begin{equation*}
J_{k}(h)=\displaystyle\int_{x_{1}}^{x_{2}}x^{k-2}\sqrt{\Delta}\mathrm{d}x
=\sqrt{1-h^{2}/4}\int_{x_{1}}^{x_{2}}x^{k-2}\sqrt{(x_{2}-x)(x-x_{1})}\mathrm{d}x.
\end{equation*}
In order to study $J_{k}(h)$, we use two different transformations.
If $\sqrt{(x_{2}-x)(x-x_{1})}=t(x-x_{1})$, then
\begin{equation*}
J_{k}(h)=\displaystyle\sqrt{4-h^{2}}\
(x_{2}-x_{1})^{2}\int_{0}^{\infty}\dfrac{t^{2}(x_{2}+t^{2}x_{1})^{k-2}}{(1+t^{2})^{k+1}}\mathrm{d}t,
\end{equation*}
and if $\sqrt{(x_{2}-x)(x-x_{1})}=t(x_2-x)$, then
\begin{equation*}
J_{k}(h)=\displaystyle\sqrt{4-h^{2}}\ (x_{2}-x_{1})^{2}\int_{0}^{\infty}\dfrac{t^{2}(x_{1}+t^{2}x_{2})^{k-2}}{(1+t^{2})^{k+1}}\mathrm{d}t.
\end{equation*}
Thus, by \eqref{Viete1}, for $k\geq2$,
\begin{equation}\begin{split}\label{Jk}
J_{k}(h)&=\displaystyle\sqrt{4-h^{2}}\ (x_{2}-x_{1})^{2}\int_{0}^{\infty}\dfrac{t^{2}[(x_{2}+t^{2}x_{1})^{k-2}+(x_{1}+t^{2}x_{2})^{k-2}]}{2(1+t^{2})^{k+1}}\mathrm{d}t\\
&=\displaystyle\sqrt{4-h^{2}}\
(x_{2}-x_{1})^{2}\sum_{i+2j=k-2}d_{ij}(x_{1}+x_{2})^{i}(x_{1}x_{2})^{j}\\
&=\displaystyle\dfrac{16(h^{2}+bch+b^{2}+c^{2}-4)}{(4-h^{2})^{3/2}}
\sum_{i+2j=k-2}d_{ij}\left[-\dfrac{2(2b+ch)}{4-h^{2}}\right]^{i}\left(\dfrac{4-c^{2}}{4-h^{2}}\right)^{j}\\
&=(4-h^{2})^{-(k-1/2)}P_{k}(h),
\end{split}\end{equation}
where $d_{ij}, i+2j=k-2$ are constants and $P_{k}(h)$ denotes a polynomial of degree $k$. By direct computations, we get that
\begin{equation}\begin{array}{ll}\label{J1J0}
J_{1}(h)=\displaystyle\int_{x_{1}}^{x_{2}}x^{-1}\sqrt{\Delta}\mathrm{d}x
=\left\{\begin{array}{ll}
-\displaystyle\dfrac{2b+ch}{2\sqrt{4-h^{2}}}\pi-\dfrac{\sqrt{4-c^{2}}}{2}\pi, & \mbox{$h\in(-2,h_{-})$,}\\[2ex]
-\displaystyle\dfrac{2b+ch}{2\sqrt{4-h^{2}}}\pi+\dfrac{\sqrt{4-c^{2}}}{2}\pi, & \mbox{$h\in(h_{+},2)$,}\\[2ex]
\end{array} \right.\\[2ex]
J_{0}(h)=\displaystyle\int_{x_{1}}^{x_{2}}x^{-2}\sqrt{\Delta}\mathrm{d}x
=\left\{\begin{array}{ll}
-\displaystyle\dfrac{2b+ch}{2\sqrt{4-c^{2}}}\pi-\dfrac{\sqrt{4-h^{2}}}{2}\pi, & \mbox{$h\in(-2,h_{-})$,}\\[2ex]
\displaystyle\dfrac{2b+ch}{2\sqrt{4-c^{2}}}\pi-\dfrac{\sqrt{4-h^{2}}}{2}\pi, & \mbox{$h\in(h_{+},2)$.}\\[2ex]
\end{array} \right.
\end{array}\end{equation}
It follows from \eqref{M1}, \eqref{Jk} and \eqref{J1J0} that
\begin{equation}\begin{split}\label{M1_9}
M_{1}(h)&=m_{0}(h)J_{0}(h)+m_{1}(h)J_{1}(h)+\displaystyle\sum_{k=2}^{n-3}m_{k}(h)J_{k}(h)\\[-6pt]
&=\hat{P}_{1}(h)+\displaystyle\dfrac{\hat{P}_{2n-6}(h)}{(4-h^{2})^{n-7/2}},
\end{split}\end{equation}
where $\hat{P}_{1}(h)$ and $\hat{P}_{2n-6}(h)$ are
polynomials with
$\deg \hat{P}_{1}(h)\leq{1}$ and $\deg\hat{P}_{2n-6}(h)\leq2n-6$, respectively.
Next we calculate $M_{2}(h)$ and $M_{3}(h)$.
\begin{equation}\begin{split}\label{M2}
M_{2}(h)&=\displaystyle\sum^{n-1}_{i=0} b_{i1}\int_{x_{1}}^{x_{2}}x^{i-2}\left[\dfrac{1}{(hx-c)/2-\sqrt{\Delta
}}-\dfrac{1}{(hx-c)/2+\sqrt{\Delta}}\right]\mathrm{d}x\\[-3pt]
&=\displaystyle \sum^{n-1}_{i=0} b_{i1}\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}\sqrt{\Delta
}}{x^{2}+bx+1}\mathrm{d}x\\[-3pt]
&=\displaystyle \sum^{3}_{i=0}\tilde{b}_{i1}S_{i}(h)+\sum^{n-1}_{i=4}\tilde{b}_{i1}\int_{x_{1}}^{x_{2}}x^{i-4}\sqrt{\Delta
}\mathrm{d}x\\[-3pt]
&=\displaystyle\sum^{3}_{i=0}\tilde{b}_{i1}S_{i}(h)+ \sum^{n-1}_{i=4}\tilde{b}_{i1}J_{i-2}(h),\\[-3pt]
\end{split}\end{equation}
where
$\tilde{b}_{i1},i=0,1,...,n-1,$ are linear combinations of $b_{i1},i=0,1,...,n-1,$ and
\begin{equation*}
S_{i}(h)=\displaystyle\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}\sqrt{\Delta
}}{x^{2}+bx+1}\mathrm{d}x, \quad i=0,1,2,3.
\end{equation*}
If $h\in(-2,h_{-})$, then
\begin{equation}\begin{array}{ll}\label{Si-}
S_{0}(h)=\dfrac{(b^{2}-2)c+bh}{\sqrt{4-b^{2}}}\pi-\dfrac{(c^{2}-2)b+ch}{\sqrt{4-c^{2}}}\pi,\\
S_{1}(h)=-\dfrac{bc+2h}{\sqrt{4-b^{2}}}\pi-\sqrt{4-c^{2}}\pi,\\
S_{2}(h)=\dfrac{2c+bh}{\sqrt{4-b^{2}}}\pi-\sqrt{4-h^{2}}\pi,\\
S_{3}(h)=-\dfrac{bc+(b^{2}-2)h}{\sqrt{4-b^{2}}}\pi-\dfrac{-2b+ch+bh^{2}}{\sqrt{4-h^{2}}}\pi,
\end{array}\end{equation}
and if $h\in(h_{+},2)$, then
\begin{equation}\begin{array}{ll}\label{Si+}
S_{0}(h)=\dfrac{(b^{2}-2)c+bh}{\sqrt{4-b^{2}}}\pi+\dfrac{(c^{2}-2)b+ch}{\sqrt{4-c^{2}}}\pi,\\
S_{1}(h)=-\dfrac{bc+2h}{\sqrt{4-b^{2}}}\pi+\sqrt{4-c^{2}}\pi,
\end{array}\end{equation}
$S_{2}(h)$ and $S_{3}(h)$ are the same as \eqref{Si-}.
By \eqref{Jk}, \eqref{M2}, \eqref{Si-} and \eqref{Si+}, we obtain
\begin{equation}\begin{split}\label{M2_9}
M_{2}(h)
=\tilde{P}_{1}(h)+\displaystyle\dfrac{\tilde{P}_{2n-6}(h)}{(4-h^{2})^{n-7/2}},\\
\end{split}\end{equation}
where $\tilde{P}_{1}(h)$ and $\tilde{P}_{2n-6}(h)$ are
polynomials with
$\deg \tilde{P}_{1}(h)\leq{1}$ and $\deg\tilde{P}_{2n-6}(h)\leq2n-6$, respectively.
The technique used in the computation of $M_{3}(h)$ is shown as follows:
\begin{equation}\begin{split}\label{M3}
M_{3}(h)&=\displaystyle\sum^{n}_{i=0}b_{i0}\int_{x_{1}}^{x_{2}}x^{i-2}\left[\dfrac{1}{((hx-c)/2-\sqrt{\Delta
})^2}-\dfrac{1}{((hx-c)/2+\sqrt{\Delta})^2}\right]\mathrm{d}x\\[-3pt]
&=\displaystyle\sum^{n}_{i=0}b_{i0}\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}(hx-c)\sqrt{\Delta
}}{(x^{2}+bx+1)^2}\mathrm{d}x\\[-3pt]
&=\displaystyle\sum^{3}_{i=0}b_{i0}R_{i}(h)+\sum^{n}_{i=4}b_{i0}\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}(hx-c)\sqrt{\Delta
}}{(x^{2}+bx+1)^2}\mathrm{d}x\\[-3pt]
&=\mu(h)S_{2}(h)+\nu(h)S_{3}(h)+\displaystyle\sum^{3}_{i=0}\bar{b}_{i0}R_{i}(h)+\sum^{n}_{i=5}\bar{b}_{i0}\omega_{i}(h)J_{i-3}(h),
\end{split}\end{equation}
where
$\bar{b}_{i0}, i=0,1,2,3,5,...,n,$ are
linear combinations of $b_{i0}, i=0,1,...,n$. $\mu(h)$, $\nu(h)$ and
$\omega_{i}(h)$ are polynomials of degree at most $1$, and
\begin{equation*}
R_{i}(h)=\displaystyle\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}(hx-c)\sqrt{\Delta
}}{(x^{2}+bx+1)^2}\mathrm{d}x, \quad i=0,1,2,3.
\end{equation*}
Further calculations show that
when $h\in(-2,h_{-})$,
\begin{equation}\begin{split}\label{Ri-}
R_{0}(h)=&\dfrac{2(-8+6c^2-6b^2(-1+c^2)+b^4(-1+c^2))\pi}{(4-b^2)^{3/2}}+\dfrac{2bc(-3+c^2)\pi}{\sqrt{4-c^2}}\\
&+h\left(\dfrac{2b(-6+b^2)c\pi}{(4-b^2)^{3/2}}+\dfrac{2(-2+c^2)\pi}{\sqrt{4-c^2}}\right)-\dfrac{4h^2\pi}{(4-b^2)^{3/2}},\\
R_{1}(h)=&\dfrac{2bc^2\pi}{(4-b^2)^{3/2}}+\dfrac{b(c^2-2)\pi}{\sqrt{4-b^2}}+c\sqrt{4-c^2}\pi+\dfrac{8ch\pi}{(4-b^2)^{3/2}}+\dfrac{2bh^2\pi}{(4-b^2)^{3/2}},\\
R_{2}(h)=&\dfrac{-4(-4+b^2+c^2+bch+h^2)\pi}{(4-b^2)^{3/2}},\\
R_{3}(h)=&\dfrac{2b(-4+b^2+c^2)\pi}{(4-b^2)^{3/2}}+\dfrac{8ch\pi}{(4-b^2)^{3/2}}
-\dfrac{b(-6+b^2)h^2\pi}{(4-b^2)^{3/2}}-h\sqrt{4-h^2}\pi,
\end{split}\end{equation}
and when $h\in(h_{+},2)$,
\begin{equation}\begin{split}\label{Ri+}
R_{0}(h)=&\dfrac{2(-8+6c^2-6b^2(-1+c^2)+b^4(-1+c^2))\pi}{(4-b^2)^{3/2}}-\dfrac{2bc(-3+c^2)\pi}{\sqrt{4-c^2}}\\
&+h\left(\dfrac{2b(-6+b^2)c\pi}{(4-b^2)^{3/2}}-\dfrac{2(-2+c^2)\pi}{\sqrt{4-c^2}}\right)-\dfrac{4h^2\pi}{(4-b^2)^{3/2}},\\
R_{1}(h)=&\dfrac{2bc^2\pi}{(4-b^2)^{3/2}}+\dfrac{b(c^2-2)\pi}{\sqrt{4-b^2}}-c\sqrt{4-c^2}\pi+\dfrac{8ch\pi}{(4-b^2)^{3/2}}+\dfrac{2bh^2\pi}{(4-b^2)^{3/2}},
\end{split}\end{equation}
$R_{2}(h)$ and $R_{3}(h)$ are the same as \eqref{Ri-}.
Hence by \eqref{Jk}, \eqref{Si-}, \eqref{Si+}, \eqref{M3}, \eqref{Ri-} and \eqref{Ri+}, we get
\begin{equation}\begin{split}\label{M3_9}
M_{3}(h)=\bar{P}_{2}(h)+\displaystyle\dfrac{{\bar{P}_{2n-5}(h)}}{(4-h^{2})^{n-7/2}},
\end{split}\end{equation}
where $\bar{P}_{2}(h)$ and $\bar{P}_{2n-5}(h)$ stand for polynomials of $h$ with
$\deg \bar{P}_{2}(h)\leq{2}$ and $\deg\bar{P}_{2n-5}(h)\leq2n-5$, respectively.
\noindent{\bfseries{Proof of Theorem 1.}} From \eqref{M29}, \eqref{M1_9}, \eqref{M2_9},
\eqref{M3_9} and Proposition 6, together with the similar form and algorithm of the two parts of
$M(h)$, we get the Abelian integral
\begin{equation}M(h)=\label{M}\left\{\begin{array}{ll}
Q_{2}(h)+\displaystyle\dfrac{{Q_{2n-5}(h)}}{(4-h^{2})^{n-7/2}} & \mbox{if $n\geq3$,}\\[2ex]
\widetilde{Q}_{2}(h) & \mbox{if $n\leq2$,}\\[2ex]
\end{array} \right.
\end{equation}
where $Q_{2}(h)$, $\widetilde{Q}_{2}(h)$ and $Q_{2n-5}(h)$ are polynomials with $\deg Q_{2}(h)\leq{2}$, $\deg \widetilde{Q}_{2}(h)\leq{2}$
and $\deg Q_{2n-5}(h)\leq2n-5$, respectively. We remark that $M(h)$ has the same form as \eqref{M} on $(-2,h_-)$ and $(h_+,2)$, respectively. But on these two intervals, they are different functions (it is clear to see from \eqref{J1J0}, \eqref{Si-}, \eqref{Si+} and so on).
Obviously, $M(h)$ is a polynomial of degree at most 2 if $n\leq2$ and it
has at most 2 zeros. Moreover, it vanishes at the value of the
center. Thus it has at most one zero in each open interval.
$M(h)$ does not have a rational form if $n\geq3$. But we can get
the estimation of the number of zeros of $M(h)$ by taking derivatives three
times. Therefore
\begin{equation*}
\#\{h|M^{(3)}(h)=0\}\leq 2n-5,
\end{equation*}
where $\#$ denotes the number of elements of a finite set.
On the other hand, note that $M(h_{\pm})=0$.
Hence
\begin{equation*}\begin{split}
&\mbox{max}\{\#\{h\in (-2,h_-)|M(h)=0\},\#\{h\in (h_+,2)|M(h)=0\}\}\\
\leq & 2n-5+3-1=2n-3,
\end{split}\end{equation*}
i.e.,
\begin{equation*}
H_{X_{2.9}}(n)\leq 2n-3.
\end{equation*}
\section{Zeros of the Abelian integral for system $X_{2.10}$}
This section is devoted to investigate the polynomial perturbations of system $X_{2.10}$, i.e., system (1.3) with $b=c$.
Firstly, we have the following result by the same argument as the proof of Proposition 6.
\begin{proposition}
The Abelian integral associated to the system
$X_{2.10}$ is
\begin{equation}\begin{split}
M(h)&=\displaystyle\oint_{\Gamma_{h}}x^{-2}y^{-2}(f\mathrm{d}y-g\mathrm{d}x)\\[-3pt]
&=\displaystyle\sum_{i+j=0}^{n}\oint_{\Gamma_{h}}(a_{ij}x^{i-2}y^{j-2}\mathrm{d}y-b_{ij}x^{i-2}y^{j-2}\mathrm{d}x)\\[-3pt]
&=\displaystyle\sum_{i+j=0}^{n}(-a_{ji}-b_{ij})I_{ij}(h), \quad h\in(2-b^2,2),
\end{split}\end{equation}
where $\Gamma_{h}$ is the closed component of the algebraic curve $1+bx+by+x^2+y^2=hxy$ with the positive orientation, and
\[
I_{ij}(h)=\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{j-2}\mathrm{d}x.
\].
\end{proposition}
Because the algorithm of $M(h)$ is the same as that in Section 2, we only give the main results.
\begin{equation}
M(h)=\sum_{i+j=0}^{n}c_{ij}I_{ij}(h)=M_{1}(h)+M_{2}(h)+M_{3}(h), \quad h\in(2-b^2,2),
\end{equation}
where $c_{ij}=-a_{ji}-b_{ij}$, and
\begin{equation*}\begin{split}
M_{1}(h)&=\displaystyle\sum_{j\geq2,\atop i+j\leq n}c_{ij}I_{ij}(h)=\sum_{j\geq2,\atop i+j\leq n}c_{ij}\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{j-2}\mathrm{d}x,\\[-3pt]
M_{2}(h)&=\displaystyle\sum^{n-1}_{i=0}c_{i1}I_{i1}(h)=\sum_{i=0}^{n-1}c_{i1}\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{-1}\mathrm{d}x,\\[-3pt]
M_{3}(h)&=\displaystyle\sum^{n}_{i=0}c_{i0}I_{i0}(h)=\sum_{i=0}^{n}c_{i0}\displaystyle\oint_{\Gamma_{h}}x^{i-2}y^{-2}\mathrm{d}x.
\end{split}\end{equation*}
Denote $\Delta=(h^{2}/4-1)x^{2}-(bh/2+b)x+b^{2}/4-1$ and let $x_{1}, x_{2}$
be the two roots of the equation $\Delta=0$ with $x_{1}<
x_{2}$. Then
\begin{equation}
x_{1}+x_{2}=-\dfrac{2b}{2-h},\ \ \
x_{1}x_{2}=\dfrac{4-b^{2}}{4-h^{2}}.
\end{equation}
Consequently,
\begin{equation}
M_{1}(h)=\displaystyle\sum_{k=0}^{n-3}m_{k}(h)J_{k}(h),
\end{equation}
where $m_{k}(h)$ is a polynomial of $h$ with
$\deg m_{k}(h)\leq{k}$ and
\begin{equation*}
J_{k}(h)=\displaystyle\int_{x_{1}}^{x_{2}}x^{k-2}\sqrt{\Delta}\mathrm{d}x.
\end{equation*}
It is easy to verify that if $k\geq2$,
\begin{equation}\begin{split}
J_{k}(h)&=\displaystyle\sqrt{4-h^{2}}\
(x_{2}-x_{1})^{2}\sum_{i+2j=k-2}\bar{d}_{ij}(x_{1}+x_{2})^{i}(x_{1}x_{2})^{j}\\
&=\displaystyle\dfrac{16(h-(2-b^2))}{(2-h)\sqrt{4-h^{2}}}
\sum_{i+2j=k-2}\bar{d}_{ij}\left(-\dfrac{2b}{2-h}\right)^{i}\left(\dfrac{4-b^{2}}{4-h^{2}}\right)^{j}\\
&=\displaystyle\dfrac{16(h-(2-b^2))}{\sqrt{4-h^{2}}(2-h)^{k-1}(2+h)^{[(k-2)/2]}}P_{[(k-2)/2]}(h),
\end{split}\end{equation}
where $\bar{d}_{ij}, i+2j=k-2$ are constants, $[(k-2)/2]$ denotes the integer part of $(k-2)/2$ and $P_{[(k-2)/2]}(h)$ is a polynomial of degree $[(k-2)/2]$.
A direct computation leads to
\begin{equation}\begin{array}{ll}
J_{1}(h)=-\displaystyle\dfrac{b(2+h)}{2\sqrt{4-h^{2}}}\pi+\dfrac{\sqrt{4-b^{2}}}{2}\pi,\\[2ex]
J_{0}(h)=\displaystyle\dfrac{b(2+h)}{2\sqrt{4-b^{2}}}\pi-\dfrac{\sqrt{4-h^{2}}}{2}\pi.
\end{array}\end{equation}
It follows from (3.4), (3.5) and (3.6) that
\begin{equation}
M_{1}(h)=\hat{P}_{1}(h)+\displaystyle\dfrac{\hat{P}_{[(3n-9)/2]}(h)}{\sqrt{4-h^{2}}(2-h)^{n-4}(2+h)^{[(n-5)/2]}},
\end{equation}
where $\hat{P}_{1}(h)$ and $\hat{P}_{[(3n-9)/2]}(h)$ are
polynomials with $\deg \hat{P}_{1}\leq{1}$ and $\deg\hat{P}_{[(3n-9)/2]}\leq[(3n-9)/2]$, respectively.
In addition,
\begin{equation}\begin{split}
M_{2}(h)&=\displaystyle\sum^{n-1}_{i=0}c_{i1}\int_{x_{1}}^{x_{2}}x^{i-2}\left(\dfrac{1}{(hx-b)/2-\sqrt{\Delta
}}-\dfrac{1}{(hx-b)/2+\sqrt{\Delta}}\right)\mathrm{d}x\\
&=\displaystyle\sum^{3}_{i=0}\tilde{c}_{i1}S_{i}(h)+\sum^{n-1}_{i=4}\tilde{c}_{i1}J_{i-2}(h),
\end{split}\end{equation}
where
$\tilde{c}_{i1}$, $i=0,1,...,n-1$ are
linear combinations of $c_{i1}$, $i=0,1,...,n-1$, and
\begin{equation*}
S_{i}(h)=\displaystyle\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}\sqrt{\Delta
}}{x^{2}+bx+1}\mathrm{d}x, \quad i=0,1,2,3.
\end{equation*}
By direct calculation we obtain that
\begin{equation}\begin{array}{ll}
S_{0}(h)=-bS_{1}(h)=\dfrac{2b(h-(2-b^2))}{\sqrt{4-b^{2}}}\pi,\\
S_{2}(h)=\dfrac{b(2+h)}{\sqrt{4-b^{2}}}\pi-\sqrt{4-h^{2}}\pi,\\
S_{3}(h)=-\dfrac{b^2+(b^{2}-2)h}{\sqrt{4-b^{2}}}\pi-\dfrac{b(-2+h+h^{2})}{\sqrt{4-h^{2}}}\pi.
\end{array}\end{equation}
Therefore, we get from (3.5), (3.8) and (3.9) that
\begin{equation}\begin{array}{ll}
M_{2}(h)=\tilde{P}_{1}(h)+\displaystyle\dfrac{\tilde{P}_{[(3n-9)/2]}(h)}{\sqrt{4-h^{2}}(2-h)^{n-4}(2+h)^{[(n-5)/2]}},
\end{array}\end{equation}
where $\tilde{P}_{1}(h)$ and $\tilde{P}_{[(3n-9)/2]}(h)$ are
polynomials with $\deg \tilde{P}_{1}\leq{1}$ and $\deg\tilde{P}_{[(3n-9)/2]}\leq[(3n-9)/2]$, respectively.
Finally, we have that
\begin{equation}\begin{split}
M_{3}(h)&=\displaystyle\sum^{n}_{i=0}c_{i0}\int_{x_{1}}^{x_{2}}x^{i-2}\left(\dfrac{1}{((hx-b)/2-\sqrt{\Delta
})^2}-\dfrac{1}{((hx-b)/2+\sqrt{\Delta})^2}\right)\mathrm{d}x\\[-3pt]
&=\displaystyle\sum^{n}_{i=0}c_{i0}\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}(hx-b)\sqrt{\Delta
}}{(x^{2}+bx+1)^2}\mathrm{d}x\\[-3pt]
&=\mu(h)S_{2}(h)+\nu(h)S_{3}(h)+\displaystyle\sum^{3}_{i=0}\bar{c}_{i0}R_{i}(h)+\sum^{n}_{i=5}\bar{c}_{i0}\omega_{i}(h)J_{i-3}(h),
\end{split}\end{equation}
where
$\bar{c}_{i0}, i=0,1,2,3,5,...,n,$ are
linear combinations of $c_{i0}, i=0,1,...,n$. $\mu(h)$, $\nu(h)$ and
$\omega_{i}(h)$ are polynomials of degree at most $1$, and
\begin{equation*}
R_{i}(h)=\displaystyle\int_{x_{1}}^{x_{2}}\dfrac{2x^{i-2}(hx-b)\sqrt{\Delta
}}{(x^{2}+bx+1)^2}\mathrm{d}x, \quad i=0,1,2,3.
\end{equation*}
A simple calculation shows that
\begin{equation}\begin{split}
R_{0}(h)&=-\dfrac{4(h-(2-b^2))(h-b^4+5b^2-2)\pi}{(4-b^2)^{3/2}},\\
R_{1}(h)&=\dfrac{2b(h-(2-b^2))(h-b^2+6)\pi}{(4-b^2)^{3/2}},\\
R_{2}(h)&=-\dfrac{4(h-(2-b^2))(h+2)\pi}{(4-b^2)^{3/2}},\\
R_{3}(h)&=-\dfrac{b(h+2)(b^2h-6h-2b^2+4)\pi}{(4-b^2)^{3/2}}-h\sqrt{4-h^2}\pi.
\end{split}\end{equation}
From (3.5), (3.9), (3.11) and (3.12), we obtain
\begin{equation}\begin{array}{ll}
M_{3}(h)=\bar{P}_{2}(h)+\displaystyle\dfrac{\bar{P}_{[(3n-7)/2]}(h)}{\sqrt{4-h^{2}}(2-h)^{n-4}(2+h)^{[(n-5)/2]}},
\end{array}\end{equation}
where $\bar{P}_{2}(h)$ and $\bar{P}_{[(3n-7)/2]}(h)$ are polynomials with $\deg \bar{P}_{2}\leq2$ and $\deg \bar{P}_{[(3n-7)/2]}\leq[(3n-7)/2]$, respectively.
\noindent{\bfseries{Proof of Theorem 2.}} It follows from Proposition 8, (3.2), (3.7), (3.10) and (3.13) that the Abelian integral is
\begin{equation}M(h)=\label{eq1}\left\{\begin{array}{ll}
Q_{2}(h)+\dfrac{Q_{[(3n-7)/2]}(h)}{\sqrt{4-h^{2}}(2-h)^{n-4}(2+h)^{[(n-5)/2]}} & \mbox{if $n\geq3$,}\\[2ex]
\widetilde{Q}_{2}(h) & \mbox{if $n\leq2$,}
\end{array} \right.
\end{equation}
where $Q_{2}(h)$, $\widetilde{Q}_{2}(h)$ and $Q_{[(3n-7)/2]}(h)$ denote polynomials with $\deg Q_{2}\leq{2}$, $\deg \widetilde{Q}_{2}\leq{2}$
and $\deg Q_{[(3n-7)/2]}\leq[(3n-7)/2]$, respectively.
If $n\leq2$, then the proof is similar to the proof of Theorem 1 and hence is omitted.
If $n\geq3$, in order to get the estimation of the number of zeros of $M(h)$, we take derivatives of $M(h)$ three
times. Noting that $M(2-b^2)=0$, it follows that
\begin{equation*}
H_{X_{2.10}}(n)\leq[\frac{3n-7}{2}]+3-1=[\frac{3n-3}{2}].
\end{equation*}
The proof is finished.
\section{$H_{X_{2.9}}(3)=3$ and $H_{X_{2.10}}(3)=3$}
In this section, we will study the sharp upper bound for the number of zeros of the Abelian integrals
with respect to the two systems for $n=3$, i.e., the determination of $H_{X_{2.9}}(3)$ and $H_{X_{2.10}}(3)$. For convenience,
we will denote
\[U^-=(-2,h_-), \quad U^+=(h_+,2).
\]
The exact expression of Abelian integral associated to system $X_{2.9}$ is given as follows:
\begin{equation}\begin{split}
M^{\pm}(h)&=\displaystyle\sum_{i+j=0}^{3}(-a_{ji}\bar{I}_{ij}(h)-b_{ij}I_{ij}(h))\\
&=a_{1}^{\pm}+a_{2}^{\pm}h+a_{3}^{\pm}h^2+a_{4}^{\pm}\sqrt{4-h^2}+a_{5}^{\pm}h\sqrt{4-h^2}, \quad h\in U^{\pm},
\end{split}\end{equation}
where
\begin{equation*}\begin{split}
a_{2}^{\pm}=&\displaystyle b_{00}\left(\pm\dfrac{2bc(6-b^2)\pi}{(4-b^2)^{3/2}}+
\dfrac{2(-2+c^2)\pi}{\sqrt{4-c^2}}\right)
+a_{00}\left(-\dfrac{2bc(-6+c^2)\pi}{(4-c^2)^{3/2}}\pm
\dfrac{2(-2+b^2)\pi}{\sqrt{4-b^2}}\right)\\
&-(b_{01}+a_{10})\left(\pm\dfrac{b\pi}{\sqrt{4-b^2}}+\dfrac{c\pi}{\sqrt{4-c^2}}\right)\pm\dfrac{2b_{11}\pi}{\sqrt{4-b^2}}+\dfrac{2a_{11}\pi}{\sqrt{4-c^2}}
\pm\dfrac{b(-b_{21}+a_{30})\pi}{\sqrt{4-b^2}}\\
&\pm\dfrac{4c(-2b_{10}+b_{20}b-2b_{30})\pi}{(4-b^2)^{3/2}}-\dfrac{4b(2a_{01}-a_{02}c+2a_{03})\pi}{(4-c^2)^{3/2}}
-\dfrac{c(a_{12}-b_{03})\pi}{\sqrt{4-c^2}},\\
a_{3}^{\pm}=&\displaystyle\pm\dfrac{4(b_{00}+b_{20})\pi}{(4-b^2)^{3/2}}+
\dfrac{4(a_{00}+a_{02})\pi}{(4-c^2)^{3/2}}\mp\dfrac{2b_{10}b\pi}{(4-b^2)^{3/2}}
-\dfrac{2a_{01}c\pi}{(4-c^2)^{3/2}}\\
&\pm\dfrac{b_{30}b(-6+b^2)\pi}{(4-b^2)^{3/2}}+\dfrac{a_{03}c(-6+c^2)\pi}{(4-c^2)^{3/2}},\\
a_{4}^{\pm}=&\pm\left(b_{21}-b_{03}+a_{12}-a_{30}\right)\pi,\\
a_{5}^{\pm}=&\pm(b_{30}+a_{03})\pi,
\end{split}\end{equation*}
and
\begin{equation}
a_{1}^{\pm}=m^{\pm}a_{2}^{\pm}+n^{\pm}a_{3}^{\pm}+p^{\pm}a_{4}^{\pm}+q^{\pm}a_{5}^{\pm}
\end{equation}
with
\begin{equation*}
m^{\pm}=-h_{\pm},\quad n^{\pm}=-h_{\pm}^{2},\quad
p^{\pm}=-\sqrt{4-h_{\pm}^{2}},\quad
q^{\pm}=-h_{\pm}\sqrt{4-h_{\pm}^{2}}.
\end{equation*}
Since $0\leq b<c<2$,
\begin{equation*}
\dfrac{\partial(a_{2}^{\pm},a_{3}^{\pm},a_{4}^{\pm},a_{5}^{\pm})}{\partial(a_{11},a_{01},a_{12},a_{03})}=-\dfrac{4c\pi^4}{(4-c^2)^2}\neq0,
\end{equation*}
which means that $a_{2}^{\pm},a_{3}^{\pm},a_{4}^{\pm},a_{5}^{\pm}$ are independent.
Denote
\[f_{0}^{\pm}(h)=h+m^{\pm},f_{1}^{\pm}(h)=h^2+n^{\pm},f_{2}^{\pm}(h)=\sqrt{4-h^2}+p^{\pm},f_{3}^{\pm}(h)=h\sqrt{4-h^2}+q^{\pm}.
\]
It follows from (4.1) and (4.2) that
\begin{equation}
M^{\pm}(h)=a_{2}^{\pm}f_{0}^{\pm}(h)+a_{3}^{\pm}f_{1}^{\pm}(h)+a_{4}^{\pm}f_{2}^{\pm}(h)+a_{5}^{\pm}f_{3}^{\pm}(h).
\end{equation}
By Theorem 1, we know that $H_{X_{2.9}}(3)\leq3$.
Now, we use \emph{Chebyshev
criterion} to show that $H_{X_{2.9}}(3)=3$, i.e., there exist the parameters $a_{i}^{\pm},
i=2,3,4,5$, such that $M^{\pm}(h)$ has exactly three zeros on
$U^{\pm}$, respectively. We introduce the following
definitions (see for instance \cite{MV}).
Let $f_{0},f_{1},...,f_{n-1}$ be analytic functions on an open
interval $L$ of $\mathbb{R}$. $(f_{0},f_{1},...,f_{n-1})$ is a
\emph{Chebyshev system} on $L$ if any nontrivial linear combination
\begin{equation*}
\lambda_{0}f_{0}(x)+\lambda_{1}f_{1}(x)+...+\lambda_{n-1}f_{n-1}(x)
\end{equation*}
has at most $n-1$ isolated zeros on $L$.
An ordered set $(f_{0},f_{1},...,f_{n-1})$ is a \emph{complete
Chebyshev system} on $L$ if
$(f_{0},f_{1},...,f_{i-1})$ is a Chebyshev system on $L$ for all
$i=1,2,...,n$.
An ordered set $(f_{0},f_{1},...,f_{n-1})$ is an \emph{extended
complete Chebyshev system} (in short, ECT-system) on $L$ if, for all
$i=1,2,...,n$, any nontrivial linear combination
\begin{equation}
\lambda_{0}f_{0}(x)+\lambda_{1}f_{1}(x)+...+\lambda_{i-1}f_{i-1}(x)
\end{equation}
has at most $i-1$ isolated zeros on $L$ counted with multiplicities.
\begin{remark}
If $(f_{0},f_{1},...,f_{n-1})$ is an ECT-system on $L$, then for
each $i=1,2,...,n$, there exists a linear of combination (4.4) with
exactly $i-1$ simple zeros on $L$ (see for instance Remark 3.7 in \cite{GV}).
\end{remark}
\begin{lemma}
\emph{(see \cite{MV})} $(f_{0},f_{1},...,f_{n-1})$ is an ECT-system on L
if, and only if, for each $i=1,2,...,n$,
\begin{equation*}
\Omega_{i}(x)=
\begin{vmatrix}
f_{0}(x) & f_{1}(x) & \cdots & f_{i-1}(x)\\
f_{0}^{\prime}(x) &f_{1}^{\prime}(x) & \cdots & f_{i-1}^{\prime}(x) \\
\vdots & \vdots & \ddots & \vdots \\
f_{0}^{(i-1)}(x) & f_{1}^{(i-1)}(x) & \cdots & f_{i-1}^{(i-1)}(x) \\
\end{vmatrix}\neq{0}
\end{equation*}
for all $x\in L$.
\end{lemma}
Simple computations show that, if $h\in U^{\pm}$,
\begin{equation*}\begin{array}{ll}
\Omega_{1}^{\pm}(h)=h-h_{\pm},\\[2ex]
\Omega_{2}^{\pm}(h)=(h-h_{\pm})^2,\\[2ex]
\Omega_{3}^{\pm}(h)=\displaystyle\dfrac{2}{(4-h^2)^{3/2}}
\left(h_{\pm}h^3-6h^2+16-2h^2_{\pm}\right)-2\sqrt{4-h^2_{\pm}},\\[2ex]
\Omega_{4}^{\pm}(h)=-\displaystyle\dfrac{24}{(4-h^2)^{3}}
\left((h_{\pm}^2-2)h^2-4h_{\pm}h+16-2h^2_{\pm}+\sqrt{4-h^2_{\pm}}\sqrt{4-h^2}(-4+h_{\pm}h)\right).
\end{array}\end{equation*}
Obviously, $\Omega_{1}^{\pm}(h)$ and $\Omega_{2}^{\pm}(h)$ do not vanish when $h\in U^{\pm}$. By direct computation, we obtain that $\Omega_{4}^{\pm}(h)$ has
only one zero $h=h_{\pm}$ with multiplicity four, which implies that
$\Omega_{4}^{\pm}(h)$ also does not change sign in each interval. However, it is a
little difficult to judge the sign of $\Omega_{3}^{\pm}(h)$.
\begin{proposition}
$(f_{0}^-(h),f_{1}^-(h),f_{2}^-(h),f_{3}^-(h))$ is an
ECT-system on $U^-$. If
$b^2+c^2\leq 4$, $(f_{0}^+(h),f_{1}^+(h),f_{2}^+(h),f_{3}^+(h))$ is an
ECT-system on $U^+$; if $b^2+c^2>4$, there is a real number
$d\in U^+$ such that $(f_{0}^+(h),f_{1}^+(h),f_{2}^+(h),f_{3}^+(h))$ is
an ECT-system on $(h_{+},d)$.
\end{proposition}
\begin{proof}
It is easy to verify that $\Omega_3^{\pm}(h_{\pm})=0$ and
\begin{equation}\label{f3}
{\Omega_3^{\pm}}^{\prime}(h)=-\displaystyle\dfrac{12h(h-h_{\pm})^2}{(4-h^2)^{5/2}}.
\end{equation}
Thus ${\Omega_3^{\pm}}^{\prime}(h)$ has zeros $h=0$ and $h=h_{\pm}$ with
multiplicity one and two, respectively.
Note that $U^-=(-2,h_-)$ and $h_{-}<0$. If $h\in U^-$, by \eqref{f3}, we find that ${\Omega_3^{-}}^{\prime}(h)>0$ holds, which implies that
$\Omega_{3}^-(h)<\Omega_3^-(h_{-})=0$. Therefore combining with the discussions above and lemma 10,
$(f_{0}^-(h),f_{1}^-(h),f_{2}^-(h),f_{3}^-(h))$ is an ECT-system on
$U^-$.
We split the proof into two cases if $h\in U^+$, i.e., $h\in(h_{+},2)$.
\emph{Case 1}. $h_{+}\geq0$, i.e., $b^2+c^2\leq4$.
By \eqref{f3}, ${\Omega_3^+}^{\prime}(h)<0$ holds on $U^+$. It follows that
$\Omega_{3}^+(h)<\Omega_3^+(h_{+})=0$. Thus $\Omega_{3}^+(h)$ does not vanish on $U^+$, which shows that
$(f_{0}^+(h),f_{1}^+(h),f_{2}^+(h),f_{3}^+(h))$ is an ECT-system on
$U^+$.
\emph{Case 2}. $h_{+}<0$, i.e., $b^2+c^2>4$.
We get ${\Omega_3^+}^{\prime}(h)>0$ for $h\in(h_{+},0)$ and ${\Omega_3^+}^{\prime}(h)<0$ for $h\in(0,2)$ by \eqref{f3}. Noting that $\Omega_3^+(h_{+})=0$,
$\Omega_3^+(0)=(\small\sqrt{4-h_{+}^2}-2)^2/2>0$ and $\Omega_{3}^+(h)\rightarrow{-\infty}$ as $h\rightarrow2^{-}$, we know that there is a simple zero of $\Omega_{3}^+(h)$
on $(0,2)$. Thus $(f_{0}^+(h),f_{1}^+(h),f_{2}^+(h),f_{3}^+(h))$ is not
an ECT-system on $U^+$ by lemma 10. But there is a real number
$d\in U^+$ such that $\Omega_{3}^+(h)>0$ for $h\in(h_{+},d)$. Consequently,
$(f_{0}^+(h),f_{1}^+(h),f_{2}^+(h),f_{3}^+(h))$ is an ECT-system on
$(h_{+},d)$.
The proof is finished.
\end{proof}
Next, let us consider system $X_{2.10}$.
Similarly, the Abelian integral associated to system $X_{2.10}$ has
the form:
\begin{equation}\begin{split}
M(h)&=\displaystyle\sum_{i+j=0}^{3}(-a_{ji}-b_{ij})I_{ij}(h)\\
&=a_{1}+a_{2}h+a_{3}h^2+a_{4}\sqrt{4-h^2}+a_{5}h\sqrt{4-h^2}, \quad h\in(2-b^2,2),
\end{split}\end{equation}
where
\begin{equation*}\begin{split}
a_{2}=&-\displaystyle\dfrac{4(b_{00}+a_{00})(4-6b^2+b^4)\pi}{(4-b^2)^{3/2}}
-\dfrac{8(b_{10}+a_{01}+b_{30}+a_{03})b\pi}{(4-b^2)^{3/2}}
-\dfrac{2(b_{01}+a_{10})b\pi}{\sqrt{4-b^2}}\\
&+\dfrac{4(b_{20}+a_{02})b^2\pi}{(4-b^2)^{3/2}}+\dfrac{2(b_{11}+a_{11})\pi}{\sqrt{4-b^2}}
-\dfrac{(b_{21}+a_{12}-b_{03}-a_{30})b\pi}{\sqrt{4-b^2}},
\end{split}\end{equation*}
\begin{equation*}\begin{split}
a_{3}&=\displaystyle\dfrac{4(b_{00}+b_{20}+a_{00}+a_{02})\pi}{(4-b^2)^{3/2}}-\dfrac{2(b_{10}+a_{01})b\pi}{(4-b^2)^{3/2}}
+\dfrac{(b_{30}+a_{03})b(-6+b^2)\pi}{(4-b^2)^{3/2}},\\
a_{4}&=(b_{21}-b_{03}+a_{12}-a_{30})\pi,\\
a_{5}&=(b_{30}+a_{03})\pi,
\end{split}\end{equation*}
and
\begin{equation}\begin{split}
a_{1}=ma_{2}+na_{3}+pa_{4}+qa_{5},
\end{split}\end{equation}
with
\begin{equation*}
m=-2+b^2, \quad n=-(-2+b^2)^2,\quad
p=-b\sqrt{4-b^2}, \quad q=b\sqrt{4-b^2}(-2+b^2).
\end{equation*}
It follows from
\begin{equation*}
\dfrac{\partial(a_{2},a_{3},a_{4},a_{5})}{\partial(a_{11},a_{01},a_{12},a_{03})}=
-\dfrac{4b\pi^4}{(4-b^2)^2}\neq0
\end{equation*}
that $a_{2},a_{3},a_{4},a_{5}$ are independent.
Denote
$f_{0}(h)=h+m, f_{1}(h)=h^2+n, f_{2}(h)=\sqrt{4-h^2}+p, f_{3}(h)=h\sqrt{4-h^2}+q$.\\
By (4.6) and (4.7),
\begin{equation}
M(h)=a_{2} f_{0}(h)+a_{3}f_{1}(h)+a_{4} f_{2}(h)+a_{5} f_{3}(h).
\end{equation}
\begin{proposition}
When $0<b\leq\sqrt{2}$, $( f_{0}(h), f_{1}(h), f_{2}(h), f_{3}(h))$ is an
ECT-system on $(2-b^2,2)$; when $\sqrt{2}<b<2$, there is a real number
$d\in(2-b^2,2)$ such that $( f_{0}(h), f_{1}(h), f_{2}(h), f_{3}(h))$ is
an ECT-system on $(2-b^2,d)$.
\end{proposition}
\begin{proof}
By direct computations,
\begin{equation*}\begin{split}
\Omega_{1}(h)=&h+b^2-2,\\
\Omega_{2}(h)=&(h+b^2-2)^2,\\
\Omega_{3}(h)=&\displaystyle\dfrac{2}{(4-h^2)^{3/2}}
\left((2-b^2)h^3-6h^2-2(-4-4b^2+b^4)\right)-2b\sqrt{4-b^2},\\
\Omega_{4}(h)=&-\displaystyle\dfrac{24}{(4-h^2)^{3}}
\left((2-4b^2+b^4)h^2-4(2-b^2)h+8+8b^2-2b^4\right.\\
&\left.+b\sqrt{4-b^2}\sqrt{4-h^2}(-4+(2-b^2)h)\right),
\end{split}\end{equation*}
where $h\in(2-b^2,2)$.
Obviously, $\Omega_{1}(h)$ and $\Omega_{2}(h)$ do not vanish on $(2-b^2,2)$. It is easy to verify that $h=2-b^2$
is a unique zero of $\Omega_{4}(h)$ with multiplicity four, which implies that $\Omega_{4}(h)\neq0$ for $h\in(2-b^2,2)$.
Therefore we just need to determine the sign of $\Omega_{3}(h)$. The discussion is similar to the proof of Proposition 11 for $h\in U^+$, thus we omit it.
\end{proof}
\noindent{\bfseries{Proof of Theorem 4.}} It follows from Theorem 1
that $H_{X_{2.9}}(3)\leq3$. By Proposition 11 and
Remark 9, we know that there exist $a_{i}^{\pm}, i=2,3,4,5$, such that the Abelian integral
$M^{\pm}(h)$ has exactly three zeros on $U^{\pm}$,
respectively. Therefore $H_{X_{2.9}}(3)=3$.
Similarly, $H_{X_{2.10}}(3)=3$ by Theorem 2, Proposition 12 and
Remark 9.
\section{Simultaneous bifurcation and distribution of limit cycles for system $X_{2.9}$}
In what follows we will consider the simultaneous bifurcation of limit cycles
bifurcating from the two period annuli of system $X_{2.9}$ under cubic polynomial perturbations. Note that $a_4^-=-a_4^+$ and $a_5^-=-a_5^+$.
Rewrite (4.3) as
\begin{equation}
M^{\pm}(h)=a_{2}^{\pm}f_{0}^{\pm}(h)+a_{3}^{\pm}f_{1}^{\pm}(h)\pm a_{4}^{+}f_{2}^{\pm}(h)\pm a_{5}^+f_{3}^{\pm}(h),\quad h\in U^{\pm}.
\end{equation}
\noindent{\bfseries{Proof of Theorem 5.}} We study the number of zeros of the two Abelian integrals simultaneously. Note that
\[
\dfrac{\partial(a_{2}^+,a_{2}^-,
a_{3}^+,a_{3}^-,a_{4}^+,a_{5}^+)}{\partial(a_{11},b_{11},b_{00},a_{01},b_{21},b_{30})}=\dfrac{128c{\pi}^6}{(4-b^2)^2(4-c^2)^2}\neq0.
\]
Thus, we can consider $a_{2}^+,a_{2}^-,a_{3}^+,a_{3}^-,a_{4}^+,a_{5}^+$ to be independent.
Firstly, let $a_{3}^+=a_{3}^-=a_{4}^+=a_{5}^+=0$ and $a_{2}^+, a_{2}^-\neq0$, then neither $M^+(h)$ has zeros on $U^+$ nor
$M^-(h)$ has zeros on
$U^-$. Hence, the distribution (0,0) is
possible.
Secondly, let $a_{4}^+=a_{5}^+=0$. By Proposition 11 and Remark 9, we can choose $a_{2}^+,
a_{3}^+$ (resp. $a_{2}^-, a_{3}^-$) such that $M^+(h)$
(resp. $M^-(h)$) has none or one zero on $U^+$
(resp. $U^-$). Thus the distributions (1,0), (0,1) and (1,1) can be
achieved.
Thirdly, take $a_{5}^+=0$. It follows from Proposition 11 and Remark 9 that there exist
$\rho_{1}, \rho_{2}, \rho_{3}(\neq0)$ and $\sigma_{1}, \sigma_{2}, \sigma_{3}(\neq0)$ such that
\[
\rho_{1}f_{0}^{+}(h)+\rho_{2}f_{1}^{+}(h)+\rho_{3}f_{2}^{+}(h)\quad \mbox{and} \quad
\sigma_{1}f_{0}^{-}(h)+\sigma_{2}f_{1}^{-}(h)-\sigma_{3}f_{2}^{-}(h)
\]
have $u$, $v$ zeros respectively, with $0\leq u, v\leq2$. Then multiplying by
$\rho_{3}/\sigma_{3}$ the
second function, it turns out that
\[
\sigma_{1}\rho_{3}/\sigma_{3}f_{0}^{-}(h)
+\sigma_{2}\rho_{3}/\sigma_{3}f_{1}^{-}(h)-
\rho_{3}f_{2}^{-}(h)
\]
has $v$ zeros.
Choose $a_{2}^+,
a_{2}^-,a_{3}^+, a_{3}^-,a_{4}^+$ such that
\begin{equation*}
a_{2}^+=\rho_{1}, \quad
a_{2}^-=\sigma_{1}\rho_{3}/\sigma_{3},\quad
a_{3}^+=\rho_{2},\quad
a_{3}^-=\sigma_{2}\rho_{3}/\sigma_{3},\quad
a_{4}^+=\rho_{3}.
\end{equation*}
It follows that the distribution $(u,v), 0\leq u, v\leq2$ is
possible.
Finally, suppose that $a_{5}^+\neq0$. It is easy to obtain that
\[
{M^{+}}^{(3)}(h)=-\dfrac{12(a_4^+h+4a_5^+)}{(4-h^2)^{5/2}} \quad \mbox{and} \quad {M^{-}}^{(3)}(h)=\dfrac{12(a_4^+h+4a_5^+)}{(4-h^2)^{5/2}}.
\]
If $a_{4}^+=0$, then $M^+(h)$ and $M^-(h)$ have at most two zeros on $U^+$ and $U^-$, respectively. Thus, in order to get more limit cycles, $a_{4}^+\neq0$.
We find that ${M^{+}}^{(3)}(h)$ and ${M^{-}}^{(3)}(h)$ have the same unique zero $h_0=-4a_5^+/a_4^+$, which implies that ${M^{+}}^{(2)}(h)$ has at most two (resp. one) zero(s) on $U^+$ and ${M^{-}}^{(2)}(h)$ has at most one (resp. two) zero(s) on $U^-$ if $h_0\in U^+$ (resp. $h_0\in U^-$), otherwise, both ${M^{+}}^{(2)}(h)$ and ${M^{-}}^{(2)}(h)$ have at most one zero on $U^+$ and $U^-$, respectively. Thus, $M^{+}(h)$ has at most three (resp. two) zeros on $U^+$ and $M^{-}(h)$ has at most two (resp. three) zeros on $U^-$, which means the distribution (3,3) is impossible. To show (3,2) is achievable, we give the following asymptotic expansions of $M^{\pm}(h)$
at $h=h_{\pm}$, respectively:
\begin{equation}\begin{split}
M^{\pm}(h)=s_1^{\pm}(h-h_{\pm})+s_2^{\pm}(h-h_{\pm})^2+s_3^{\pm}(h-h_{\pm})^3+s_4^{\pm}(h-h_{\pm})^4+\cdots,
\end{split}\end{equation}
where
\begin{equation*}\begin{split}
s_1^{\pm}&=a_2^{\pm}+2a_3^{\pm}h_{\pm}\mp\frac{a_4^+h_{\pm}+2a_5^+(h_{\pm}^2-2)}{(4-h_{\pm}^2)^{1/2}}, \quad
s_2^{\pm}=a_3^{\pm}\mp\frac{2a_4^+-a_5^+h_{\pm}(h_{\pm}^2-6)}{(4-h_{\pm}^2)^{3/2}},\\
s_3^{\pm}&=\mp\dfrac{2(a_4^+h_{\pm}+4a_5^+)}{(4-h_{\pm}^2)^{5/2}},\quad \quad \quad \quad \quad \quad \quad \quad \quad \,
s_4^{\pm}=\mp\dfrac{2(a_4^++5a_5^+h_{\pm}+a_4^{+}h_{\pm}^2)}{(4-h_{\pm}^2)^{7/2}}.
\end{split}\end{equation*}
Since
\[
\dfrac{\partial(s_{1}^+,s_{1}^-,
s_{2}^+,s_{2}^-,s_{3}^+,s_{4}^+)}{\partial(a_{2}^+,a_{2}^-,
a_{3}^+,a_{3}^-,a_{4}^+,a_{5}^+)}=-\dfrac{4}{(4-h_{+}^2)^5}\neq0,
\]
we consider $s_{1}^+,s_{1}^-,
s_{2}^+,s_{2}^-,s_{3}^+,s_{4}^+$ as the new independent parameters. Denote
\[
M^+(h)=M^+(h,s_{1}^+,s_{2}^+,s_{3}^+,s_{4}^+) \quad \mbox{and} \quad M^-(h)=M^-(h,s_{1}^-,s_{2}^-,s_{3}^-,s_{4}^-).
\]
Without loss of
generality, suppose $s_{4}^+>0$. To get more zeros of $M^+(h)$, we choose $s_{i}^+$ and $h_{i}\in U^+$, $i=
4, 3, 2, 1$, such that
\begin{equation}\begin{split}\label{M^+}
&M^+(h_4, 0, 0, 0, s_{4}^+)>0, \quad \quad \,M^+(h_3,0,0,s_{3}^+,s_{4}^+)<0,\\
&M^+(h_2,0,s_{2}^+,s_{3}^+,s_{4}^+)>0, \quad M^+(h_1,s_{1}^+,s_{2}^+,s_{3}^+,s_{4}^+)<0,
\end{split}\end{equation}
and $0 < |s_{1}^+|\ll|s_{2}^+|\ll |s_{3}^+|\ll |s_{4}^+|$, $h_+ < h_1 < h_2< h_3 < h_4 < 2$. It is
easy to show that $M^+(h)$ has three zeros which tend to $h_+$. Once $s_{3}^+,s_{4}^+$ are chosen, the sign of $s_3^-$ is
determined. For example, from the analysis above, we know that $s_{4}^+>0$ and $s_{3}^+<0$ by \eqref{M^+}. The result that $M^+(h)$ has three zeros on $U^+$ implies $h_0=-4a_5^+/a_4^+\in U^+$. Thus $a_4^+<0$ and $s_{3}^->0$.
There exists $h_5\in U^-$ such that $M^-(h_5,0,0,s_{3}^-,s_{4}^-)<0$. Take $s_{i}^-, i=2,1$ and
$h_{i}\in U^-, i=6,7$, such that
\[
M^-(h_6,0,s_{2}^-,s_{3}^-,s_{4}^-)>0, \quad M^-(h_7,s_{1}^-,s_{2}^-,s_{3}^-,s_{4}^-)<0,
\]
and $0 < |s_{1}^-|\ll|s_{2}^-|\ll \mbox{min}\{|s_{3}^-|,|s_{4}^-|\}$, $-2 < h_5 < h_6< h_7 < h_-$. It follows that
$M^-(h)$ has at least two zeros on $U^-$, which tend to $h_-$. Thus, (3,2) is realizable. Similarly, other configurations of limit cycles $(u,v)$ with $0\leq u, v\leq 3, u+v\leq5$
can be realized in this way.
This completes the proof.
\section*{Acknowledgements}
Research is supported by the International Program of Project 985, Sun Yat-Sen University, the NSF of China (No. 11171355) and the Ph.D. Programs Foundation of Ministry of Education of China (No. 20100171110040).
|
\section[]{INTRODUCTION}
Flat rotation-curve observations from spiral galaxies have been the smoking gun for theories of missing mass or new laws of gravity since the 1970's ~\citep{1978Rubin,Bosma78}. There are two fundamental observables involved in the flat rotation-curve problem: photometry and spectra.
Photometry is the measurement of the total light, in a specific wavelength band, which is interpreted as mass through a population synthesis model. The visible mass then yields orbital velocities by Newton's second law and the expected Keplerian fall-off at large radii. The second observable is Doppler-shifted spectra from characteristic atomic transitions, which imply the dynamical
mass by the relation between the orbital velocities (from the Lorentz Doppler-shift formula) and the Poisson equation. These two observables predict very different mass content, since rotation curves from spectra do not fall off in a Keplerian sense at large radii. The most popular theories to explain the conflicting observations are missing mass or new laws of gravity.
The missing mass, commonly called dark matter, invokes a new particle which does not interact electromagnetically, and hence is ``dark.'' New laws of physics commonly work on the principle that our understanding of gravity, via General Relativity, is incomplete on extra-galactic distance scales.\\
Examples of dark matter theories include the psuedo-isothermal dark matter halo model (ISO)~\citep{ISO} and the Navarro, Frenk \& White (NFW)~\citep{NFW} dark matter halo model. Examples of new laws of gravity include modified Newtonian dynamics (MOND)~\citep{Milgrom}, non-local gravity ~\citep{Rahvar} and conformal relativity ~\citep{Mannheim}. For a full review of NFW or MOND approaches see ~\citet{SanMcGa,Gianfranco}. MOND successfully predicts core densities, the baryonic Tully-Fisher relation, and rotation curves. Dark matter theories are the standard for interpreting flat rotation-curves and form the basis of a comprehensive
cosmology. But the fact remains that neither class of theory, dark particles or deviations from General Relativity, has been observed in decisive terrestrial experiments and neither has resulted in a full and rigorously computed set of population synthesis models~\citep{Blok}. Dark matter may have been observed in the DAMA ~\citep{DAMA}, CoGeNT ~\citep{CoGeNT}, and CDMS ~\citep{CDMS} experiments, although these results are in some conflict with the limits of XENON100 ~\citep{XENON}. \\
It is a common feature of rotation curves that galaxies smaller than the Milky Way are `dark-matter dominated' (ascending rotation curves) and galaxies larger than the Milky Way have minimal dark matter halos (descending rotation curves). Classification of galaxies based on the relative size of the dark matter halo is called the universal rotation curve~\citep{Rub,PS88,PS90}. Since a priori there appears to be no physical reason for the Milky Way to occupy the median point in this dark matter halo distribution, we interpret the universal rotation curve as indicating frame-dependent effects in rotation curve observations. \\
Currently the Doppler-shifted spectra are interpreted to imply only frame-dependent effects from relative velocity. However, it is known from classical electrodynamic theory that effects on light naturally divide mathematically between those from relative velocity and those from relative acceleration\footnote{of the emitter with respect to the receiver frame}~\citep[see][Eqs.~(14.13) \& (14.14)]{Jack}. In the context of Relativistic kinematics we replace the word ``acceleration'' with ``curvature.'' The Luminous Convolution Model (LCM) is an empirical formula for predicting flat rotation curves based on interpreting the Doppler-shifted spectra as having two contributions: relative velocity \emph{and} relative curvature. LCM curvature effects are quantified via the gravitational redshifts for the respective luminous matter content of the emitter and receiver galaxies, and rephrased in kinematical terms using equivalent Doppler-shifts.\\
A sample of twenty-three $(23)$ galaxies from forty-two $(42)$ different data sets, and four $(4)$ different Milky Way luminous mass models will be used to demonstrate the utility of the LCM to constrain luminous mass modeling. Comparisons are made to reported mass-to-light ratios from dark matter and MOND models, and differing distance indicators are noted. The highest resolution rotation curve data are presented here along with the older standards in the field of spiral-galaxy rotation-curve studies.
The LCM has one free parameter $\tilde{a}$ that is dimensionless and does not yet have a physical interpretation. The LCM successfully predicts rotation curves across a broad range of galaxy sizes and morphologies. In this paper we allow the fits to scale the given mass-to-light ratios and gas fractions reported by the reference data sets, within a fixed range. Our analysis is focused on the high symmetry case of spiral galaxies in the plane of the galactic disk. While the LCM mapping can be extended to an arbitrary metric~\citep{Cisneros:2013vha} to predict other astrophysical observations, such studies are beyond the scope of the current paper. On average for the data studied in this paper, the LCM rotation curve fits are better than MOND or dark matter model fits, returning reasonable estimates of stellar and gas masses.
The paper is divided as follows: section~\ref{sec:stationTotal} describes the rotation curve formalism and associated mass modeling; section~\ref{sec:DERIVE} derives the LCM; section~\ref{sec:RESULTS} describes the sample and LCM results, and section~\ref{sec:conclusion} gives our conclusions.
\section[]{ROTATION CURVES OF SPIRAL GALAXIES}
\label{sec:stationTotal}
\subsection{Velocity addition formulae}
In dark matter halo theories the total dynamical mass $M'$ as a function of radius
is a sum of the luminous mass $M_{l}$ and the
dark matter halo mass $M_{dm}$:
\begin{equation}
M'(r)=M_{l}(r)+M_{dm}(r).
\label{eq:mass1}
\end{equation}
The mass components are then related to Poisson law forces for the appropriate geometry.
Stellar disks and gas distributions are generally modeled as exponential profiles and dark matter
halos and stellar bulges are modeled as spherical. The associated sum of forces then
yields a sum of centripetal accelerations, $v^2/r$, attributed to each component. Orbital velocities are generally treated
as those experienced by a test particle at some radius $r$ from the center of the mass distribution, assumed to be in a stable, circular orbit.
The total velocity is taken to be a quadratic sum of the contributions from the luminous $v^2_{l}$ and the dark halo
$ v^2_{halo}$ velocities:
\begin{equation}
v_{rot}^2 = v_{l}^2 + v_{halo}^2.
\label{eq:zonte1}
\end{equation}
All velocities are functions of the radius. The predicted total velocity $v_{rot}$ is fit to the observed rotation curve $v_{obs}$.
For a rotation curve observation made using a well-known atomic transition,
having a frequency $\omega_o$ in the observer's rest-frame and
Doppler-shifted frequency $\omega'$, the observed rotation curve is:
\begin{equation}
\frac{v_{obs}(r)}{c}=
\frac{ \frac{\omega'(r)}{\omega_o}- \frac{\omega_o}{\omega'(r)}}{
\frac{\omega'(r)}{\omega_o}+ \frac{\omega_o}{\omega'(r)}}.
\label{eq:dataLorentz}
\end{equation}
This is the standard Lorentz Doppler-shift formula.
\subsubsection{Assumption of spherical symmetry as a final approximation}
\label{spherical}
Forces associated with each of the mass components in Eq.~\ref{eq:mass1} (gas, stellar disk and bulge, and dark matter halo) come individually from the solution to a classical Poisson equation for a given geometry. Dark matter halos and stellar bulges are generally modeled as spherical, whereas stellar disks and gas distributions are modeled as exponential disks with Bessel functions. Since the mass sum and the resulting sum of centripetal accelerations are not functionally dependent on $r$ in the same way~\citep{Binney}, the quadratic velocity sum in Eq.~\ref{eq:zonte1} is a simplifying assumption of spherical symmetry.
Rotation curve derivations based on concentric rings also employ the assumption of spherical symmetry, so that the Gaussian shell technique can be used~\citep{Frat}. Luminous mass modeling relies on the assumption of spherical symmetry as a final approximation as well, to sum the contributions to the gravitational potential from the terms calculated in their respective geometries~\citep{Xue,Klypin}. The logic of invoking this assumption is
that the error introduced by assuming spherical geometry for gravitational potentials calculated with disk geometries is a difference in magnitudes, not functional line shape with respect to radius~\citep{Chatterjee}.
The spherical assumption is beneficial computationally and conceptually. Since the flat rotation-curve problem is clearly a functional difference at large radii, this approximation cannot be responsible for the rotation curve velocities. Since the assumption of spherical symmetry is commonly used as a final approximation in evaluating spiral galaxies, we will use this assumption in the LCM derivation (section~\ref{sec:DERIVE}) in the same ways .
\subsubsection{LCM velocity addition formula}
\label{LCMfunc}
In the LCM, the observed Doppler-shifted frequencies $\omega'$ in Eq.~\ref{eq:dataLorentz} are posited to include two contributions: relative velocity and relative curvature. Since these frequency shifts are reported as the total rotation velocity, $v_{obs}$, we phrase our arguments kinematically, including contributions to the relative curvature in terms of equivalent Doppler-shifts. It is clear from the work of ~\citet{Cisn} and ~\citet{Radosz} that
curvature effects on light can be cast in purely kinematical terms.
Since dark matter theories are based on classical Newtonian and Relativistic gravity, we co-op the dark matter halo velocity addition formula in Eq.~\ref{eq:zonte1}, to represent the contributions from relative velocity and relative curvature to the total frequency shifts by replacing the dark matter contribution $v^2_{dm}$ with a relative curvature term $\tilde{v}^2_{lcm}$:
\begin{equation}
v_{rot}^2 = v_{l}^2+ \tilde{a} \tilde{v}_{lcm}^2 ,
\label{eq:zonteLCM}
\end{equation}
for $ v_{l}^2$, the relative velocity contribution; $\tilde{a}$ is
a dimensionless, free parameter (see Fig.~\ref{fig:alphaDistrib1}), and $v_{rot}$ is the total rotation velocity.
Relative velocity and relative curvature effects are calculated using only the luminous matter as reported in the cited literature, consistent with photometry and population synthesis modeling. In this way, by recasting the curvature contributions, using
the associated gravitational redshifts as a function of radius to get the equivalent Doppler-shifts, the relative velocity and relative curvature are represented by the kinematics of the Lorentz group.
\subsection{Luminous mass modeling}
\label{sec:POTENTIALS}
Luminous matter modeling in spiral galaxies is an under-constrained
field of study ~\citep{Conroy}. Dark matter models and MOND are commonly used as constraints on luminous matter modeling. Fig.~\ref{fig:massmodels18} compares original reported luminous profiles, fit with dark matter and MOND models, from fourteen ($14$) of the galaxies in our sample. As can be seen in the figure, neither
class of model demonstrates consistent predictions of the luminous mass in a given galaxy.
Estimates of the total luminous galaxy mass, $M_l$, commonly include
the masses of gas and dust, $M_{g}$, stellar bulge, $M_{b} $ and disk, $M_{d} $,
\begin{equation}
M_{l}=M_{g}+M_{d} +M_{b}.
\label{eq:lumMasses}
\end{equation}
Measurements of the individual component come from a variety of observations for the
surface brightness profiles at many wavelengths. The dominant gas
mass is contributed by atomic hydrogen, H\,{\sevensize\bf I}, observed at $21$-cm and by
molecular gas, observed through CO rotational lines in the
mm-regime. Generally, it is assumed that the $21$-cm line is optically
thin in which case it can be translated directly into a column
density of H\,{\sevensize\bf I}. Estimates of the effects of optical depth (e.g. ~\citet{Braun}) imply that the actual H\,{\sevensize\bf I} column density could be higher
by up to $30\%$, but this likely is dependent on the inclination of the
galaxy. In the case of CO, various elaborate calibration methods have
led to a conversion factor of CO intensity to a molecular hydrogen
column density. This has an associated uncertainty of up to a factor of two.
Generally, molecular gas surface densities are higher in the inner
disks of galaxies and H\,{\sevensize\bf I} column densities in the outer disks of
galaxies. In terms of ionized gas, the only substantial mass is
contributed by a diffuse ionized medium. The column density of this
ionized medium is about $30\%$ of that of the H\,{\sevensize\bf I} column in the solar
neighborhood. Because it is only possible to obtain column density
estimates in the Milky Way for this gas (from pulsar dispersion
measurements), it is usually ignored in rotation curve
modeling. Other ionized gas phases in H\,{\sevensize\bf II} regions and hot X-ray
emitting plasmas contribute little mass. The derived gas column
densities are generally corrected for the presence of helium, by
a factor of approximately 1.4. Lastly, the dust-to-gas ratio in the
interstellar gas is only about $1\%$ so dust masses may be ignored in
rotation curve modeling. ~(R.\,A.\,M.\, Walterbos, personal communication, 2014)
Associating a mass surface density with the observed light profiles
of the stellar components involves a suite of modeling assumptions
used to reproduce the observed light intensity in various bands. The
conversion of light into stellar mass generally depends on the
age(s) of the stellar populations present, their metallicities,
corrections for obscuration by dust and the inclination of the
galaxy. The stellar populations present are a consequence of the
star formation history, the initial mass function, and the
evolutionary history of the galaxy. The most important of these
factors are included in a population synthesis model (PSM) that is
matched to the observed light distribution in various colors to yield
a mass-to-light ratio (M/L). Conversion of light into masses by the M/L is the largest uncertainty in luminous matter modeling~\citep{Toky} as evidenced by the variations in Fig.~\ref{fig:massmodels18}.
After the appropriate Poisson's law forces are calculated for each galaxy component in
Eq.~(\ref{eq:lumMasses}), the
orbital velocities are added in quadrature by the assumption of spherical symmetry:
\begin{equation}
v^2_{l} = v^2_{g} + ( \rmn{M/L} )_d v^2_{d} +( \rmn{M/L} )_b v^2_{b}.
\label{eq:quadrature}
\end{equation}
The orbital rotation velocity associated with the gas mass is $v^2_{g}$, with the stellar bulge mass is $(\rmn{M/L} )_b v^2_{b} $ and with the stellar disk mass is $(\rmn{M/L})_d v^2_{d}$.
\begin{figure*}
\centering
\subfigure[]{\includegraphics[width=0.3\textwidth]{m31LumMassmodels.pdf}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{n7814LumMassmodels.pdf}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{n891LumMassmodels.pdf}}\\
\subfigure[]{\includegraphics[width=0.3\textwidth]{n2841LumMassmodels.pdf}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{n7331LumMassmodels.pdf}}
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n5055LumMassmodels.pdf}}\\
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n2903LumMassmodels.pdf}}
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n6946LumMassmodels.pdf}}
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n3198LumMassmodels.pdf}}\\
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n2403LumMassmodels.pdf}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{m33LumMassmodels.pdf}}
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{f563_1LumMassmodels.pdf}}\\
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n925LumMassmodels.pdf}}
\subfigure[ ]{\includegraphics[width=0.3\textwidth]{n7793LumMassmodels.pdf}}
\caption{Original reported rotation curve velocities due to the posited luminous mass for spiral galaxies in this sample; the originating model context is indicated in each figure legend. Variations in the originating luminous mass profiles reported in the literature illustrate the under-constrained nature of population synthesis modeling, both in magnitude and geometry. Originating model M/L ratios, wavelength bands and references are in Table~\ref{sumRESULTS}.}
\label{fig:massmodels18}
\end{figure*}
\section[]{Lorentz kinematics and the LCM derivation }
\label{sec:DERIVE}
The LCM construction is based upon a careful identification of the relationship between the local flat frames
where physical measurements are made and the underlying curved manifolds. It is known from
the differential geometry of General Relativity, that sufficiently precise measurements in the local flat frames can completely specify the underlying curved manifold's metric.
In the flat rotation-curve problem, we assume spectra are such sufficiently precise measurements, and that our local observations of the visible extent of spiral galaxies gives enough information
to determine the underlying global curvatures of the emitter and receiver galaxies from their baryonic masses.
\subsection{Curvatures from luminous mass}
\label{sec:curvatures}
Curvatures in Schwarzschild metric space-times are indicated by gravitational redshift effects on photon frequencies:
\begin{equation}
\frac{\omega_o }{ \omega ( r) }=\left(\frac{1}{\sqrt{-g_{tt}} }\right)_r,
\label{eq:Clone}
\end{equation}
where $\omega_o$ is the characteristic photon frequency as defined in section~\ref{sec:stationTotal}, $\omega (r)$ is the shifted frequency\footnote{ as
received by a stationary observer at asymptotic infinity}. The time coefficient of the Schwarzschild metric is $g_{tt}$:
\begin{equation}
g_{tt}(r)=-\left(1- 2\frac{G M}{c^2r}\right),
\label{eq:timeportion}
\end{equation}
where $G$ is Newton's constant of gravity, $M$ is the enclosed mass at some radial distance $r$ from the center of
the mass distribution, and
$c$ is the vacuum light speed.
In the weak field limit ~\citep{Hartle}
the metric coefficient becomes:
\begin{equation}
g_{tt}(r) \approx -1 + 2\frac{\Phi(r)}{c^2}.
\label{eq:weakfield}
\end{equation}
where $\Phi$ is the Newtonian scalar gravitational potential. \\
Spiral galaxies are treated as weak fields due to the diffuse nature of the luminous mass distributions, such that dark matter theories are based on Newtonian kinematics and Special Relativistic interpretations of spectra. To extend these flat space-time concepts to the slightly curved frames of spiral galaxies, we will make the following two caveats.
\subsubsection{Assumption $1$ }
The exterior Schwarzschild metric is a vacuum solution to Einstein's equations,
intended for use outside the central mass that is generating the curvature. By invoking the simplifying assumption of spherical symmetry generally used in galaxy analysis (see section \ref{sec:stationTotal}), Gauss's law for spherical mass distributions~\citep{Fowles} becomes a valid approximation. Therefore, at each radius $r$,
all mass elements external to $r$ cancel by symmetry, such that the exterior Schwarzschild metric exactly satisfies the Einstein equations as a vacuum solution. Note that any theory using this approximation necessarily loses information at small radii where tidal forces dominate or in the presence of severe symmetry breaking features. Since the flat rotation-curve problem is one of large radii, we find this to be an acceptable approximation to first order. Each galaxy manifold can be viewed as foliations of Schwarzschild solutions with increasing radii.
In practice this assumption results in summing the stellar and gas contributions to the gravitational potential $\Phi$ in Eq.~\ref{eq:weakfield} under the assumption of spherical symmetry, though the contributions are individually calculated in the correct geometric Poisson equation. This approach is consistent with the standard treatment of gravitational potentials in rotation curve calculations~\citep{Xue,Klypin} because, to
good approximation, deviations from spherical symmetry are higher order corrections to the gravitational potential~\citep{Binney,Chatterjee}.
\subsubsection{Assumption $2$ }
The gravitational potential, $\Phi(r)$, which parametrizes the curvatures of interest in Eq.~\ref{eq:weakfield}, is defined as an integral over the Newtonian force $F(r)$:
\begin{equation}
\Phi(r)+\Phi_o=-\int \frac{F(r) }{m} dr,
\label{eq:potentialgeneral}
\end{equation}
where $F(r) /m$ is the force per unit mass for each individual luminous mass component, and $\Phi_o$ is the integration constant of interest.
This integration constant is generally set such that the potential $\Phi(r)\to 0$ as $r\to \infty$. However, when considering two arbitrary galaxies\footnote{the emitter galaxy and the receiver galaxy (Milky Way)}, connected by a single photon, it is a violation of energy conservation to set the respective integration constants to different values. We select a single universal value for the integration constant, taken to be zero, though a more physical choice may be found in future dark energy research. Physically, this means that at large $r$ the gravitational potentials go to small but non-zero values.
\subsection{The convolution function}
The LCM mapping term, $ v_{lcm}^2$, is composed
of three terms:
\begin{equation}
v_{lcm}^2= \kappa v_{1} v_{2},
\label{eq:convolutionFunc}
\end{equation}
where $\kappa$ is the curvature ratio, and $v_1$ and $v_2$ are successive Lorentz transformations.
We then normalize the mapping, $ v_{lcm}^2$, by its value as $r\to \infty$, $(v^2_{lcm})_\tau$, such that the relative curvature term in Eq.~\ref{eq:zonteLCM} is:
\begin{equation}
\tilde{v}^2_{lcm}=\frac{v^2_{lcm}}{(v^2_{lcm})_\tau}.
\label{eq:domain}
\end{equation}
\subsubsection{The LCM curvature ratio: $\kappa$}
\label{kappa}
In traditional Lorentz transformations all frames are symmetric with respect to the coordinate time $t=\cosh \xi$ for the rapidity angle, $\xi$. Since Lorentz transformation are used to interpret spectra with respect to the coordinate time of the Milky Way, we need a rule to re-express the coordinate time of the emitting galaxy in terms of that of the Milky Way.
This is done with the curvature ratio $\kappa$, which is based upon the idea of coordinate light speeds $\tilde{c} $. Fig.~\ref{kappaClocks} shows how time is affected by the underlying curvature of space.
Coordinate light speeds are a physical indicator of curvature; the degree to which
$\tilde{c}<c $ indicate the increase in path length due to curvature \citep{Narayan}. As viewed by an external observer at asymptotic infinity, who can not \emph{see} manifest curvature but has knowledge of the line-of-sight travel distance of the light, curvature is indicated by the difference between the vacuum light speed and the measured coordinate light speed,
$c-\tilde{c}(r)$.
Coordinate light speeds arise in the solution to the wave equation for light~\citep{Cisn}. In the Schwarzschild case, the
effective index of refraction, $n(r)$, relates the two light speeds:
\begin{equation}
n(r) \tilde{c }=\left(\frac{1}{\sqrt{-g_{tt}} }\right)_r \tilde{c}=c,
\label{eq:index}
\end{equation}
by the gravitational redshift (Eq.~\ref{eq:Clone}).
The ratio $\kappa$ scales the coordinate time of the emitter galaxy relative to that of the Milky Way in terms of deviations from flatness as a function of radius:
\begin{equation}
\kappa(r)=\frac{c-\tilde{c}_{gal}(r)}{c-\tilde{c}_{mw}(r)},
\label{eq:kappa}
\end{equation}
where $\tilde{c}_{gal} (r)$ and $\tilde{c}_{mw} (r)$ are the respective coordinate light speeds of the emitter and receiver galaxies. The value of the ratio at the limit of the data, $\kappa_\tau$, approximates the value when $r\to \infty$ when all the luminous mass is enclosed.
\begin{figure}
\includegraphics[scale=0.25]{kappaClocks}
\caption{A cartoon of emitter and receiver galaxies. The vertical arrows represent the unit times at each point in the respective galaxy. The relative length of the unit time is indicated by the length of the arrow; a longer unit time indicates that time runs slower. Unit times within the enclosed mass density are longer than those outside, due to the mass dependence of the time coefficient of the space-time metric. $\kappa(r)$ is proportional to the ratio of the arrow lengths at radius $r$, because the observed frequencies are measured with respect to the receiver's coordinate time. \label{kappaClocks}}
\end{figure}
\subsubsection{Equivalent Lorentz Doppler-shift formula}
The most general form of the Lorentz transformation is the exponential mapping:
\begin{equation}
\Lambda= e^{\chi}=\sum_{n=0}^{\infty} \frac{\chi^n}{n!}
\end{equation}
where $\chi=-\xi S$ is the product of the rapidity angle $\xi$ and the generator of the rotation $S$. The rapidity
angle defines the relationship between two frames in the hyperbolic space-time of Special Relativity.
The Doppler-shift formula in Eq.~\ref{eq:dataLorentz} comes from such a Lorentz transformation.
In hyperbolic form, the Lorentz transformation matrix for a boost in the $x$ direction is
\begin{align}
\Lambda_x&=
\left( \begin{array}{cccc}
\cosh \xi &-\sinh \xi & 0&0 \\
-\sinh \xi & \cosh \xi & 0&0\\
0&0& 1&0\\
0&0& 0&1
\end{array} \right) \\
&= exp \left[- \xi \left(\begin{array}{cccc}
0&1 & 0&0 \\
1 & 0 & 0&0\\
0&0& 0&0\\
0&0& 0&0
\end{array} \right)\right] .
\end{align}
This form yields the Lorentz Doppler-shift formula by rotating a photon's 4-vector $(\omega, k_i)$ through $\xi$, where $k_i=2\pi/\lambda_i$, $\lambda$ is the wavelength, and the indices of the spatial basis are $i=1,2,3$. The hyperbolic form of the Lorentz
Doppler-shift formula for line of sight photons is:
\begin{equation}
\frac{v}{c}= \tanh \xi= \frac{e^\xi - e^{-\xi}}{e^\xi + e^{-\xi}}.
\label{eq:LorentzDefine}
\end{equation}
We can extend this formalism to the mapping of the small relative curvatures between a pair of emitter
and receiver galaxies.
The resulting parameter $v$ from Eq.~\ref{eq:LorentzDefine} will relate the emitter to receiver galaxy, and should not to be confused with a physical speed.
To generalize the Doppler formula as given in Eq.~\ref{eq:LorentzDefine}, we identify the mapping factor $e^\xi$ with the ratio of the received to emitted frequencies, $\omega_s/\omega_o$. In Special Relativity the two frames are symmetric, so it is meaningless to pin frequencies to specific frames.
However, transitioning to gently curved frames it becomes necessary to consistently identify the mapping factor.
In what follows, we always identify the numerator with the receiver's reference frame
and the denominator with the emitter's reference frame:
\begin{equation}
e^\xi(r)=\omega_{\rm{receiver}}(r)/\omega_{\rm{emitter}}(r).
\label{eq:array}
\end{equation}
\subsubsection{ Mapping $v_1$ }
The first LCM term, $v_1$, looks at the gravitational redshift frequencies $\omega(r)$ (Eq.~\ref{eq:Clone}) and rephrases them in terms of an equivalent Doppler-shift formula via Lorentz transformations. This term is a 2-frame map of the emitting and receiving galaxies, created by
their respective gravitational potentials from luminous mass.
By the convention in Eq.~\ref{eq:array}, the curved 2-frame
mapping factor is:
\begin{equation}
e^{\xi_{c}}(r)=\frac{\omega_{mw}(r)}{\omega_{gal}(r)},
\label{eq:specific}
\end{equation}
for the redshift frequencies of the emitter galaxy, $\omega_{gal}(r)$, and the receiver galaxy, $\omega_{mw}(r)$.
Consistent with the form of Eq.~\ref{eq:LorentzDefine}, the first LCM mapping is then
\begin{equation}
\frac{v_{1}}{c}=\frac{ e^{\xi_{c}} - e^{-\xi_{c}}}{e^{\xi_{c}}+ e^{-\xi_{c}}}.
\label{eq:prime1}
\end{equation}
All quantities are functions of radius except for the vacuum light speed. Fig.~\ref{v_1graphic} gives a visual representation of such a mapping.
\subsubsection{Mapping $v_2$ }
The second LCM term, $v_2$, looks at the requisite transformation from the curved 2-frame (indicated by Eq.~\ref{eq:specific}) to the flat 2-frame where physical measurements
are made. Fig.~\ref{v_2graphic} gives a visual of such a mapping. The term $v_2$ is also phrased in terms of an equivalent Doppler-shift via Lorentz transformations. That observations are always made in flat frames is demonstrated by the constancy of the local speed of
light as measured by all observers.
We define the flat frames by those frequencies which wouldhave been measured if the Keplerian rotation curve velocities $v_{l}(r)$ from the luminous mass in Eq.~\ref{eq:quadrature} were observed. Keplerian velocities are calculated in a purely Newtonian context and so describe our best understanding of the absence of all curvature. The shifted frequencies $\omega_{l}(r)$ are defined by:
\begin{equation}
\frac{v_{l}(r)}{c}=
\frac{
\frac{\omega_{l}(r) }{\omega_{o}}
- \frac{\omega_{o} }{\omega_{l}(r)}}{
\frac{\omega_{l}(r)}{\omega_{o}}
+ \frac{\omega_{o} }{\omega_{l}(r)}} .
\label{eq:MPflat}
\end{equation}
for the characteristic frequency $\omega_o$ defined in Eq.~\ref{eq:dataLorentz}.
Consistent with in Eq.~\ref{eq:array}, the flat 2-frame mapping factor is:
\begin{equation}
e^{\xi_{f}}(r) = \frac{\omega_{l} (r)}{\omega_{o}}.
\label{eq:flatflat}
\end{equation}
The $v_2$ mapping involves four frames instead of two, since the two curved frames and the two flat frames have already been mapped onto each other. To generalize the mapping factor used in the standard Lorentz boost formula (Eq.~\ref{eq:array}) for the $v_2$ mapping factor, we write the ratio of the receiver 2-frame (flat) to emitter 2-frame (curved) as a square:
\begin{equation}
(e^{ \xi_2} )^2= \frac{e^{\xi_{f}}}{ e^{\xi_{c}} }
\label{eq:FCFtwo}
\end{equation}
where $e^{\xi_{f}}$ and $ e^{\xi_{c}}$ are functions of $r$.
Before we write the final form of the 4-frame map, we consider the idea of relative motion invoked in the Lorentz transformation. In Special Relativity, Lorentz boosts are always defined as positive rotations away from the rest frame, which is
aligned with the vertical time axis of the light cone (see Fig.~\ref{lightconegraphic}).
Since all frames are symmetric in Special Relativity, specific identification of which frame is the rest frame is meaningless.
However in the LCM we have included small deviations from flatness, and so we identify the moving frame with the
curved 2-frame mapping factor $e^{\xi_{c}}$ and the rest frame with the flat 2-frame mapping factor $e^{\xi_{f}}$. This identification is made based upon
the respective clocks (i.e. the coordinate times in each map), which reflect our concepts of the moving and rest frames.
Once this identification is made, we note that we are transforming from the moving frame
to the rest frame --- effectively a
reverse boost. We characterize the reverse boost by the reciprocal of the Lorentz transformation in Eq.~\ref{eq:LorentzDefine}:
\begin{equation}
\frac{v_2 }{c}= \coth \xi_2= \frac{e^{\xi_2 } +e^{-\xi_2 }}{e^{\xi_2 } - e^{-\xi_2 }}.
\label{eq:hyperbolicreference}
\end{equation}
Re-written for algebraic convenience, to accommodate a 4-frame map, the final
LCM mapping is:
\begin{equation}
\frac{v_2 }{c}= \frac{e^{ 2\xi_2 }+1}{e^{ 2\xi_2 } - 1},
\label{eq:hyperbolico}
\end{equation}
where all quantities are functions of radius except for the vacuum speed of light and the characteristic frequency $\omega_o$. \\
\begin{figure}
\includegraphics[scale=0.25]{v_1graphic}
\caption{The respective small curvatures of two galaxies are mapped from emitting galaxy to receiving galaxy (Milky Way), as a function of radius. This is done with $v_1$ using Lorentz-group kinematics.\label{v_1graphic}}
\includegraphics[scale=0.25]{v_2graphic}
\caption{The final step in adjusting the received frequency shifts to the frame of the observer is mapping the curved 2-frame onto the flat 2-frame where observations are made. This conversion is done with $v_2$ as a function of radius, using Lorentz group kinematics. \label{v_2graphic}}
\end{figure}
\begin{figure}
\includegraphics[scale=0.25]{lightconegraphic}
\caption{The identification of the rest frame and the moving frame in the light cone. \label{lightconegraphic}}
\end{figure}
\section[]{ LCM Sample and Results}
\label{sec:RESULTS}
In this section we describe
the four Milky Way luminous mass distributions (section~\ref{MW}), the rotation curve fitting protocol (section~\ref{fitting}), the sample of emitter galaxies (section~\ref{sample}),
and the results of the LCM fits (section~\ref{results}).
\subsection{Milky Way luminous mass models}
\label{MW}
As previously describe, in the LCM we map each emitter galaxy onto the receiver galaxy (i.e. the Milky Way) to derive the relative curvature term $v_{lcm}$. The luminous matter profile of the Milky Way is notoriously difficult to determine because we are observing the system from the inside (e.g. interpreting H\,{\sevensize\bf I} outside the solar radius is particularly difficult~\citep{Car}). We compare each of the emitter galaxies to four $(4)$ different Milky Way luminous mass models to accommodate for the variations in the Milky Way luminous mass modeling.
The four models differ in key features: inner rise rate, asymptotic velocity at large $r$, and coverage. Mass components and total coverage are described in Table~\ref{tab:MWlum} and the total luminous profiles of each Milky Way model are plotted in Fig.~\ref{MWlum}, in terms of the resulting orbital velocities for a test particle.
The Milky Ways from ~\citet{Sofue} and ~\citet{Xue} differ in the inner rise rates of the orbital velocities, and the Keplerian fall-offs at large $r$ differ only in magnitude. The Sofue model has the largest luminous mass of all four Milky Ways. The two Milky Ways from ~\citet{Klypin} (models A and B) differ in that model B includes angular momentum sharing for in-falling baryons to the disk, and so the Keplerian rotation curve from model B is globally higher than model A.
\begin{table}
\centering
\begin{minipage}{140mm}
\caption{Milky Way Luminous Mass Models \label{tab:MWlum}}
\begin{tabular}{@{}llcc@{}}
\hline
Galaxy &coverage &$M_{bulge} $&$M_{disk} $\\
\hline
Sofue & 30 kpc & $1.8$ &$6.8$ \\
Xue &60 kpc &$1.5$ &$ 5.0$\\
Klypin, A &15 kpc & $0.8$ &$4.0$ \\
Klypin, B &15 kpc &$1.0$ &$5.0$\\
\hline
\end{tabular}\\
Masses in units of $M_{\sun} \times 10^{10}$
\end{minipage}
\end{table}
\begin{figure}
\includegraphics[scale=0.45]{MWlum}
\caption{The Milky Way orbital velocities due to the luminous mass profiles reported in each of the four references used in this work.\label{MWlum}}
\end{figure}
\subsection{LCM rotation curve fitting protocol}
\label{fitting}
In this paper we fit the predicted LCM rotation curve velocity $v_{rot}$ (Eq.\ref{eq:zonteLCM}) to the reported rotation curve data $v_{obs}$ (Eq.~\ref{eq:dataLorentz}). The LCM prediction and fit are calculated using the MINUIT minimization software as implemented in the ROOT data-analysis package~\citep{ROOT}.
The fit procedure is as follows:
\begin{enumerate}
\item The luminous mass components reported in each reference are digitized using the software package Graph Click \citep{Graphclick};
\item The associated Newtonian gravitational potential $\Phi$ is calculated for each component and the components are summed (Eq~\ref{eq:potentialgeneral});
\item We parametrize the Schwarzschild metric as a function of radius and calculate the convolution function, $v_{lcm}$, for the selected Milky Way mass model, (Eq.~\ref{eq:domain})
\item The minimization procedure explores the parameter space to find an optimal fit of $v_{rot}$ to $v_{obs}$; the parameter space includes the stellar disk and bulge M/L constrained (separately) to be between $0.00$ and $5.00$, reported gas fraction constrained between $0.75$ and $1.5$, and $\tilde{a}$ free;
\item The best-fit values for: M/L (bulge and disk), gas scalings, $\tilde{a}$, and the associated $\chi^2$ values for the fit are reported in Tables~\ref{sumRESULTS} and \ref{sumRESULTSbigger}.
\end{enumerate}
The initial luminous mass profiles used in the LCM fits come from four different fitting models: NFW, ISO, MOND and Maximum Light (MaxLight). In all of these model fits M/L was free, with the notable exception of those galaxies sampled from Ref.~8 \citep{Blok}.
Comparisons between the original fitting models' reported M/L and the $\chi^2$ values to those of the LCM fits are reported in Table~\ref{sumRESULTS}. The NFW, ISO, MOND or MaxLight $\chi^2$ values reported here have been recalculated using the same MINUIT software used for the LCM fits so that the $\chi^2$ values can be compared appropriately.
The ranges used to constrain M/L are motivated by significant variations for a given galaxy's luminous mass profile, even when fit reports for the M/L are done with the same model, see Fig.~\ref{fig:massmodels18}. Reported gas profiles endure less variation than stellar M/L, but still vary significantly for a single system~\citep{Frat,Blok}. The M/L reported in the literature, which are used as the LCM starting mass estimates, are reported in Table~\ref{sumRESULTS}.
\subsubsection{Error estimates and reduced $\chi^2_r$ values}
All figures reported here indicate the uncertainties as reported in the literature.
There is currently no standard practice as to how to quantify the uncertainties associated with rotation curve data ~\citep{Gent,Blok,JNav,San96}, such that resulting $\chi^2_r$ values can not be interpreted in a uniform way. Often the uncertainty comes from a fit of the tilted ring model to the H\,{\sevensize\bf I} velocity field. In other analyses uncertainty estimates come from differences between the approaching- and receding-side velocity fields ~\citep{Blok,Gent,Toky} .
\subsection{The sample}
\label{sample}
The LCM sample reported in this paper represents twenty-three $(23)$ galaxies, in forty-two $(42)$ different data sets. The galaxies were randomly selected, with the only selection criteria being some effort to represent both galaxies smaller and larger than the Milky Way and to compare different model-based assumptions of the luminous mass profiles. We report the different galaxies in roughly decreasing order of size. In Table~\ref{sumRESULTS} average results (over the four $(4)$ different Milky Way luminous mass models) are reported as well as the original model contexts, M/L values, associated wavelength bands, and $\chi^2$ values. The $\chi^2$ values reported here for the originating models (NFW, MOND, ISO or MaxLight, from which we use the reported stellar and gas profiles as starting estimates for LCM luminous profiles) have been calculated using the same points and in the same fitting code as the associated LCM fits.
In sections~\ref{Blok}-\ref{Bottema} we describe some features and basic assumptions from some of the references for the emitter galaxies.
\begin{table*}
\centering
\begin{minipage}{140mm}
\caption{Summary of results and M/L for originating model \& LCM \label{sumRESULTS}}
\begin{tabular}{@{}lllcccccccc@{}}
\hline
Galaxy &Ref.~& Band M/L &Model & \multicolumn{3}{c}{\underline{ Model Fit Results}} & & \multicolumn{3}{c}{\underline{LCM Fit Results (average)}} \\
\hline
& & & & $\rmn{M/L}_{disk}$ &$\rmn{M/L}_{bulge}$ & $\chi^2_r$ & &$\rmn{M/L}_{disk}$&$\rmn{M/L}_{bulge}$&$\chi^2_r$ \\
\hline
F 563-1 &2 &I &NFW &2.3 &-- &0.05& &12.92 &-- &0.16 \\
F 563-1 &13 &B & NFW &2.6 &-- &0.14& &4.40 &-- &0.22 \\
M 31 & 12 &B &ISO &7.5 &3.5 &0.36 & &13.0 &2.17 &0.30 \\
M 31 & 15 &R &NFW &0.93 &3.0 &13.45& &1.26 &2.10 &10.6 \\
M 33 & 5 & K & NFW &0.70 & -- &2.46& &0.46 &-- & 0.84 \\
M 33 &6 & B &MIdm$^{a}$&0.80 & -- &0.16& & 0.70 &-- &0.18 \\
NGC 891* &11 & 3.6$\,\umu$m&MOND & 0.5 & 2.0 &5.30 & &0.92 & 1.40 &1.06 \\
NGC 891 & 11 &3.6$\,\umu$m &MaxLight&0.9 &1.64 &1.10 & & 1.03 &1.49 &1.15 \\
NGC 925* &3 &3.6$\,\umu$m &NFW &0.65 &-- &2.94& &0.36 &-- &6.45 \\
NGC 925* &3 &3.6$\,\umu$m &ISO &0.18 &-- &2.40& &0.003 &-- &3.81 \\
NGC 2403 & 3 & 3.6$\,\umu$m& NFW &0.41 & -- &4.56& &0.41 &-- & 0.88 \\
NGC 2403 &9 & B & MOND &1.60 & -- &3.67 & &0.99 &-- & 2.30 \\
NGC 2841 &3 & 3.6$\,\umu$m& NFW &0.74 & 0.84 &0.45 & & 0.99 &1.29 &0.44 \\
NGC 2841*& 8 &3.6$\,\umu$m&MOND &0.89 &1.04 &0.87& & 0.84 &1.09 & 0.43\\
NGC 2841 &9 & B & MOND &8.3 & 0.83 &34.7& & 5.10 &2.91 & 1.94 \\
NGC 2903 &3 & 3.6$\,\umu$m&NFW &0.61 &1.3 &0.55& & 2.07 &0.65 &0.75\\
NGC 2903 &8 &3.6$\,\umu$m&MOND &1.71 & -- & 0.61& & 2.03 &-- &0.76\\
NGC 2903 &10 &B &MOND &3.60 & -- &10.71& & 2.94 &-- &2.10\\
NGC 3198 & 1 &r &NFW & 3.8 &-- & 1.34 & &3.26 & -- &1.65 \\
NGC 3198 & 2 &r &NFW & 1.4 &-- & 0.67 & &2.87 & -- &1.38 \\
NGC 3198 & 3 &3.6$\,\umu$m&NFW &0.80 & -- &5.40 & &0.64 & -- &4.61 \\
NGC 3198 & 4 &3.6$\,\umu$m&MOND & 0.33 & -- &3.50 & & 0.61 & -- &3.39 \\
NGC 3521 & 8 &3.6$\,\umu$m &MOND &0.71 &-- &0.97 & & 0.72 &-- &0.80 \\
NGC 3726 & 10 &B &MOND &1.00 &-- &3.57& &1.15 &-- &3.12 \\
NGC 3953 & 10 &B &MOND &2.7 &-- &1.35& &1.72 &-- &0.62 \\
NGC 3992 & 10 &B &MOND &4.93 &-- &0.50& &4.04 &-- &0.37 \\
NGC 4088 & 10 &B &MOND &1.16 &-- &1.70& &1.64 &-- &1.39 \\
NGC 4138 & 10 &B &MOND &3.5 &-- &2.12& &3.21 &-- &2.34 \\
NGC 5055 & 3 &3.6$\,\umu$m & NFW &0.79 &0.11 &17.23& &0.49 &0.44 &2.85 \\
NGC 5055 &7 &B & NFW &3.20 & -- & 3.27 & & 2.51 &-- &2.55 \\
NGC 5055 & 8 &3.6$\,\umu$m & MOND &0.55 &0.56 &1.12& & 0.56 &0.36 &1.16\\
NGC 5533 & 10 &B &MOND &0.6 &7.2 &1.57 & &3.00 &6.67 &0.73 \\
NGC 5907 & 10 &B &MOND &1.6 &6.8 &0.44& &1.41 &4.29 &0.24 \\
NGC 6946 & 3 & 3.6$\,\umu$m&NFW &0.64 & 1.00 & 4.34 & & 0.35 &0.66 & 1.42 \\
NGC 6946 & 10 & B & MOND & 0.5 & -- & 3.03& &0.70 &-- & 0.46 \\
NGC 7331 &8 & 3.6$\,\umu$m& MOND &0.4 &1.22 & 0.45& & 0.40 &0.95 &0.27 \\
NGC 7331*& 9 & B & MOND &2.0 & 1.8 &9.32 & & 1.41 &1.29 &4.47 \\
NGC 7793 &8 &3.6$\,\umu$m &MOND &0.28 &-- &4.61& &0.13 &-- &7.99 \\
NGC 7793*&14 &B &ISO &2.6 &-- &1.08& &2.55 &-- &0.59 \\
NGC 7814 &11 & 3.6$\,\umu$m&ISO & 0.68 &0.71 & 0.25& &3.17 &0.54 &0.44 \\
NGC 7814* & 11 & 3.6$\,\umu$m &MaxLight&9.25 &0.64 &6.11 & & 3.81 &0.56 & 0.50 \\
UGC 6973 & 10 &B &MOND &2.7 &-- &23.5 & &1.49 &-- &0.07 \\
\hline
\end{tabular}
{\bf Notes:}\\
Values in this table are averaged over the four emitter-receiver (Milky Way) galaxy pairings. Galaxies for which one LCM fit-pair failed is indicated by (*). MIdm$^{a}$ is a model independent dark matter fit.
Mass-to-light ratios in units of $M_{\sun} / L_{\sun}$. Reduced $\chi^2$ per degree of freedom indicated by $ \chi^2_r$. \\
{\bf References:}
1.~\citet{Bege}, 2.~\citet{JNav}, 3.~\citet{Blok} , 4.~\citet{Maria},
5.~\citet{Cor03}, 6.~\cite{CoSa00}, 7.~\citet{Batt}, 8.~\citet{Gent}, 9.~\citet{Bot}, 10.~\citet{SanMcGa},
11.~\citet{Frat}, 12.~\citet{Car}, 13.~\citet{giraud2000universal}, 14.~\citet{Dicaire}, 15.~\citet{Klypin}. \\
\end{minipage}
\end{table*}
\subsubsection{Ref.~ 3 parameters ~\citet{Blok}, NFW }
\label{Blok}
From Ref.~3 ~\citep{Blok} we sample seven galaxies: NGC 3198, NGC 5055, NGC 2403, NGC 2841, NGC 6946, NGC 2903 and NGC 925. These galaxies are reported in the context of
NFW fits with fixed M/L, consistent with the PSM of ~\citet{BelldYong} with a
Salpeter ~\citep{Salpeter} initial mass function. For NGC 925 we sample both the NFW and the ISO fits, both with M/L fixed.
The Ref.~3 rotation curve data comes from the
H\,{\sevensize\bf I}\, THINGS survey ~\citep{Walter}. The THINGS data is generally considered to be the highest resolution rotation curve data available. The Ref.~3 sample is chosen based on favorable inclination angles and regular rotations. Rotation curve derivations are based on tilted-ring rotation curve analysis, fit with third-order Gauss-Hermite polynomials. Gas estimates are based on a thin gas disk geometry from the H\,{\sevensize\bf I}\, flux maps reported by ~\citet{Walter}, scaled by $1.4$ to account for helium and metals. Molecular gas is included as a slight increase in the M/L of the stellar disk. Stellar M/L are reported in the $3.6\,\umu$m band to avoid the uncertainties prevalent in B-band from hot dust and young stars. Ref.~3 decomposes the stellar light into a bulge and disk with a double exponential disk model, avoiding the canonical $R^{1/4}$ bulge distribution.
\subsubsection{ Ref.~8 parameters ~\citet{Gent}, MOND}
\label{TaM}
From Ref.~8~\citep{Gent} we sample six galaxies: NGC 2903, NGC 2841, NGC 5055, NGC 7331, NGC 7793 \& NGC 3521. These galaxies are reported in the context of
MOND fits and all galaxy rotation curves come from Ref.~3~\citep{Blok}. Ref.~8 uses the Ref.~3 geometry and gas assumptions (above), with stellar M/L of the disk and bulge free, distances constrained, and the $\umu$ ``simple'' interpolating function. They report M/L in the $3.6\,\umu$m band.
Of those galaxies sampled here, they report that all are consistent with the PSMs of ~\citet{BelldYong} with the exception of NGC 2903 and NGC 7331. As evidenced in Fig.~\ref{fig:massmodels18}, while NGC 2841 velocities associated with the luminous profiles reported in Ref.~3 and Ref.~8 differ by approximately $30\, \rmn{km\, s^{-1}}$ and those for NGC 5055 by about $60\, \rmn{km\, s^{-1}}$, both sets of luminous models are reported as consistent with the same PSMs, demonstrating the degree of freedom within current PSMs.
\subsubsection{Ref.~10 parameters ~\citet{SanMcGa}, MOND}
\label{SanMc02}
From Ref.~10~\citep{SanMcGa} we sample ten galaxies: NGC 5533, NGC 4138, NGC 5907, NGC 3992, NGC 2903, NGC 3953, UGC 6973, NGC 4088, NGC 3726, and NGC 6946. These galaxies are
reported in the context of MOND fits, at Ursa Major distances. In Ref.~10 the distance to the Ursa Major system is reported as $15.5$ Mpc, but Cepheid based recalibration indicates $18.6$ Mpc~\citep{Tully}.
They model the luminous mass with thin stellar and gaseous disks and a spheroidal stellar bulge, for M/L free. Most systems are treated with coplanar circular motions, but, when appropriate, complexities for bars and warping of the gas layer are modeled. M/L is reported in the B-band. Neutral hydrogen is scaled by $1.4$ to account for primordial helium.
\subsubsection{Ref.~9 parameters ~\citet{Bot}, MOND}
\label{Bottema}
From ~\citet{Bot} we sample three galaxies: NGC 2841, NGC 7331, and NGC 2403. These galaxies are reported in the context of MOND fits at Cepheid distances. The only geometry assumption they report is a thing stellar disk. M/L is free and reported in the B-band. Ref.~9 gas is reported from a surface density distribution equal to the observed H\,{\sevensize\bf I} scaled by a factor of $1.3$ to account for primordial helium. The contribution of the gas to the rotation curve is fixed and does not depend on the distance to the galaxy.
\subsection{Results}
\label{results}
In this section we make notes on individual galaxies and groups of galaxies with interesting features that constrain PSM and imply an interpretation of the LCM free parameter $\tilde{a}$.
In Table~\ref{sumRESULTS} we report the LCM average results for M/L and reduced $\chi^2_r$ values from the LCM fits plus a comparison to fits using the originating models (NFW, ISO, MOND, or MaxLight).
Individual emitter galaxy fit results for each Milky Way luminous mass model are reported in Table~\ref{sumRESULTSbigger}. LCM fits for seven ($7$) out of $165$ emitter-receiver galaxy pairings failed; these failures indicate problems with the initial luminous mass profiles.
Figs.~\ref{fig:resultssmall}-\ref{galaxiesSmallest} show LCM fits using the ~\citet{Xue} Milky Way. We selected the Xue Milky Way for the figures because it has the largest radial coverage. In the one case where the LCM fit with the Xue Milky Way failed, we instead show the ~\citet{Sofue} Milky Way pairing, as indicated in the figure caption.
The distribution of values for $\tilde{a}$,
has a mean value of $\tilde{a}=0.09\pm0.05$ and does not exhibit an obvious dependence on the relative curvature ratio, $\kappa_\tau$, as indicated in Fig.~\ref{fig:alphaDistrib1}. In section~\ref{n891n7814} we present a conjecture regarding a possible physical correlation between $\tilde{a}$ and density gradients in the luminous matter.
\begin{figure}
\includegraphics[scale=0.45]{tilde_a_Kap.pdf}
\caption{Free parameter $\tilde{a}$ versus $\kappa_\tau$. Each dot represents one
emitter/receiver galaxy pair. Mean value for the sample presented here $\tilde{a}=0.09\pm0.05$. \label{fig:alphaDistrib1}}
\end{figure}
\begin{table*}
\centering
\begin{minipage}{160mm}
\caption{LCM emitter-galaxy pairs Results \label{sumRESULTSbigger}}
\begin{tabular}{@{}llcccccc@{}}
\hline
emitter&receiver Milky Way &$\kappa_\tau$&$\tilde{a}$ &gas scaling & $ \rmn{M/L}_{disk}$&$ \rmn{M/L}_{bulge}$&$\chi^2_{R}$ \\
\hline
F 563-1 Ref.~2 &Sofue &0.04 &0.03 &1.5 &9.59 &-- &0.15 \\
&Klypin A &0.07 &0.06 &1.5 &9.37 &-- &0.16\\
&Klypin B &0.08 &0.05 &1.5 &9.45 &-- &0.16 \\
&Xue &0.10 &0.07 &1.5 &9.33 &-- &0.16 \\
F 563-1 Ref.~13 &Sofue &0.05 &0.02 &1.5 &4.48 &-- &0.21\\
&Klypin A &.10 &0.04 &1.5 &4.36 &-- &0.22\\
&Klypin B &0.09 &0.04 &1.5 &4.40 &-- &0.22\\
&Xue &0.11 &0.05 &1.5 &4.34 &-- &0.22 \\
M 31 Ref.~12 &Sofue &0.89 &0.02 & 0.75 & 8.98 &1.88 &0.30 \\
&Klypin A &2.16 &-0.07 &0.75 &17.90 &1.94 &0.28 \\
&Klypin B &2.16 &-0.06 &0.75 &17.67 &2.22 &0.28 \\
&Xue &2.59 &0.04 &0.75 &7.41 &2.63 &0.33 \\
M 31 Ref.~15 &Sofue &0.94 &0.10 &-- &0.55 &3.18 &9.52 \\
&Klypin A &1.88 &0.04 &-- &2.02 &2.44 &13.8 \\
&Klypin B &1.10 &0.13 &-- &1.61 &1.39 &9.48 \\
&Xue &1.43 &0.13 &-- &0.85 &1.38 &9.53\\
M 33 Ref.~5 &Sofue &0.04 &0.04 & 1.5 &0.47 &-- &0.78\\
&Klypin A &0.09 &0.09 & 1.5 &0.45 &-- &0.87 \\
&Klypin B &0.08 &0.08 & 1.5 &0.46 &-- &0.87\\
&Xue &0.10 &0.10 & 1.5 &0.46 &-- &0.83\\
M 33 Ref.~6 &Sofue &0.04 &0.05 &0. 89 &0.72 &-- &0.19\\
&Klypin A &0.07 &0.10 &1.00 &0.70 &-- &0.18\\
&Klypin B &0.07 &0.09 &1.02 &0.70 &-- &0.18\\
&Xue &0.09 &0.12 &1.05 & 0.69 &-- &0.18\\
NGC 891 Ref.~11 \footnote{Citation's M/L assumptions from MOND fit}
&Sofue & 0.65 &0.02 &0.75 &0.86 &1.24 &0.95 \\
&Klypin A* &-- & -- & -- & -- &-- &-- \\
&Klypin B &1.36 &0.01 &0.75 &0.93 &1.40 &1.06\\
&Xue &1.73 &0.01 &0.75 &0.92 &1.39 &1.06 \\
NGC 891 Ref.~11 \footnote{Citation's M/L assumptions from MaxLight fit}
&Sofue & 0.79 &0.01 &0.75 &0.98 &1.37 &1.08\\
&Klypin A &1.78 &0.01 &0.75 &1.04 &1.55 &1.16\\
&Klypin B &1.62 &0.01 &0.75 &1.05 &1.53 & 1.17\\
&Xue &2.08 &0.01 &0.75 &1.04 &1.53 &1.17 \\
NGC 925 Ref.~3\footnote{original citation M/L assumptions: NFW} &Sofue*
&-- &-- &-- &-- &-- &--\\
&Klypin A &0.03 & 0.04 & 1.5 & 0.36 &-- &6.44 \\
&Klypin B &0.03 &0.04 &1.5 &0.37 &-- &6.43\\
&Xue &0.04 &0.05 &1.5 & 0.36 &-- &6.48 \\
NGC 925 Ref.~3\footnote{original citation M/L assumptions: ISO} &Sofue* &--
&-- &-- &-- &-- &--\\
&Klypin A &0.01 &0.07 & 1.5 &0.003 &-- &3.80 \\
&Klypin B &0.01 &0.07 &1.5 &0.003 &-- &3.80\\
&Xue &0.01 &0.09 &1.5 &0.003 &-- &3.83 \\
NGC 2403 Ref.~3&Sofue &0.06 &0.05 & 1.5 &0. 42 &-- &0.87\\
&Klypin A &0.11 &0.10 &1.5 &0.40 &-- &0.88 \\
&Klypin B &0.10 &0.10 &1.5 & 0.41 &-- &0.88\\
&Xue &0.13 &0.12 &1.5 & 0.41 &-- &0.88\\
NGC 2403 Ref.~9&Sofue & 0.07 &0.05 & 1.5 & 1.01 &-- &2.24\\
&Klypin A &0.15 &0.09 & 1.5 & 0.98 &-- &2.30\\
&Klypin B &0.14 &0.09 & 1.5 &0. 99 &-- &2.30\\
&Xue &0.17 &0.11 & 1.5 & 0.98 &-- &2.35 \\
\hline
\end{tabular}\\
References as in Table~\ref{sumRESULTS}. Mass-to-light ratios in units of $M_{\sun} / L_{\sun}$.
Milky Way: ~\citet{Sofue}, ~\citet{Klypin}, \& ~\citet{Xue}. The quantity $\kappa_\tau$ is defined in section~\ref{kappa}.
(*) indicates an LCM fit which failed.
\end{minipage}
\end{table*}
\begin{table*}
\addtocounter{table}{-1}
\centering
\begin{minipage}{160mm}
\caption{Continued: LCM emitter-receiver galaxy pairs}
\begin{tabular}{@{}llcccccc@{}}
\hline
emitter &receiver Milky Way&$\kappa_\tau$&$\tilde{a}$ &gas scaling & $ \rmn{M/L}_{disk}$&$\rmn{M/L}_{bulge}$&$\chi^2_{R}$ \\
\hline
NGC 2841 Ref.~3 &Sofue &1.13 &0.08 &0.75 & 1.40 & 2.50 &0.48\\
&Klypin A &0.73 &0.26 &1.5 & 0.63 & 0.60 &0.19 \\
&Klypin B &0.73 &0.23 &1.5 & 0.67 & 0.69 &0.16\\
&Xue &1.91 &0.16 &1.5 & 1.24 & 1.36 &0.96\\
NGC 2841 Ref.~8 &Sofue & 0.51 &0.13 &1.50 &0.76 &1.12 &0.10\\
&Klypin A* &-- &-- &-- &-- &-- &-- \\
&Klypin B &0.67 &0.23 &1.50 &0.55 &0.95 &0.10\\
&Xue &1.74 &0.17 &1.50 &1.20 &1.21 &1.10\\
NGC 2841 Ref.~ 9 &Sofue &0.58 &0.11 &1.50 & 3.73 &2.56 &0.28 \\
&Klypin A &1.46 &0.06 &0.75 & 5.46 &3.13 &1.28 \\
&Klypin B &1.36 &0.06 &0.75 & 5.53 &3.17 &1.36\\
&Xue &1.91 &0.16 &1.50 & 5.70 &2.80 &4.82\\
NGC 2903 Ref.~3 &Sofue &0.28 &0.07 & 1.5 &2.27 &0 &1.14\\
&Klypin A &0.49 &0.12 & 1.5 &2.00 &0 &0.72 \\
&Klypin B &0.46 &0.11 & 1.5 &2.04 &0 &0.74\\
&Xue &1.06 &0.10 & 0.75 &1.97 &2.57 &0.38\\
NGC 2903 Ref.~8&Sofue & 0.24 &0.07 & 1.5 &2.17 &-- &1.02\\
&Klypin A &0.41 &0.12 & 1.5 &1.94 &-- &0.53\\
&Klypin B &0.38 &0.12 & 1.5 &1.96 &-- &0.54\\
&Xue &0.53 &0.13 & 1.5 &2.07 &-- &0.95\\
NGC 2903 Ref.~10&Sofue &0.30 &0.07 & 1.5 &3.16 &-- &2.99 \\
&Klypin A &0.52 &0.12 & 1.5 & 2.78 &-- &1.00 \\
&Klypin B &0.49 &0.11 & 1.5 & 2.85 &-- &1.04\\
&Xue &0.65 &0.12 & 1.5 &2.95 &-- &3.39 \\
NGC 3198 Ref.~1&Sofue &0.15 &0.05 &1.5 &3.34 &--&1.45\\
&Klypin A &0.27 &0.08 &1.5 &3.21 &--&1.83 \\
&Klypin B &0.25 &0.07 &1.5 &3.24 &--&1.84 \\
&Xue &0.33 &0.09 &1.5 &3.24 &--&1.47\\
NGC 3198 Ref.~2&Sofue & 0.15 &0.05 &1.5 &2.91 &--&1.27 \\
&Klypin A &0.30 &0.08 &1.5 &2.84 &--&1.42\\
&Klypin B &0.27 &0.08 &1.5 &2.87 &--&1.43\\
&Xue &0.35 &0.10 &1.5 &2.84 &--&1.39 \\
NGC 3198 Ref.~3&Sofue &0.12 &0.05 &1.5 &0.73 &--&3.94\\
&Klypin A &0.16 &0.10 &1.5 &0.53 &--&5.44 \\
&Klypin B &0.15 &0.09 &1.5 &0.55 &--&5.35 \\
&Xue &0.29 &0.09 &1.5 &0.75 &--&3.69 \\
NGC 3198 Ref.~4&Sofue & 0.13 &0.04 &1.5 &0.65 &--&2.94 \\
&Klypin A &0.22 &0.08 &1.5 &0.57 &--&4.36\\
&Klypin B &0.20 &0.07 &1.5 &0.58 &--&4.31\\
&Xue &0.32 &0.10 &1.5 &0.64 &--&1.95 \\
NGC 3521 Ref.~8 & Sofue &0.42 &0.05 &1.50 &0.79 &-- &0.68\\
&Klypin A &0.74 &0.09 &1.50 &0.69 &-- &0.85\\
&Klypin B &0.72 &0.08 & 1.50 &0.74 &-- &0.86\\
&Xue &0.81 &0.11 &1.50 &0.65 &-- &0.83\\
NGC 3726 Ref.~10&Sofue &0.12 &0.06 &0.75 &1.15 &--&2.40 \\
&Klypin A &0.23 &0.08 &1.5 & 1.17 &--&3.86\\
&Klypin B &0.20 &0.08 &0.75 & 1.17 &--&3.46\\
&Xue &0.27 &0.12 &1.50 & 1.10 &--&2.76\\
NGC 3953 Ref.~10&Sofue & 0.20 &0.08 &0.75 & 1.86 &-- &0.63 \\
&Klypin A &0.37 &0.16 &0.75 & 1.68 &-- &0.61 \\
&Klypin B &0.35 &0.15 &0.75 & 1.73 &-- &0.62 \\
&Xue &0.42 &0.19 &0.75 & 1.63 &-- &0.61 \\
NGC 3992 Ref.~10&Sofue & 0.36 & 0.08 & 1.00 & 4.52 & -- &0.29 \\
&Klypin A &0.54 & 0.13 & 0.75 & 3.86 & -- &0.45\\
&Klypin B &0.51 & 0.12 & 0.75 & 3.99 & -- &0.48\\
&Xue &0.71 & 0.16 & 1.50 & 3.79 & -- &0.27\\
\hline
\end{tabular}\\
References as in Table~\ref{sumRESULTS}. Mass-to-light ratios in units of $M_{\sun} / L_{\sun}$. Milky Way: ~\citet{Sofue}, ~\citet{Klypin}, \& ~\citet{Xue}. The quantity $\kappa_\tau$ is defined in section~\ref{kappa}.
(*) indicates an LCM fit which failed.
\end{minipage}
\end{table*}
\begin{table*}
\addtocounter{table}{-1}
\centering
\begin{minipage}{160mm}
\caption{Continued:LCM emitter-receiver galaxy pairs}
\begin{tabular}{@{}llcccccc@{}}
\hline
emitter&receiver Milky Way &$\kappa_\tau$&$\tilde{a}$ &gas scaling & $ \rmn{M/L}_{disk}$&$ \rmn{M/L}_{bulge}$&$\chi^2_{R}$ \\
\hline
NGC 4088 Ref.~10&Sofue &0.17 &0.05 &0.75 &1.15 &-- &1.38 \\
&Klypin A &0.33 &0.09 &1.11 &1.10 &-- &1.39\\
&Klypin B &0.31 &0.08 &1.06 &1.12 &-- &1.40\\
&Xue &0.38 &0.10 &1.20 & 1.09 &-- &1.40\\
NGC 4138 Ref.~10&Sofue &0.22 &0.05 &1.50 &3.24 &-- &2.40 \\
&Klypin A &0.45 &0.08 &1.50 &3.20 &-- &2.26 \\
&Klypin B &0.41 &0.07 & 1.50 &3.23 &-- &2.34\\
&Xue &0.52 &0.09 & 1.50 &3.19 &-- &2.31\\
NGC 5055 Ref.~3 &Sofue & 0.34 &0.07 &0.90 & 0.39 &4.83 &1.46\\
&Klypin A &1.03 &0.004 &1.50 &0.66 &6.50 &5.21 \\
&Klypin B &0.55 &0.10 &0.75 &0.35 &4.40 & 1.87\\
&Xue* &-- & -- &-- &-- &-- &--\\
NGC 5055 Ref.~7 &Sofue &0.35 &0.05 &1.5 &2.77 &-- &2.93\\
&Klypin A &0.57 &0.10 &1.5 &2.31 &-- &1.76\\
&Klypin B &0.55 &0.09 &1.5 &2.43 &-- &1.68\\
&Xue &0.75 &0.11 &1.5 &2.51 &-- &3.83\\
NGC 5055 Ref.~8 &Sofue & 0.39 &0.07 & 1.5 &0.62 & 0.42 &1.26 \\
&Klypin A &0.63 &0.12 & 1.5 &0.50 & 0.33 &0.88 \\
&Klypin B &0.63 &0.11 & 1.5 &0.54 & 0.37 &0.96\\
&Xue &0.80 &0.12 & 1.5 &0.58 & 0.31 &1.53\\
NGC 5533 Ref.~10 &Xue &2.04 &0.08 &1.50 &3.0 &6.67 &0.73 \\
NGC 5907 Ref.~10&Sofue & 0.34 &0.08 &0.75 &1.54 &4.72 &0.20 \\
&Klypin A &0.57 &0.11 &0.75 & 1.34 &4.27 &0.30\\
&Klypin B &0.55 &0.10 &0.75 &1.39 &4.43 &0.29\\
&Xue &0.69 &0.15 &0.95 &1.37 &3.75 &0.19\\
NGC 6946 Ref.~3&Sofue &0.18 &0.08 &1.5 & 0.38 &0.70 &1.37\\
&Klypin A &0.34 &0.14 &1.5 & 0.34 & 0.65 &1.44 \\
&Klypin B &0.32 &0.13 &1.50 & 0.35 & 0.67 &1.42 \\
&Xue &0.39 &0.16 &1.50 & 0.33 & 0.64 &1.46 \\
NGC 6946 Ref.~10&Sofue &0.12 &0.06 &0.75 & 0.71 &-- &0.46 \\
&Klypin A &0.22 &0.08 &0.75 & 0.71 &-- &0.36\\
&Klypin B &0.20 &0.07 &0.75 & 0.72 &-- &0.38\\
&Xue &0.28 &0.10 &1.5 & 0.67 &-- &0.62 \\
NGC 7331 Ref.~8 &Sofue &0.40 &0.08 &0.75 &0.44 & 1.04 &0.34\\
&Klypin A &0.69 &0.12 &1.07 &0.41 &0.88 &0.17 \\
&Klypin B &0.67 &0.11 &0.75 &0.44 &0.88 &0.17 \\
&Xue &0.74 &0.18 &0.75 &0.31 &1.01 &0.40 \\
NGC 7331 Ref.~9 & Sofue &0.48 &0.08 &0.75 &1.90 & 0.25 &6.63 \\
&Klypin A*
&-- &-- &-- &-- &-- &--\\
&Klypin B &1.09 &0.11 &0.75 &0.21 &1.64 &1.50 \\
&Xue &1.48 &0.11 &0.75 &2.13 &1.32 &5.28 \\
NGC 7793 Ref.~8 &Sofue &0.02 &0.04 &0.75 & 0.12 &-- &7.93\\
&Klypin A &0.03 &0.09 &0.75 & 0.12 &-- &8.12 \\
&Klypin B &0.03 &0.08 &0.75 & 0.12 &-- &8.11 \\
&Xue &0.05 &0.09 &1.50 & 0.15 &-- &7.81 \\
NGC 7793 Ref.~14&Sofue* &-- &-- &-- &-- &-- &--\\
&Klypin A & 0.13 &0.02 &1.5 &2.55 &-- &0.56 \\
&Klypin B &0.12 &0.02 &1.5 &2.55 &-- &0.58 \\
&Xue &0.16 &0.03 &1.5 &2.55 &-- &0.60 \\
NGC 7814 Ref.~11\footnote{Citation's M/L assumptions from ISO fit }
&Sofue &0.77 &0.05 &0.75 & 3.32 &0.55 &0.23 \\
&Klypin A &1.67 &0.07 &0.75 & 2.81 & 0.60 &0.41 \\
&Klypin B &1.41 &0.07 &1.5 & 3.4 & 0.53 &0.82 \\
&Xue &1.70 &0.09 &0.75 &3.17 &0.51 &0.32\\
NGC 7814 Ref.~11 \footnote{Citation's M/L assumptions from MaxLight fit }
&Sofue* &-- &-- & -- & -- &-- &--\\
&Klypin A &1.71 &0.07 &0.75 &3.28 &0.61 &0.45\\
&Klypin B &1.46 &0.06 &1.50 &4.50 &0.54 &0.70\\
&Xue &1.78 &0.08 &0.75 &3.63 &0.53 &0.36\\
UGC 6973 Ref.~10&Sofue &0.15 &0.08 &0.75 &1.51 &-- & 0.07 \\
&Klypin A &0.32 &0.16 &0.75 &1.47 &-- &0.06\\
&Klypin B &0.31 &0.15 &0.91 &1.50 &-- &0.07 \\
&Xue &0.39 &0.18 &1.13 &1.48 &-- &0.07\\
\hline
\end{tabular}\\
References as in Table~\ref{sumRESULTS}. Mass-to-light ratios in units of $M_{\sun} / L_{\sun}$.
Milky Way: ~\citet{Sofue}, ~\citet{Klypin}, \& ~\citet{Xue}. The quantity $\kappa_\tau$ is defined in section~\ref{kappa}.
(*) indicates an LCM fit which failed.
\end{minipage}
\end{table*}
\subsubsection{M 31}
M 31 is an early-type spiral galaxy of large angular size, in our Local group.
The proximity of M 31 to the Milky Way presents difficulties in disentangling the gas kinematics of the two galaxies. It remains an open question if M 31 or the Milky Way is the largest member of the Local Group. There is general agreement on the M 31 rotation curve between $10$ and $30$ kpc ~\citep{Sof81}. The bulge of M 31 is thought to be an old stellar population which is almost twice as massive and more compact than that of the Milky Way ~\citep{Klypin}.
We compare rotation curves as reported in Ref.~12 ~\citep{Car} and Ref.~15~\citep{Klypin}.
Both Ref.~12 and Ref.~15 luminous profiles are from dark matter fits. The luminous matter profile in Ref.~12 is from an ISO fit and that in Ref.~15 is from an NFW fit with concentration taken from N-body simulation (consistent with $\Lambda\rmn{CDM}$).
Ref.~12 disentangles M 31 and Milky Way gas velocities by using only the approaching side kinematics of the rotation curve data reported in ~\citet{1985ApJ...299...59C}.
Both references use the stellar parameters as reported in ~\citep{walt87,walt88}.
Comparison of the Ref.~12 and Ref.~15 M/L results reflect the under-constrained nature of PSM, reflected in the two reported luminous profiles shown in Fig.~\ref{fig:massmodels18}.
The Ref.~12 ISO fit returns an anomalously high disk $ \rmn{M/L}_B =7.5$ and a bulge consistent with ~\citet{walt87,walt88} of $3< M/L_B <4$. Most fit results for M 31 return a M/L of the disk twice that of the bulge ~\citep{1989Kent}, however Ref.~ 15 disputes this trend with their varying concentration treatment, which returns M/L respectively of $0.93$ and $3.0$ for the disk and bulge, consistent with PSM of ~\citet{BelldYong}.
Ref.~15 attributes the difference to dark matter treatments which use constant density dark halos and do not account for absorption. ~\citet{Widrow} also notes extinction effects from dust in the disk.
While the LCM fits for M 31 in Fig.~\ref{fig:resultssmall} do not obviously favor one luminous model over the other, the stability of individual pairing results reported in Table~\ref{sumRESULTSbigger} indicate the
Ref.~15 luminous profile is more physically acceptable. As will be discussed in section~\ref{n891n7814}, the M 31 $\tilde{a}$ values are consistent with a conjecture relating $\tilde{a}$ to density gradients in the luminous matter. Based on that conjecture, the Ref.~15 luminous profile is again more physically representative an old, heavy stellar bulge.
\subsubsection{ NGC 5533}
NGC 5533 is an early-type galaxy at an estimated distance of $54$ Mpc. NGC 5533 has significant side-to-side asymmetries and a warp.
We sample NGC 5533 as reported in Ref.~10~\citep{SanMcGa} with an associated luminous profile from a MOND fit.
The Ref.~10 rotation curve data is from ~\citet{San96}.
The luminous profile geometry reported by Ref.~10 is calculated using a double exponential model similar to those used in all Ref.~3 and Ref.~8 galaxies.
The Ref.~10 gas is from H\,{\sevensize\bf I} surface densities, reported as $ 3.0 \times 10^{10} M_{\sun}$. Ref.~10 notes that a large fraction of total light is in the central bulge and that higher bulge M/L are consistent with expectations for older, less actively star forming spheroidal sub-systems.
Unfortunately, only one Milky Way luminous mass model in our sample has sufficient radial coverage to adequately fit this galaxy ~\citep{Xue}. However, NGC 5533 is retained
because there are so few galaxies larger than the Milky Way. We return M/L of the bulge a little lower than reported in Ref.~10, and maximal gas scalings of $1.5$. The Ref.~10 bulge M/L is about ten times that of the disk. Our LCM fits reduce this proportion to about a factor of two ($2$). Fit results are shown in Fig.~\ref{fig:resultssmall}. As will be discussed in the conjecture in section~\ref{n891n7814}, the NGC 5533 $\tilde{a}$ values (in Table~\ref{sumRESULTSbigger}) indicate a stellar bulge for NGC 5533 which is less dense than that of M 31 as reported in Ref.~15.
\subsubsection{NGC 891 \& NGC 7814 \label{n891n7814}}
The comparisons between the NGC 891 and NGC 7814 as reported in Ref.~11
~\citep{Frat} are probably the most important in our sample. NGC 891 and NGC 7814 are edge-on spiral galaxies at respective distances of $9.5$ and $14.6$ Mpc. The Ref.~11 rotation curves are from \citet{Oosterloo:2007se}, with a velocity resolution of $4.12\, \rmn{km\, s^{-1}}$, derived from concentric rings and Gaussian fits to the velocity field. M/L is reported in the $3.6\,\umu$m band, and the geometries of an exponential disk and spherical bulge are used. Because these two galaxies have similar total enclosed luminous mass at the limit of the data as indicated by
$\kappa_\tau$ (see section~\ref{kappa}),
very similar rotation curves beyond $3-4~\rmn{kpc}$, and similar inclinations, the direct comparison of their
free parameter $\tilde{a}$ values in Table~\ref{sumRESULTSbigger} can be tied to morphology.
NGC 891 is disk-dominated and NGC 7814 is bulge-dominated.
The comparison of the LCM results for NGC 891 and NGC 7814 are the most direct indication of a physical interpretation for $\tilde{a}$. The lower density gradient in the luminous matter associated with the disk-dominated galaxy NGC 891 is associated with low $\tilde{a}$ values as compared to those for the higher density gradient bulge-dominated NGC 7814. We see this trend throughout the sample, with one possible exception; therefore we make the following conjecture: $\tilde{a}$ increases with increasing density gradients in the luminous matter.
We note that it remains possible that $\tilde{a}$ instead implies a relation between the dark matter halo and the luminous matter profile, because NGC 891 is commonly referred to as having a minimal dark matter halo, consistent with vanishingly small $\tilde{a}$ values.
Ref.~11 reports rotation curves fits for NGC 7814 and NGC 891 in three different model contexts: Maximum light (MaxLight), ISO and MOND. For NGC 7814, we sample the ISO and MaxLight luminous profiles. For NGC 891, we sample the MOND and MaxLight luminous profiles. Ref.~11 MOND fits are based on the $\umu$-simple interpolating function as described in ~\citet{Gent}. The LCM results for both galaxies demonstrate convergence of the reported M/L from very different starting estimates. The NGC 7814 MaxLight fit result M/L of the bulge was unrealistically high, taken to indicate the necessity of a dark matter halo. The LCM lowers that M/L without a halo. In Fig.~\ref{fig:resultssmall} we report only one of the NGC 891 LCM fits (from MOND starting estimates of M/L) in the interest of space, given that the two LCM luminous profiles and fits were essentially the same, and for NGC 7814 LCM results from the starting assumptions of M/L from ISO are indicated by (A) and those from MaxLight are indicated by (B).
\subsubsection{NGC 2841}
NGC 2841 is an early-type, large spiral galaxy larger in radial extent than the Milky Way. The controversial distance estimates to this galaxy are the Cepheid distance of $14.1$ Mpc versus that indicated by MOND fits of $23$ Mpc. The larger MOND distance is supported by the lowest luminosity supernova type IA event ever observed~\citep{Gent,Bot}, which would be of standard luminosity at the MOND distance. The Cepheid based distance is considered the most precise distance measure, though some unresolved issues persist~\citep{HubbleCepheid}.
We sample three different references for NGC 2841: an NFW fit from Ref.~3~\citep{Blok} and MOND fits from Ref.~8~\citep{Gent} and Ref.~9~\citep{Bot}. The fits in Ref.~ 3 and 9 are at the
Cepheid distance, and that from Ref.~8 is at the constrained Cepheid distance of $15.60$ Mpc. Ref.~3 and Ref.~8 rotation curve derivations and modeling assumptions are described in sections~\ref{Blok} and \ref{TaM}. The Ref.~3 and Ref.~8 rotation curves agree with that in Ref.~9 within the error. The Ref.~9 rotation curves is from ~\citet{BegeTH}. Ref.~3 notes excess surface brightness in the central radii of NGC 2841, not accounted for by their exponential model.
The major difference between these three luminous profiles as reported is the stellar mass decomposition. In Ref.~3 and Ref.~8 the disk M/L is slightly smaller than that of the bulge. In Ref.~9, the disk M/L is ten times that of the bulge. The M/L reported in Ref.~3 are fixed consistent with ~\citet{BelldYong} PSM. As shown in Fig.~\ref{fig:resultssmall}, the LCM returns three luminous profiles and associated rotation curve fits which are very similar in all features. However, what distinguishes these three LCM fits are the resulting $\tilde{a}$ values.
The luminous assumptions in Ref.~3 and Ref.~8 return the largest $\tilde{a}$ values in the sample, versus those from Ref.~9 which are very close to the mean value (Fig.~\ref{fig:alphaDistrib1}). The scaling of $\tilde{a}$ with steeper density gradients in the luminous matter (section~\ref{n891n7814}) indicates that future directed photometric investigation of NGC 2841 can identify which model is most physically representative.
\subsubsection{NGC 7331}
NGC 7331 is an early-type spiral with strong spiral arm structure.
We sample two NGC 7331 rotation curves and associated MOND fits, using different distance estimates and associated luminous profiles. Ref.~9~\citep{Bot} uses the cepheid based distance ($14.7$ Mpc) and
and Ref.~8~\citep{Gent} the constrained distance ($13.43$ Mpc). The Ref.~8 rotation curve is higher resolution, and shows a rise in the final radii which is not observed in Ref.~9. The difference is attributed to differing assumptions of the galaxy's inclination ~\citep{Blok} .
The major difference between these two data sets are the luminous profiles; in Ref.~9 the M/L of the bulge and disk are approximately equal, and in Ref.~8 the M/L of the bulge is about three ($3$) times that of the disk.
PSM ~\citep{BelldYong} for this galaxy indicate a M/L of the bulge which is bigger by a factor of approximately $1.5$~\citep{Blok}. As can be seen in Fig.~\ref{fig:resultssmall}, the two LCM fits are roughly equivalent past $10$ kpc, though differing coverage makes the two data sets difficult to directly compare, though the LCM fits from Ref.~8 assumptions are more stable. Consistent with the conjecture in section~\ref{n891n7814}, the larger $\tilde{a}$ values are associated with fits using the larger bulge assumptions in Ref.~8. The total gas reported in Ref.~9 is $1.4\times 10^{10}M_{\sun}$, with associated rotation curve velocities approximately $30\, km\, s^{-1}$ higher than those for the gas reported in Ref.~8. The majority of LCM pairings for both data sets return minimal gas scalings for this galaxy. A comparison to NGC 3521 is made in the next section.
\subsubsection{NGC 3521}
NGC 3521 was one of the first cases of a genuinely declining H\,{\sevensize\bf I} rotation curve, although not a Keplerian descent. However, we sample data from Ref.~8~\citep{Gent} which does not confirm the descending rotation curve. There is a small decrease in circular velocity at high radii, but not as previously reported.
Ref.~8 reports MOND fit results for M/L free and the constrained distance of $7.50$\, Mpc, from the
Hubble flow distance of $10.7\, \pm\, 3.2$ Mpc. See section~\ref{TaM} for a description of the
associated assumptions for rotation curve derivations and luminous mass modeling.
In ~\citet{Blok} the fixed PSM of the disk M/L is $0.73$. Ref.~8 returns a value of $0.71$. The LCM fit return an average value of $0.72$. LCM fits returns maximal gas scalings of $1.5$.
An interesting feature for this analysis is the distribution of the luminous mass. While the total enclosed luminous mass at the farthest extent of the data (indicated by $\kappa_\tau$) for NGC 3521 and NGC 7331 (as reported in Ref.~8) are approximately equal, the average values of the free parameter $\tilde{a}$ for NGC 3521 are about $40\%$ smaller. With no reported bulge component for NGC 3521 as compared to NGC 7331, this is consistent with the conjecture in section~\ref{n891n7814}. Fit results are shown in Fig.~\ref{fig:resultssmall}.
\subsubsection{NGC 5055}
NGC 5055 is a rotationally dominated spiral galaxy at a favorable inclination for kinematic study. It has a
nearby companion galaxy UGC 8313, at a projected distance of $50$ kpc~\citep{Batt}.
We sample three rotation curves: Ref.~3~\citep{Blok}, Ref.~7~\citep{Batt}, and Ref.~8~\citep{Gent}.
The rotation curves in Ref.~3 and Ref.~8, as compared to Ref.~7, have a
more pronounced rise rate and higher velocities at large radii. Ref.~3 and Ref.~8 derive their rotation curves as described in Sections ~\ref{Blok} and \ref{TaM}.
Ref.~7 derives their rotation curve velocity using the conventional intensity-weighted mean Gaussian profile fits.
Distance indicators for the three authors are: Ref.~3 uses the Hubble flow distance of $10.1$ Mpc, Ref.~7 uses a Cepheid distance of $7.2$ Mpc~\citep{Pierce} and Ref.~8 uses a constrained fit distance of $7.07$ Mpc. The gas mass reported in Ref.~7.~\citep{Batt} is the total H\,{\sevensize\bf I} mass of $6.2 \times 10^9 M_ {\sun}$, scaled for helium abundances.
The luminous profile reported in Ref.~ 3 is fixed by PSM, that in
Ref.~7 is from a NFW fits and in Ref.~8 from a MOND fit. All three authors report their resulting M/L are consistent with PSM, though in Fig.~\ref{fig:massmodels18} they appear geometrically different. In Ref.~3, the M/L of the bulge is seven times smaller than that of the disk. In Ref.~8, the M/L of the bulge and disk are approximately equal. In Ref.~7 there is no stellar bulge.
LCM fits from the Ref.~3 luminous profile are highly unstable, see Table~\ref{sumRESULTSbigger}. The LCM fits using the Ref.~7 and Ref.~8 luminous profiles (both at essentially the same estimated distance) are visually (see Fig.~\ref{galaxiesSmaller}) difficult to compare given the higher resolution data from Ref.~8 and different extents of the coverage. However, if we hope to address the question of whether or not mass models should include a bulge, it is necessary to agree on a set of rotation curves, associated uncertainties, and luminous M/L in the same wavelength bands. Consistent with the conjecture in section~\ref{n891n7814}, the LCM fit results from Ref.~8 assumptions (with a bulge component) have larger $\tilde{a}$ values than those from Ref.~7 assumptions.
The LCM fits (from Ref.~8) do return a bulge M/L which is smaller than that of the disk, by a factor of approximately $1.6$.
\subsubsection{NGC 4138, NGC 5907 and NGC 3992}
The rotation curves for Ursa Major galaxies NGC 4138, NGC 5907 and NGC 3992 as reported in Ref.~10~\citep{SanMcGa}, are associated with luminous profiles from MOND fits. The rotation curve data is from ~\citet{Ver98,San96}. The Ref.~10 rotation curve derivation and luminous mass modeling assumptions are described in section~\ref{SanMc02}. These three galaxies are grouped because they have similar enclosed mass at the limit of the data (indicated by $\kappa_\tau$).
Comparison of these three galaxies, and their associated $\tilde{a}$, may challenge the conjecture in section~\ref{n891n7814}. The $\tilde{a}$ ranking, in increasing order is: NGC 4138, NGC 5907, NGC 3992. Correlating increasing $\tilde{a}$ with increasing density gradients in the luminous matter is challenged by the fact that in the Ref.~10 analysis, NGC 5907 has a bulge, whereas the high surface brightness galaxy NGC 3992 does not. The NGC 5907 bulge M/L is proportionally smaller than that of the disk, by about a factor of two, so it remains possible that NGC 3992 has a steeper density gradient. ~\citet{Bott3992} reports that NGC 3992 is a highly non-exponential, extreme Freeman type II galaxy. The LCM fits for NGC 4138 indicates that Ref.~10 over-estimates the luminous mass at small radii. Fits for NGC 4138, NGC 5907 and NGC 3992 are shown in Fig.~\ref{galaxiesSmaller}.
\subsubsection{NGC 2903}
NGC 2903 is a high surface brightness galaxy, for which we sample the reported rotation curves and associated luminous profiles from Ref.~3~\citep{Blok}, Ref.~8~\citep{Gent} and Ref.~10~\citep{SanMcGa}.
Distance indicators for the three authors are: Ref.~3 (NFW fit) uses the brightest stars distance of $8.9\,\pm 2.2$ Mpc, Ref.~8 (MOND fit) uses the constrained brightest stars distance of $ 9.56 $ Mpc, and Ref.~10 (MOND fit) uses the Hubble flow distance $6.4$ Mpc from ~\citet{BegeTH}. The
Ref.~3 and Ref.~8 rotation curve agree with that reported in ~\citet{BegeTH}, though Begeman's rotation curve rises more steeply in the inner radii. Ref.~3 attributes this difference to Begeman not using a full tilted-ring model in the inner parts of the galaxy, therefore not accounting for the sudden changes in the P.A. and inclination in his analysis. See sections~\ref{Blok}, \ref{TaM} and \ref{SanMc02} for details of the total light decomposition and rotation curve derivations for these three references. Due to non-circular motions and the presence of a bar, constraints on the inner radii of NGC 2903 remain open questions. The gas profile used in Ref.~ 10 is from ~\citet{BegeTH}, of $2.4\,\pm 0.1 \times 10^{9}M_{\sun}$. Out to a radius of $14.5$ kpc, the Ref.~3 and Ref.~8 gas profiles agree with Ref.~10. However at larger radii, the gas profiles of Ref.~3 and 8 are far larger than that in Ref.~10.
The major distinguishing feature between the Ref.~3 luminous profile and those in Ref.~8 and Ref.~10 is that Ref.~3 includes a bulge. Since the errors and rotation curves from Ref.~3 and Ref.~8 are essentially identical, it is telling that the two separate luminous profiles returned essentially equivalent LCM $\chi^2_r$ values. However, visually the Ref.~3 luminous profile with a bulge, as compare to Xue Milky Way in Fig.~\ref{galaxiesSmaller} is slightly better than that for Ref.~8. However, the Ref.~3 LCM results are unstable, as all the other Milky Way pairings returned a bulge M/L value of zero. This is an ambiguous result, because the chi-square for the Xue pairing is easily twice as good as any of the other pairs, see Table~\ref{sumRESULTSbigger}. Additionally, this is the only pairing in the group of nine ($9$) pairs which returns a minimal gas scaling.
These three data sets do indicate consistency with the conjecture in section~\ref{n891n7814} because the total light profile reported from Ref.~8 is steeper in the density gradient in the inner $10$ kpc than either Ref.~3 or Ref.~10. This can be seen in the Keplerian profiles in Fig.~\ref{galaxiesSmaller}.
\subsubsection{NGC 3953, UGC 6973, NGC 4088 and NGC 3726}
The rotation curves for Ursa Major Galaxies NGC 3953, UGC 6973 NGC 4088 and NGC 3726 are reported in Ref.~10~\citep{SanMcGa}, with luminous profiles resulting from MOND fits. Rotation curve data is reported from ~\citet{Ver98,San96}. Rotation curve and modeling assumptions are described in section~\ref{SanMc02}. UGC 6973 and NGC 4088 are reported as being kinematically disturbed. Fit results for all three galaxies are reported as consistent with PSM predictions. Reported gas masses, in H\,{\sevensize\bf I} are: for N3953 $0 .27\times 10^{10}M_{\sun}$, for U6973 $0 .17\times 10^{10}M_{\sun}$, for NGC 4088 $0 .79\times 10^{10}M_{\sun}$ and for NGC 3726 $0 .62\times 10^{10}M_{\sun}$.
These four galaxies are found to have approximately the same enclosed mass at the limit of the data ($\kappa_\tau\approx 0.28$), but two distinct sets of $\tilde{a}$. Averaged over the Milky Way pairings, these are: for
NGC 3953 \& UGC 6973 $\tilde{a}\approx 0.14$ and for N4088 \& N3726 $\tilde{a}\approx 0.08$. As is clear in Fig.~\ref{fig:results3}, the more central mass distributions of NGC 3953 and UGC 6973, as compared to NGC 3953 and UGC 6973, are consistent with the conjecture in section~\ref{n891n7814}.
\subsubsection{NGC 6946}
NGC 6946 is a late-type spiral at the limit of low inclinations which can feasibly be studied kinematically.
The two rotation curves we sample are from Ref.~3~\citep{Blok} and Ref.~10~\citep{SanMcGa}. Their rotation curves differ enormously,
both in magnitude and functional line shape with respect to $r$, see Fig.~\ref{galaxiesSmaller}. The rotation curve data in Ref.~10 comes from ~\citet{Carig90} and they report a MOND fit with luminous matter modeling estimates reported in ~\citet{San96}. Ref.~3 notes that the differences in their reported rotation curves is most likely due to the assumption of inclination angle; Ref.~3 uses $i=32.6^{\circ}$ versus that of Ref.~10 $i=38^{\circ}$. They also use different distance estimates: Ref.~3 uses $5.9$ Mpc ~\citep{Walter} and Ref.~10 uses $10.1$ Mpc
~\citep{Carig90}. Ref.~3 notes for NGC 6946 the large galaxy size relative to the beam size.
Since the error budget on the Ref.~3 rotation curve is easily twice that of Ref.~10, it is significant that the LCM fits to from Ref.~10 assumptions are three times better than those from the Ref.~3 assumptions. Physically, Ref.~3 and Ref.~10 provide very different NGC 6946s. Ref.~3 gives a fixed bulge and disk M/L from PSM, and Ref.~10 gives only a stellar disk, with M/L free. The LCM decreases the M/L reported in Ref.~3 by approximately $40\%$, and increases that reported by Ref.~10 by about the same margin. The galaxy, as reported by Ref.~3, yields maximal gas scalings, whereas the galaxy as reported by Ref.~10 returns minimal gas scalings.
\subsubsection{NGC 3198}
\label{n3198}
For the high surface brightness spiral galaxy NGC 3198 we sample four different data sets.
NGC 3198 is a moderately-inclined, late-type galaxy with a small central bar and is one of the standards for flat rotation-curve studies due to its high symmetry and regular rotations. We compare the best gas measurements from Ref.~4~\citep{Maria}, the
best rotation curve data Ref.~ 3~\citep{Blok}, the older standard from Ref.~1~\citep{Bege}, and an alternate interpretation of the Ref.~1 rotation curve in Ref.~2~\citep{JNav}. Ref.~1-3 report NFW fits and Ref.~4 a MOND fit. Ref.~1 and 2 use the distance of $9.4$ Mpc (Hubble Flow), Ref.~3 uses $13.8$ Mpc (Cepheids), and Ref.~4 uses $12.3$ Mpc (constrained from the Cepheids). Ref.~1 and Ref.~ 2 use the ~\citet{BegeTH} rotation curve, based on a tilted-ring model derivation. The Ref.~3 rotation curve derivation is described in section~\ref{Blok}. The Ref.~4 rotation curve is from the HALOGAS ~\citep{Heald} survey of the extra-planar gas in 22 spiral galaxies using very deep H\,{\sevensize\bf I} observations, , augmented by $\rmn{ H}\alpha$ observations. The rotation curve derivation reported in Ref.~4 is from a tilted-ring model with intensity-weighted velocity fields, treating each side of the galaxy separately. Ref.~4 uses a 3D model of the H\,{\sevensize\bf I} layers, giving attention to variable inclinations for the rings, position angles, and the features of variable rotation speeds as a function of distance from the plane. They find previously unreported gas features and show a lag on the rotation velocities of the extra planar gas. They analyze the H$\alpha$ emission to detect stellar light out to the end of the extra planar gas. Ref.~4 reports H\,{\sevensize\bf I} gas mass of $1.08\times 10^{10} M_{\sun}$ which is $6 \%$ higher than reported in Ref.~3. They attribute the difference to amplitude calibrations and detection of extra planar gas. Ref.~1 and 2 report gas mass of $5.0\times 10^9 M_{\sun}$. All four citations return maximal LCM gas scalings of $1.5$. In Ref.~4, the authors show all the NGC 3198 rotation curves compared here, plotted on an arc second scale. On that scale, all the rotation curves are similar, modulo the noise. However, as can be seen in Fig.~\ref{fig:results3}, the conversion to the assumed distance scales causes projection differences in the rotation curves.
As can be seen in Fig.~\ref{fig:massmodels18}, the reported luminous mass profiles vary significantly for this galaxy in geometry and magnitude. Modulo fixed geometry assumptions, the LCM returns similar luminous profiles for all four data sets, as shown in Fig.~\ref{fig:results3}. Ref.~4 uses the same geometry as Ref.~3, but a lower M/L ratio.
Ref.~ 1 and Ref.~2 respectively report M/L of $3.8$ and $1.39$ in the r-band, in solar mass units. The resulting LCM average M/L were respectively $3.3$ and $2.9$, demonstrating convergence. Ref.~3 and Ref.~4 respectively report M/L of $0.8$ and $0.3$ in the $3.6 \,\umu \rmn{m}$ band, in solar mass units. The resulting LCM average M/L were respectively $0.64$ and $0.61$, again demonstrating convergence. Visually, the LCM fit values for Ref.~2 are basically equivalent to those of Ref.~1 and those of Ref.~4 are superior to Ref.~3. The Ref.~3 assumptions over-predict the luminous profile at small radius.
The $\tilde{a}$ values are very consistent across all four of these posited luminous NGC 3198 profiles, despite differences in the posited distances.
\subsubsection{NGC 2403
Fig.~\ref{fig:results3}
NGC 2403 is a late-type Sc spiral in the M81 group, dominated by regular rotations. We sample two data sets:
Ref.~3~\citep{Blok} and
Ref.~9~\citep{Bot}. The Ref.~3 luminous profile is fixed by PSMs, and that of Ref.~9 is from a MOND fit; both using the Cepheid based distance of $ 3.22$ Mpc. Ref.~3 reports an NFW fit to the rotation curve data. The reported rotation curves are identical, modulo noise, as are the gas profiles. Ref.~9 reports a total gas mass of
$0.4\,\times 10^{10}\,M_{\sun}$, which is H\,{\sevensize\bf I} multiplied by $1.3$ to account for primordial Helium abundances. Ref.~3 notes that the velocity in the inner kiloparsecs are suspect to beam smearing effects, given the steep increase in velocity.
Both Ref.~3 and 9 use exponential disk geometries, respectively reporting stellar disk M/L in Ref.~9 of $\rmn{M/L}_B=1.6$ and Ref.~3
of $\rmn{M/L}_{3.6\,\umu \rmn{m}}=0.41$. As can be seen in Fig.~\ref{fig:massmodels18}, this results in higher Keplerian velocities due to the stellar disk in Ref.~9 than in Ref.~3, by approximately $30\, \rmn{km\, s^{-1}}$. The LCM scales down the stellar disk M/L from Ref.~9 and maintains that reported in Ref.~3. Both sets of fits return maximal gas scalings.
That the $\tilde{a}$ values from Ref.~3 assumptions are slightly higher than those for Ref.~9, consistent with the conjecture in section~\ref{n891n7814}, given that the geometry assumptions for the exponential disk in Ref.~3 indicates the same inner rise rate as in Ref.~10, but a much faster decline in the light, and associated mass, profile -- and thereby a steeper density gradient.
\subsubsection{M 33}
The galaxy M 33 in our Local group is used as a standard distance calibrator, measured independent of the Hubble flow to be
at a distance of $0.84$ Mpc. It is a low-luminosity, very blue, active star forming spiral galaxy with major implications for galaxy evolution and formation models. M33 is smaller than the Milky Way and hence tends to be denoted as being dark matter dominated ~\citep{Salucci}. The LCM successfully fits this galaxy without a dark matter halo in two reports: Ref.~5~\citep{Cor03}and Ref.~6~\citep{CoSa00}. These two rotation curve are respectively observed in: molecular gas (Ref.~5) and $21$-cm (Ref.~6). The two rotation curves differ only in
fine detail and have essentially the same error budgets. Ref.~5 reports their luminous profile in the context of an NFW fit, and Ref.~6 in the context of a model independent dark matter fit, consistent with CDM models.
Including molecular gas in Ref.~5 results in total gas mass of Ref.~5 $3.2\times 10^9 M_{\sun}$ as compared to that previously reported in Ref.~ 6 of $1.8\times 10^9 M_{\sun}$). This mass is predominantly added to M 33 past the radius of $8$\,kpc, which taken with the two different scale lengths used for the exponential disk geometries, results in slight modifications in the overall geometries implied in the two models. The
LCM $\chi^2_r$ values indicate that the Ref.~6 luminous distribution is approximately five times better than that of Ref.~5. Since the gas observations are more complete in Ref.~5, we assume the difference between these two fits is based upon the choice of scale length. Ref.~5 uses an exponential scale length of $R_d\approx 1.3\pm0.2$ kpc and Ref.~6 uses $R_d\approx 1.2\,\pm\,0.2$\,kpc, both in the K-band.
Consistent with the conjecture in section~\ref{n891n7814}, values for $\tilde{a}$ are larger for Ref.~6 assumptions with a steeper density gradient. Ref.~5 notes that after subtracting the disk profile from the total light, there remains excess central emission which motivates the interpretation given here.
\subsubsection{F 563-1 and NGC 925 and NGC 7793 }
The smallest galaxies in our sample are F 563-1, NGC 925 and NGC 7793. For each galaxy we compare two reports in the literature (rotation curve and luminous assumptions).
For F 563-1, we compare NFW fits in Ref.~2~\citep{JNav} and Ref.~13~\citep{giraud2000universal}.
Ref.~2 is rotation curve data reported by ~\citet{deBlok:1997ut} and Ref.~13 is rotation curve data reported by ~\citet{mcgaugh1998testing}.
For NGC 925, we compare the fits from Ref.~3~\citep{Blok}, reported in the context of NFW and ISO models and both reporting M/L fixed by PSM. For NGC 7793, we compare a MOND fit in Ref.~8~\citep{Gent} and an ISO fit in Ref.~14~\citep{Dicaire}.
The low surface brightness galaxy F 563-1 reported in Ref.~2 and Ref.~13 is at a distance of 45 Mpc. Both references use an exponential disk. Visually, in Fig.~\ref{galaxiesSmallest}, the LCM fits using the luminous mass assumptions from Ref.~2 are slightly better than those from Ref.~13, but the resulting M/L from Ref.~13 assumptions are more physically acceptable. Ref.~13 uses a similar disk profile to that reported in Ref.~2, but on average resulting in associated orbital velocities which are about $30 \rmn{km\, s^{-1}}$ higher(see Fig.~\ref{fig:massmodels18}). Both sets of fits require maximal LCM gas scalings.
NGC 925, late-type barred spiral at a reported distance of $9.2$ Mpc, is a very interesting example, as it indicates something physical about the LCM mapping. From Ref.~3 we sample two very different luminous profiles, consistent with the ISO and NFW, bot fits for M/L fixed by PSM. The ISO fit returns a significantly lower M/L than reported for the NFW fits, based on observations of the rotation curve. The LCM fit lowers the ISO M/L substantially, but does not set it to zero as is done in Ref.~3 for their NFW fit with M/L free. The fascinating feature from the ISO starting assumptions is that the luminous mass features (in ISO) are constructed to mimic the bumps in the rotation curve, but in the LCM fits shows it is actually the mirror image of these features that is reflected in the rotation curve. We interpret this as a lens-type effect, where small luminous mass features in galaxies that map mostly onto the bulge of the Milky Way are magnified and inverted. This key feature could lead to refinement of the luminous mass modeling of low luminosity and dwarf galaxies, based exclusively on the rotation curve data. LCM fits using
both the ISO and the NFW luminous profiles failed when paired to the \citet{Sofue} Milky Way.
NGC 7793 is a late-type spiral in the Sculptor group. The two rotation curves compared here come from $21$-cm observations reported in Ref.~8 and $\rmn{H}\alpha$ observations in Ref.~14. They differ outside of $4$ kpc, where Ref.~14 rotation velocities begin to decline faster than those reported in Ref.~8. The two luminous profiles use slightly different distance estimates, which may account for the differing rotation curves. Ref.~8 uses distance of $3.91$ Mpc from the tip of the red giant branch indicator and Ref.~14 uses the distance of $3.38$ Mpc from ~\citet{Puche}. Ref.~14 fails when compared to Milky Way ~\citet{Sofue}, but is otherwise very stable in the other three pairings. Ref.~8 notes that redder galaxies generally require higher stellar M/L, but they report MOND best-fit
$ (\rmn{M/L})_{disk}=0.28$. The M/L reported in Ref.~14 is much larger, but in a different wavelength band. However, visually the LCM fits indicate that the luminous assumptions in Ref.~14 are clearly preferable (Fig.~\ref{galaxiesSmallest}).
The average $\tilde{a}$ values for F 563-1, NGC 925 and NGC 7793 are consistent with the conjecture in section~\ref{n891n7814}.
\begin{figure*}
\centering
\subfigure[M 31, Ref.~15]{\includegraphics[width=0.33\textwidth]{m31klypxue.pdf}}
\subfigure[M 31, Ref.~12]{\includegraphics[width=0.33\textwidth]{m31xue.pdf}}
\subfigure[NGC 5533, Ref.~10]{\includegraphics[width=0.33\textwidth]{n5533xue.pdf}}\\
\subfigure[ NGC 7814, Ref.~11 from A ]{\includegraphics[width=0.33\textwidth]{n7814ISOxue.pdf}}
\subfigure[ NGC 7814, Ref.~11 from B ]{\includegraphics[width=0.33\textwidth]{n7814xue.pdf}}
\subfigure[ NGC 891, Ref.~11 ]{\includegraphics[width=0.33\textwidth]{n891MONDxue.pdf}}\\
\subfigure[ NGC 2841, Ref.~3 ]{\includegraphics[width=0.33\textwidth]{n2841bNFWxue.pdf}}
\subfigure[ NGC 2841, Ref.~8 ]{\includegraphics[width=0.33\textwidth]{n2841TaMxue.pdf}}
\subfigure[ NGC 2841, Ref.~9 ]{\includegraphics[width=0.33\textwidth]{n2841MONDxue.pdf}}\\
\subfigure[NGC 7331,~Ref.~8 ]{\includegraphics[width=0.33\textwidth]{n7331TaMxue.pdf}}
\subfigure[NGC 7331,~Ref.~9 ]{\includegraphics[width=0.33\textwidth]{n7331MONDxue.pdf}}
\subfigure[NGC 3521, Ref.~8 ]{\includegraphics[width=0.33\textwidth]{n3521TaMxue.pdf}}
\caption{LCM rotation curve fits. In all panels: black circles represent the observed rotation velocities, thin bars represent the reported uncertainties, dotted curves show the best-fit LCM Newtonian contributions from the luminous matter and LCM resulting rotation curve is the red dotted-dashed line. References are as in Table~\ref{sumRESULTS}.}
\label{fig:resultssmall}
\end{figure*}
\begin{figure*}
\centering
\subfigure[ NGC 5055, Ref.~3 vs. Sofue]{\includegraphics[width=0.33\textwidth]{n5055dNFWsofue}}
\subfigure[NGC 5055, Ref.~7]{\includegraphics[width=0.33\textwidth]{n5055batt2xue}}
\subfigure[NGC 5055, Ref.~ 8 ]{\includegraphics[width=0.33\textwidth]{n5055TaMxue}}
\subfigure[ NGC 4138, Ref.~10]{\includegraphics[width=0.33\textwidth]{n4138xue}}
\subfigure[ NGC 5907, Ref.~10 ]{\includegraphics[width=0.33\textwidth]{n5907xue}}
\subfigure[NGC 3992, Ref.~10 ]{\includegraphics[width=0.33\textwidth]{n3992xue}}
\subfigure[ NGC 2903, Ref.~3 ]{\includegraphics[width=0.33\textwidth]{n2903bNFWxue}}
\subfigure[NGC 2903, Ref.~8]{\includegraphics[width=0.33\textwidth]{n2903TaMxue}}
\subfigure[NGC 2903, Ref.~10 ]{\includegraphics[width=0.33\textwidth]{n2903xue}}\\
\subfigure[NGC 6946, Ref.~3 ]{\includegraphics[width=0.33\textwidth]{n6946bNFWxue}}
\subfigure[NGC 6946, Ref.~10]{\includegraphics[width=0.33\textwidth]{n6946xue}}
\caption{LCM rotation curve fits. In all panels: lines are as in Fig.~\ref{fig:resultssmall}. References are as in Table~\ref{sumRESULTS}.}
\label{galaxiesSmaller}
\end{figure*}
\begin{figure*}
\centering
\subfigure[ NGC 3953, Ref.~10 ]{\includegraphics[width=0.33\textwidth]{n3953xue}}
\subfigure[ UGC 6973, Ref.~10 ]{\includegraphics[width=0.33\textwidth]{u6973xue}}
\subfigure[ NGC 4088, Ref.~10]{\includegraphics[width=0.33\textwidth]{n4088xue}} \\
\subfigure[ NGC 3726, Ref.~10]{\includegraphics[width=0.33\textwidth]{n3726xue}}
\subfigure[NGC 3198, Ref.~1]{\includegraphics[width=0.33\textwidth]{n3198aNFWxue}}
\subfigure[NGC 3198, Ref.~2]{\includegraphics[width=0.33\textwidth]{n3198bNFWxue}} \\
\subfigure[ NGC 2403, Ref.~3]{\includegraphics[width=0.33\textwidth]{n2403bNFWxue}}
\subfigure[ NGC 3198, Ref.~3]{\includegraphics[width=0.33\textwidth]{n3198cNFWxue}}
\subfigure[ NGC 3198, Ref.~4]{\includegraphics[width=0.33\textwidth]{n3198dMONDxue}}\\
\subfigure[ NGC 2403, Ref.~9]{\includegraphics[width=0.33\textwidth]{n2403MONDxue}}
\subfigure[M 33, Ref.~5]{\includegraphics[width=0.33\textwidth]{m33xue}}
\subfigure[ M 33, Ref.~6]{\includegraphics[width=0.33\textwidth]{m33bNFWxue}}
\caption{LCM rotation curve fits. In all panels:lines are as in Fig.~\ref{fig:resultssmall}. References are as in Table~\ref{sumRESULTS}.}
\label{fig:results3}
\end{figure*}
\begin{figure*}
\centering
\subfigure[ F 563-1, Ref.~2]{\includegraphics[width=0.33\textwidth]{f563_1NFWxue}}
\subfigure[NGC 925, Ref.~3 (from NFW mass model)]{\includegraphics[width=0.33\textwidth]{n925NFWxue}}
\subfigure[ NGC 7793, Ref.~8]{\includegraphics[width=0.33\textwidth]{n7793TaMxue}}
\subfigure[ F 563-1, Ref.~13]{\includegraphics[width=0.33\textwidth]{f563_1bNFWxue}}
\subfigure[ NGC 925, Ref.~3 (from ISO mass model)]{\includegraphics[width=0.33\textwidth]{n925ISOxue}}
\subfigure[ NGC 7793, Ref.~14]{\includegraphics[width=0.33\textwidth]{n7793ISOxue}}
\caption{LCM rotation curve fits. In all panels:lines are as in Fig.~\ref{fig:resultssmall}. References are as in Table~\ref{sumRESULTS}.}
\label{galaxiesSmallest}
\end{figure*}
\section[]{Conclusions}
\label{sec:conclusion}
It has been noted by ~\citet{Bot} that a credible, empirical alternative to dark matter must:
\begin{itemize}
\item successfully predict rotation curves with reasonable estimates of the stellar mass-to-light ratios and gas fractions,
\item make sense within our physics framework,
\item predict various other astrophysical observations,
\item and have one universal parameter.
\end{itemize}
At this stage, the LCM does the following:
\begin{itemize}
\item Successfully predict rotation curves for the sample of galaxies reported here as evidenced by the average reduced $<\chi^2_r>$. For the MOND subset of 21 data sets: the MOND $<\chi^2_r>=5.41$ vs. the LCM $<\chi^2_r>=1.71$. For the dark matter model subset of 19 data sets: the dark matter model
$<\chi^2_r>= 3.22$ whereas that for the LCM fits was $<\chi^2_r>= 2.07$. LCM fits resulted in reasonable estimates of the luminous matter reported in Table~\ref{sumRESULTS}.
\item Makes sense in our physical framework: the LCM is based on classical physics and luminous matter, and it fits galactic rotation curves based on the postulate that observed frequency shifts include both relative velocity and relative curvature components. Both effects are phrased in terms of the equivalent Lorentz Doppler-shifts, by the kinematics of the Lorentz group.
\end{itemize}
In future investigations, the LCM can be extended to:
\begin{itemize}
\item Predict other astrophysical phenomena by analytical extension to other metric geometries and dynamics;
\item Include a physical interpretation of the free parameter $\tilde{a}$. There is some evidence of a correlation between $\tilde{a}$ and the steepness of the density gradient (section~\ref{n891n7814}), but further theoretical developments and tests on a larger sample of galaxies are needed. This would constitute the single
universal parameter discussed by ~\citet{Bot}.
\end{itemize}
Our fits indicate that it may be possible to constrain population synthesis modeling directly from rotation curve data in the cases of M 33 and NGC 925. For M 33 the LCM makes a direct constraint to the two reported disk scale lengths, which differ by about $8\%$. In the case of NGC 925, the resulting LCM fits demonstrate the galaxy's luminous profile appears to be magnified and flipped in small spatial features in the rotation curve, much as would be done by an optical lens. In both cases, the defining feature that allows the LCM to constrain population synthesis modeling is the compatibility of rotation curves, uncertainties, distances, M/L bands, scale lengths, etc.
In many of the LCM fits it is clear that the gas scalings are maximal or minimal, which is not to be taken as indication that the gas observations are at fault, but rather that the gas profiles are compensating for stellar models that are not appropriate. In the view of the authors, once the stellar profiles and rotation curves have been constrained to best approximation, gas scalings will be indicative of real features.
This is already the case for the M 33 fit using Ref.~6.
In the future, as the LCM sample size increases, trends in the fits will indicate which Milky Way luminous mass model most closely represents the true mass distribution. At this stage of the analysis there is not a statistical preference in the fits for one model over another. One particularly interesting question for future work is how to construct the LCM for the rotation curve of the Milky Way itself; in this case the receiver frame is imbedded in the emitter frame. Weak lensing problems are another place to which investigations can easily be extended, as weak lensing has already been phrased in the geometric terms of Schwarzschild coordinate light speeds and effective indices of refraction by \citet{Narayan}.
Given the role that the luminous matter of galaxies plays in the our understanding of dark energy and the Hubble flow, indications from free parameters are critical to the big picture. Initial investigations into relations between the free parameter and the scale lengths are beyond the scope of the present paper. Once the physical significance of the parameter $\tilde{a}$ has been identified, it will be possible to directly report the luminous matter distribution from the observed shifted spectra.
It remains possible that the LCM free parameter indicates scaling relations between the dark matter halo and the luminous matter distribution; if that is so, the LCM offers an alternate method for luminous matter modeling with fewer parameter constraints on such a relationship.
\section[]{Acknowledgments}
The authors would like to thank R.\,A.\,M.\, Walterbos, V.\,P.\, Nair, M.\, Inzunza II, J.\, Conrad, V.\, Papavassiliou, T.\, Boyer, P.\, Fisher, E.\, Bertschinger and I.\, Cisneros. \\
S.\, Cisneros is supported by the MIT Martin Luther King Jr. Fellowship, while J.\,A. Formaggio and N.\,A. Oblath are supported by the United States Department of Energy under Grant No. DE-FG02-06ER- 41420.
|
\section{Introduction}
Studies of star-forming galaxies at a wide range of cosmic epochs have revealed a strong correlation at fixed redshift between star formation rate (SFR) and stellar mass ($M_*$), of the form
\begin{equation}
\log SFR (M_\odot/\textrm{yr}) = \alpha \times (\log M_*/M_\odot - 10.5) + \beta,
\label{eq:sfr}
\end{equation}
with $\alpha$ and $\beta$ likely time-dependent. This relationship has been shown to hold over 4\,--\,5 orders of magnitude in mass \citep{santini+09} and from $z = 0$ to $z \sim 6$ \citep{noeske+07,daddi+07,karim+11,kashino+13,Speagle2014}. It is a tight correlation, with only $\sim .25$\,--\,$.35$\,dex of scatter at any redshift \citep{daddi+07,Whitaker2012}.
This paper extends the star-forming main sequence (SFMS) to massive galaxies at higher redshifts than previous studies by using the first data from the Spitzer Large Area Survey with Hyper-Suprime-Cam (SPLASH). This survey is obtaining 2475h ($>$6h/pointing) of Spitzer IRAC 3.6$\mu$m and 4.5$\mu$m observations on the two Hyper-Suprime-Cam \citep{Takada2010} ultra-deep fields COSMOS \citep{scoville+07,Sanders2007} and SXDS \citep{SXDS}. Here we use the first 50\% of the data on the COSMOS field ($\sim4$h/pointing) along with previously published 0.15-2.5$\mu$m data \citep{ilbert+13} to probe star forming galaxies with masses $>10^{10}M_\odot$ to z$\sim6$.
SPLASH is is a photometric survey that improves upon existing studies at $4 < z < 6$. For these galaxies, the Lyman break allows quality photometric redshift determination, while SPLASH is complete and deep enough to greatly reduce biases that occur in studies specifically selecting for Lyman break galaxies \citep{lee+12,Speagle2014}.
In \S~\ref{sec:dataset}, we describe the SPLASH catalog, as well as the spectral energy distribution (SED) template fitting we use to produce photometric redshifts (photo-z's), stellar masses, and SFRs. \S~\ref{sec:mainseq} describes the SPLASH view of the SFMS, including a lack of observed quenching at the high-mass end. Possible explanations for this lack of a turnoff are discussed in \S~\ref{sec:discussion}, and in more detail in \citet{Steinhardt2014}.
Throughout this work, we assume a standard $(H_0,\Omega_m,\Omega_\Lambda)$ = $(70,0.3,0.7)$ cosmology, AB magnitudes, and a Chabrier \citep{Chabrier2003} IMF (integrated from $0.1 M_\odot$ to $100 M_\odot$).
\section{The SPLASH Dataset}
\label{sec:dataset}
This work uses a revised version of the Subaru i$^+$ band selected Cosmological Evolution Survey (COSMOS) catalog from \citet{Capak2007} to provide $0.15-2.5\mu$m photometry, with $10^6$ spurious sources removed \citep{salvato+09}, updated with intermediate band data, photo-z's, and physical parameters as described in \citet{ilbert+10}. The catalog was further augmented with significantly (1 magnitude) deeper z$^+$ band data taken with an updated Subaru-Suprime-Cam, Ultra-Vista \citep{McCracken2012} imaging in Y, J, H, and K bands, and importantly the SPLASH IRAC $3.6\mu$m and $4.5\mu$m data. A full description of SPLASH is provided in Capak et al. (in preparation).
In this paper we use the first 50\% of the SPLASH IRAC data, which covers a 1.2 degree diameter circle centered on the COSMOS field to 5$\sigma$ $\sim 25.3$ mag$_{\textrm{AB}}$ at both wavelengths. To reduce AGN contamination, X-ray detected sources \citep{Brusa2010,Civano2012} were removed from the sample. To overcome source confusion (blending) we extracted the photometry using the $i^+$ band catalog as input for the IRACLEAN procedure described in \citet{Hsieh2012}. The IRACLEAN photometry was compared with a T-FIT \citep{Laidler2007} catalog in the CANDELS \citep{Grogin2011,Koekemoer2011} area using an HST-WFC3 H band image as a prior and an EMPHOT \citep{Conseil2011} catalog of the whole COSMOS field using the Ultra-Vista K band image as a prior. No systematic bias as a function of flux or position was found due to the photometric method.
The primary selection effect in our sample is the i-band catalog, which is used as a prior for the Spitzer photometry and as the basis for the photometric catalog. The use of this catalog limits us to $z \lesssim 6$, when the Lyman break leaves the i-band. The depth limits us to unobscured star formation rates $>5-10 M_\odot$/yr at $z\sim4-6$ (typical excitations are $A_V \sim 0.4$ mag for objects in our sample; cf. \citet{Bouwens2012}). The typical depth of i-band selection is 26.5 mag \citep{Capak2007,Ilbert2009}. The second limit is the depth of the IRAC data, which introduces an age-dependent mass cutoff. For zero-age galaxies with a flat spectral energy distribution the IRAC data will limit the lowest measurable masses. For older galaxies, however, the depth of the Subaru $i^+$ image will limit the lowest measurable masses since they would not be in the i-band catalog. For stellar populations with approximately the age of the universe, our joint mass cutoff is noted in Figure 2. A third, more subtle effect is the method used to extract the IRAC photometry. If the photometry is not properly deblended, fainter galaxies will have systematically increased fluxes due to photometric crowding. The IRACLEAN algorithm corrects for this at least as well as other methods (TFIT, EMPHOT), but we note that this may be affecting a small number of objects crowded by a source not in the i-band catalog.
\subsection{Photometric Redshifts}
\label{subsec:photoz}
We determine photo-z's using Le\_PHARE \citep{Arnouts2011} with the methodology described in \citet{ilbert+13}. Stellar masses were estimated with \citet{Bruzual2003} (BC03) models including strong emission lines with a mix of exponentially-declining ($\tau/\textrm{100 Myr}=[0.1,0.3,1,3,10,30]$) and delayed-$\tau$ ($\Delta t_{\textrm{peak}}/\textrm{Gyr}=[1,3]$) star formation histories (SFHs) with solar and half-solar metallicities at a range of ages spanning 0.05-13.5 Gyr to account for the effects of low metallicities and increasing SFHs on estimated physical parameters. Dust attenuation is modelled using the starburst \citet{calzetti+00} curve and the $\lambda^{-0.9}$ curve from \citet{Arnouts2013} with $E(B-V)$ up to 0.7 mag.
Although the presence of a Lyman break should result in precise photo-z's, the 4000\AA\ break and Lyman break can be confused resulting in catastrophic redshift failures. To check the quality of the photo-z's we cross-reference them with spectroscopically confirmed sources (spec-z's), primarily from DEIMOS observations selected to be representative of $z>4$ galaxies \citep{Capak2010,Mallery2012}. Out of 139 galaxies with spec-$z > 4$ and robust IRAC fluxes, 87 (63\%) have successful photo-z determinations (Fig. \ref{fig:DEIMOScomp}) to within 15\%, with most of the rest mistakenly fit with at much lower redshift. Howvever, there are very few false positives in the SPLASH catalog. 6\% of objects with photo-$z > 4$ and measured spec-$z$s lie at spec-$z < 2$ (indicating confusion between the two breaks), and after correcting for the significantly larger fraction of objects with measured spectroscopic redshifts at $z<2$ (10\% at $z<2$ vs 4\% at z$>4$ with ch1$<25.3$), this implies a contamination rate of 2\%. So, the current photo-z analysis is preferentially scattering objects down from $z \sim 4-6$ to $z \sim 0.5$, (Fig. \ref{fig:DEIMOScomp}) with approximately 40\% of SPLASH star-forming galaxies erroneously fit as low-redshift and removed from the sample. Tests using the two-point correlation and cross-correlation functions confirm these fractions (Coupon et al., private communication).
We note that typical SFRs are far lower at $z \sim 0.5$ than $z > 4$, so up-scattered objects fall well off the SFMS at $z>4$ where they will not affect our further analysis.
\begin{figure}[!ht]
\plotone{fig1.ps}
\caption{A comparison of the best-fit photometric redshift and DEIMOS spectroscopic redshift for objects in SPLASH at $z_{phot} > 4$ or $z_{spec} > 4$. Due to confusion between the Balmer break and Lyman break, many objects with catastrophic errors are mistakenly excluded from the sample (scattered to $z \sim 0.5$), but very few objects are scattered up, even though there are an order of magnitude more spectroscopic galaxies at low redshift.}
\label{fig:DEIMOScomp}
\end{figure}
\subsection{Stellar Masses and Star Formation Rates}
Once the best-fit photo-z's and extinctions are determined, Le\_PHARE fixes both and fits for a stellar mass and SFR based upon the BC03 templates with emission lines added at the best-fit redshift \citep{Ilbert2009}. We compared these masses to those with several different SFHs, dust extinction laws, and emission line prescriptions similar to \citet{Arnouts2013} and found the results varied by less than 0.2 dex. Furthermore, we found consistent stellar mass estimates for objects at $z>4$ when comparing these results with \citet{ilbert+13}, with the exception of heavily blended objects where our improved IRAC photometry had an effect. We note that although stellar mass estimates come from SED fits at all redshifts, at lower redshifts it is possible to measure SFRs independently. Although beyond the scope of this Letter, this is discussed extensively in \citet{Speagle2014}.
At $z\sim0-3$, several studies have investigated SED-based SFRs and found there is generally good agreement between the SED and other SFR indicators \citep{salim+07,Arnouts2013,Carollo2013}. To verify this at the higher-redshifts in SPLASH we tested full SED fitting-based SFR indicators against UV- \citep{Meurer+99} and FIR- \citep{Lee2013} based SFRs.
It was possible to derive reasonable UV SFRs using the IR excess - UV slope (IRX-$\beta$) relation \citep{Meurer+99} for objects with detections in at least three bands in the rest-frame 1600-3000 \AA~ range. Galaxies with either negative $A_{1600}$ or unconstrained $\beta$ were removed, leaving a sample of 404 galaxies. There is good agreement in the mean values of these UV and SED-based SFRs over more than two orders of magnitude in SFR, but with a large scatter of 0.8 dex. This scatter is consistent with expectations at these high redshifts \citep{Bouwens2013,Finkelstein2012} due to both the intrinsic properties of these galaxies and measurement error.
For galaxies with very high SFRs ($\gtrsim500 M_\odot/\textrm{yr}$) we use Herschel fluxes (Lee et al., in prep.) to estimate FIR-based SFRs and find generally good agreement. The scatter between the SED and FIR based SFRs is also large but consistent with other studies (e.g., \citet{Reddy2008}). We conclude that our SED based SFRs are reasonable, but may be under-estimating the actual scatter in the SFMS. One additional concern is that we only use one dust extinction law, but in reality there are likely many, and we cannot disambiguate at these redshifts.
\subsection{Sample Selection}
There are 2,002,871 sources in the SPLASH catalog, of which 1,319,197 are classified as stars or galaxies with well-constrained redshifts, masses, and SFRs. Of these, 7583 have a best-fit photo-z $> 4$ (\S~\ref{subsec:photoz}). We removed any objects where quasar or stellar SEDs had better fits than the galaxy templates. Assuming conservatively that the stellar population has the same age as the universe, SPLASH is mass complete at $\log M_*/M_\odot > 10$ at these redshifts, a limit that will be improved by over 1 mag with upcoming Hyper-Suprime-Cam (HSC) data in the visible.
To obtain a clean sample we chose a quality cut based on the SED fitting uncertainty, $\chi^2/\textrm{DOF} < 5$, rather than a single band signal-to-noise (S/N). In practice, this is primarily a selection for having many observed bands with S/N $>3$. This leaves a sample of 3398 objects between redshifts 4 and 6 (2541/857 above/below $z = 4.8$, dividing the range into two bins of equal cosmic time). Limiting the sample to the mass-complete regime leaves 2152 objects (1513/639). We will use this smaller, mass-complete (sub)sample when deriving relationships between physical quantities to avoid biases. We include in our sample approximately 50 objects with no IRAC detection but otherwise well-constrained SED fits that pass the remainder of our cuts to avoid introducing a bias against particular SED shapes. Removing these objects has a negligible effect on the conclusions drawn in \S~\ref{sec:mainseq}.
Objects selected by the $\chi^2$ cut are preferentially brighter in the optical, and may introduce a bias in the sample distribution. To test for this we did a Kolmogorov-Smirnov comparison between the inferred SFR and stellar mass of our $\chi^2/\textrm{DOF} < 5$ sample and the original sample. The two are consistent with being drawn from the same distribution.
As discussed in \S~\ref{subsec:photoz}, approximately 80 objects ($\sim 2\%$) of our sample are expected to be low-redshift galaxies fit with $z > 4$ due to confusion between the 4000\AA~ and Lyman breaks. At $z > 4$, it is expected that all galaxies are star-forming \citep{Brammer2011,Behroozi2013}, and $z < 2$ galaxies of any type scattered up to $z>4$ will be assigned much lower SFR or even be selected as quiescent due to the shape of the typical galaxy SED. 73/3398 galaxies have SFRs below 10 $M_\odot$/yr (70 with $z < 4.8$), while 46/3398 are found to be quiescent according to the \citet{ilbert+13} criteria based on dust-corrected NUV$rJ$ criteria. Because these two selections have little overlap and comprise a negligible fraction of the catalog, we choose to use all 3398 objects in our final sample.
\section{The Star-Forming Main Sequence}
\label{sec:mainseq}
The main result of this paper, the SFMS from SPLASH at $4 < z < 4.8$ and $4.8 < z < 6$ (bins of equal time interval), are shown in Fig. \ref{fig:SFfig}.
\begin{figure}[!ht]
\plotone{fig3.ps}
\caption{The ``main sequence'' for star-forming galaxies at photometric (a) $4 < z < 4.8$ and (b) $4.8 < z < 6$. A best-fit linear relationship is indicated by the blue dashed line in each panel. Mass and SFR completeness are indicated by the solid black lines. The magenta shaded region reflects an estimate of the increased sensitivity of SPLASH due to the addition of IRAC channels to existing multiwavelength data. Contours are drawn are equal intervals in number density.}
\label{fig:SFfig}
\end{figure}
In both redshift ranges, there is a strong correlation between stellar mass and SFR, qualitatively similar to that seen in approximately two dozen previous studies \citep{Speagle2014} but with a different locus than at lower redshift. Almost all objects lie near this SFMS, with a small number of clear outliers at low SFR. Because of the limits of broad-band SED fitting, it is unclear without followup observations whether these outliers are candidates for quenched galaxies, catastrophic errors in photo-z's, and/or SED fitting, or have some alternative explanation.
The main sequence in each panel is divided by stellar mass into bins of width $0.1$ dex, with the median for each bin shown in Fig. \ref{fig:SFturnoff}. For ease of comparison with other studies, medians above the mass completeness threshold that encompass $> 25$ objects are then fit with a power law (Eq. \ref{eq:sfr}),
with best-fit $\alpha$ and $\beta$ indicated in Table \ref{table:fits}. Using both fits, individual objects exhibit 1$\sigma$ scatter of $\sim 0.24$ dex about the main sequence, and agree well with Lyman-break selected objects from \citet{lee+12} at similar/slightly lower redshifts and lower stellar masses.
\begin{table*}
\caption{Best-fit Parameters for the Main Sequence at $4<z<6$}
\centering
\begin{tabular}{l|c|c|c|c}
\label{table:fits}
Source & $\alpha(4-4.8)$ & $\beta(4-4.8)$ & $\alpha(4.8-6)$ & $\beta(4.8-6)$ \\
\hline
Measured & $0.78 \pm 0.02$ & $1.976 \pm 0.005$ & $0.78 \pm 0.02$ & $2.110 \pm 0.003$ \\
Lee+12 & $0.79$ & $1.955$ & $0.73$ & $1.975$ \\
Whitaker+12(z) & 0.13 & 1.72 & 0.02 & 1.17 \\
Speagle+14(t) & $0.79 \pm 0.03$ & $2.11 \pm 0.04$ & $0.80 \pm 0.03$ & $2.16 \pm 0.04$ \\
\end{tabular}
\\
{\scriptsize Best fit slope $\alpha$ and SFR at $\log M/M_\odot = 10.5$ $\beta$ for the star-forming main sequence in SPLASH. The expected fits according to the redshift and time dependence given by \citet{Whitaker2012} and \citet{Speagle2014}, respectively, are included for comparison, as well as the best fit values for the Lyman-break selected samples of \citet{lee+12} at $z \sim 3.9$ and $5.0$ corrected for extinction via \citet{bouwens+12}.}
\end{table*}
Previous studies have disagreed on whether $\alpha$ is increasing (e.g., \citet{santini+09}), constant (e.g., \citet{karim+11}), or decreasing (e.g., \citet{Whitaker2012}) with redshift. We find a high-redshift slope of $0.78 \pm 0.02$ (Table \ref{table:fits}), in good agreement with \citet{lee+12} and the increasing-$\alpha$ prediction from the \citet{Speagle2014} literature compilation. The uncertainties given are statistical, but this good agreement may imply that the unknown systematic uncertainties are smaller than might otherwise have been expected.
\subsection{High-Mass Galaxies and Quenching}
As star formation is quenched, individual galaxies should turn off the SFMS, either lying at lower SFR or becoming quiescent and disappearing entirely due to selection. It has been previously suggested that high-mass quenching is observed as a sub-linear SFR-$M_*$ relationship at high masses, but this may be due to selection effects \citep{noeske+07,karim+11,Whitaker2012}.
At the higher redshifts observed in SPLASH, the SFR main sequence is mass- and SFR-complete for the portion of the distribution indicated in Fig. \ref{fig:SFfig}. The SFR main sequence is reasonably approximated by a linear relationship for all masses above mass completeness and exhibits no evidence of a high-mass turnoff (Fig. \ref{fig:SFturnoff}). However, due to the depth of the current optical/NIR data, SPLASH is not sensitive to a population of fully quenched galaxies at this redshift.
\begin{figure}[!ht]
\plotone{fig4.ps}
\caption{The star-forming galaxy main sequence median values (red dots) at (a) $4 < z < 4.8$ and (b) $4.6 < z < 6$, with contours from Fig. 2 superimposed. There is no decrease in star formation rate or any other evidence of quenching even for the highest-mass star forming galaxies. That is, the power-law form holds at all masses.}
\label{fig:SFturnoff}
\end{figure}
\section{Discussion}
\label{sec:discussion}
Several previous studies have examined the redshift evolution of the SFMS, finding that the slope $\alpha = d(\log SFR)/d(\log M_*)$ may decrease towards higher redshift. SPLASH allows us to check whether this trend continues to $z \sim 6$. At lower redshift the evolution of $\alpha$ has been fit as a function of redshift \citep{Whitaker2012} and of time \citep{Speagle2014}. We find the time-based fit is better at high redshift as $(1+z) \propto t^{2/3}$ becomes more clearly non-linear (Table \ref{table:fits}), which makes sense because the physical quantity affecting SFR is likely time, not redshift.
It is surprising that the SFMS at $4 < z < 6$ has almost exactly the properties that one would extrapolate from lower-redshift behavior, with increasing specific SFRs ($\equiv \psi/M_*$) for $M>10^{11} M_\odot$ galaxies. This trend cannot continue indefinitely because the virialization time for a galaxy depends upon its initial overdensity, and a typical galaxy with total baryonic mass of $M > 10^{11} M_\odot$ ($M \geq M_*$ ) will not virialize until $z \lesssim 10$, with more massive galaxies having even lower virialization redshifts \citep{Haiman1997}. Furthermore, there is an effective Eddington limit on star formation for a given galaxy mass before star formation destroys the interstellar medium \citep{Younger2008} of
\begin{equation}
\psi_{\textrm{max}} = 600 \sigma_{400}^2D_{\textrm{kpc}}\kappa_{100}^{-1}M_\odot\textrm{yr}^{-1},
\end{equation}
where the line-of-sight velocity dispersion in units of 400 km/s ($\sigma_{400}$), the interstellar medium opacity in units of 100 cm$^2$/g ($\kappa_{100}$), and characteristic physical scale of the starburst in kpc ($D_{\textrm{kpc}}$) are likely of order unity. Thus, even under ideal conditions, it takes at least an additional $\sim 10^8-10^9$ yr beyond virialization for the largest protogalaxies observed in SPLASH to attain $M_* > 10^{11.5} M_\odot$. The universe at $z\sim 6$ is slightly less than $10^9$ yr old.
At $z \sim 6$ the formation of these galaxies is possible if they spend most of their time at near-maximum SFRs. Such galaxies would then not have a history of sporadic starbursts, but rather near-continuous high-rate star formation from the moment of collapse until $z=6$. This calculation, however, is only constrained to order of magnitude, so some variability in their SFHs is possible. However, even under ideal conditions this trend cannot continue far beyond $z > 6$, as there is insufficient time to build up stellar mass. Thus, we should expect that upcoming Hyper-Suprime-Cam (HSC) data, which is anticipated to extend the SPLASH catalog to $z \sim 8$, should find a turnover redshift where the most massive galaxies are not observed, indicating we are seeing the rapid buildup of galaxies that have just completed their primordial formation and virialization. Otherwise, a substantially modified theory for early-universe structure formation will be required.
HSC observations will also provide a substantial improvement in completeness in selecting high-mass, low-SFR galaxies. Such galaxies lying only slightly below the SFMS would be present in the current data, but improved Y- and z-band detection using HSC is necessary to select dusty and/or more quiescent high-mass, high-redshift galaxies. The current SPLASH catalog is sufficient to demonstrate that there is no population of turnoff galaxies lying just below the main sequence, but cannot determine whether this is because no galaxies have completed their star formation by $z\sim 6$ or because these galaxies have rapidly changed their SED and are lost due to selection. Such galaxies have been observed to $z\sim 3.5$ \citep{ilbert+13} and likely form in Eddington limited starbursts \citep{Toft2014}, but it is unclear when the quenching process began. Finding these first turnoff galaxies is essential to understanding when and how the most massive systems in the universe are quenched.
Taken together, our results provide strong hints that the most massive galaxies at high redshift all have a very similar history, one where they are forming stars at a nearly maximal rate from the first moment of virialization until a rapid turnoff at high redshift. This might turn out to be incompatible with the current consensus view that most galaxies form their stars through a series of individual starbursts, likely triggered by local events and environment and individual to the history of each galaxy. However, SPLASH is reporting only on the most massive galaxies, formed at the highest redshifts and very overdense initial fluctuations, so such galaxies are not necessarily typical. A natural next step is to follow up on this analysis, combining it with observational evidence about the formation process for galactic nuclei to try and build a consistent framework for galaxy evolution (cf. \citet{Steinhardt2014}), but such a model is beyond the scope of this Letter.
\acknowledgments
The authors would like to thank Steve Bickerton, Sean Carey, Martin Elvis, and Brian Feldstein for helpful comments.
\bibliographystyle{apj}
|
\section{Introduction}
The first Fermi-LAT GRB catalog (Ackermann et al 2013) identifies a "temporally extended"
emission at 100 MeV--10 GeV for eight bursts, with four other having LAT detections well after
the end of the Fermi-GBM prompt phase. The LAT emission of those eight afterglows peaks at
10--20 s after trigger, having a fluence $\Phi = 10^{-5\pm 1} \ergcm2$, followed by a
flux decay $\nu F_\nu \propto t^{-1.3\pm 0.3}$ until up to 1 ks, with a photon spectrum
$C_\nu \propto \nu^{-2.1\pm 0.2}$. The energetic output of those LAT afterglows, $E_{100 MeV} =
10^{53 \pm 1}$ erg, is 10--100 percent of their GRB output (at $\sim 1$ MeV).
The condition of optical thinness to pair-formation for the highest-energy LAT photons yields
lower limits on the Lorentz factor $\Gamma$ of their source. Assuming that the prompt LAT emission
has same origin as the GBM burst and using the burst variability timescale to determine the source
radius, it was inferred that $\Gamma_{grb} > 200-1000$ (e.g. Abdo et al 2009). The general lack
of pair-formation signatures in the light-curves and spectra of LAT afterglows indicates that
the afterglow source also has $\Gamma_{ag} > 200$ (Panaitescu, Vestrand \& Wozniak 2014), with
a detection bias against afterglows with $\Gamma \mathrel{\rlap{\raise 0.511ex \hbox{$<$}}{\lower 0.511ex \hbox{$\sim$}}} 75$, for which pair-formation attenuates
the intrinsic afterglow emission too much and the emergent afterglow is too dim to be detected by LAT.
In this work, we calculate the synchrotron and inverse-Compton emission from pairs formed in
LAT afterglows, in a simplified set-up. The temporal and spectral properties of LAT afterglows
being consistent with those at lower photon energies (optical and X-ray), indicates that the LAT
emission arises in the forward-shock driven by the GRB ejecta into the circumburst medium
(Kumar \& Barniol Duran 2009). An important simplification is that we consider only the pairs
formed {\sl behind} the forward-shock, whose energy distribution is set primarily by the
spectrum of the high-energy photons, and ignore the pairs formed {\sl ahead} of the forward-shock,
whose energy is set by their shock-acceleration (given that half of the emitted photons travel
ahead of the forward-shock, it follows that a comparable number of pairs form behind and ahead
of the shock).
For ease of calculating the number of pairs, we assume that the single power-law spectrum of the
high-energy LAT photons extends well below 100 MeV and above 100 GeV, and that it has a spectral
index -2 (in photon number). Furthermore, we consider only the pairs formed from high-energy
photons for which the photon front is optically thick to pair-formation. Another simplification
made is that the threshold energy for pair-formation is set only by the relativistic collimation of
the seed LAT photons, i.e. we ignore the scattering/decollimation of the high-energy photons by the
already formed pairs.
Section \S\ref{npairs} calculates the number of pairs with the above approximations, and their
minimal energy in the shock frame; \S\ref{syic} presents the calculation of the spectral
breaks of the pair emission, the regions in the $nb-\Gamma$ corresponding to various orderings
of the spectral breaks being identified in \S\ref{regions}.
The calculation of the received synchrotron self-Compton emission, taking into account synchrotron
self-absorption, radiative cooling, first inverse-Compton scattering if the optical thickness to
scattering by pairs is $\tpm < 1$, and higher orders inverse-Compton if $\tpm > 1$, is presented
in \S\ref{moresyic}. Optical and X-ray pair light-curves are discussed in \S\ref{results}.
\vspace*{2mm}
\section{Number of pairs and their distribution with energy}
\label{npairs}
A photon of {\sl observer}-frame energy $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb$ MeV forms a pair when interacting with another
photon of energy above the {\sl source}-frame threshold
\begin{equation}
\eps_t (\varepsilon} \def\eps{\epsilon} \def\nnbb{nb) = \frac{ 4\Gamma^2 (m_e c^2)^2 }{(z+1) \varepsilon} \def\eps{\epsilon} \def\nnbb{nb} =
35\, \z^{-1} \frac{\Gamma_2^2}{\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_8} \; {\rm MeV}
\label{ethr}
\end{equation}
with $\Gamma$ the Lorentz factor of the source that produced both photons, and using the notations
\begin{equation}
X_n = \frac{X (cgs)}{10^n} \;, \quad \z \equiv \frac{z+1}{3}
\end{equation}
For an afterglow fluence $\Phi$ at 0.1--10 GeV, with a photon-number distribution with energy
$C_\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \propto \varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$, the number of photons with energy above $\eps$ is
\begin{equation}
N_\gamma (> \eps) = \frac{4 \pi d_l^2}{z+1} \frac{\Phi}{4.6\, \eps} =
1.14 \times 10^{56} \z^3 \frac{\phi5}{\eps_8}
\label{NN}
\end{equation}
where $d_l = 5.10^{27} (z+1)^2$ cm is the luminosity distance for redshift $z$.
Thus, the number of photons above the threshold energy $\eps_t$ is
\begin{equation}
N_\gamma [> \eps_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb)] = 3.3 \times 10^{56} \z^4 \frac{\phi5}{\Gamma_2^2} \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_8
\end{equation}
A relativistic source moving at constant $\Gamma$ has radius
\begin{equation}
R = \frac{4}{3} \frac{ct}{z+1} \Gamma^2
\label{R}
\end{equation}
at observer-frame time $t$, the factor 4/3 corresponding to photons emitted from the "edge"
of the source, i.e. from the fluid flowing at an angle $\Gamma^{-1}$ relative to the direction
toward the observer. Thus, the optical thickness to pair-formation for a $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb$ photon is
\begin{equation}
\tau_{\gamma\gamma} (\varepsilon} \def\eps{\epsilon} \def\nnbb{nb) = \bar{\sigma}_{\gamma\gamma} \frac{N_\gamma [> \eps_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb)]}{4 \pi R^2}
= 1.8\, \z^6 \frac{\phi5}{\Gamma_2^6 t_1^2} \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_8
\end{equation}
where $\bar{\sigma}_{\gamma\gamma} = 0.18 \sigma_e$ is the pair-formation cross-section
averaged over the $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$ photon distribution and $\sigma_e$ is the lepton scattering
cross-section.
Thus, the afterglow photon front is optically thick to pair-formation ($\tau_{\gamma\gamma} >1$)
for photons of energy above
\begin{equation}
\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm = 56\, \z^{-6} \frac{\Gamma_2^6 t_1^2}{\phi5} \; {\rm MeV}
\label{epm2}
\end{equation}
We approximate the number of pairs formed as that of photons with energy $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb > \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$.
A fraction $\tau_{\gamma\gamma} (\varepsilon} \def\eps{\epsilon} \def\nnbb{nb) < 1$ of the photons with $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb < \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ also form pairs;
these pairs are roughly a factor $\ln (\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm/\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_m) > 1$ more numerous than those for
which $\tau_{\gamma\gamma} (\varepsilon} \def\eps{\epsilon} \def\nnbb{nb) > 1$, $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ being the peak energy of the $\nu F_\nu$ LAT spectrum.
This approximation is made for two reasons. One is to avoid carrying unknown observables --
$\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_m$ and the spectral slope below that peak -- in the following calculations, the other
is to work with a single power-law pair distribution with energy.
The ensuing underestimation of the true pair number increases with $\Gamma$ because $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm
\propto \Gamma^6$. However, for high Lorentz factors, pairs are cooling much faster than they
are created, and only the pairs formed from photons with $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb > \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ radiate synchrotron
emission at the frequencies of interest (optical and X-rays).
Given that each photon with energy above $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ yields two leptons (an electron and a positron),
it follows that the number of leptons formed is
\begin{equation}
N = 2 N_\gamma(> \eps_\pm) = 1.34 \times 10^{56} \z^8 \frac{\phi5^2}{\Gamma_2^6 t_1^2}
\label{N2}
\end{equation}
The above results hold when the pair-formation threshold energy for $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ photons
\begin{equation}
\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) = 21\, \z^4 \frac{\phi5}{\Gamma_2^4 t_1^2} \; {\rm MeV}
\label{et}
\end{equation}
is below $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$, i.e. when there are enough absorbing photons above $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm)$.
The condition $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) = \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ defines a Lorentz factor
\begin{equation}
\Gamma_c \equiv 91\, \z \frac{\phi5^{0.2}}{t_1^{0.4}}
\label{Gc}
\end{equation}
such that $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) < \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ for $\Gamma > \Gamma_c$.
For $\Gamma < \Gamma_c$, we have $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) > \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$, i.e. there there are fewer
photons above $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm)$ than above $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ and, consequently, not all photons
above $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ can form pairs, even though $\tau_{\gamma \gamma}(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) > 1$.
In this case, the energy $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ above which all photons form pairs is given by
$\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) = \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$. Then, equation (\ref{ethr}) leads to
\begin{equation}
\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm = \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm) = \frac{2 \Gamma m_e c^2}{z+1} = 34\, \z^{-1} \Gamma_2 \; {\rm MeV}
\quad (\Gamma < \Gamma_c)
\label{epm1}
\end{equation}
Because the $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ form pairs mostly with other $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ photons, the number of
leptons formed is just the number of photon with energy above $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$:
\begin{equation}
N = N_\gamma(> \eps_\pm) = 1.11 \times 10^{56} \z^3 \frac{\phi5}{\Gamma_2} \quad (\Gamma < \Gamma_c)
\label{N1}
\end{equation}
using equation (\ref{NN}).
Most pairs form at threshold, with a typical {\sl lab}-frame energy $\gamma m_e c^2 = (z+1)
[\varepsilon} \def\eps{\epsilon} \def\nnbb{nb + \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb)]/2$, thus
\begin{equation}
\gamma m_e c^2 \simeq (z+1) \left\{ \begin{array}{ll}
\varepsilon} \def\eps{\epsilon} \def\nnbb{nb & [\Gamma < \Gamma_c, \varepsilon} \def\eps{\epsilon} \def\nnbb{nb = \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb)] \\
\varepsilon} \def\eps{\epsilon} \def\nnbb{nb/2 & [\Gamma \gg \Gamma_c, \varepsilon} \def\eps{\epsilon} \def\nnbb{nb \gg \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb)] \\
\end{array} \right.
\label{lab}
\end{equation}
This approximate one-to-one correspondence between the absorbed photon energy and the
electron-positron pair energy implies that the distribution with energy of pairs
is that of the high-energy photons
\begin{equation}
\frac{dN}{d\gamma} (>\gamma_i) \propto \gamma^{-2}
\label{plaw}
\end{equation}
The above pair distribution with energy holds above a {\sl shock}-frame energy $\gamma'_i$
that can be determined from the pair lab-frame energy $\gamma_i$ corresponding to the minimum
energy $(z+1)\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm$ above which all photons form pairs, and from the angle $\theta_\pm$
at which the pairs move (in the lab-frame) relative to the radial direction of shock's motion.
Taking into account that the high-energy photons are collimated (by the relativistic motion of
their source, the shock) within an angle $\theta_\gamma \simeq \Gamma^{-1}$ of the radial direction,
it follows that
$i)$ the center-of-momentum of the colliding photons moves at angle $\theta_{cm} \sim \Gamma^{-1}$
relative to the radial direction and
$ii)$ the pairs emerge at a typical angle $\theta_{out} = (\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t)^{1/2}/
[\Gamma(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb + \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t)]$ relative to the direction of motion of the center-of-momentum.
In the lab-frame, the emerging pair moves at an angle $\theta_\pm \simeq \max (\theta_{cm},
\theta_{out})$ relative to the shock's direction of motion.
For $\Gamma < \Gamma_c$, we have $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm = \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm)$, thus $\theta_{out} =
1/(2\Gamma)$, from where $\theta_\pm \simeq \theta_{cm} \simeq \Gamma^{-1}$, which implies that,
in the shock-frame (moving outward at Lorentz factor $\Gamma$), the minimal pair energy is
$\gamma'_i \simeq \gamma_i/\Gamma$. From equations (\ref{epm1}) and (\ref{lab}),
$\gamma_i m_e c^2 = (z+1)\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm = 2 \Gamma m_e c^2$, thus
\begin{equation}
\gamma'_i = 2 \quad (\Gamma < \Gamma_c)
\label{gi1}
\end{equation}
For $\Gamma \gg \Gamma_c$: $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm \gg \varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm)$, thus $\theta_{out} \simeq
(\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t/\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm)^{1/2}/\Gamma \ll \Gamma^{-1}$, from where $\theta_\pm \simeq \theta_{cm}
\simeq \Gamma^{-1}$, which implies that $\gamma'_i \simeq \gamma_i/\Gamma$, as for $\Gamma < \Gamma_c$.
From equation (\ref{lab}), $\gamma_i m_e c^2 = (z+1)\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm/2$, which, together with equation
(\ref{epm2}), leads to
\begin{equation}
\gamma'_i = \frac{(z+1)\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_\pm}{2\Gamma m_e c^2} = 1.6\, \z^{-5} \frac{\Gamma_2^5 t_1^2}{\phi5}
= \left( \frac{\Gamma}{\Gamma_c} \right)^5 \quad (\Gamma \gg \Gamma_c)
\label{gi2}
\end{equation}
\vspace*{2mm}
\section{Synchrotron and inverse-Compton spectral characteristics}
\label{syic}
For a power-law distribution with energy of the radiating particles, the synchrotron spectrum
is a sequence of power-laws with breaks at frequencies
\begin{equation}
\nu_{i,a,c} = \frac{e}{2 m_e c} \frac{\gamma_i^2 B \Gamma}{z+1} =
3.0 \times 10^8 \z^{-1} \gamma_{i,a,c}^2 B_0 \Gamma_2 \; {\rm Hz}
\label{nui}
\end{equation}
where $\gamma$ is the pair random Lorentz factor in the shock's frame (prime notation dropped).
$\nu_i$ is the {\sl injection} frequency corresponding to the $\gamma_i$ (equations \ref{gi1}
and \ref{gi2}). $\nu_a$ is the synchrotron {\sl self-absorption} frequency and $\nu_c$ is the
{\sl cooling} frequency; both are calculated below.
$B$ is the magnetic field in the forward-shock, parametrized by the fraction $b$ of the
post-shock energy density $u' = \Gamma n' m_p c^2$ that it contains ($u'_B = B^2/8 \pi$),
with $n' = 4\Gamma n$ the post-shock proton density and $n$ the external medium proton density
at the location $R(t)$ of the forward-shock. Thus
\begin{equation}
B = (32\pi b \Gamma^2 n m_p c^2)^{1/2} = 39\,(n b)^{1/2} \Gamma_2 \; {\rm G}
\label{B}
\end{equation}
The synchrotron and inverse-Compton emissions depend on $n$ and $b$ only through the product $nb$.
The continuous creation of pairs in the shocked medium, with the power-law distribution given
in equation (\ref{plaw}), and their radiative cooling leads to an effective pair distribution
with energy
\begin{equation}
\frac{dN}{d\gamma} \propto \left\{ \begin{array}{ll}
\gamma^{-2} & (\gamma_p < \gamma < \gamma_b) \\
\gamma^{-3} & (\gamma_b < \gamma) \\
\end{array} \right.
\label{dndg}
\end{equation}
with
\begin{equation}
\gamma_p \equiv \min(\gamma_i,\gamma_c) \;,\; \gamma_b \equiv \max(\gamma_i,\gamma_c)
\end{equation}
where $\gamma_c$ is the pair {\sl cooling} Lorentz factor, defined as the energy to which a pair
cools on the dynamical timescale $t'_d=R/(c\Gamma) = (4/3) \Gamma [t/(z+1)]$. The pair number $N$
and minimum Lorentz factor $\gamma_i$ change on a dynamical timescale.
Pairs created with an energy $\gamma m_e c^2$ in the shock-frame cool radiatively at power
$P' (\gamma) = (4/3)\sigma_e c (B^2/8 \pi) (Y+1) \gamma^2 \equiv c_r \gamma^2 $, where $Y$ is
the Compton parameter (the inverse-Compton to synchrotron power ratio). Integrating the equation
for pair cooling, $c_r \gamma^2 = - d(\gamma m_e c^2)/dt'$, one obtains that a pair of high
initial energy reaches a random Lorentz factor $\gamma (t') = m_e c^2/c_r t'$ after a time $t'$
since its creation.
Therefore, the cooling Lorentz factor is $\gamma_c \equiv \gamma (t'_d) = m_e c^2/c_r t'_d$
\begin{equation}
\gamma_c = \frac{9\pi}{2} (z+1) \frac{m_e c}{\sigma_e} \frac{1}{\Gamma t B^2 (Y+1)}
\end{equation}
\begin{equation}
\gamma_c = \frac{1150\,\z}{(nb) t_1 \Gamma_2^3} \frac{1}{Y+\Theta (\gamma_c-\gamma_a)}
\label{gc}
\end{equation}
where $\Theta (x) = 0$ for $x < 0$, $\Theta (x) = 1$ for $x > 0$, and $\gamma_a$ is the random Lorentz
factor of the pairs that radiate at the synchrotron self-absorption frequency $\nu_a$ (see below).
Equation (\ref{gc}) approximates the effect of synchrotron self-absorption by switching
off synchrotron cooling when $\gamma_c$ has decreased to $\gamma_a$. After that, the radiative
cooling continues only through inverse-Compton scatterings.
This means that, if $\gamma_c$ calculated from equation (\ref{gc}) for $Y < 1 $ and $\Theta = 0$
is larger than $\gamma_a$, then the correct cooling Lorentz factor is $\gamma_c \mathrel{\rlap{\raise 0.511ex \hbox{$<$}}{\lower 0.511ex \hbox{$\sim$}}} \gamma_a$.
From equation (6.53) of Rybicki \& Lightman (1979) for the synchrotron self-absorption coefficient
for a power-law distribution of particles, it can be shown that the pair optical-thickness to
{\sl self-absorption} at the peak energy $\nu_p = \min(\nu_i,\nu_c)$ of the intrinsic synchrotron
spectrum $F_\nu$ is
\begin{equation}
\tau_p = \frac{5e \tpm}{\sigma_e B\gamma_p^5} = 3.6 \times 10^{15} \frac{\tpm}{B_0\gamma_p^5}
\label{taua}
\end{equation}
where
\begin{equation}
\tpm = \frac{\sigma_e N}{4\pi R^2}
\label{taupm}
\end{equation}
is the pair optical-thickness to photon {\sl scattering}. For the pair distribution given in equation
(\ref{dndg}), the optical-thickness to self-absorption at frequency $\nu$ is
\begin{equation}
\tau_a (\nu) = \tau_p \left\{ \begin{array}{ll}
\displaystyle{ \left( \frac{\nu_p}{\nu} \right)^{5/3} } & (\nu < \nu_p) \\
\displaystyle{ \left( \frac{\nu_p}{\nu} \right)^3 } & (\nu_p < \nu < \nu_b) \\
\displaystyle{ \left( \frac{\nu_p}{\nu_b} \right)^3 \left( \frac{\nu_b}{\nu} \right)^{7/2} } &
(\nu_b < \nu) \\ \end{array} \right.
\label{taua1}
\end{equation}
where $\nu_b = \max (\nu_p,\nu_c)$.
From here, the self-absorption frequency $\nu_a$ defined by $\tau_a (\nu_a) = 1$ has a corresponding
pair Lorentz factor $\gamma_a$ given by
\begin{equation}
\gamma_a = \left\{ \begin{array}{ll}
\gamma_p \tau_p^{3/10} & (\gamma_a < \gamma_p) \\
\gamma_p \tau_p^{1/6} & (\gamma_p < \gamma_a < \gamma_b) \\
\left( \gamma_p^6 \gamma_b \tau_p \right)^{1/7} & (\gamma_b < \gamma_a) \\
\end{array} \right.
\label{ga}
\end{equation}
The Compton parameter $Y$ is the ratio of the pairs energy output in inverse-Compton emission
to that in synchrotron, hence $Y = P_{ic}/P_{sy} = u'_{sy}/u'_B$, where $u'_{sy}$ is the energy
density of the synchrotron photons received by a scattering lepton. Synchrotron self-absorption
reduces $u'_{sy}$ because photons of energy less than $h\nu_a$ are absorbed before being scattered.
However, that reduction is not substantial for the particle distribution given in equation
(\ref{dndg}) because most of the synchrotron energetic output is at frequencies above $\nu_a$,
for which the pairs medium is transparent (to self-absorption):
$i)$ for $\nu_a < \nu_{max}$ with $\nu_{max} \equiv \max(\nu_i,\nu_c)$, the synchrotron output
above $\nu_a$ is $\nu F_\nu \propto \nu^{1/2}$ at $\nu < \nu_{max}$ and $\nu F_\nu \propto \nu^0$
at $\nu_{max} < \nu$, thus all the synchrotron output is above $\nu_a$;
$ii)$ for $\nu_{max} < \nu_a$, the synchrotron output is $\nu F_\nu \propto \nu^{7/2}$
for $\nu_{max} < \nu < \nu_a$ and $\nu F_\nu \propto \nu^0$ at $\nu_a < \nu$,
hence the reduction of the synchrotron output due to self-absorption is a factor
$\ln (\nu_{br}/\nu_a)/\ln (\nu_{br}/\nu_{max})$ with $\nu_{br}$ the high-energy end of
the otherwise diverging $F_\nu (>\nu_{max}) \propto \nu^{-1}$ synchrotron spectrum, thus the
reduction is of order unity for $\nu_{br} \gg \nu_a$.
For a single photon scattering, the ratio $u'_{sy}/u'_B$ is the product of $(4/3)\overline{\gamma^2}$
(the average increase in photon energy due to scattering) and the fraction $\min(\tpm,1)$ of
photons that are upscattered. $\overline{\gamma^2} \simeq \gamma_i \gamma_c$ for the pair distribution
of equation (\ref{dndg}), thus
\begin{equation}
Y_1 = \frac{4}{3} \gamma_i \gamma_c \min(\tpm, 1)
\label{Y}
\end{equation}
for the first scattering.
In the lab-frame, the relative velocity between a photon (moving at $c$) and the pair-front
(moving at Lorentz factor $\Gamma$) is $v_r = c/(2\Gamma^2)$, thus the photon crosses the
pair-front of geometrical thickness $\Delta = R/(2 \Gamma^2)$ in a time $t_+ = \Delta/v_r = R/c = t_d$.
This means that, for $\tpm < 1$, only the first scattering takes place within a dynamical
timescale $t_d$, and higher order scatterings occur on a longer timescale.
Because the effective pair distribution with energy is that resulting from pair creation
and cooling over one $t_d$, higher order scatterings (taking longer than $t_d$) are ignored
for $\tpm < 1$. However, higher order scatterings should be considered for $\tpm > 1$ because,
in that case, the time between scatterings is less than $t_d$.
\vspace*{2mm}
\section{Regions in the ($\nnbb-\Gamma$) parameter space}
\label{regions}
To obtain the pair synchrotron self-Compton emission at some observing frequency $\nu$, one must
calculate first the spectral breaks of the previous section. The calculation of the injection break
$\nu_i$ is trivial, the analytical expression of the cooling break $\nu_c$ depends on whether
the Compton parameter $Y$ is below or above unity, that of $Y$ depends on whether the pair
optical thickness $\tpm$ is above/below unity. In general, the self-absorption break $\nu_a$
is not needed for the synchrotron emission, as the optical thickness to self-absorption
$\tau_a(\nu)$ suffices. However, the location of $\nu_a$ relative to $\nu_c$ is useful for the
calculation of $\nu_c$ when $Y < 1$ (see equation \ref{gc}), and is required for determining
the upscattered absorption break of the inverse-Compton spectrum.
Given the observables $\Phi$, $t$, and $z$, all quantities needed depend on only two parameters:
$nb$ and $\Gamma$. The conditions $\tpm = 1$, $Y=1$, and $\gamma_c = 1$ define {\sl lines} in this
$nb-\Gamma$ plane while the equality of two spectral breaks defines {\sl boundaries}. Expressions
for these lines and boundaries are derived below and are useful for a correct calculation of the
synchrotron/inverse-Compton spectral breaks and of the peak flux for each spectral component.
The first separation of the $nb-\Gamma$ plane is provided by $\Gamma = \Gamma_c$ (equation \ref{Gc}),
across which $N$ changes from equation (\ref{N2}) to (\ref{N1}) and $\gamma_i$ from equation
(\ref{gi1}) to (\ref{gi2}). A similarly simple separation is provided by the condition $\tpm = 1$.
Using equations (\ref{R}), (\ref{N2}), (\ref{N1}), and (\ref{taupm}), the pair optical-thickness to
photon scattering is
\begin{equation}
\tpm = \left\{ \begin{array}{ll}
3.3\, \z^5 \phi5 t_1^{-2} \Gamma_2^{-5} & \Gamma < \Gamma_c \\
4.0\, \z^{10} \phi5^2 t_1^{-4} \Gamma_2^{-10} & \Gamma_c \ll \Gamma \\
\end{array} \right.
\label{tpm0}
\end{equation}
which, after using equation (\ref{Gc}), can be written as
\begin{equation}
\tpm = \frac{\sigma_e}{\bar{\sigma}_{\gamma\gamma}} \left\{ \begin{array}{ll}
( \Gamma_c/\Gamma )^{5} & \Gamma < \Gamma_c \\
2( \Gamma_c/\Gamma )^{10} & \Gamma_c \ll \Gamma \\ \end{array} \right.
\label{tpm}
\end{equation}
This implies that $\tpm < 1$ for $\Gamma > \Gamma_\tau$ where
\begin{equation}
\Gamma_\tau \equiv \left( \frac{2\sigma_e}{\bar{\sigma}_{\gamma\gamma}} \right)^{1/10} \Gamma_c =
1.27\; \Gamma_c = 115\, \z \frac{\phi5^{0.2}}{t_1^{0.4}}
\end{equation}
The extrapolations of the $\Gamma < \Gamma_c$ and $\Gamma \gg \Gamma_c$ branches of equations
(\ref{gi2}) and (\ref{tpm}) intersect at $\tGc \equiv 2^{1/5} \Gamma_c$ for both
equations. For that reason, we approximate them by
\begin{equation}
\tpm = \frac{\sigma_e}{2\bar{\sigma}_{\gamma\gamma}} \left\{ \begin{array}{ll}
(\tGc/\Gamma )^{5} & \Gamma < \tGc \\ (\tGc/\Gamma )^{10} & \tGc < \Gamma \\ \end{array} \right.
\label{tpm1}
\end{equation}
\begin{equation}
\gamma_i = \left\{ \begin{array}{ll} 2 & (\Gamma < \tilde{\Gamma}_c) \\
2 (\Gamma/\tGc )^5 & (\tGc < \Gamma) \\ \end{array} \right.
\quad \tGc \equiv 104\, \z \frac{\phi5^{0.2}}{t_1^{0.4}}
\label{gi3}
\end{equation}
The range $\tGc < \Gamma < \Gamma_\tau$, where $\gamma_i > 2$ and $\tpm > 1$,
is narrow, its 10 percent extent in $\Gamma$ corresponding to a 30 percent increase in observer
time, for a shock decelerating as $\Gamma \propto t^{-3/8}$ (homogeneous external medium),
i.e. it lasts less than one dynamical timescale.
\begin{figure*}
\vspace*{-5mm}
\centerline{\psfig{figure=lines.eps,height=90mm,width=190mm}}
\figcaption{ {\sl Boundaries} (solid curves, labeled in right panel) and {\sl Lines} (dotted and
dashed curves, labeled in the left panel) in the $nb - \Gamma$ plane, and regions corresponding
to all possible orderings of spectral breaks (separated by the boundaries identified in right panel),
for the synchrotron emission from pairs formed at $z=2$ from high-energy photons of fluence
$\Phi (>100\,{\rm MeV})=10^{-5} \ergcm2$ at observer time $t = 10$ s.
$\Gamma$ is the Lorentz factor of the pair-source at that time, and $nb$ is its magnetic field parameter.
The boundaries intersect at the "triple point" $(nb,\Gamma) = (0.035,250)$, where all three breaks
are equal.
For $\Gamma < \Gamma_\tau$, the pairs are optical thick to photon scattering $\tau > 1$, but
for $nb > (nb)_\tau$, the hot pairs ($\gamma > 1$) are optically thin. }
\end{figure*}
\vspace*{2mm}
\subsection{Optically-thin pairs: $\Gamma > \Gamma_\tau$}
Equation (\ref{Y}) is simply
\begin{equation}
Y = \frac{4}{3} \gamma_i \gamma_c \tpm (= Y_1)
\label{Y1}
\end{equation}
retaining only the first scattering, together with equation (\ref{gc}) leading to
\begin{equation}
Y(Y+1) = \frac{(nb)_Y}{nb} \;, \quad (nb)_Y \equiv 10^4\, \z^6 \frac{\phi5}{t_1^3} \Gamma_2^{-8}
\label{nbY}
\end{equation}
The line $nb = (nb)_Y$ separates the $nb-\Gamma$ plane in two regions, with $Y < 1$ for
$nb > (nb)_Y$ and $Y > 1$ for $nb < (nb)_Y$ (see Figure 1).
\subsubsection{\boldmath $nb < (nb)_Y$ \unboldmath (iC cooling)}
In this case $Y > 1$, and equations (\ref{gc}) and (\ref{nbY}) yield
\begin{equation}
\gamma_c = 11\, \z^{-2} \left( \frac{t_1 \Gamma_2^2}{(nb) \phi5} \right)^{1/2} \;, \quad
Y \simeq \left( \frac{(nb)_Y}{nb} \right)^{1/2}
\label{Y2}
\end{equation}
\centerline{ $(\tau < 1, Y > 1, \gamma_c > 1)$ }
From equations (\ref{gi2}) and (\ref{Y2}), the condition $\gamma_i = \gamma_c$ defines the boundary
\begin{equation}
(nb)_{ci} \equiv 47\, \z^6 \frac{\phi5}{t_1^3} \Gamma_2^{-8}
\label{nbci}
\end{equation}
so that $\gamma_i < \gamma_c$ for $nb < (nb)_{ci}$. From equations (\ref{nbY}) and (\ref{nbci}),
it follows that
\begin{equation}
\frac{(nb)_{ci}}{(nb)_Y}=\frac{9}{64}\left(\frac{\bar{\sigma}_{\gamma\gamma}}{\sigma_e}\right)^2 =
\frac{1}{219}
\end{equation}
thus $(nb)_{ci}$ is just a shift of $(nb)_Y$.
For $nb < (nb)_{ci}$, we have $\gamma_i < \gamma_c$ and the condition $\gamma_a = \gamma_i$
is equivalent to $\tau_p \equiv \tau_a (\gamma_i) = 1$ which, after using equations (\ref{gi2}),
(\ref{taua}), and (\ref{tpm0}), is satisfied on the boundary
\begin{equation}
(nb)_{ia} \equiv 0.014\, \z^{70} \frac{\phi5^{14}}{t_1^{28}} \Gamma_{2.4}^{-72}
\end{equation}
with $\gamma_i < \gamma_a$ for $nb < (nb)_{ia}$.
In this regime, $\gamma_a = \gamma_i \tau_p^{1/6}$; using equations (\ref{gi2}), (\ref{taua}),
(\ref{tpm0}), and (\ref{Y2}), $\gamma_a = \gamma_c$ is satisfied on the boundary
\begin{equation}
(nb)_{ca} \equiv 0.011 \z^{-6.8} \frac{t_1^2}{\phi5^{1.6}} \Gamma_{2.3}^{4.8}
\end{equation}
with $\gamma_a < \gamma_c$ if $nb < (nb)_{ca}$.
For $nb > (nb)_{ci}$, we have $\gamma_c < \gamma_i$ and the $\gamma_a = \gamma_c$ boundary
is defined by $\tau_p \equiv \tau_a (\gamma_c) = 1$ which, together with equations (\ref{taua}),
(\ref{tpm0}), and (\ref{Y2}), leads to
\begin{equation}
(nb)_{ac} \equiv 8.8\, \z^{-10} \frac{t_1^{3.25}}{\phi5^{2.25}} \Gamma_{2.7}^8
\end{equation}
with $\gamma_a > \gamma_c$ for $nb > (nb)_{ac}$. In this case, $\gamma_a = \gamma_c \tau_p^{1/6}$;
using equations (\ref{gi2}), (\ref{taua}), (\ref{tpm0}), and (\ref{Y2}), it follows that
$\gamma_a = \gamma_i$ on the boundary
\begin{equation}
(nb)_{ai} \equiv 200\, \z^{38} \frac{\phi5^{7.5}}{t_1^{15.5}} \Gamma_{2.3}^{-40}
\label{ai}
\end{equation}
such that $\gamma_i < \gamma_a$ if $nb < (nb)_{ai}$.
\vspace*{3mm}
\subsubsection{\boldmath $(nb)_Y < nb < (nb)_{a/c}$ \unboldmath (sy cooling)}
In this case $Y < 1$; for $\gamma_a < \gamma_c$, pairs cool mostly through synchrotron
emission and equations (\ref{gc}) and (\ref{nbY}) yield
\begin{equation}
\gamma_c \simeq \frac{1.15\,\z}{(nb) t_1 \Gamma_3^3} \;,\; Y = \frac{(nb)_Y}{nb} \quad
(\tau < 1, Y < 1, \gamma_c > \gamma_a)
\label{Y3}
\end{equation}
Because $(nb)_Y > (nb)_{ci}$, we have $\gamma_c < \gamma_i$, and the condition $\gamma_a = \gamma_c$
is equivalent to $\tau_p \equiv \tau_a (\gamma_c) = 1$ which, using equations (\ref{taua}),
(\ref{tpm0}), and (\ref{Y3}), is satisfied on the boundary
\begin{equation}
(nb)_{a/c} \equiv 0.19\, \frac{\z^{-1.11}}{t_1^{0.22} \phi5^{0.44}} \Gamma_3^{-0.89}
\end{equation}
\subsubsection{\boldmath $ nb > \max[(nb)_Y, (nb)_{a/c}]$ \unboldmath}
For $nb > (nb)_{a/c}$, the $\gamma_c < \gamma_a$ regime occurs; without synchrotron
cooling, $\gamma_c \propto Y^{-1}$, and the expressions for $Y$ and $\gamma_c$ are the
same as in equation (\ref{Y2}) for $Y > 1$, although $Y < 1$ now.
Consequently, the $\gamma_a = \gamma_i$ and $\gamma_a = \gamma_c$ boundaries are the same
as for $Y>1$: $(nb)_{ai}$ and $(nb)_{ac}$, respectively.
In contrast with that case,
a new region appears now, defined by $(nb)_{a/c} < nb < (nb)_{ac}$, for which $\gamma_c$
of equation (\ref{Y2}) does not satisfy the $\gamma_c < \gamma_a$ condition. This is the
case where inverse-Compton cooling, operating alone after the epoch $t'_{ac} < t'_d$ when
the cooled pair Lorentz factor $\gamma(t'_{ac})$ has reached $\gamma_a(t'_{ac})$, does not
decrease $\gamma(t')$ significantly until $t'_d$. In this case, we impose $\gamma_c \equiv
\gamma (t'_d) \simeq \gamma_a$, thus
\begin{equation}
\gamma_c = \gamma_a = \left( \frac{5e \tpm}{\sigma_e B} \right)^{1/5} =
5.2\, \z^2 \left( \frac{\phi5^2}{(nb)^{1/2} t_1^4 \Gamma_3^{11}} \right)^{1/5}
\label{gc1}
\end{equation}
\centerline{ $(\tpm < 1, Y < 1, \gamma_c=\gamma_a)$ }\\
\subsubsection{\boldmath $nb > (nb)_{c1}$ \unboldmath or
\boldmath $nb > \max[(nb)_{c2},(nb)_{c3}]$ \unboldmath}
\centerline{\bf $(\gamma_c=1)$ }
In some of the regions identified above, radiative cooling can be strong enough that
$\gamma (t') = 1$ at some time $t' < t'_d$. In this case, $\gamma_c \equiv \gamma (t'_d) = 1$,
and $\gamma_a$ and $Y$ are those given by equations (\ref{ga}) and (\ref{Y1}) with $\gamma_c = 1$.
From equations (\ref{Y2}), (\ref{Y3}), and (\ref{gc1}), the condition $\gamma_c = 1$ defines
three lines:
\begin{equation}
(nb)_{c1} \equiv 1270\, \z^{-4} \frac{t_1}{\phi5^2} \Gamma_{2.5}^2
\end{equation}
\begin{equation}
(nb)_{c2} \equiv 3.6\, \z^{20} \frac{\phi5^4}{t_1^8} \Gamma_{3.3}^{-22}
\label{c2}
\end{equation}
\begin{equation}
(nb)_{c3} \equiv 0.043\, \frac{\z}{t_1} \Gamma_{3.5}^{-3}
\end{equation}
such that $\gamma_c = 1$ if $nb > (nb)_{c1}$ or $nb > \max[(nb)_{c2},(nb)_{c3}]$.
The pair cooling to $\gamma_c=1$ means that a pair of high initial energy loses radiatively
all its energy over a dynamical timescale. From the equation for radiative cooling,
$\gamma (t') = m_e c^2/c_r t'$, it follows that the time for complete cooling is
$t'_c(\gamma \sim 1) \equiv m_e c^2/c_r$. When $t'_c (\gamma \sim 1) < t'_d$, a fraction
$f_h = t'_c (\gamma \sim 1)/t'_d$ of the pairs created in the last dynamical timescale
are hot ($\gamma > 1$) and radiate synchrotron and upscatter/absorb that emission, and
a fraction $1 - f_h $ are cold ($\gamma = 1$), scatter photons without a significant
change in energy and absorb the emission below their characteristic synchrotron frequency.
Noting that $t'_c(\gamma \sim 1)/t'_d = m_e c^2/(c_r t'_d) = \gamma_c^{({\rm eq}. \ref{gc})}$,
with $\gamma_c^{({\rm eq}. \ref{gc})}$ the cooling Lorentz factor of equation (\ref{gc}),
we find that
\begin{equation}
f_h = \gamma_c^{(eq \ref{gc})} = \frac{0.144\,\z}{(nb)_3 t_1 \Gamma^3_{2.3}}
\frac{1}{Y+\Theta (1-\gamma_a)}
\label{fh}
\end{equation}
The parameters of interest -- $\gamma_a,Y$ and peak flux $F_p \propto N$ --
are those of equations (\ref{ga}) and (\ref{Y}) with $\gamma_c = 1$ and for a number of
hot pairs $N_h$ or optical thickness $\tau_h$ satisfying $N_h/N = \tau_h/\tpm = f_h$.
A first consequence of $\tau_h = f_h \tpm = \tpm \gamma_c^{({\rm eq}. \ref{gc})}$ for
$\gamma_c = 1$ is that $Y = (4/3)\gamma_i \tau_h = (4/3)\gamma_i \gamma_c^{(eq \ref{gc})} \tpm$,
thus the Compton parameter given in equation (\ref{Y1}) applies whether pairs cool completely
($f_h < 1$) or not ($f_h = 1$) during one dynamical timescale. That entails the conclusions
that, for $\gamma_c = 1$, $Y=1$ on the $(nb)_Y$ line given in equation (\ref{Y2}), and that
$Y$ is as in equation (\ref{Y2}) for $\gamma_c < \gamma_a$ and as in equation (\ref{Y3}) for
$\gamma_a < \gamma_c$.
Owing to the accumulation of pairs at $\gamma_c = 1$, the synchrotron self-absorption thickness
satisfies $\tau_a (\nu < \nu_c) \propto N$ (all pairs absorb synchrotron emission below $\nu_c$)
and $\tau_a (\nu > \nu_c) \propto N_h$ (only hot pairs absorb above $\nu_c$), yielding a
discontinuity of $\tau_a (\nu)$ at $\nu_c$ and a slight ambiguity on the boundary between
the $\nu_c < \nu_a < \nu_i$ and $\nu_a < \nu_c < \nu_i$ regions, as following.
The $\gamma_a = \gamma_c$ boundary of the $\gamma_c < \gamma_a$ region satisfies
$\tau_a(\nu_c+) = (5e \tau_h /\sigma_e B)^{1/6} = 1$, with $\tau_h = f_h \tpm \propto \tpm/Y$.
Using equations (\ref{tpm0}), (\ref{Y2}), and (\ref{fh}), we find that $\gamma_a = \gamma_c$
along the boundary
\begin{equation}
(nb)_{ac1} \equiv 4.4 \times 10^5\, \z^8 \frac{\phi5^{1.5}}{t_1^{3.5}} \Gamma_3^{-10}
\end{equation}
The $\gamma_a = \gamma_c$ boundary of the $\gamma_a < \gamma_c$ region satisfies
$\tau_a(\nu_c-) = (5e \tau /\sigma_e B)^{1/6} = 1$, leading to the same condition as for
$\gamma_c = 1$ for the $\nu_a = \nu_c < \nu_i$ region (see Figure 1), therefore the $(nb)_{c2}$
line of equation (\ref{c2}) is also a $\gamma_a = \gamma_c$ boundary.
For $\gamma_c < \gamma_a < \gamma_i$, equations (\ref{taua}) and (\ref{ga}) give
$\gamma_a = \gamma_c [\tau_a(\nu_c+)]^{1/6} = (5e \tau_h/\sigma_e B)^{1/6}$; then,
$\tau_h = \tpm \gamma_c^{({\rm eq}. \ref{gc})}$ implies that $\gamma_a = \gamma_i$ along the
$(nb)_{ai}$ boundary given in equation (\ref{ai}), extending into the $\gamma_c =1$ region.
\subsubsection{Klein-Nishina scattering}
The Compton parameter $Y$ of equation (\ref{Y1}) is for inverse-Compton scatterings
in the Thomson regime. The pair synchrotron flux is $\nu F_\nu \propto \nu^{1/2}$ for
$\nu_p < \nu < \nu_b$ and flat above $\nu_b$ for a $C_\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \propto \varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$ spectrum
of the high-energy photons that form pairs. However, for a typical LAT spectrum, which
is slightly softer than $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$, the synchrotron flux $\nu F_\nu$ peaks at $\nu_b$.
For $nb > (nb)_{ci}$, we have $\nu_c < \nu_i$, hence $\nu F_\nu$ peaks at $\nu_i$.
Pairs of energy $\gamma$ scatter synchrotron photons at the $\nu_i$ peak of $\nu F_\nu$
in the Klein-Nishina (KN) regime if $\gamma (h \nu'_i) = \gamma (eh/2 m_e c) \gamma_i^2 B >
m_e c^2$. Staying in the $\gamma_c < \gamma_i$ case, the KN regime will reduce the Compton
parameter significantly if the $\gamma_i$ pairs satisfy the above inequality, i.e. if
$\gamma_i > \gamma_{KN} \equiv (2 m_e^2 c^3/heB)^{1/3}$. Using equations (\ref{gi2}) and
(\ref{B}), the condition $\gamma_i > \gamma_{KN}$ becomes $nb > (nb)_{KN}$ where
\begin{equation}
(nb)_{KN} \equiv 0.29\, \z^{30} \frac{\phi5^6}{t_1^{12}} \Gamma_{2.7}^{-32}
\label{nkb}
\end{equation}
For $nb > 10^{-5}$, we have $(nb)_{KN} > (nb)_{ci}$ (Figure 1), thus $nb > (nb)_{KN}$
implies $nb > (nb)_{ci}$ and the derivation of equation (\ref{nkb}) is self-consistent.
The KN scattering effect on the calculation of the Compton parameter of equation (\ref{Y1})
is as following.
For $nb < (nb)_{KN}$, most pairs, being between $\gamma_c$ and $\gamma_i$, scatter the
synchrotron input in Thomson regime and the KN effect is negligible.
For $nb > (nb)_{KN}$, pairs above an energy $\tilde{\gamma} \equiv \gamma_{KN}^3/\gamma_i^2$
with $\gamma_c < \tilde{\gamma} < \gamma_i$ scatter synchrotron photons in the KN regime
and the Compton $Y$ is reduced; however, $Y < 1$ is very likely because $nb > (nb)_{KN}$
implies $nb > (nb)_Y$ (Figure 1), hence that reduction of $Y$ by the KN effect is
largely inconsequential.
For $nb < 10^{-5}$, we have $\gamma_{KN} < \gamma_i < \gamma_c$; in this case, all pairs
upscatter in the KN regime the $\nu_i$ synchrotron photons and also the $\nu_c$ synchrotron
photons (where $\nu F_\nu$ peaks), thus equation (\ref{Y1}) significantly overestimates the
true Compton parameter, leading to an overestimation of the inverse-Compton flux and an
underestimation of $\gamma_c \propto Y^{-1}$. The latter leaves unchanged the synchrotron
flux below $\nu_c$ but underestimates the synchrotron flux above $\nu_c$. For the rest
of this paper, we will avoid the $nb < 10^{-5}$ region, so that the KN effect can be ignored.
\vspace*{2mm}
\subsection{Optically-thick pairs: $\Gamma < \Gamma_\tau$}
Over one dynamical timescale, a photon undergoes $\tpm$ scatterings and diffuses a distance
$L'_s = \sqrt{\tpm} l'_s$ in the pair-shell comoving frame, where $l'_s = \Delta'/\tpm$ is the
mean free-path between scatterings. Thus, over one dynamical timescale, the observer receives
photons from a geometrical depth $L'_s = \Delta'/\sqrt{\tpm}$, corresponding to a scattering optical
depth $\sqrt{\tpm}$.
A synchrotron photon undergoes $\tpm^2$ scatterings before escaping the pair medium.
In the lab-frame, if the pair medium were stationary, the photon mean free-path between scatterings
would be $l_s = \Delta/\tpm$, with $\Delta = R/(2 \Gamma^2)$ the geometrical thickness of the pair
front. Because the pair medium moves at Lorentz factor $\Gamma$, the lab-frame photon-pair
relative velocity is $v_r = c/(2\Gamma^2)$, hence each scattering takes a time $t_s = 2\Gamma^2 l_s/c=
R/(c\tpm) = t_d/\tpm$. A photon starting from a geometrical depth $x' = \tau_\o (\Delta'/\tpm)$,
corresponding to a scattering optical depth $\tau_\o$, will diffuse through the pair medium a distance
$L'_d = \sqrt{n_s} l'_s$ (measured relative to the forward-shock) after $n_s$ scatterings and will exit
the pair medium in the direction toward the observer when $L'_d = x'$, i.e. after $n_s = (x'/l'_s)^2=
\tau_\o^2$ scatterings, in a lab-frame time $t_{esc} = n_s t_s = \tau_\o^2 t_d/\tpm$.
Therefore, only photons up to an optical thickness depth $\sqrt{\tpm}$ cross the pair shell
within one dynamical timescale, and these photons undergo up to $\tpm$ scatterings.
Because we approximate the effective pair distribution with energy as that established by injection
and cooling over $t_d$, we will consider only $\tpm$ scatterings for the calculation of inverse-Compton
parameter, cooling, and emission.
Hence, the pair cooling $\gamma_c$ is that of equation (\ref{gc}) with the Compton parameter
corresponding to $\tpm$ scatterings: $Y = Y_1 + ... + Y_{\tpm}$. As long as scatterings
occur in the Thomson regime, the Compton parameter of $i$-th scattering is $Y_i = Y_1^i$,
where $Y_1 = (4/3) \gamma_i \gamma_c$ is the Compton parameter of the first scattering
for $\gamma_c > 1$. Because $Y_1 > 1$, we can approximate $Y_1 + ... + Y_{\tpm} =
(Y_1^{\tpm +1} -1)/(Y_1-1) \simeq Y_1^{\tpm}$, thus
\begin{equation}
Y = Y_1^\tpm \;,\; Y_1 = \frac{4}{3} \gamma_i \gamma_c \quad (\tau > 1, \gamma_c > 1)
\end{equation}
Adding equation (\ref{gc}) with $Y > 1$, we find that
\begin{equation}
\gamma_c = \left[ \left( \frac{3}{4\gamma_i} \right)^\tpm
\frac{1150\, \z}{(nb) t_1 \Gamma_2^3} \right]^{1/(\tpm+1)} \quad (\tau > 1, \gamma_c > 1)
\label{gc2}
\end{equation}
After $i$ scatterings, a synchrotron photon of initial energy $h \nu$ reaches an energy
$(\overline{\gamma^2})^i h\nu$, where $\overline{\gamma^2} = \gamma_i \gamma_c$ is the average
$\gamma^2$ for pairs. The $i$-th scattering occurs at the end of the Thomson regime if,
in the pair frame, the $(i-1)$-th scattered photon is below $m_e c^2$ (Thomson scattering)
and the $i$-th scattered photon is above $m_e c^2$ (Klein-Nishina scattering):
$\bar{\gamma} (\overline{\gamma^2})^{i-1} h\nu < m_e c^2 < \bar{\gamma}
(\overline{\gamma^2})^i h\nu$, where $\bar{\gamma} =(\overline{\gamma^2})^{1/2} =
(\gamma_i \gamma_c)^{1/2}$. Thus the $i$-th scattering is borderline in Thomson regime
if $(\gamma_i \gamma_c)^{i-1/2} h\nu < m_e c^2 < (\gamma_i \gamma_c)^{i+1/2} h\nu$,
and we approximate the number of Thomson scatterings $n_T$ by
$(\gamma_i \gamma_c)^{n_T} h\nu = m_e c^2$.
The synchrotron photon of energy $h \nu$ considered here should be that where most of the
synchrotron output $\nu F_\nu$ lies, which is $\max (\nu_i, \nu_c)$, leading to
\begin{equation}
\max(\gamma_i^2,\gamma_c^2) (\gamma_i \gamma_c)^{n_T} = \frac{2 m_e^2 c^3}{heB} =
\frac{3.6 \times 10^{11}}{(nb)^{1/2} \Gamma_2}
\label{heb}
\end{equation}
Substituting here $\gamma_c$ from equation (\ref{gc2}) and using equations (\ref{tpm1})
and (\ref{gi3}), one can determine the $nb-\Gamma$ region where $\tpm < n_T$, i.e. for
which all $\tpm$ scatterings occurring within one dynamical timescale are in the Thomson
regime. The result is that, for the $nb$ range shown in Figure 1, all $\tpm$ upscatterings
occur in the Thomson regime, and that result can be illustrated easier if we use the more
stringent condition that $\nu_c$ photons are upscattered by $\gamma_c$ pairs. Then, from
equation (\ref{heb}) with $\gamma_i$ substituted by $\gamma_c$ and with $\gamma_c$ from
equation (\ref{gc2}), it follows that $\tpm < n_T$ is equivalent to $nb > (nb)_{kn}$ with
\begin{equation}
(nb)_{kn} \equiv 2.4 \times 10^{-4} \left[ \left( \frac{3}{4\gamma_i} \right)^{\tpm}
\frac{\z}{t_1 \Gamma_2^{2.5}} \right]^{4/3}
\end{equation}
Taking into account the dependence on $\Gamma$ of $\tpm$ and $\gamma_i$, we find that $(nb)_{kn}$
above rises with $\Gamma$, reaching a maximum value at $\Gamma_\tau$, where $\tpm = 1$ and
$\gamma_i = 3.3$, hence that maximum value is $(nb)_{kn}(\Gamma_\tau) \simeq 2.10^{-5}$.
Thus, the $nb$ range shown in Figure 1 satisfies $nb > (nb)_{kn}$ and all $\tpm$ upscatterings
occur in the Thomson regime.
\vspace*{2mm}
\subsubsection{\boldmath $\gamma_c > 1$ \unboldmath }
From equation (\ref{gc2}), it follows that $\gamma_c = \gamma_i$ along the boundary
\begin{equation}
(nb)_{ic} \equiv \left( \frac{3}{4\gamma_i^2} \right)^{\tpm}
\frac{1150\, \z}{\gamma_i t_1 \Gamma_2^3}
\end{equation}
that $\gamma_c = \gamma_a (= \gamma_i \tau_p^{1/6})$ for
\begin{equation}
(nb)_{c/a} \equiv \left[ 3.9\, \frac{\z^{-(5\tpm-1)/6}}{(390 \gamma_i)^\tpm}
\frac{t_1^{(\tpm-2)/3} \Gamma_2^{\tpm-2}} {\phi5^{(\tpm +1)/6}} \right]^{12/(11-\tpm)}
\end{equation}
(this boundary exists only for $\tpm < 11$, i.e. $\Gamma > \Gamma_c$),
and that $\gamma_c = 1$ on
\begin{equation}
(nb)_c \equiv \left( \frac{3}{4\gamma_i} \right)^{\tpm} \frac{1150\, \z}{t_1 \Gamma_2^3}
= \gamma_i^{\tpm +1} (nb)_{ic}
\end{equation}
with $\gamma_i$ given in equation (\ref{gi3}). These results become simpler for $\Gamma < \tGc$,
where $\gamma_i = 2$ and the $(nb)_c$ boundary is just a shift of $(nb)_{ic}$.
\vspace*{2mm}
\subsubsection{\boldmath $\gamma_c = 1$ \unboldmath }
For $nb > (nb)_c$, only a fraction
\begin{equation}
f_h = \frac{1150\,\z}{(nb)t_1 \Gamma_2^3} \frac{1}{Y} \;,\;
Y=\sum_{i=1}^{\max(1,\tau_h)} Y_1^i \;,\; Y_1 = \frac{4}{3} \gamma_i \min(1,\tau_h)
\label{fh1}
\end{equation}
\centerline {$(\tau > 1, \gamma_c=1)$} \\ \\
of all pairs {\sl upscatter} synchrotron photons. This result is equation (\ref{fh}) for
$\tau_h = f_h \tpm$ upscatterings (a synchrotron photon undergoes $\tpm$ scatterings in one
dynamical timescale, out of which a fraction $f_h < 1$ are upscatterings by hot pairs) and
for $\gamma_a > \gamma_c = 1$ (as Figure 1 shows for $\Gamma < \Gamma_\tau$).
To find $f_h$, $\tau_h$, and $Y$, equations (\ref{fh1}) with $\tau_h = f_h \tpm$ must be solved.
For $\tau_h > 1$, we have $Y_1>1$ and $Y=\sum_{i=1}^{\tau_h} Y_1^i \simeq Y_1^{\tau_h}$,
thus $\tau_h$ satisfies
\begin{equation}
\left( \frac{4}{3} \gamma_i \right)^{\tau_h} \tau_h = \frac{1150\,\z \tpm}{(nb)t_1 \Gamma_2^3}
\;,\; Y = \left( \frac{4}{3}\gamma_i \right)^{\tau_h} \; (\tau_h > 1, \gamma_c =1)
\label{th}
\end{equation}
From here, $\tau_h$ can be determined numerically, a rough approximation being
$\tau_h \simeq \log [1150\,\z \tpm/(nb)t_1 \Gamma_2^3]/ \log (4 \gamma_i/3)$.
From equation (\ref{th}), the line corresponding to $\tau_h = 1$ is
\begin{equation}
(nb)_\tau = \frac{870\, \z \tpm}{t_1 \Gamma_2^3 \gamma_i} = \left\{ \begin{array}{ll}
2200\, \z^{16} \displaystyle{ \frac{\phi5^3}{t_1^7} \Gamma_2^{-18}} &
\tilde{\Gamma}_c < \Gamma < \Gamma_\tau \\
1470\, \z^6 \displaystyle{ \frac{\phi5}{t_1^3} \Gamma_2^{-8} } & \Gamma < \tilde{\Gamma}_c \\
\end{array} \right.
\end{equation}
For $nb < (nb)_\tau$, the hot pairs are optically thick ($\tau_h > 1$) to photon scattering
while for $nb > (nb)_\tau$, they are optically thin ($\tau_h < 1$) despite that $\tpm > 1$.
For $\tau_h < 1$, at most one upscattering occurs within a dynamical timescale, thus
$Y = Y_1 = (4/3)\gamma_i \tau_h$, and
\begin{equation}
\tau_h = \left( \frac{(nb)_\tau}{nb}\right)^{1/2} \;,\;
Y = \frac{4}{3} \gamma_i \tau_h \quad (\tau_h < 1, \gamma_c =1)
\end{equation}
which implies that $Y > 1$ for $nb < (nb)_Y$ with
\begin{equation}
(nb)_Y = \left( \frac{4}{3} \gamma_i \right)^2 (nb)_\tau =
1540\, \z \frac{\gamma_i \tpm}{t_1 \Gamma_2^3} = (nb)_Y^{({\rm eq}.\ref{nbY})}
\end{equation}
with the last equality resulting from equations (\ref{tpm1}) and (\ref{gi3}).
Thus the $Y=1-(nb)_Y$ line given in equation (\ref{nbY}) for $\Gamma > \Gamma_\tau$ extends
in the $\Gamma < \Gamma_\tau$ region.
We also note that $Y = (4/3)\gamma_i \tau_h$ implies that, for $\Gamma < \tilde{\Gamma}_c$
(i.e. for $\gamma_i = 2$), the $\tau_h = 1$ line is a shift of the $Y=1$ line.
\vspace*{2mm}
\section{Pair emission}
\label{moresyic}
\subsection{Synchrotron}
The intrinsic synchrotron spectrum peaks at $\nu_p = \min(\nu_i,\nu_c)$, where the flux
density is
\begin{equation}
F_p = \frac{z+1}{4\pi d_l^2} \frac{e^3}{m_e c^2} N_h B \Gamma =
6.8\, \z^5 \frac{\phi5^2 (nb)^{1/2}}{t_1^2 \Gamma_{2.5}^4} \; {\rm Jy} \;(\Gamma > \tGc)
\label{Fp}
\end{equation}
with $N_h = N$ if $\gamma_c > 1$ and $N_h = f_h N$ if $\gamma_c = 1$.
From equations (\ref{nui}), (\ref{B}),(\ref{gi2}), (\ref{Y2}), (\ref{Y3}), the injection and
cooling frequencies scale as
\begin{equation}
\nu_i \propto n^{1/2} \z^{-11} \Gamma^{12} \Phi^{-2} t^4 \;,\;
\nu_c \propto \left\{ \begin{array}{ll} \z^{-3} n^{1/2} \Gamma^{-4} t^{-2} & Y < 1 \\
\z^{-5} n^{-1/2} \Gamma^{4} \Phi^{-1} t & Y > 1 \end{array} \right.
\label{nuic}
\end{equation}
For the pair distribution with energy given in equation (\ref{dndg}), the spectrum of the
{\sl unabsorbed} synchrotron emission is
\begin{equation}
F_\nu^{(o)} = F_p \left\{ \begin{array}{ll}
\displaystyle{ \left(\frac{\nu}{\nu_p}\right)^{1/3} } & (\nu<\nu_p) \\
\displaystyle{ \left(\frac{\nu_p}{\nu}\right)^{1/2} } & (\nu_p<\nu<\nu_b) \\
\displaystyle{ \left(\frac{\nu_p}{\nu_b}\right)^{1/2} \frac{\nu_b}{\nu} } & (\nu_b<\nu) \\
\end{array} \right.
\label{Fnu}
\end{equation}
where $\nu_p = \min(\nu_i,\nu_c)$ and $\nu_b = \max(\nu_i,\nu_c)$.
For a LAT photon spectrum $C_\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \propto \varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$, the corresponding pair synchrotron
$\nu F_\nu$ spectrum is flat above $\nu_b$, but peaks at $\nu_b$ for softer LAT photon spectra.
1) $\Gamma > \Gamma_\tau$.
Then $\tpm < 1$ and the received/emergent synchrotron flux is that of equation (\ref{Fnu})
corrected only for self-absorption:
\begin{equation}
F_\nu = \frac{F_\nu^{(o)}}{\max[1,\tau_a(\nu)]} \quad (\tpm < 1)
\label{Fnu1}
\end{equation}
where $\tau_a (\nu)$ is the synchrotron self-absorption optical thickness (equation \ref{taua1})
at observing frequency $\nu$, for all pairs (if $\gamma_c > 1$ or if $\gamma_c =1$ and $\nu < \nu_c$)
or only for hot pairs (if $\gamma_c =1$ and $\nu > \nu_c$).
Equation (\ref{Fnu1}) simply means that the observer receives emission from the entire pair medium
at frequencies where $\tau_a (\nu) < 1$ and from a layer of optical depth (to self-absorption) unity
if $\tau_a (\nu) > 1$.
Substituting equations (\ref{Fp}) and (\ref{nuic}) in (\ref{Fnu}), we arrive at the synchrotron
light-curves listed in Table 1 for $\nu > \nu_a$, $\gamma_c > 1$, and for two types of ambient medium:
{\sl homogeneous}, for which $n$= const and the shock deceleration is given by $\Gamma^2 n R^3 =
\Gamma^2 n (\Gamma^2 t)^3=$ const, hence $\Gamma \propto t^{-3/8}$;
{\sl wind}-like, where $n \propto r^{-2}$, thus $\Gamma \propto t^{-1/4}$, $R \propto t^{1/2}$, and
$n \propto t^{-1}$.
\begin{table*}[t]
\caption{Synchrotron light-curves at frequencies above self-absorption, for optically-thin pairs,
for forward-shock (external) electrons (eqs in Appendixes B and C of Panaitescu \& Kumar 2000),
and for reverse-shock (ejecta) electrons undergoing adiabatic cooling, i.e. after the reverse shock
has crossed the ejecta shell and when there is no further injection of fresh electrons
(the $\nu < \nu_c$ flux scalings are inferred from eqs 52 and 53 of Panaitescu \& Kumar 2004,
the $\nu_c < \nu$ flux decay is that of the "large-angle emission" derived by Kumar \& Panaitescu 2000).
The injected electrons and formed pairs have a power-law distribution with energy of index -2
(equation \ref{plaw}).
$\Phi$ is the fluence at time $t$ of the high-energy photons that form pairs, $Y$ is the
Compton parameter, $\nu_i$ and $\nu_c$ are the injection and cooling frequencies.
The synchrotron spectrum is given in equation (\ref{Fnu}). }
\vspace*{2mm}
\begin{tabular}{cccccccccccccc}
\hline \hline
type of
& $F(\nu<\nu_i<\nu_c)$ & \multicolumn{2}{c}{$F(\nu<\nu_c<\nu_i)$} & $F(\nu_i<\nu<\nu_c)$ &
\multicolumn{2}{c}{$F(\nu_c<\nu<\nu_i)$} & \multicolumn{2}{c}{$F(\nu_i,\nu_c<\nu)$} \\
medium & & $Y < 1$ & $Y > 1$ & & $Y < 1$ & $Y > 1$ & $Y < 1$ & $Y > 1$ \\
\hline \hline
\multicolumn{9}{c}{PAIRS} \\
\hline
homogeneous &
$\Phi^{8/3} t^{-1/3}$ & $\Phi^2 t^{-1/3}$ & $\Phi^{7/3} t^{-1/3}$ & $\Phi t^{-3/4}$ &
$\Phi^2 t^{-3/4}$ & $\Phi^{3/2} t^{-3/4}$ & $\Phi t^{-1}$ & $\Phi^{1/2} t^{-1}$ \\
$n \propto r^{-2}$ wind &
$\Phi^{8/3} t^{-5/3}$ & $\Phi^2 t^{-1}$ & $\Phi^{7/3} t^{-5/3}$ & $\Phi t^{-5/4}$ &
$\Phi^2 t^{-9/4}$ & $\Phi^{3/2} t^{-5/4}$ & $\Phi t^{-2}$ & $\Phi^{1/2} t^{-1}$ \\
\hline \hline
\multicolumn{9}{c}{FORWARD SHOCK} \\
\hline
homogeneous &
$t^{1/2}$ & \multicolumn{2}{c}{$t^{1/6}$} & $t^{-3/4}$ & \multicolumn{2}{c}{$t^{-1/4}$} &
\multicolumn{2}{c}{$t^{-1}$} \\
$n \propto r^{-2}$ wind &
$t^0$ & \multicolumn{2}{c}{$t^{-2/3}$} & $t^{-5/4}$ & \multicolumn{2}{c}{$t^{-1/4}$} &
\multicolumn{2}{c}{$t^{-1}$} \\
\hline \hline
\multicolumn{9}{c}{REVERSE SHOCK (adiabatic cooling)} \\
\hline
homogeneous &
$t^{-0.31}$ & \multicolumn{2}{c}{$t^{-0.31}$} & $t^{-1.11}$ & \multicolumn{2}{c}{$t^{-2.5}$} &
\multicolumn{2}{c}{$t^{-2.5}$} \\
$n \propto r^{-2}$ wind &
$t^{-0.28}$ & \multicolumn{2}{c}{$t^{-0.28}$} & $t^{-1.46}$ & \multicolumn{2}{c}{$t^{-2.5}$} &
\multicolumn{2}{c}{$t^{-2.5}$} \\
\hline \hline
\end{tabular}
\end{table*}
Table 1 illustrates the un-surprising correlation of the pair flux with the fluence $\Phi$ of
the high-energy photons that form the pairs. For the same distribution of leptons with energy,
the pair synchrotron emission always decays faster than the forward-shock's (taking into account
that $\Phi \propto t^{-1.3\pm 0.3}$). Compared with the synchrotron emission from the reverse-shock,
after that shock crossed the ejecta shell and electrons cool adiabatically, the pair emission
decay is steeper at $\nu < \nu_p$ and slower at $\nu > \nu_b$.
2) $\Gamma < \Gamma_\tau$ and $nb > (nb)_\tau$.
In this case, $\tpm > 1$, $\gamma_c = 1$, the scattering optical thickness of hot pairs is $\tau_h < 1$,
i.e. most scatterings are on cold pairs and leave the synchrotron photon energy unchanged.
Scatterings by cold leptons increase the optical thickness to self-absorption to an effective
value $\ttau_a = \sqrt{\tau_a(\tau_a+\tpm)}$. The observer receives emission from a layer of
geometrical thickness $l'_a$ corresponding to one self-absorption optical thickness,
$l'_a = \Delta'/\ttau_a$. Because we take into account only photons that escape the pair medium
in one dynamical timescale (on which the number of pairs changes), i.e. only photons from a
scattering optical depth $\sqrt{\tpm}$, the received synchrotron flux is
\begin{equation}
F_\nu = \frac{F_\nu^{(o)}}{ \max [\sqrt{\tpm}, \ttau_a (\nu) ] }
\quad \ttau_a(\nu) \equiv \sqrt{\tau_a(\nu) [\tau_a(\nu)+\tpm]}
\label{Fnu2}
\end{equation}
or
\begin{equation}
F_\nu = F_\nu^{(o)} \left\{ \begin{array}{ll}
\tpm^{-1/2} & (\tau_a < 1) \\ (\tau_a \tpm)^{-1/2} & (1 < \tau_a < \tpm) \\
\tau_a^{-1} & (\tpm < \tau_a) \\ \end{array} \right.
\end{equation}
3) $\Gamma < \Gamma_\tau$ and $(nb)_\tau > nb > (nb)_c$.
In this case, $\tpm > 1$, $\gamma_c =1$, and $\tau_h > 1$. Because the upscattering of a synchrotron
photon by hot pairs means the "destruction" of the synchrotron photon (which will be counted as an
inverse-Compton photon), upscatterings reduce the emergent synchrotron flux in a fashion similar to
self-absorption
\begin{equation}
F_\nu = \frac{F_\nu^{(o)}}{ \max[ \sqrt{\tpm}, \ttau_a(\nu), \ttau_h ] } \quad (\tpm > 1)
\label{Fnu3}
\end{equation}
\begin{equation}
\ttau_h \equiv \sqrt{\tau_h \tpm} \;,\;
\ttau_a(\nu) \equiv \sqrt{\tau_a(\nu) [\tau_a(\nu)+\tpm_c]}
\end{equation}
where $\tau_c = \tau -\tau_h$ is the scattering optical thickness of cold pairs.
Equation (\ref{Fnu3}) reduces to (\ref{Fnu2}) for $\tau_h < 1$.
The self-absorption frequency $\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a$ defined by $\ttau_a (\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a)= 1$ satisfies
$[\tpm_c + \tau_a(\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a)] \tau_a(\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a) = 1$, which implies that $\tau_a (\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a) < 1$.
Adding that $\tau_a (\nu)$ decreases with $\nu$, this means that $\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a > \nu_a$, where
$\nu_a$ is the self-absorption frequency without scatterings, defined by $\tau_a (\nu_a) = 1$.
Obviously, scatterings by cold leptons (without changing the energy of the synchrotron
photon) increase the self-absorption frequency. For $\tpm_c < \tau_a (\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a)$,
we have $\tau_a (\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a) \simeq 1$, thus $\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a \simeq \nu_a$ (trivial).
For $\tpm_c > \tau_a (\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a)$, we get $\tpm_c \tau_a(\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a) \simeq 1$, which can be solved
easily: the solution is that of equation (\ref{ga}) but with $\tpm_c \tau_p$ instead of $\tau_p$.
The boundaries involving $\nu_a$ shown in Figure 1 remain unchanged for $\tpm > 1$ because
$\gamma_c > 1$, hence $\tpm_c = 0$ (there are not any cold pairs).
4) $\Gamma < \Gamma_\tau$ and $nb < (nb)_c$.
In this case, $\gamma_c > 1$, $\tau_c = 0$, and $\tau_h = \tpm > 1$, thus
\begin{equation}
F_\nu = \frac{F_\nu^{(o)}}{\max [\tau_a(\nu),\tpm] }
\label{Fnu4}
\end{equation}
Note that equation (\ref{Fnu3}) reduces to (\ref{Fnu4}) for $\tau_c = 0$, thus equation (\ref{Fnu3})
applies to all $\tau > 1$ cases.
For the sake of generality, equations (\ref{Fnu1}) and (\ref{Fnu3}) can be combined as
\begin{equation}
F_\nu = \frac{F_\nu^{(o)}}{ \max[ 1, \sqrt{\tpm}, \ttau_a(\nu), \ttau_h, \tpm/3 ] }
\label{Fnu5}
\end{equation}
valid for any $\tpm$ and $\tau_h$. The last term in the denominator accounts for the formation
of pairs ahead of the diffusing photon. As discussed before, a photon starting from a scattering
optical depth $\tau_\o$ will undergo $\tau_\o^2$ scatterings before it diffuses over a distance
$L'_d = \tau_\o l'_s$ (in the comoving-frame) and exits the pair shell. These scatterings take a
time $\delta t' = \tau_\o^2 l'_s/c$, hence the photon diffuses at average speed $v'_d = L'_d/\delta t'
= c/ \tau_\o$ relative to the shocked fluid. For the photon to really escape the pair medium,
that diffusion speed should exceed the velocity $v'_{rel}$ of the outer edge of the pair-medium
relative to the shocked fluid.
When the pair medium is optically thick, most pairs form within the shocked fluid that produced
the pair-forming photons, whose outer edge is the forward-shock, which moves at speed $v'_{sh} = c/3$
relative to the post-shock fluid. Thus, the condition $v'_d > v'_{sh}$ implies that only photons
from optical depth $\tau_\o < 3$ escape the pair medium.
\vspace*{2mm}
\subsection{Inverse-Compton}
\begin{figure*}
\centerline{\psfig{figure=xoG.eps,height=70mm,width=170mm}}
\figcaption{ Optical (left panel) and hard X-ray (right panel) emission from pairs formed
behind the forward-shock from high-energy photons of fluence $\Phi = 10^{-5} \ergcm2$
above 100 MeV, with a $C_\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \propto \varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$ photon spectrum, at observer time $t_\o=10$ s
(during the burst), for various magnetic field parameters $nb$ (legend), and for a range of
source Lorentz factors $\Gamma (t_\o)$.
Dotted lines identify the inverse-Compton emission (iC).
For an optically-thin ($\tau < 1$) pair-medium, the optical emission is only synchrotron;
for $\tau > 1$, it is mostly iC. In contrast, the 100 keV emission can be iC for $\Gamma >
\Gamma_\tpm$. The wiggly X-ray inverse-Compton flux for $\tau > 1$ is due to variations
of the order of inverse-Compton components that escape the optically-thick medium in one
dynamical timescale.
A low $\Gamma$ entails a higher $\tau$ and a low flux from the pairs in the layer of
optical depth $\sqrt{\tau}$, whose emission escapes within one dynamical timescale.
A high $\Gamma$ leads to a high threshold for pair-formation, a low number of pairs formed,
hence a dimmer pair emission.}
\end{figure*}
\vspace*{2mm}
\subsubsection{$\tau < 1 \; (\Gamma > \Gamma_\tau)$}
Only one upscattering occurs during one dynamical timescale, hence $Y=Y_1=(4/3)\gamma_i\gamma_c\tau_h$.
The upscattering of a fraction $\tpm$ of synchrotron photons implies that the peak of the first
inverse-Compton spectrum is
\begin{equation}
F_p^{(1)} = \tau_h F_p^{(sy)}
\end{equation}
with $F_p^{(sy)}$ the flux at the peak of the {\sl emergent} synchrotron spectrum:
\begin{equation}
F_p^{(sy)} = F_p \left\{ \begin{array}{ll}
1 & (\nu_a <\nu_p) \\
\displaystyle{ \left(\frac{\nu_p}{\nu_a}\right)^{1/2} } & (\nu_p<\nu_a<\nu_b) \\
\displaystyle{ \left(\frac{\nu_p}{\nu_b}\right)^{1/2} \frac{\nu_b}{\nu_a} } & (\nu_b<\nu_a) \\
\end{array} \right.
\label{Fpsy}
\end{equation}
with $F_p$ the peak flux of the intrinsic synchrotron spectrum, given in equation (\ref{Fp}),
$\nu_a$ being the self-absorption frequency, $\nu_p = \min(\nu_i,\nu_c)$ is the peak energy of the
$F_\nu$ synchrotron spectrum, and $\nu_b = \max(\nu_i,\nu_c)$ is the peak of the $\nu F_\nu$ synchrotron
spectrum. The emergent synchrotron spectrum peaks at $\max (\nu_p, \nu_a)$.
The spectrum of the first inverse-Compton emission has breaks at
\begin{equation}
\nu_a^{(1)} \equiv \frac{4}{3} \gamma_p^2 \nu_a ,\;
\nu_p^{(1)} \equiv \frac{4}{3} \gamma_p^2 \max (\nu_p,\nu_a) ,\;
\nu_b^{(1)} \equiv \frac{4}{3} \gamma_b^2 \max (\nu_b,\nu_a)
\end{equation}
The shape of the {\sl intrinsic} inverse-Compton spectrum is the same as for the synchrotron
spectrum (equation \ref{Fnu}), but upscattering of the self-absorbed synchrotron emission
$F_\nu \propto \nu^{1/3}/\tau_a(\nu) \propto \nu^2$ yields a softer {\sl emergent} inverse-Compton
spectrum $F_\nu^{(1)} \propto \nu$, that spectrum being
\begin{equation}
F_\nu^{(1)} = F_p^{(1)} \left\{ \begin{array}{ll}
\left( \nu/\nu_a^{(1)} \right) \left( \nu_a^{(1)}/\nu_p^{(1)} \right)^{1/3}
& (\nu<\nu_a^{(1)}<\nu_p^{(1)}) \\
\nu/\nu_a^{(1)} & (\nu<\nu_a^{(1)} , \nu_p^{(1)} < \nu_a^{(1)}) \\
\left( \nu/\nu_p^{(1)} \right)^{1/3} & (\nu_a^{(1)}<\nu<\nu_p^{(1)}) \\
\left( \nu_p^{(1)}/\nu \right)^{1/2} & (\nu_p^{(1)} < \nu < \nu_b^{(1)}) \\
\sqrt{\nu_p^{(1)}\nu_b^{(1)}}/\nu & (\nu_a^{(1)} < \nu_b^{(1)} < \nu) \\
\nu_a^{(1)}/\nu & (\nu_b^{(1)} < \nu_a^{(1)} < \nu) \\ \end{array} \right.
\label{Fic}
\end{equation}
\vspace*{2mm}
\subsubsection{$\tau > 1 \; (\Gamma < \Gamma_\tau)$}
If the effective optical thickness to scattering by hot pairs $\ttau_h = \sqrt{\tau_h \tpm} < 1$
then the above equations for $\tpm < 1$ apply, but with the self-absorption frequency $\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a$
accounting for scatterings by cold pairs instead of $\nu_a$.
For $\ttau_h >1$, the cooling $\gamma_c$ and Compton $Y$ are those calculated in \S3.2 for $\tau_h$
upscatterings that a synchrotron photon, starting from optical depth $\sqrt{\tpm}$, undergoes during
one dynamical timescale.
Taking into account than only photons from a scattering optical depth $\tau_\o < 3$ catch up with
the forward-shock, and that such photons undergo up to 9 scatterings before escaping the pair medium,
out of which $9 \tau_h/\tpm$ are upscatterings, it follows that the maximal inverse-Compton order
to be considered is $iC_{max} = \min(\tau_h,9\tau_h/\tpm)$.
The emergent upscattered emission is the sum of $iC_{max}$ inverse-Compton components
peaking at progressively higher energies. After $i$ upscatterings, the scattered photon has diffused,
on average, a distance $\sqrt{i} l'_s$, where $l'_s = \Delta'/\ttau_h$ is the photon free mean-path
between upscatterings. Therefore, most of the $i$-th inverse-Compton emission arises from a
upscatterings of seed photons produced by a layer located at upscattering optical depth
$\sqrt{i-1}\div \sqrt{i}$, which we will call the "$i$-th iC layer". With increasing distance from it,
inner and outer layers yield a lesser and lesser contribution to the $i$-th inverse-Compton emission.
\begin{figure*}
\centerline{\psfig{figure=xos0.eps,height=70mm,width=170mm}}
\figcaption{ Optical (left) and soft X-ray (right) light-curves from pairs produced by high-energy
photons with fluence $\Phi = 10^{-5} (t/t_\o)^{-1/3} \ergcm2$ and $C_\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \propto \varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$
photon spectrum, for a fixed magnetic field parameter $nb = 10^{-2}$, and a shock deceleration
$\Gamma = \Gamma_\o (t/t_\o)^{-3/8}$ corresponding to a {\sl homogeneous} medium (of particle
density $n$), starting from various $\Gamma_\o$ at $t_\o = 10$ s.
The dominant emission process is indicated, as well as the origin of light-curve breaks
("c=o" means cooling frequency crossing the optical). Only for the lowest $\Gamma_\o = 100$,
the pair medium is optically thick to photon scattering, until the time of the inverse-Compton
light-curve. }
\end{figure*}
For ease of calculations and accounting of the seed synchrotron photons, we assume that the $i$-th
inverse-Compton emission arises from upscatterings of synchrotron photons produced {\sl only} by the
$i$-th iC layer. Given that all seed photons from $i>1$ layers are upscattered, this one-to-one
correspondence between inverse-Compton order and optical depth should entail that {\sl all} the
synchrotron photons produced by the $i$-th iC layer become $i$-th inverse-Compton photons,
thus the peak flux of the $i$-th inverse-Compton spectrum is equal to the peak flux of the
synchrotron spectrum for the $i$-th iC layer:
\begin{equation}
F_p^{(i)} = \frac{\sqrt{i}-\sqrt{i-1}}{\ttau_h} F_p^{(sy)} \quad
[i=1 \div \min(\tau_h,9\tau_h/\tpm)]
\end{equation}
where $F_p^{(sy)}$ is the flux at the peak of the synchrotron spectrum for the entire pair medium,
given in equation (\ref{Fpsy}), but with $\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a$ corresponding to self-absorption only in the
$i$-th iC layer:
\begin{equation}
\ttau_a^{(i)} (\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a ) = 1 \;,\;
\ttau_a^{(i)} \equiv \sqrt{\tau_a^{(i)} \left(\tpm_c^{(i)} + \tau_a^{(i)} \right)}
\end{equation}
where $\tpm_c^{(i)} \equiv (\sqrt{i}-\sqrt{i-1}) \tpm_c/\ttau_h$ is the optical thickness
to scattering by cold pairs in the $i$-th iC layer (although the $i$-th iC layer is
optically thin to upscatterings by hot pairs, it is not necessarily thin to scatterings
by cold pairs), and with $\tau_a^{(i)}(\nu)$ as given in equation (\ref{taua1}) but
for the scattering optical thickness $\tpm^{(i)} \equiv (\sqrt{i}-\sqrt{i-1}) \tpm/\ttau_h$
of the $i$-th iC layer.
The flux of the $i$-th inverse-Compton emission is as given in equation (\ref{Fic}) but with
$F_p^{(i)}$ instead of $F_p^{(1)}$ and with spectral breaks at
\begin{equation}
\nu_a^{(i)} \equiv \left( \frac{4}{3} \gamma_p^2 \right)^i \tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a
\end{equation}
\begin{equation}
\nu_p^{(i)} \equiv \left( \frac{4}{3} \gamma_p^2 \right)^i \max (\nu_p,\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a)
\end{equation}
\begin{equation}
\nu_b^{(i)} \equiv \left( \frac{4}{3} \gamma_b^2 \right)^i \max (\nu_b,\tilde{\nu}} \def\tGc{\tilde{\Gamma}_c} \def\ttau{\tilde{\tau}_a)
\end{equation}
To summarize the accounting of photons for $\tpm > 1$: the outermost layer of one upscattering
optical depth ($i=1$) yields the synchrotron emission and the first inverse-Compton scattering,
layers of geometrical depth $(\sqrt{i-1} \div \sqrt{i})(\Delta/\ttau_h)$ produce all the seed
photons for the $i$-th inverse-Compton emission, but we ignore the inverse-Compton emission of
order higher than $\tau_h$, produced by pairs at geometrical depth larger than
$\sqrt{\tau_h} \Delta/\ttau_h = \Delta/\sqrt{\tpm}$ because that emission is trapped in the
pair medium for longer than one dynamical timescale, on which change the number of pairs and their
distribution with energy.
\vspace*{2mm}
\section{Optical and X-ray light-curves}
\label{results}
\begin{figure*}
\centerline{\psfig{figure=xos2.eps,height=70mm,width=170mm}}
\figcaption{ As for Figure 3 but for a source deceleration $\Gamma = \Gamma_\o (t/t_\o)^{-1/4}$
corresponding to a {\sl wind-like} medium of proton density $n(R) = 3.10^{35}\, A_*/R^2$,
and a magnetic field parameter $A_* b = 10^{-2}$.
Light-curve breaks originate from a spectral break crossing the optical ("a=o" for absorption,
"i=o" for injection), from the occurrence of the ($\gamma_a = \gamma_c$, $Y < 1$) regime
("a=c"), or from an inverse-Compton to synchrotron transition ("iC$\rightarrow$sy",
in the X-ray). }
\end{figure*}
The formalism presented so far allows the calculation of the optical and X-ray synchrotron
self-Compton flux from pairs at the observer-time $t$ when LAT measured a fluence $\Phi (t)$,
for an assumed source Lorentz factor $\Gamma$, and a shock magnetic field $B$ (parametrized
and tied to $\Gamma$ by the $nb$ parameter).
The "pseudo light-curves" of Figure 2 show the synchrotron and inverse-Compton flux at 2 eV
(optical) and 100 keV (hard X-ray) as function of the source Lorentz factor $\Gamma$ at $t = 10$ s
(during the prompt emission phase), for a LAT fluence $\Phi = 10^{-5} \ergcm2$, and a
few values for the magnetic field parameter $nb$.
The brightest optical flux is obtained when the peak frequency $\nu_p$ of the synchrotron
spectrum is close to the optical range. We note that the peak flux of equation (\ref{Fp})
means magnitude $V=6.6$, comparable to that shown in Figure 2 for $nb =1$ and with the
brightest optical counterpart (emission during the prompt/burst phase) ever observed
(GRB 080319B, reached peak magnitude $V = 5.3$ at 20 s after trigger -- Greco et al 2009).
The 10 keV--1 MeV fluence of Fermi-GBM bursts is about ten times larger than the 100 MeV--10 GeV
fluence of the corresponding LAT afterglows, thus the average MeV flux of a 10 s burst is
$F_{grb} \sim 10^{-5} {\;\rm erg\, cm^{-2}\, s^{-1}}$.
Figure 2 shows that, during the prompt phase, the pair emission
at 100 keV is dimmer by about two orders of magnitude than the burst.
Figure 3 shows proper pair-emission optical and soft X-ray light-curves for a source deceleration
corresponding to a blast-wave interacting with a homogeneous ambient medium (uniform $n$) --
$\Gamma \propto t^{-3/8}$ -- a slowly-decreasing high-energy fluence $\Phi$, a constant magnetic field
parameter $nb$, and starting from various initial Lorentz factors $\Gamma_\o$ at $t_\o = 10$ s.
The shock deceleration corresponds to moving from right to left, on a horizontal line in Figure 1.
The pair flux is calculated at each epoch in same way as for Figure 2, adjusting the fluence $\Phi$.
For Figure 4, the ambient medium has a wind-like structure, $n \propto R^{-2}$, as for a massive star
GRB progenitor, hence the shock deceleration is $\Gamma \propto t^{-1/4}$ and the magnetic field
parameter evolves as $nb \propto t^{-1}$, the deceleration corresponding
to moving from right to left on a $nb \propto \Gamma^4$ line in Figure 1.
Because the pre-deceleration phase of the forward-shock and the rise of the high-energy flux $\Phi$
before $t_\o$ are not accounted for, the calculated pair synchrotron light-curves miss the initial
rising part, but they can still display peaks when the self-absorption $\nu_a$ break or the $F_\nu$
peak $\nu_p = \min(\nu_i,\nu_c)$ fall below the observing band. Otherwise, crossing by the
$\nu F_\nu$ peak frequency $\nu_b = \max(\nu_i,\nu_c)$ yields a light-curve break.
For a typical LAT light-curve, $\Phi(t) = Ft \propto t^{-1/3}$, the pair flux has a power-law decay
$F_\nu \propto t^{-\alpha}$ with $\alpha \in (1.0,1.4)$ for a homogeneous medium and $\alpha \in
(1.2,2.0)$ for a wind, the latter range being more compatible with the observations of GRB optical
counterparts (flashes): $\alpha =1.8$ for GRB 990123 (Akerlof et al 1999),
$\alpha =2.0$ for GRB 061126 (Perley et al 2008),
$\alpha =2.5$ for GRB 080319B (Wozniak et al 2009),
$\alpha =1.7$ for GRB 130427A (Vestrand et al 2014).
Thus, a wind-like medium is favored when interpreting the optical flashes as emission from pairs;
still, we note that steeper decays of the pair emission result for a fluence $\Phi$ decreasing faster
than $t^{-1/3}$.
While Figures 3 and 4 show that the X-ray flux from pairs at 0.1--10 ks can be compatible with
that measured for Swift X-ray afterglows -- $F_\nu = 10^{-11}-10^{-10} {\;\rm erg\, cm^{-2}\, s^{-1}}$
(O'Brien et al 2006) -- the above pair light-curve decays are too steep compared with X-ray plateau
measurements -- $\alpha \in (0.2,0.8)$ -- thus, only the faster-decaying plateaus can be accommodated
by pair emission, provided that the fluence $\Phi$ is nearly constant, and that the ambient medium is
homogeneous.
\vspace*{2mm}
\subsection{Brightest optical flash from pairs}
Figure 5 shows the maximal optical flux from pairs at $t = 10$ s (i.e. during the prompt emission phase),
for a range of high-energy fluence, obtained by searching a reasonable range of the $(\Gamma,nb)$
parameter space: $\Gamma \in (10,10^4)$ and $nb \in (10^{-5},10^5)$. As illustrated in Figure 2,
one expects that $\Gamma \in (100,1000)$ for the brightest optical flash, because too low Lorentz factors
lead to optically-thick pairs and less radiation escapes in a dynamical timescale, while too high
Lorentz factors increase the pair-formation threshold energy, leading to the formation of fewer pairs,
and reducing the flux produced by pairs. As suggested by Figure 2, Figure 5 shows that the maximal
optical synchrotron emission from pairs is larger than that from inverse-Compton, at any fluence.
The important result shown in Figure 5 is the existence of an upper limit $R_{\min}$ on the brightness
of the optical flash from pairs, which offers test of this model: any flash brighter than shown in Figure 5,
for the measured LAT fluence, cannot originate from pairs. In more detail, the brightest observed
LAT afterglows ($\Phi \sim 10^{-4} \ergcm2$) could yield optical flashes as bright as $R = 5$, the dimmest
LAT afterglows ($\Phi \sim 10^{-6} \ergcm2$) may produce $R = 9$ optical flashes, with reasonably bright
flashes originating from the pairs produced in GRB afterglows that are not detectable by LAT above 100 MeV.
We note that the existence of $R_{\min}$ is not a immediate consequence of equation (\ref{Fp}),
which leaves the possibility of a brighter optical flash than shown in Figure 5, if the peak frequency
$\nu_p$ of the synchrotron spectrum fell in the optical and if it were not self-absorbed.
Instead, the brightest optical flash from pairs shown in Figure 5 corresponds to $\nu_p$ being
below optical ($\nu_o$); more exactly, $\nu_c < \nu_a = \nu_o < \nu_i$ (and $Y > 1$, $\tau < 1$)
is satisfied everywhere on the $R_{\min} (\Phi)$ line, a condition also fulfilled by the brightest
two "peaks" shown in Figure 2 (left panel).
\begin{figure*}
\centerline{\psfig{figure=rpeak.eps,height=75mm}}
\figcaption{ {\sl Left panel}: distribution (number of events in $\delta R = 1$ bin, relative to maximum)
of pair optical flash magnitude $R$ at $t = 10$ s, for three fixed GeV fluences $\Phi$ (corresponding
to a bright, dim, and undetectable LAT emission) and for a
GeV fluence distributed uniformly in log-space (solid line). The source magnetic field parameter $nb$
and Lorentz factor $\Gamma$ at time $t$ were assumed uniformly distributed in log-space, spanning
the range $10^{-4}-10^4\, {\rm cm^{-3}}$ and $10^2-10^4$, respectively (the shape of the distributions
and its peak location depend on the assumed range and distribution of these two parameters, but the
brightest end magnitude depends only on the fluence).
{\sl Right panel}: brightest pair optical flash flux $R_{\min}$ (i.e. the bright end of the distributions
shown in the left panel) as a function of fluence $\Phi$, with the two parameters $(\Gamma,nb)$ left
free. The $R_{\min}$ is reached for $\nu_c < \nu_a < \nu_i$, $Y > 1$, $\tau < 1$, and when the
self-absorption frequency is in the optical.
For this ordering of synchrotron break frequencies, $R_{\min}$ is independent of the burst redshift
and very weakly dependent on the epoch $t$, thus the upper limit $R_{\min}$ on the optical flash
flux shown here is a robust prediction for the synchrotron emission from pairs. }
\end{figure*}
The surprising aspect of the brightest optical flash shown in Figure 5 is that $R_{\min} (\Phi)$ is
{\sl independent} of the burst redshift $z$. Perhaps a moderate dependence of $R_{\min}$ on redshift
is expected because $R_{\min}$ is calculated for a fixed fluence $\Phi$, thus a higher $z$ implies
a larger afterglow output above 100 MeV, a larger number of formed pairs, and a larger synchrotron
luminosity that compensates for the larger luminosity distance. That $R_{\min} (\Phi)$ is $z$-independent
can be proven in the following way. From equations (\ref{Fp}) -- (\ref{Fnu}), it follows that,
for $\nu_c < \nu_a < \nu_o < \nu_i$, the optical flux satisfies above self-absorption is
\begin{equation}
F_o = F_p \left( \frac{\nu_c}{\nu_o} \right)^{1/2} \propto
\z^{5/2} \frac{\Phi^{3/2} (nb)^{1/4}}{\Gamma^2 t^{3/2}} \propto \Gamma^{-2}
\label{Fo1}
\end{equation}
For $\nu_c < \nu_a < \nu_i$, equations (\ref{nui}) and (\ref{taua})-- (\ref{ga}) lead to
\begin{equation}
\gamma_a = \gamma_c \tau_p^{1/6} \propto \left( \frac{\tau \gamma_c}{B} \right)^{1/6} ,
\quad \nu_a \propto \z^{5/3} \frac{\Phi^{1/2} (nb)^{1/6}}{\Gamma^{4/3} t^{7/6}}
\label{nua}
\end{equation}
From equations (\ref{taua1}) and (\ref{Fp}) -- (\ref{Fnu1}), the optical flux below $\nu_a$ is
\begin{equation}
F_o = F_p \left( \frac{\nu_c}{\nu_a} \right)^{1/2} \left( \frac{\nu_o}{\nu_a} \right)^{5/2}
\propto \z^{-5/2} \frac {\Gamma^2 t^2} {\Phi^0 (nb)^{1/4}}
\label{Fo2}
\end{equation}
Defining $\Gamma_p$ by $\nu_a (\Gamma_p) = \nu_o$, it follows from equation (\ref{nua}) that
$i)$ $\nu_o < \nu_a$ for $\Gamma < \Gamma_p$, hence $F_o \propto \Gamma^2$ (equation \ref{Fo2}),
the optical flux increasing with $\Gamma$, and
$ii)$ $\nu_a < \nu_o$ for $\Gamma > \Gamma_p$, hence $F_o \propto \Gamma^{-2}$ (equation \ref{Fo1}),
the optical flux decreasing with $\Gamma$.
Consequently, the optical flux is maximal for $\Gamma = \Gamma_p$, with $\Gamma_p \propto
\z^{5/4} \Phi^{3/8} (nb)^{1/8} t^{-7/8}$ following from the defining condition $\nu_a (\Gamma_p) =
\nu_o$. Substituting $\Gamma_p$ in either equation (\ref{Fo1}) or (\ref{Fo2}), the maximal
optical flux satisfies
\begin{equation}
F_o^{\max} \equiv F_o (\Gamma_p) \propto \z^0 (nb)^0 \Phi^{3/4} t^{1/4}
\label{Fo}
\end{equation}
being independent of redshift and also of the magnetic field parameter $nb$ (meaning that,
for any $nb$ that allows $\nu_c < \nu_a = \nu_o < \nu_i$, $Y > 1$, and $\tau < 1$, there is
a $\Gamma_p$ that maximizes the optical flux to the same value $F_o (\Phi,t)$).
The coefficient missing from equation (\ref{Fo}) can be determined by carrying the
coefficients of all equations involved in its derivation. From Figure 5, the maximal optical
flux of equation (\ref{Fo}) is
\begin{equation}
R_{\min} = 8.7 - 2.5 \log \left( \frac{F_o^{\max}}{\rm Jy} \right) =
7.0 - \frac{15}{8} \log \Phi_{-5} - \frac{5}{8} \log t_1
\label{Rmin}
\end{equation}
For GRB 130427A, the optical flash peaked at $R \simeq 7.4$ at 10--20 s after trigger (Wren et al 2013),
when the LAT 0.1--100 GeV fluence was $\Phi \simeq 4.10^{-5} \ergcm2$ (Tam et al 2013). The upper limit
given in equation (\ref{Rmin}), $R_{\min} \simeq 5.8$, is brighter than measured, thus a pair origin
for the optical flash of GRB 130427A is not ruled out.
\vspace*{2mm}
\subsection{Caveats}
Our assumption that the power-law spectrum of the high-energy photons (measured by LAT at
100 MeV--10 GeV) extends well outside that range could lead to an {\sl overestimation} of
the emission from pairs in two ways.
Equations (\ref{nui}) and (\ref{B}) show that the synchrotron emission at photon energy $h\nu$
is produced by pairs of shock-frame energy $\gamma (\nu) = 155\, (\z h\nu/1 eV)^{1/2} (nb)^{-1/4}
\Gamma_2^{-1}$. Pairs of this energy are formed from photons of observer-frame energy $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb (\nu)
\simeq 2 \gamma \Gamma m_e c^2/(z+1) = 5.3\, (h\nu/1 eV)^{1/2} \z^{-1/2} (nb)^{-1/4}$ GeV. Thus,
optical synchrotron emission requires pairs formed from seed photons of $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \mathrel{\rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}} 10$ GeV
(which have been occasionally detected by LAT), 1 keV synchrotron emission requires photons of
$\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \mathrel{\rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}} 200$ GeV (the highest-energy photon detected by LAT had $\sim 100$ GeV, for the
$z=0.3$ GRB 130427A -- Fan et al 2013, Tam et al 2013), while 100 keV synchrotron emission
requires photons of energy $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb \mathrel{\rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}} 3$ TeV.
At the other end, if the high-energy spectrum
has a break not far below 100 MeV, the assumption of a single power-law overestimates the
number of target photons for the $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb (\nu)$ photon, which leads to an overestimation of the
optical thickness to pair-formation and of the number of $\gamma (\nu)$ pairs that are formed
(if the true optical thickness $\tau_{\gamma\gamma} [\varepsilon} \def\eps{\epsilon} \def\nnbb{nb(\nu)]$ is below unity).
From equation (\ref{ethr}), the pair-formation threshold-energy for a $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb (\nu)$ photon is
$\varepsilon} \def\eps{\epsilon} \def\nnbb{nb_t (\nu) = 0.22 (h\nu/1 eV)^{-1/2} \z^{-3/2} (nb)^{1/4} \Gamma_2^2$ MeV, which, for both
optical and X-ray photons, is well below the lower edge of the LAT window.
Not accounting for the decollimation introduced by the scattering of the seed photons on pairs
leads to an {\sl underestimation} of true number of pairs. An estimate of the importance of that pair-cascade
can be obtained by first noting that most pairs are formed by the more numerous, lower energy photons
above the threshold for pair-formation, i.e. by the $\sim 1$ MeV photons of the LAT spectrum extrapolation.
For a flat $\nu F_\nu$ LAT spectrum, the fluence of the 1 MeV photons is comparable to the LAT fluence,
hence the 1 MeV photons should have an energy output of $10^{53 \pm 1}$ erg. For a $10^{53}$ erg burst
lasting for 10 s, the outflow pair-loading through a cascade process was shown (Kumar \& Panaitescu 2004)
to produce a significant pair-enrichment up to a radius $R_{max} = 5\times 10^{15}$ cm. For an
observer-frame time $t$, this radius corresponds to source Lorentz factor $\Gamma_{max} = [(z+1)
R_{max}/ct]^{1/2} \simeq 220\, (\z/t_1)^{1/2}$. Thus, ignoring the pair-cascade process implies an
underestimation of the pair number for a source with $\Gamma < \Gamma_{max} = 1.5 (\z^{0.7} \phi5^{0.2}
t_1^{0.1})^{-1} \Gamma_\tau$, i.e. the pair-cascade is effective mostly when the pair-medium produced
by the unscattered seed photons is optically-thick ($\tau > 1$). As illustrated in Figure 2, a larger
number of pairs and the associated larger optical thickness $\tau$ lead to a dimmer emergent emission,
hence the pair-cascade would reduce even more the pair emission.
\vspace*{2mm}
\section{Conclusions}
The above investigation of the broadband emission from pairs shows that the synchrotron emission
from pairs formed from $>100$ MeV afterglow photons can accommodate the brightest optical counterparts
(flashes) that were observed (in a few cases) during the prompt (GRB) phase, with the fast decay of
optical flash pointing to a wind-like circumburst medium or to a faster decaying fluence of the high-energy
photons. The inverse-Compton emission from pairs may yield an afterglow brightening, as seldom seen
in optical afterglow light-curves.
A brighter Fermi-LAT afterglow implies more pairs that can form and, thus, a brighter synchrotron
and inverse-Compton emission from those pairs. The light-curve scalings given in Table 1 quantify
the positive correlation between the synchrotron pair flux and the high-energy photon fluence $\Phi$.
In addition to the 100 MeV fluence, the pair emission depends also on the source Lorentz factor
$\Gamma$ and on the magnetic field in the pair medium. A very high Lorentz factor ($\Gamma \mathrel{\rlap{\raise 0.511ex \hbox{$>$}}{\lower 0.511ex \hbox{$\sim$}}} 1000$)
raises the threshold energy for pair-formation, reduces the number of pairs and, implicitly, their
emission. A very low Lorentz factor ($\Gamma \mathrel{\rlap{\raise 0.511ex \hbox{$<$}}{\lower 0.511ex \hbox{$\sim$}}} 50$) leads to a larger number of pairs, an
optically-thick pair medium, which traps the emission and, consequently, yields a dim pair flux.
Leaving free the source Lorentz factor $\Gamma$ and the magnetic field parameter $nb$, we find
that the brightest optical flash produced by pairs satisfies equation (\ref{Rmin}), which gives
an upper limit to that optical flash that depends mostly on the high-energy fluence, is weakly
dependent on the epoch of observation, and, most important, is independent of the source redshift.
We note that the brightest optical flash from pairs for a given LAT fluence is obtained when the
self-absorption frequency is in the optical, thus the intrinsic synchrotron optical spectrum of
the brightest optical flash from pairs should be flat.
Given that a powerful source of high-energy photons is needed to produce enough pairs that can
account for optical flashes/counterparts and that the source of high energy photons is, most likely,
the forward-shock (Kumar \& Barniol Duran 2009), the multiwavelength data of GRB afterglows at early
times should be interpreted/modeled as the sum of synchrotron and inverse-Compton emissions from
the forward-shock, the reverse-shock, and from pairs, at least in those cases where LAT measures a
a bright high-energy afterglow. To calculate the pair emission requires three parameters: the magnetic
field parameter $nb$ (which is also constrained by fitting the multiwavelength afterglow data with
the reverse/forward-shock emission), the LAT afterglow fluence $\Phi$ (which is the blast-wave emission),
and the source Lorentz factor $\Gamma$ (which is constrained directly by the double-shock emission
fits at the deceleration time, and indirectly, through the ratio of blast-wave energy to external
density $E/n$, after deceleration onset). Thus, the pair emission dose not entail any new parameters.
We note that lower limits on $\Gamma$ during the burst phase (pre-deceleration) can be obtained
from the photon-photon opacity of the prompt LAT emission (e.g. Abdo et al 2009).
The high LAT fluence of GRB 130427A (Ackermann et al 2014), its bright
optical counterpart (Vestrand et al 2014), and broadband coverage would make it a good candidate
for such a study. The low circumburst medium density, inferred by modeling its 10 s--10 d radio,
optical, X-ray, and 100 MeV measurements, with synchrotron emission from the reverse and forward
shocks (Laskar et al 2013, Panaitescu et al 2013, Perley et al 2014), implies a very high afterglow
Lorentz factor ($\Gamma_\o \sim 800$ at 10 s) and, consequently, a low number of formed pairs,
leading to an optical emission from pairs that is well below the bright optical flash of GRB 130427A.
The pair emission discussed here offers an alternate explanation (to the reverse-shock) for GRB optical
counterparts. Both "mechanisms" can yield bright optical flashes with a fast decay after the peak.
The pair optical flux should be correlated with the GeV contemporaneous fluence, but that feature
may exist also for reverse-shock flashes, if most of the GeV flux arises from the reverse-shock.
The post-peak decay of the optical pair flux, given in Table 1 for a $\varepsilon} \def\eps{\epsilon} \def\nnbb{nb^{-2}$ LAT spectrum,
can be generalized and used to test a pair origin for the optical flash, although a continuous
reverse-shock (i.e. not one that ceases at the optical peak time, followed by adiabatic cooling
of ejecta electrons) may yield a similar decay index -- spectral slope closure relation. Perhaps,
modeling of the broadband afterglow data will be the best way to disentangle the pair emission
and reverse-shock emissions.
\acknowledgments{This work was supported by an award from the Laboratory Directed Research and
Development program at the Los Alamos National Laboratory}
|
\section{Introduction and notation}
Let $G=(V,E)$ be a graph and consider the following chip firing game on $G$. At any stage, we have a chip configuration consisting of a nonnegative number of chips at each vertex of $G$. We may go from one chip configuration to another by a sequence of moves. A move consists of choosing a subset $U\subseteq V$ and moving one chip from $u$ to $v$ for every edge $uv$ with $u\in U$ and $v\in V\setminus U$. For the move to be possible, every vertex $u\in U$ must have at least as many chips as it has edges to vertices in $V\setminus U$. Observe that a move corresponding to a subset $U$ can be reversed by subsequently making the move corresponding to the complementary set $V\setminus U$.
A chip configuration is winning if for every vertex $v$ there is a sequence of moves that results in a configuration with at least one chip on $v$. The gonality $\dgon(G)$ of $G$ is the smallest number of chips in a winning chip configuration.
The main result of this paper is to show that the gonality of a connected graph is lower bounded by its treewidth. This result was conjectured in \cite{vanDobbendeBruyn2012}.
The remainder of Section 1 is devoted to preliminaries, including basic notation and terminology related to graphs, divisors and treewidth. In Section 2, we state and prove the main theorem. In Section 3, we consider some families of graphs for which treewidth equals gonality. These include: trees, grids and complete multipartite graphs. In Section 4, we briefly review divisor theory for \emph{metric} graphs. We show that the gonality of a metric graph is lower bounded by the gonality of a subdivision of the underlying graph. Hence, the tweewidth is also a lower bound for metric graphs. In Section 5, we discuss some related notions of gonality defined in terms of harmonic morphisms, and show that there the treewidth is also a lower bound.
\subsection{Graphs}
The graphs in this paper will be finite and undirected (unless stated otherwise). We allow our graphs to have multiple (parallel) edges, but no loops. We will almost exclusively consider \emph{connected} graphs. For a graph $G$, we denote by $V(G)$ and $E(G)$ the set of vertices and edges of $G$, respectively. By an edge $uv$, we mean an edge with ends $u$ and $v$. For (not necessarily disjoint) subsets $U,W\subseteq V(G)$, we denote by $E(U,W)$ the set of edges with an end in $U$ and an end in $W$. For vertices $u$ and $v$, we use the abbreviations $E(u,v):=E(\{u\},\{v\})$ and $E(u):=E(\{u\},V\setminus \{u\})$. The degree of a vertex $v$ equals the number of edges with $v$ as an endpoint and is denoted by $d_G(v):=|E(v)|$. For a subset $U\subseteq V(G)$, we denote by $G[U]:=(U,E(U,U))$ the subgraph of $G$ \emph{induced by $U$}. That is, $G[U]$ is the graph with vertex set $U$ and edge set consisting of the edges of $G$ with both ends in $U$.
Let $G=(V,E)$ be a \emph{connected} graph. We can make $G$ into an \emph{oriented graph} by, for every edge $e$, assigning one end to be the \emph{head} of $e$ and the other end to be the \emph{tail} of $e$. We view the edge $e$ as oriented from its tail to its head. For a cycle $C$ in $G$, we then denote by $\chi_C\in \ensuremath{\mathbb{R}}^{E}$ the signed incidence vector defined by
\begin{equation}
\chi_C(e)=\begin{cases}
1& \text{if $e$ is traversed in forward direction by $C$,}\\
-1&\text{if $e$ is traversed in backward direction by $C$,}\\
0&\text{otherwise.}
\end{cases}
\end{equation}
Similarly, we write $\chi_P$ for the signed incidence vector of a path $P$.
The \emph{incidence matrix} $M=M(G)\in \ensuremath{\mathbb{R}}^{V\times E}$ of $G$ is defined by, for every $v\in V$ and $e\in E$, setting
\begin{equation}
M_{v,e}:=
\begin{cases}
1&\text{if $v$ is the head of $e$,}\\
-1&\text{if $v$ is the tail of $e$,}\\
0&\text{otherwise.}
\end{cases}
\end{equation}
\noindent The matrix $Q=Q(G):=MM^\mathsf{T}$ is the \emph{Laplacian} of $G$ and it is independent of the chosen orientation. Indeed, for any two vertices $u$ and $v$, $Q_{vv}$ equals the degree of vertex $v$ and $Q_{uv}$ equals $-|E(u,v)|$. The \emph{cut lattice} of $G$ is the set $\mathcal{C}(G):=\ensuremath{\mathbb{Z}}^{E}\cap \Col(M^\mathsf{T})$ of integral vectors in the column space of the transpose of $M$.
The following two lemma's are well-known, see for example \cite{GodsilRoyle2001}. For the sake of the reader, we will give the short proofs.
\begin{lemma}\label{clattice}
Let $f\in \ensuremath{\mathbb{Z}}^{E}$. Then the following are equivalent:
\begin{itemize}
\item[i)] $f$ is in the cut lattice of $G$,
\item[ii)] $f^\mathsf{T}\chi_C=0$ for every cycle $C$ in $G$,
\item[iii)] $f=M^\mathsf{T} x$ for some $x\in \ensuremath{\mathbb{Z}}^{V}$.
\end{itemize}
\end{lemma}
\begin{proof}
The implication from iii) to i) is trivial. The implication from i) to ii) follows since $M\chi_C=0$ for every cycle $C$. For the implication from ii) to iii), let $f\in \ensuremath{\mathbb{Z}}^{E}$ satisfy the condition in ii). Let $T$ be a spanning tree in $G$ with root $r$, and define $x\in \ensuremath{\mathbb{Z}}^V$ by $x(v):=f^\mathsf{T}\chi_{P_v}$, where $P_v$ is the path in $T$ from $r$ to $v$. Now for every edge $e=uv$ oriented from $u$ to $v$, we have $x(v)-x(u)=f(e)$. Hence, $f=M^\mathsf{T} x$.
\end{proof}
\begin{lemma}\label{Laplacian}
The null space of $Q$ is spanned by the all-one vector $\ensuremath{\mathbf{1}}$.
\end{lemma}
\begin{proof}
Since the row sums of $Q$ equal zero, it is clear that $Q\ensuremath{\mathbf{1}}=0$. Conversely, let $x$ be in the null space of $Q$ and suppose, for contradiction, that $x$ is not a multiple of $\ensuremath{\mathbf{1}}$. Since $G$ is connected, we may choose $v\in V$ for which $x(v)$ is maximal and such that $v$ has a neighbour $u$ with $x(u) < x(v)$. From $Qx=0$ it follows that $d_G(v)x(v)=\sum_{w\in V\setminus\{v\}} |E(v,w)|\cdot x(w)$. On the other hand,
\begin{equation*}
\sum_{w\in V\setminus\{v\}}|E(v,w)|\cdot x(w)\ < \sum_{w\in V\setminus\{v\}}|E(v,w)|\cdot x(v)= d_G(v)x(v)
\end{equation*}
by our choice of $v$. This is a contradiction.
\end{proof}
\subsection{Divisors}
We will largely adopt notation from \cite{BakerNorine2007}. Let $G=(V,E)$ be a connected graph. A vector $D\in \ensuremath{\mathbb{Z}}^V$ is called a \emph{divisor} on $G$. The set $\Div(G):=\ensuremath{\mathbb{Z}}^V$ denotes the set of all divisors on $G$. For a divisor $D\in \Div(G)$ we call $\deg(D):=\sum_{v\in V}D(v)$ the \emph{degree} of $D$. A divisor $D$ is said to be \emph{effective} if it is nonnegative. We denote by $\Div_+(G)$ the set of effective divisors on $G$ and by $\Div_+^k(G)$ the set of effective divisors of degree $k$. We denote by $\supp(D):=\{v\in V\mid D(v)\neq 0\}$ the \emph{support} of $D$.
We call two divisors $D$ and $D'$ \emph{equivalent} and write $D\sim D'$ if there is an integer vector $x\in \ensuremath{\mathbb{Z}}^V$ such that $D-D'=Q(G)x$. Clearly, this is indeed an equivalence relation. Observe that equivalent divisors have equal rank as $Q(G)$ has column sums equal to zero.
We will often consider the situation where $x$ is the incidence vector of a subset $U$ of $V$, that is $D'=D-Q(G)\ensuremath{\mathbf{1}}_U$. Observe that is this case
\begin{equation}
D'(v)=\begin{cases}D(v)-|E(\{v\},V\setminus U)|&\text{if $v\in U$},\\D(v)+|E(\{v\},U)|&\text{if $v\in V\setminus U$.}\end{cases}
\end{equation}
In particular, $D'(v)\leq D(v)$ if $v\in U$ and $D'(v)\geq D(v)$ if $v\in V\setminus U$. In terms of chip firing, we move one chip along each edge in the cut $E(U,V\setminus U)$. The following lemma shows that for equivalent effective divisors $D$ and $D'$, we can obtain $D'$ from $D$ by a sequence of steps of this form and each intermediate divisor being effective.
\begin{lemma}\label{chain}
Let $D_0$ and $D$ be equivalent effective divisors satisfying $D \neq D_0$. There is a chain of sets $\emptyset\subsetneq U_1\subseteq U_2\subseteq \cdots\subseteq U_k\subsetneq V$ such that $D_t:=D-Q(G)(\sum_{i=1}^t \ensuremath{\mathbf{1}}_{U_i})$ is effective for every $t=1,\ldots,k$ and $D_k=D$. Moreover, this chain is unique.
\end{lemma}
\begin{proof}
Since $D_0\sim D$, there exists an $x\in \ensuremath{\mathbb{Z}}^V$ such that $D_0-Q(G)x=D$. By Lemma \ref{Laplacian}, $x$ is unique up to integral multiples of $\ensuremath{\mathbf{1}}$. Hence, there is a unique such $x$ with the additional property that $x\geq 0$ and $\supp(x)\neq V$. Let $k:=\max\{x(v)\mid v\in V\}$ and define $U_i:=\{v\in V\mid x(v)\geq k-i+1\}$ for $i=1,\ldots, k$. It follows that $\sum_{i=1}^k \ensuremath{\mathbf{1}}_{U_i}=x$.
Now consider any $v\in V$ and any $t\in \{1,\ldots, k\}$. If $v\not\in U_t$, then $D_0(v)\leq D_1(v)\leq \cdots\leq D_t(v)$, hence $D_t(v)\geq 0$. If $v\in U_t$, then $D_t(v)\geq D_{t+1}(v)\geq \cdots \geq D_k(v)$, hence $D_t(v)\geq 0$. It follows that $D_1,\ldots,D_{k-1}$ are effective.
Uniqueness follows directly from the uniqueness of an $x\in \ensuremath{\mathbb{Z}}^V$ for which $x\geq 0$, $\supp(x)\neq V$ and $D_0-Q(G)x=D$, in combination with the uniqueness of the decomposition $x=\sum_{i=1}^k\ensuremath{\mathbf{1}}_{U_i}$ as a sum of characteristic vectors of a chain $\emptyset\subsetneq U_1\subseteq U_2\subseteq \cdots\subseteq U_k\subsetneq V$.
\end{proof}
The \emph{rank} of a divisor $D$ is defined as
\begin{equation}
\rank(D):=\max\{k\mid \text{$D-D'$ is equivalent to an effective divisor for every $D'\in\Div_+^k$}\}.
\end{equation}
\noindent Observe that equivalent divisors have equal rank and that $\rank(D)\leq \deg(D)$. Also observe that the restriction of $D'$ to effective divisors in the definition is immaterial.
Following Baker \cite{Baker2008}, we define the \emph{gonality} of $G$ by
\begin{equation}
\dgon(G):=\min\{k\mid \text{there is a divisor of degree $k$ on $G$ with positive rank}\}.
\end{equation}
An effective divisor $D$ is called \emph{$v$-reduced} if for any nonempty subset $U\subseteq V\setminus \{v\}$ the divisor $D-Q(G)\ensuremath{\mathbf{1}}_U$ is not effective. In other words, for every nonempty $U\subseteq V\setminus \{v\}$ there is a $u\in U$ with $D(u)<|E(\{u\}, V\setminus U)|$.
\begin{lemma}\label{reduction}
Let $v\in V$ and let $D$ be an effective divisor on $G$. Then there is a unique $v$-reduced divisor equivalent to $D$.
\end{lemma}
\begin{proof}
For any divisor $D' \sim D$, there is a unique $x_{D'}\in \{x\in \ensuremath{\mathbb{Z}}^V\mid x\geq 0,\ x(v)=0\}$ such that $D'=D-Q(G)x_{D'}$ by Lemma \ref{Laplacian}. Let
\begin{equation}
S:=\{x_{D'}\mid D'\text{ is effective and equivalent to $D$}\}.
\end{equation}
The set $S$ is finite since the number of effective divisors equivalent to $D$ is finite. Choose $x_{D'}\in S$ maximizing $\sum_{u\in V} x_{D'}(u)$. Then $D'$ is $v$-reduced because for any nonempty $U\subseteq V\setminus \{v\}$, the vector $x_{D'}+\ensuremath{\mathbf{1}}_U$ is not in $S$ by the choice of $x_{D'}$.
To show uniqueness, let $D$ and $D'$ be two different, but equivalent effective divisors. It suffices to show that $D$ and $D'$ are not both $v$-reduced. By Lemma \ref{chain} there are sets $\emptyset\subsetneq U_1\subseteq\cdots\subseteq U_k\subsetneq V$ such that $D-Q(G)\ensuremath{\mathbf{1}}_{U_1}$ and $D'+Q(G)\ensuremath{\mathbf{1}}_{U_k}=D'-Q(G)\ensuremath{\mathbf{1}}_{V\setminus U_k}$ are effective. If $v\not\in U_1$, then $D$ is not $v$-reduced. If $v\in U_1$, then $v\not\in V\setminus U_k\subseteq V\setminus U_1$ and hence $D'$ is not $v$-reduced.
\end{proof}
\noindent Observe that if $D$ is $v$-reduced and $\rank(D)\geq 1$, then we have $D(v)\geq 1$.
We say that a divisor $D$ \emph{covers} $v\in V$ if there is an effective divisor $D'$ equivalent to $D$ with $v\in \supp(D')$. A nonempty subset $S\subseteq V$ is called a \emph{strong separator} if for each component $C$ of $G[V\setminus S]$ we have that $C$ is a tree and $|E(\{s\},V(C))|\leq 1$ for every $s\in S$. The folowing lemma is similar to a theorem of Luo \cite{Luo2011} on rank determining sets in the context of metric graphs.
\begin{lemma}\label{Luotype}
Let $S$ be a strong separator of $G$ and let $D$ be a divisor covering every $s\in S$. Then $D$ has positive rank.
\end{lemma}
\begin{proof}
Since any superset of a strong separator is again a strong separator, we may assume that $S=\{s\in V\mid \text{ $s$ is covered by $D$}\}$. We have to show that $S=V$.
Suppose not. Let $C$ be a component of $G[V\setminus S]$ and let $S':=\{s\in S: |E(\{s\},V(C))|=1\}$. Since $G$ is connected, $S'$ is not empty, so we may take $s\in S'$ and assume that $D$ is $s$-reduced. If $S'\subseteq \supp(D)$, then $D+Q(G)\ensuremath{\mathbf{1}}_{V(C)}$ is effective and has support on at least one vertex in $V(C)\subseteq V\setminus S$, a contradiction.
Hence, we may assume that there is a $t\in S'\setminus \supp(D)$. In particular, $D$ is not $t$-reduced. Let $a$ and $b$ be the unique neighbours of $s$ and $t$ in $V(C)$, respectively, and let $P=(s,a,\ldots,b,t)$ be the path from $s$ to $t$ with its interior points in $V(C)$. Since $D$ is $s$-reduced, but not $t$-reduced, there is a set $U\subseteq V$ with $s\in U$, $t\not\in U$ such that $D':=D-Q(G)\ensuremath{\mathbf{1}}_U$ is effective. The cut $E(U,V\setminus U)$ must intersect some edge $e=uv$ of the path $P$, and we find that $D(u)\geq 1$ and $D'(v)\geq 1$. Since at least one of $u$ and $v$ is in $V(C)\subseteq V\setminus S$, we obtain a contradiction.
\end{proof}
\begin{corollary}\label{rankdetermining}
If $H$ is a subdivision of $G$ and $D$ is a divisor on $H$ that covers all $v\in V(G)$, then $D$ has positive rank.
\end{corollary}
\subsection{Treewidth}
The notion of \emph{treewidth} was first introduced by Halin \cite{Halin1976} and later rediscovered by Robertson and Seymour \cite{RobertsonSeymour1990} as part of their graph minor theory. There are several equivalent definitions of treewidth. The most natural one is perhaps in terms of tree-decompositions of a graph. However, for reasons of brevity and since we will not need tree-decompositions here, we use the following definition in terms of chordal extensions.
A graph $H$ is called \emph{chordal} if it has no induced cycle of length at least 4. If $G=(V,E)$ is a subgraph of a chordal graph $H=(V,F)$, then $H$ is called a \emph{chordal extension} of $G$. We denote the maximum size of a clique in a graph $H$ by $\omega(H)$. The treewidth $\tw(G)$ of a graph $G$ can now be defined by
\begin{equation}
\tw(G):=-1+\min\{\omega(H)\mid \text{ $H$ is a chordal extension of $G$}\}.
\end{equation}
\noindent Observe that the treewidth of a (nontrivial) tree equals 1 and the treewidth of a complete graph on $n$ nodes equals $n-1$ as these graphs are chordal and have clique number $2$ and $n$, respectively.
It is NP-complete to determine for a given graph $G$ and a given integer $k$ whether $\tw(G)\leq k$ (see \cite{ArnborgEtal1987}). The fact that this problem is in NP follows directly from the definition by using a suitable chordal extension $H$ as a certificate. Indeed, a perfect elimination order for $H$ certifies chordality of $H$ and provides $\omega(H)$.
In order to use treewidth as a lower bound, we will need a way to lower bound treewidth. For this, we will utilize the notion of bramble. Let $G=(V,E)$ be a graph, and let $2^V$ denote the power set of $V$. A set $\mathcal{B}\subseteq 2^V\setminus\{\emptyset\}$ is called a \emph{bramble} if for any $B,B'\in \mathcal{B}$ the induced subgraph $G[B\cup B']$ is connected. In particular, $G[B]$ is connected for every $B\in \mathcal{B}$. For any $B,B'\in \mathcal{B}$, either $B\cap B'\neq \emptyset$, or $B\cap B'=\emptyset$ and there is an edge in $E(B,B')$. In the latter case, we say that $B$ and $B'$ \emph{touch}.
A set $S\subseteq V$ is called a \emph{hitting set} for $\mathcal{B}$ if it has nonempty intersection with every member of $\mathcal{B}$. The \emph{order} of $\mathcal{B}$, denoted $\order{\mathcal{B}}$, is the minimum size of a hitting set for $\mathcal{B}$. That is:
\begin{equation}
\order{\mathcal{B}}:=\min \{|S| : S\subseteq V, S\cap B\neq\emptyset \text{ for all $B\in \mathcal{B}$}\}.
\end{equation}
We will use the following characterization of treewidth due to Seymour and Thomas~\cite{SeymourThomas1993}.
\begin{theorem}[treewidth duality]
Let $k\geq 0$ be an integer. A graph $G$ has treewidth at least $k$ if and only if it has a bramble of order at least $k + 1$.
\end{theorem}
\begin{remark}
Observe that the treewidth of a graph is equal to the treewidth of the underlying simple graph. It is well-known that treewidth is monotone under taking minors (see for example \cite{DiestelGraphTheory}). That is, removing edges or contracting edges can only decrease treewidth. This also follows easily from the definition.
In particular, if $H$ is a subdivision of $G$, then $\tw(G)\leq \tw(H)$. It is not hard to see that if $G$ has treewidth at least 2, then in fact $\tw(G)=\tw(H)$ holds. Indeed, it suffices to consider the case that $H$ is obtained from $G$ by subdividing an edge $uv$. Let $G'$ be a chordal extension of $G$ with $\omega(G')=\tw(G)+1$. By adding to $G'$ a new node $w$ and edges $uw$ and $vw$, we obtain a chordal extension $H'$ of $H$. Clearly, $\omega(H')=\max(3,\omega(G'))$. Hence,
\begin{equation}
\tw(H)\leq \omega(H')-1=\max(2,\omega(G')-1)=\max(2,\tw(G))=\tw(G).
\end{equation}
If $\tw(G)=1$ and $G$ has two parallel edges, then subdividing such an edge yields a graph of treewidth 2.
\end{remark}
We refer the interested reader to Chapter 12 in \cite{DiestelGraphTheory} for an excellent exposition of treewidth and its role in the graph minor theory.
\section{Proof of the main theorem}
In this section we prove our main theorem.
\begin{theorem}\label{main}
Let $G=(V,E)$ be a connected graph. Then $\dgon(G)\geq \tw(G)$.
\end{theorem}
We start by stating and proving two lemmas.
\begin{lemma}\label{oneside}
Let $D,D'$ be effective divisors such that $D'=D-Q(G)\ensuremath{\mathbf{1}}_U$ for some subset $U\subseteq V$. Let $B\subseteq V$ be such that $G[B]$ is connected. Suppose that $B\cap \supp(D)$ is nonempty, but $B\cap \supp(D')$ is empty. Then $B\subseteq U$.
\end{lemma}
\begin{proof}
Clearly, $B$ cannot be a subset of $V\setminus U$, because otherwise $D'(v)\geq D(v)$ for every $v\in B$. Now suppose that $B\cap U$ and $B\setminus U$ are both nonempty. Since $G[B]$ is connected, there is an edge $uv$ with $u\in B\cap U$ and $v\in B\setminus U$. But then $D'(v)=(D-Q(G)\ensuremath{\mathbf{1}}_U)(v)\geq D(v)+1\geq 1$ since $u\in U$ is a neighbour of $v\in V\setminus U$. This is a contradiction as well, so we see that $B\subseteq U$ must hold.
\end{proof}
\begin{lemma}\label{brambleseparation}
Let $\mathcal{B}$ be a bramble in $G$ and let $U\subseteq V$. Suppose that there exist $B,B'\in \mathcal{B}$ such that $B\subseteq V\setminus U$ and $B'\subseteq U$. Then $|E(U,V\setminus U)|+1\geq \order{\mathcal{B}}$.
\end{lemma}
\begin{proof}
We will construct a hitting set for $\mathcal{B}$ of size at most $|E(U,V\setminus U)|+1$.
Let $F:=E(U,V\setminus U)$ be the cut determined by $U$ and let $H:=(V,F)$. Let $$X:=\{v\in U\mid d_H(v)\geq 1\}\qquad\text{and}\qquad Y:=\{v\in V\setminus U\mid d_H(v)\geq 1\}$$
be the `shores' of the cut $F$. Let $\mathcal{B}':=\{B'\in \mathcal{B}\mid B'\subseteq U\}$. By assumption, $\mathcal{B}'$ is nonempty. Choose $B'\in \mathcal{B}'$ for which $B'\cap X$ is inclusionwise minimal. Let $B\in \mathcal{B}$ be such that $B\subseteq V\setminus U$. Observe that $B'\cap X$ is nonempty, since $B'$ must touch $B$.
We now define a hitting set $S$ for $\mathcal{B}$ as follows. Add an arbitrary element $s$ from $B'\cap X$ to $S$. For each edge $xy\in E(X,Y)$ with $x\in X, y\in Y$, we add $x$ to $S$ if $x\not\in B'$, and otherwise we add $y$ to $S$. Hence $|S|\leq 1+|F|$. See Figure \ref{hittingset} for a depiction of the situation.
\begin{figure}
\centering
\begin{tikzpicture}[punt/.style={draw,circle,inner sep=.9pt},
grootpunt/.style={fill,circle,inner sep=1.3pt}]
\draw[rounded corners=5pt,fill=green!30!gray!6] (0,0) rectangle (5.5,3.5);
\node[anchor=base west] at (0.1,3) {$U$};
\draw[rounded corners=5pt,fill=green!30!gray!6] (6,0) rectangle (11.5,3.5);
\node[anchor=base east] at (11.4,3) {$V\setminus U$};
\draw[rounded corners=5pt,fill=purple!60!gray!10] (4.5,0.1) rectangle (5.3,3.4);
\node[anchor=base] at (4.9,0.3) {$X$};
\draw[rounded corners=5pt,fill=purple!60!gray!10] (6.2,0.1) rectangle (7,3.4);
\node[anchor=base] at (6.6,0.3) {$Y$};
\draw[fill=blue,fill opacity=0.05] (3.5,1.5) ellipse (16mm and 9mm);
\node at (2.6,1.9) {$B'$};
\draw[fill=blue,fill opacity=0.05] (8.3,1.7) ellipse (19mm and 9mm);
\node at (9.2,2.1) {$B$};
\node[grootpunt] (A1) at (4.9,3) {};
\node[punt] (B1) at (6.6,3) {};
\draw (A1) -- (B1);
\node[grootpunt] (B2) at (6.6,2.6) {};
\node[punt] (A2) at (4.7,1.9) {};
\draw (B2) -- (A2);
\node[grootpunt] (B3) at (6.8,2) {};
\node[punt] (A3) at (4.9,1.55) {};
\draw (B3) -- (A3);
\node[grootpunt] (A4) at (4.9,2.5) {};
\node[punt] (B4) at (6.7,1.5) {};
\draw (A4) -- (B4);
\node[grootpunt,label=below:$s$,outer sep=-1pt] (s) at (4.65,1.3) {};
\node[grootpunt] (B5) at (6.6,1.0) {};
\draw (s) -- (B5);
\end{tikzpicture}
\caption{The hitting set $S$ for the bramble $\mathcal{B}$ is formed by the black nodes.\label{hittingset}}
\end{figure}
\medskip
To prove that $S$ covers $\mathcal{B}$, consider any $A\in \mathcal{B}$. First observe that $A$ intersects $X\cup Y$. Otherwise, we would have $A\subseteq U\setminus X$ or $A\subseteq (V\setminus U)\setminus Y$ as $G[A]$ is connected. In the first case $G[A\cup B]$ is not connected and in the second case $G[A\cup B']$ is not connected. In both cases, this contradicts the fact that $\mathcal{B}$ is a bramble.
We consider the following three cases.
\begin{itemize}
\item Case $A\cap Y=\emptyset$.\quad In this case $A\subseteq U$. By the choice of $B'$, we have either $B'\cap X\subseteq A\cap X$ and hence $s\in A$, or there exists an $x\in (X\cap A)\setminus B'$, which implies that $x\in S$. In both situations $S$ intersects $A$.
\item Case $A\cap X=\emptyset$.\quad In this case $A\subseteq (V\setminus U)$. Since $A$ touches $B'$, there must be an edge $e=xy$ with $x\in B'\cap X$ and $y\in A\cap Y$. By construction of $S$ we have $y\in S$. Hence, $S$ intersects $A$.
\item Case $A\cap X\neq \emptyset$ and $A\cap Y\neq \emptyset$.\quad Since $G[A]$ is connected, there is an edge $e=xy$ with $x\in X$, $y\in Y$ and $x,y\in A$. Since $S$ contains at least one endpoint from each edge in $F$, the set $S$ must intersect $A$.
\end{itemize
We conclude that $S$ is a hitting set for $\mathcal{B}$ of size at most $|E(U,V\setminus U)|+1$, which proves the lemma.
\end{proof}
\noindent We now prove the main theorem.
\begin{proof}[Proof of Theorem \ref{main}]
Let $\mathcal{B}$ be a bramble in $G$ of maximum order. That is, $\order{\mathcal{B}} = \tw(G)+1$. Let $D'\geq 0$ be a divisor of positive rank and degree $\dgon(G)$. Among the effective divisors equivalent to $D'$, we choose $D$ such that $\supp(D)$ intersects a maximum number of sets in $\mathcal{B}$. If $\supp(D)$ is a hitting set for $\mathcal{B}$, then we are done:
\begin{equation}
\dgon(G)=\deg(D) \geq \supp(D) \geq \order{\mathcal{B}} > \tw(G).
\end{equation}
We may therefore suppose that $B\in\mathcal{B}$ is not intersected by $\supp(D)$ and let $v\in B$. Since $D$ has positive rank and $D(v)=0$, it follows that $D$ is not $v$-reduced. Hence, by Lemma~\ref{chain}, there exist a chain $\emptyset\subsetneq U_1\subseteq\ldots\subseteq U_k\subseteq V\setminus{\{v\}}$ and a sequence of equivalent effective divisors $D_0:=D,D_1,\ldots,D_k$ such that $D_k$ is $v$-reduced and for every $i=1,\ldots, k$ we have $D_i=D_{i-1}-Q(G)\ensuremath{\mathbf{1}}_{U_i}$. Since $D$ has positive rank, $\supp(D_k)$ contains $v$ and hence intersects $B$.
Let $i\leq k$ be the smallest index such that there is a $B'\in \mathcal{B}$ that is covered by $\supp(D_0)$ but not by $\supp(D_i)$. Such an index exists, since otherwise $\supp(D_k)$ intersects more members of $\mathcal{B}$ then $\supp(D_0)$, contradicting our choice of $D=D_0$. From $B'\cap \supp(D_{i-1})\neq \emptyset$ and $B'\cap \supp(D_{i})=\emptyset$ it follows by Lemma \ref{oneside} that $B'\subseteq U_i$.
Again by our choice of $D$, the set $\supp(D_{i-1})$ does not intersect $B$. Since $\supp(D_k)$ does intersect $B$, there is an index $j\geq i$ such that $B\cap \supp(D_{j-1})=\emptyset$ and $B\cap \supp(D_{j})\neq\emptyset$. Hence, since $D_{j-1}=D_j-Q(G)\ensuremath{\mathbf{1}}_{V\setminus U_j}$, we have $B\subseteq V\setminus U_j\subseteq V\setminus U_i$ by Lemma \ref{oneside}.
Since $B\subseteq V\setminus U_i$ and $B'\subseteq U_i$, it follows by Lemma \ref{brambleseparation} that $|E(U_i,V\setminus U_i)|\geq \order{\mathcal{B}}-1$. Since
\begin{equation}
\deg (D_{i-1})\geq \sum_{u\in U}D_{i-1}(u)\geq |E(U_i,V\setminus U_i)|,
\end{equation}
it follows that $\dgon(G)=\deg(D) = \deg(D_{i-1}) \geq \order{\mathcal{B}} - 1 = \tw(G)$.
\end{proof}
\section{Examples}
We first discuss some classes of graphs for which equality holds in $\tw(G)\leq \dgon(G)$.
\begin{example}
Let $G=(V,E)$ be a \emph{simple} graph with at least one edge. Let $g:=|E|-|V|+1$ be its circuit rank. If $g=0$, then $G$ is a tree and $\tw(G)=\dgon(G)=1$. If $g\in\{1,2\}$, we have $\tw(G)=\dgon(G)=2$.
\end{example}
\begin{example}[Complete $k$-partite graph]
Let $G=(V,E)$ be a complete $k$-partite graph, $k\geq 2$, with partition $V=V_1\cup\cdots\cup V_k$, where $n_i:=|V_i|\geq 1$. We may assume that $n_1\leq n_2\leq \cdots\leq n_k$.
For $i=1,\ldots, k$ let $s_i\in V_i$ and consider the bramble $\mathcal{B}:=\{\{s_1\},\ldots,\{s_k\}\}\cup\{\{u,v\}\mid uv\in E\}$. A set $S\subseteq V$ is a hitting set for $\mathcal{B}$ if and only if $s_1,\ldots,s_k\in S$ and there is at most one index $i$ such that $V_i\not\subseteq S$. Hence a hitting set of minimal cardinality is given by $S:=V_1\cup\cdots\cup V_{k-1}\cup\{s_k\}$. Hence $\tw(G)\geq \order{\mathcal{B}}-1=n_1+\cdots+n_{k-1}$.
Let $D:=\ensuremath{\mathbf{1}}_{V_1\cup\cdots\cup v_{k-1}}$. For every $v\in V_k$, the divisor $D+Q(G)\ensuremath{\mathbf{1}}_{\{v\}}$ is effective. Hence $D$ has rank at least one and therefore $\dgon(G)\leq n_1+\cdots+n_{k-1}$.
We conclude that $\tw(G)=\dgon(G)=n_1+\cdots+n_{k-1}$. In particular we have $\dgon(K_n)=n-1$ for the complete graph on $n$ vertices, and $\dgon(K_{m,n})=m$ for the complete bipartite graph with colour classes of sizes $m\leq n$. For the octahedron $K_{2,2,2}$ we find $\dgon(K_{2,2,2})=4$.
\end{example}
\begin{example}[Rectangular grid]
Let $m\leq n$ be integers and let $G=(V,E)$ be the $(m+1)\times (n+1)$ rectangular grid. That is, $V:=[m+1]\times [n+1]$ and two vertices $(a,b)$ and $(a',b')$ form an edge if $|a-a'|+|b-b'|=1$.
Let $A:=[m+1]\times \{n+1\}$ and $B:=\{m+1\}\times [n]$. For $i\in [m]$ and $j\in [n]$ consider the `cross'
$$C_{ij}:=\{(a,b)\in [m]\times [n]\mid \text{$a=i$ or $b=j$}\}.$$
It is easy to see that $\mathcal{B}:=\{A,B\}\cup\{C_{ij}\}_{i\in [m], j\in [n]}$ is a bramble. Any hitting set for the $C_{ij}$ contains at least $m$ elements from $[m]\times [n]$ (one from each row). Hence, since $A$, $B$, and $[m]\times [n]$ are disjoint, the order of $\mathcal{B}$ is at least $m+2$.
On the other hand, take the divisors $D_i:=\ensuremath{\mathbf{1}}_{[m+1]\times\{i\}}$ for $i=1,\ldots, m+1$. These divisors are equivalent, since $D_{i+1}=D_i-Q(G)\ensuremath{\mathbf{1}}_{[m+1]\times [i]}$ for $i=1,\ldots, n$. Hence, since every $(a,b)\in V$ is in the support of some $D_b$, the rank of $D_1$ is at least one. Hence, we can conclude that $m+1\leq \tw(G)\leq \dgon(G)\leq m+1$, and hence $\dgon(G)=\tw(G)=m+1$.
\end{example}
An interesting family for which we do not know the answer is the following. Let $Q_n$ be the $n$-dimensional cube. That is $Q_n$ is the graph with vertex set $\{0,1\}^n$ and two vertices $x,y$ are connected by an edge if $x$ and $y$ differ in exactly one coordinate. It is clear that $\dgon(Q_n)\leq 2^{n-1}$ and we believe that equality holds. On the other hand, $\tw(Q_n)=\Theta(\frac{2^n}{\sqrt{n}})$, see \cite{SunilKavitha2006}.
\section{Metric graphs}
In this section, we show that for any metric graph $\Gamma$ with underlying connected graph $G$, there is a subdivision $H$ of $G$, such that $\dgon(H)\leq \dgon(\Gamma)$. Hence, the treewidth is also a lower bound for metric graphs:
\begin{equation}\tw(G)\leq \tw(H)\leq \dgon(H)\leq \dgon(\Gamma).\end{equation}
Let $G$ be a connected graph with vertex set $V$ and edge set $E$. Let $l:E\to \ensuremath{\mathbb{R}}_{>0}$ be a length function on the edges. Associated to the pair $(G,l)$ is the \emph{metric graph} $\Gamma$ which is the compact connected metric space obtained by identifying edge $e=\{u,v\}$ with a segment of length $l(e)$. The free abelian group on the points of $\Gamma$ is denoted $\Div(\Gamma)$ and the elements of $\Div(\Gamma)$ are called \emph{divisors} on $\Gamma$. For $D=c_1v_1+\cdots+c_kv_k\in \Div(\Gamma)$, with $c_1,\ldots, c_k\in \ensuremath{\mathbb{Z}}$ and $v_1,\ldots, v_k\in \Gamma$, the \emph{degree} of $D$ is defined as $\deg(D):=c_1+\cdots+c_k$. The divisor is \emph{effective} if $c_1,\ldots, c_k\geq 0$. The support of $D$ is denoted $\supp(D)$.
Let $f$ be a piecewise linear continuous function $f$ on $\Gamma$ with integral slopes. For each $v\in \Gamma$, let $c_v$ be the sum of the outgoing slopes of $f$ at $v$. So $c_v\neq 0$ only for breakpoints of $f$. The associated divisor is denoted $\todiv(f):=\sum_{v\in \Gamma} c_vv$ and is called a \emph{principal} divisor. The set of principal divisors is denoted $\Prin(\Gamma)$ and is a subgroup of $\Div(\Gamma)$. Two divisors are \emph{equivalent} if their difference is a principal divisor.
For a point $s\in \Gamma$, we say that a divisor $D$ \emph{covers} $s$ if there exists an effective divisor equivalent to $D$ with $v$ in its support. The \emph{gonality} $\dgon(\Gamma)$ is defined as the minimum degree of a divisor that covers every point $v\in \Gamma$. It was proven in \cite{Luo2011} that if $D$ covers every $v\in V$, then $D$ covers every $v\in \Gamma$. However, we will not use that result here.
We denote by $\Div_V(\Gamma)$ the subgroup of divisors with support contained in $V$. We identify the elements of $\Div_V(\Gamma)$ with the corresponding elements on $\ensuremath{\mathbb{Z}}^V$. Hence, the divisors in $\Div_V(\Gamma)$ can also be seen as divisors on $G$. By $\mathcal{C}(\Gamma)$ we denote the set of continuous piecewise linear functions $f$ on $\Gamma$ with integral slopes and $\todiv(f)\in \Div_V(\Gamma)$. This last condition simply means that $f$ is linear on each edge of $\Gamma$. Observe that any two divisors $D,D'\in \Div_V(\Gamma)$ are equivalent if and only if $D-D'=\todiv(f)$ for some $f\in \mathcal{C}(\Gamma)$.
We fix an arbitrary orientation on $G$. We define a map $\phi:\mathcal{C}(\Gamma)\to \ensuremath{\mathbb{Z}}^E$ by setting $\phi(f)(e)$ to be the slope of $f$ on edge $e$ (in the forward direction). Let $g:E\to \ensuremath{\mathbb{Z}}$. It is easy to see that $g$ is in the image of $\phi$ if and only if
\begin{equation}
\sum_{e\in E} g(e)l(e)\chi_C(e)=0\quad \text{for every cycle $C$ in $G$}.
\end{equation}
Observe that $\todiv(f)=-M \phi(f)$, where $M$ is the signed vertex-edge incidence matrix of $G$. Also observe that for $l=\ensuremath{\mathbf{1}}$, the function that is identically one, $g$ is in the image of $\phi$ if and only if $g=M^\mathsf{T} x$ for some $x\in \ensuremath{\mathbb{Z}}^V$ by Lemma \ref{clattice}. Hence $\todiv(f)=-M\phi(f)=-MM^\mathsf{T} x=-Q(G)x$ for some $x\in \ensuremath{\mathbb{Z}}^v$. In other words, two divisors in $\Div_V(\Gamma)$ are equivalent if and only if they are equivalent as divisors on $G$.
\begin{theorem}
Let $\Gamma$ be the metric graph associated to $(G,l)$. Then there is a subdivision $H$ of $G$ such that $\dgon(\Gamma)\geq \dgon(H)$.
\end{theorem}
\begin{proof}
Let $D$ be a minimum degree divisor covering $\Gamma$. In particular, $D$ covers every $v\in V$. Hence, for every $v\in V$, there is an effective divisor $D_v$ equivalent to $D$ with $v$ in its support. Let $V':=V\cup \supp(D)\cup\bigcup_{v\in V}\supp(D_v)$. Let $\Gamma'$ be obtained by subdividing $\Gamma$ at the points in $V'\setminus V$. Denote by $G'$ and $l'$ the corresponding underlying graph and length function so that $\Gamma'$ is the metric graph associated with $(G',l')$. The divisor $D$ and the divisors $D_v$ can now be seen as equivalent elements of $\Div_{V'}(\Gamma')$.
For all $v\in V$, let $f_v\in \mathcal{C}(\Gamma')$ be such that $D-\todiv(f_v)=D_v$. It follows that $y=l'$ is a solution to the system
\begin{equation}\label{cycle}
\sum_{e\in G'}y(e)\phi(f_v)(e)\chi_C(e)=0\quad \text{for every cycle $C$ in $G$ and every $v\in V$}.
\end{equation}
Since (\ref{cycle}) is a (finite) rational linear system in $y$, and since $l'>0$ is a solution, the system also has a solution $l''\in \ensuremath{\mathbb{Z}}_{>0}^E$. It follows that the $D_v$ are equivalent divisors on the metric graph associated with $(G', l'')$. Subdividing every edge $e$ of $G'$ into $l''$ parts to obtain a graph $H$, we can view the $D_v$ as equivalent divisors in $\Div_{V'}(\Gamma'')$, where $\Gamma''$ is the metric graph associated to $(H,\ensuremath{\mathbf{1}})$ in which all edges have length one. Finally, this implies that the $D_v$ are also equivalent as divisors of $H$. It follows that for any $v\in V$, the divisor $D_v\in \Div(H)$ covers $V$, and hence by Corollary \ref{rankdetermining} has positive rank.
\end{proof}
The following corollary is immediate.
\begin{corollary}
Let $\Gamma$ be a metric graph with underlying connected graph $G$. Then $\tw(G)\leq \dgon(\Gamma)$.
\end{corollary}
\section{Other notions of gonality}
Other notions of gonality of a graph $G$ have been proposed by Caporaso \cite{Caporaso2012} and by Cornelissen, Kato, and Kool in \cite{CornelissenKatoKool2012}. These notions are based on harmonic morphisms from $G$ to a tree. Here we will show that treewidth is also a lower bound for the gonality in these cases. Again, we assume that our graphs are connected, finite, and loopless (but possibly with multiple edges).
We follow terminology from \cite{BakerNorine2009}.
A \emph{morphism from $G=(V,E)$ to $G'=(V',E')$}, is a map $\phi:V\cup E\to V'\cup E'$ such that
\begin{itemize}
\item[(i)] $\phi(V)\subseteq V'$,
\item[(ii)] if $e\in E(u,v)$, then either $\phi(e)=\phi(u)=\phi(v)$, or $\phi(e)\in E'(\phi(u),\phi(v))$.
\end{itemize}
If $\phi(E)\subseteq E'$, then $\phi$ is called a \emph{homomorphism}. We call a morphism $\phi$ \emph{harmonic} if
\begin{itemize}
\item[(iii)] for every $v\in V$ there exists a nonnegative integer $m_\phi(v)$ such that
\begin{equation}\label{harm}
m_\phi(v)=|\phi^{-1}(e')\cap E(v)|\quad\text{for every $e'\in E'(\phi(v))$,}
\end{equation}
\end{itemize}
\noindent and \emph{non-degenerate} if in addition
\begin{itemize}
\item[(iv)] $m_\phi(v)\geq 1\quad$ for every $v\in V$.
\end{itemize}
\noindent If $\phi$ is harmonic, then there is a number $\deg(\phi)$ such that for every edge $e'\in E'$ and every $v'\in V'$
$$
\deg(\phi)=|\phi^{-1}(e')|=\sum_{v\in \phi^{-1}(v')}m_\phi(v).
$$
\begin{lemma}
Let $G=(V,E)$ and $G'=(V',E')$ be graphs and let $\phi:G\to G'$ be a non-degenerate harmonic morphism. Then $\dgon(G)\leq \dgon(G')\deg(\phi)$. In particular, $\dgon(G)\leq \deg(\phi)$ when $G'$ is a tree.
\end{lemma}
\begin{proof}
For any divisor $D\in \Div(G')$, define the divisor $\phi^*(D)\in \Div(G)$ by $\phi^*(D)(v):=m_\phi(v)D(\phi(v))$. Observe that $\deg(\phi^*(D))=\deg(D)\deg(\phi)$, and that its support is $\phi^{-1}(\supp(D))$ by non-degeneracy of $\phi$. When $D$ is effective, then so is $\phi^*(D)$.
It is easy to see that if $D,D'\in \Div(G')$ are equivalent, then $\phi^*(D)$ and $\phi^*(D')$ are equivalent as well. Indeed, for any $y\in \ensuremath{\mathbb{Z}}^{V'}$ we have $\phi^*(Q(G')y)=Q(G)x$, where $x(u):=y(\phi(u))$.
Hence, if $D\in\Div(G')$ is an effective divisor of positive rank in $G'$, then $\phi^*(D)$ is an effective divisor of positive rank in $G$ with $\deg(\phi^*(D))=\deg(D)\deg(\phi)$.
\end{proof}
The notion of harmonic morphism can be extended to \emph{indexed harmonic morphism} by associating to every edge $e\in \phi^{-1}(E')$ a positive integer $r_e$ and counting in (\ref{harm}) every edge $e\in \phi^{-1}(e')$ with multiplicity $r_e$. Hence, an indexed harmonic morphism $G\to G'$ corresponds to a harmonic morphism $H\to G'$, where $H$ is obtained from $G$ by replacing every edge $e$ by $r_e$ parallel edges which are mapped to the same edge as the original edge $e$.
In \cite{Caporaso2012}, Caporaso defined the gonality of a graph $G$ as the minimum degree of a non-degenerate indexed harmonic morphism (with some additional restriction) from $G$ to a tree. Hence it follows that this measure of gonality is lower bounded by $\dgon(H)$ for some $H$ obtained from $G$ by adding parallel edges, and hence by $\tw(H)=\tw(G)$.
In \cite{CornelissenKatoKool2012}, Cornelissen, Kato and Kool define the \emph{stable gonality} $\sgon(G)$ of $G$ to be the minimum degree of an indexed harmonic homomorphism from a \emph{refinement} of $G$ to a tree $T$. Note that a harmonic homomorphism is automatically non-degenerate. A refinement of $G$ is a graph obtained from $G$ by subdividing edges and adding leaves (nodes of degree 1). Therefore $\sgon(G)$ is lower bounded by $\dgon(H)$ for some graph $H$ obtained from $G$ by subdividing edges, adding leaves and adding parallel edges. Hence, $\sgon(G)\geq \dgon(H)\geq \tw(H)\geq \tw(G)$.
For a comparison of the different notions of gonality, we refer the reader to \cite{CornelissenKatoKool2012}.
\section{Acknowledgements}
We would like to thank Maarten Derickx. He was the first to prove that the gonality of the $n\times m$ grid equals $\min(m,n)$ (unpublished). His method inspired us to conjecture and prove that $\tw(G)\leq \dgon(G)$. The second author would also like to thank Jan Draisma for some stimulating discussions on graph gonality.
\bibliographystyle{plain}
|
\section{Introduction}
In this paper we study a class of self-adjoint operators in a Kre\u{\i}n space which turns out to be similar to self-adjoint operators in a Hilbert space. Recall that a Kre\u{\i}n space is a complex vector space $\cK$ with a sesquilinear form $\kip$ such that there exist subspaces $\cK_+$ and $\cK_-$ of $\cK$ with the following three properties: $\bigl[\cK_+,\cK_-\bigr] = \{0\}$, $\cK = \cK_+ \dot{+} \cK_-$, a direct sum, and the spaces $(\cK_+, \kip)$ and $(\cK_-, -\kip)$ are Hilbert spaces. A pair of subspaces $\cK_+$ and $\cK_-$ with the preceding three properties is called a {\em fundamental decomposition} of a Kre\u{\i}n space $\cK$. The projections $P_+$ and $P_-$ associated with the direct sum $\cK = \cK_+ \dot{+} \cK_-$ are called {\em fundamental projections} and the operator $J = P_+ - P_-$ is called a {\em fundamental symmetry} of a Kre\u{\i}n space. The space $\cK$ with the inner product $\la x, y \ra = [Jx,y]$, $x,y \in \cK$, is a Hilbert space. Neither a fundamental decomposition nor a fundamental symmetry of a Kre\u{\i}n space is unique. However, the Hilbert space norms generated by different fundamental decompositions via the corresponding fundamental symmetries are equivalent. All topological notions in a Kre\u{\i}n space refer to the topology of the Hilbert space $(\cK, \ahip)$. For the general theory of Kre\u{\i}n spaces and operators acting in them we refer to the monographs \cite{AI, B74}.
Unlike the spectrum of a self-adjoint operator in a Hilbert space, the spectrum of a general self-adjoint operator in a Kre\u{\i}n space can be quite arbitrary. Therefore it is of interest to look for conditions that would guarantee good spectral properties of a self-adjoint operator in a Kre\u{\i}n space.
The ultimate task in this direction is to provide sufficient conditions for a self-adjoint operator in a Kre\u{\i}n space to be similar to a self-adjoint operator in a Hilbert space. A simple characterization of similarity is as follows. A self-adjoint operator $A$ in a Kre\u{\i}n space $(\cK,\kip)$ is similar to a self-adjoint operator in the Hilbert space $(\cK,\ahip)$ if and only if $A$ is fundamentally reducible in $(\cK,\kip)$; where {\em fundamentally reducible} means that there exists a fundamental decomposition $\cK = \cK_+ \dot{+} \cK_-$ of $(\cK,\kip)$ such that $A$ is the direct sum of its restrictions to $\cK_+\!\cap (\dom A)$ and $\cK_- \cap (\dom A)$. Equivalently, an operator $A$ is fundamentally reducible in a Kre\u{\i}n space $(\cK,\kip)$ if and only if there exists a fundamental symmetry $J$ on $(\cK,\kip)$ such that $J \dom(A) \subseteq \dom(A)$ and $JAx = AJx$ for all $x \in \dom(A)$. The above characterization of similarity for bounded operators appears in \cite[Theorem~1, Section~2]{McEn82} and in \cite[Proposition~2.2]{KuTr11}, where an equivalent terminology of $\cC$-symmetry is used.
Another kind of a self-adjoint operator in a Kre\u{\i}n space whose spectral properties resemble those of a self-adjoint operator in a Hilbert space is a nonnegative self-adjoint operator with a nonempty resolvent set. The spectrum of such an operator is real and, excluding arbitrary neighborhoods of $0$ and $\infty$, the operator $A$ has a projector valued spectral function whose properties resemble the properties of the spectral function of a self-adjoint operator in a Hilbert space; for details see \cite{La} and Subsection~\ref{ss:non-Ks} below for a short review. If the spectrum of $A$ accumulates on both sides of $0$ ($\infty$), then $0$ ($\infty$, respectively) is called a {\em critical point} of $A$. If the spectral function of $A$ is bounded in a neighborhood of a critical point, then that critical point is said to be {\em regular}. Otherwise, it is said to be a {\em singular} critical point. The set of all singular critical points of $A$ is denoted by $c_s(A)$. Here, by definition, $c_s(A)\subseteq \{0,\infty\}$. These concepts are closely related to the similarity of $A$ to a self-adjoint operator in a Hilbert space: A nonnegative self-adjoint operator $A$ in a Kre\u{\i}n space such that $\rho(A) \neq \emptyset$ is similar to a self-adjoint operator in a Hilbert space if and only if $\ker(A) = \ker(A^2)$ and $0, \infty \not\in c_s(A)$.
Our first step in studying the similarity question is to introduce a new concept related to the fundamental reducibility in Kre\u{\i}n spaces. For the definition of defect numbers of symmetric operators which we use in the next definition and for general Hilbert space theory see \cite{AG}.
\begin{definition}\label{def:PartFR}
A self-adjoint operator $A$ in a Kre\u{\i}n space $(\cK,\kip )$ is called {\em partially fundamentally reducible} if there exists a fundamental decomposition $\cK = \cK_+[\dot{+}]\cK_-$ of $(\cK,\kip)$ such that the subspaces
\begin{equation*}
{\mathcal D}_+\!\!=\!\bigl\{ f \in \cK_+\!\cap\!(\dom A)\! : \! Af \in \cK_+ \!\bigr\}
\ \ \text{and} \ \
{\mathcal D}_-\!\!=\!\bigl\{ f \in \cK_-\!\cap\!(\dom A)\! : \! Af \in \cK_-\!\bigr\}
\end{equation*}
are dense in $\cK_+$ and $\cK_-$ and the restrictions
\begin{equation}\label{eq:Spm}
S_+ = A|_{{\mathcal D}_+}\quad\text{and} \quad S_- = - A|_{{\mathcal D}_-}
\end{equation}
are symmetric operators with defect numbers $(1,1)$
in the Hilbert spaces $(\cK_+, \kip)$ and $(\cK_-, -\kip)$, respectively.
\end{definition}
The idea behind the above definition is to introduce a desirable restriction of an operator $A$ and then utilize those desirable properties of the restriction to study $A$. This method resembles the Glazman decomposition method from \cite{Gl}. A similar idea was also used to study definitizability of differential operators with indefinite weights in \cite[Section~3]{Beh07} and definitizability of self-adjoint operators in Kre\u{\i}n spaces in \cite[Theorem~3.5]{BehPhi10}.
Our objective in this paper is to give sufficient conditions under which a nonnegative partially fundamentally reducible operator in a Kre\u{\i}n space is similar to a self-adjoint operator in a Hilbert space.
To this end we will use a boundary triple approach to extension theory developed in \cite{Koch79, GG, DM91}; see Subsection~\ref{SecW} below for a brief review. We will apply this theory to the symmetric operators $S_+$ and $S_-$ associated via~\eqref{eq:Spm} with a partially fundamentally reducible operator $A$.
Specifically, let $\bigl(\nC,\Gamma^+_0,\Gamma^+_1\bigr)$ be a boundary triple of the operator $S_+^*$, the adjoint of $S_+$ in the Hilbert space $(\cK_+, \kip)$, and let $m_+$ be the corresponding Weyl function. Then there exists a unique boundary triple $\bigl(\nC,\Gamma^-_0,\Gamma^-_1\bigr)$ for $S_-^*$ such that the operator $A$ is a coupling of $S_+$ and $S_-$ relative to the boundary triples $\bigl(\nC,\Gamma^+_0,\Gamma^+_1\bigr)$ and $\bigl(\nC,\Gamma^-_0,\Gamma^-_1\bigr)$. That is, $f \in \dom(A)$ if and only if there exist $f_+ \in \dom\bigl(S_+^*\bigr)$ and $f_- \in \dom\bigl(S_-^*\bigr)$ such that
\[
f = f_+ + f_- \quad \text{and} \quad \Gamma_0^+ f_+ = \Gamma_0^- f_-, \quad \Gamma_1^+ f_+ = - \Gamma_1^- f_-
\]
(see Theorems~\ref{tHsc} and~\ref{tKsc} below). Let $m_-$ be the Weyl function of $S_-$ corresponding to the boundary triple $\bigl(\nC,\Gamma^-_0,\Gamma^-_1\bigr)$. The Weyl functions $m_+$ and $m_-$ belong to the class of Nevanlinna functions, for the definition and basic properties see Subsection~\ref{sect:NS} below. These functions completely characterize the simple (non-self-adjoint) parts of the symmetric operators $S_+$ and $S_-$ acting in the Hilbert spaces $\cK_+$ and $\cK_-$. Therefore, it is natural to look for conditions for the fundamental reducibility of $A$ in terms of the local behavior of the associated Weyl functions $m_+$ and $m_-$ at $0$ and $\infty$.
The coupling method used here is a combination of the Glazman decomposition method and the boundary triple approach to the extension theory. The coupling method was worked out for operators in Hilbert spaces in~\cite{DHMS} and it was used in \cite{Ka}, and also in \cite{KaMal07, KaKost08, KaKoMal09, Ka09, Kost13}, to study the problem of similarity of differential operators with indefinite weights to self-adjoint operators in Hilbert spaces.
Using this method, see \cite{KaMal07, KaKost08}, it was proved that the boundedness of the function
\begin{equation} \label{eq:D0-Dinf}
y \mapsto \frac{\im m_+(iy)+\im m_-(iy)}{m_+(iy)+m_-(-iy)}, \quad y > 0,
\end{equation}
on $(0,1)$ (on $(1,\infty)$, respectively) is necessary for $0 \not\in c_s(A)$ ($\infty \not\in c_s(A)$, respectively). Since we use these necessary conditions in an essential way, we introduce the following terminology. A pair of functions $(m_+,m_-)$ is said to have $D_0$-{\em property} ($D_\infty$-{\em property}) if the function in \eqref{eq:D0-Dinf} is bounded on $(0,1)$ (on $(1,\infty)$, respectively).
Next we introduce a different kind of local behavior of a Nevanlinna function. A Nevanlinna function $m$ with the integral representation \eqref{eq:IntRep+} in Subsection~\ref{SubSec-B-prop-def} is said to have $B_0$-property ($B_\infty$-property, respectively) if the mapping
\begin{equation*}
f \mapsto \int_{0}^{+\infty}\! \frac{f(x)}{x+y}\,d\sigma(x),
\end{equation*}
is a bounded mapping from $L_{\sigma}^2({\mathbb R}_+)$ into $L_{w_m}^2\!(0,1)$ ($L_{w_m}^2\!(1,\infty)$, respectively). Here $w_m(y) = \bigl(\im m(iy)\bigr)^{-1}$, the reciprocal of the imaginary part of $m$.
With $D$- and $B$-properties our main results are as follows. If $A$ is a nonnegative partially fundamentally reducible operator and if associated Weyl functions $m_+$ and $m_-$ have $B_\infty$-property, then $\infty \not\in c_s(A)$ if and only if the pair $(m_+, m_-)$ has $D_\infty$-property (Theorem~\ref{Main}). Analogously, if $A$ is a nonnegative partially fundamentally reducible operator and if the associated Weyl functions $m_+$ and $m_-$ have $B_0$-property, then $0 \not\in c_s(A)$ if and only if the pair $(m_+, m_-)$ has $D_0$-property and $\ker(A) = \ker(A^2)$ (Theorem~\ref{Main0}). Together these two results give sufficient conditions for a nonnegative partially fundamentally reducible operator in a Kre\u{\i}n space $(\cK,\kip)$ to be similar to a self-adjoint operator in a Hilbert space.
Furthermore, in Subsection~\ref{sSecAsy} we define the asymptotic class $\cA_\infty$ of Nevanlinna functions which all satisfy $B_\infty$-property and pairs of which have $D_\infty$-property and the analogous asymptotic class $\cA_0$ for $B_0$-property and $D_0$-property. In Theorem~\ref{t:AllPos} we prove that if the Weyl functions $m_+$ and $m_-$ associated with the partially fundamentally reducible operator $A$ both belong to $\cA_\infty \cap \cA_0$, then, not only $A$, but all the nonnegative self-adjoint extensions of $S_+\oplus(-S_-)$ are similar to a self-adjoint operators in a Hilbert space.
Finally, in Section~\ref{SecExe} we apply our results to indefinite Sturm-Liouville differential operators. In some cases our results lead to a new point of view at some results from \cite{CL, CN, F, FN98, Ka, KaMal07, KaKost08, KaKoMal09}. We also get some new results for the case of nonsymmetric coefficients and the case when $A$ is a coupling of two differential operators of different order.
Since a self-adjoint operator in a Kre\u{\i}n space is similar to a self-adjoint operator in a Hilbert space if and only if it is fundamentally reducible, throughout the rest of the paper we will use the shorter phrase fundamentally reducible to relate to this property.
The main results of this paper were presented at the 21st International Workshop on Operator Theory and Applications held in July of 2010 in Berlin, Germany.
The authors would like to thank Victor Katsnelson and Mark Malamud for useful discussions and relevant literature suggestions.
\subsection{Notation.} We use the standard notation $\nC$ for the set of complex numbers and $\nR$ for the set of real numbers. By $\nC_+$ we denote the set of all $z \in \nC$ with positive imaginary part. Similarly, $\nR_+$ ($\nR_-$) stands for the set of all positive (negative, respectively) reals. For $z \in \nC$, $\co{z}$, $\re z$, $\im z$ and $\Arg z$ denote the complex conjugate, real, imaginary part of $z$ and the principal value of the argument of $z$ with $\Arg z \in (-\pi, \pi]$, respectively.
For a noninteger real $\alpha$ and $z \in \nC\setminus[0,+\infty)$ we designate the principal branch of $z^\alpha$ to be $|z|^\alpha \exp\bigl(i \alpha \Arg(z)\bigr)$, where $\Arg(z) \in (-\pi, \pi)$.
All operators in this paper are closed densely defined linear operators. For such an operator $T$, we use the common notation $\rho(T)$, $\dom(T)$, $\ran(T)$ and $\ker(T)$ for the resolvent set, the domain, the range and the null-space, respectively, of $T$.
We use the symbols $\pm$ and $\mp$ in a very specific way. Each sentence in which one or both of these symbols appear should be read twice, the first time with the top symbols and the second time with the bottom symbols. A good example of this is the sentence containing \eqref{CritTpm} and \eqref{CritTpm-+} in Subsection~\ref{ss:RegCPs}. Reading this sentence the first time defines $T_{++}$ and $T_{+-}$; reading it the second time defines $T_{--}$ and $T_{-+}$. The only exception to this rule is the symbol $\operatorname{SL}^\pm(2,\nR)$ which is the common symbol for one object: the group of all real matrices with determinant equal to $-1$ or $1$.
\section{Preliminaries}
\subsection{Weyl functions of symmetric operators} \label{SecW}
Let $S$ be a closed densely defined symmetric operator in the
Hilbert space $\bigl(\cH,\ahip_\cH\bigr)$ and let $\wh\rho(S)$ denotes the set of points of regular type of $S$, see~\cite{AG}. The subspace
\begin{equation*}
\sN_z=\cH\ominus\ran(S-\bar z)=\ker(S^*-z), \quad z\in\wh\rho(S)
\end{equation*}
is called the defect subspace of the operator $S$. The dimension $\dim(\sN_z)$ is constant on each of the open half-planes $\nC^+$ and $\nC^-$ and is denoted by $d_+,$ for $z \in \nC^+$ and $d_-,$ for $z \in \nC^-$. The numbers $d_+$ and $d_-$ are called the upper and lower defect numbers of $S$. In this paper we
assume that $d_+ = d_- = 1$.
Since the space $(\dom S^*) /(\dom S)$ is two dimensional there exist
(unbounded) non-zero linear functionals $\Gamma_j: S^* \to \nC, \ i,
j \in \{0,1\}$, such that
\[
\ker(\Gamma_0)\cap \ker(\Gamma_1) = \dom(S)
\]
and the abstract Green's identity
\begin{equation} \label{eGreen}
\la S^* f , g \ra_\cH - \la f , S^* g \ra_\cH = \Gamma_0(g)^* \,
\Gamma_1(f) - \Gamma_1(g)^* \, \Gamma_0(f),
\end{equation}
holds for all $f,g\in \dom(S^*)$, see \cite[Section~3.1.4]{GG} for much
more general setting. In \cite{DM95} the triple $\bigl(\nC,
\Gamma_0, \Gamma_1\bigr)$ is called the {\em boundary triple} of a
symmetric operator $S$.
It follows from \eqref{eGreen} that the extensions $S_0$, $S_1$ of
$S$ defined as restrictions of $S^*$ to the domains
\begin{equation*}
\dom(S_0) := \ker(\Gamma_0) \qquad \text{and} \qquad \dom(S_1) :=
\ker(\Gamma_1),
\end{equation*}
are self-adjoint linear operators in $\bigl(\cH,\ahip_\cH\bigr)$.
For all $z \in \rho(S_0)$ we have
\begin{equation} \label{eS*ds}
\dom(S^*) = (\dom S_0) \dotplus \sN_z \quad \text{ direct sum in }
\quad \cH.
\end{equation}
Since $\sN_z$ is one-dimensional and $\Gamma_0 \neq 0$, it follows
that the restriction $\Gamma_0|_{\sN_z}$ of $\Gamma_0$ to $\sN_z$ is
a bijection between $\sN_z$ and $\nC$. Then
\[
\bigl( \Gamma_0|_{\sN_z} \bigr)^{-1}: \nC \longrightarrow \sN_z
\subset \cH, \quad z \in \rho(S_0),
\]
and we define the following two functions:
\[
\psi: \nCR\longrightarrow \cH \qquad \text{and} \qquad m: \nCR
\longrightarrow \nC
\]
by
\begin{equation} \label{eWeyl}
\psi(z) := \bigl( \Gamma_0|_{\sN_z} \bigr)^{-1}(1)\quad \text{and}
\quad m(z) := \Gamma_1 \bigl( \Gamma_0|_{\sN_z} \bigr)^{-1}(1) ,
\quad z \in \rho(S_0).
\end{equation}
Clearly,
\begin{equation}\label{esubg1}
m(z) = \Gamma_1\psi(z) .
\end{equation}
The functions $\psi$ and $m$ are called the {\em Weyl solution} and
the {\em Weyl function}, respectively, of the symmetric operator
$S$ relative to the boundary triple $\bigl(\nC, \Gamma_0,
\Gamma_1\bigr)$. For a fixed $z \in \rho(S_0)$ the vector
$f=\psi(z) $ is the solution of the boundary value problem
\[
S^*f = z f, \quad \Gamma_0 f = 1, \quad f \in \dom(S^*).
\]
With $z,w \in \rho(S_0)$, substituting
\begin{equation*}
f = \psi(z), \quad S^*f = z\psi(z), \quad g = \psi(w), \quad S^* g = w\psi(w)
\end{equation*}
in \eqref{eGreen} and using $\Gamma_0 f = \Gamma_0 g = 1$, $m(z) = \Gamma_1 f$, $m(w) = \Gamma_1 g$ we get
\begin{equation} \label{eQfun}
m(z) - \co{m(w)} = (z - \co{w})\,\la \psi(z),\psi(w) \ra_\cH \quad \text{for all} \quad
z,w \in \rho(S_0).
\end{equation}
With $w = \co{z}$
the identity
\eqref{eQfun} yields that the Weyl function $m$ satisfies the
symmetry condition
\begin{equation} \label{esym}
m(\co{z}) = \co{m(z)} \quad \text{for all} \quad z \in \rho(S_0).
\end{equation}
The identity \eqref{eQfun} was used in \cite{KL1} as a definition
of the $Q$-function. It follows from \eqref{eQfun} and \eqref{esym} that $m$
is a Nevanlinna function; for the definition and basic properties see Subsection~\ref{sect:NS} below.
\begin{remark}
It turns out (see \cite[page 8]{DM91}) that the Weyl solution can be
used to evaluate $\Gamma_1(S_0-z)^{-1}h$ for arbitrary $h \in \cH$
and $z \in \rho(S_0)$:
\begin{equation} \label{eG1A0}
\Gamma_1(S_0-z)^{-1}h = \bigl\la h,\psi(\co{z}) \bigr\ra_\cH.
\end{equation}
Indeed, substituting in~\eqref{eGreen} $f=(S_0-z)^{-1}h$,
$g=\psi(\co{z})$, and using $\Gamma_0f = 0$, $\Gamma_0\psi(\co{z})=1$,
we obtain the equality
\[
\bigl\la (S_0-z)f,\psi(\co{z}) \bigr\ra_\cH
= (\Gamma_1 f)(\Gamma_0 \psi(\co{z}))^*- (\Gamma_0 f)(\Gamma_1 \psi(\co{z}))^* = \Gamma_1f,
\]
which proves \eqref{eG1A0}.
\end{remark}
\begin{proposition}\label{prop:S_01}
{\rm\cite{AG,DM91}}
For every $z\in\rho(S_0)$ the following equivalence hold:
\begin{equation*}
z\in\rho(S_1)\quad \Longleftrightarrow \quad m(z)\ne 0
\end{equation*}
and the resolvent of $S_1$ can be found by the formula
\begin{equation*}
(S_1-z)^{-1} h =(S_0-z)^{-1} h - \frac{\bigl\langle h,\psi(\co{z})\bigr\rangle_{{\cH}}}{m(z)} \,
\psi(z)
\end{equation*}
for all $h \in \cH$ and all $z \in \rho(S_0) \cap\rho(S_1)$.
\end{proposition}
\begin{remark} \label{rem:transp}
Any two boundary triples $\bigl(\nC,
\Gamma_0, \Gamma_1\bigr)$ and $\bigl(\nC,
\wh{\Gamma}_0, \wh{\Gamma}_1\bigr)$ of the same symmetric operator $S$ are related by
\begin{equation*}
\begin{pmatrix}
\wh{\Gamma}_0\\
\wh{\Gamma}_1
\end{pmatrix} = W
\begin{pmatrix}
\Gamma_0\\
\Gamma_1
\end{pmatrix},
\end{equation*}
where $W$ is a complex $2\!\times\!2$ matrix satisfying $W^* \V{J} W= \V{J}$, with
$\V{J}=\begin{pmatrix}
0 & -i\\
i & 0
\end{pmatrix}$. The condition $W^* \V{J} W= \V{J}$ is equivalent to $W$ being a unitary matrix in the Kre\u{\i}n space $\bigl(\nC^2, \kip_2\bigr)$, where $[\mb{x},\mb{y}]_2 = \mb{y}^*\! \V{J} \mb{x}$ for $\mb{x},\mb{y} \in \nC^2$. A direct calculation shows that a unitary matrix $W$ in $\bigl(\nC^2, \kip_2\bigr)$ allows a factorization
\[
W = e^{i \theta} \begin{pmatrix}
a & b\\
c & d
\end{pmatrix} \quad \text{where} \quad \begin{pmatrix}
a & b\\
c & d
\end{pmatrix} \in \operatorname{SL}(2,\nR) \quad \text{and} \quad \theta \in (-\pi, \pi].
\]
Here $\operatorname{SL}(2,\nR)$ is the special linear group of all real $2\!\times\!2$ matrices with determinant one.
The Weyl function $m$ relative to the boundary triple $\bigl(\nC,
\Gamma_0, \Gamma_1\bigr)$ and the Weyl function $\wh{m}$ relative to the boundary triple $\bigl(\nC, \wh{\Gamma}_0, \wh{\Gamma}_1\bigr)$ are related by
\[
\wh{m}(z) = \frac{d\rule{1.2pt}{0pt}m(z)+c}{b\rule{1.2pt}{0pt}m(z)+a}, \qquad z \in \rho(S_0) \cap\rho(S_1).
\]
In particular, the boundary triple $\bigl(\nC,-\Gamma_1, \Gamma_0\bigr)$ is said to be a {\it transpose} of the boundary triple $\bigl(\nC, \Gamma_0, \Gamma_1\bigr)$. The Weyl function $m^{\!\top}$ of $S^*$ relative to $\bigl(\nC,-\Gamma_1, \Gamma_0\bigr)$ is given by
$m^{\!\top}(z)=-{1}/{m(z)}$, $z \in \nCR$.
\end{remark}
\subsection{Nonnegative operators in Hilbert spaces} \label{sect:NonnegSO}
Recall that a symmetric operator $S$ in a
Hilbert space $\bigl(\cH,\ahip_\cH\bigr)$ is called nonnegative if $\langle Sf,f\rangle_{\cH}\ge 0$ for all $f\in\dom(S)$. By a result of Friedrichs, every nonnegative symmetric operator $S$ admits a nonnegative self-adjoint extension. Moreover, as was shown by Kre\u{\i}n \cite{Kr47}, in the set $\operatorname{Ext}_+(S)$ of all nonnegative self-adjoint extensions of a nonnegative symmetric operator $S$ there are two extremal extensions $S_F$ and $S_K$. The extensions $S_F$ and $S_K$, which are called the Friedrichs extension and the Kre\u{\i}n extension, respectively, are maximal and minimal in the following sense: for all $\wt S \in \operatorname{Ext}_+(S)$ and all $a > 0$ we have,
\[
(S_F+a)^{-1}\le(\wt S+a)^{-1}\le(S_K+a)^{-1}.
\]
The extensions $S_F$ and $S_K$ can be characterized in terms of boundary triples and Weyl functions. If $S$ is a symmetric nonnegative operator and $\bigl( \nC,\Gamma_0,\Gamma_1\bigr)$ is a boundary triple for $S^*$, then the extension $S_0$ of $S$ is also nonnegative if and only if the corresponding Weyl function $m$ is holomorphic on $\nR_-$.
Moreover (see~\cite[Proposition~10]{DM91}), the following two equivalences hold:
\begin{align} \nonumber
S_0=S_F \quad &\Longleftrightarrow \quad \lim_{x\downarrow -\infty} m(x) =-\infty,\\
\label{eq:m_K0}
S_0=S_K \quad &\Longleftrightarrow \quad \rule{2pt}{0pt}\lim_{x\uparrow 0} m(x) = +\infty.
\end{align}
\subsection{Nevanlinna, Stieltjes and inverse Stieltjes functions} \label{sect:NS}
A complex function $m$ is called a {\em
Nevanlinna function} if $m$ is holomorphic at least on $\nCR$ and satisfies the following two conditions
\begin{equation*}
m(\co{z}) = \co{m(z)} \quad \text{and} \quad \im m(z) \ge 0, \qquad \text{for all} \qquad z \in \nC_+.
\end{equation*}
Equivalently, $m$ is a Nevanlinna function if and only if there exist $a, b \in \nR$, $b \geq 0$, and a nondecreasing function $\sigma:\nR \to \nR$ such that
\begin{equation*}
m(z) = a + b z +\int_{\nR} \left( \frac{1}{t-z} -
\frac{t}{1+t^2} \right) d\sigma(t)
\end{equation*}
and
\begin{equation}\label{eq:sigmaCond}
\int_{\nR} \dfrac{d\sigma(t)}{1+t^2} < + \infty.
\end{equation}
If, additionally, $\sigma$ is normalized by
\begin{equation} \label{eq:sigma-nor}
\sigma(t) = \frac{\sigma(t-0)+\sigma(t+0)}{2} \quad \text{and} \quad \sigma(0) = 0,
\end{equation}
then it is uniquely determined by $m$. For these and other facts on Nevanlinna functions see \cite{KaKr74} and \cite[Chapter~II]{Do}.
We consider that a Nevanlinna function is defined on its {\em domain of holomorphy}. That is, the domain of a Nevanlinna function $m$ coincides with the union of $\nCR$ and the set of all those real points to which $m$ admits a holomorphic continuation.
A Nevanlinna function $m$ is called a {\em Stieltjes} (an {\em inverse Stieltjes}) function if it is holomorphic on $\nC\setminus[0,+\infty)$ and it takes nonnegative (nonpositive, respectively) values on $\nR_-$.
The following proposition is a direct consequence of the definitions.
\begin{proposition}\label{p:SiS}
Let $m$ be a Nevanlinna function. The following statements are equivalent.
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item
$m$ is an inverse Stieltjes function.
\item
The function $z \mapsto -m(1/z)$, $z \in \nCR$, is a Stieltjes function.
\item
The function $z \mapsto -1/m(z)$, $z \in \nCR$, is a Stieltjes function.
\end{enumerate}
\end{proposition}
A Stieltjes function $m$ admits the integral representation
\begin{equation}\label{eq:int_S}
m(z) = \gamma+\int_{0}^{+\infty} \frac{d\sigma(t)}{t-z}
\end{equation}
with $\gamma \ge 0$ and with a non-decreasing function $\sigma(t)$, such that
\begin{equation} \label{eq:int_S_gc}
\int_{0}^{+\infty} \dfrac{d\sigma(t)}{1+t} < + \infty.
\end{equation}
Clearly, for every Stieltjes function function $m$
\begin{equation*}
\re m(iy)\ge 0 \quad\mbox{for all}\quad y\in\nR_+.
\end{equation*}
\subsection{Reproducing kernel Hilbert spaces} \label{sect:RKHS}
In this section we review the basic properties of reproducing
kernel Hilbert spaces associated with Nevanlinna functions.
A function $m$ defined on
$\nCR$, is a Nevanlinna function if and only if
the kernel
\begin{equation} \label{eKm}
K_m(z,w) : = \dfrac{m(z) - \co{m(w)}}{z - \co{w}}, \quad z \neq
\co{w}, \quad z,w \in \nCR,
\end{equation}
is non-negative.
With the kernel $K_{m}(z,w)$ in \eqref{eKm} we associate the
reproducing kernel Hilbert space $\bigl(
\cH(m),\ahip_{\cH(m)} \bigr)$ defined as follows:
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item
the elements of $\cH(m)$ are holomorphic functions defined on
$\nCR$,
\item \label{rkfs}
for every $w \in \nCR$ the function $z \mapsto K_m(z,w), \
z\in\nCR$, belongs to $\cH(m)$,
\item
the set of all linear combinations of the functions in (\ref{rkfs})
is dense in the reproducing kernel Hilbert space $\bigl(
\cH(m),\ahip_{\cH(m)} \bigr)$,
\item
for every $f \in \cH(m)$ and every $w \in \nCR$ we have
\begin{equation*}
\bigl\la f(\cdot), K_m(\cdot,w) \bigr\ra_{\cH(m)} = f(w).
\end{equation*}
\end{enumerate}
The following theorem, presented in \cite[Proposition 5.3]{DM95} (see also~\cite[Theorem 2.2]{LT77}) assures
that each Nevanlinna function is a Weyl function of a closed simple
symmetric operator.
\begin{theorem} \label{mz}
Let $m$ be a Nevanlinna function, $m(z)\not\equiv 0$. The operator $S_{m}$ of
multiplication by the independent variable in the reproducing kernel
Hilbert space $\cH(m)$ is a closed simple symmetric operator
with defect numbers $(1,1)$. The operator $S_m$ is densely defined if and only if
\begin{equation*}
\lim_{y\uparrow +\infty}\frac{\im m(iy)}{y}=0 \quad \text{and} \quad
\lim_{y\uparrow +\infty}y\im m(iy)=+\infty.
\end{equation*}
In the graph notation, the adjoint of $S_m$ is given by
\begin{align*}
S_{m}^* &= \cspan\bigl\{ \{ K_{m}(\cdot,w), \co{w} K_{m}(\cdot,w)
\} : w\in\nCR \bigr\} \\
&= \bigl\{ \{f,g\} \in \cH(m)^2 :
g(z) - z f(z) = c_0 - c_1 m(z) \ \, \text{for some} \, \ c_0, c_1
\in \nC \bigr\}.
\end{align*}
The numbers $c_0,c_1 \in \nC$ in the last displayed equality are uniquely
determined by $f \in \dom(S_{m}^{*})$ and $\bigl(\nC,\Gamma_{m,0}, \Gamma_{m,1}\bigr)$ defined by
\begin{equation*}
\Gamma_{m,0}(f) := c_0, \qquad \Gamma_{m,1}(f) := c_1, \qquad f \in \dom(S_{m}^{*}),
\end{equation*}
is a boundary triple for $S_{m}^*$. The Weyl function of $S_m$
relative to the boundary triple $\bigl(\nC,\Gamma_{m,0}, \Gamma_{m,1}\bigr)$ of $S_{m}^*$ is
$m$.
\end{theorem}
The boundary triple $\bigl(\nC,\Gamma_{m,0}, \Gamma_{m,1}\bigr)$ is
called the {\em canonical boundary triple} for the operator $S_m^*$.
The following theorem is in some sense a converse of Theorem \ref{mz}.
\begin{theorem} \label{tFt}
Let $S$ be a closed simple symmetric operator in a Hilbert space
$\bigl(\cH,\ahip_{\cH}\bigr)$ with defect numbers $(1,1)$, let
$\bigl(\nC,\Gamma_0, \Gamma_1\bigr)$ be a boundary triple for $S$
and let $m$ be the corresponding Weyl function. Then the operator
$F_m: \cH \to \cH(m)$ defined by
\begin{equation}\label{eq:F_trans}
(F_m f)(z) = \bigl\la f, \psi(z^*) \bigl\ra_{\cH}, \quad f \in \cH,
\ \ \ z \in \nC\setminus\nR,
\end{equation}
is an isometry between $\bigl(\cH,\ahip_{\cH}\bigr)$ and $\bigl(
\cH(m),\ahip_{\cH(m)} \bigr)$ and
\begin{equation*}
F_m S = S_m F_m.
\end{equation*}
\end{theorem}
\subsection{Nevanlinna functions and M\"{o}bius transformations}
It is clear that a composition of two Nevanlinna functions is a Nevanlinna function. However, the reproducing kernel space corresponding to the composition can be significantly different from the reproducing kernel Hilbert space of the composed functions. In this subsection we show that the situation is different when a Nevanlinna function is composed with a linear fractional (or M\"{o}bius) transformations, which are itself Nevanlinna functions. Then the corresponding reproducing kernel Hilbert spaces are isomorphic.
It is straightforward to verify that the mapping
\begin{equation} \label{eq:MobHom}
\operatorname{GL}(2,\nC) \ni \left(\!\!\! \begin{array}{cc}
a & b \\ c & d \end{array} \!\!\!\right) \mapsto \mu(z) = \frac{a\rule{1pt}{0pt}z +b}{c\rule{1pt}{0pt}z+d}
\end{equation}
is a group homomorphism from the matrix group $\operatorname{GL}(2,\nC)$ of all $2\!\times\!2$ complex matrices with nonzero determinant onto the M\"{o}bius group.
The kernel of this homomorphism is the subgroup of all nonzero multiples of the identity matrix $I_2$.
We already encountered the subgroup $\operatorname{SL}(2,\nR)$ of $\operatorname{GL}(2,\nC)$. Below we will encounter its another subgroup $\operatorname{SL}^{\pm}(2,\nR)$ of all $2\!\times\!2$ real matrices with determinant $1$ or $-1$. The following lemma is standard. A short proof based on the homomorphism in \eqref{eq:MobHom} is included for completeness.
\begin{lemma} \label{lMobNev}
For a M\"{o}bius transformation $\mu$ the condition $\mu(z^*) = \mu(z)^*$ is equivalent to
\begin{equation} \label{eq:MobNev}
\mu(z) = \frac{a\rule{1pt}{0pt}z +b}{c\rule{1pt}{0pt}z+d} \quad \text{for some}
\quad \left(\!\!\! \begin{array}{cc}
a & b \\ c & d \end{array} \!\!\!\right) \in \operatorname{SL}^\pm(2,\nR).
\end{equation}
A M\"{o}bius transformation $\mu$ in \eqref{eq:MobNev} is a Nevanlinna function whenever $ad-bc >0$; its opposite $-\mu$ is a Nevanlinna function whenever $ad-bc <0$.
\end{lemma}
\begin{proof}
Let
\begin{equation*}
\mu(z) = \frac{a_1z + b_1}{c_1z+d_1} \qquad \text{where} \qquad \left(\!\!\! \begin{array}{cc}
a_1 & b_1 \\ c_1 & d_1 \end{array} \!\!\!\right) \in \operatorname{GL}(2,\nC).
\end{equation*}
Assume that $\mu(z^*) = \mu(z)^*$, that is $\mu(z^*)^* = \mu(z)$. Then the homomorphism in \eqref{eq:MobHom} implies
\begin{equation} \label{eq:conmult}
\delta \left(\!\!\! \begin{array}{cc}
a_1^* & b_1^* \\ c_1^* & d_1^* \end{array} \!\!\!\right) =\left(\!\!\! \begin{array}{cc}
a_1 & b_1 \\ c_1 & d_1 \end{array} \!\!\!\right) \quad \text{for some} \quad \delta \in \nC\setminus\{0\}.
\end{equation}
Calculating determinants of both sides yields,
\begin{equation*}
\delta^2 = e^{2 i \theta} \qquad \text{where} \qquad\theta = \operatorname{Arg}(a_1d_1-b_1c_1).
\end{equation*}
Since $\delta$ in \eqref{eq:conmult} is uniquely determined, we have
\begin{equation} \label{eq:choose-pm}
\delta = e^{i \theta} \qquad \text{or} \qquad \delta = -e^{i \theta}.
\end{equation}
It follows from \eqref{eq:conmult} and \eqref{eq:choose-pm} that
\[
\frac{\sqrt{\delta}}{\sqrt{|a_1d_1 - b_1c_1|}} \left(\!\!\! \begin{array}{cc}
a_1 & b_1 \\ c_1 & d_1 \end{array} \!\!\!\right) = \left(\!\!\!\begin{array}{cc}
a & b \\ c & d \end{array} \!\!\!\right) \in \operatorname{SL}^\pm(2,\nR)
\]
and whether $ad-bc = 1$ or $ad-bc = -1$ depends on the choice of the root in \eqref{eq:choose-pm}. The converse is straightforward.
The last claim follows from the identity
\begin{equation*}
\frac{\im \mu(z)}{\im z} = \frac{ad - bc}{|c\rule{1pt}{0pt}z+d|^2}. \qedhere
\end{equation*}
\end{proof}
For a M\"{o}bius transformation $\mu$ given in \eqref{eq:MobNev} a direct calculation confirms the following identity
\begin{equation} \label{eq:MobId}
\mu(z) - \mu(w^*) = \frac{ad-bc}{(c\rule{1pt}{0pt}z+d)(c\rule{1pt}{0pt}w^*+d)} (z - w^*) \quad \text{for all} \quad z, w \in \nCR,
\end{equation}
which will be used in the next proof.
\begin{theorem} \label{tV}
Let $\mu_1$ and $\mu_2$ be M\"{o}bius transformations given by
\begin{equation} \label{eq:MobNev12}
\mu_j(z) = \frac{a_jz +b_j}{c_j z+d_j} \quad \text{where} \quad \left(\!\!\! \begin{array}{cc}
a_j & b_j \\ c_j & d_j \end{array} \!\!\!\right) \in \operatorname{SL}^\pm(2,\nR), \quad j \in \{1,2\}.
\end{equation}
For $j \in \{1,2\}$ set $\epsilon_j := a_j d_j-b_j c_j \in \{-1,1\}$.
Let $m$ be a complex function defined on $\nCR$ and set
\begin{equation} \label{eq:MobNev-whm}
\wh{m}(z) = \epsilon_1 \epsilon_2 \, \mu_2\bigl(m(\mu_1(z))\bigr), \qquad z \in \nC\setminus\nR.
\end{equation}
Then $m$ is a Nevanlinna function if and only if $\wh{m}$ is a Nevanlinna function.
If $m$ is a Nevanlinna function the mapping $V: \cH(\wh{m}) \to \cH\bigl(m\bigr)$ defined by
\begin{equation*}
(Vf)(z) : = \frac{c_1 \mu_1(z) - a_1}{c_2 m(\mu_1(z))+d_2} \, f\bigl(\mu_1(z)\bigr), \ \ \ f \in \cH(m), \ \ \ z \in \nCR,
\end{equation*}
is an isomorphism between the Hilbert spaces $\cH(m)$ and $\cH(\wh{m})$.
\end{theorem}
\begin{proof}
First notice that the M\"{o}bius transformations in
\eqref{eq:MobNev12} have inverses which are also M\"{o}bius transformations of the same kind. For example
\[
\mu_1^{-1}(z) = \frac{d_1 z - b_1}{-c_1 z + a_1} \quad \text{and} \quad \left(\!\!\! \begin{array}{cc}
d_1 & -b_1 \\ -c_1 & a_1 \end{array} \!\!\!\right) \in \operatorname{SL}^\pm(2,\nR).
\]
Now the first statement follows from Lemma~\ref{lMobNev} and the fact that a composition of Nevanlinna functions is a Nevanlinna function.
Assume that $m$ is a Nevanilnna function. In the following identities we assume that $z, w \in \nCR$ and we use \eqref{eq:MobId}, first for $\mu_2$, then for $\mu_1^{-1}$. For convenience we set $z_1 = \mu_1(z)$ and $w_1 = \mu_1(w)$ and calculate:
\begin{align*}
K_{\wh{m}}(z,w) & = \epsilon_1 \epsilon_2 \,\frac{\mu_2\bigl(m(\mu_1(z))\bigr) - \mu_2\bigl(m(\mu_1(w^*))\bigr)}{z-w^*} \\
& = \frac{\epsilon_1 \epsilon_2\epsilon_2}{(c_2 m(z_1)+d_2)(c_2m(w_1)^*+d_2)} \ \frac{m(z_1) - m(w_1^*)}{\mu_1^{-1}(z_1)-\mu_1^{-1}(w_1^*)} \\
& = \frac{(-c_1 z_1 + a_1)(-c_1 w_1^*+ a_1)}{(c_2 m(z_1)+d_2)(c_2m(w_1)^*+d_2)} \
\frac{m(z_1) - m(w_1^*)}{z_1 -w_1^*} \\
& = \frac{(c_1 \mu_1(z) - a_1)(c_1 \mu_1(w)^* - a_1)}{(c_2 m(\mu_1(z))+d_2)(c_2 m(\mu_1(w^*))+d_2)} \
K_m\bigl(\mu_1(z),\mu_1(w)\bigr).
\end{align*}
Consequently, substituting $w$ with $\mu_1^{-1}(w)$, we get
\begin{equation*}
\frac{c_1 \mu_1(z) - a_1}{c_2 m(\mu_1(z))+d_2} \
K_m\bigl(\mu_1(z),w\bigr) = \frac{c_2 m(w^*)+d_2}{c_1 w^* - a_1} K_{\wh{m}}(z,\mu_1^{-1}(w)),
\end{equation*}
that is
\begin{equation} \label{eqUl}
V\bigl(K_m(\cdot,w)\bigr) = \frac{c_2 m(w^*)+d_2}{c_1 w^* - a_1} K_{\wh{m}}(\cdot,\mu_1^{-1}(w)) \quad \text{for all} \quad w \in \nCR.
\end{equation}
Therefore, for arbitrary $v, w \in \nCR$ we have
\begin{align*}
\bigl\la V K_{m}(\cdot,v), V K_{m}(\cdot,w) \bigr\ra_{\cH(\wh{m})} & \\
& \hspace*{-4cm}
= \frac{c_2 m(v^*)+d_2}{c_1 v^* - a_1} \
\frac{c_2 m(w)+d_2}{c_1 w - a_1} \
\Bigl\la K_{\wh{m}}(\cdot,\mu_1^{-1}(v)), K_{\wh{m}}(\cdot,\mu_1^{-1}(w)) \Bigr\ra_{\cH(\wh{m})} \\
&\hspace*{-4cm} = \frac{c_2 m(v^*)+d_2}{c_1 v^* - a_1} \
\frac{c_2 m(w)+d_2}{c_1 w - a_1} \
K_{\wh{m}} \bigl(\mu_1^{-1}(w),\mu_1^{-1}(v) \bigr) \\
& \hspace*{-4cm} = K_m(w,v) \\
&\hspace*{-4cm} = \bigl\la K_m(\cdot,v), K_m(\cdot,w) \bigr\ra_{\cH(m)}.
\end{align*}
Thus, for arbitrary $v, w \in \nCR$ we have
\begin{equation} \label{eqUlu}
\bigl\la V K_{m}(\cdot,v), V K_{m}(\cdot,w) \bigr\ra_{\cH(\wh{m})} = \bigl\la K_m(\cdot,v), K_m(\cdot,w) \bigr\ra_{\cH(m)}.
\end{equation}
Set
\begin{align*}
\cL(m) & := \lspan \bigl\{ K_m(\cdot,w) \, : \, w \in \nCR \bigr\}, \\
\cL(\wh{m}) & := \lspan \bigl\{ K_{\wh{m}}(\cdot,w) \, : \, w \in
\nCR \bigr\}.
\end{align*}
As both functions $m$ and $\wh{m}$ are Nevanlinna functions, $\cL(m)$ is dense in the Hilbert space $\cH(m)$ and $\cL(\wh{m})$ is dense in the Hilbert space $\cH(\wh{m})$.
Since $V$ is linear, it follows from \eqref{eqUl} and \eqref{eqUlu}
that the restriction $V\bigl|_{\cL(m)}\bigr.$ is a bijection and an
isomorphism of pre-Hilbert spaces $\cL(m)$ and $\cL(\wh{m})$.
Denote by $V_1$ the extension by continuity of $V\bigl|_{\cL(m)}
\bigr.$ to $\cH(m)$. Then $V_1$ is an isomorphism between
$\cH(m)$ and $\cH(\wh{m})$. Let $g \in \cH(\wh{m})$ be arbitrary
and let $f \in \cH(m)$ be such that $V_1 f = g$. For $z \in
\nCR$ we calculate:
\begin{align*}
g(z) & = \bigl\la g(\cdot), K_{\wh{m}}(\cdot,z) \bigr\ra_{\cH(\wh{m})} \\
& = \Bigl\la
(V_1f)(\cdot), \frac{c_1 \mu_1(z^*) - a_1}{c_2 m(\mu_1(z^*))+d_2} V K_m(\cdot,\mu_1(z)) \Bigr\ra_{\cH(\wh{m})} \\
& = \frac{c_1 \mu_1(z) - a_1}{c_2 m(\mu_1(z))+d_2} \, \Bigl\la f(\cdot), K_m(\cdot,\mu_1(z)) \Bigr\ra_{\cH(m)} \\
& = \frac{c_1 \mu_1(z) - a_1}{c_2 m(\mu_1(z))+d_2} \, f\bigl( \mu_1(z) \bigr).
\end{align*}
Thus $V = V_1$.
\end{proof}
\subsection{Stieltjes functions and M\"{o}bius transformations}
Assigning to a meromorphic function $m$ the value $\infty$ at the poles we can consider $m$ to be defined on its whole domain of meromorphy. In particular, if the domain of meromorphy of a Nevanlinna function $m$ includes $\nR_-$, then the image $m(\nR_-)$ of $\nR_-$ under $m$ is contained in the one point compactification $\overline{\nR} = \nR\cup \{\infty\}$ of $\nR$. In the next lemma by $\overline{m(\nR_-)}$ we denote the closure of $m(\nR_-)$ in $\overline{\nR}$.
\begin{lemma} \label{l:Ms}
Let $m$ be a Nevanlinna function. The following statements are equivalent.
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{i:Ms-1}
There exists a
M\"{o}bius transformation $\mu$ of the form~\eqref{eq:MobNev} such that $\mu\circ m$ is a Stieltjes function.
\item \label{i:Ms-2}
There exists a M\"{o}bius transformation $\mu$ of the form~\eqref{eq:MobNev} such that $\mu\circ m$ is an inverse Stieltjes function.
\item \label{i:Ms-3}
The function $m$ is meromorphic on $\nC\setminus[0,+\infty)$ and
$\overline{m(\nR_-)}$ is a proper subset of $\overline{\nR}$.
\end{enumerate}
\end{lemma}
\begin{proof}
The equivalence of (\ref{i:Ms-1}) and (\ref{i:Ms-2}) follows from Proposition~\ref{p:SiS}. The implication (\ref{i:Ms-2}) $\Rightarrow$ (\ref{i:Ms-3}) is clear. To complete the proof, assume (\ref{i:Ms-3}). Then the complement of $\overline{m(\nR_-)}$ in $\overline{\nR}$ contains a finite open interval. By shifting $m$ we can assume that that interval is $(0,2/c)$ with $c > 0$. Set $\mu(z) = -z/(c z-1)$. Then $\mu\circ m$ is a Nevanlinna function which is holomorphic on $\nC\setminus[0,+\infty)$. Since $\mu(x) < 0$ whenever $x \in \nR \setminus[0,2/c]$ and $m(\nR_-) \subset \nR \setminus[0,2/c]$ we conclude that $(\mu\circ m)(x) < 0$ for all $x \in \nR_-$. Thus $\mu\circ m$ is an inverse Stieltjes function. This proves (\ref{i:Ms-3}) $\Rightarrow$ (\ref{i:Ms-2}).
\end{proof}
Denote by $\SM$ the set of Nevanlinna functions that satisfy the equivalent conditions in Lemma~\ref{l:Ms}. The set $\SM$ appears in \cite{BHSWW13} as the union of Nevanlinna functions of type I, II, III and V$^{\prime}$, see \cite[Definition~2.4]{BHSWW13} and also \cite[Corollary~5.6]{BHSWW13}.
Notice that a function $m \in \SM$ can have at most one pole in $\nR_-$ and that the following limits exist
\[
m(-\infty):= \!\!\lim_{x\downarrow -\infty} m(x) \in \{-\infty\}\cup\nR \quad \text{and} \quad m(0-):=\lim_{x\uparrow 0} m(x) \in \nR \cup\{+\infty\},
\]
with at least one of them being finite. In addition, $m(-\infty) \leq m(0-)$ if and only if $m$ is holomorphic on $\nC\setminus[0,+\infty)$ and $m(-\infty) > m(0-)$ if and only if $m$ has a pole in $\nR_-$.
\begin{remark} \label{r:MsInfZer}
Let $m \in \SM$. It follows from Lemma~S1.3.1 and Remark~5.1 in \cite{KaKr74} that $m(-\infty) \in \nR$ if and only if $\lim_{y\uparrow +\infty} m(i y)$ exists as a real number; in this case $\lim_{y\uparrow +\infty} m(i y)=m(-\infty)$. Also, $m(-\infty) = -\infty$ if and only if $\lim_{y\uparrow +\infty} m(i y) = \infty$.
Since $m \in\SM$ if and only if the function $z \mapsto -m(1/z)$, $z\in \nCR$, is in $\SM$, we similarly have that $m(0-) \in \nR$ if and only if $\lim_{y\downarrow 0} m(i y)$ exists as a real number and in this case $\lim_{y\downarrow 0} m(i y)=m(0-)$. Also, $m(0-) = +\infty$ if and only if $\lim_{y\downarrow 0} m(i y) = \infty$.
\end{remark}
\subsection{Two asymptotic classes of Nevanlinna functions} \label{sSecAsy}
For functions $f$ and $g$ defined on $\nR_+$ the expression
\begin{alignat*}{2}
f(t) \sim g(t) \quad &\text{as} \quad t \to +\infty & \qquad
\bigl( f(t) \sim & g(t) \quad \text{as} \quad t \to 0+ \bigr)
\intertext{
means
}
\lim_{t\uparrow +\infty} \frac{f(t)}{g(t)} &= 1 & \qquad
\Bigl( \lim_{t\downarrow 0+} \frac{f(t)}{g(t)} &= 1, \quad \text{respectively} \Bigr).
\end{alignat*}
We define the set of functions $\cA_\infty$ as follows: a Nevanlinna function $m$ belongs to $\cA_\infty$ if and only if there exist $\alpha\in(0,1)$, $C > 0$ and a M\"{o}bius transformation $\mu$ of the form~\eqref{eq:MobNev} such that $\mu \circ m$ is a Stieltjes function and for all $z \in \nCR$ we have
\begin{equation}\label{eq:mAsymInf-defAInf}
(\mu \circ m)(r z) \sim \frac{C}{ (- r z)^{\alpha} } \qquad \text{as} \qquad r \to +\infty.
\end{equation}
Similarly, we define the set of functions $\cA_0$ as follows: a Nevanlinna function $m$ belongs to $\cA_0$ if and only if there exist $\alpha \in (0,1)$, $C > 0$ and a M\"{o}bius transformation $\mu$ of the form~\eqref{eq:MobNev} such that $\mu \circ m$ is an inverse Stieltjes function and for all $z \in \nCR$ we have
\begin{equation} \label{eq:mAsymZer-defAZer}
(\mu\circ m)( r z) \sim - C (- r z)^{\alpha} \qquad \text{as} \qquad r \to 0\!+.
\end{equation}
In the notation of Theorem~{\rm\ref{tV}}, the following proposition holds.
\begin{proposition} \label{prop:AInf-AZer}
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{i:AInf-AZer-1}
Assume that $\mu_1(z) = z$ and let $\mu_2$ be a M\"{o}bius transformation as in \eqref{eq:MobNev12}. Then, $m \in \cA_\infty$ {\rm (}$m \in \cA_0${\rm )} if and only if $\wh{m} \in \cA_\infty$ {\rm (}$\wh{m} \in \cA_0$, respectively{\rm )}.
\item \label{i:AInf-AZer-2}
Let $\mu_1(z) = 1/z$ and let $\mu_2$ be a M\"{o}bius transformation as in \eqref{eq:MobNev12}. Then, $m \in \cA_\infty$ {\rm (}$m \in \cA_0${\rm )} if and only if $\wh{m} \in \cA_0$ {\rm (}$\wh{m} \in \cA_\infty$, respectively{\rm )}.
\end{enumerate}
\end{proposition}
\begin{proof}
We will prove one implication from each (\ref{i:AInf-AZer-1}) and (\ref{i:AInf-AZer-2}). The other implications are proved similarly.
Let $m \in \cA_\infty$. Then there exists a M\"{o}bius transformation $\mu$ of the form~\eqref{eq:MobNev} such that $\mu \circ m$ is a Stieltjes function and \eqref{eq:mAsymInf-defAInf} holds for all $z \in \nCR$.
Recall that, according to \eqref{eq:MobNev-whm}, in (\ref{i:AInf-AZer-1}) we have $\wh{m} = \epsilon_2 \mu_2\circ m$. Set $\nu(z) = \epsilon_2 z$, $z \in \nC$. Since $\mu\circ\mu_2^{-1}\circ \nu \circ \wh{m} = \mu\circ m$ and $\mu\circ\mu_2^{-1}\circ \nu$ is a a M\"{o}bius transformation of the form~\eqref{eq:MobNev}, we have $\wh{m} \in \cA_\infty$. This proves the implication $m \in \cA_\infty \Rightarrow \wh{m} \in \cA_\infty$ from (\ref{i:AInf-AZer-1}).
Again, according to \eqref{eq:MobNev-whm}, in (\ref{i:AInf-AZer-2}) we have $\wh{m} = -\epsilon_2 \mu_2\circ m \circ \mu_1$. Set $\nu(z) = -\epsilon_2 z$, $z \in \nC$. Clearly,
\[
-\mu \circ \mu_2^{-1} \circ \nu \circ \wh{m} = - \mu \circ m \circ \mu_1. \]
Since $\mu \circ m$ is a Stieltjes function that satisfies \eqref{eq:mAsymInf-defAInf} for all $z \in \nCR$, by Proposition~\ref{p:SiS} the function
$- \mu \circ m \circ \mu_1$ is an inverse Stieltjes function that satisfies \eqref{eq:mAsymZer-defAZer} for all $z \in \nCR$.
As $-\mu \circ \mu_2^{-1} \circ \nu$ is a M\"{o}bius transformation of the form~\eqref{eq:MobNev}, we have $\wh{m} \in \cA_0$. This proves the implication $m \in \cA_\infty \Rightarrow \wh{m} \in \cA_0$ from (\ref{i:AInf-AZer-2}).
\end{proof}
In the next proposition we characterize the functions in $\cA_\infty$ and $\cA_0$ by their asymptotic behavior.
\begin{proposition} \label{prop:0-inf}
The following equivalences hold.
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{i:P-0-inf-1}
$m \in \cA_\infty$ if and only if $m \in \SM$ and there exist $C_0 \in \nR\!\setminus\!\{0\}$ and $\alpha_0 \in (-1,0)\cup(0,1)$ such that for all $z \in \nCR$ we have
\begin{equation}\label{eq:mAsymInf-noC}
m( r z) \sim C_0 (- r z)^{\alpha_0} \quad \text{as} \quad r \to +\infty,
\end{equation}
or, corresponding to $\alpha_0 =0$, there exist $C_0\in\nR\!\setminus\!\{0\}$, $C_1 > 0$ and $\alpha_1 \in (-1,0)$ such that for all $z \in \nCR$ we have
\begin{equation}\label{eq:mAsymInf-C}
m( r z) - C_0 \sim C_1( - r z )^{\alpha_1} \quad \text{as} \quad r \to +\infty.
\end{equation}
\item \label{i:P-0-inf-2}
$m\in \cA_0$ if and only if $m \in \SM$ and there exist $C_0 \in \nR\!\setminus\!\{0\}$ and $\alpha_0 \in (-1,0)\cup(0,1)$ such that for all $z \in \nCR$ we have
\begin{equation*}
m( r z) \sim C_0 (- r z)^{\alpha_0} \quad \text{as} \quad
r \to 0\!+,
\end{equation*}
or, corresponding to $\alpha_0 =0$, there exist $C_0, \in\nR\!\setminus\!\{0\}$, $C_1 > 0$ and $\alpha_1 \in (0,1)$ such that for all $z \in \nCR$ we have
\begin{equation*}
m( r z) - C_0 \sim -C_1( - r z )^{\alpha_1} \quad \text{as} \quad r \to 0\!+.
\end{equation*}
\end{enumerate}
\end{proposition}
\begin{proof}
We first prove the direct implication in (\ref{i:P-0-inf-1}). Clearly $\cA_\infty \subset \SM$. Assume that $m \in \cA_\infty$ is such that \eqref{eq:mAsymInf-defAInf} holds and let $\mu^{-1}(z) = (az+b)/(cz+d)$ with $ad - bc = 1$. We distinguish two cases.
If $d = 0$, then $\mu^{-1}(z) = a/c + b/(cz)$, and therefore, for all $z \in \nCR$,
\[
m( r z) \sim \frac{b}{c C} (-r z)^{\alpha} \quad \text{as} \quad r \to +\infty.
\]
Thus \eqref{eq:mAsymInf-noC} holds.
If $d \neq 0$, then $\mu^{-1}(z) = b/d + z/d^2 + O(z^2)$ as $z\to 0$, and therefore, for all $z \in \nCR$
\[
m( r z) - \frac{b}{d} \sim \frac{C}{d^2}(-r z)^{-\alpha} \quad \text{as} \quad r \to +\infty.
\]
Consequently, \eqref{eq:mAsymInf-noC} holds if $b=0$ and \eqref{eq:mAsymInf-C} holds if $b \neq 0$. This proves the direct implication in (\ref{i:P-0-inf-1}).
To prove the converse let $m \in \SM$. First assume that $m$ satisfies \eqref{eq:mAsymInf-noC} with $\alpha_0 \in (-1,0)$ and $C_0 > 0$. By Remark~\ref{r:MsInfZer} we have $m(-\infty) = 0$. If $m$ is holomorphic on $\nC\setminus[0,+\infty)$, then it is a Stieltjes function; so $m \in \cA_\infty$ in this case. If $m$ has a pole in $\nR_-$, then, since $m \in \SM$, $m(0-) < 0$ and $m(x) \not\in\bigl(m(0-),0\bigr)$ for all $x\in\nR_-$. Setting $c = -m(0-)/2$ and $\mu(z) = z/(c z+1)$, as in the proof of Lemma~\ref{l:Ms} we have that $\mu\circ m$ is a Stieltjes function. Since $\mu(z) = z + O(z^2)$ as $z \to 0$, we have that
\[
(\mu\circ m)(r z) \sim C_0 (- r z)^{\alpha_0} \qquad \text{as} \qquad r \to 0.
\]
Hence $m \in \cA_\infty$.
Second, assume that $m$ satisfies \eqref{eq:mAsymInf-noC} with $\alpha_0 \in (0,1)$ and $C_0 < 0$. Then, the already proven part implies $m^{\!\top} \in \cA_\infty$ and Proposition~\ref{prop:AInf-AZer} yields $m \in \cA_\infty$.
Third, assume that $m$ satisfies \eqref{eq:mAsymInf-C}. Then, with $\mu(z) = z - C_0$, we have $\mu\circ m \in \cA_\infty$, by the second part of this proof. Now Proposition~\ref{prop:AInf-AZer} implies that $m \in \cA_\infty$.
The proof of (\ref{i:P-0-inf-2}) is similar.
\end{proof}
\subsection{Nonnegative operators in Kre\u{\i}n spaces} \label{ss:non-Ks}
Let $\bigl(\cK,\kip\bigr)$ be a Kre\u{\i}n space and let $A$ be a densely defined operator in $\cK$. The {\em adjoint} $A^{[*]}$ of $A$ with respect to $\kip$ is defined analogously as in a Hilbert space. In fact, if $J$ is a fundamental symmetry on $\bigl(\cK,\kip\bigr)$ and $\ahip$ is the corresponding Hilbert space inner product, then $A^{[*]} = JA^{\la*\ra}J$. In analogy with definitions in a Hilbert space, $A$ is {\em symmetric in} $\bigl(\cK,\kip\bigr)$ if $A^{[*]}$ is an extension of $A$ and $A$ is {\em self-adjoint in} $\bigl(\cK,\kip\bigr)$ if $A = A^{[*]}$.
Since our main interest in this paper is similarity of a self-adjoint operator in a Kre\u{\i}n space to a self-adjoint operator in a Hilbert space, in the next proposition we recall a known characterization of similarity. This characterization is proved in \cite[Theorem~1, Section~2]{McEn82} for bounded self-adjoint operators in a Kre\u{\i}n space. In \cite[Proposition~2.2]{KuTr11} an equivalent statement in terms of $\cC$-stable symmetries is given and Phillips theorem \cite[Chapter 2, Corollary 5.20]{AI} is cited for a proof. Below we give a simple direct proof.
\begin{proposition} \label{p:Sim-iff-fd}
A self-adjoint operator $A$ in a Kre\u{\i}n space $(\cK,\kip)$ is similar to a self-adjoint operator in the Hilbert space $(\cK,\ahip)$ if and only if $A$ is fundamentally reducible in $(\cK,\kip)$.
\end{proposition}
\begin{proof}
Let $A$ be a self-adjoint operator in a Kre\u{\i}n space $(\cK,\kip)$ and assume that $A$ is similar to a self-adjoint operator in the Hilbert space $(\cK,\ahip)$. That is, $A$ is self-adjoint in a Hilbert space $(\cK,\hip)$ whose norm is equivalent to that of $(\cK,\ahip)$. The equivalence of the norms of $(\cK,\ahip)$ and $(\cK,\hip)$ and the identity $\la x,y \ra = [Jx,y]$, for all $x,y \in \cK$, imply the existence of a bounded and boundedly invertible operator $G:\cK \to \cK$ such that $(G x,y)= [x,y]$ for all $x,y \in \cK$. The last identity yields that $G$ is self-adjoint in $\hip$ and the operator $\sgn G$ is a fundamental symmetry on $(\cK,\kip)$. As $A$ is self-adjoint in both $\kip$ and $\hip$, $A$ commutes with $G$. Hence, $A$ commutes with the fundamental symmetry $\sgn G$ on $(\cK,\kip)$; proving that $A$ is fundamentally reducible. The converse is clear.
\end{proof}
A densely defined operator $A$ is {\em nonnegative in} $\bigl(\cK,\kip\bigr)$ if $[Af,f] \geq 0$ for all $f \in \dom(A)$. A nonnegative self-adjoint operator in a Kre\u{\i}n space can have an empty resolvent set; a specific example is given in \cite[1.2]{La71} and \cite[Example~VII.1.5]{B74}. A modification of this example leads to a positive self-adjoint operator in a Kre\u{\i}n space with an empty resolvent set, see \cite[Example~3.4]{Ko14}.
However, if $A$ is a nonnegative self-adjoint operator in $\bigl(\cK,\kip\bigr)$ and $\rho(A) \neq \emptyset$, then the spectrum of $A$ is real and $A$ has a spectral function $E$, see \cite[Theorem~II.3.1]{Le}. The domain of $E$ is the ring $\cI$ which consists of finite unions of (bounded) intervals whose endpoints are nonzero real numbers and their complements in $\overline{\nR} = \nR\cup\{\infty\}$. The values of $E$ are bounded operators on $\cK$ with the following properties. For all $\Delta, \Delta_1, \Delta_2 \in \cI$ we have
\begin{enumerate}
\item[(E1)]
$E(\emptyset) = 0$, $E(\overline{\nR}) = I$,
\item[(E2)]
$E(\Delta) = E(\Delta)^{[*]}$
\item[(E3)]
$E(\Delta_1 \cap \Delta_2) = E(\Delta_1) E(\Delta_2)$,
\item[(E4)]
$E(\Delta_1 \cup \Delta_2) = E(\Delta_1) + E(\Delta_2)$ whenever $\Delta_1 \cap \Delta_2 = \emptyset$,
\item[(E5)]
the space $\bigl(E(\Delta)\cK, \pm\kip \bigr)$ is a Hilbert space whenever $\Delta \subset \nR_\pm$,
\item[(E6)]
$E(\Delta)$ is in the double commutant of the resolvent of $A$,
\item[(E7)]
for a bounded $\Delta$ we have $E(\Delta)\cK \subseteq \dom(A)$ and the restriction of $A$ to $E(\Delta)\cK$ is a bounded operator whose spectrum is contained in $\overline{\Delta}$.
\end{enumerate}
For $\lambda \in \nR\setminus\{0\}$, it follows from (E5) that in a neighborhood of $\lambda$ the spectral function $E$ behaves as a spectral function of a self-adjoint operator in a Hilbert space. In particular, with $\lambda_1, \lambda_2 \in \nR\setminus\{0\}$ such that $\lambda_1 < \lambda < \lambda_2$ the limits
\begin{equation} \label{eq:sotl}
\lim_{t\uparrow \lambda}E\bigl([\lambda_1,t]\bigr) \qquad \text{and} \qquad \lim_{t\downarrow \lambda}E\bigl([t,\lambda_2]\bigr)
\end{equation}
exist in the strong operator topology. The existence of these limits is a consequence of the following property of nonzero real numbers: For $\lambda \in \nR\setminus\{0\}$ there exists a $\Delta \in \cI$ such that $\lambda$ is in the interior of $\Delta$ and $\bigl(E(\Delta)\cK, (\sgn \lambda)\kip \bigr)$ is a Hilbert space.
A possible absence of the just mentioned property for $0$ or $\infty$ is a motivation for the following definition. The point $0$ ($\infty$) is said to be a {\em critical point of} $A$ if $\kip$ is indefinite on $E(\Delta)\cK$ for every $\Delta \in \cI$ such that $0 \in \Delta$ ($\infty \in \Delta$, respectively).
However, even if $0$ or $\infty$ is a critical point the limits analogous to \eqref{eq:sotl} may exist. If $0$ is a critical point of $A$ and the limits in \eqref{eq:sotl} exist with $\lambda = 0$, then $0$ is called a {\em regular critical point of} $A$. If $\infty$ is a critical point of $A$ and the limits
\begin{equation*}
\lim_{t\uparrow +\infty}E\bigl([\lambda_1,t]\bigr) \qquad \text{and} \qquad \lim_{t\downarrow -\infty}E\bigl([t,\lambda_2]\bigr)
\end{equation*}
exist in the strong operator topology for some $\lambda_1, \lambda_2 \in \nR\setminus\{0\}$, then $\infty$ is called {\em regular critical point of} $A$. A critical point of $A$ which is not regular is called {\em singular critical point of} $A$. The set of all singular critical points of $A$ is denoted by $c_s(A)$.
The following proposition is a part of folklore in this setting. For a bounded $A$ it was proved in \cite{Bay78}, and, in a more general form in \cite{McEn82}. For unbounded $A$ with a bounded inverse it appears in \cite{Cur} where \cite{Bay78} was cited for a proof by taking an inverse. For completeness we include a simple proof.
\begin{proposition} \label{p:fd-iff-rcp}
Let $A$ be a nonnegative operator in a Kre\u{\i}n space $\bigl(\cK,\kip\bigr)$. Then $A$ is fundamentally reducible in $\bigl(\cK,\kip\bigr)$ if and only if $\rho(A) \neq \emptyset$, $\ker(A) = \ker(A^2)$ and $\infty, 0 \not\in c_s(A)$.
\end{proposition}
\begin{proof}
Since a fundamentally reducible operator is similar to a self-adjoint operator in a Hilbert space, the direct implication follows. For the converse, notice that $\infty, 0 \not\in c_s(A)$ implies the existence of the operators $E(\nR_-):=\lim_{t\downarrow-\infty} E([t,1/t])$ and $E(\nR_+):=\lim_{t\uparrow +\infty} E([1/t,t])$. It follows from (E2), (E3), (E5) and (E6) that $E(\nR_\pm) = E(\nR_\pm)^2 = E(\nR_\pm)^{[*]}$, $\bigl(E(\nR_\pm)\cK, \pm\kip \bigr)$ is a Hilbert space and $E(\nR_\pm)$ is in the double commutant of the resolvent of $A$. Since $\ker(A) = \ker(A^2)$, Propositions~5.1 and~5.6 in~\cite{La} yield
\[
\cK = E(\nR_+)\cK [\dot{+}] \ker(A) [\dot{+}] E(\nR_-)\cK,
\]
direct and orthogonal sum in $\bigl(\cK,\kip\bigr)$. Since $\bigl(E(\nR_-)\cK [\dot{+}]E(\nR_+)\cK,\kip\bigr)$ is a Kre\u{\i}n space, \cite[Theorem~5.2]{La} implies that $\bigl(\ker(A),\kip\bigr)$ is a Kre\u{\i}n space. Let $\ker(A) = \cK_{+}^0[\dot{+}]\cK_{-}^0$ be its fundamental decomposition. Then
\[
\cK = \bigl(E(\nR_+)\cK [\dot{+}] \cK_{+}^0\bigr)[\dot{+}]\bigl(\cK_{-}^0 [\dot{+}] E(\nR_-)\cK\bigr)
\]
is a fundamental decomposition of $\cK$ which reduces $A$.
\end{proof}
In Subsection~\ref{ss:RegCPs} we essentially use the following resolvent criterion of K.~Veseli\'{c} \cite{Ves} for $\infty\not\in c_s(A)$. We state a special case of this criterion as it has appeared in~\cite[Corollary 1.6]{Jo}.
\begin{theorem} \label{Veselic}
Let $A$ be a nonnegative self-adjoint operator with a nonempty resolvent set in a Kre\u{\i}n space $({\cH}, [\,\cdot\,,\,\cdot\, ])$. Then:
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{iVeselic-1}
$\infty\not\in c_s(A)$ if and only if the operators
\[
\int_{-i\eta}^{-i}+ \int_{i}^{i\eta}(A-{z})^{-1}d{z}
\]
are uniformly bounded for $\eta\in(1,\infty)$.
\item \label{iVeselic-2}
$0\not\in c_s(A)$ if and only if $\ker(A)=\ker(A^2)$ and the operators
\[
\int_{-i}^{-i\varepsilon}+ \int_{i\varepsilon}^{i}(A-{z})^{-1}d{z}
\]
are uniformly bounded for $\varepsilon\in(0,1)$.
\end{enumerate}
\end{theorem}
\section{$B$-properties of Nevanlinna functions.} \label{Sec-B-prop}
\subsection{The definition and basic properties.} \label{SubSec-B-prop-def}
\begin{definition} \label{defBinf}
Let $m$ be a Nevanlinna function. Set
\[
w_m(y) := \frac{1}{\im m(i\,\! y)}, \qquad y > 0.
\]
Let $G_{m,\infty}$ be the mapping defined for all $h \in \cH(m)$
by
\[
\bigl(G_{m,\infty} h\bigr)(y) := h(i\,\!y), \ \ \ y > 1.
\]
A Nevanlinna function $m$ is said to have a $B_\infty$-{\em
property} if $\ran\bigl(G_{m,\infty}\bigr) \subset L_{w_m}^2\!(1,\infty)$ and the operator $G_{m,\infty}: \cH(m) \to L_{w_m}^2\!(1,\infty)$ is bounded.
Let $G_{m,0}$ be the mapping defined for all $h \in
\cH(m)$ by
\[
\bigl(G_{m,0} h\bigr)(y) := h(i\,\!y), \ \ \ 0 < y < 1.
\]
A Nevanlinna function $m$ is said to have a $B_0$-{\em property}
if $\ran\bigl(G_{m,0}\bigr) \subset L_{w_m}^2\!(0,1)$ and the operator $G_{m,0}: \cH(m) \to L_{w_m}^2\!(0,1)$ is bounded.
\end{definition}
The following proposition is a straightforward consequence of
Theorem~\ref{tFt} and the above definition.
\begin{proposition} \label{pGF}
Let $S$ be a closed simple symmetric operator with defect numbers $(1,1)$ in a Hilbert space $\bigl(\cH,\ahip_{\cH}\bigr)$. Let
$\bigl(\nC,\Gamma_0, \Gamma_1\bigr)$ be a boundary triple for $S^*$
and let $m$ be the corresponding Weyl function. Then $m$ has
$B_\infty$-property {\rm (}$B_0$-property{\rm )} if and only if $G_{m,\infty} F_m$ is a bounded mapping from $\cH$ into $L_{w_m}^2\!(1,\infty)$ {\rm (}$L_{w_m}^2\!(0,1)$, respectively{\rm )}. Moreover, if $m$ has a $B_\infty$-property {\rm (}$B_0$-property{\rm )}, then $\|G_{m,\infty}\| = \|G_{m,\infty} F_m\|$ {\rm (}$\|G_{m,0}\| = \|G_{m,0} F_m\|$, respectively{\rm )}.
\end{proposition}
In the next lemma, for a Nevanlinna function $m$, we introduce the operator $G_{m,\infty}^-$.
\begin{lemma}\label{prop:4.5}
Let $m$ be a Nevanlinna function. Then:
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{p45i}
$m$ has $B_\infty$-property if and only if
\[
\bigl(G_{m,\infty}^- h\bigr)(y) := h(-i\,\!y), \ \ \ y > 1.
\]
is a bounded operator from $\cH(m)$ into $L_{w_m}^2\!(1,\infty)$.
\item \label{p45ii}
$m$ has $B_0$-property if and only if
\[
\bigl(G_{m,0}^- h\bigr)(y) := h(-i\,\!y), \ \ \ y \in(0,1).
\]
is a bounded operator from $\cH(m)$ into $L_{w_m}^2\!(0,1)$.
\end{enumerate}
\end{lemma}
\begin{proof}
For a Nevanlinna function $m$ define the anti-linear mapping
$W: \cH({m}) \to \cH({m})$ by
\[
(Wf)(z) = f(z^*)^*, \quad z \in \nC\setminus\nR, \quad f \in \cH({m}).
\]
Then $W$ is a bijection and
\begin{equation} \label{eq:Wh}
\bigl\la Wf,Wg \bigr\ra = \la f,g \ra^* \qquad \text{for all} \qquad f,g \in \cH({m}).
\end{equation}
Further, for every $\omega\in\nCR$ and $y > 0$ we have
\begin{align*}
\bigl(G_{m,\infty}^- K_\omega\bigr)(y)&=K_\omega(-iy)=\frac{m(-iy)-m(\omega)^*}{-iy-\omega^*}\\
&=K_{\omega^*}(iy)^*=\left(\bigl(G_{m,\infty} K_{\omega^*}\bigr)(y)\right)^*=\left(\bigl(G_{m,\infty}W K_\omega\bigr)(y)\right)^*.
\end{align*}
Together with~\eqref{eq:Wh}, this implies statements~(\ref{p45i}) and~(\ref{p45ii}).
\end{proof}
M\"{o}bius transformations of Nevanlinna functions preserve $B_\infty$-property and $B_0$-property. In the notation of Theorem~{\rm\ref{tV}} we have the following proposition.
\begin{proposition} \label{thm:Binv}
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{iBinv-1}
Assume that $\mu_1(z) = z$ or $\mu_1(z) = -z$ and let $\mu_2$ be a M\"{o}bius transformation as in \eqref{eq:MobNev12}. A Nevanlinna function $m$ has $B_\infty$-property {\rm (}$B_0$-property{\rm )} if and only if the Nevanlinna function $\wh{m}$ has $B_\infty$-property {\rm (}$B_0$-property, respectively{\rm )}.
\item \label{iBinv-2}
Let $\mu_1(z) = 1/z$ or $\mu_1(z) = -1/z$ and let $\mu_2$ be a M\"{o}bius transformation as in \eqref{eq:MobNev12}. A Nevanlinna function $m$ has $B_\infty$-property {\rm (}$B_0$-property{\rm )} if and only if the Nevanlinna function $\wh{m}$ has $B_0$-property {\rm (}$B_\infty$-property, respectively{\rm )}.
\end{enumerate}
\end{proposition}
\begin{proof}
To prove (\ref{iBinv-1}), let $\mu_1(z) = z$ and calculate
\[
w_{\wh{m}}(y) = |c_2 m(iy)+ d_2|^2 \, w_m(y), \quad y > 0.
\]
For $f \in \cH(m)$ we have
\begin{align*}
\| G_{\wh{m},\infty} Vf \|^2_{L_{w_{\wh{m}}}^2\!(1,\infty)}
& = \int_{1}^{\infty} \frac{1}{|c_2m(iy)+d_2|^2}
\left| f(iy)\right|^2 \, w_{\wh{m}}(y) dy \\
&=\int_{1}^{\infty} \left| f(iy)\right|^2 \, w_{m}(y) dy \\
& = \| G_{m,\infty} f \|^2_{L_{w_m}^2\!(1,\infty)}.
\end{align*}
The above equality, Theorem~\ref{tV}, and elementary arguments
yield the proposition. The statement for $\mu_1(z) = -z$ follows from Lemma~\ref{prop:4.5}.
For (\ref{iBinv-2}), let $\mu_1(z) = -1/z$ and calculate
\[
w_{\wh{m}}(y) = |c_2 m(i/y)+ d_2|^2 \, w_m( 1/y), \quad y > 0.
\]
For $f \in \cH(m)$ we have
\begin{align*}
\| G_{\wh{m},\infty} V f \|^2_{L_{w_{\wh{m}}}^2\!(1,\infty)}
& = \int_{1}^{\infty} \frac{1}{y^2} \frac{1}{|c_2m(i/y)+d_2|^2}
\left| f(i/y)\right|^2 \, w_{\wh{m}}(y) dy \\
& = \int_{1}^{\infty} \frac{1}{y^2} | f(i/y)|^2 w_m(1/y) dy \\
& = \int_{0}^{1} | f(i s )|^2 w_m(s) ds \\
& = \| G_{m,0}f \|^2_{L_{w_m}^2\!(0,1)}.
\end{align*}
This identity, Theorem~\ref{tV} and elementary arguments yield claim (\ref{iBinv-2}). The statement for $\mu_1(z) = 1/z$ follows from Lemma~\ref{prop:4.5}.
\end{proof}
Next, we characterize $B$-properties in terms of a bounded operator between weighted $L^2$ spaces.
\begin{proposition} \label{prop:B0prop}
Let $m$ be a Nevanlinna function with the integral representation
\begin{equation} \label{eq:IntRep+}
m(z) = a +\int_{0}^{+\infty} \left( \frac{1}{x-z} - \frac{x}{1+x^2}
\right) d\sigma(x)
\end{equation}
where $a$ is a real number and $\sigma(x)$ is a non-decreasing
function such that~\eqref{eq:sigmaCond} holds.
The following statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{B0prop-i}
The function $m$ has $B_\infty$-property if and only if the mapping $H_{m,\infty}$ defined by
\begin{equation*}
\bigl(H_{m,\infty} f \bigr)(y) : = \int_{0}^{+\infty}\frac{f(x)}{x+y}d\sigma(x),
\quad f \in L_{\sigma}^2({\mathbb R}_+), \quad y > 1,
\end{equation*}
is a bounded operator from $L_{\sigma}^2({\mathbb R}_+)$ into
$L_{w_m}^2\!(1,\infty)$.
\item \label{B0prop-ii}
The function $m$ has $B_0$-property if and only if the mapping $H_{m,0}$ defined by
\begin{equation*}
\bigl(H_{m,0} f \bigr)(y) : = \int_{0}^{+\infty}\frac{f(x)}{x+y}d\sigma(x),
\quad f \in L_{\sigma}^2({\mathbb R}_+), \quad 0 < y < 1,
\end{equation*}
is a bounded operator from $L_{\sigma}^2({\mathbb R}_+)$ into
$L_{w_m}^2\!(0,1)$.
\end{enumerate}
\end{proposition}
\begin{proof}
It follows from~\cite[Theorem~3.2]{AG04} that the mapping
\begin{equation}\label{eq:Lm}
f \mapsto \int_{0}^{+\infty}\frac{f(x)}{x-z}d\sigma(x)
\end{equation}
is an isomorphism between $L^2_{\sigma}({\mathbb R}_+)$ and $\cH(m)$. Therefore, by Definition~\ref{defBinf}, $m$ has $B_\infty$-property if and only if the composition of \eqref{eq:Lm} and $G_{m,\infty}$, that is, \begin{equation}\label{eq:Tm}
f \mapsto \int_{0}^{+\infty}\frac{f(x)}{x-iy}d\sigma(x)
\end{equation}
is a bounded operator from $L^2_{\sigma}({\mathbb R}^+)$ into
$L^2_{w_m}(1,\infty)$. The inequalities
\[
\frac{1}{x+y}\le \frac{1}{|x-iy|}\le \frac{\sqrt{2}}{x+y}, \quad
(x,y\in{\mathbb R}_+)
\]
imply that the operator in \eqref{eq:Tm} is a bounded operator from
$L^2_{\sigma}({\mathbb R}^+)$ into $L^2_{w_m}(1,\infty)$ if and only if
$H_{m,\infty}$ is a bounded operator from
$L^2_{\sigma}({\mathbb R}^+)$ into $L^2_{w_m}(1,\infty)$. This proves (\ref{B0prop-i}).
The second statement is proved similarly.
\end{proof}
\subsection{Sufficient conditions for $B$-properties of Nevanlinna functions}
We will need the following Schur test, see \cite[Theorem~5.9.2]{Gar}, for boundedness of integral operators.
\begin{lemma}\label{lem:Ap1}
Let $(X,\Sigma_1,\sigma_1)$ and $(Y,\Sigma_2,\sigma_2)$ be $\sigma$-finite measure spaces and let $K(x,y)$ be a nonnegative measurable function on a product space $(X,\Sigma_1,\sigma_1)\times(Y,\Sigma_2,\sigma_2)$. Suppose that there exist strictly positive measurable functions $q_1$ on $(X,\Sigma_1,\sigma_1)$ and $q_2$ on $(Y,\Sigma_2,\sigma_2)$ such that the function
\begin{equation*}
y \mapsto \frac{1}{q_2(y)}\int_X K(x,y)q_1(x)d\sigma_1(x)
\end{equation*}
is essentially bounded on $(Y,\Sigma_2,\sigma_2)$ and the function
\begin{equation*}
x \mapsto \frac{1}{q_1(x)} \int_Y K(x,y)q_2(x)d\sigma_2(y)
\end{equation*}
is essentially bounded on $(X,\Sigma_1,\sigma_1)$.
Then for all $f \in L_{\sigma_1}^2\!(X)$ we have that
\[
(Af)(y) = \int_X K(x,y)f(x)d\sigma_1(x)
\]
exists for $\sigma_2$-almost all $y \in Y$, $Af \in L_{\sigma_2}^2\!(Y)$ and $A$ is a bounded operator from $L_{\sigma_1}^2\!(X)$ to $L_{\sigma_2}^2\!(Y)$.
\end{lemma}
Combining Proposition~\ref{prop:B0prop} and Lemma~\ref{lem:Ap1} one
obtains the following sufficient conditions for $B_\infty$ and
$B_0$-properties.
\begin{corollary} \label{cor:Binfprop}
Let $m$ have the integral representation \eqref{eq:IntRep+} and
assume that there are strictly positive measurable functions $q_1$
on ${\mathbb R}_+$ and $q_2$ on $(1,\infty)$ such that the function
\begin{equation}\label{eq:CorSchur1}
y \mapsto \frac{1}{q_2(y)} \int_{0}^{+\infty}\frac{q_1(x)}{x+y}d\sigma(x)
\end{equation}
is essentially bounded on $y\in(1,\infty)$ and the function
\begin{equation}\label{eq:CorSchur2}
x \mapsto \frac{1}{q_1(x)} \int_{1}^{+\infty}\frac{q_2(y)}{x+y}w_m(y)dy
\end{equation}
is $\sigma$-essentially bounded on $\nR_+$. Then the function $m$ has $B_\infty$-property.
\end{corollary}
\begin{proof}
The assumptions of the corollary and Lemma~\ref{lem:Ap1} imply that the operator $H_{m,\infty}$ is bounded, so, by Proposition~\ref{prop:B0prop}, $m$ has $B_\infty$-property.
\end{proof}
In Theorem~\ref{th:Binfprop} and its corollaries below we give sufficient conditions for $B_\infty$-property and \mbox{$B_0$-property} in terms of the asymptotics of a Nevanlinna function $m$. The key step in the proof of this theorem is the following Abelian and Tauberian theorem for the Stieltjes transform at the point $\infty$ which is essentially contained in \cite{Ben89} and~\cite{CarHay81}, see also~\cite{Car11} for a more general theorem, or \cite{Ka31} for the classical Karamata Tauberian theorem.
\begin{theorem} \label{thm:TaubFinal}
Let $m$ be a Stieltjes function with the integral representation
\begin{equation} \label{eq:IntRep0+}
m({z}) = \int\limits_{0}^{+\infty} \frac{d\sigma(t)}{t-{z}}
\end{equation}
where $\sigma(t)$ is a non-decreasing function satisfying \eqref{eq:sigma-nor} and \eqref{eq:int_S_gc}. Let $C > 0$ and $\alpha \in (0,1)$. Then
\begin{equation}\label{eq:mAsympInf-A}
\text{for all} \ \ z \in \nCR \quad \text{we have} \quad m(r z) \sim \frac{C}{(-r z)^\alpha} \quad \text{as} \quad r \to +\infty,
\end{equation}
if and only if
\begin{equation}\label{eq:SpAsympInf-A}
\sigma(t)\sim \frac{C\sin(\pi\alpha)}{\pi}\frac{t^{1-\alpha}}{1-\alpha} \quad \text{as} \quad t \to +\infty.
\end{equation}
\end{theorem}
\begin{proof}
The direct part of Theorem~\ref{thm:TaubFinal} is a
special case of the Tauberian Theorem~7.5 from~\cite{Ben89}.
Assume now that $\sigma$ satisfies~\eqref{eq:SpAsympInf-A}. Integrating by parts
in~\eqref{eq:IntRep0+} and using~\eqref{eq:SpAsympInf-A} one obtains
\[
m(z) = \int\limits_{0}^{+\infty} \frac{\sigma(t)dt}{(t-z)^2}.
\]
Now the asymptotic~\eqref{eq:mAsympInf-A} is implied by the Abelian
Theorem~3.2 from~\cite{CarHay81}.
\end{proof}
\begin{theorem} \label{th:Binfprop}
Let $m$ be a Stieltjes function. If there exist $\alpha\in(0,1)$ and $C > 0$ such that for all $z \in\nCR$ we have
\begin{equation}\label{eq:AsymInf}
m(rz) \sim \frac{C}{(-r z)^\alpha} \quad \text{as} \quad r \to +\infty,
\end{equation}
then $m$ has $B_\infty$-property.
\end{theorem}
\begin{proof}
Let $m$ be a Stieltjes function with the integral representation \eqref{eq:int_S} for which \eqref{eq:int_S_gc} and \eqref{eq:sigma-nor} hold. Assume \eqref{eq:AsymInf}. Since by \cite[Lemma~S1.3.1]{KaKr74} and \eqref{eq:AsymInf} for $\gamma$ in \eqref{eq:int_S} we have
\[
\gamma = \lim\limits_{y\uparrow+\infty} m(iy) = 0,
\]
Theorem~\ref{thm:TaubFinal} applies.
Consequently, there exists $C_1 = \frac{C\sin(\pi\alpha)}{\pi(1-\alpha)} > 0$ such that
\begin{equation} \label{eqasigma}
\sigma(t) \sim C_1 \, t^{1-\alpha } \quad \text{as} \quad t\to +\infty.
\end{equation}
Let $\beta > 0$ be such that $\alpha + \beta < 1$. Set
\begin{align*}
q_1(x) &:= \frac{1}{(1+x)^\beta} \\
\intertext{and}
q_2(y) &:= \int_{0}^{\infty}\frac{d\sigma_1(x)}{x+y}, \quad \text{where} \quad
\sigma_1(x)=\int_{0}^{x}\frac{d\sigma(t)}{(1+t)^\beta}.
\end{align*}
Notice that by definitions of $q_1$ and $q_2$ we have
\begin{equation*}
\int_0^\infty \frac{q_1(x)}{x+y} d\sigma(x) = q_2(y) \quad \text{for all} \quad y > 1.
\end{equation*}
Hence, $q_1$ and $q_2$ satisfy \eqref{eq:CorSchur1} in Corollary~\ref{cor:Binfprop}. The rest of the proof is a verification of \eqref{eq:CorSchur2}.
Integration by parts in the formula for $\sigma_1$ yields
\begin{equation} \label{eq2ndsig}
\sigma_1(x)=\frac{\sigma(x)}{(1+x)^\beta} + \beta \int_{0}^{x} \frac{\sigma(t)}{(1+t)^{\beta+1}} dt .
\end{equation}
It follows from \eqref{eqasigma} that
\[
\frac{\sigma(t)}{(1+t)^{\beta+1}} \sim \frac{C_1}{t^{\alpha + \beta}}
\quad \text{as} \quad t\to +\infty,
\]
and, by l'H\^{o}pital's rule,
\[
\int_{0}^{x} \frac{\sigma(t)}{(1+x)^{\beta+1}} dt \sim
C_2 \, x^{1- \alpha - \beta} \quad \text{as} \quad x\to +\infty.
\]
Since also
\[
\frac{\sigma(t)}{(1+t)^{\beta}} \sim C_1 \, t^{1-\alpha - \beta}
\quad \text{as} \quad t\to +\infty,
\]
the equality \eqref{eq2ndsig} implies
\begin{equation} \label{eqsig1as}
\sigma_1(x) \sim C_3 \, x^{1-\alpha - \beta} \quad \text{as} \quad x\to +\infty,
\end{equation}
Now \eqref{eqsig1as} and Theorem~\ref{thm:TaubFinal} yield
\begin{equation}\label{eq:4.12}
q_2(y) \sim \frac{C_4}{y^{\alpha+\beta}} \quad \text{as} \quad y\to\infty.
\end{equation}
for some $C_4 > 0$. Next consider the function
\[
x \mapsto \int_{1}^{\infty}\frac{q_2(y)}{x+y}w_m(y)dy, \quad x > 0.
\]
Since by \eqref{eq:4.12} and assumption \eqref{eq:AsymInf} we have
\[
q_2(y)w_m(y) \sim \frac{C_5}{y^{\beta}} \quad \text{as} \quad y\to\infty,
\]
for some $C_5 > 0$, Theorem~\ref{thm:TaubFinal} yields
\[
\int_{1}^{\infty}\frac{q_2(y)}{x+y}w_m(y)dy \sim \frac{C_6}{x^{\beta}} \quad \text{as} \quad x\to\infty.
\]
The last displayed relationship implies that the function
\[
x \mapsto \frac{1}{q_1(x)} \int_{1}^{\infty}\frac{q_2(y)}{x+y}w_m(y)dy, \quad x > 0,
\]
is bounded; that is \eqref{eq:CorSchur2} holds. Now Corollary~\ref{cor:Binfprop} implies that $m$ has $B_\infty$- property.
\end{proof}
\begin{corollary} \label{co:Binfprop}
Every function in $\cA_\infty$ has $B_\infty$-property.
\end{corollary}
\begin{proof}
Let $m \in \cA_\infty$ be arbitrary. Then there exists a M\"{o}bius transformation $\mu$ such that $\mu\circ m$ is a Stieltjes function and \eqref{eq:mAsymInf-defAInf} holds. By Theorem~\ref{th:Binfprop} the Nevanlinna function $\mu\circ m$ has $B_\infty$-property. In turn, Proposition~\ref{thm:Binv}(\ref{iBinv-1}) yields that $m = \mu^{-1}\circ\mu\circ m$ also has $B_\infty$-property.
\end{proof}
\begin{corollary} \label{th:B0prop}
Every function in $\cA_0$ has $B_0$-property.
\end{corollary}
\begin{proof}
Let $m \in \cA_0$ be arbitrary. Set $\mu_1(z) = 1/z$ and $\mu_2(z) = -z$. By Proposition~\ref{prop:AInf-AZer}(\ref{i:AInf-AZer-2}) the function $\wh{m} = \mu_2\circ m\circ \mu_1$ belongs to $\cA_\infty$, so $\wh{m}$ has $B_\infty$-property by Corollary~\ref{co:Binfprop}. Now, Proposition~\ref{thm:Binv}(\ref{iBinv-2}) implies that $m$ has $B_0$-property.
\end{proof}
\section{Coupling of symmetric operators} \label{SecCoup}
\subsection{Coupling of symmetric operators in a Hilbert space.} \label{SubS:CoupleHS}
In this section we consider Hilbert spaces $\bigl( \cH_+, \ahip_{\cH_+} \bigr)$ and $\bigl( \cH_-, \ahip_{\cH_-} \bigr)$ and their (external) direct sum ${\cH} = \cH_+\oplus \cH_-$ with the natural inner product $\ahip_{\cH}$. If $T_+$ is an operator in $\cH_+$ and $T_-$ is an operator in $\cH_-$, then $T_+ \oplus T_-$ denotes their direct sum, that is
\begin{equation*} \label{edefwAHC}
\bigl(T_+ \oplus T_-\bigr) \begin{pmatrix} f_+ \\ f_-
\end{pmatrix}:=
\begin{pmatrix} T_+ f_+ \\ T_- f_- \end{pmatrix}, \quad \ f_+ \in \dom(T_+), \ f_- \in \dom(T_-).
\end{equation*}
The following assumptions apply to all four statements in this subsection. We assume that $S_\pm$ is a closed symmetric densely defined operator with defect numbers $(1,1)$ in the Hilbert space $\bigl( \cH_\pm, \ahip_{\cH_\pm} \bigr)$. We let $\bigl(\nC,\Gamma_{0}^\pm, \Gamma_{1}^\pm\bigr)$ be a boundary triple for $S_\pm^*$ and $m_\pm$ and $\psi_\pm$ are the corresponding Weyl function and the Weyl solution. By $S_0^\pm$ we denote the self-adjoint extension of $S_\pm$ which is defined on $\dom(S_0^\pm) = \ker(\Gamma_{0}^\pm)$. That is
$S_0^\pm=S_\pm^*|_{\ker(\Gamma_{0}^\pm)}$.
In the following theorem we reformulate results from~\cite{DHMS}, (see also~\cite{CDR}) in a form which is convenient for further use in this paper.
\begin{theorem} \label{tHsc}
Under the general assumptions of this subsection we have:
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item
The linear operator $S$ defined as the restriction of $S_+^* \oplus S_-^*$ to the domain
\begin{equation*}
\dom(S) = \left\{
\begin{pmatrix}\!f_+ \\ f_-\!\end{pmatrix}: \!
\begin{array}{l}
\Gamma_{0}^+(f_+) = \Gamma_{0}^- (f_-)=0, \\
\Gamma_{1}^+(f_+) + \Gamma_{1}^- (f_-)=0,
\end{array} \ \begin{array}{l} f_+ \in \dom(S_+^*), \\ f_-\in \dom(S_-^*)
\end{array}
\right\}
\end{equation*}
is, closed, densely defined and symmetric with defect
numbers $(1,1)$ in the Hilbert space ${\cH}$.
\item \label{H2b}
The adjoint ${S}^*$ of ${S}$ is the restriction of $S_+^* \oplus S_-^*$ to the domain
\begin{equation*}
\dom(S^*) = \left\{\!
\begin{pmatrix}\! f_+ \\ f_- \!\end{pmatrix} :
\Gamma_{0}^+(f_+) - \Gamma_{0}^- (f_-) = 0, \ \begin{array}{l} f_+ \in \dom(S_+^*), \\ f_-\in \dom(S_-^*)
\end{array} \right\}.
\end{equation*}
\item \label{ibatr}
A boundary triple $\bigl(\nC,\Gamma_0,\Gamma_1\bigr)$ for $S^*$ is
given by
\begin{equation} \label{Htripl}
\Gamma_0 f = \Gamma_0^+ f_+, \quad
\Gamma_1 f = \Gamma_1^+ f_+ + \Gamma_1^-f_-, \quad
f = \begin{pmatrix} f_+ \\ f_- \end{pmatrix} \in \dom(S^*).
\end{equation}
\item \label{H2d}
The Weyl function of $S$ relative to the boundary triple $\bigl(\nC,\Gamma_0,\Gamma_1\bigr)$ is
\begin{equation*} \label{mH2}
m(z) = m_+(z) + m_-(z), \quad z \in \nC \setminus \nR.
\end{equation*}
\item \label{ietHsc}
The self-adjoint extension $S_1$ of $S$ such that $\dom(S_1) = \ker(\Gamma_1)$ coincides with the restriction of $S_+^* \oplus S_-^*$
to the domain
\begin{equation} \label{eq:domwAHC}
\dom(S_1) = \left\{\begin{pmatrix} f_+ \\ f_- \end{pmatrix}:
\begin{array}{l}
\Gamma_{0}^+(f_+) - \Gamma_{0}^- (f_-)=0, \\
\Gamma_{1}^+(f_+) + \Gamma_{1}^- (f_-)=0,
\end{array} \ \begin{array}{l} f_+ \in \dom(S_+^*), \\ f_-\in \dom(S_-^*)
\end{array}
\right\}.
\end{equation}
\item \label{ietHsc0}
The self-adjoint extension $S_0$ of $S$ such that $\dom(S_0) = \ker(\Gamma_0)$ coincides with the restriction of $S_+^* \oplus S_-^*$
to the domain
\begin{equation*}
\dom(S_0) = \left\{\begin{pmatrix} f_+ \\ f_- \end{pmatrix}:\,
\Gamma_{0}^+(f_+) = \Gamma_{0}^- (f_-)=0, \ \begin{array}{l} f_+ \in \dom(S_+^*), \\ f_-\in \dom(S_-^*)
\end{array} \right\}
\end{equation*}
and thus $S_0=S_0^+\oplus S_0^-$.
\end{enumerate}
\end{theorem}
The operator $S_1$ defined in Theorem~\ref{tHsc}(\ref{ietHsc}) is called the {\it coupling} of the operators $S_+$ and $S_-$ in the Hilbert space $\bigl( {\cH}, \ahip_{{\cH}} \bigr)$ relative to the triples $\bigl(\nC,\Gamma_{0}^+, \Gamma_{1}^+\bigr)$ and $\bigl(\nC,\Gamma_{0}^-, \Gamma_{1}^-\bigr)$.
In the next proposition we give a criterion for nonnegativity of the coupling $S_1$ of two nonnegative operators $S_+$ and $S_-$ in the Hilbert space $\bigl( {\cH}, \ahip_{{\cH}} \bigr)$.
\begin{proposition}\label{prop:nonneg}
In addition to the general assumptions of this subsection assume that $S_0^\pm$ is a nonnegative self-adjoint extension of $S_\pm$ in $\cH_\pm$. Then the coupling ${S_1}$ is a nonnegative self-adjoint operator in $\bigl(\cH, \ahip_{\cH}\bigr)$ if and only if
\begin{equation} \label{eqN}
m_+(x) + m_-(x) \neq 0 \qquad \text{for all} \qquad x \in \nR_-.
\end{equation}
\end{proposition}
\begin{proof}
Since the self-adjoint extensions $S_0^+$ and $S_0^-$ are nonnegative, then the Weyl functions $m_+$ and $m_-$ are holomorphic on $\nR_-$ and $(-\infty,0)\subset\rho(S_0)$, where $S_0=S_0^+\oplus S_0^-$. As, by Theorem~\ref{tHsc}(\ref{H2d}), $m(z)=m_+(z)+m_-(z)$ is the Weyl function of $S$ relative to the boundary triple~\eqref{Htripl}, it follows from Proposition~\ref{prop:S_01} that~\eqref{eqN} holds if and only if $(-\infty,0)\subset\rho(S_1)$, or, equivalently, if and only if $S_1$ is nonnegative.
\end{proof}
\begin{corollary} \label{col:nonneg}
In addition to the general assumptions of this subsection assume that the operator $S_0^+$ is the Kre\u{\i}n extension of $S_+$. Then the coupling ${S_1}$ is a nonnegative self-adjoint operator in the Hilbert space $\bigl( {\cH}, \ahip_{{\cH}} \bigr)$ if and only if the following two conditions are satisfied:
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item
$S_0^-:=S_-^*|_{\ker(\Gamma_0^-) }$ is a non-negative operator in the Hilbert space $\cH_-$.
\item \label{eq:m_+-inf}
$\lim_{x\downarrow -\infty} \bigl( m_+(x) + m_-(x) \bigr) \geq 0$.
\end{enumerate}
If the coupling ${S_1}$ is a nonnegative, then there exists a constant $c\in\nR$ such that $m_+-c$ and $m_-+c$ are Stieltjes functions.
\end{corollary}
\begin{proof} Since $S_0^+$ is the Kre\u{\i}n extension of $S_+$, then by~\eqref{eq:m_K0}
\begin{equation}\label{eq:m_+inf}
\lim_{x\uparrow 0}m_+(x)=+\infty.
\end{equation}
If the coupling ${S_1}$ is nonnegative, then the operator $S_-$ is also nonnegative and hence its extension $S_0^-$ has at most one negative eigenvalue (see~\cite{Kr47}). Then the Weyl function $m_-$ has at most one pole on $\nR_-$ and by~\eqref{eq:m_+inf}
\begin{equation}\label{eq:m+-K0}
\lim_{x\uparrow 0}\bigl(m_+(x)+m_-(x)\bigr)=+\infty.
\end{equation}
Assume that $m_-$ has a pole at $x_0\in(-\infty,0)$. Then
\[
\lim_{x\downarrow x_0}\bigl(m_+(x)+m_-(x)\bigr)=-\infty,\quad
\]
and, hence $m_+ + m_-$ has a zero in the interval $(x_0,0)$.
By Proposition~\ref{prop:nonneg} the operator $S_1$ has a negative eigenvalue which is impossible, if the operator $S_1$ is nonnegative.
Therefore, the function $m_-$ is holomorphic on $\nR_-$ and $S_0^-$ is a non-negative operator in the Hilbert space $\cH_-$.
Since the restrictions of $m_+$ and $m_-$ to $\nR_-$ are continuous and monotonically increasing functions which satisfy the condition~\eqref{eq:m+-K0}, the condition~\eqref{eqN} of nonnegativity of $S_1$ can be rewritten as~\eqref{eq:m_+-inf}.
Conversely, if (i) and (ii) hold, then the operators $S_0^+$ and $S_0^-$ are nonnegative in $\cH_+$ and $\cH_-$, respectively, $m_-$ is holomorphic on $\nR_-$ and~\eqref{eqN} holds. Therefore, the coupling $S_1$ is nonnegative by Proposition~\ref{prop:nonneg}.
To prove the last statement, set
$
c = \lim_{x\downarrow -\infty}m_+(x).
$
Then $m_+(x)-c \ge 0$ and by (ii)
\[
m_-(x)+ c \ge \lim_{x\downarrow -\infty}m_-(x)+\lim_{x\downarrow -\infty}m_+(x)\ge 0
\]
for all $x\in\nR_-$. Therefore $m_+-c$ and $m_-+c$ are Stieltjes functions.
\end{proof}
The next corollary is proved similarly.
\begin{corollary} \label{col:nonnegF}
In addition to the general assumptions of this subsection assume that the operator $S_0^+$ is the Friedrichs extension of $S_+$.
Then the coupling ${S_1}$ is a nonnegative self-adjoint operator in the Hilbert space $\bigl( {\cH}, \ahip_{{\cH}} \bigr)$ if and only if the following two conditions are satisfied:
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item
$S_0^-:=S_-^*|_{\ker(\Gamma_0^-) }$ is a non-negative operator in $\cH_-$.
\item
$m_+(0) + m_-(0) =\lim_{x\uparrow 0-} \bigl(m_+(x)+m_-(x) \bigr) \leq 0.$
\end{enumerate}
If the coupling $S_1$ is nonnegative, then there exists a constant $c\in\nR$ such that $m_+-c$ and $m_-+c$ are inverse Stieltjes functions.
\end{corollary}
\subsection{Non-separated extensions as couplings.}
In Theorem~\ref{tHsc}(\ref{ietHsc}) we defined the coupling $S_1$ of symmetric operators $S_+$ and $S_-$ relative to the boundary triples $\bigl(\nC,\Gamma_0^+,\Gamma_1^+\bigr)$ for $S_+^*$ and $\bigl(\nC,\Gamma_0^-,\Gamma_1^-\bigr)$ for $S_-^*$. Using different boundary triples would result in a different coupling operator. In this subsection we characterize all self-adjoint extensions of $S_+\oplus S_-$ which can be obtained as couplings of $S_+$ and $S_-$.
Let, as before, $\bigl(\nC,\Gamma_0^\pm,\Gamma_1^\pm\bigr)$ be a boundary triple for $S_\pm^*$. It is not difficult to see (cf.~\cite[Lemma~3.4]{CDR}, \cite{DHMS} for a more general setting or \cite[Appendix~II,~125]{AG78} for the case of differential operators) that a restriction $\wh{S}$ of $S_+^*\!\oplus\!S_-^*$ is a self-adjoint extension of $S_+\oplus S_-$ in $\bigl(\cH, \ahip_{\cH}\bigr)$ if and only if
\[
\dom(\wh{S}) = \left\{ \left(\!\!\!\begin{array}{c} f_+ \\ f_-
\end{array}\!\!\!\right) \in \bigl(\dom S_+^*\bigr) \oplus \bigl(\dom S_-^*\bigr) :
M\! \left(\!\!\!\begin{array}{c} \Gamma_1^+ f_+ \\ \Gamma_1^- f_-
\end{array}\!\!\!\!\right) = N\! \left(\!\!\!\begin{array}{c} \Gamma_0^+ f_+ \\ \Gamma_0^- f_-
\end{array}\!\!\!\!\right) \right\},
\]
where $M$ and $N$ are $2\times 2$ matrices with complex entries such that the block matrix $\bigl( M \ N \bigr)$ has rank $2$ and the matrix $MN^*$ is self-adjoint.
If $\wh{S}$ is a self-adjoint extension of $S_+\oplus S_-$ in $\bigl(\cH, \ahip_{\cH}\bigr)$ then either
\begin{equation} \label{eq:nssa}
(\gr \wh{S}) \cap \bigl(\cH_+ \oplus \{0\}\bigr)^2 = \gr(S_+) \quad \text{and}
\quad (\gr \wh{S}) \cap \bigl(\{0\}\oplus \cH_-\bigr)^2 = \gr(S_-)
\end{equation}
or
\begin{equation*}
(\gr \wh{S}) \cap \bigl(\cH_+ \oplus \{0\}\bigr)^2 \quad \text{and}
\quad (\gr \wh{S}) \cap \bigl(\{0\}\oplus \cH_-\bigr)^2
\end{equation*}
are graphs of self-adjoint operators in $\bigl(\cH_+, \ahip_{\cH_+}\bigr)$ and $\bigl(\cH_-, \ahip_{\cH_-}\bigr)$, respectively.
In the later case we say that the self-adjoint operator $\wh{S}$ is a {\em separated} extension of $S_+\oplus S_-$ and the corresponding boundary conditions are called {\em separated} boundary conditions. If \eqref{eq:nssa} holds, then we say that $\wh{S}$ is a {\em non-separated} extension of $S_+\oplus S_-$ and the corresponding boundary conditions are called {\em non-separated} boundary conditions.
Notice that multiplying matrices $M$ and $N$ from the left by the same invertible matrix does not change the domain of $\wh{S}$. Therefore, we can assume that the block matrix $\bigl( M \ N \bigr)$ is in reduced row echelon form, see \cite[Subsection~0.3.4]{HJ}. As the rank of $\bigl( M \ N \bigr)$ is $2$, the reduced row echelon form of $\bigl( M \ N \bigr)$ takes one of the following six forms:
\begin{alignat*}{3}
&
\left(\!\!\!\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \end{array}\!\!\!\right) & \qquad
&
\left(\!\!\!\begin{array}{cccc}
0 & 1 & * & 0 \\
0 & 0 & 0 & 1 \end{array}\!\!\!\right) & \qquad
&
\left(\!\!\!\begin{array}{cccc}
1 & * & * & 0 \\
0 & 0 & 0 & 1 \end{array}\!\!\!\right) \\
&
\left(\!\!\!\begin{array}{cccc}
0 & 1 & 0 & * \\
0 & 0 & 1 & * \end{array}\!\!\!\right) & \qquad
&
\left(\!\!\!\begin{array}{cccc}
1 & * & 0 & * \\
0 & 0 & 1 & * \end{array}\!\!\!\right) & \qquad
&
\left(\!\!\!\begin{array}{cccc}
1 & 0 & * & * \\
0 & 1 & * & * \end{array}\!\!\!\right),
\end{alignat*}
where $*$ stands for an arbitrary complex number.
A straightforward calculations show that the only matrices of the above 6 types for which the corresponding matrix $MN^*$ is self-adjoint are of the following four types:
\begin{alignat}{2} \label{eq:types-a-b}
&
\left(\!\!\!\begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \end{array}\!\!\!\right) & \qquad
&
\left(\!\!\!\begin{array}{cccc}
1 & 0 & \alpha & 0 \\
0 & 0 & 0 & 1 \end{array}\!\!\!\right) \\ \label{eq:types-c-d}
&
\left(\!\!\!\begin{array}{cccc}
1 & \rho e^{i \theta} & 0 & \sigma e^{i \theta} \\
0 & 0 & 1 & -e^{i \theta}/\rho \end{array}\!\!\!\right) & \qquad
&
\left(\!\!\!\begin{array}{cccc}
1 & 0 & \alpha & \omega \\
0 & 1 & \omega^* & \beta \end{array}\!\!\!\right),
\end{alignat}
where $\omega \in \nC$, $\alpha, \beta, \sigma, \theta \in \nR$, and $\rho > 0$. Clearly, the matrices in \eqref{eq:types-a-b} give rise to separated boundary conditions, while the first matrix in \eqref{eq:types-c-d} leads to non-separated boundary conditions. The second matrix in \eqref{eq:types-c-d} leads to separated boundary conditions if and only if $\omega = 0$. A similar classification of boundary conditions for a regular Sturm-Liouville problem was established in \cite{CFK13}.
\begin{theorem}\label{Thm:5.5}
Let $S_\pm$ be a closed symmetric densely defined operator with defect numbers $(1,1)$ in the Hilbert space $\bigl(\cH_\pm,\ahip_{\cH_\pm}\bigr)$. A self-adjoint extension $\wh{S}$ of $S_+\oplus S_-$ in $\bigl(\cH, \ahip_{\cH}\bigr)$ is a coupling of $S_+$ and $S_-$ relative to some boundary triples of $S_+^*$ and $S_-^*$ if and only if $\wh{S}$ is a non-separated extension of $S_+\oplus S_-$.
\end{theorem}
\begin{proof}
Assume that $\wh{S}$ is a coupling of $S_+$ and $S_-$ relative to boundary triples $\bigl(\nC,\Gamma_0^+,\Gamma_1^+\bigr)$ and $\bigl(\nC,\Gamma_0^-,\Gamma_1^-\bigr)$. Let
\[
\left\{\left(\!\!\!\begin{array}{c}
f_+ \\ 0
\end{array}\!\!\!\right), \left(\!\!\!\begin{array}{c}
S_+^* f_+ \\ 0
\end{array}\!\!\!\right) \right\} \in \gr \wh{S}.
\]
Then, by definition of $\wh{S}$ as a coupling, we have $\Gamma_0^+ f_+ - \Gamma_0^- 0 = 0$ and $\Gamma_1^+ f_+ + \Gamma_1^- 0 = 0$. That is, $\Gamma_0^+ f_+ = 0$ and $\Gamma_1^+ f_+ = 0$, proving that $(\gr \wh{S}) \cap \bigl(\cH_+ \oplus \{0\}\bigr)^2 = \gr S_+$. Similarly, $(\gr \wh{S}) \cap \bigl(\{0\}\oplus\cH_- \bigr)^2 = \gr S_-$. Hence, $\wh{S}$ is a non-separated extension.
To prove the converse, assume that $\wh{S}$ is a non-separated self-adjoint extension of $S_+\oplus S_-$ in $\bigl(\cH, \ahip_{\cH}\bigr)$. Let $\bigl(\nC,\Gamma_0^\pm,\Gamma_1^\pm\bigr)$ be a boundary triple for $S_\pm^*$. Then the block matrix $\bigl(M \ N\bigr)$ corresponding to the boundary conditions that determine the domain of $\wh{S}$ are of two forms in \eqref{eq:types-c-d} with $\omega \neq 0$.
First we consider the boundary conditions corresponding to the first matrix in \eqref{eq:types-c-d}
\begin{alignat*}{3}
&\Gamma_0^+ f_+ & &- \bigl(e^{i \theta}/\rho\bigr) \Gamma_0^- f_- & & = 0 \\
&\Gamma_1^+ f_+ & &+ \bigl(-\sigma e^{i \theta}\bigr) \Gamma_0^- f_- + \rho e^{i \theta} \Gamma_1^- f_- & & = 0.
\end{alignat*}
These boundary conditions can be rewritten as:
\begin{equation}\label{eq:Coupl_Cond}
\wh{\Gamma}_{0}^+(f_+) - \wh\Gamma_{0}^- (f_-)=0, \quad
\wh{\Gamma}_{1}^+(f_+) + \wh\Gamma_{1}^- (f_-)=0,
\end{equation}
where
\begin{equation*}
\begin{pmatrix}
\wh\Gamma_{0}^+ (f_+) \\ \wh\Gamma_{1}^+ (f_+)
\end{pmatrix}=
\begin{pmatrix}
\Gamma_{0}^+ (f_+) \\ \Gamma_{1}^+ (f_+)
\end{pmatrix}
,\quad
\begin{pmatrix}
\wh\Gamma_{0}^- (f_-) \\ \wh\Gamma_{1}^- (f_-)
\end{pmatrix}=
e^{i\theta}
\begin{pmatrix}
1/\rho & 0\\
-\sigma & \rho
\end{pmatrix}\begin{pmatrix}
\Gamma_{0}^- (f_-) \\ \Gamma_{1}^- (f_-)
\end{pmatrix}
\end{equation*}
By Remark~\ref{rem:transp} the triple $\bigl(\nC,\wh{\Gamma}_0^-,\wh{\Gamma}_1^- \bigr)$ is a boundary triple for $S_-^*$. Therefore, the coupling of $S_+$ and $S_-$ relative to the boundary triples $\bigl(\nC,\wh{\Gamma}_0^+,\wh{\Gamma}_1^+ \bigr)$ and $\bigl(\nC,\wh{\Gamma}_0^-,\wh{\Gamma}_1^- \bigr)$ coincides with the operator $\wh{S}$.
Next we consider the boundary conditions corresponding to the second matrix in \eqref{eq:types-c-d} with $\omega \neq 0$. These boundary conditions are
\begin{alignat}{3} \label{eq:B_Cond21}
&\omega^* \Gamma_0^+ f_+ & &- \bigl( \Gamma_1^- f_- - \beta \Gamma_0^- f_- \bigr) & & = 0 \\
\label{eq:B_Cond22}
&\frac{1}{\omega}\Gamma_1^+ f_+ - \frac{\alpha}{\omega} \Gamma_0^+ f_+
& & + (- \Gamma_0^- f_-) & & = 0
\end{alignat}
Setting
\begin{equation*}
\begin{pmatrix}
\wh\Gamma_{0}^+ (f_+) \\ \wh\Gamma_{1}^+ (f_+)
\end{pmatrix}=
\begin{pmatrix}
\omega^* & 0\\
-\frac{\alpha}{\omega} & \frac{1}{\omega}
\end{pmatrix}
\begin{pmatrix}
\Gamma_{0}^+ (f_+) \\ \Gamma_{1}^+ (f_+)
\end{pmatrix},\
\begin{pmatrix}
\wh\Gamma_{0}^- (f_-) \\ \wh\Gamma_{1}^- (f_-)
\end{pmatrix}=
\begin{pmatrix}
-\beta & 1\\
-1 & 0
\end{pmatrix}\begin{pmatrix}
\Gamma_{0}^- (f_-) \\ \Gamma_{1}^- (f_-)
\end{pmatrix}
\end{equation*}
one can rewrite boundary conditions~\eqref{eq:B_Cond21},\eqref{eq:B_Cond22} in the form~\eqref{eq:Coupl_Cond}. By Remark~\ref{rem:transp} the triple $\bigl(\nC,\wh{\Gamma}_0^+,\wh{\Gamma}_1^+ \bigr)$ is a boundary triple for $S_+^*$ and the triple $\bigl(\nC,\wh{\Gamma}_0^-,\wh{\Gamma}_1^- \bigr)$ is a boundary triple for $S_-^*$. Therefore the coupling of $S_+$ and $S_-$ relative to the boundary triple $\bigl(\nC,\wh{\Gamma}_0^+,\wh{\Gamma}_1^+ \bigr)$ for $S_+^*$ and the boundary triple $\bigl(\nC,\wh{\Gamma}_0^-,\wh{\Gamma}_1^- \bigr)$ for $S_-^*$ coincides with the operator $\wh{S}$.
This completes the proof.
\end{proof}
\subsection{A partially fundamentally reducible operator as a coupling} \label{SubS:CoupleKS}
Given two symmetric operators, $S_+$ in the Hilbert space $\cH_+$ and $S_-$ in the Hilbert space $\cH_-$, in Subsection~\ref{SubS:CoupleHS} we constructed a coupling $S_1$ of these two operators which is a self-adjoint operator in the Hilbert space $\cH=\cH_+\oplus\cH_-$.
Next we introduce a Kre\u{\i}n space structure on $\cH$. Let $J$ be a self-adjoint involution on $\bigl( \cH, \ahip_{\cH} \bigr)$ defined by
\begin{equation} \label{eq:defJ}
J \begin{pmatrix} f_+ \\ f_- \end{pmatrix} := \begin{pmatrix}\! f_+\! \\ \!- f_- \!\end{pmatrix}, \quad f_+ \in \cH_+, \ f_- \in \cH_-.
\end{equation}
This involution induces an indefinite inner product on $\cH$:
\[
[f,g]_{{\cH}} := \bigl\langle {J} f,g\bigr\rangle_{{\cH}}, \quad f,g \in {\cH},
\]
and with this inner product $\bigl( {\cH}, \kip_{{\cH}} \bigr)$ is a Kre\u{\i}n space.
With $S_0$ and $S_1$ from Theorem~\ref{tHsc} we define $A_0 := J S_0$ and $A_1 := J S_1$. Since $S_0$ and $S_1$ are self-adjoint in the Hilbert space $({\cH},\ahip_\cH)$, $A_0$ and $A_1$ are self-adjoint in the Kre\u{\i}n space $(\cH,\kip_\cH)$. In fact, $A_0$ and $A_1$ are given by
\begin{equation}\label{eq:A0}
A_0 \begin{pmatrix} f_+ \\ f_- \end{pmatrix}=\begin{pmatrix} S_0^+ f_+ \\ -S_0^-f_-\end{pmatrix}, \quad
\begin{pmatrix} f_+ \\ f_- \end{pmatrix}\in \dom(A_0)=\dom(S_0),
\end{equation}
\begin{equation}\label{eq:A1}
A_1 \begin{pmatrix} f_+ \\ f_- \end{pmatrix}=\begin{pmatrix} S_+^* f_+ \\ -S_-^*f_-\end{pmatrix}, \quad
\begin{pmatrix} f_+ \\ f_- \end{pmatrix}\in \dom(A_1)=\dom(S_1).
\end{equation}
The operator $A_1$ will be called the {\it coupling} of the operators $S_+$ and $-S_-$ in the Kre\u{\i}n space $\bigl( {\cH}, \kip_{{\cH}} \bigr)$.
In the rest of this subsection we will proceed in the opposite direction: Given a partially fundamentally reducible self-adjoint operator $A$ in the Kre\u{\i}n space $({\cK}, \kip_{\cK})$ we will prove that $A$ is a coupling of two operators.
Recall that Definition~\ref{def:PartFR} associates a fundamental decomposition
$\cK = \cK_+ [\dot{+}]\cK_-$ of $(\cK, \kip_{\cK})$ and symmetric operators $S_+$ and $S_-$
with a partially fundamentally reducible operator $A$ in a Kre\u{\i}n space $(\cK, \kip_{\cK})$.
As before, $S_\pm^*$ denotes the adjoint of $S_\pm$ in the Hilbert space $(\cK_\pm, \pm\kip)$. By $P_+$ and $P_-$ we denote the orthogonal projections and by $J$ the fundamental symmetry corresponding to the fundamental decomposition $\cK = \cK_+[\dot{+}]\cK_-$, while $\ahip_\cK$ denotes the corresponding Hilbert space inner product and $\|\cdot\|_\cK$ denotes the norm induced by $\ahip_\cK$. The notation related to a partially fundamentally reducible operator $A$ introduced in this paragraph is used throughout the rest of the paper.
\begin{theorem} \label{prop:AasCoupl}
Let $A$ be a partially fundamentally reducible operator in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$. The following statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{pAasCoupl-1}
We have $\dom(S_\pm^*)=P_\pm\dom(A)$ and
\begin{equation*}
\pm S_\pm^*P_\pm f=P_\pm Af\quad\mbox{for all}\quad f\in\dom(A).
\end{equation*}
\item \label{pAasCoupl-2}
Let $\bigl(\nC,\Gamma_0^+,\Gamma_1^+\bigr)$ be a boundary triple for $S_+^*$. Then the equalities
\begin{equation}\label{eq:Gamma-}
\Gamma_0^- P_-f = \Gamma_0^+ P_+f, \qquad \Gamma_1^- P_-f = - \Gamma_1^+ P_+f, \qquad f\in\dom(A),
\end{equation}
define a boundary triple $\bigl(\nC,\Gamma_0^-,\Gamma_1^-\bigr)$ for $S_-$.
\item \label{pAasCoupl-3}
We have $A = JS_1$ where $S_1$ is the coupling of the operators $S_+$ and $S_-$ in the Hilbert space $(\cK, \ahip_\cK)$ relative to the boundary triples $\bigl(\nC,\Gamma_0^+,\Gamma_1^+\bigr)$ and $\bigl(\nC,\Gamma_0^-,\Gamma_1^-\bigr)$.
\item
Let $A$ be a nonnegative operator in $(\cK, \kip_{\cK})$ and assume that the operator $S_0^+:=S_+^*|_{\ker(\Gamma_0^+) }$ coincides with Kre\u{\i}n's or Friedrichs' extension of $S_+$. Then $S_0^-:=S_-^*|_{\ker(\Gamma_0^-) }$ is a nonnegative operator in the Hilbert space $(\cK_-, -\kip_{{\cK}})$.
\end{enumerate}
\end{theorem}
\begin{proof}
The statements (\ref{pAasCoupl-1}) and (\ref{pAasCoupl-2}) are proved in~\cite[Lemma 5.1]{DHMS}.
(\ref{pAasCoupl-3}) It follows from~\eqref{eq:domwAHC} and \eqref{eq:Gamma-} that $\dom(A)=\dom(S_1)$. The equality $A=J\!A_1$ follows from~\eqref{eq:A1}.
The last statement is implied by Corollaries~\ref{col:nonneg} and~\ref{col:nonnegF}.
\end{proof}
The results of the next theorem can be derived from~\cite{Der95}. However, we prefer to present a direct proof for this elementary case when defect numbers of $S_\pm$ are $(1,1)$. In the case of indefinite Sturm-Liouville operator similar construction has been used in
\cite[Proposition~2.5]{KaMal07}.
\begin{theorem} \label{tKsc}
Let $A$ be a partially fundamentally reducible operator in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$ and let $\bigl(\nC,\Gamma_0^+, \Gamma_1^+\bigr)$ and $\bigl(\nC,\Gamma_0^-, \Gamma_1^-\bigr)$ be the boundary triples from Theorem~{\rm\ref{prop:AasCoupl}~\!\!(\ref{pAasCoupl-2})}. Let $m_\pm$ be the Weyl function and let $\psi_\pm$ be the Weyl solution of $S_\pm$ relative to the boundary triple $\bigl(\nC,\Gamma_0^\pm, \Gamma_1^\pm\bigr)$. Let the operator $A_0$ be given by~\eqref{eq:A0}. Then
\begin{equation} \label{eRwA}
\rho(A) \setminus\nR
= \bigl\{ z \in \nCR : m_+(z) + m_-(-z) \neq 0 \bigr\}.
\end{equation}
If $\rho(A) \setminus\nR $ is a nonempty set, then for every $\ z \in \rho(A) \setminus\nR$ and every $h \in \cK$ the resolvent of $A$ is given by
\begin{equation} \label{ereswA}
\bigl(A - z \bigr)^{-1} h = \bigl(A_0 - z \bigr)^{-1} h
- \frac{\bigl[h,\psi(\co{z})\bigr]_{{\cK}}}{m_+(z)+m_-(-z)} \,
\psi(z),
\end{equation}
where
\begin{equation*}
\psi(z)= \psi_+(z) + \psi_-(-z).
\end{equation*}
For $h\in\cK_+$ the vector $f=P_{+}\bigl(A - z \bigr)^{-1} h$ is the solution of the $z$-dependent boundary value problem
\begin{equation}\label{eq:z_depend}
(S_+^*-z)f=h,\quad \Gamma_1^+f+m_-(-z)\Gamma_0^+f=0.
\end{equation}
\end{theorem}
\begin{proof}
Recall that the operator $A$ coincides with the coupling $A_1$ defined by~\eqref{eq:A1}.
Let $z$ be an arbitrary non-real number. By the definition of $A_1$ the equation $(A_1-z)f=h$ with $f \in \dom(S_1)$ and $h \in \cK$ is equivalent to the system
\begin{equation}\label{eq:syst}
(S_+^*-z)f_+ = h_+, \ \ -(S_-^*+z)f_-=h_-,
\end{equation}
where $f_+ = P_+ f$, $f_- = P_- f$, $h_+ = P_+ h$ and $h_- = P_- h$.
It follows from \eqref{eS*ds} and \eqref{eq:syst} that $f_+ \in
\dom(S_+^*)$ can be expressed as
\begin{equation} \label{efdecom+}
f_+ = \bigl(S_0^+ - z\bigr)^{-1}h_+ + c_+ \psi_+(z) \quad
\text{with some} \quad c_+ \in\nC.
\end{equation}
Similarly, $f_- \in \dom(S_-^*)$ can be expressed as
\begin{equation} \label{efdecom-}
f_- = -\bigl(S_0^- + z\bigr)^{-1}h_- + c_- \psi_-(-z) \quad
\text{with some} \quad c_- \in\nC.
\end{equation}
By~\eqref{esubg1} and~\eqref{eG1A0} we have
\begin{align*}
\Gamma_0^+f_+ & = c_+, \\
\Gamma_0^-f_- & = c_-, \\
\Gamma_1^+f_+ & = \Gamma_1^+\bigl(S_0^+ - z\bigr)^{-1}h_+ + c_+ m_+(z) \\
& = \bigl[ h_+, \psi_+(\co{z}) \bigr]_{\cK} + c_+ m_+(z), \\
\Gamma_1^-f_- & = -\Gamma_1^-\bigl(S_0^- + z\bigr)^{-1}h_- + c_- m_-(-z) \\
& = \bigl[ h_-,\psi_-(-z^*) \bigr]_{\cK} + c_- m_-(-z).
\end{align*}
As $f_+$ and $f_-$ satisfy the boundary conditions~\eqref{eq:Gamma-} in the definition of $\dom (A_1) = \dom (S_1)$ we get
\begin{align*}
c_+ - c_- & = 0, \\
\Bigl(
\bigl[ h_+, \psi_+(\co{z}) \bigr]_{\cK} + c_+ m_+(z) \Bigr) +
\Bigl(
\bigl[ h_-,\psi_-(-\co{z}) \bigr]_{\cK} + c_- m_-(-z)\Bigr)& = 0.
\end{align*}
For each $z$ such that $m_+(z)+m_-(-z)\ne 0$ the above system has a unique solution for $c_+, c_-$:
\begin{equation*}
c_+ = c_-
=-\frac
\bigl[ h_+, \psi_+(\co{z})\bigr]_{\cK} +
\bigl[ h_-, \psi_-(-\co{z})\bigr]_{\cK}
{m_+(z)+m_-(-z)} = \frac{-\bigl[ h, \psi(\co{z})\bigr]_\cK}{m_+(z)+m_-(-z)}.
\end{equation*}
Now, the preceding equations, \eqref{efdecom+} and \eqref{efdecom-} imply that whenever $m_+(z)+m_-(-z)\ne 0$, for arbitrary $h\in {\cK}$, the system \eqref{eq:syst} has a unique solution $f\in \dom(A_1)$ and \eqref{ereswA} holds. This also proves that the right hand side of~\eqref{eRwA} is a subset of the left hand side.
To prove the converse inclusion in~\eqref{eRwA} assume that $m_+(z)+m_-(-z)=0$ for some $z\in\nC\setminus\nR$. Then~\eqref{eWeyl} and \eqref{esubg1} imply that
\[
\begin{split}
\Gamma_0^+\psi_+(z)-\Gamma_0^-\psi_-(-z)&=1-1=0\\
\Gamma_1^+\psi_+(z)+\Gamma_1^-\psi_-(-z)&=m_+(z)+m_-(-z)=0.
\end{split}
\]
In view of~\eqref{eq:domwAHC} this means that $\psi_+(z) + \psi_-(-z) \in\dom(A_1)=\dom(S_1)$. Since
\[
(A_1-z)\psi(z) = (S_+^*-z)\psi_+(z) + -(S_-^*+z)\psi_-(-z) = 0,
\]
$z$ is an eigenvalue of $A_1$. That is $z \notin \rho(A_1)\setminus\nR$.
To prove~\eqref{eq:z_depend}, set $h=h_+\in\cK_+$. Then the first equality in ~\eqref{eq:z_depend} is implied by~\eqref{eq:syst}. Since~\eqref{efdecom-} takes the form $f_-=c_-\psi_-(-z)\in\ker(S_-^*+z)$, then
\[
\Gamma_1^-f_-=m_-(-z)\Gamma_0^-f_-=m_-(-z)\Gamma_0^+f_+.
\]
In view of~\eqref{eq:Gamma-} this proves the second equality in~\eqref{eq:z_depend}.
\end{proof}
\begin{corollary}\label{c:m+-}
Let the assumptions of Theorem~{\rm\ref{tKsc}} hold. Then
\begin{equation}\label{eq:m+-}
m_+(z)+m_-(-z) = 0 \quad \text{for all} \quad z \in \nC \setminus \nR
\end{equation}
if and only if the resolvent set of $A$ is empty. In this case
$\nC\setminus\nR\subset\sigma_p(A)$.
\end{corollary}
As we pointed out in Subsection~\ref{ss:non-Ks}, a nonnegative self-adjoint operator in Kre\u{\i}n space can have an empty resolvent set. Next we show that this cannot happen for a nonnegative partially fundamentally reducible operator.
\begin{corollary} \label{c:non-none}
The spectrum of a nonnegative partially fundamentally reducible operator $A$ in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$ is real.
\end{corollary}
\begin{proof}
Assume that $\rho(A)=\emptyset$. Since $A$ is nonnegative in $({\cK}, \kip_{{\cK}})$, then $S_+$ and $S_-$ are nonnegative in the Hilbert spaces $(\cK_+, \kip_{{\cK}})$ and $(\cK_-, -\kip_{{\cK}})$, respectively. Hence each of the Weyl functions $m_+$ and $m_-$ has at most 1 pole on $\nR_-$. By Corollary~\ref{c:m+-} the equality~\eqref{eq:m+-} holds, which implies that $m_+$ is a rational function with at most three poles on $\nR$. But this contradicts the assumption that $S_+$ is a densely defined operator.
\end{proof}
The claim in the last corollary can also be deduced from \cite[Corollary~2.5]{ABT}. We notice that for a class of indefinite Sturm-Liouville operators the nonemptiness of the resolvent set was proved in~\cite[Proposition~4.5]{KaMal07}, \cite[Proposition~3.1]{KaKost08} and \cite[Theorem~5.3]{KuTr11} using the known asymptotic behavior of the Titchmarsh-Weyl coefficients.
\section{Regularity of critical points} \label{sect:3}
Throughout this section we use the notation introduced in the paragraph preceding Theorem~\ref{prop:AasCoupl} and in Theorem~\ref{tKsc}. By Corollary~\ref{c:non-none} a nonnegative partially fundamentally reducible operator $A$ in a Kre\u{\i}n space has a nonempty resolvent set. Therefore $A$ has a spectral function with critical points $0$ and $\infty$. In this section we study these critical points in terms of the Weyl functions $m_+$ and $m_-$.
\subsection{$D$-properties of pairs of Nevanlinna functions}
The following necessary conditions for the regularity
of the critical points $0$ and $\infty$ of an indefinite
Sturm-Liouville operator in terms of the Titchmarsh-Weyl
coefficients were proved in~\cite[Theorem 3.4]{KaKost08} as an extension of the corresponding result for the real line established in~\cite[Corollary 5.3]{{KaMal07}}.
\begin{theorem}\label{thm:NesReg}
{\rm (\cite{KaMal07}, \cite{KaKost08})} Let $A$ be a nonnegative partially fundamentally reducible operator
in a Kre\u{\i}n space $(\cK, \kip_{\cK})$ and let $m_+$ and $m_-$ be Weyl functions introduced in Theorem~{\rm~\ref{tKsc}}. Then the following two statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{thm:NesReg-1}
If $\infty\not\in c_s(A)$, then the functions
\begin{equation}\label{eq:ratio2}
z \mapsto \frac{\im m_+({z})}{|m_+({z})+m_-(-{z})|}, \qquad z \mapsto
\frac{\im m_-({z})}{|m_+({z})+m_-(-{z})|}
\end{equation}
are defined on $\nC_+$ and are bounded on
$\Omega_R^\infty = \{{z}\in\nC_+:\,|{z}|>R\}$ for each $R>0$;
\item \label{thm:NesReg-2}
If $0\not\in c_s(A)$, then the functions in~\eqref{eq:ratio2} are defined on $\nC_+$ and are bounded on $\Omega_R^0 = \{{z}\in\nC_+:\,|{z}|<R\}$ for each $R>0$.
\end{enumerate}
\end{theorem}
Let us sketch the proof of (\ref{thm:NesReg-1}) from~\cite{KaKost08} using the notation of Theorem~\ref{tKsc}. Since $\infty\not\in c_s(A_1)$
the functions
$z \mapsto (\im{z})\|(A-{z})^{-1}\|$,
$z \mapsto (\im{z})\|(S_+ -{z})^{-1}\|$ and
$z \mapsto (\im{z})\|(S_- - {z})^{-1}\|$
are bounded on $\Omega_R^\infty$. Then, it follows
from~\eqref{ereswA} and the equalities
\[
\|\psi_+(z)\|_{\cK_+}^2
= \frac{\im m_+(z)}{\im {z}} \quad \text{and} \quad
\|\psi_-(z)\|_{\cK_-}^2
= \frac{\im m_-(z)}{\im z},
\]
that the ratio in~\eqref{eqDinf} is also bounded on
$\Omega_R^\infty$.
Clearly, this proof works for arbitrary nonnegative coupling $A$ in Theorem~{\rm\ref{prop:AasCoupl}}.
\begin{definition}
We say that a pair $(m_+,m_-)$ of two Nevanlinna functions has
$D_\infty$-property ($D_0$-property), if the function
\begin{equation} \label{eqDinf}
y \mapsto \frac{\im m_+(iy)+\im m_-(iy)}{\bigl|m_+(iy)+m_-(-iy)\bigr|},
\end{equation}
is bounded on the set $(1,\infty)$ ($(0,1)$, respectively). In particular,
a Nevanlinna functions $m_+$ is said to have
$D_\infty$-property ($D_0$-property), if the pair $(m_+,m_+)$ has $D_\infty$-property ($D_0$-property, respectively), that is, if
the function
\[
y \mapsto \frac{\im m_+(iy)}{\re m_+(iy)},
\]
is bounded on the set $(1,\infty)$ ($(0,1)$, respectively).
\end{definition}
It follows from Theorem~\ref{thm:NesReg} that $D_\infty$-property
($D_0$-property) is necessary for the conditions $\infty\not\in
c_s(A_1)$ ($0\not\in c_s(A_1)$, respectively).
\begin{corollary}\label{Dinf}{\rm (\cite{KaMal07}, \cite{KaKost08})}
Let $A$ be a nonnegative partially fundamentally reducible operator
in a Kre\u{\i}n space $(\cK, \kip_{\cK})$ and let $m_+$ and $m_-$ be Weyl functions introduced in Theorem~{\rm~\ref{tKsc}}.
Then the following statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item
If $\infty\not\in c_s(A)$, then the pair $(m_+, m_-)$ has the $D_\infty$-property.
\item
If $0\not\in c_s(A)$, then the pair $(m_+, m_-)$ has the $D_0$-property.
\end{enumerate}
\end{corollary}
For a nonnegative densely defined symmetric operator $S$ with defect numbers $(1,1)$ in a Hilbert space we can always choose a boundary triple relative to which the Weyl function of $S$ is a Stieltjes function. Therefore the following proposition is of interest.
\begin{proposition}\label{prop:DInf}
Let $m_+$ be a Stieltjes function which has $D_\infty$-property {\rm (}$D_0$-property{\rm )}. Then for every Stieltjes function $m_-$ the pair $(m_+,m_-)$ has $D_\infty$-property {\rm (}$D_0$-property, respectively{\rm )}.
\end{proposition} \begin{proof}
Since $m_\pm$ is a Stieltjes function, we have $\re m_\pm(iy)\ge 0$ for all $y\in\nR_+$. Therefore,
\begin{equation}\label{eq:Re_miy}
\re(m_+(iy)) \le \bigl|m_+(iy)+m_-(-iy)\bigr| \qquad \text{for all} \qquad y\in\nR_+.
\end{equation}
As $m_+$ has $D_\infty$-property there exists $C > 0$ such that
\[
\frac{\im m_+(iy)}{\re m_+(iy)} \le C \qquad \text{for all} \qquad y > 1.
\]
By \eqref{eq:Re_miy} we have
\[
\frac{\im m_+(iy)}{\bigl|m_+(iy)+m_-(-iy)\bigr|}
\le \frac{\im m_+(iy)}{\re m_+(iy)} \le C \qquad \text{for all} \qquad y > 1.
\]
Further, the inequality
\begin{equation*}\label{eq:Im_miy}
\bigl|\im \bigl(m_+(iy)- m_-(iy)\bigr) \bigr| \le \bigl|m_+(iy)+m_-(-iy)\bigr| \quad \text{for all} \quad y\in\nR_+
\end{equation*}
yields that for all $y > 1$ we have
\[
\begin{split}
\frac{\im
m_-(iy)}{\bigl|m_+(iy)+m_-(-iy)\bigr|}
&
\le \frac{\bigl|\im \bigl( m_-(iy)- m_+(iy) \bigr) \bigr|
{\bigl|m_+(iy)+m_-(-iy)\bigr|}
+ \frac{\im m_+(iy)}{\bigl|m_+(iy)+m_-(-iy)\bigr|} \\
& \le C+1 .
\end{split}
\]
The proof of the $D_0$-property is similar.
\end{proof}
Asymptotic behavior of Weyl functions has been investigated for many kinds of special symmetric operators. Some specific examples appear in Section~\ref{SecExe} below. The next proposition deduces $D$-properties from the asymptotic behavior of Nevanlinna functions introduced in Subsection~\ref{sSecAsy}.
\begin{proposition}\label{prop:DpsAsym}
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{i:DpsAsym-1}
If Nevanlinna functions $m_+$ and $m_-$ belong to $\cA_\infty$, then the corresponding pair $(m_+, m_-)$ has $D_\infty$-property.
\item \label{i:DpsAsym-2}
If Nevanlinna functions $m_+$ and $m_-$ belong to $\cA_0$, then the corresponding pair $(m_+, m_-)$ has $D_0$-property.
\end{enumerate}
\end{proposition}
\begin{proof}
Let $m_\pm \in \cA_\infty$. Proposition~\ref{prop:0-inf} applies. Let $C_0^\pm \in \nR\setminus\{0\}$, $\alpha_0^\pm \in (-1,1)$, and, as appropriate, $C_1^\pm \in \nR\setminus\{0\}$ and $\alpha_1^\pm \in (-1,0)$ be the corresponding coefficients appearing in \eqref{eq:mAsymInf-noC} and \eqref{eq:mAsymInf-C}. We first notice that as a consequence of the asymptotics in \eqref{eq:mAsymInf-noC} or \eqref{eq:mAsymInf-C} we have
\begin{align} \label{eq:Limfrm}
\lim_{y\uparrow +\infty} \frac{m_\pm(iy)}{y^{\alpha_0^\pm}} & = C_0^\pm i^{\alpha_0^\pm}, \\
\intertext{
which, in turn, implies
} \label{eq:LimfrIm}
\lim_{y\uparrow+\infty} \frac{\im m_\pm(iy)}{y^{\alpha_0^\pm}}
&= C_0^\pm \sin\bigl(\pi\alpha_0^\pm/2\bigr).
\end{align}
Let
\[
\alpha_0 = \begin{cases}
\max\bigl\{ \alpha_0^+ , \alpha_0^- \bigr\} & \ \text{if} \quad \max\{\alpha_0^+, \alpha_0^-\} \geq 0, \\[1pt]
\min\bigl\{ \alpha_0^+ , \alpha_0^- \bigr\} & \ \text{if} \quad \max\{\alpha_0^+, \alpha_0^-\} < 0.
\end{cases}
\]
and, further, set $\epsilon_\pm = 1 - \sgn(|\alpha_0 - \alpha_0^\pm|)$. By definition $\epsilon_+, \epsilon_- \in \{0,1\}$, and at least one of $\epsilon_+$, $\epsilon_-$ equals $1$.
Next we will calculate the limit
\begin{equation} \label{eq:DinfLimit}
\lim_{y\uparrow+\infty}
\frac{\im m_+(iy)+\im m_-(iy)}{\bigl|m_+(iy)+m_-(-iy)\bigr|}
\end{equation}
and doing so we prove (\ref{i:DpsAsym-1}). Using \eqref{eq:Limfrm} and \eqref{eq:LimfrIm} this limit can be calculated as follows:
\begin{equation} \label{eq:DinfLimitL}
\lim_{y\uparrow+\infty} \frac{\frac{\im m_+(iy)}{y^{\alpha_0}}
+ \frac{\im m_-(y)}{y^{\alpha_0}}}{\left|\frac{m_+(iy)}{y^{\alpha_0}}
+ \frac{m_-(-iy)}{y^{\alpha_0}}\right|}
=
\frac{
\epsilon_+ C_0^+ \sin\bigl(\pi\alpha_0^+/2\bigr)
+ \epsilon_- C_0^-\sin\bigl(\pi\alpha_0^-/2\bigr)
{\bigl| \epsilon_+ C_0^+i^{\alpha_0^+} + \epsilon_- C_0^-(-i)^{\alpha_0^-}\bigr|}.
\end{equation}
If exactly one of $\epsilon_+, \epsilon_-$ equals $1$, then the denominator in the last fraction is clearly positive. Otherwise, that is, if $\epsilon_+ = \epsilon_- =1$, we have $\alpha_0^+ = \alpha_0^-$ and the denominator is again clearly positive provided that $\alpha_0^+ = \alpha_0^- \neq 0$. If $\alpha_0^+ = \alpha_0^- = 0$, the denominator equals $\bigl|C_0^+ + C_0^-\bigl|$.
Thus, the limit in \eqref{eq:DinfLimit} is calculated to be the right-hand side of \eqref{eq:DinfLimitL} with the exception of the case when $\alpha_0^+ = \alpha_0^- = 0$ and $C_0^+ + C_0^- = 0$.
To calculate the limit in \eqref{eq:DinfLimit} when $\alpha_0^+ = \alpha_0^- = 0$ and $C_0^+ + C_0^- = 0$ we use the second term of the asymptotics in \eqref{eq:mAsymInf-C}. First we deduce
\begin{align*}
\lim_{y\uparrow+\infty} \frac{m_\pm(iy)-C^\pm}{y^{\alpha_1^\pm}} & = C_1^\pm i^{\alpha_1^\pm}, \\
\intertext{which clearly yields}
\lim_{y\uparrow+\infty} \frac{\im m_\pm(y)}{y^{\alpha_1^\pm}}
& = C_1^\pm \sin\bigl(\pi\alpha_1^\pm/2\bigr).
\end{align*}
As before, we introduce $\alpha_1 = \min\{\alpha_1^+, \alpha_1^-\}$ and $\epsilon_1^\pm = 1 - \sgn(|\alpha_1 - \alpha_1^\pm|) \in \{0,1\}$.
Since we assume that $C_0^+ + C_0^- = 0$, the limit in \eqref{eq:DinfLimit} equals
\begin{equation*}
\lim_{y\uparrow+\infty}\! \frac{\frac{\im m_+(iy)}{y^{\alpha_1}} + \frac{\im m_-(y)}{y^{\alpha_1}}}{\left|\frac{m_+(iy)-C_0^+}{y^{\alpha_1}}
+ \frac{m_-(-iy)-C_0^-}{y^{\alpha_1}}\right|}
=
\frac{
\epsilon_1^+ C_1^+ \sin\bigl(\pi\alpha_1^+/2\bigr) + \epsilon_1^- C_1^-\sin\bigl(\pi\alpha_1^-/2\bigr)
{\bigl| \epsilon_1^+ C_1^+i^{\alpha_1^+} + \epsilon_1^- C_1^-(-i)^{\alpha_1^-}\bigr|}.
\end{equation*}
Now recall that at least one of $\epsilon_1^-$, $\epsilon_1^+$ equals $1$, $\alpha_1^-,\alpha_1^+ \in (-1,0)$ and $C_1^-,C_1^+ \in \nR\!\setminus\!\{0\}$, to deduce that the denominator of the last fraction is always positive.
Thus, in each case we calculated the limit in \eqref{eq:DinfLimit}. This proves that the pair $(m_+,m_-)$ has the $D_\infty$-property.
The proof of (\ref{i:DpsAsym-2}) is similar.
\end{proof}
\subsection{Regularity of the critical points} \label{ss:RegCPs}
Let $A$ be a nonnegative partially fundamentally reducible operator
in a Kre\u{\i}n space $(\cK, \kip_{\cK})$. In this subsection we provide sufficient conditions in terms of $m_+$ and $m_-$ for the points $\infty$ or $0$ not to be singular critical points of the operator $A$.
In the next lemma we use Theorem~\ref{Veselic} to obtain criteria for the regularity of the critical points $0$ and $\infty$ for the operator $A_1$ defined by~\eqref{eq:A1} and~\eqref{eq:domwAHC}. To formulate these results we need the following family of operators. For arbitrary $\varepsilon, \eta$ such that $0< \varepsilon < \eta$ we define the operator
$
T_{\pm\pm}(\epsilon,\eta):\cK_\pm\to\cK_\pm
$
by
\begin{equation}\label{CritTpm}
\bigl( T_{\pm\pm}(\epsilon,\eta) \bigr)f_{\pm} := \int\limits_{\epsilon\le|y|\le\eta}
\frac{\bigl\langle f_\pm ,\psi_\pm(\mp iy) \bigr\rangle_\cK \,
\psi_\pm(\pm iy)}{m_+(iy)+m_-(-iy)} \, dy, \quad f_\pm \in \cK_\pm,
\end{equation}
and the operator $T_{\pm\mp}(\epsilon,\eta):\cK_\mp\to\cK_\pm$
by
\begin{equation}\label{CritTpm-+}
\bigl( T_{\pm\mp}(\epsilon,\eta) \bigr)f_\mp := \int\limits_{\epsilon\le|y|\le\eta}
\frac{\bigl\langle f_\mp ,\psi_\mp(\pm iy)\bigr \rangle_\cK \,
\psi_\pm(\pm iy)}{m_+(iy)+m_-(-iy)} \, dy, \quad f_\mp \in \cK_\mp.
\end{equation}
\begin{lemma}\label{CritDinf}
Let $A$ be a nonnegative partially fundamentally reducible operator in the Kre\u{\i}n space $(\cK, \kip_{\cK})$. Then the following statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{iCritDinf-1}
$\infty\not\in c_s(A)$ if and only if the operators $T_{++}(1,\eta)$, $T_{--}(1,\eta)$, $T_{+-}(1,\eta)$ and $T_{-+}(1,\eta)$
are uniformly bounded for $\eta\in(1,\infty)$;
\item \label{iCritDinf-2}
$0\not\in c_s(A)$ if and only if $\ker(A)=\ker(A^2)$ and the operators $T_{++}(\varepsilon,1)$, $T_{--}(\varepsilon,1)$, $T_{+-}(\varepsilon,1)$ and $T_{-+}(\varepsilon,1)$
are uniformly bounded for $\varepsilon\in(0,1)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Notice first that by Corollary~\ref{c:non-none} the operator $A$ has a nonempty resolvent set. Set $A_0 = S_{0}^+\oplus (-S_{0}^-)$, see Theorem~\ref{tHsc}(\ref{ietHsc0}) and \eqref{eq:A0}. Since $A_0$ is self-adjoint in the Hilbert space $(\cK, \ahip_{\cK})$, the family of operators
\[
\int\limits_{1\le|y|\le\eta} (A_0 - iy)^{-1}dy, \qquad \eta >1,
\]
is uniformly bounded. Recall from Theorem~\ref{tKsc} that $\psi(z) = \psi_+(z) + \psi_-(-z)$, $z \in \nCR$, and let $f \in \cK$ be arbitrary. By \eqref{ereswA} we have
\begin{equation*}
\begin{split}
&
\int\limits_{1\le|y|\le\eta}\left(\bigl(A - iy \bigr)^{-1}f -\bigl(A_0 - iy \bigr)^{-1}f\right)dy \\
& \rule{4cm}{0pt} = \int\limits_{1\le|y|\le\eta}
\frac{-\bigl[ f, \psi(-{iy})\bigr]_{\cK} \, \psi(iy)}{m_+(iy)+m_-(-iy)} dy \\
& \rule{4cm}{0pt} =\begin{pmatrix} T_{++}(1,\eta) & T_{+-}(1,\eta)
\\ T_{-+}(1,\eta) & T_{--}(1,\eta) \end{pmatrix}
\begin{pmatrix} -P_+ f \\ P_- f \end{pmatrix}.
\end{split}
\end{equation*}
Therefore, the uniform boundedness of the family of operators
\[
\int\limits_{1\le|y|\le\eta}
({A_1} - {iy})^{-1}d{y}, \qquad \eta > 1,
\]
is equivalent to the uniform boundedness of the families of
operators
\[
T_{++}(1,\eta), \ \ T_{--}(1,\eta), \ \ T_{+-}(1,\eta),
\ \ T_{-+}(1,\eta), \quad \eta\in(1,\infty).
\]
Now (\ref{iCritDinf-1}) follows from Theorem~\ref{Veselic}(\ref{iVeselic-1}).
The statement (\ref{iCritDinf-2}) is implied by the formula~\eqref{ereswA}
and Theorem~\ref{Veselic}(\ref{iVeselic-2}).
\end{proof}
\begin{theorem}\label{Main}
Let $A$ be a nonnegative partially fundamentally reducible operator in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$ and assume that Weyl functions $m_+$ and $m_-$ introduced in Theorem~{\rm~\ref{tKsc}} have $B_\infty$-property (see Definition~\ref{defBinf}). Then $\infty\not\in c_s(A)$ if and only if the pair $(m_+,m_-)$ has $D_\infty$-property.
\end{theorem}
\begin{proof}
For the necessity see Corollary~\ref{Dinf}.
To prove the sufficiency consider first $f_+$, $g_+\in\cK_+$ and find an upper bound for
\[
\Bigl|\bigl\la T_{++}(1,\eta)f_+,g_+\bigr\ra_{\cK} \Bigr|.
\]
By the definition of the generalized Fourier transform $F_+=F_{m_+}$ we have
\begin{equation*} \label{eqesT++}
\bigl\la T_{++}(1,\eta)f_+,g_+\bigr\ra_{\cK}
= \int\limits_{1\le|y|\le\eta}
\frac{(F_{+}f_+)(iy) \bigl((F_{+}g_+)(-iy)\bigr)^*}{m_+(iy)+m_-(-iy)}dy.
\end{equation*}
Since the pair $(m_+,m_-)$ has $D_\infty$-property, there exists $C_1 > 0$ such that
\begin{equation}\label{eq:D_infty+}
\bigl| m_+(iy)+m_-(-iy) \bigr| \geq C_1^{-1} \im m_+(iy) \quad \text{for all} \quad y > 1.
\end{equation}
Set $w_\pm(y)=\bigl(\im m_\pm(iy)\bigr)^{-1}$, $y\in\nR_+$.
Since $m_+$ has $B_\infty$-property, then in view of Definition~\ref{defBinf} and Proposition~\ref{prop:4.5} the mappings
\begin{align*}
G_{m_+,\infty}^+F_{+} &: f_+\mapsto (F_{+}f_+)(iy), \quad f_+\in\cK_+, \\
G_{m_+,\infty}^-F_{+} &: f_+\mapsto (F_{+}f_+)(-iy), \quad f_+\in\cK_+,
\end{align*}
are bounded from $\cK_+$ to $L^2_{w_{+}}(1,\infty)$ and hence there exists $C_2>0$, such that
\begin{equation}\label{eq:B_infty+}
\bigl\|G_{m_+,\infty}^+F_{+}f_+ \bigr\|_{L^2_{w_{+}}\!(1,\infty)}
\le C_2 \bigl\|f_+\bigr\|_{\cK}
\end{equation}
and
\begin{equation}\label{eq:B_infty-}
\bigl\|G_{m_+,\infty}^-F_{+}f_+\bigr\|_{L^2_{w_{+}}\!(1,\infty)}
\le C_2 \bigl\|f_+\bigr\|_{\cK}
\end{equation}
for all $f_+\in\cK_+$.
Using \eqref{eq:D_infty+}, \eqref{eq:B_infty+}, \eqref{eq:B_infty-} and~\eqref{eq:F_trans} we obtain
\begin{align*}
\Bigl|\bigl\la T_{++}(1,\eta)f_+,g_+\bigr\ra_{\cK} \Bigr|
& \leq C_1 \int\limits_{1\le|y|\le\eta} \
\bigl|(F_{+}f_+)(iy) \bigr| \, \bigl| (F_{+}g_+)(-iy) \bigr| w_{+}(y) dy \\
& \leq C_1
\left(\int\limits_{1\le|y|\le\eta} \bigl|(F_{+}f_+)(iy)\bigr|^2 w_{+}(y) dy\right)^{\!1/2} \\
& \rule{26mm}{0pt} \times \left(\int\limits_{1\le|y|\le\eta} \bigl|(F_{+}g_+)(iy)\bigr|^2 w_{+}(y) dy\right)^{\!1/2} \\
&\le 2C_1C_2^2 \bigl\|f_+\bigr\|_{\cK}\bigl\|g_+\bigr\|_{\cK},
\end{align*}
for all $f_+,g_+ \in \cK_+$. This proves that the family $T_{++}(1,\eta)$ is uniformly bounded for $\eta\in(1,\infty)$.
The proof for the families $T_{--}(1,\eta)$ and $T_{+-}(1,\eta)$ is similar.
Next, we consider $f_+\in\cK_+$ and $g_-\in\cK_-$ and give an upper bound for
\[
\Bigl|\bigl\la T_{-+}(\epsilon,\eta)f_+,g_-\bigr\ra_{\cK} \Bigr|.
\]
Since the function in \eqref{eqDinf} is bounded, then the function
\begin{equation*}
y \mapsto \frac{\sqrt{\im m_+(iy)} \sqrt{\im m_-(iy)}}{\bigl|m_+(iy)+m_-(-iy)\bigr|} \qquad \text{for all} \qquad y > 1,
\end{equation*}
is bounded as well. Therefore, there exists $C_3 >0$ such that
\begin{equation*}
\bigl|m_+(iy)+m_-(-iy)\bigr| \geq C_2 \sqrt{\im m_+(iy)} \sqrt{\im m_-(iy)} \quad \text{for all} \quad y > 1,
\end{equation*}
Consequently, with $F_- = F_{m_-}$, we have
\begin{align*}
\Bigl|\bigl\la T_{-+}(1,\eta)f_+,g_- \bigr\ra_{\cK} \Bigr|
& \le \int\limits_{1\le|y|\le\eta} \
\biggl| \frac{(F_{+}f_+)(iy) \bigl((F_{-}g_-)(-iy)\bigr)^*
{m_+(iy)+m_-(-iy)} \biggr| \ dy \\
& \rule{-5mm}{0pt}
\leq C_3 \!\!\!\!\! \int\limits_{1\le|y|\le\eta} \!\!\!\!
\bigl|(F_{+}f_+)(iy) \bigr| \bigl|(F_{-}g_-)(-iy)\bigr| \sqrt{w_{+}(y)} \sqrt{w_{-}(y)} dy \\
& \rule{-5mm}{0pt}
\leq 2C_2C_3 \ \bigl\| G_{m_+} F_{+}f_+ \bigr\|_{L_{w_{+}}^2\!(1,\infty)} \ \bigl\| G_{m_-} F_{-}f_- \bigr\|_{L_{w_{-}}^2\!(1,\infty)} \\
& \rule{-5mm}{0pt}
\le 2C_2C_3 \ \bigl\|f_+ \bigr\|_{\cK} \ \bigl\|g_-\bigr\|_{\cK}
\end{align*}
for all $f_+ \in \cK_+$ and all $g_- \in \cK_-$. This proves that the family $T_{-+}(1,\eta)$ is uniformly bounded for $\eta\in(1,\infty)$.
The uniform boundedness of $T_{+-}(1,\eta)$ is proved similarly. Now the statement is implied by Lemma~\ref{CritDinf}.
\end{proof}
\begin{theorem}\label{Main0}
Let $A$ be a nonnegative partially fundamentally reducible operator in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$ and assume that Weyl functions $m_+$ and $m_-$ introduced in Theorem~{\rm~\ref{tKsc}} have $B_0$-property. Then $0\not\in c_s(A)$ if and only if $\ker(A)=\ker(A^2)$ and the pair $(m_+,m_-)$ has $D_0$-property.
\end{theorem}
\begin{proof}
We prove sufficiency. For necessity see Corollary~\ref{Dinf}.
Since $m_+$ and $m_-$ have $D_0$-property, there exists $C_4 > 0$ such that
\begin{equation}\label{eq:D_infty+A}
\bigl| m_+(iy)+m_-(-iy) \bigr| \geq C_4 \im m_+(iy) \qquad \text{for all} \qquad y\in(0,1).
\end{equation}
Since $m_+$ has $B_0$-property, the mappings
\[
G_{m_+,0}^+F_{+}^+:f_+\mapsto (F_{+}f_+)(iy)\quad (f_+\in\cK_+)
\]
\[
G_{m_+,0}^-F_{+}^-:f_+\mapsto (F_{+}f_+)(-iy)\quad (f_+\in\cK_+)
\]
are bounded from $\cK_+$ to $L_{w_{+}}^2\!(0,1)$ and hence there exists $C_5>0$, such that
\begin{align}\label{eq:B_0+}
\bigl\|G_{m_+,0}^+F_{+}f_+\bigr\|_{L_{w_{+}}^2\!(0,1)}
& \le C_5 \bigl\|f_+ \bigr\|_{\cK}, \\
\label{eq:B_0-}
\bigl\|G_{m_+,0}^-F_{+}f_+\bigr\|_{L_{w_{+}}^2\!(0,1)}
& \le C_5 \bigl\|f_+\bigr\|_{\cK}
\end{align}
for all $f_+\in\cK_+$.
Using \eqref{eq:D_infty+A}, \eqref{eq:B_0+}, \eqref{eq:B_0-} and~\eqref{eq:F_trans} we obtain
\begin{align*}
\Bigl|\bigl\la T_{++}(\epsilon,1)f_+,g_+\bigr\ra_{\cK} \Bigr|
& \leq C_4 \int\limits_{\epsilon\le|y|\le 1} \bigl|(F_{+}f_+)(iy)
\bigr| \bigl|(F_{+}g_+)(-iy)\bigr| w_{+}(y) dy \\
&\le C_4\, \bigl\| G_{m_+,0}^+ F_{+}f_+ \bigr\|_{L^2_{w_{+}}\!(0,1)} \ \bigl\| G_{m_+,0}^- F_{+}g_+ \bigr\|_{L^2_{w_{+}}\!(0,1)}\\
&\le 2C_4C_5^2 \, \bigl\|f_+ \bigr\|_{\cK} \, \bigl\|g_+\bigr\|_{\cK}
\end{align*}
for all $f_+,g_+\in\cK_+$. This proves that the family $T_{++}(\epsilon,1)$ is uniformly bounded for $\epsilon\in(0,1)$.
The proof for the family $T_{--}(\epsilon,1)$ is similar. Now the statement is implied by Lemma~\ref{CritDinf}.
\end{proof}
Proposition~\ref{p:fd-iff-rcp} together with Theorems \ref{Main} and \ref{Main0} yield the following statement.
\begin{corollary}\label{cor:Main0}
Let $A$ be a nonnegative partially fundamentally reducible operator in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$ and let $m_+$ and $m_-$ be Weyl functions introduced in Theorem~{\rm~\ref{tKsc}}. If each of the functions $m_+$ and $m_-$ has both $B_\infty$-property and $B_0$-property, then $A$ is fundamentally reducible in the Kre\u{\i}n space $(\cK, \kip_{\cK})$ if and only if the pair $(m_+,m_-)$ has $D_\infty$-property and $D_0$-property and $\ker(A)=\ker(A^2)$.
\end{corollary}
\begin{remark}\label{rem:Reflex}
The coupling $A_1$ in~\eqref{eq:A1} will be called {\it
reflexive} if $m_+(z)=m_-(z)$.
In the reflexive case analogs of Theorems~\ref{Main} and \ref{Main0} for indefinite Sturm-Liouville operator were proven in~\cite{Kost13}. It was shown there that the $B$-property for $m_+$ is automatically satisfied if the corresponding $D$-property holds.
\end{remark}
This remark leads to the following sufficient conditions of regularity.
\begin{corollary}\label{cor:Main1}
Let $A$ be a nonnegative partially fundamentally reducible operator in a Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$ and assume that Weyl functions $m_+$ and $m_-$ introduced in Theorem~{\rm~\ref{tKsc}} are Stieltjes functions. Then the following statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\roman{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{i:cor:Main1-1}
If each function $m_+$ and $m_-$ has $D_\infty$-property, then $\infty\not\in c_s(A)$.
\item \label{i:cor:Main1-2}
If each function $m_+$ and $m_-$ has $D_0$-property and $\ker(A) = \ker(A^2)$, then $0\not\in c_s(A)$.
\item \label{i:cor:Main1-3}
If each of the functions $m_+$ and $m_-$ has both $D_\infty$-property and $D_0$-property, and $\ker(A) = \ker(A^2)$, then $A$ is fundamentally reducible in the Kre\u{\i}n space $({\cK}, \kip_{{\cK}})$.
\end{enumerate}
\end{corollary}
\begin{proof}
(\ref{i:cor:Main1-1}) Since each function $m_+$ and $m_-$ has $D_\infty$-property then also each $m_+$ and $m_-$ has $B_\infty$-property (see Remark~\ref{rem:Reflex}). By Proposition~\ref{prop:DInf} the pair $(m_+,m_-)$ has $D_\infty$-property and hence $\infty\not\in c_s(A)$ by Theorem~\ref{Main}.
(\ref{i:cor:Main1-2}) Similarly, if each $m_+$ and $m_-$ has $D_0$-property then also each $m_+$ and $m_-$ has $B_0$-property (see Remark~\ref{rem:Reflex}). By Proposition~\ref{prop:DInf} the pair $(m_+,m_-)$ has $D_0$-property and thus $0\not\in c_s(A)$ by Theorem~\ref{Main0}.
\end{proof}
In the next theorem we use the notation introduced at the beginnings of Subsections~\ref{SubS:CoupleHS} and \ref{SubS:CoupleKS}.
\begin{theorem} \label{t:AllPos}
Let $S_\pm$ be a closed symmetric densely defined nonnegative operator with defect numbers $(1,1)$ in the Hilbert space $\bigl(\cK_\pm, \ahip_{\cK_\pm}\bigr)$. Let $\bigl(\nC,\Gamma_0^\pm,\Gamma_1^\pm\bigr)$ be a boundary triple for $S_\pm^*$ and let $m_\pm$ be the corresponding Weyl function. Let $A$ be a nonnegative self-adjoint extension of $S_+\oplus(-S_-)$ in the Kre\u{\i}n space $\bigl(\cK,\kip_\cK\bigr)$. Then the following statements hold.
\begin{enumerate}
\renewcommand*\theenumi{\alph{enumi}}
\renewcommand*\labelenumi{{\rm (\theenumi)}}
\item \label{i:AllPos-1}
If $m_+, m_- \in \cA_\infty$, then $\infty\not\in c_s(A)$.
\item \label{i:AllPos-2}
If $m_+, m_- \in \cA_0$ and $\ker(A)=\ker(A^2)$, then $0\not\in c_s(A)$.
\item \label{i:AllPos-3}
If $m_+, m_- \in \cA_0 \cap \cA_\infty$ and $\ker(A)=\ker(A^2)$, then $A$ is fundamentally reducible.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $A$ be an arbitrary nonnegative self-adjoint extension of $S_+\oplus(-S_-)$ in the Kre\u{\i}n space $\bigl(\cK,\kip_\cK\bigr)$. Then $A$ is a nonnegative fundamentally reducible operator in $\bigl(\cK,\kip_\cK\bigr)$, so $\rho(A) \neq \emptyset$ by Corollary~\ref{c:non-none}.
The operator $S = J\,A$ is a self-adjoint extension of $S_+\oplus S_-$ in the Hilbert space $\bigl(\cK,\ahip_\cK\bigr)$. Assume first that $S$ is a separated extension of $S_+\oplus S_-$. By the definition of a separated extension, there exists a self-adjoint extension $T_+$ of $S_+$ in $\bigl(\cK_+,\kip_{\cK_+}\bigr)$ and a self-adjoint extension $T_-$ of $S_-$ in $\bigl(\cK_-,\kip_{\cK_-}\bigr)$ such that $S = T_+\oplus T_-$. Then, $S$ commutes with the fundamental symmetry $J$ introduced in \eqref{eq:defJ}. Therefore, $A = JS$ is fundamentally reducible, so all three claims are trivial in this case.
Next assume that $S$ is a non-separated extension of $S_+\oplus S_-$. By Theorem~\ref{Thm:5.5} there exist a boundary triple $\bigl(\nC,\wh{\Gamma}_0^+, \wh{\Gamma}_1^+\bigr)$ for $S_+^*$ and
a boundary triple $\bigl(\nC,\wh{\Gamma}_0^-, \wh{\Gamma}_1^-\bigr)$ for $S_-^*$, such that $S$ is a coupling of $S_+$ and $S_-$ relative to the triples $\bigl(\nC,\wh{\Gamma}_0^+, \wh{\Gamma}_1^+\bigr)$ and $\bigl(\nC,\wh{\Gamma}_0^-, \wh{\Gamma}_1^-\bigr)$. By Remark~\ref{rem:transp} there exists a M\"{o}bius transformation $\mu_\pm$ such that the Weyl function $\wh{m}_\pm$ of $S_\pm$ relative to $\bigl(\nC,\wh{\Gamma}_0^\pm, \wh{\Gamma}_1^\pm\bigr)$ is given by
\begin{equation*}
\wh{m}_\pm = \mu_\pm \circ m_\pm.
\end{equation*}
We proceed with a proof of (\ref{i:AllPos-1}). Assume that $m_+, m_- \in \cA_\infty$. Proposition~\ref{prop:AInf-AZer}(\ref{i:AInf-AZer-1}) implies that $\wh{m}_+, \wh{m}_- \in \cA_\infty$ and, in turn, Corollary~\ref{co:Binfprop} yields that $\wh{m}_+$ and $\wh{m}_-$ have $B_\infty$-property. Since by Proposition~\ref{prop:DpsAsym}(\ref{i:DpsAsym-1}) the pair $(\wh{m}_-,\wh{m}_+)$ has $D_\infty$-property, the claim follows from Theorem~\ref{Main}.
Next assume that $m_+, m_- \in \cA_0$. As before, Proposition~\ref{prop:AInf-AZer}(\ref{i:AInf-AZer-1}) implies that $\wh{m}_+, \wh{m}_- \in \cA_\infty$. Now Corollary~\ref{th:B0prop} yields that $\wh{m}_+$ and $\wh{m}_-$ have $B_0$-property and by Proposition~\ref{prop:DpsAsym}(\ref{i:DpsAsym-2}) the pair $(\wh{m}_-,\wh{m}_+)$ has $D_0$-property. The claim now follows from Theorem~\ref{Main0}.
Statement (\ref{i:AllPos-3}) is a consequence of Corollary~\ref{cor:Main0}.
\end{proof}
\section{Examples} \label{SecExe}
\begin{example} \label{ex2ndq}
Consider the singular differential expression
\begin{equation} \label{eq2ndq}
\ell(f)(t):= -\frac{\sgn t}{w(t)} \Bigl( \bigl(p(t)\,f'(t)\bigr)' + q(t) f(t)\Bigr) \quad \text{for a.a.} \quad t \in \nR,
\end{equation}
where the coefficients $p$, $q$ and $w$ are real functions on $\nR$ satisfying the conditions
\begin{enumerate}
\item[(C1)]
$1/p, q, w \in L^1_{\rm loc}(\nR)$ and $p , w > 0$ a.e. on $\nR$,
\item[(C2)]
the expression $\ell$ is in the limit point case at $-\infty$ and at $+\infty$.
\end{enumerate}
It is natural to consider the expression in \eqref{eq2ndq} in the Kre\u{\i}n space $\bigl(L_{w}^2(\nR), \kip\bigr)$. Here $\bigl(L_{w}^2(\nR),\ahip\bigr)$ is the standard weighted $L^2$-space with the positive definite inner product $\ahip$ and the indefinite inner product is given by $[f,g] = \la Jf,g \ra, f, g \in L_{w}^2(\nR)$, where
\begin{equation*}
(Jf)(t)=(\sgn t) f(t), \qquad f \in L_{w}^2(\nR),
\end{equation*}
is a fundamental symmetry on $\bigl(L_{w}^2(\nR),\kip\bigr)$. Set
\begin{equation} \label{eq:L2pm}
\cK_\pm = \bigl\{ f \in L_{w}^2(\nR) : f = 0 \ \ \text{a.e. on} \ \ \nR_\mp \bigr\}.
\end{equation}
Then $L_{w}^2(\nR) = \cK_+ [\dot{+}]\cK_-$ is the fundamental decomposition corresponding to $J$.
Let $A$ be the operator associated with the expression in \eqref{eq2ndq} in the Hilbert space $L_{w}^2(\nR)$; that is the operator defined by $Af = \ell(f)$ for all
\[
f \in \dom(A) =
\bigl\{ f \in L_{w}^2(\nR) : f, pf' \in AC_{\rm loc}(\nR), \ \ell(f) \in L_{w}^2(\nR) \bigr\}.
\]
The differential operator $A$ is partially fundamentally reducible in the Kre\u{\i}n space $\bigl(L_{w}^2(\nR),\kip\bigr)$. To see this, consider the range restriction $S_\pm$ of $\pm A$ to $\cK_\pm$ which are defined on
\[
\bigl\{ f \in \dom(A) : Af \in \cK_\pm \bigr\} = \cK_\pm \cap \dom(A).
\]
Then $S_\pm$ is the minimal operator associated in $L_{w_\pm}^2\!(\nR_\pm)$ with the restriction of $\pm\ell$ to $\nR_\pm$; here $w_\pm$ denotes the restriction of $w$ to $\nR_\pm$. In fact we have,
\begin{align} \nonumber
\dom(S_\pm^*) & =
\bigl\{ f \in L_{w_\pm}^2\!(\nR_\pm): f, pf' \in AC_{\rm loc}[0,\pm\infty), \ \ell(f) \in L_{w_\pm}^2\!(\nR_\pm) \bigr\}, \\
\nonumber
\dom(S_\pm) & = \bigl\{ f \in \dom(S_\pm^*) : f(0) = f'(0) = 0\bigr\}, \\
\intertext{and}
\label{edefS}
S_\pm f &:= \pm \ell(f), \quad f \in \dom(S_\pm).
\end{align}
Since we assume that $\ell$ is in the limit point case at
$\pm\infty$, the operator $S_\pm$ is a densely defined symmetric operator with defect numbers $(1, 1)$ in the Hilbert space $L_{w_\pm}^2\!(\nR_\pm)$.
Let $z \in \nCR$ and denote by $\vartheta(\cdot,z)$ and
$\varphi(\cdot,z)$ the unique solutions of the equation
\begin{equation*}
-(p\,f')' + q f =z w f
\end{equation*}
satisfying the boundary conditions
\[
\vartheta(0,z) = 1, \ (p\,\vartheta')(0,z) = 0, \quad \text{and}
\quad \varphi(0,z) = 0, \ (p\,\varphi')(0,z) = 1,
\]
respectively. Since we assume that $\ell$ is in the limit point case at
$\pm\infty$, for each $z\in\nCR$ there is a unique solution
\begin{equation} \label{eTWeyl}
\psi_\pm(t,z) = \varphi(t,z) \mp m_\pm(z) \vartheta(t,z), \qquad t \in \nR_\pm,
\end{equation}
of the restriction of $\pm\ell(f) = z f$ to $\nR_\pm$ which belongs to $L_{w_\pm}^2\!(\nR_\pm)$. Relation
\eqref{eTWeyl} defines the function $m_\pm: \nCR\to \nC$ uniquely. The function $m_\pm$ is called {\em Titchmarsh-Weyl coefficient} of the restriction of the expression $\pm\ell$ to $\nR_\pm$.
A boundary triple for $S_\pm^*$ is $\bigl(\nC,\Gamma_0^\pm,
\Gamma_1^\pm\bigr)$, where
\begin{equation} \label{ebtS}
\Gamma_0^\pm f := (p\,f')(0\pm), \quad \Gamma_1^\pm(f) = \mp
f(0\pm), \quad f \in \dom(S_\pm^*).
\end{equation}
It follows from \eqref{eWeyl} and \eqref{ebtS} that the
Titchmarsh-Weyl coefficient defined by \eqref{eTWeyl} coincides with
the Weyl function of the operator $S_\pm$ in \eqref{edefS}
relative to the boundary triple in \eqref{ebtS}.
Assume additionally, that the coefficients $p$, $q$ and $w$ satisfy the
conditions:
\begin{enumerate}
\item[(C3)] $p> 0$ and $p'\in AC[-b,b]$ for all $b>0$;
\item[(C4)] $w = 1$.
\end{enumerate}
The following asymptotic for the Titchmarsh-Weyl coefficient $m_\pm$ has been established in~\cite{E72} (see also~\cite{Le}, where the asymptotic of the spectral function was found). For all $z \in \nCR$
\begin{equation}\label{eq_Friedr_Asymp}
m_\pm(r{z})=\frac{1}{\sqrt{p(0)}}\frac{1}{\sqrt{-{z} r}}+O\left(\frac{1}{r}\right)\qquad \text{as} \qquad r\to +\infty.
\end{equation}
It is clear that this asymptotic implies \eqref{eq:AsymInf}.
By Theorem~\ref{prop:AasCoupl}, the coupling $A_1$ of $S_+$ and $S_-$ relative to the boundary triples given in \eqref{ebtS} coincides with the differential operator $A$ associated with the expression $\ell$ in $L_{w}^2(\nR)$.
In addition, assume that the operator $S_\pm$ is nonnegative in the Hilbert space $L_{w_\pm}^2\!(\nR_\pm)$. Then the Friedrichs extension $S_F^\pm$ of $S_\pm$ defined on
\begin{equation*}
\dom\bigl(S_F^\pm\bigr) = \bigl\{ f \in \dom\bigl(S_\pm^*\bigr) : f(0) = 0 \bigr\}
\end{equation*}
is also nonnegative in the Hilbert space $L_{w_\pm}^2\!(\nR_\pm)$.
Since $\dom\bigl(S_F^\pm\bigr) = \ker(\Gamma_1^\pm)$, the function $m_\pm^{\!\top} = -1/m_\pm$ is holomorphic on $\nR_-\subset\rho(S_F^\pm)$ and satisfies
\begin{equation*}
\lim_{x\downarrow-\infty} m_{\pm}^{\!\top}(x) = -\infty.
\end{equation*}
By Corollary~\ref{col:nonneg} the self-adjoint operator $A = A_1$ is nonnegative in the Kre\u{\i}n space $\bigl(L_{w}^2(\nR),\kip\bigr)$ if and only if
\begin{equation}\label{eq:Noneg_A_1}
\lim_{x\uparrow 0}\bigl(m_+^{\!\top}(x)+m_-^{\!\top}(x)\bigr)\le 0.
\end{equation}
Assuming that $A = A_1$ is nonnegative, \eqref{eq:Noneg_A_1} yields $m_+^{\!\top}(0-), m_-^{\!\top}(0-) \in \nR$. Therefore, if $A_1$ is nonnegative, then $m_+^{\!\top}, m_-^{\!\top} \in \SM$, and consequently, $m_+, m_- \in \SM$. Now, \eqref{eq_Friedr_Asymp} and Proposition~\ref{prop:0-inf} imply that $m_+, m_- \in \cA_\infty$.
Further, by Corollary~\ref{co:Binfprop} the functions $m_+$ and $m_-$ have $B_\infty$-property and by Proposition~\ref{prop:DInf} $m_+$ and $m_-$ have $D_\infty$-property. Therefore, by Theorem~\ref{Main}, we have $\infty\not\in c_s(A)$. This result also follows from \cite[Theorem~3.6]{CL}, which was proved by completely different methods.
Moreover, Theorem~\ref{t:AllPos} yields that not only $A$, but an arbitrary nonnegative self-adjoint extension of $S_+\oplus(\!-S_-\!)$ in the Kre\u{\i}n space $\bigl(L_{w}^2(\nR),\kip\bigr)$ has a nonempty resolvent set and $\infty$ is not its singular critical point.
\end{example}
\begin{example}\label{ex:5.2}
Consider the differential expression \eqref{eq2ndq} with $p = w = 1$ and assume that the potential $q \in L^1_{\rm loc}(\nR)$ satisfies
\[
\int_\nR (1+|t|)|q(t)|dt < +\infty.
\]
These functions clearly satisfy (C1); (C2) follows from \cite[Problem~9.4]{CoLe} or \cite[Theorem~1]{Re}. Let the operator $S_\pm$ and the corresponding Titchmarsh-Weyl coefficient $m_\pm$ be as in Example~\ref{ex2ndq}. Assume that $S_\pm$ is a nonnegative operator in the Hilbert space $L^2(\nR_{\pm})$ and
that~\eqref{eq:Noneg_A_1} holds. Then the coupling $A_1$, which coincides with the differential operator $A$ associated with the expression \eqref{eq2ndq}, is a nonnegative self-adjoint operator in the Kre\u{\i}n space $\bigl(L^2(\nR),\kip\bigr)$.
The asymptotic behavior of $m_\pm$ at $0$ has been established in~\cite{KaKoMal09} as follows. It was shown that either there exists $k_\pm > 0$ such that for all $z \in\nC\setminus\nR$ we have
\begin{equation} \label{eq:as1+}
m_\pm(z) \sim - k_\pm \sqrt{-r z} \qquad \text{as} \qquad r \to 0,
\end{equation}
or there exist $a_\pm>0$ and $b_\pm\in\nR$ such that for all $z \in\nC\setminus\nR$ we have
\begin{equation} \label{eq:as2+}
m_\pm(rz) \sim \frac{a_\pm}{b_\pm + \sqrt{- r z}} \qquad \text{as} \qquad r \to 0.
\end{equation}
Assumption~\eqref{eq:Noneg_A_1} implies that the case \eqref{eq:as1+} is not possible. Thus \eqref{eq:as2+} holds. The asymptotic in \eqref{eq:as2+} implies that for all $z \in\nC\setminus\nR$ we have
\begin{equation} \label{eq:as3+}
m_\pm(rz) - \frac{a_\pm}{b_\pm} \sim -\frac{a_\pm}{b_\pm^2} \sqrt{- r z} \qquad \text{as} \qquad r \to 0.
\end{equation}
Since in Example~\ref{ex2ndq} we proved that $m_+, m_- \in \SM$, \eqref{eq:as3+} and Proposition~\ref{prop:0-inf} yield $m_+, m_- \in \cA_0$. Recall that in Example~\ref{ex2ndq} we proved that $m_+, m_- \in \cA_\infty$ and in \cite[Proposition~4.3]{KaKoMal09} it was proved that $\ker(A) = \{0\}$. Now Theorem~\ref{t:AllPos} implies that $A$ is fundamentally reducible. This has been proved in~\cite{KaKoMal09} and Theorem~\ref{t:AllPos} provides an alternative proof of this result. Under a stronger assumption on $q$ the fundamental reducibility of $A$ was proved in \cite{FS00} using a different approach.
However, Theorem~\ref{t:AllPos} implies more. Let $\wt{A}$ be an arbitrary nonnegative self-adjoint extension of $S_+\oplus(-S_-)$ in the Kre\u{\i}n space $\bigl(L^{2}(\nR),\kip\bigr)$. As in \cite[Proposition~4.3]{KaKoMal09}, it follows that $\ker(\wt{A}) = \{0\}$. Since $m_+, m_- \in \cA_0 \cap\cA_\infty$, Theorem~\ref{t:AllPos} implies that $\wt{A}$ is fundamentally reducible.
\end{example}
\begin{example}\label{ex:5.3}
Consider the differential expression \eqref{eq2ndq} with $q = 0$ and
\[
w(t) = \begin{cases}
(-t)^{\alpha_-}, & \text{if} \ \ t < 0, \\
\phantom{(}\rule{5pt}{0pt}t^{\alpha_+}, & \text{if} \ \ t > 0, \\
\end{cases}
\qquad \text{and} \qquad
p(t) = \begin{cases}
(-t)^{\beta_-}, & \text{if} \ \ t < 0, \\
\phantom{(}\rule{5pt}{0pt} t^{\beta_+}, & \text{if} \ \ t > 0, \\
\end{cases}
\]
with $\alpha_+, \alpha_- > -1$ and $\beta_+, \beta_- < 1$. In this example we use the notation introduced in Example~\ref{ex2ndq}. It is clear that the coefficients $p, q$ and $w$ satisfy condition (C1). In~\cite[Theorem~1]{EvZ78} it was proved that (C2) is also satisfied and that the function $m_\pm$ has the form
\[
m_\pm({z}) = C_\pm{(-{z})^{-\nu_\pm}},
\]
where
\[
\nu_\pm=\frac{1-\beta_\pm}{\alpha_\pm-\beta_\pm+2}, \quad C_\pm=\frac{(2k_\pm)^{2\nu_\pm}\Gamma(1+\nu_\pm)}{(1-\beta_\pm)\Gamma(1-\nu_\pm)}, \quad
k_\pm=\frac{\alpha_\pm-\beta_\pm + 2}{2}.
\]
As in Example~\ref{ex2ndq} the coupling $A_1$ of $S_+$ and $S_-$ relative to the boundary triples $\bigl(\nC, \Gamma_0^+,\Gamma_1^+\bigr)$ and $\bigl(\nC, \Gamma_0^-,\Gamma_1^-\bigr)$ given in \eqref{ebtS} coincides with the differential operator associated in $L_{w}^2(\nR)$ with the expression \eqref{eq2ndq} with $q=0$ and above $p$ and $w$.
It follows from Proposition~\ref{prop:nonneg} that $A_1$ is a nonnegative self-adjoint operator in the Kre\u{\i}n space $\bigl(L_{w}^2(\nR),\kip\bigr)$. Moreover, since the equation $\bigl(py'\bigr)'= 0$ does not have nontrivial solutions in $L_{w}^2(\nR)$, we have $\ker(A_1) = \{0\}$.
As $\nu_\pm\in(0,1)$ we have $m_+, m_- \in \cA_\infty \cap \cA_0$. Now, Theorem~\ref{t:AllPos} implies that $A_1$ is fundamentally reducible in the Kre\u{\i}n space $L_{w}^2(\nR)$.
In the case $\alpha_\pm = \beta_\pm = 0$ this result was proved in \cite{CN} and for $\alpha_+ = \alpha_-$ and $\beta_\pm = 0$ in~\cite{FN98}.
However, Theorem~\ref{t:AllPos} implies more. Let $\wt{A}$ be an arbitrary nonnegative self-adjoint extension of $S_+\oplus(-S_-)$ in the Kre\u{\i}n space $\bigl(L_{w}^{2}(\nR),\kip\bigr)$. Then, clearly, $\ker(\wt{A}) = \{0\}$ and Theorem~\ref{t:AllPos} yields that $\wt{A}$ is fundamentally reducible in the Kre\u{\i}n space $\bigl(L_{w}^2(\nR),\kip\bigr)$. In \cite{Kost06} similar results were obtained for the case $\alpha_+ = \alpha_-$ and $\beta_\pm = 0$.
\end{example}
\begin{example}
In this example we present a nonnegative coupling in a Kre\u{\i}n space which has a singular critical point at $0$. This example is modeled after the example in \cite[Section~5]{KaKost08}. It is proved in \cite{KaKost08} that for the expression in \eqref{eq2ndq} with $p=1$, $q = 0$ and $w(t) = (3|t|+1)^{-4/3}$ the corresponding Titchmarsh-Weyl coefficients $m_+$ and $m_-$ are $m_+(z) = m_-(z) = 1/\sqrt{-z}-1/z$. From this it was deduced in \cite{KaKost08} that $0$ is a singular critical point of the associated nonnegative differential operator in $(L_{w}^2(\nR),\kip)$. We think that it will be instructive to present a similar example as a coupling.
First define the Hilbert space $\cK_\pm = L^2(\nR_\pm)\oplus\nC$ with the inner product
\[
\left\la\! \left(\!\!\begin{array}{c} f_\pm \\ u_\pm \end{array}\!\! \right),
\left(\!\! \begin{array}{c} g_\pm \\ v_\pm \end{array}\!\!\right) \!\right\ra_\pm
\! := \! \int_{\nR_\pm} \!\! f_\pm(t) g_\pm^*(t) dt + u_\pm v_\pm^*, \ \begin{array}{l}
f_\pm, g_\pm \in L^2(\nR_\pm), \\ u_\pm, v_\pm \in \nC.
\end{array}
\]
In this space we consider the operator $S_\pm$ given by its graph
\[
\gr(S_\pm) = \left\{
\left\{
\left(\!\!\!\begin{array}{c} f_\pm \\ \mp f(0\pm) \end{array}\!\! \right),
\left(\!\!\!\begin{array}{c} - f_\pm'' \\ 0 \end{array}\!\!\right)
\right\} \, : \, \begin{array}{c}
f_\pm \in W^{2,2}(\nR_\pm) \\[3pt]
f'(0\pm) = 0
\end{array} \right\}.
\]
It is easy to see that $S_\pm$ is densely defined and positive. The graph of its adjoint is
\[
\gr(S_\pm^*) = \Biggl\{
\left\{
\left(\!\!\begin{array}{c} f_\pm \\ u_\pm \end{array}\!\! \right),
\left(\!\! \begin{array}{c} - f_\pm''(t) \\ f_\pm'(0\pm) \end{array}\!\!\right)
\right\} \, : \,
f_\pm \in W^{2,2}(\nR_\pm) \Biggr\}.
\]
A boundary triple $\bigl(\nC, \Gamma_0^\pm, \Gamma_1^\pm\bigr)$ for $S_\pm$ is
\[
\Gamma_0^\pm\left(\!\!\!\begin{array}{c}
f_\pm \\
u_\pm \\
\end{array}\!\!\!\right) = f_\pm'(0\pm),\qquad \Gamma_1^\pm \left(\!\!\!\begin{array}{c}
f_\pm \\
u_\pm \\
\end{array}\!\!\!\right)= \mp f_\pm(0\pm)-u_\pm .
\]
The corresponding Weyl solution is
\[
\psi_\pm(t,z) = \left(\!\!\begin{array}{c}
\mp \exp\bigl(\mp t \sqrt{-z} \bigr)/\sqrt{-z} \\
1/z \\
\end{array}\!\!\right), \qquad z \in \nC^+.
\]
That is,
\[
\Gamma_0^\pm \psi_\pm(\cdot,z) = 1 \quad \text{and} \quad S_\pm^* \psi(\cdot, z) = z \psi(z), \qquad
z \in \nC^+.
\]
Therefore the corresponding Weil function is
\[
m_\pm(z) = \frac{1}{\sqrt{-z}}-\frac{1}{z}, \qquad z \in \nC^+.
\]
The graph of the coupling $S_1$ of the symmetric operators $S_-$ and $S_+$ in the direct sum Hilbert space
\[
\cK = \cK_+ \oplus \cK_- = L^2(\nR) \oplus \nC^2
\]
is
\begin{equation*}
\left\{ \left\{ \begin{pmatrix} f \\ u_+ \\u_-
\end{pmatrix},
\begin{pmatrix} - f'' \\ f'(0+)\\ f'(0-) \end{pmatrix} \right\} :
\,\begin{array}{c}
f \in W^{2,2}(\nR_-)\oplus W^{2,2}(\nR_+) \\
f'(0+)=f'(0-), \\
f(0-)-f(0+)=u_-+u_+, \\
\end{array} \right\}.
\end{equation*}
By Corollary \ref{col:nonneg} $S_1$ is a nonnegative operator in $(\cK, \ahip)$.
The following indefinite inner product turns the Hilbert space $\cK = L^2(\nR) \oplus \nC^2$ into a Kre\u{\i}n space:
\[
\left[
\left(\!\!\begin{array}{c} f \\ u_+ \\ u_- \end{array}\!\! \right),
\left(\!\! \begin{array}{c} g \\ v_+ \\ v_- \end{array}\!\!\right) \right] := \int_\nR (\sgn t) f(t)g(t)^* dt + u_+v_+^* -u_-v_-^*
\]
for all $f, g \in L^2(\nR)$ and all $u_+,u_-,v_+,v_- \in \nC$. The direct sum $\cK = \cK_+\oplus \cK_-$ is a corresponding fundamental decomposition of $(\cK,\kip)$.
The graph of the coupling $A_1$ of the operators $S_-$ and $S_+$ is
\begin{equation} \label{edefwAHC4pm}
\left\{ \left\{ \begin{pmatrix} f(t) \\ u_+ \\u_-
\end{pmatrix},
\begin{pmatrix} - (\sgn t) f''(t) \\ f'(0+) \\ - f'(0-) \end{pmatrix} \right\} :
\,\begin{array}{c}
f \in W^{2,2}(\nR_-)\oplus W^{2,2}(\nR_+) \\
f'(0+)=f'(0-), \\
f(0-)-f(0+)=u_-+u_+, \\
\end{array} \right\}.
\end{equation}
It follows from~\eqref{edefwAHC4pm} that
$\ker(A_1)=\ker(A_1^2)$ and $\ker(A_1)$ is spanned by $( 0 \, \ 1 \ -1 )^{\!\top}$.
Since the pair of Weyl functions $m_+(z) = m_-(z) = 1/\sqrt{-z}-1/z$ does not have $D_0$-property, it follows from Corollary~\ref{Dinf} that $0$ is a singular critical point for $A_1$.
\end{example}
\begin{example}
Consider the following boundary value problem. For an arbitrary $g \in L^2(\nR_+)$ find $f \in L^2(\nR_+)$ such that
\begin{equation}\label{24Example}
-f''-zf=g,
\quad f'(0)-\sqrt{2}\sqrt[4]{z}f(0)=0.
\end{equation}
In this example we will show that a linearization of this boundary value problem in the Kre\u{i}n space $\bigl(L^2(\nR),\kip\bigr)$ is fundamentally reducible.
Here the Kre\u{\i}n space $\bigl(L^2(\nR),\kip\bigr)$ is the $L^2$ space introduced in Example~\ref{ex2ndq} with $w=1$. We also use the fundamental decomposition $L^2(\nR) = \cK_+ [\dot{+}]\cK_-$ with $\cK_+$ and $\cK_-$ from \eqref{eq:L2pm}.
Let $S_+$ be the minimal operator associated with the differential expression $-\frac{d^2}{dt^2}$ in the Hilbert space $\cK_+$ and let $\bigl(\nC,\Gamma^+_0,\Gamma^+_1\bigr)$ be the boundary triple for $S_+^*$ given by
\[
\Gamma^+_0f_+=f_+(0),\quad \Gamma^+_1f_+=f'_+(0).
\]
The Weyl function of $S_+$ relative to this boundary triple is $m_+(z)=-\sqrt{-z}$.
Let $S_-$ be the differential operator in the Hilbert space $\bigl(\cK_-,-\kip\bigr)$ defined on
\[
\dom(S_-) = \bigl\{ f_- \in W^{4,2}(\nR_-)\, : \, f_-(0) = f_-'(0) =
f_-''(0) = 0\bigr\}
\]
by $S_-f_- = f_-^{(4)}$. Then the adjoint operator $S_-^*$ is defined on
\[
\dom(S_-^*) = \bigl\{ f_- \in W^{4,2}(\nR_-)\, : \, f_-(0) = 0\bigr\}
\]
by the same expression $S_-^*f_- = f_-^{(4)}$. With
\[
\Gamma_0^-f_- = f_-'(0), \ \ \ \Gamma_1f_- = -f_-''(0), \ \ \ f_- \in \dom(S_-^*),
\]
the triple $\bigl(\nC,\Gamma_0^-,\Gamma_1^- \bigr)$ is a boundary
triple for $S_-^*$. It turns out that
\[
m_-({z})= -\sqrt{2}\sqrt[4]{-{z}}
\]
is the corresponding Titchmarsh-Weyl coefficient.
By Theorem~\ref{tKsc} the coupling $A_1$ of the operators $S_+$ and $S_-$ in the graph notation takes the form
\begin{equation*}
A_1=\left\{\left\{\begin{pmatrix}
f_+ \\
f_-
\end{pmatrix},
\begin{pmatrix}
-f_+'' \\
f_-^{(4)}
\end{pmatrix}
\right\}: \!\! \begin{array}{l}
f_+ \in W^{2,2}(\nR_+),\,f_+(0)-f_-'(0)=0\\
f_- \in W^{4,2}(\nR_-),\,
f_+'(0)-f''_-(0)=f_-(0)=0
\end{array}
\!\right\}
\end{equation*}
and the compressed resolvent $P_{+}(A_1-z)^{-1}|_{\cK_+}$ of $A_1$ gives a solution of the boundary value problem~\eqref{24Example} by the formula
\[
f=P_{+}(A_1-z)^{-1}g.
\]
The operator $A_1$ is called a linearization of the boundary value problem~\eqref{24Example}.
By Corollary~\ref{col:nonnegF} the operator $A_1$ is nonnegative. Since clearly $\ker(A_1)=\{0\}$ and the functions $m_+$ and $m_-$ belong to $\cA_\infty \cap \cA_0$, Theorem~\ref{t:AllPos} yields that the linearization $A_1$ of~\eqref{24Example} is fundamentally reducible.
\end{example}
|
\section{Random variables}
Broadly speaking---at the risk of haziness and looseness!---a \emph{random variable} is a mathematical representation of a
measurement (an experiment, trial, or observation) of some uncertain or unpredictable occurrence or value.
For instance, the random variable $Z$ could represent a single throw of a die,
so $Z$ represents possible outcomes $\{z=1, \ldots , z=6\}$ with probabilities $\{ \frac 1 6,\ldots ,\frac 16\}$.
Or it could represent measurement of a standard normal variable whose possible values are the real numbers $z \in \mathbf{R}$,
with standard normal probability distribution $\mathbf{N}(0,1)$.
Suppose the throw of the die yields a payoff or outcome $y=f(z)$ obtained by the following deterministic calculation:
\[
y = \left\{
\begin{array}{rl}
-1 & \mbox{if } z=1 \\
+1 & \mbox{if } z=6 \\
0 & \mbox{otherwise.}
\end{array}
\right.
\]
This particular experiment or game depends on (is \emph{contingent} on) the outcome of the experiment $Z$,
and can be denoted by $Y=f(Z)$. Where $Z$ has six possible outcomes, with a uniform probability distribution,
$Y$ has three possible outcomes whose probability distribution can easily be deduced by means of the deterministic calculation $f$.
The probability distribution\footnote{The probability distribution (``margin of error'')
carries the essential information specifying the character of the random variable or experiment.
It is often convenient to include other ``potential'' values or outcomes which are \textbf{not} actually possible or ``potential''.
For instance, in the die-throwing experiment we can declare that every real number is a potential outcome. In that case we assign probability
zero to the impossible outcomes. This does not change the random variable or its probability distribution in any essential way
that affects its mathematical meaning} of $Y$ is $y=-1$ with probability $\frac 1 6$, $y=+1$ with probability $\frac 1 6$, and $y=0$
with probability $\frac 2 3$.
We can easily invent such \emph{contingent} random variables or gambling games using more than one throw of the die,
and with payoff $Y$ dependent on some calculation based on the joint outcome of the successive throws.
This intuitive formulation is compatible with the formal and rigorous conception of a random variable as a $P$-measurable function
whose domain is a $P$-measurable sample space $\Omega$. This twentieth century injection of mathematical rigor brought
about a great extension of the depth and scope of the theory of probability and random variation,
including the development of many new spheres of application of the theory.
These applications often involve \emph{stochastic processes}.
Suppose $T$ is some set of indexing elements $\{s\}$. For instance, $T$ could be an interval of real numbers $[a,b]$.
A stochastic process $Y=Y_T$ is a family $Y=\left(Y(s)\right)_{s \in T}$, for which each element $Y(s), = Y_s$, is a
random variable. A sample path $\left(y(s)\right)_{s \in T}$ of the process $Y=Y_T$ can be thought of as a function
$y: T \mapsto \mathbf{R}$ in which, for each $s$, $y(s)$ (or $y_s$) is a possible outcome of the random variable (measurement, experiment, trial) $Y(s)$.
\section{Stochastic integrals}
Take $T=[0,t]$.
Equation (\ref{ito2}) above appears to be the result of applying an integration operation $\int_T$ to the equation (\ref{ito1}).
If this is the case, and if this step is justified, then comparison of
(\ref{ito1}) and (\ref{ito2}) implies (without delving into their actual meaning) that
\begin{equation}\label{most basic stochastic integral}
\int_T dY_s,=\int_0^t dY_s,=Y_t - Y_0;\;\;\mbox{or }\;\int_T dY(s),=\int_0^t dY(s),=Y(t) - Y(0).
\end{equation}
On the face of it, a clear and precise understanding of this simplest of all possible stochastic integrals would seem to be the \emph{sine
qua non} of this subject. Expressed as a stochastic differential equation, it is the tautology $dY_s = dY_s$.
Whatever (\ref{most basic stochastic integral}) actually means, it seems consistent enough with more familiar forms of integration
of the Stieltjes kind, in the somewhat loose and uncritical sense that the integral (or sum) of increments $dY$ gives an overall increment.
Advancing a little bit further, take a deterministic function $f$, and
consider $\int_T f(Y_s)dY_s$ (or $\int_T f(Y(s))dY(s)$), which is a more general version of $\int_T dY_s$.
If $y$ is a sample path of the process $Y$, the expression
\begin{equation}\label{first stochastic integral}
\int_T f(y(s)) dy(s)\;\;\;\mbox{ or }\;\;\int_T f(y_s) dy_s
\end{equation}
is a Stieltjes-type integral, which, if it exists, may be thought of as some limit of Riemann sums
$\sum f(y(s)) \Delta y(s)$ or $\sum f(y(s_j)) \left(y(t_j) - y(t_{j-1})\right)$,
where the finite set of points $t_j$ form a partition of the interval $T=[0,t]$, with
$t_{j-1} \leq s_j \leq t_j$ for each $j$.
From the point of view of basic mathematical analysis, unlike (\ref{most basic stochastic integral}) which is about ``margins of error'' in
probabilistic measurement,
there is nothing problematic about (\ref{first stochastic integral})---this Riemann-Stieltjes-type integral may or may not exist
for particular functions $y$ and $f$, but it is a fairly familiar subject for anyone who has studied basic Riemann-type integration.
In the Riemann sums for (\ref{first stochastic integral}),
{some applications require that $s_j =t_{j-1}$ for each $j$. Cauchy's approach to the theory of integration
used approximating sums with $s_j = t_{j-1}$ or $s_j = t_{j}$, so such sums can be called \emph{Cauchy sums} rather than Riemann sums.}
In any event, there are various ways, including the Lebesgue method,
in which we can seek to define an integral $\int_T f(y(s)) dy(s)$ for sample paths $y_T=(y(s))_{s \in T}$ of a stochastic process $Y = Y_T$.
Suppose a Stieltjes-type integral of $f(y(s))$ is calculated with respect to the increments $y(I) : = y(t_j) - y(t_{j-1})$ of the function $y_T$.
For instance, if $f$ is a function taking some fixed, real, constant value such as $1$, then a ``naive'' Riemann sum calculation
on the domain $T=[0,t]$, with $t_0=0$ and $t_n=1$
gives
$
\sum f(y(s)) y(I) \;\;=\;\;\sum_{j=1}^n y(I) \;\;=$
\[
= (\left(y(t_1) - y(0)\right) + (\left(y(t_2) - y(t_1)\right) + \cdots + (\left(y(1) - y(t_{n-1})\right)
=y(1) - y(0)\]
for \textbf{every} sample outcome $y_T$ of the process $Y_T$.
So it is reasonable---in some ``naive'' way---to claim that, for this particular function $f$,
the Riemann-Stieltjes integral exists for all outcomes $y_T$:
\[
\int_T f(y_s) dy_s =\int_0^t dy(s) = y(t)-y(0).
\]
One might then be tempted\footnote{A warning about this temptation is provided in Example \ref{Dirichlet 1} below.}
to apply such an argument to step functions $f$, and perhaps to try to extend it to some class of continuous functions $f$,
especially if we are only concerned with sample paths $y_T$ which are continuous.
But the key point here is that, given a stochastic process $Y=Y_T$, and given certain
deterministic functions $f$,
real values $\int_T f(y(s)) dy(s) $ can be obtained for each sample path $y=y_T$ by means of a recognizable Stieltjes integration procedure.
Can this class of real numbers or outcomes be related somehow to some identifiable random variable $Z$
which possesses some identifiable probability distribution (or ``margin of error'' estimates)?
If so, then $Z$ might reasonably be considered to be the random variable obtained by integrating,
in some Stieltjes fashion, the random variable $f(Y_s)$ with respect to the increments $Y(I) = Y(t_j) - Y(t_{j-1})$ of the stochastic process $Y_T$.
In other words, $Z$ is the stochastic integral
$\int_T f(Y_s) dY_s$.
To justify the latter step, a probability distribution (or ``margin of error'' data)
for $Z$ must be determined. But, in the case of the constant function $f$
given above ($f(y_s)=1$), this is straightforward. Because, with $f(y_s)=1$ for all outcomes $y_s$ in all sample paths (or joint outcomes) $y_T$,
the distribution function obtained for the Riemann sum values
$\sum f(y_s) y(I)$ is simply the known distribution function of the outcomes $y(t) - y(0)$ of the random variable $Y(t) - Y(0)$.
This distribution is the same for all partitions of $T=[0,t]$.
So it is reasonable to take it to be the distribution function of the stochastic integral $Z=\int_T f(Y_s) dY_s$.
For constant $f$ this seems to provide meaning and rationale for (\ref{first stochastic integral}).
What this amounts to is a naive or intuitive interpretation of stochastic integration which seems to hold for some elementary functions $f$.
This approach can be pursued further to give a straightforward interpretation---indeed, a ``proof''---of It\^{o}'s formula,
at least for the unchallenging functions $f$ mentioned above.
But what of the standard or rigorous theory of stochastic integration?
\section{Standard theory of stochastic integration}
Unfortunately, this theory
cannot accommodate the naive or intuitive construction of the simple stochastic integrals described in the preceding section.
Broadly speaking, the elementary Riemann sum type of calculation is not adequate for the kinds of analysis needed in this subject.
It is not possible, for instance, to apply a monotone convergence theorem, or a dominated convergence theorem,
to simple Riemann and Riemann-Stieltjes integrals. Historically, these kinds of analysis and proof have been supplied by Lebesgue-type integrals
which, while requiring a measure function as integrator, cannot be simply defined by means of the usual arrangement\footnote{But Section
\ref{-Complete integration} shows that Lebesgue integrals are essentially Riemann-Stieltjes integrals.}
of Riemann sums.
And this is where the difficulty is located. Suppose, for instance, that the stochastic process $Y_T$ that we are dealing with is a
standard Brownian motion. In that case any sample path $y_T$ is, on the one hand, almost surely continuous---which is ``nice'';
but, on the other hand, it is almost surely
\emph{not} of bounded variation in every interval $J$ of the domain $T=[0,1]$. And the latter is ``nasty''.
This turns out to be very troublesome if we wish to construct a Lebesgue-Stieltjes integral
using the increments $y(I), = y(t_j)-y(t_{j-1})$, of a sample path which is continuous but not of bounded variation in any interval.
The problem is that, in order to construct a Lebesgue-Stieltjes measure from the increments $y(I)$, we must separate the
non-negative increments $y_+(I)$ from the negative-valued increments $y_-(I)$,
\[
y(I) = y_+(I) - |y_-(I)|,
\]
and try to construct a non-negative measure from each of the components. But, because $y$ is not of bounded variation,
the construction for each component diverges to infinity on every interval $J$. Thus the standard theory of stochastic
integration encounters a significant difficulty at the very first step (\ref{first stochastic integral}).
To summarize:
\begin{itemize}
\item In the standard It\^{o} or Lebesgue integral approach, the most basic calculation of the integral of a constant function $f(Y_T)$,
with respect to the increments $dY$ of a Brownian process, fails because the Lebesgue-Stieltjes measure does not exist.
\item
On the other hand, if Riemann sums of the increments of the process $Y_T$ are used, then, by cancelation, a finite result is obtained
for each Riemann sum---a result which agrees with what is intuitively expected.
\end{itemize}
In the standard Lebesgue (or It\^{o}) theory of stochastic integration---in \cite{O} for instance---this
problem is evaded by \emph{postulating} a finite measure
$\mu_y(J)$ for each sample path, and then constructing a weak form of integral which, in the case of Brownian motion,
is based on certain helpful properties of this process.
The trouble with this approach is that it produces a quite difficult theory
which does not lend itself to the natural, intuitive interpretation described above.
However, elementary Riemann-sum-based integration is not generally considered to have the analytical power possessed by Lebesgue-style integration.
And a great deal of analytical power is required in the theory of stochastic processes. So at first sight it seems that
we are stuck with the standard theory of stochastic integration, along with all its baggage of subtlety and complication.
But this is not really the case. The good news is that is actually possible to formulate the theory of stochastic integrals
using Riemann sums instead of the measures of Lebesgue theory.
\section{Integration of functions}
To see this, it is first necessary to review the various kinds of integration which are available to us.
First consider the basic Riemann integral, $ \int_a^b f(s) ds$, of a real-valued, bounded, continuous function $f(s)$ on an interval $[a,b]$.
Let $\mathcal{P}$ be a partition of $[a,b]$;
\[
\mathcal{P}: \;\;\;a=t_0<t_1<t_2< \cdots < t_n=b,
\]
for any choice of positive integer $n$ and any choice of $t_j$, $1\leq j <n$.
For any $u<v$ and any interval $I$ with end-points $u$ and $v$, write $|I| =v-u$. Denoting intervals $]t_{j-1}, t_j]$ by $I_j$ let
\[
U_\mathcal{P}=\sum_{j=1}^n P_j|I_j|,\;\;\;\;\;\;L_\mathcal{P}=\sum_{j=1}^n p_j|I_j|
\]
where
\[
P_j = \sup \{f(s): s \in I_j \},\;\;\;\;\;\; p_j = \inf \{f(s): s \in I_j \}.
\]
\begin{definition}\label{def riemann}
Define the \emph{upper Riemann integral} of $f$ by
\[
U:=\inf\{L_\mathcal{P}: \mbox{ all partitions } \mathcal{P} \mbox{ of }[a,b]\},
\]
and the \emph{lower Riemann integral} of $f$ by
\[
L:=\sup\{l_\mathcal{P}: \mbox{ all partitions } \mathcal{P} \mbox{ of }[a,b]\}.
\]
Then $U_\mathcal{P} \geq L_\mathcal{P}$ for all $\mathcal{P}$, and if $U= L$ we say that $f$ is \emph{Riemann integrable}, with
\[
\int_a^b f(s) ds :=U=L.
\]
\end{definition}
Write the partition $\mathcal{P} $ as $ \{I\}$ where each $I$ has the form $I_j =\, ]t_{j-1}, t_j]$, with $|I_j| = t_j-t_{j-1}$, and Riemann sum
\[
(\mathcal{P})\sum f(s)|I| = \sum_{j=1}^n f(s_j)|I_j|.
\]
Suppose $g(s)$ is a real-valued, monotone increasing function of $s \in [a,b]$, so $g(s) \geq g(s')$ for $s>s'$.
For any interval $I$ with end-points $u$ and $v$ ($u<v$), define the increment or interval function $g(I)$ to be $g(v)-g(u)$.
\begin{definition}\label{def riemann-stieltjes 1}
If $|I|$ and $|I_j|$ are replaced by $g(I)$ and $g(I_j)$ in Definition \ref{def riemann} of the Riemann integral, then the resulting integral is called
the \emph{Riemann-Stieltjes integral} of $f$ with respect to $g$, $\int_a^b f\;dg$ or $\int_a^b f(s) dg(s)$.
\end{definition}
In fact if we start with the latter definition the Riemann integral is a special case of it,
obtained by taking the point function $g$ to be the identity function $g(s) =s$.
If $g(s)$ has bounded variation it can be expressed as the difference of two monotone increasing, non-negative point functions,
\[
g(s) = g_+(s) - (-g_-(s)),
\]
and the Riemann-Stieltjes integral of $f$ with respect to $g$ can then be defined as the
difference of the Riemann-Stieltjes integrals of $f$ with respect to $g_+$ and $-g_-$, respectively.
The following result is well known:
if real-valued, bounded $f$ is continuous and if real-valued $g$ has bounded variation then $\int_a^b f\;dg$ exists.
As suggested earlier, the Lebesgue integral of a real-valued point function $k$ with respect to a measure $\mu$ can be viewed,
essentially, as a Riemann-Stieltjes integral in which the
point-integrand $k(\omega)$ satisfies the condition of \emph{measurability}.
To explain this statement further,
consider a measure space $(\Omega, \mathcal{A}, \mu)$ with non-negative measure $\mu$ on a sigma-algebra $\mathcal{A}$
of $\mu$-measurable subsets of the arbitrary measurable space $\Omega$. Thus, if $\mu(\Omega) =1$, the measure space is a probability space.
Suppose the point-integrand $k$ is a bounded real-valued $\mu$-measurable function on the domain $\Omega$.
Then there exist real numbers $c$ and $d$ for which
\[
c \leq k(\omega) \leq d \;\;\;\mbox{ for all }\;\;\;\omega \in \Omega.
\]
Also, for each sub-interval $J$ of $[c,d]$, measurability of $k$ implies $\mu(k^{-1}(J))$ is defined.
The basic definition of the Lebesgue integral of $k$ with respect to $\mu$ on $\Omega$ is as follows.
\begin{definition}\label{def lebesgue}
Let $\mathcal{Q} = \{J_j\} =\{]v_{j-1},v_j]\}$
be a partition of $[c,d]$,
\[
\mathcal{Q}: \;\;\;c=v_0<v_1<v_2 < \cdots < v_n=d,
\]
and let
\[
L_\mathcal{Q}= \sum_{j=1}^n v_{j-1}\mu(k^{-1}(J_j)),\;\;\;\;\;\;U_\mathcal{Q}= \sum_{j=1}^n v_{j}\mu(k^{-1}(J_j)).
\]
Let $L:= \sup\{L_\mathcal{Q}: \mathcal{Q}\}$, $U:= \inf\{U_\mathcal{Q}: \mathcal{Q}\}$, the supremum and infimum being taken over all
partitions $\mathcal{Q}$ of $[c,d]$.
If $L=U$, then their common value is the \emph{Lebesgue integral} $\int_\Omega k(\omega)d\mu$.
\end{definition}
An advantage of Lebesgue integration over Riemann integration is that the former has theorems,
such as the dominated and monotone convergence theorems which,
under certain condition, make it possible for instance to change the order of integration and differentiation.
Also, Fubini's and Tonelli's theorems allow exchange of order of multiple integrals.
What makes ``good'' properties such as these possible is \emph{measurability} of the integrand $k$.
But the Lebesgue integral itself is, by definition, a Riemann-Stieltjes-type integral.
To see this, for each $u \in [c,d]$ define the monotone increasing function
\begin{equation}\label{lebesgue as riemann-stieltjes (1)}
g(u) = \mu\left(k^{-1}([c,u])\right),
\end{equation}
and take the point function $h(u)$ to be the identity function $h(u)=u$. Then the construction\footnote{The integral of a point function $h(u)$ with respect to a point function $g(u)$ can be addressed either as a Riemann-Stieltjes construction or as a Lebesgue-Stieltjes construction. When $h(u)=u$ and $g(u) = \mu\left(k^{-1}([c,u])\right)$ the former approach gives the Lebesgue integral $\int_\Omega k(\omega) d\mu$. On the other hand, if the Lebesgue-Stieltjes construction is attempted with $h(u)=u$ and $g(u) = \mu\left(k^{-1}([c,u])\right)$, we simply replicate the Riemann-Stieltjes construction of the Lebesgue integral $\int_\Omega k(\omega) d\mu$, and nothing new emerges.}
in Definition \ref{def lebesgue}
shows that
\begin{equation}\label{lebesgue as riemann-stieltjes (2)}
\int_\Omega k(\omega) d\mu \;\;\;=\;\;\;\int_c^d h(u) \;dg(u),\;\;\;=\;\;\;\int_c^d u\;dg.
\end{equation}
In other words, when combined with the measurability property of the point-integrand,
this particular Riemann-Stieltjes construction gives the ``good'' properties required in the integration of functions.
\section{Riemann definition}\label{Riemann definition}
But in fact a Riemann construction can give these ``good'' properties \textbf{without} postulating measurability in the
definition\footnote{And if measurability is
redundant in the definition, then so is the measure space structure.} of the integral.
To see this, we start again by considering a more general and more flexible definition of basic Riemann and Riemann-Stieltjes integration
which generalizes the construction of these integrals as given above in Definitions \ref{def riemann} and \ref{def riemann-stieltjes 1}.
The proposed, more general, definition of the Riemann-Stieltjes integral is applicable to real- or complex-valued functions $f$ (bounded or not);
and to real- or complex-valued functions $g$, with or without bounded variation.
\begin{definition}\label{def riemann-stieltjes 2}
The function $f$ is Rie\-m\-ann-Stieltjes integrable with respect to $g$, with integral $\alpha$, if, given $\varepsilon>0$,
there exists a constant $\delta>0$ such that, for every partition $\mathcal{P} = \{I\}$ of $[a,b]$ satisfying $|I| < \delta$ for each $I \in \mathcal{P}$,
the corresponding Riemann sum satisfies
\[
\left| \alpha - (\mathcal{P})\sum f(s)g(I) \right| < \varepsilon,
\]
so $\alpha = \int_a^b f\;dg$.
\end{definition}
If $g$ is the identity function $g(s)=s$ then Definition \ref{def riemann-stieltjes 2} reduces to the ordinary Riemann integral of $f$, $\int_a^b f(s)ds$.
{Definition \ref{def riemann-stieltjes 2} does not embody conditions which
ensure the existence of the integral. Such integrability conditions are not postulated but are deduced, in the form of theorems,
from the definition of the integral.}
Thus, if the function properties specified, respectively, in Definitions \ref{def riemann},
\ref{def riemann-stieltjes 1}, and \ref{def lebesgue} above are assumed, the integrability in each case follows from
Definition \ref{def riemann-stieltjes 2}; and Definitions \ref{def riemann},
\ref{def riemann-stieltjes 1}, and \ref{def lebesgue} become theorems of Riemann, Riemann-Stieltjes, and Lebesgue integration, respectively.
Definition \ref{def lebesgue} can now be expressed in terms of Definition \ref{def riemann-stieltjes 2}, using the formulations
(\ref{lebesgue as riemann-stieltjes (1)}) and
(\ref{lebesgue as riemann-stieltjes (2)}), and assuming measurability of the integrand $f$ with respect to measure space $(\Omega, \mathcal{A}, \mu)$.
\begin{definition}\label{def lebesgue 2}
The function $f$ is Lebesgue integrable with respect to measure $\mu$, with integral $\int_\Omega f(\omega)d\mu=\alpha$, if, given $\varepsilon>0$,
there exists a constant $\delta>0$ such that, for every partition $\mathcal{Q} = \{J\}$ of $[c,d]$ satisfying $|J| < \delta$ for each $J \in \mathcal{Q}$,
the corresponding Riemann sum satisfies
\[
\left| \alpha - (\mathcal{Q})\sum h(u)g(J) \right| < \varepsilon,
\]
where $h(u)=u$ is the identity function on $[c,d]$;
so $\alpha = \int_c^d h(u)dg(u)= \int_c^d u\;dg$.
\end{definition}
Thus, by definition, the Lebesgue integral $\int_\Omega f(\omega)d\mu$, with domain $\Omega$, is the Riemann-Stieltjes integral $\int_c^d u\;dg$,
with domain $[c,d]$.
The following result is an obvious
consequence of {Definition} \ref{def riemann-stieltjes 2}.
If $f$ has constant value $\beta$ and if $g$ is an arbitrary real- or complex-valued function,
then $\int_a^b f\;dg$ exists and equals $\beta(g(b) - g(a))$.
This follows directly from Definition \ref{def riemann-stieltjes 2} since, for every partition $\mathcal{P}$ of $[a,b]$, cancelation of terms gives
\[
(\mathcal{P})\sum f(s)g(I) = \beta\sum_{j=1}^n g(t_j) - g(t_{j-1}) = \beta\left(g(b) - g(a)\right).
\]
This result does not in general hold for Lebesgue-Stieltjes integration, as the latter
requires that $g(s)$ be resolved into its negative and non-negative components, $g(s) = g_+(s) - (-g_-(s)$,
and convergence may fail when the integral is calculated with respect to each of these components separately.
Example \ref{Dirichlet 1} below shows that, though constant functions $f$ are Riemann-Stieltjes integrable with respect to any integrator function $g$,
this does not necessarily extend to step functions $f$.
Definition \ref{def riemann-stieltjes 2} of the Riemann or Riemann-Stieltjes integral
does not postulate any boundedness, continuity, measurability
or other conditions for the integrand $f$. But, as already stated, in the absence of integrand measurability
and the construction in Definition \ref{def lebesgue}, this method of integration does not deliver
good versions of monotone and dominated convergence theorems, or Fubini's theorem.
\section{-Complete integration}\label{-Complete integration}
Developments in the subject since the 1950's---developments which were originated independently by R.~Henstock
and J.~Kurzweil---have made good this deficit in the basic Riemann and Riemann-Stieltjes construction. In this new development of the subject,
Definition \ref{def riemann-stieltjes 2} of the Riemann-Stieltjes integral is amended
as follows.
\begin{definition} \label{def stieltjes-complete}
A function $f$ is \emph{Stieltjes-complete} integrable with respect to a function $g$, with integral $\alpha$ if, given $\varepsilon>0$,
there exists a function $\delta(s) >0$ such that
\[
\left| \alpha - (\mathcal{P})\sum f(s)g(I) \right| < \varepsilon
\]
for every partition $\mathcal{P}$ such that, in each term $f(s)g(I)$ of the Riemann sum, we have $s-\delta(s) <u \leq s \leq v < s+\delta(s)$,
where $u$ and $v$ are the end-points of the partitioning interval $I$.
\end{definition}
In other words, where $|I|$ is less than a constant $\delta$ in the basic Riemann-Stieltjes definition,
we have $|I| < \delta (s)$ in the new definition.
Write $\alpha = \int_{[a,b]} f(s)g(I)$, or $\int_{[a,b]} f\;dg$, for the Stieltjes-complete integral whenever it exists.
Again, if the integrator function $g$ is the identity function $g(s)=s$, the resulting integral (corresponding to the basic Riemann integral), is the
\emph{Riemann-complete} integral of $f$, written $\alpha = \int_{[a,b]}f(s)|I|$, or $\int_{[a,b]}f(s)ds$.
The latter is also known as the Henstock integral,
the Kurzweil integral, the Henstock-Kurzweil, the generalized Riemann integral, or the \emph{gauge} integral since in this context the function
$\delta(s)>0$ is called a gauge.
It is obvious that every Riemann (Riemann-Stieltjes) integrable integrand is also Riemann-complete (Stieltjes-complete) integrable,
as the gauge function $\delta(s)>0$ of Definition \ref{def riemann-stieltjes 2} can be taken to be the constant $\delta >0$ of
Definition \ref{def riemann} and Definition \ref{def riemann-stieltjes 1}.
This argument indicates a \emph{Lebesgue-complete} extension of the Lebesgue integral, by replacing the
constant $\delta >0$ of Definition \ref{def lebesgue 2} with a variable gauge $\delta(u)>0$:
\begin{definition}\label{def lebesgue-complete}
Let $h(u)=u$ be the identity function on $[c,d]$.
The function $f$ is \textbf{Lebesgue-complete} integrable with respect to measure $\mu$, with integral $\int_\Omega f\;d\mu=\alpha$, if, given $\varepsilon>0$,
there exists a gauge $\delta(u)>0$ for $c \leq u \leq d$, such that
\[
\left| \alpha - (\mathcal{Q})\sum h(u)g(J) \right| < \varepsilon,
\]
for every partition $\mathcal{Q} = \{J\}$ of $[c,d]$ satisfying
\[u-\delta(u)<v_{j-1} \leq u \leq v_j < u+\delta(u)\]
for each $J =\,]v_{j-1},v_j]\in \mathcal{Q}$.
\end{definition}
In that case $\alpha = \int_{[c,d]} h(u)g(J)= \int_{[c,d]} u\;g(J)$, and the
Lebesgue-complete integral is a special case of the Stieltjes-complete integral---a special case in which a measure space structure exists and
for which the integrand is measurable.
So it is again clear that every Lebesgue integrable integrand is Lebesgue-complete integrable; since the former is, in effect,
a Riemann-Stieltjes integral, the latter is a Stieltjes-complete integral,
and every Riemann-Stieltjes integrable function is also Stieltjes-complete integrable. (No special notation has been introduced here
to distinguish the Lebesgue integral $\int_\Omega f\;d\mu$ from its Lebesgue-complete counterpart.)
If the measurable domain $\Omega$ is a real interval such as $[a,b]$, then some ambiguity arises in the interpretation of the Lebesgue integral
as an integral of the gauge, or generalized Riemann, kind. The reason for the ambiguity is as follows. Assuming the existence of the
Lebesgue integral $\int_{\Omega}f(\omega) d\mu, =\int_{[a,b]}f(\omega) d\mu$,
where $\omega$ now represents real numbers in the domain $[a,b]$, then we are assured of
the existence of the Stieltjes and Stieltjes-complete (or Lebesgue-complete) integrals $\int_c^d u\;dg$ and $\int_{[c,d]}u\;g(J)$,
respectively, with
\[
\int_{[a,b]}f(\omega) d\mu = \int_c^d u\;dg = \int_{[c,d]}u\;g(J),
\]
where the values $u, = h(u)$, are elements of $[c,d]$ and $h$ is the identity function on $[c,d]$.
But in this case, letting $\omega = s$ denote points of the domain $[a,b]$ and with $I$ denoting subintervals of $[a,b]$,
the function $\mu(I)$ is defined on intervals $I$, and two different Stieltjes-type constructions are possible.
First, there is the Riemann-Stieltjes integral $\int_c^du \;dg$ which defines the Leb\-esgue integral $\int_{\Omega}f(\omega) d\mu,
= \int_{[a,b]}f(\omega) d\mu$. Secondly, there is the gauge integral $\int_{[a,b]}f(s) \mu(I)$ which has a Stieltjes-complete construction.
It is then meaningful to consider whether, with $f$ measurable, existence of the Lebesgue integral
$\int_{[a,b]}f(\omega) d\mu$ implies existence of the Stieltjes-complete integral $\int_{[a,b]} f(s) \mu (I)$, and whether
\[
\int_c^du \;dg = \int_{[a,b]}f(s) \mu(I)
\]
holds,\footnote{There is a considerable literature on this question, which is usually answered as:
\emph{``Every Lebesgue integrable function on an interval of the real numbers $\mathbf{R}$ is also
Henstock-Kurzweil integrable.''} If the domain of the integrand is a measurable space $\Omega$ which is \textbf{not}
a subset of $\mathbf{R}$ or $\mathbf{R}^n$, then the appropriate way to formulate the corresponding Henstock-Kurzweil (or -complete)
integral is in the form $\int_{[c,d]} u\,g(J)$
described in Definition \ref{def lebesgue-complete}.}
the first of these integrals being the Lebesgue integral
$\int_{[a,b]}f(\omega) d\mu $, which, by Definition \ref{def lebesgue 2}, is interpreted as the Riemann-Stieltjes integral $\int_c^du \;dg$.
To see that these two integrals coincide, take $f$ to be a bounded, measurable function on $[a,b]$. This can be expressed as the difference of two
non-negative, bounded, measurable functions $f_+$ and $f_-$. Accordingly, and without loss of generality,
take $f$ to be non-negative, bounded, measurable. Then the Lebesgue integrable
function $f$ is the $\mu$-almost everywhere point-wise limit of a monotone increasing sequence of step functions $f_j$.
With $\omega = s$,
each step function $f_j$ is
Lebesgue integrable, with Lebesgue integral $\int_{[a,b]}f_j(\omega) \;d\mu$; and
each step function $f_j$ is Stieltjes-complete integrable, with Stieltjes-complete integral
$\int_{[a,b]}f_j(s) \mu(I)$, and
\[
\int_{[a,b]}f_j(\omega) \;d\mu = \int_{[a,b]}f_j(s) \mu(I)
\]
for each $j$. (This statement is also true if ``Lebesgue integral'' and ``Lebesgue integrability'' are replaced by ``Lebesgue-complete integral'' and ``Lebesgue-complete integrability''.)
By the monotone convergence theorem of Lebesgue integration (or, respectively, by the monotone convergence theorem of Lebesgue-complete integration),
\[
\int_{[a,b]}f_j(\omega) d\mu \rightarrow \int_{[a,b]}f(\omega) d\mu
\]
as $j \rightarrow \infty$. By the monotone convergence theorem of Stieltjes-complete integration, $f(s)\mu(I)$ is Stieltjes-complete integrable and
\[
\int_{[a,b]}f_j(s) \mu(I) \rightarrow
\int_{[a,b]}f(s) \mu(I)
\]
as $j \rightarrow \infty$. Since corresponding integrals of the pair of sequences are equal, their limits are equal:
\[
\int_{[a,b]}f(\omega) d\mu = \int_{[a,b]}f(s)\mu(I).
\]
This is the gist of a proof that existence of a Lebesgue integral (or of a Lebesgue integral) on a real domain implies existence of the corresponding Stieltjes-complete integral
on the same domain,
and equality of the two.
Thus the above argument can be applied to either the Lebesgue or the Lebesgue-complete integral on $\Omega = [a,b]$ in conjunction, respectively, with the corresponding Stieltjes-complete integral
on the same domain. In effect, if the domain $\Omega$ is a subset of $\mathbf{R}$, and if $f$ is Lebesgue integrable or Lebesgue-complete integrable with respect to $\mu$,
then $f(s)\mu(I)$ is also Stieltjes-complete integrable and the two integrals are equal.
The specific properties of the Lebesgue-complete integral have not been investigated.
As mentioned earlier, constant functions $f$ are Riemann-Stieltjes integrable,
and hence Stieltjes-complete integrable, with respect to any integrator function $g$.
But as the following counter-example shows, this does not necessarily extend to step functions $f$, or any other functions which are not constant.
\begin{example}
\label{Dirichlet 1}
\textbf{Dirichlet function:} For $0 \leq s \leq 1$ let $D(s)$ be $1$ if $s$ is rational, and $0$ otherwise. For $I=\,]u,v]$ let
$D(I) = D(v) - D(u)$. Let $D([0,v])= D(v) - D(0)$. The point function $D(s)$ is discontinuous everywhere, and has infinite variation on every interval
$J \subseteq [0,1]$. If $f(s)$ is constant for $0\leq s \leq 1$, then the Riemann-Stieltjes integral $\int_0^1 f(s)\,dD$ exists and equals
$D(1)-D(0)$; that is, $\int_0^1 f(s)\,dD=0$. \textbf{But if $f$ is not constant on $[0,1]$, then the Riemann-Stieltjes integral of $f$
with respect to $D$
does not exist.}
What about Stieltjes-complete integrability of $f(s)D(I)$?
In fact, if $f$ is not constant on $[0,1]$, then the Stieltjes-complete integral of $f$
with respect to $D$
does not exist.
This is proved in Theorem 1 of \cite{Henstock 1955}, and the proof is reproduced in Theorem 67 of \cite{MTRV}.
\textbf{Thus $f(s) D(I)$ is Riemann-Stieltjes integrable and
Stieltjes-complete integrable on $[0,1]$ if and only if $f(s)$ is constant for $0 \leq s \leq 1$.}
\end{example}
Historically this is the first published result (Theorem 1 of \cite{Henstock 1955}) in the theory of -com\-plete integration.
\section{-Complete approach to stochastic integrals}
Returning to stochastic integrals, the -complete method of integration allows us to construct Stieltjes-type Riemann sums for highly oscillatory
expressions which include both positive
and negative terms. Cancelation of terms can occur in the Riemann sum approximations, so the possibility of convergence is
preserved by this construction.
The Lebesgue construction, on the other hand, requires integral convergence, separately and independently, of the positive and negative components
of the integrand. The difficulty this presents is illustrated in the alternating or oscillating series $\sum_{j=1}^\infty (-1)^{-j} j^{-1}$. If the positive
and negative terms of the series are considered as two separate series then each of them diverges. But the series itself is conditionally
(or non-absolutely) convergent. Similarly, for sample paths $y(s)$ of a stochastic process $Y_T$ the integral $\int_{[0,t]}dy(s)$
does not generally exist when considered as a Lebesgue-Stieltjes integral. But it exists for all sample paths $y_T$, with value $y(t) - y(0)$, when
considered as a Stieltjes-complete integral.
There is no analytical cost or disadvantage in relinquishing the Lebesgue construction in favor of the -complete method. This is because
the important theorems of Lebesgue integration, such as monotone and dominated convergence, are also valid for the -complete approach.
Furthermore, there are other convergence theorems of a similar kind, specifically designed to deal with highly oscillatory functions
such as those which occur in the theory of stochastic processes but which are beyond the scope of the Lebesgue method.
See \cite{MTRV} for details of these.
However, stochastic integration includes novelties and challenges which have not yet been addressed in this essay.
For Brownian motion processes $X_T$, one of the most important stochastic integrals is $\int_0^t dX_s^2 =t$.
The corresponding integral for a sample path $x(s)$ ($0\leq s \leq t$) is ``$\int_0^t (dx(s))^2$''.
But this expression does not have the familiar form of a Stieltjes-type integral: $\int_a^b f(s)dg$,
which, when $g$ is the identity function, reduces to the even more familiar $\int_a^b f(s)ds$.
In Riemann sum approximation we are dealing with expressions $\sum (x(I))^2$, where, for $I=\,]u,v]$, $x(I) = x(v)-x(u)$.
But traditionally, while a Riemann sum for a Stieltjes integral involves terms $f(s)x(I)$ with integrator function $x(I)$
(in which $f(s)$ can be identically $1$),
we do not usually expect to see integrators such as $(x(I))^2$ or $dX_s^2$.
Another important stochastic integral Brownian motion theory is
\[
\int_0^t X_s dX_s = \frac 12 X_t^2 - \frac 12 t.
\]
For a sample path $x(s)$ of Brownian motion, this involves $\int_0^t x(s) dx(s)$, or, in Riemann sum terms, $\sum x(s)x(I)$.
The latter, as it stands, is a finite sum of terms $x(s)(x(v) - x(u))$ where $I=\,]u,v]$ and $u \leq s \leq v$.
And if we are using the Stieltjes-complete approach as described above, then we might suppose that each $s$ in the Riemann sum
is the special point used in partitions which are constrained by a gauge $\delta(s)$,
\[
s-\delta(s) < u \leq s \leq v < s+\delta(s).
\]
But in fact this is not what is required in the stochastic integral $\int_0^t X_s dX_s$. In Riemann sum format, what is required is
\[
\sum x(u) x(I),\;\;\;\mbox{ or }\;\;\;
\sum x(u) \left(x(v) - x(u)\right),
\]
where the first factor $x(u)$ in the integrand is a point function evaluated at the left hand end-point $u$ of the interval $I=\,]u,v]$.
Sometimes the form $\sum x(w)(x(v)-x(u))$ is used, with $w=u + \frac 12 (v-u)$.
In a way, integrands of form $x(I)^2$, $x(u)x(I)$, or $x(w)x(I)$, are an unexpected innovation.
Their value is calculated from the numbers $u$ and $v$ which specify the interval $I$.
So they can be thought of as functions $h(I)$ of intervals $I$.
But these functions are \textbf{not additive} on intervals.\footnote{If $h(I)$ were finitely additive on intervals $I$ it could
be used to define a point function $h(s):=h([0,s])$, and vice versa. Integrals with respect to finitely additive integrators are
therefore representable as Stieltjes-type integrals, and vice versa.}
In that regard they are unlike the integrators $|I|$ and $x(I)$ which are themselves functions of $I$ but are
finitely additive on intervals, in the sense that, if $J=I_1 \cup \ldots \cup I_n$ is an interval, then
\[
|J| = \sum_{j=1}^n |I_j|,\;\;\;\;\;\;x(J) = \sum_{j=1}^n x(I_j).
\]
Broadly speaking, integration is a summation process in which the summed terms involve functions of intervals.
Up to this point in this essay, the only integrands to be considered included a factor which was an additive function of intervals $I$, such as
the length function $|I|$
or the Stieltjes-type functions $g(I)$ or $x(I)$. But there is nothing inherent in the definition of -complete integrals that requires any
$I$-dependent factor in the integrand
to be additive.
With this in mind, consider again the definition of the -complete integral on an interval $[a,b]$.
Firstly, a \emph{gauge} is a function $\delta(s)>0$, $a\leq s\leq b$. Given $s$, an interval $I=\,]u,v]$
for which $s$ is either an end-point or an interior point, is
$\delta(s)$\emph{-fine}
if $s-u<\delta(s)$ and $v-s<\delta(s)$. A finite collection $\mathcal{D} = \{(s_1,I_1), \ldots , (s_n,I_n)$ is a \emph{division} of $[a,b]$
if each $s_j$ is either an interior point or end-point of $I_j$ and the intervals $I_j$ form a partition of $[a,b]$.
Given a gauge $\delta$, a division $\mathcal{D}$ is $\delta$-fine if each $(s_j,I_j) \in \mathcal{D}$ is $\delta$-fine.
Now suppose $h$ is a function of elements $(s,I)$. Examples include:
\[
h(s,I) =h_1(I)= |I|,\;\;\;h(s,I) =h_2(s)=s,\;\;\;h_3(s,I)=s^2|I|,\;\;\;h_4(I) = |I|^2.
\]
Given a division $\mathcal{D}=\{(s,I)\}$ of $[a,b]$ whose intervals $I$ form a partition $\mathcal{P}$, the corresponding Riemann sum is
\[
(\mathcal{D})\sum h(s,I), = \sum \{h(s,I): I \in \mathcal{P} \}.
\]
\begin{definition}
\label{def Burkill integral}
A function $h(s,I)$ is integrable on $[a,b]$, with integral $\int_{[a,b]}h(s,I) = \alpha$, if, given $\varepsilon>0$.
there exists a gauge $\delta(s)>0$ so that, for each $\delta$-fine division $\mathcal{D}$ of $[a,b]$,
\[
\left| \alpha - (\mathcal{D})\sum h(s,I) \right| < \varepsilon.
\]
\end{definition}
Applying this definition to the examples, $h_1$ is integrable with integral $b-a$, $h_2$ is not integrable, $h_3$ is integrable with integral
$\frac 13(b^3 - a^3)$, and $h_4$ is integrable with integral $0$. If $h(s,I) = h_5(I) = u^2|I|$ where, for each $I$, $u$ is the left hand end-point of $I$,
then it is not too hard to show that $h_5$ is integrable with integral $\frac 13(b^3 - a^3)$.
Actually, it is the traditional
custom and practice in this subject to only consider integrands $h(s,I) = f(s)p(I)$
where the integrator function $p(I)$ is a measure function or, at least, finitely additive on intervals $I$;
and where the evaluation point $s$
of the point function integrand $f(s)$ is the point $s$ of $(s,I)$ for each $(s,I)\in \mathcal{D}$.
When $p(I) = |I|$, this convention is needed in order to prove the Fundamental Theorem of Calculus.\footnote{The Fundamental Theorem of Calculus
states that if $F'(s)=f(s)$ then $f(s)$ is integrable on $[a,b]$ with definite
integral equal to $F(b)-F(a)$}.
But, while the Fundamental Theorem of Calculus is important in subjects such as differential equations, it hardly figures at all in some other branches
of mathematics such as probability theory or stochastic processes. And we have seen that stochastic integration often requires point integrands
$f(s)$ to be evaluated, not at the points $s$ of $(s,I) \in \mathcal{D}$, but at the left hand end-points of the partitioning intervals $I$.
So, with $I=\,]u,v]$, $f(u)$ is, in fact, an integrand function which depends, not on points $s$ but on intervals $]u,v]$.
These are a few of the ``unexpected innovations'' to be encountered in stochastic integration, giving it a somewhat alien and counter-intuitive feel
to anyone versed in the traditional methods of calculus.
For instance, the stochastic integral $\int_0^t X dX$ is given the value $\frac 12 X(t)^2 - \frac 12 t$ when the process $X(s)$ (with $X(0)=0$)
is a Brownian motion. Introductory treatments of this problem sometimes contrast the expression $\int_0^t X dX$
with the elementary calculus integral $\int x dx$ whose indefinite integral is $\frac 12 x^2$, in which the use of symbols $X$ and $x$
can, in the mind of an inexperienced reader, set up an inappropriate and misleading analogy.
In terms of sample paths, the stochastic integral $\int_0^t X(s) dX(s)$ has representative sample form $\int_0^t x(s) dx(s)$
which is a Stieltjes-type integral with integrator function $x(I), = x(v)-x(u)$, formed from a typically ``zig-zag'' Brownian path $x(s)$,
$0<s\leq t$, with $x(0)=0$. Then the notation for the contrasting elementary calculus integral is not $\int x \,dx$,
but $\int s \,ds$, with value $\frac 12 s^2$.
Putting the latter in Stieltjes terms, $\int s \,ds$ is the Stieltjes integral $\int_0^t x(s) dx(s)$ where the sample path path or function $x$ is the
identity function $x(s)=s$, $0\leq s\leq t$.
Clearly a Stieltjes integral involving a ``typical'' Brownian path $x(s)$ (which though continuous is, typically, nowhere differentiable)
is a very different beast from a Stieltjes integral involving the straight line path $x(s)=s$.
So in reality it is not surprising that there is a very big difference between the two integrals
\begin{equation}\label{weak stoch int 1}
\int X(s)dX(s)^2, =\frac 12 X(t)^2 - \frac 12 t,\;\;\;\mbox{ and }\;\;\;\int s\,ds, = \frac 12 s^2.
\end{equation}
The first integral typically involves Stieltjes integrals using very complicated and difficult Brownian paths $x(s)$.
It should be distinguished sharply from the more familiar and simpler Stieltjes integrals in which, for instance, the point function
component of the integrand is a continuous function, and the integrator or interval function is formed from increments of a monotone increasing
or bounded variation function.
It is easy to overlook this distinction.
Example 60 of \cite{MTRV} illustrates the potential pitfall. In this Example, $X_T$ is an arbitrary stochastic process
and, with a fixed partition of $T=\,]0,t]$,
$0=\tau_0<\tau_1< \cdots < \tau_m = t$, the function $\sigma(s)$ is constant for $\tau_{j-1} < s \leq \tau_j$. Example 60
claims, in effect, that the stochastic integral $\int_{\tau_{j-1}}^{\tau_j} \sigma(s)dX_s$ exists for each $j$ in the same way that, for constant $\beta$,
$\int_{\tau_{j-1}}^{\tau_j} \beta \;dX_s$ exists and equals $\beta(X(\tau_j) - X(\tau_{j-1}))$.
But Example \ref{Dirichlet 1} above shows that this claim is false. As a step function, $\sigma(\tau_{j-1})$ is not generally equal to the constant
$\beta = \sigma (s)$ when $s>\tau_{j-1}$. So if the sample path $x(s)$ is the Dirichlet function $D(s)$, the Stieltjes integral
$\int_{\tau_{j-1}}^{\tau_j} \sigma(s)dx(s)$ does not exist, and the claim in Example 60 is invalid.
However, if $X_T$ is a Brownian motion process, then each of the significant sample paths $x(s)$ satisfies a condition of uniform continuity.
In that case Example 60 is valid. But it requires some proof, similar to the proof of Theorem 229 on the succeeding page.
So what is truly surprising in (\ref{weak stoch int 1}) is, not that the two integrals give very different results,
but that any convergence at all can be found for the first integral.
Why is this so? This essay has avoided giving any precise meaning to expr\-essions such as $\int_0^t X dX$---or even to
a random variable $X_s$.
But the meaning of the random variable $\int_0^t X dX$ is somehow representative of a Stieltjes-type integral which can be formulated for \textbf{every}
sample path $\{x(s): 0 < s \leq t\}$. These sample paths may consist of joined-up straight line segments (as in the archetypical
jagged-line Brownian motion diagram), or smooth paths, or everywhere discontinuous paths (like the Dirichlet function). Thus any claim that
all of the separate and individual
Stieltjes integrals $\int_0^t x(s) dx(s)$ of the class of such sample paths $x$---a very large class indeed---have integral values
$\frac 12 x(t)^2 - \frac 12 t$ must be somehow
challenging and dubious.
The integrals $\int_0^t dX(s) = X(t)$, $\int_0^t dx(s) = x(t)$, show that each member of a large class of Stieltjes integrals \textbf{can}
indeed yield a common, single,
simple result. Our discussion of the Riemann sum calculation of these integrals illustrates how this happens:
regardless of the values of $x(s)$
for $s<t$, adding up increments ensures that \textbf{all} values $x(s)$ cancel out, except the terminal value $x(t)$.
Thus, if $f(s)$ takes constant value $\beta$ for $0 \leq s \leq t$, then, for every sample path $x(s)$, the Riemann-Stieltjes
(and Stieltjes-complete) integral $\int_0^t f(s) dx(s)$ exists, and $\int_0^t f(s) dx(s)= \beta x(t)$ (or $\beta(x(t) - x(0))$ if $x(0)\neq 0$.
This is the basis of the claim that the stochastic integral $\int_0^t f(s) dX(s)$ exists, and is the random variable $\beta X(t)$.
However, Example \ref{Dirichlet 1}
demonstrates that caution must be exercised in pursuing further the logic of Riemann sum
cancelation.
Because, if
the sample path
$x(s), = D(s)$, the expression $f(s)D(I)$ is not integrable on $[0,t]$,
in either the Riemann-Stieltjes sense or the Stieltjes-complete sense,
even when $f(s)$ is a step function (non-constant).
It is indeed possible to take the Riemann sum
cancelation idea further. Theorem 229 of \cite{MTRV} shows how this can be done.
But many important stochastic integrands are not actually integrable in the basic sense of the Definition \ref{def Burkill integral}.
If various sample paths $x(s)$ are experimented with in the integral $\int_0^t dX_s^2$, many different results will be found.
So what is the meaning of the result $\int_0^t dX_s^2 = t$?
While, for different sample paths $x$, $\int_0^t dx_s^2$ is not generally convergent to any definite value,
there is a weak sense of convergence of the integral which makes
``$\int_0^t dX_s^2 = t$'' meaningful. Most importantly in this case, the weak limit $t$ is a fixed quantity
rather than a random or unpredictable quantity
such as $x(t)$.
But this question goes beyond the scope of the present essay, whose aim is to explore some of the basic concepts of this subject, and hopefully to
illuminate them a little. A more extensive exploration is given in \cite{MTRV}.
|
\section{Introduction}
The classical Grunwald--Wang theorem is an example of a local--global (or Hasse) principle stating that except in some \textit{special} cases which are precisely determined, an element $m$ in a number field $\bK$ is an $a$-th power in $\bK$ if and only if it is an $a$-th power in the completion $\bK_{\wp}$ of $\bK$ for all but finitely many primes $\wp$ of $\bK$. In more concrete terms, the equation $x^a = m$ has a solution $x_{\bK}$ in $\bK$ if and only if it has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for all but finitely many primes $\wp$ of $\bK$. The aim of this article is to prove a Carlitz module analogue of the Grunwald--Wang theorem.
Throughout the paper, let $A = \bF_q[T], k = \bF_q(T)$, where $q$ is a power of a prime $p$, and let $k^{\alg}$ be the algebraic closure of $k$. Let $\tau$ be the map defined by $\tau(x) = x^q$. Let $k\langle \tau \rangle $ denote the twisted polynomial ring equipped with twisted multiplication, namely, $\tau a = a^q \tau$ for all $a \in k$. Let $C$ be the Carlitz module, i.e., $C : A \rightarrow k\langle \tau \rangle$ be an $\bF_q$-algebra homomorphism such that $C_T = T + \tau$. Note that for every commutative $k$-algebra $\D$, the action of $\sum_{j}a_i \tau^j \in k\langle \tau \rangle$ on an element $d \in \D$ is defined by
\begin{align*}
(\sum_{j}a_i \tau^j)(d) = \sum_{j}a_i d^{q^j}.
\end{align*}
A special case of an analogue of the Grunwald--Wang theorem for function fields was proved by G.-J. van der Heiden \cite{van-der-Heiden}. We now describe the result of van der Heiden. Let $\cX$ be a projective, smooth, geometrically irreducible curve over the finite field $\bF_q$, and let $\infty$ be a fixed closed point on $\cX$. Let $\bF_q(\cX)$ be the function field of $\cX$, and let $\cO_{\bF_q(\cX)}$ be the ring of functions in $\bF_q(\cX)$ that are regular outside $\infty$. Let $\cK$ be a finite separable extension of $\bF_q(\cX)$, and let $\iota : \cO_{\bF_q(\cX)} \rightarrow \cK$ be the natural embedding of $\cO_{\bF_q(\cX)}$ into $\cK$. Let $\Dr : \cO_{\bF_q(\cX)} \rightarrow \cK\langle \tau \rangle$ be a Drinfeld module over $\cK$ of rank 1, that is, $\Dr$ is an $\bF_q$-algebra homomorphism such that for all $f \in \cO_{\bF_q(\cX)}$, we have
\begin{align*}
\Dr_f = \sum_{j = 0}^{\deg(f)}f_i\tau^i,
\end{align*}
where the $f_i$ are elements in $\cK$ with $f_{\deg(f)} \in \cK^{\times}$ and $f_0 = \iota(f)$. In \cite{van-der-Heiden}, Van der Heiden proved the following result.
\begin{theorem}
\label{Theorem-Van-der-Heiden-theorem}
$(\text{van der Heiden, \cite[Proposition {\bf5}]{van-der-Heiden}})$
Let $P$ be a prime element of $\cO_{\bF_q(\cX)}$, and let $m$ be an element in $\cK$. Then the equation $\Dr_P(x) = m$ has a solution $x_{\cK}$ in $\cK$ if and only if it has a solution $x_{\wp}$ in the completion $\cK_{\wp}$ for every place $\wp$ of $\cK$.
\end{theorem}
Since $P$ is a prime element of $\cO_{\bF_q(\cX)}$ and the equation $\Dr_P(x) = m$ in Theorem \ref{Theorem-Van-der-Heiden-theorem} is considered over the field of definition of $\Dr$, namely $\cK$, we see that Theorem \ref{Theorem-Van-der-Heiden-theorem} might be viewed as analogous to the following special case of the classical Grunwald--Wang theorem: \textit{the equation $x^p = m$ with prime $p$ in $\bZ$ and $m \in \bQ$ has a solution $x_{\bQ}$ in $\bQ$ if and only if it has a solution $x_l$ in the $l$-adic field $\bQ_l$ for every prime $l$.}
To better understand the analogies between Theorem \ref{Theorem-Van-der-Heiden-theorem} and the above statement, one can consider the case when $\cX = \bP_1/\bF_q$. Thus $\cK = k, \cO_{\bF_q(\cX)} = A$, and the Drinfeld module $\Dr$ is the Carlitz module $C$ that was introduced earlier. There are well-known analogies \cite{Carlitz} \cite{Goss} \cite{Hayes} \cite{Rosen} \cite{Thakur-book} between the map $x \mapsto x^a$ with $a \in \bZ$ and the map $x \mapsto C_a(x)$ with $a \in A$; for example, adjoining torsion points of the Carlitz module to the field $k$ generates cyclotomic function fields in a similar manner as cyclotomic extensions of $\bQ$ obtained by adjoining roots of unity to $\bQ$. Hence Theorem \ref{Theorem-Van-der-Heiden-theorem} can be regarded as a special case of the classical Grunwald--Wang theorem.
In the hope of obtaining the best possible analogy with the classical Grunwald--Wang theorem, we will have to be content with working with the Carlitz module over $k$ because, among other reasons, the ring $\cO_{\bF_q(\cX)}$ might have non-principal ideals and the analogies between $(A, k)$ and $(\bZ, \bQ)$ are stronger. It is natural to search for a generalization of Theorem \ref{Theorem-Van-der-Heiden-theorem} that might be viewed as the full Grunwald--Wang theorem in the function field setting.
\begin{question}
\label{Question-An-analogue-of-the-Grunwald-Wang-theorem-for-function-fields}
Let $\bK$ be a finite separable extension of $k$. Let $a$ be an element in $A$ of positive degree, and let $m$ be an element of $\bK$. Consider the equation
\begin{align*}
C_a(x) = m.
\end{align*}
If $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for all but finitely many primes $\wp$ of $\bK$, is it true that the equation $C_a(x) = m$ has a solution $x_{\bK}$ in $\bK$?
\end{question}
An affirmative answer to Question \ref{Question-An-analogue-of-the-Grunwald-Wang-theorem-for-function-fields} should give rise to a result that is analogous to the classical Grunwald--Wang theorem in the number field case, and is a generalization of the result of van der Heiden in the Carlitz module setting. Indeed, in contrast to Theorem \ref{Theorem-Van-der-Heiden-theorem}, although the field of definition of the Carlitz module is $k$, we consider the equation $C_a(x) = m$ over any finite separable extension of $k$. Furthermore $a$ can be any element in $A$ of positive degree, and we \textit{only} require that the equation $C_a(x) = m$ is locally solvable at \textit{all but finitely many primes} of $\bK$. In comparison with the number field case, the classical Grunwald--Wang theorem considers global solvability of the equation $x^a = m$ over a number field although the power map $x^a$ with $a \in \bZ$ is defined over $\bQ$.
Let $\Lambda_a \subset k^{\alg}$ be the cyclic $A$-module defined by $\Lambda_{a} = \left\{\lambda \in k^{\alg} \; | \; C_a(x) = 0 \right\}$. The following theorem is a summary of the main results in this paper.
\begin{theorem}
\label{Theorem-A-summary-of-the-main-results}
Question \ref{Question-An-analogue-of-the-Grunwald-Wang-theorem-for-function-fields} has an affirmative answer in the following cases:
\begin{itemize}
\item [(i)] $\bK$ contains a primitive generator of the cyclic $A$-module $\Lambda_a$.
\item [(ii)] $\bK$ is a Galois extension of $k$ such that every prime $P$ dividing $a$ is unramified in $\bK$.
\end{itemize}
\end{theorem}
Part $(i)$ of Theorem \ref{Theorem-A-summary-of-the-main-results} is the content of Corollary \ref{Corollary-A-special-case-of-a-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}, and part $(ii)$ is the content of Theorem \ref{Theorem-A-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}. Note that Theorem \ref{Theorem-A-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem} and Corollary \ref{Corollary-A-special-case-of-a-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem} actually are more general than Theorem \ref{Theorem-A-summary-of-the-main-results}. More precisely, in the statements of Theorem \ref{Theorem-A-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem} and Corollary \ref{Corollary-A-special-case-of-a-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}, instead of assuming local solvability of the equation $C_a(x) = m$ at almost all primes in $\bK$, we only assume that $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for all primes $\wp$ in a set of primes of $\bK$ of Dirichlet density greater than $1 - 1/\phi(a)q^{\deg(a)}$. Here $\phi(a)$ is the function field analogue of the Euler $\phi$-function.
Let us say a few words about the ideas of the proof of Theorem \ref{Theorem-A-summary-of-the-main-results}. In \cite{Ono-Terasoma}, T. Ono and T. Terasoma attempted to give a different proof of a special case of the classical Grunwald--Wang theorem. Unfortunately the statement of a special case of the Grunwald--Wang theorem as well as its attempted proof given in \cite{Ono-Terasoma} were erroneous as pointed out by T. Ono \cite{Ono}. Despite the errors in \cite{Ono-Terasoma}, the underlying ideas in that paper can be modified to apply to the function field setting. We will carry out such modifications in the proof of Theorem \ref{Theorem-A-summary-of-the-main-results} whose main ingredients are the use of Tchebotarev density theorem and basic facts in class field theory over function fields. Thus the ideas of this paper are completely different from those used in \cite{van-der-Heiden} in which G.-J. van der Heiden extensively used Galois cohomology.
\section{The splitting field of the polynomial $C_a(x) - m$}
\label{Section-The-splitting-field-of-the-polynomial-C-a(x)-=-m}
Let $\bK$ be a finite, separable extension of $k$. Let $a$ be an element in $A$, and let $m$ be an element in $\bK$. In this section, we will describe the splitting field of the polynomial $C_a(x) - m \in \bK[x]$ over $\bK$ when the equation $C_a(x) = m$ is locally solvable at all primes in a set of primes of $\bK$ of Dirichlet density greater than $1 - 1/\phi(a)q^{\deg(a)}$. The case that the equation $C_a(x) = m$ has a solution $x_{\wp}$ in $\bK_{\wp}$ for almost all primes in $\bK$ follows immediately as a special case. The following result is very elementary and can be derived from the results in \cite[Chapter {\bf12}]{Rosen}. For the sake of self-containedness, we also include a proof here.
\begin{lemma}
\label{Lemma-Description-of-elements-in-C-a-lambda}
Let $a$ be a polynomial of positive degree $s$ in $A$. Let $\Lambda_a \subset k^{\alg}$ be the $A$-module defined by
\begin{align*}
\Lambda_a = \left\{\lambda \in k^{\alg} \; | \; C_a(\lambda) = 0 \right\},
\end{align*}
and let $\lambda_a$ be a generator for the cyclic $A$-module $\Lambda_a$ $(\text{see \cite[Chapter {\bf12}]{Rosen}})$. Then
\begin{itemize}
\item [(i)] $C_b(\lambda_a) \ne C_c(\lambda_a)$ for every $b, c \in A$ with $b \ne c$ and $\max(\deg(b), \deg(c)) < s$; and
\item [(ii)] $\Lambda_a = \{C_b(\lambda_a) \; | \; \text{$b \in A$ with $\deg(b) < s$}\}$.
\end{itemize}
\end{lemma}
\begin{proof}
Part $(ii)$ follows immediately from part $(i)$. We now prove part $(i)$. Let $b, c$ be distinct elements in $A$ with $\max(\deg(b), \deg(c)) < s$. Assume that $C_b(\lambda_a) = C_c(\lambda_a)$. It then follows that
\begin{align*}
C_{b - c}(\lambda_a) = 0.
\end{align*}
Since $b \ne c$, we see that $b - c \ne 0$, and hence $0 \le \deg(b - c) < s$. If $\deg(b - c) = 0$, we see that $b - c$ belongs to $\bF_q^{\times}$, and hence $C_{b - c}(\lambda_a) = (b - c)\lambda_a = 0$. Thus $\lambda_a = 0$, which is a contradiction.
If $\deg(b - c) > 0$, we know from \cite[Proposition {\bf12.4}]{Rosen} that $\Lambda_{b - c}$ has exactly $q^{\deg(b - c)}$ elements. Since $C_{b - c}(\lambda_a) = 0$ and $\lambda_a$ is a generator of the $A$-module $\Lambda_a$, we deduce that $\Lambda_a \subseteq \Lambda_{b - c}$. Thus
\begin{align*}
q^s = q^{\deg(a)} = \#\Lambda_a \le \#\Lambda_{b - c} = q^{\deg(b - c)},
\end{align*}
and therefore $s \le \deg(b - c)$, which is a contradiction. Therefore $C_b(\lambda_a) \ne C_c(\lambda_a)$.
\end{proof}
We now prove the main result in this section.
\begin{theorem}
\label{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}
Let $\bK$ be a finite, separable extension of $k$. Let $m$ be an element in $\bK$, and let $a$ be a polynomial of positive degree in $A$. Let $\bL$ be the splitting field of the polynomial $C_a(x) - m \in \bK[x]$ over $\bK$. Let $\lambda_a$ be a generator of the $A$-module $\Lambda_a$, where
\begin{align*}
\Lambda_a = \left\{\lambda \in k^{\alg} \; | \; C_a(\lambda) = m \right\}.
\end{align*}
Assume that the equation $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for all but finitely many primes $\wp$ of $\bK$. Then $\bL = \bK(\lambda_a)$.
\end{theorem}
\begin{proof}
Note that $\bK \subset k^{\alg}$ since $\bK$ is a finite extension of $k$. Let $h$ be an element in $k^{\alg}$ such that
\begin{align*}
C_a(h) = m.
\end{align*}
By Lemma \ref{Lemma-Description-of-elements-in-C-a-lambda}, we see that the set of all solutions of the equation $C_a(x) = m$ is given by
\begin{align}
\label{Equation-The-set-S-of-all-solutions-to-the-equation-C-a-(x)=m}
\cS = \{C_b(\lambda_a) + h \; | \; \text{$b \in A$ with $\deg(b) < \deg(a)$}\}.
\end{align}
Since the polynomial $C_a(x) - m$ is separable and $\bL$ is the splitting field of the polynomial $C_a(x) - m \in \bK[x]$, we deduce that $\bL = \bK(\cS)$ and $\bL$ is a Galois extension of $\bK$. Since $\lambda_a = (C_1(\lambda_a) + h) - (C_0(\lambda_a) + h)$, we see that $\lambda_a \in \bL$. We see that $h = C_0(\lambda) + h \in \bL$, and thus
\begin{align*}
\bL = \bK(\lambda_a, h).
\end{align*}
Let $\Gal(\bL/\bK)$ denote the Galois group of the Galois extension $\bL/\bK$. An element $\psi \in \Gal(\bL/\bK)$ is uniquely determined by its action on $\lambda_a$ and $h$. Hence to each element $\psi \in \Gal(\bL/\bK)$, we can associate a unique ordered pair $(b_{\psi}, u_{\psi}) \in (A/aA)^{\times} \times A/aA$ such that $\psi(\lambda_a) = C_{b_{\psi}}(\lambda_a)$ and $\psi(h) = C_{u_{\psi}}(\lambda_a) + h$. Let $\Psi : \Gal(\bL/\bK) \rightarrow \GL_2(A/aA)$ be the map defined by
\begin{align}
\label{Definition-The-representation-Psi-of-the-Galois-group-of-L-over-K}
\Psi(\psi) = \begin{pmatrix} b_{\psi} & u_{\psi} \\ 0 & 1 \end{pmatrix}
\end{align}
for each $\psi \in \Gal(\bL/\bK)$. Since $C$ is the Carlitz module, it is not difficult to see that for any elements $\sigma, \psi \in \Gal(\bL/\bK)$, the action of $\psi \circ \sigma$ on $\lambda_a$ and $h$ are given by
\begin{align*}
(\psi\circ\sigma)(\lambda_a) = C_{b_{\psi}b_{\sigma}}(\lambda_a) \quad \text{and} \quad (\psi\circ\sigma)(h) = C_{b_{\psi}u_{\sigma} + u_{\psi}}(\lambda_a) + h,
\end{align*}
where $(b_{\sigma}, u_{\sigma}), (b_{\psi}, u_{\psi}) \in (A/aA)^{\times} \times A/aA$ are uniquely determined by $\sigma, \psi$, respectively in the same way as above. Hence $\Psi$ is an injective homomorphism.
Let $\Sigma$ be the image of $\Gal(\bL/\bK)$ under $\Psi$. Set
\begin{align}
\label{Equation-The-set-E-in-GL2-A/aA}
\cE = \left\{\begin{pmatrix} \star & \star \\ 0 & 1 \end{pmatrix} \in \GL_2(A/aA) \right\}
\end{align}
and
\begin{align}
\label{Equation-The-set-M-in-GL2-A/aA}
\cM = \left\{\begin{pmatrix} 1 & \star \\ 0 & 1 \end{pmatrix} \in \GL_2(A/aA) \right\}.
\end{align}
We see that $\Sigma$ is a subgroup of $\cE$ and $\cE/\cM \cong (A/aA)^{\times}$. Hence we obtain the embedding
\begin{align}
\label{Equation-The-first-form-of-the-embedding}
\Sigma/\Sigma \cap \cM \hookrightarrow \cE/\cM \cong (A/aA)^{\times}.
\end{align}
We contend that $\Sigma \cap \cM = \left\{\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} \right\}$. Indeed, let $M = \begin{pmatrix} 1 & u \\ 0 & 1 \end{pmatrix}$ be an arbitrary element in $\Sigma \cap \cM$ for some element $u \in A/aA$. Then there exists an element $\psi \in \Gal(\bL/\bK)$ such that $\Psi(\psi) = M$. Let $(b_{\psi}, u_{\psi}) \in (A/aA)^{\times} \times (A/aA)$ be the ordered pair uniquely determined by
\begin{align}
\label{Equation-The-action-of-psi-on-lambda-a}
\psi(\lambda_a) = C_{b_{\psi}}(\lambda_a)
\end{align}
and
\begin{align}
\label{Equation-The-action-of-psi-on-h}
\psi(h) = C_{u_{\psi}}(\lambda_a) + h.
\end{align}
By the definition of $\Psi$, we see that
\begin{align*}
\begin{pmatrix} 1 & u \\ 0 & 1 \end{pmatrix} = M = \Psi(\psi) = \begin{pmatrix} b_{\psi} & u_{\psi} \\ 0 & 1 \end{pmatrix},
\end{align*}
and it thus follows that
\begin{align}
\label{Equation-The-value-of-b-psi}
b_{\psi} = 1
\end{align}
and
\begin{align}
\label{Equation-The-value-of-u-psi}
u_{\psi} = u.
\end{align}
By assumption, we know that there exists a finite set of primes of $\bK$, say, $\cP$ such that $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for every prime $\wp$ of $\bK$ with $\wp \not\in \cP$.
By the Tchebotarev density theorem (see \cite[Theorem {\bf9.13A}]{Rosen}), there exist infinitely many primes $\wp$ in $\bK$ such that $\wp$ is unramified in $\bL$ and $\psi$ is equal to the Frobenius automorphism $(\fB, \bL/\bK)$ for some prime $\fB$ in $\bL$ lying above $\wp$. Since the set $\cP$ is finite, one can choose a prime $\wp$ in $\bK$ with $\wp \not\in \cP$ such that $\wp$ is unramified in $\bL$ and $\psi = (\fB, \bL/\bK)$ for some prime $\fB$ in $\bL$ lying above $\wp$.
Let $\cD(\fB/\wp)$ be the decomposition group of $\fB$ over $\wp$. It is well-known \cite[Proposition {\bf9.8}]{Neukirch} that the fixed field of $\cD(\fB/\wp)$, or equivalently the decomposition field of $\fB$ over $\wp$, is $\bL \cap \bK_{\wp}$. Since $\wp \not\in \cP$, we know that there exists an element $x_{\wp}$ in $\bK_{\wp}$ such that $C_a(x_{\wp}) = m$. Hence $x_{\wp}$ belongs to $\cS$, where $\cS$ is given by $(\ref{Equation-The-set-S-of-all-solutions-to-the-equation-C-a-(x)=m})$. Thus $x_{\wp} = C_b(\lambda_a) + h$ for some $b \in A$ with $\deg(b) < \deg(a)$. In particular this implies that $x_{\wp} \in \bL \cap \bK_{\wp}$.
Since $\psi = (\fB, \bL/\bK) \in \cD(\fB/\wp)$ and $\bL \cap \bK_{\wp}$ is the fixed field of $\cD(\fB/\wp)$, we deduce that
\begin{align*}
\psi(x_{\wp}) = (\fB, \bL/\bK)(x_{\wp}) = x_{\wp}.
\end{align*}
On the other hand, we deduce from $(\ref{Equation-The-value-of-b-psi})$ and $(\ref{Equation-The-value-of-u-psi})$ that
\begin{align*}
C_b(\lambda_a) + h = x_{\wp} = \psi(x_{\wp}) &= \psi(C_b(\lambda_a) + h) \\
&= \psi(C_b(\lambda_a)) + \psi(h) \\
&= C_b(\psi(\lambda_a)) + C_u(\lambda) + h \\
&= C_b(C_1(\lambda_a)) + C_u(\lambda) + h \\
&= C_b(\lambda_a) + C_u(\lambda) + h,
\end{align*}
and thus $C_u(\lambda_a) = 0$.
If $u \ne 0$, then we deduce from Lemma \ref{Lemma-Description-of-elements-in-C-a-lambda} that $C_u(\lambda_a) \ne C_0(\lambda_a)$, and thus $C_u(\lambda_a) \ne 0$, which is a contradiction. Therefore $u = 0$, and hence
\begin{align}
\label{Equation-The-intersection-between-Sigma-and-cM-is-trivial}
\Sigma \cap \cM = \left\{\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} \right\}.
\end{align}
Let $\I$ be the subgroup of $\Gal(\bL/\bK)$ that corresponds to the subfield $\bK(\lambda_a)$ of $\bL$. In other words, $\bK(\lambda_a)$ is the fixed field of $\I$. Take any element $\sigma \in \I$. We know that there exists a unique ordered pair $(b_{\sigma}, u_{\sigma}) \in (A/aA)^{\times} \times A/aA$ such that $\sigma(\lambda_a) = C_{b_{\sigma}}(\lambda_a)$ and $\sigma(h) = C_{u_{\sigma}}(\lambda_a) + h$. Since $\sigma$ belongs to $\I$, we know that $C_{b_{\sigma}}(\lambda_a) = \sigma(\lambda_a) = \lambda_a = C_1(\lambda_a)$, and it thus follows from Lemma \ref{Lemma-Description-of-elements-in-C-a-lambda} that $b_{\sigma} = 1$. Hence we deduce that
\begin{align*}
\Psi(\sigma) = \begin{pmatrix} b_{\sigma} & u_{\sigma} \\ 0 & 1 \end{pmatrix} = \begin{pmatrix} 1 & u_{\sigma} \\ 0 & 1 \end{pmatrix},
\end{align*}
and hence $\Psi(\sigma) \in \Sigma \cap \cM$. By $(\ref{Equation-The-intersection-between-Sigma-and-cM-is-trivial})$, we deduce that $\Psi(\sigma) = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}$, and thus $u_{\sigma} = 0$. Hence $\sigma$ is the identity map, and therefore $\I =\{\text{id}_{\bL}\}$. Thus by Galois theory, we deduce that $\bL = \bK(\lambda_a)$, which proves our contention.
\end{proof}
\begin{remark}
\label{Remark-The-conditions-can-be-simplified-to-a-certain-positive-density-of-primes}
One of the key ingredients in proving Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation} is the use of Tchebotarev density theorem to prove that $\Sigma \cap \cM = \left\{\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}\right\}$. For an element $\psi \in \Gal(\bL/\bK)$, if we use the precise formula for the Dirichlet density of all unramified primes $\fp$ in $\bK$ such that the set $\left(\fp, \bL/\bK\right)$ of all Frobenius automorphisms equals the conjugacy class of $\psi$ in $\Gal(\bL/\bK)$, then we obtain a stronger result than Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. We will shortly prove a refinement of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation} using the Dirichlet density that leads us to the same conclusion as Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation} with less \textit{restrictive} assumptions.
\end{remark}
Let us first recall some basic notions and the precise statement of Tchebotarev density theorem as presented in \cite[Chapter {\bf9}]{Rosen}. Let $\bL/\bK$ be a Galois extension of global function fields. Let $\cP_{\bK}$ be the set of all primes in $\bK$, and let $\cA$ be a set of primes of $\bK$. If the limit given by
\begin{align}
\label{Equation-The-limit-in-the-definition-of-the-Dirichlet-density}
\lim_{s \rightarrow 1^{+}}\dfrac{\sum_{\wp \in \cA}{N\wp^{-s}}}{\sum_{\wp \in \cP_{\bK}}{N\wp^{-s}}}
\end{align}
exists, we define the \textit{Dirichlet density of $\cA$} to be the value of the limit given by $(\ref{Equation-The-limit-in-the-definition-of-the-Dirichlet-density})$, and denote it by $\delta(\cA)$. If the limit given by $(\ref{Equation-The-limit-in-the-definition-of-the-Dirichlet-density})$ does not exist, we say that $\cA$ does not have the Dirichlet density.
We now recall the statement of Tchebotarev density theorem.
\begin{theorem}
\label{Theorem-Tchebotarev-density-theorem}
$(\text{see \cite[Theorem {\bf9.13A}]{Rosen}})$
Let $\bL/\bK$ be a Galois extension of global function fields, and let $\Gal(\bL/\bK)$ denote the Galois group. Let $\cP^{\text{un}}_{\bK}$ be the set of primes of $\bK$ that are unramified in $\bL$. Let $\C \subset \Gal(\bL/\bK)$ be a conjugacy class in $\Gal(\bL/\bK)$. Let $\cH$ be the subset of $\cP^{\text{un}}_{\bK}$ defined by
\begin{align*}
\cH = \left\{\wp \in \cP^{\text{un}}_{\bK} \; | \; (\wp, \bL/\bK) = \C \right\}.
\end{align*}
Then
\begin{align*}
\delta(\cH) = \dfrac{\Card(\C)}{\Card(\Gal(\bL/\bK))},
\end{align*}
where $\Card(\cdot)$ denotes the number of elements in a set $(\cdot)$.
\end{theorem}
We now prove a refinement of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation} in which we only assume that the equation $C_a(x) = m$ is locally solvable at all primes $\wp$ in a set of the Dirichlet density greater than $1 - 1/\Card(\Gal(\bL/\bK))$.
\begin{theorem}
\label{Theorem-A-refinement-of-the-theorem-about-the-splitting-field-of-C-a-x-equals-m}
We maintain the same notation as in Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Let $\cH$ be a set of primes in $\bK$ such that the Dirichlet density of $\cH$ satisfies
\begin{align}
\label{Equation-The-lower-bound-of-the-Dirichlet-density-of-H}
\delta(\cH) > 1 - \dfrac{1}{\Card(\Gal(\bL/\bK))}.
\end{align}
Assume that the equation $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for each prime $\wp \in \cH$. Then $\bL = \bK(\lambda_a)$.
\end{theorem}
\begin{remark}
\label{Remark-The-absolute-lower-bound-for-the Dirichlet-density-of-H}
We can replace the lower bound of $\delta(\cH)$ in $(\ref{Equation-The-lower-bound-of-the-Dirichlet-density-of-H})$ by an absolute lower bound that only depends on $a$. Indeed, following the first part of the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}, we know that $\bL = \bK(\lambda_a, h)$. Note that this follows directly from the fact that $\bL$ is the splitting field of the polynomial $C_a(x) - m \in \bK[x]$ over $\bK$. By \cite[Theorem {\bf12.3.16}]{Villa-Salvador}, we know that $\lambda_a$ is a root of the $a$-th cyclotomic polynomial $\Phi_a(x) \in k[x] \subset \bK[x]$. We also know from \cite[Theorem {\bf12.3.16}]{Villa-Salvador} that the degree of $\Phi_a(x)$ is $\phi(a)$, where $\phi(a)$ is the number of nonzero polynomials in $A$ of degree less than $\deg(a)$ and relatively prime to $a$. (For a precise formula for $\phi(a)$, see \cite[Proposition {\bf1.7}]{Rosen}.) Thus
\begin{align*}
[\bK(\lambda_a) : \bK] \le \phi(a).
\end{align*}
Since the polynomial $C_a(x) - m \in \bK[x]$ is of degree $q^{\deg(a)}$ and $C_a(h) = m$, we deduce that
\begin{align*}
[\bL : \bK(\lambda_a)] = [\bK(\lambda_a)(h) : \bK(\lambda_a)] \le \deg(C_a(x) - m) = q^{\deg(a)},
\end{align*}
and therefore
\begin{align*}
\Card(\Gal(\bL/\bK)) = [\bL : \bK] = [\bL : \bK(\lambda_a)][\bK(\lambda_a) : \bK] \le \phi(a)q^{\deg(a)}.
\end{align*}
Thus
\begin{align*}
1 - \dfrac{1}{\Card(\Gal(\bL/\bK))} \le 1 - \dfrac{1}{\phi(a)q^{\deg(a)}}.
\end{align*}
We see from the above inequality that if $\delta(\cH) > 1 - \dfrac{1}{\phi(a)q^{\deg(a)}}$, then the condition $(\ref{Equation-The-lower-bound-of-the-Dirichlet-density-of-H})$ follows trivially.
\end{remark}
\begin{proof}[Proof of Theorem \ref{Theorem-A-refinement-of-the-theorem-about-the-splitting-field-of-C-a-x-equals-m}]
We maintain the same notation as in the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. As noted in Remark \ref{Remark-The-conditions-can-be-simplified-to-a-certain-positive-density-of-primes}, a crucial ingredient in the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation} is the use of Tchebotarev density theorem to prove that
\begin{align}
\label{Equation-The-intersection-of-Sigma-and-M-is-trivial-in-the-refinement-theorem}
\Sigma \cap \cM = \left\{\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}\right\}.
\end{align}
Looking closely at the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}, we see that once we establish $(\ref{Equation-The-intersection-of-Sigma-and-M-is-trivial-in-the-refinement-theorem})$, then it follows immediately that $\bL = \bK(\lambda_a)$. Hence it suffices to verify $(\ref{Equation-The-intersection-of-Sigma-and-M-is-trivial-in-the-refinement-theorem})$.
Take any element $M = \begin{pmatrix} 1 & u \\ 0 & 1 \end{pmatrix} \in \Sigma \cap \cM$ for some $u \in A/aA$. There exists an element $\psi \in \Gal(\bL/\bK)$ such that $\Psi(\psi) = M = \begin{pmatrix} 1 & u \\ 0 & 1 \end{pmatrix}$, where $\Psi$ is the injective homomorphism defined by $(\ref{Definition-The-representation-Psi-of-the-Galois-group-of-L-over-K})$. Since $\Psi$ is injective, we see from the definition of $\Psi$ that $\psi(\lambda) = C_1(\lambda_a) = \lambda_a$ and $\psi(h) = C_u(\lambda_a) + h$.
Let $\C_{\psi}$ be the conjugacy class of $\psi$ in $\Gal(\bL/\bK)$. Since $\psi \in \C_{\psi}$, we see that $\Card(\C_{\psi}) \ge 1$. Let $\cB_{\psi}$ be the set defined by
\begin{align*}
\cB_{\psi} = \left\{\wp \in \cP^{\text{un}}_{\bK} \; | \; (\wp, \bL/\bK) = \C_{\psi} \right\},
\end{align*}
where $\cP^{\text{un}}_{\bK}$ is the set of all primes of $\bK$ that are unramified in $\bL$. We prove that $\cH \cap \cB_{\psi} \ne \emptyset$. Assume the contrary, that is, $\cH \cap \cB_{\psi} = \emptyset$. Then
\begin{align}
\label{Equation-The-indentity-between-Dirichlet-densities}
\delta(\cH \cup \cB_{\psi}) = \delta(\cH) + \delta(\cB_{\psi}).
\end{align}
On the one hand, we see that $\delta(\cH \cup \cB_{\psi}) \le \delta(\cP_{\bK}) = 1$, where $\cP_{\bK}$ denotes the set of all primes in $\bK$. On the other hand, since $\Card(\C_{\psi}) \ge 1$, we deduce from Theorem \ref{Theorem-Tchebotarev-density-theorem} that
\begin{align*}
\delta(\cH) + \delta(\cB_{\psi}) &= \delta(\cH) + \dfrac{\Card(\C_{\psi})}{\Card(\Gal(\bL/\bK))} \\
&> \left(1 - \dfrac{1}{\Card(\Gal(\bL/\bK))}\right) + \dfrac{1}{\Card(\Gal(\bL/\bK))} \\
&= 1 \ge \delta(\cH \cup \cB_{\psi}),
\end{align*}
which is a contradiction to $(\ref{Equation-The-indentity-between-Dirichlet-densities})$. Therefore $\cH \cap \cB_{\psi} \ne \emptyset$, and it thus follows that there exists a prime $\fq$ in $\bK$ such that $\fq \in \cH$, $\fq$ is unramified in $\bL$ and the Frobenius automorphism $(\fB/\fq, \bL/\bK)$ equals $\psi$ for some prime $\fB$ in $\bL$ lying above $\fq$.
By assumption, we know that there exists an element $x_{\fq}$ in the completion $\bK_{\fq}$ of $\bK$ such that $C_a(x_{\fq}) = m$. Repeating the same arguments as in the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}, we see that
\begin{align*}
x_{\fq} = (\fB/\fq, \bL/\bK)(x_{\fq}) = \psi(x_{\fq}).
\end{align*}
From the above equation, one can follow the arguments in the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation} to prove that $C_u(\lambda_a) = 0$, and thus $u = 0$. Therefore $(\ref{Equation-The-intersection-of-Sigma-and-M-is-trivial-in-the-refinement-theorem})$ holds, and as remarked in the first paragraph of this proof, we deduce immediately that $\bL = \bK(\lambda_a)$.
\end{proof}
The following result follows immediately from Remark \ref{Remark-The-absolute-lower-bound-for-the Dirichlet-density-of-H}.
\begin{corollary}
\label{Corollary-The-description-of-the-splitting-field-in-terms-of-the-Dirichlet-density}
We maintain the same notation as in Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Let $\cH$ be a set of primes in $\bK$ such that the Dirichlet density of $\cH$ satisfies
\begin{align}
\label{Equation-The-lower-bound-of-the-Dirichlet-density-of-H-in-a-special-case-of-the-Grunwald-Wang-theorem}
\delta(\cH) > 1 - \dfrac{1}{\phi(a)q^{\deg(a)}}.
\end{align}
Assume that the equation $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for each prime $\wp \in \cH$. Then $\bL = \bK(\lambda_a)$.
\end{corollary}
By the above corollary, we obtain a Carlitz module analogue of the Grunwald--Wang theorem in the case that $\bK$ contains a generator of the $A$-module $\Lambda_a$.
\begin{corollary}
\label{Corollary-A-special-case-of-a-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}
We maintain the same notation as in Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Assume that $\bK$ contains a generator $\lambda_{\star}$ of the module $\Lambda_a$. Let $\cH$ be a set of primes in $\bK$ such that the Dirichlet density of $\cH$ satisfies
\begin{align*}
\delta(\cH) > 1 - \dfrac{1}{\phi(a)q^{\deg(a)}}.
\end{align*}
Then the equation $C_a(x) = m$ has a solution $x_{\bK}$ in $\bK$ if and only if it has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ of $\bK$ for every prime $\wp \in \cH$.
\end{corollary}
\begin{proof}
If $C_a(x) = m$ has a solution $x_{\bK}$ in $\bK$, then it is obvious that it is locally solvable at every prime of $\bK$.
If $C_a(x) = m$ is locally solvable at every prime in $\cH$, we deduce from Corollary \ref{Corollary-The-description-of-the-splitting-field-in-terms-of-the-Dirichlet-density} that $\bL = \bK(\lambda_a)$. Since $\lambda_{\star}, \lambda_a$ are generators of $\Lambda_a$, it follows that $\bL = \bK(\lambda_a) = \bK(\lambda_{\star}) = \bK$. Since $\bL$ is the splitting field of the polynomial $C_a(x) - m \in \bK[x]$ over $\bK$, we deduce that all solutions to the equation $C_a(x) = m$ belong to $\bK$.
\end{proof}
\begin{remark}
It is known (see \cite[Chapter {\bf12}]{Rosen}) that a generator of the $A$-module $\Lambda_a$ is a function field analogue of a primitive $a$-th root of unity in the number field setting. It is well-known (see \cite{Carlitz}, \cite{Goss}, \cite{Hayes}, \cite{Rosen}, and \cite{Thakur-book}) that the map $x \mapsto C_a(x)$ with $a \in A$ is a Carlitz module analogue of the map $x \mapsto x^a$ with $a \in \bZ$ in the number field setting. Thus Corollary \ref{Corollary-A-special-case-of-a-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem} is a Carlitz module analogue of the classical Grunwald--Wang theorem in the case that $\bK$ contains a generator of the $A$-module $\Lambda_a$. For an account of a similar result in the number field setting, see \cite[Theorem {\bf1.1}, Chapter {\bf VIII}]{Milne-CFT}.
\end{remark}
\section{An analogue of the Grunwald--Wang theorem for certain Galois extensions $\bK/k$}
\label{Section-An-analogue-of-the-Grunwald-Wang-theorem-for-the-intersection-of-K-and-k-lambda-a-being-equal-to-k}
In this section, we prove a Carlitz module analogue of the classical Grunwald--Wang theorem for certain Galois extensions $\bK/k$. The following theorem is the main result in this section.
\begin{theorem}
\label{Theorem-A-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}
We maintain the same notation as in Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Assume that $\bK$ is a Galois extension of $k$ such that every prime $P$ dividing $a$ is unramified in $\bL$. Let $\cH$ be a set of primes of $\bK$ satisfying the following conditions:
\begin{itemize}
\item [(i)] the Dirichlet density of $\cH$ satisfies
\begin{align*}
\delta(\cH) > 1 - \dfrac{1}{\phi(a)q^{\deg(a)}}; \; \text{and}
\end{align*}
\item [(ii)] $\cH$ contains all primes $\wp$ of $\bK$ that divides $a$.
\end{itemize}
Then the equation $C_a(x) = m$ has a solution $x_{\bK}$ in $\bK$ if and only if it has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ of $\bK$ for every prime $\wp$ of $\bK$ with $\wp \in \cH$.
\end{theorem}
\begin{proof}
If $C_a(x) = m$ has a solution $x_{\bK}$ in $\bK$, then it follows immediately that $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $K_{\wp}$ for every prime $\wp$ in $\bK$.
Suppose that $C_a(x) = m$ has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ for each prime $\wp \in \cH$. We will prove that the equation $C_a(x) = m$ has a solution $x_{\bK} \in \bK$.
By assumption, we know from Corollary \ref{Corollary-The-description-of-the-splitting-field-in-terms-of-the-Dirichlet-density} that $\bL = \bK(\lambda_a)$.
Let $h$ be an element in $k^{\alg}$ such that $C_a(h) = m$, and let $\cS$ be the set defined by $(\ref{Equation-The-set-S-of-all-solutions-to-the-equation-C-a-(x)=m})$ in the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Since $a$ is a polynomial of positive degree, we can write $a$ in the form
\begin{align}
\label{Equation-The-representation-of-a}
a = \epsilon P_1^{e_1}\cdots P_n^{e_n},
\end{align}
where $\epsilon$ is in $\bF_q^{\times}$, the $P_i$ are distinct monic primes in $A$, and the $e_i$ are positive integers. For each $1 \le i \le n$, set
\begin{align}
\label{Equation-The-equation-of-a-i-dividing-a}
a_i := \dfrac{a}{P_i^{e_i}} \in A,
\end{align}
We prove that for each $1 \le i \le n$, there exists an element $b_i \in A$ with $\deg(b_i) < \deg(a)$ such that $C_{a_i}(C_{b_i}(\lambda_a) + h) \in \bK$. Indeed, take an integer $1 \le i \le n$. Set
\begin{align*}
\Lambda_{P_i^{e_i}} = \left\{\gamma \in k^{\alg} \; | \; C_{{P_i}^{e_i}}(\gamma) = 0 \right\}.
\end{align*}
We know from \cite[Proposition {\bf12.7}]{Rosen} that $k(\Lambda_{P_i^{e_i}})$ is a Galois extension of $k$.
Let $\fq$ be a prime of $\bK$ lying above $(P_i) = P_iA$, and let $\fQ$ be a prime of $\bK(\Lambda_{P_i^{e_i}})$ lying above $\fq$. Let $\fB$ be a prime of $k(\Lambda_{P_i^{e_i}})$ lying below $\fQ$. Then $\fB$ lies above $(P_i)$.
\begin{center}
\begin{tikzpicture}[node distance = 2cm, auto]
\node (k) {$(P_i) \subset k$};
\node (K) [above of=k, left of=k] {$\fq \subset \bK$};
\node (k1) [node distance = 1cm, above of=k, right of=k, right of =k, right of =k] {$\fB \subset k(\Lambda_{P_i^{e_i}})$};
\node (K1) [above of=k, node distance = 4cm] {$\fQ \subset \bK(\Lambda_{P_i^{e_i}})$};
\draw[-] (k) to node {} (K);
\draw[-] (k) to node [swap] {} (k1);
\draw[-] (K) to node {} (K1);
\draw[-] (k1) to node [swap] {} (K1);
\end{tikzpicture}
\end{center}
We prove that $\fq$ is totally ramified in $\bK(\Lambda_{P_i^{e_i}})$. Let $\e(\fQ/\fq)$, $\e(\fq/(P_i))$ be the ramification indices of $\fQ$ over $\fq$ and $\fq$ over $(P_i)$, respectively. Let $\e(\fQ/\fB)$, $\e(\fB/(P_i))$ be the ramification indices of $\fQ$ over $\fB$ and $\fB$ over $(P_i)$, respectively. Let $\e(\fQ/(P_i))$ be the ramification index of $\fQ$ over $(P_i)$. We know from \cite[Proposition {\bf12.7}]{Rosen} that the prime ideal $(P_i) = P_iA$ is totally ramified in the extension $k(\Lambda_{P_i^{e_i}})/k$, and thus
\begin{align*}
\e(\fB/(P_i)) = [k(\Lambda_{P_i^{e_i}}) : k] = \phi(P_i^{e_i}),
\end{align*}
where $\phi(\cdot)$ denotes the function field analogue of the Euler $\phi$-function (see, for example, \cite[Proposition {\bf1.7}]{Rosen} for the definition of $\phi(\cdot)$). By assumption, we know that $(P_i)$ is unramified in $\bK$, and thus $\e(\fq/(P_i)) = 1$. Since $K(\Lambda_{P_i^{e_i}}) = K\cdot k(\Lambda_{P_i^{e_i}})$, it thus follows from \cite[Lemma {\bf4.6.3}]{Koch} that $\fB$ is unramified in $\bK(\Lambda_{P_i^{e_i}})$. Hence $\e(\fQ/\fB) = 1$, and we deduce that
\begin{align}
\label{Equation-The-ramification-index-of-fQ/fq}
\e(\fQ/\fq) = \e(\fQ/\fq)\e(\fq/(P_i)) = \e(\fQ/(P_i)) = \e(\fQ/\fB)\e(\fB/(P_i)) = \e(\fB/(P_i)) = \phi(P_i^{e_i}).
\end{align}
We contend that the degree of $\bK(\Lambda_{P_i^{e_i}})$ over $\bK$ is less than or equal to $\phi(P_i^{e_i})$. Indeed, let $\chi$ be a primitive generator of $\Lambda_{P_i^{e_i}}$. We know from \cite[Theorem {\bf12.3.16}]{Villa-Salvador} that $\chi$ is a root of the $P_i^{e_i}$--th cyclotomic polynomial $\Phi_{P_i^{e_i}}(x) \in k[x] \subset \bK[x]$ of degree $\phi(P_i^{e_i})$. Hence we deduce that
\begin{align*}
[\bK(\Lambda_{P_i^{e_i}}) : \bK] = [\bK(\chi) : \bK] \le \deg\left(\Phi_{P_i^{e_i}}(x)\right) = \phi(P_i^{e_i}).
\end{align*}
Therefore it follows from $(\ref{Equation-The-ramification-index-of-fQ/fq})$ that
\begin{align*}
\phi(P_i^{e_i}) = \e(\fQ/\fq) \le [\bK(\Lambda_{P_i^{e_i}}) : \bK] \le \phi(P_i^{e_i}),
\end{align*}
and hence
\begin{align*}
\e(\fQ/\fq) = [\bK(\Lambda_{P_i^{e_i}}) : \bK] = \phi(P_i^{e_i}).
\end{align*}
Thus $\fq$ is totally ramified in $\bK(\Lambda_{P_i^{e_i}})$.
We now consider the equation
\begin{align}
\label{Equation-The-equation-C-P^e-x-m}
C_{P_i^{e_i}}(z) = m.
\end{align}
Let $\bL_i$ be the splitting field of the polynomial $C_{P_i^{e_i}}(z) - m \in \bK[z]$ over $\bK$. Since the polynomial $C_{P_i^{e_i}}(z) - m$ is separable, we see that $\bL_i$ is a Galois extension of $\bK$.
We prove that the equation given by $(\ref{Equation-The-equation-C-P^e-x-m})$ has a solution $z_{\wp}$ in the completion $\bK_{\wp}$ for each prime $\wp \in \cH$. Indeed take any prime $\wp \in \cH$. We know that there exists an element $x_{\wp}$ in $\bK_{\wp}$ such that $C_a(x_{\wp}) = m$. Since $a = P_i^{e_i}a_i$, we deduce that
\begin{align}
\label{Equation-The-equation-C-P^e(x)-equals-m-is-locally-solvable}
C_{P_i^{e_i}}(C_{a_i}(x_{\wp})) = C_{P_i^{e_i}a_i}(x_{\wp}) = C_a(x_{\wp}) = m.
\end{align}
Since $C_{a_i}(x) \in k[x] \subset \bK[x]$ and $\bK \subset \bK_{\wp}$, we see that $C_{a_i}(x_{\wp}) \in \bK_{\wp}$. Upon letting $z_{\wp} = C_{a_i}(x_{\wp}) \in \bK_{\wp}$, we see from $(\ref{Equation-The-equation-C-P^e(x)-equals-m-is-locally-solvable})$ that the equation $C_{P_i^{e_i}}(z) = m$ has a solution $z_{\wp}$ in $\bK_{\wp}$. Therefore the equation $C_{P^e}(z) = m$ has a solution $z_{\wp}$ in the completion $\bK_{\wp}$ for each prime $\wp \in \cH$. Applying Corollary \ref{Corollary-The-description-of-the-splitting-field-in-terms-of-the-Dirichlet-density} for the equation $C_{P_i^{e_i}}(z) = m$ with $P_i^{e_i}, m, \bL_i, \chi$ in the roles of $a, m, \bL, \lambda_a$, respectively, where we recall that $\chi$ is a primitive generator of $\Lambda_{P_i^{e_i}}$, we deduce that $\bL_i = \bK(\chi) = \bK(\Lambda_{P_i^{e_i}})$, and thus
\begin{align}
\label{Equation-The-Galois-group-of-L-i-over-k}
\Gal(\bL_i/\bK) = \Gal(\bK(\Lambda_{P_i^{e_i}})/\bK).
\end{align}
Recall that we have shown that $\fq$ is totally ramified in $ \bK(\Lambda_{P_i^{e_i}})$ and the prime $\fQ$ lies above $\fq$. Let $\cD(\fQ/\fq) \le \Gal(\bK(\Lambda_{P_i^{e_i}})/\bK)$ be the decomposition group of $\fQ$ over $\fq$. Since $\fq$ is totally ramified in $\bK(\Lambda_{P_i^{e_i}})$, we deduce from $(\ref{Equation-The-Galois-group-of-L-i-over-k})$ that
\begin{align}
\label{Equation-The-decomposition-group-of-fQ-over-fq}
\cD(\fQ/\fq) = \Gal(\bK(\Lambda_{P_i^{e_i}})/\bK) = \Gal(\bL_i/\bK).
\end{align}
By assumption $(ii)$ in the statement of Theorem \ref{Theorem-A-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}, there exists an element $x_{\fq}$ in the completion $\bK_{\fq}$ such that $C_a(x_{\fq}) = m$. Hence $x_{\fq}$ belongs to the set $\cS$, where $\cS$ is the set given by $(\ref{Equation-The-set-S-of-all-solutions-to-the-equation-C-a-(x)=m})$ in the proof of Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Thus $x_{\fq} = C_{b_i}(\lambda_a) + h$ for some element $b_i \in A$ with $\deg(b_i) < \deg(a)$. We see that
\begin{align*}
C_{P_i^e}(C_{a_i}(x_{\fq})) = C_{P_i^{e_i}a_i}(x_{\fq}) = C_a(x_{\fq}) = m,
\end{align*}
and thus $C_{a_i}(x_{\fq})$ is a root of the polynomial $C_{P_i^{e_i}}(z) - m \in \bK[z]$. Since $\bL_i = \bK(\Lambda_{P_i^{e_i}})$ is the splitting field of the polynomial $C_{P_i^e}(z) - m$ over $\bK$, we deduce that
\begin{align}
\label{Equation-The-field-of-definition-of-C-ai-x-fq}
m_i := C_{a_i}(x_{\fq}) = C_{a_i}(C_{b_i}(\lambda_a) + h) \in \bK_{\fq} \cap \bL_i.
\end{align}
Since $\bK_{\fq} \cap \bL_i$ is the fixed field of $\cD(\fQ/\fq)$ (see \cite[Proposition {\bf9.8}]{Neukirch}), we see from $(\ref{Equation-The-field-of-definition-of-C-ai-x-fq})$ that $\psi(m_i) = m_i$ for every $\psi \in \cD(\fQ/\fq)$. By $(\ref{Equation-The-decomposition-group-of-fQ-over-fq})$, we know that $\cD(\fQ/\fq) = \Gal(\bL_i/\bK)$, and hence $m_i$ belongs to $\bK$.
In summary, we have shown that each $1 \le i \le n$, there exists an element $b_i \in A$ with $\deg(b_i) < \deg(a)$ such that $m_i = C_{a_i}(C_{b_i}(\lambda_a) + h) \in \bK$.
Since the elements $P_i^{e_i}$ are pairwise coprime in the principal ideal domain $A$, it follows from the Chinese Remainder Theorem that there exists an element $\epsilon$ in $A$ such that
\begin{align}
\label{Equation-The-element-epsilon-in-A}
\epsilon \equiv b_i \pmod{P_i^{e_i}}
\end{align}
for each $1 \le i \le n$. Set
\begin{align}
\label{Equation-The-equation-of-x-bK}
x_{\bK} := C_{\epsilon}(\lambda_a) + h.
\end{align}
We prove that $x_{\bK}$ belongs to $\bK$, and is a solution to the equation $C_a(x) = m$. We first show that $C_{a_i}(x_{\bK}) = m_i$ for each $1 \le i \le n$.
Take any integer $1 \le i \le n$. By $(\ref{Equation-The-element-epsilon-in-A})$, there exists an element $\delta_i \in A$ such that $\epsilon = b_i + \delta_iP_i^{e_i}$. It thus follows from $(\ref{Equation-The-equation-of-x-bK})$ and $(\ref{Equation-The-equation-of-a-i-dividing-a})$ that
\begin{align*}
C_{a_i}(x_{\bK}) &= C_{a_i}(C_{\epsilon}(\lambda_a) + h) \\
&= C_{a_i}(C_{b_i + \delta_iP_i^{e_i}}(\lambda_a) + h) \\
&= C_{a_i}(C_{b_i}(\lambda_a) + h) + C_{a_i}(C_{\delta_iP_i^{e_i}}(\lambda_a))\\
&= m_i + C_{\delta_i a_iP_i^{e_i}}(\lambda_a) \\
&= m_i + C_{\delta_i a}(\lambda_a) \\
&= m_i + C_{\delta_i}(C_a(\lambda_a)) \\
&= m_i.
\end{align*}
In particular, this implies that $C_{a_i}(x_{\bK}) = m_i \in \bK$ for every $1 \le i \le n$.
By $(\ref{Equation-The-equation-of-a-i-dividing-a})$, we deduce that $\gcd(a_1, a_2, \ldots, a_n) = 1$, and thus there exist elements $\upsilon_1, \upsilon_2, \ldots, \upsilon_n$ in $A$ such that
\begin{align*}
\upsilon_1 a_1 + \upsilon_2 a_2 + \ldots + \upsilon_n a_n = 1.
\end{align*}
Since $C_{\upsilon_i}(x) \in k[x] \subset \bK[x]$ for every $1 \le i \le n$, we deduce that
\begin{align*}
x_{\bK} = C_1(x_{\bK}) = C_{\upsilon_1 a_1 + \upsilon_2 a_2 + \cdots + \upsilon_n a_n}(x_{\bK}) = C_{\upsilon_1}(C_{a_1}(x_{\bK})) + \cdots + C_{\upsilon_n}(C_{a_n}(x_{\bK})) \in \bK.
\end{align*}
Furthermore since $m_1 = C_{a_1}(C_{b_1}(\lambda_a) + h)$ is a solution to the equation $C_{P_1^{e_1}}(z) = m$, we deduce that
\begin{align*}
C_{a}(x_{\bK}) = C_{P_1^{e_1}a_1}(x_{\bK}) = C_{P_1^{e_1}}(C_{a_1}(x_{\bK})) = C_{P_1^{e_1}}(m_1) = m,
\end{align*}
which proves that the equation $C_a(x) = m$ has a solution $x_{\bK} \in \bK$. Thus our contention follows.
\end{proof}
Let $\cP_{\bK}$ be the set of all primes in $\bK$. If $\cH$ is a set of primes of $\bK$ such that $\cP_{\bK}\setminus \cH$ is finite and $\cH$ contains all primes of $\bK$ that divide $a$, then $\cH$ is of Dirichlet density $1$. Hence we obtain the following result that is a special case of Theorem \ref{Theorem-A-Carlitz-module-analogue-of-the-Grunwald-Wang-theorem}.
\begin{corollary}
We maintain the same notation as in Theorem \ref{Theorem-The-description-of-the-splitting-field-of-the-Carlitz-equation}. Assume that $\bK$ is a Galois extension of $k$ such that every prime $P$ dividing $a$ is unramified in $\bK$. Then the equation $C_a(x) = m$ has a solution $x_{\bK}$ in $\bK$ if and only if it has a solution $x_{\wp}$ in the completion $\bK_{\wp}$ of $\bK$ for all primes $\wp \in \cH$, where $\cH$ is a set of primes of $\bK$ such that $\cP_{\bK}\setminus \cH$ is finite and $\cH$ contains all primes of $\bK$ that divide $a$.
\end{corollary}
|
\section{Introduction}\label{aba:sec1}
Recent observations provide strong evidence for the acceleration of our universe. The explanation of this remarkable phenomenon is rather difficult: one needs a dark energy field, which could be the cosmological constant.
However, there is a huge discrepancy between the observed value of the cosmological constant and the value obtained from the vacuum energy density. Moreover, there is an incredibly fine-tuning between the energy density of this dark energy and the energy density of matter.
It is a great challenge to find other explanations for the late-time acceleration, without the need for a cosmological constant. Physicists speculate that extra spatial dimensions could exist in addition to our ordinary 4-dimensional spacetime.
These theories can be used to explain several of the shortcomings of the Standard Model, i.e., the unknown origin of dark energy and the weakness of gravity (hierarchy problem). Recently there is growing interest in the so-called brane world models\cite{IA:1998,RS:1999}. In these models, the weakness of gravity might be fundamental.
Compact objects, such as black holes and cosmic strings, could have tremendous mass in the bulk, while their warped manifestations in the brane can be consistent with observations. The energy scale of the cosmic string, $G\mu$, could be warped down to GUT scale, even if its value was at the Planck scale.
Cosmological cosmic strings can also be investigated on a FLRW background. However, the string-cosmology spacetime essentially looks like a scaled version of a string in a vacuum spacetime.\cite{Greg:1989}
Static and time-dependent solutions of the U(1) gauge string on a warped FLRW spacetime show significant deviation from this classical solution in 4D\cite{Slag2:2014,Slag3:2015}.
While the inflaton field plays a crucial role in the early stage of our universe, it could play a comparable role at much later times, if we modify gravity by considering warped brane-world models. It could be possible that there exists a correlation between the accelerating universe and large extra dimensions in brane-world models.
Wavelike disturbances triggered by the huge mass of the cosmic string in the bulk, could have observational effects on the brane. One conjectures that these disturbances could act as an effective dark energy field.
In this paper we examine the possibility that cosmological cosmic strings, in contrast with the 4-dimensional counterpart model,\cite{Greg:1989} could contribute to a dark radiation and triggers the acceleration or deceleration of the universe. This conjecture is founded by a multiple-scale analysis.
\section{The 5-dimensional Model}
We consider here the 5-dimensional Einstein equations with only a bulk cosmological constant $\Lambda_5$\cite{shir:2000}
\begin{equation}
{^{5}\!G}_{\mu\nu}=-\Lambda_5{^{5}\!g}_{\mu\nu}+\kappa_5^2 \delta(y)\Bigl(-\Lambda_4 {^{4}\!g}_{\mu\nu}+{^{4}\!T}_{\mu\nu}\Bigr),\label{eqn1}
\end{equation}
with $\kappa_5= 8\pi {^{5}\!G}= 8\pi/{^{5}\!M}_{pl}^3$, $\Lambda_4$ the brane tension, ${^{4}\!g}_{\mu\nu}={^{5}\!g}_{\mu\nu}-n_\mu n_\nu$,
and $n^\mu$ the unit normal to the brane. We consider here the matter field ${^{4}\!T}_{\mu\nu}$ confined to the brane, i.e., the U(1) scalar-gauge field, written in the form \cite{Garf:1985} $\Phi=\eta X(t, r)e^{i\varphi}, A_\mu =\frac{1}{e}(P(t, r)-1)\nabla_\mu\varphi$,
with $\eta$ the vacuum expectation value of the scalar field and potential $V=\frac{1}{8}\beta(X^2-\eta^2)^2$.
Let us consider the cylindrically symmetric warped spacetime
\begin{equation}
ds^2 = {\cal W}(t, r, y)^2\Bigl[e^{2(\gamma(t, r)-\psi(t, r))}(-dt^2+ dr^2)+e^{2\psi(t, r)}dz^2+ r^2 e^{-2\psi(t, r)}d\varphi^2\Bigr]+ dy^2,\label{eqn2}
\end{equation}
with ${\cal W}$ the warp factor and $y$ the bulk space coordinate.
One can analyse the behavior of gravitational waves generated by the local strings on the expansion of the universe. One usually considers the zero-thickness limit of non-singular spacetimes containing a cylindrical distribution of stress-energy embedded on a cosmological background. In general, however, the late-time behavior of non-vacuum spacetimes, show conical behavior, not desirable. It was found\cite{Greg:1989} that a U(1) cosmic string can be embedded into a flat FLRW spacetime along the polar axis.
It turns out that when the width of the strings is smaller than the Hubble radius, the disturbances are negligible.
On a warped 5-dimensional spacetime, they could survive and the effective brane spacetime is non-conical.\cite{Slag2:2014}
From the 5D equations one obtains the solution for ${\cal W}(t, r, y)$ with two branches:\cite{Slag3:2015}
\begin{equation}
{\cal W}(t, r, y)=\pm\frac{e^{\sqrt{-\frac{\Lambda_5}{6}}(y- y_0)}}{\sqrt{\tau r}}\sqrt{\Bigl(d_1 e^{(\sqrt{2\tau})t}-d_2e^{-(\sqrt{2\tau})t}\Bigr)\Bigl(d_3 e^{(\sqrt{2\tau})r}-d_4e^{-(\sqrt{2\tau})r}\Bigr)}, \label{eqn3}
\end{equation}
with $\tau, d_i$ some constants. In general, ${\cal W}$ can possess a saddle-point or extremal values.
The modified Einstein field equations induced on the brane become\cite{shir:2000}
\begin{equation}
{^{4}\!G}_{\mu\nu}=-\Lambda_{eff}{^{4}\!g}_{\mu\nu}+\kappa_4^2 {^{4}\!T}_{\mu\nu}+\kappa_5^4{\cal S}_{\mu\nu}-{\cal E}_{\mu\nu}.\label{eqn4}
\end{equation}
${\cal S}_{\mu\nu}$ is the quadratic term in the energy-momentum tensor and ${\cal E_{\mu\nu}}$
is a part of the 5D Weyl tensor and carries information of the gravitational field outside the brane.
From the Einstein equations and scalar-gauge field equations one obtains a set of PDE's, which can be solved numerically.\cite{Slag2:2014}
The time dependent part of the warp factor causes disturbances of the order much larger than the expected values in the 4D case.
Moreover, it was found that for an extremum of $W_1$, the solution diverges. This could have a significant influence on a transition from acceleration to a deceleration or vice versa.\cite{Slag3:2015}
\section{The High-Frequency Approximation}
We consider fields $V_i$ in point x on a manifold M dependent on different scales $(x_\mu, \xi, \chi , ...)$:\cite{Choc:1977,Slag4:1986}
\begin{equation}
V_i=\sum_{0}^{\infty} \frac{1}{\omega^n} F_i^{(n)}(x_\mu,\xi,\chi ,...)\label{eqn5},
\end{equation}
with $\omega>>1$, $\xi =\omega \Theta(x_\mu)$, $\chi =\omega \Pi(x_\mu), ...$ and $\Theta, \Pi , ...$ scalar (phase) functions on M. The small parameter $\frac{1}{\omega}$ can be the ratio of the characteristic wavelength of the perturbation to the characteristic dimension of the background (or the ratio of the extra dimension y to the background dimension).
We expand
\begin{eqnarray}
g_{\mu\nu}=\bar g_{\mu\nu}(x_\mu)+ \frac{1}{\omega}h_{\mu\nu}(x_\mu,\xi,\chi)+\frac{1}{\omega^2}k_{\mu\nu}(x_\mu,\xi,\chi) + ... \cr \Phi=\bar\Phi(x_\mu) +\frac{1}{\omega}\Psi(x_\mu, \xi, \chi)+\frac{1}{\omega^2}\Xi(x_\mu, \xi, \chi)+...
\cr A_\mu=\bar A_\mu (x_\mu)+\frac{1}{\omega}B_\mu (x_\mu,\xi ,\chi) +\frac{1}{\omega^2}C_\mu (x_\mu,\xi,\chi) +...,\label{eqn6}
\end{eqnarray}
with $\bar g_{\mu\nu}$ the background metric and $\bar\Phi, \bar A_\mu$ the background scalar and gauge fields.
Let us now define
\begin{eqnarray}
\frac{d g_{\mu\nu}}{d x^\sigma}=g_{\mu\nu,\sigma}+\omega l_\sigma \dot g_{\mu\nu} \qquad g_{\mu\nu,\sigma}\equiv \frac{\partial g_{\mu\nu}}{\partial x^\sigma}\qquad
\dot g_{\mu\nu}\equiv \frac{\partial g_{\mu\nu}}{\partial \xi},\label{eqn7}
\end{eqnarray}
with $l_\mu \equiv \frac{\partial \Theta}{\partial x^\mu}$. So we consider, for the time being, only rapid variation in the direction of $l_\mu$ transversal to the sub-manifold $\Theta$ = constant.
We can expand the several relevant tensors, for example,
$\Gamma_{\mu\nu}^\alpha =\bar \Gamma_{\mu\nu}^\alpha +\Gamma_{\mu\nu}^{\alpha (0)} +\frac{1}{\omega}\Gamma_{\mu\nu}^{\alpha (1)}+...$ and
$R^\sigma_{\mu\tau\nu}=\omega R^{(-1)\sigma}_{\mu\tau\nu}+\bar R^\sigma_{\mu\tau\nu}+R^{(0)\sigma}_{\mu\tau\nu} +.... $,
with
$\Gamma^{\sigma (0)}_{\mu\nu}=\frac{1}{2}\bar g^{\beta\sigma}\bigl( l_\mu \dot h_{\beta\nu}+l_\nu\dot h_{\beta\mu}-l_\beta \dot h_{\mu\nu}\bigr)$ and
$\Gamma^{\sigma (1)}_{\mu\nu}=\frac{1}{2}\Bigl(h^\sigma_{\mu :\nu}+h^\sigma_{\nu :\mu}-h_{\mu\nu}^{: \sigma}\Bigr)-\frac{1}{2}\Bigl(l_\nu \dot k_\mu^\sigma + l_\mu \dot k_\nu^\sigma -l^\sigma \dot k_{\mu\nu}\Bigr)-h_\rho^\sigma \Gamma^{\rho (0)}_{\mu\nu}$.
We substitute the expansions into the effective brane Einstein equations Eq.~(\ref{eqn4}) and subsequently put equal zero the various powers of $\omega$. We then obtain a system of partial differential equations for the fields $\bar g_{\mu\nu}, h_{\mu\nu}, k_{\mu\nu}$ and the scalar gauge fields $\bar \Phi, \Psi, \Xi, \bar A_\mu, B_\mu \equiv[B_0,B_1,0,B,0]$ and $C_\mu$. The $\omega^{(-1)}$ equation becomes
\begin{equation}
{^{4}\!G_{\mu\nu}^{(-1)}}=-{\cal E}_{\mu\nu}^{(-1)},\label{eqn8}
\end{equation}
and the $\omega^{(0)}$ equation
\begin{equation}
{^{4}\!\bar G_{\mu\nu}}+{^{4}\!G_{\mu\nu}^{(0)}}=-\Lambda_{eff}{^{4}\!\bar g_{\mu\nu}}+\kappa_4^2 \bigl({^{4}\!\bar T_{\mu\nu}}+{^{4}\!T_{\mu\nu}^{(0)}}\bigr)
+\kappa_5^4\bigl(\bar {\cal S}_{\mu\nu}+{\cal S}_{\mu\nu}^{(0)}\bigr)-\bigl(\bar{\cal E}_{\mu\nu}+{\cal E}_{\mu\nu}^{(0)}\bigr).\label{eqn9}
\end{equation}
The interesting contribution comes from the projected Weyl tensor. This contribution from the bulk space, ${\cal E}_{\mu\nu}^{(-1)}$, must be calculated with the 5D Riemann tensor
${^{5}\!R^{(-1)\sigma}_{\mu\tau\nu}}=l_\tau {^{5}\!\dot \Gamma^{(0)\sigma}_{\mu\nu}-l_\nu }{^{5}\!\dot\Gamma^{(0)\sigma}_{\mu\tau}}$.
If we consider $l_\mu l^\mu =0$, i.e., the eikonal equation, then one obtains from Eq.~(\ref{eqn8}) $l^\alpha\bigl(\ddot h_{\alpha\nu}-\frac{1}{2}\bar g_{\alpha\nu}\ddot h\bigr)=0,$
which in other context is used as gauge conditions. It turns out that the contribution from the ${\cal E}_{\mu\nu}^{(-1)}$ don't change these conditions on $h_{\mu\nu}$.
Let us consider the zero-order Eq.~(\ref{eqn9}). We need
\begin{eqnarray}
{\cal E}_{\mu\nu}^{(0)}&=&n^\gamma n^\delta {^{4}\!g_\mu^\alpha}{^{4}\!g_\nu^\beta}\Bigl[{^{5}\!R^{(0)}_{\alpha\gamma\beta\delta}}-\frac{1}{3}\Bigl({^{5}\!\bar g_{\alpha\gamma}} {^{5}\!R^{(0)}_{\delta\beta}}-{^{5}\!\bar g_{\alpha\delta}} {^{5}\!R^{(0)}_{\gamma\beta}}
-{^{5}\!\bar g_{\beta\delta}} {^{5}\!R^{(0)}_{\gamma\alpha}}\cr &+&{^{5}\!\bar g_{\beta\delta}} {^{5}\!R^{(0)}_{\gamma\alpha}}\Bigr)
+\frac{1}{12}\Bigl({^{5}\!\bar g_{\alpha\gamma}}{^{5}\!\bar g_{\delta\beta}}-{^{5}\!\bar g_{\alpha\delta}}{^{5}\!\bar g_{\gamma\beta}}\Bigr){^{5}\!R}\Bigr].\label{eqn10}
\end{eqnarray}
Now we take as an example for the radiative coordinates $\Pi(x_\mu)=t+r$ the simplified case $l_\mu =[1,1,0,0,0]$. Then we obtain from the gauge condition that only $h_{11}, h_{22}, h_{44}, h_{55}, h_{13}, h_{14}, h_{15}, h_{34}, h_{45}$ and $h_{35} $ survive.
From the Einstein equations (\ref{eqn9}), one obtains a set of PDE's. When we use the additional gauge conditions: $h_{22}=-h_{11}, h_{34}=h_{35}=h_{45}=h_{14}=h_{15}=0$ (leaving 4
independent $h_{\mu\nu}$ terms), we have 7 unknown functions for the background and first order perturbations: $\bar W_1, \bar\psi, \bar \gamma, \dot h_{13}, \dot h_{11}, \dot h_{44}$ and $\dot h_{55}$
One can also integrate the equation Eq.(\ref{eqn9}) with respect to $\xi$. If we suppose that the perturbations are periodic in $\xi$, we then obtain the Einstein equations with back-reaction terms:
\begin{equation}
{^{4}\!\bar G_{\mu\nu}}=\kappa_4^2{^{4}\!\bar T_{\mu\nu}}+\kappa_5^4 \bar {\cal S}_{\mu\nu}-\bar{\cal E}_{\mu\nu}
+\frac{1}{\tau}\int \Bigl(\kappa_4^2T_{\mu\nu}^{(0)}+\kappa_5^4S_{\mu\nu}^{(0)}-{^{4}\!G_{\mu\nu}^{(0)}}-{\cal E}_{\mu\nu}^{(0)}\Bigr ) d\xi\label{eqn11}
\end{equation}
where we took $\Lambda_{eff}=0$ for the RS fine-tuning and $\tau$ de period of the high-frequency components.
One can say that the term $-\int {\cal E}_{\mu\nu}^{(0)}d\xi$ in Eq.(\ref{eqn11})is the KK-mode contribution of the perturbative 5D graviton. It is an extra back-reaction term, which contain $\dot h_{55}$ amplified by the warp factor and with opposite sign with repect to the $\kappa_4^2$-term. So it can play the role of an effective cosmological constant.
By substituting back these equations into the original equations, one gets propagations equations for the first order perturbations. In this way we obtain the set PDE's
\begin{eqnarray}
\partial^2_{tt}\bar W_1=-\partial^2_{rr}\bar W_1+\frac{2}{\bar W_1}(\partial_t\bar W_1^2+\partial_r\bar W_1^2)-\bar W_1(\partial_t\bar \psi^2+\partial_r\bar \psi^2)
+\frac{\bar W_1}{r}(\partial_r\bar \gamma-\partial_t\bar\gamma)\cr
+2(\partial_r\bar W_1-\partial_t\bar W_1)(\partial_t\bar\psi-\partial_r\bar\psi+\partial_r\bar\gamma-\partial_t\bar\gamma)
+2\bar W_1\partial_r\bar\psi\partial_t\bar\psi-4\frac{\partial_r\bar W_1\partial_t\bar W_1}{\bar W_1}
+ 2\partial_{rt}\bar W_1\cr
-\frac{3}{4}\kappa_4^2\Bigl(e^{2\bar\psi}\frac{(\partial_t\bar P-\partial_r\bar P)^2}{\bar W_1 r^2 e^2}
+\bar W_1(\partial_t\bar X-\partial_r\bar X)^2\Bigr),\qquad\qquad\label{eqn12}
\end{eqnarray}
\begin{eqnarray}
\partial_{tt}\bar\psi=\partial_{rr}\bar\psi +\frac{\partial_r\bar\psi}{r}+\frac{2}{\bar W_1}(\partial_r \bar W_1\partial_r\bar\psi -\partial_t\bar W_1\partial_t\bar\psi)-\frac{\partial_r\bar W_1}{r\bar W_1}\cr
+\frac{3 e^{2\bar\psi}}{4\bar W_1^2r^2e^2}\kappa_4^2\Bigr(\partial_t\bar P^2-\partial_r\bar P^2-\bar W_1^2 e^2\bar X^2\bar P^2 e^{2\bar\gamma-2\bar\psi}\Bigr),\label{eqn13}
\end{eqnarray}
\begin{eqnarray}
\partial_{t}\bar\gamma =\partial_r\bar\gamma+\frac{1}{\partial_t\bar W_1-\partial_r\bar W_1-\frac{\bar W_1}{2r}}\Bigr\{\tfrac{1}{2}\bar W_1(\partial_t\bar\psi -\partial_r\bar\psi)^2+\frac{\partial_r\bar W_1}{r}
-\partial_{tr}\bar W_1+\partial_{rr}\bar W_1 \cr+\frac{2\partial_r\bar W_1\partial_t\bar W_1}{\bar W_1}+(\partial_r\bar W_1-\partial_t\bar W_1)(\partial_r\bar \psi-\partial_t\bar \psi)
-\frac{\partial_r\bar W_1^2+3\partial_t\bar W_1^2}{2\bar W_1}+\kappa_4^2\frac{\bar W_1}{16}\Bigl(7\partial_t\bar X^2
+5\partial_r\bar X^2\cr-12\partial_r\bar X\partial_t\bar X +5\frac{e^{2\bar\gamma}\bar X^2\bar P^2}{r^2}
+6e^{2\bar\psi}\frac{(\partial_r\bar P-\partial_t\bar P)^2}{\bar W_1^2r^2e^2}+\bar W_1^2\beta e^{2\bar\gamma-2\bar\psi}(\bar X^2-\eta^2)^2\Bigr)\Bigr\},\qquad\label{eqn14}
\end{eqnarray}
\begin{equation}
\partial_t\dot h_{13}=\partial_r\dot h_{13}+\ddot k_{13}-\ddot k_{23}+2\Bigl(\frac{\partial_t\bar W_1 -\partial_r\bar W_1}{\bar W_1}+\partial_t\bar\psi -\partial_r\bar\psi\Bigr)\dot h_{13},\label{eqn15}
\end{equation}
\begin{eqnarray}
\partial_t\dot h_{11}=\partial_r\dot h{11}+\frac{e^{2\bar\gamma}}{r^2}\Bigl(\partial_r\bar\psi-\partial_t\bar\psi-\frac{1}{2r}\Bigr)\dot h_{44}+\frac{1}{2}(\ddot k_{22}+\ddot k_{11})-\ddot k_{12}\cr
\frac{2}{\bar W_1}\Bigl(\partial_t\bar W_1-\partial_r\bar W_1+\bar W_1(\partial_r\bar\psi-\partial_t\bar\psi+\partial_t\bar\gamma-\partial_r\bar\gamma)\Bigr)\dot h_{11}\cr
+\frac{1}{2}e^{2\bar\gamma-2\bar\psi}\bar W_1^2\Bigr(\frac{1}{2r}+\frac{\partial_r\bar W_1-\partial_t\bar W_1}{\bar W_1}\Bigr)\dot h_{55}
+\kappa_4^2e^{2\bar\gamma-2\bar\psi}\bar W_1^2(\partial_t\bar X-\partial_r\bar X) \dot\Psi\cos\varphi, \qquad\qquad\label{eqn16}
\end{eqnarray}
\begin{eqnarray}
\partial_\varphi\Bigl(\dot h_{11}+\frac{e^{2\bar\gamma}}{r^2}\dot h_{44}-\bar W_1^2 e^{2\bar\gamma-2\bar\psi}\dot h_{55}\Bigr)+\ddot k_{24}-\ddot k_{14}=
-2\kappa_4^2\bar X\bar Pe^{2\bar\gamma-2\bar\psi}\bar W_1^2\sin\varphi\dot\Psi,\label{eqn17}
\end{eqnarray}
\begin{eqnarray}
\partial_t\dot h_{44}=\partial_r\dot h_{44}+\Bigl(2\partial_r\bar\psi-2\partial_t\bar\psi-\frac{3}{2r}+\frac{\partial_t\bar W_1-\partial_r\bar W_1}{\bar W_1}\Bigr)\dot h_{44}\cr
+\frac{\kappa_4^2}{\epsilon}(\partial_r\bar P-\partial_t\bar P)\dot B+\frac{1}{2}\bar W_1^2 r^2e^{-2\bar\psi}\Bigl(\partial_t\bar\psi-\partial_r\bar\psi+\frac{1}{2r}\Bigr)\dot h_{55}.\label{eqn19}
\end{eqnarray}
We notice that in our simplified case the equations for the background metric separates from the perturbations. So this example is very suitable to investigate the perturbation equations.
For the first order gauge field perturbation $B_\mu$ we used the condition $l^\mu B_\mu =0$, which is a consequence, as we will see, of the gauge field equations. So $B_\mu$ can be parameterized as $B_\mu =[B_0,B_0,0,B,0]$.
The propagation equation for $\dot h_{55}$ yields $\dot h_{55}=F_1(t+r)F_2(\varphi,y,\xi)$, which is expected, because the brane part of $\dot h_{55}$ must be separable from the bulk part.
We omitted for the time being, the $\kappa_5^4$ contribution.
It is manifest that to zero order there is an interaction between the high-frequency perturbations from the bulk, the matter fields on the brane and the evolution of $\dot h_{ij}$, also found in the numerical solution{\cite{Slag2:2014,Slag3:2015}. We observe again that the bulk contribution $\dot h_{55}$ is amplified with $\bar W_1^2$. It is a reflection of the massive KK modes felt on the brane.
The equations for the matter fields can be obtained in a similar way. The equation for the background $\bar \Phi$ becomes
\begin{equation}
\bar D^\alpha\bar D_\alpha\bar\Phi-\frac{1}{2}\beta\bar\Phi(\bar\Phi\bar\Phi^*-\eta^2)=\frac{1}{\tau}\int\Bigl(h^{\mu\nu}l_\mu l_\nu\ddot\Psi +\bar g^{\mu\nu}\Gamma^{\alpha(0)}_{\mu\nu}\dot\Psi\Bigr)d\xi.\label{eqn20}
\end{equation}
The equation for $\bar A_\mu$ is the same as in the unperturbed situation. For the first order perturbations we obtain ( for $l^\alpha C_\alpha =0$)
\begin{equation}
\partial_t \dot\Psi=\partial_r\dot\Psi+\frac{\dot\Psi}{\bar W_1}(\partial_r\bar W_1 -\partial_t\bar W_1)+\frac{1}{2r}\dot\Psi,\label{eqn21}
\end{equation}
\begin{eqnarray}
\partial_t\dot B=\partial_r\dot B+\Bigl(\partial_r\bar\psi-\partial_t\bar\psi-\frac{1}{2r}\Bigr)\dot B+
\frac{e^{2\bar\psi}(\partial_t\bar P-\partial_r\bar P)}{2r^2\bar W_1^2 e}\dot h_{44},\label{eqn22}
\end{eqnarray}
\begin{equation}
\partial_t \dot B_0 =\partial_r \dot B_0-\frac{e^{2\bar\gamma}}{r^2}\partial_\varphi \dot B -ee^{2\bar\gamma-2\bar\psi}\dot\Psi \bar X\bar W_1^2\sin\varphi.\label{eqn23}
\end{equation}
For these matter field equations one needs the condition $l^\alpha \bar A_\alpha =0$, otherwise the real and imaginary parts of $\dot \Psi$ interact as the propagation progresses.
From Eq.(\ref{eqn16}), Eq.(\ref{eqn17}) and Eq.(\ref{eqn23}) we observe on the right hand side $\varphi$-dependent terms, amplified by $W_1^2$. So the approximate wave solution is no longer axially symmetric, also found by Choquet-Bruhat\cite{Choc:1968}. After integration with respect to $\varphi$, we obtain from Eq.(\ref{eqn17}) ( for $k_{14}=k_{24}$)
\begin{equation}
\dot h_{11}=e^{2\bar\gamma -2\bar\psi}\bar W_1^2\dot h_{55}-\frac{e^{2\bar\gamma}}{r^2}\dot h_{44}-2\kappa_4^2e^{2\bar\gamma-2\bar\psi}\bar X\bar P\bar W_1^2\int(\dot\Psi\sin\varphi) d\varphi.\label{eqn24}
\end{equation}
We notice that the first order disturbance $\dot h_{11}$ (and so $\dot h_{22}=-\dot h_{11}$)) could have its maximum for fixed angle $\varphi$ amplified by the warp factor $W_1^2$.
If we choose for example
$\Psi=\cos(\varphi) \tilde\Psi(t,r,\xi)$, then the last term in Eq.(\ref{eqn24}) becomes $\kappa_4^2\bar X\bar P e^{2\bar\gamma-2\bar\psi}\bar W_1^2\cos(2\varphi )\dot{\tilde\Psi}$, which has two extremes on $[0,\pi]$ $\mod(\frac{1}{2}\pi )$. This could be an explanation of the recently found spooky alignment of the rotation axes of quasars over large distances in two perpendicular directions.
\section{Conclusions}
A nonlinear approximation of the field equations of the coupled Einstein-scalar-gauge field equations on a warped 5D spacetime is investigated. To zeroth order in the expansion parameter it is found that the evolution
of the perturbations on the brane is triggered by the electric part of the 5D Weyl tensor and carries information of the gravitational field outside the brane. The warpfactor in the nominator in front of the bulk contributions will cause a huge disturbance on the brane and could act as dark energy.
|
\section{Introduction}
Sandage-Loeb (SL) test is a unique method to directly measure the expansion history of the universe in the ``redshift desert'' of $2\lesssim z\lesssim 5$. It was firstly proposed by Sandage~\cite{sandage} to directly measure the variation of redshift of distant sources. Then Loeb~\cite{loeb} pointed out the possibility of detecting redshift drift in the spectra of Lyman-$\alpha$ forest of distant quasars (QSO) in decades. The 39-meter European Extremely Large Telescope (E-ELT) equipped with a high-resolution spectrograph called CODEX (COsmic Dynamics EXperiment) is in built to achieve this goal. The SL test is of great significance for cosmology because it is a direct geometric measurement of the expansion history of the universe and covers the high redshift range of $2\lesssim z\lesssim 5$, which is almost unaccessible with existing probes.
The effect of the SL test on parameter estimation has been studied by enormous works~\cite{sl1,sl2,sl3,sl4,sl5,sl6,sl7,Darling,Zhang21}, however, many works incorrectly assumed 240 or 150 quasars to be observed. In fact, according to a Monte Carlo simulation analyzed in depth, using a telescope with a spectrograph like CODEX, only about 30 quasars are bright enough and/or lying at a high enough redshift for the actual observation~\cite{Liske}. Moreover, as far as we know, in almost all the existing papers, the best-fit $\Lambda$CDM model to current observational data is usually chosen as the fiducial model in simulating the mock future SL test data. In such a way, when these simulated data are combined with other actual data to constrain some dynamical dark energy models (or modified gravity models), tension between the simulated SL data and other actual data may occur, leading to an inappropriate joint constraint. Thus, such a method may not give convincing conclusion on the potential impact of the future SL test data on parameter estimation.
In our recent work~\cite{msl1}, we suggested that to avoid the potential inconsistency in data the best-fitting model (in study) to current actual data is taken to be the fiducial model in
producing the simulated SL test data, and 30 mock data are then produced with this procedure. In such a way, the simulated mock data are well consistent with the current actual data no matter what dark energy
models are considered. The conclusion of the impact of SL test on future parameter estimation is thus rather convincing. In Ref.~\cite{msl1}, as a typical example, we only focused on the
dark energy model with constant $w$ (i.e., the $w$CDM model).
It was shown that compared to the current combined data of type Ia supernovae (SN), baryon acoustic oscillation (BAO), cosmic microwave background (CMB), and Hubble constant, the 30-yr observation of SL
test could improve the constraint on $\Omega_m$ by about 80\% and the constraint on $w$ by about 25\%.
Furthermore, if the interaction between dark energy and dark matter is considered, the SL test 30-yr data could also improve the constraint on the coupling $\gamma$ by about 30\% and 10\%
for the $Q=\gamma H\rho_{\rm c}$ and $Q=\gamma H\rho_{\rm de}$ models, respectively, as shown in Ref.~\cite{msl1}.
In this paper, we will further extend the discussions in Ref.~\cite{msl1} and investigate the parameter estimation with the SL test in depth.
We will consider the case of time-evolving dark energy model, and show how the SL test impacts on the constraints on the equation of state of
such a dark energy. As usual, we adopt the most commonly used parametrization $w(z)=w_0 + w_az/(1+z)$, and call the corresponding model the $w_0w_a$CDM model.
A comprehensive comparison among the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models with the SL test will be performed.
Another important issue is about the determination of the Hubble constant by using the future SL test data.
It is well known that in the current data there is a strong degeneracy between $\Omega_m$ and $H_0$ (they are in an anti-correlation).
Breaking this degeneracy is extremely important for cosmology.
In this work, we will show that the SL test is very helpful in breaking the degeneracy between $\Omega_m$ and $H_0$, and thus
is very helpful in determining the value of the Hubble constant.
Furthermore, we will also discuss what accuracy would be achieved when using the SL test to directly measure the high-redshift $H(z)$ values.
In fact, a more meaningful question is to ask how the SL test would impact on the dark energy constraints in the future geometric measurements.
We will also address this issue in the present work. As a concrete example, we simulate the future SN and BAO data based on the long-term space-based project JDEM.
We wish to quantify the constraining power of the SL test in the future high-precision joint geometric constraints on dark energy.
\section{Methodology}
First, we briefly describe the current observational data used in the analysis. Actually, the current data used in this work are the same to those in Ref.~\cite{msl1},
in order to make a direct comparison. The most typical geometric measurements are chosen, i.e., the observations of SN, BAO, CMB, and $H_0$.
The combination of SN, BAO, CMB, and $H_0$ is, actually, the most commonly used data combination in parameter estimation studies of dark energy models.
For the SN data, the SNLS compilation~\cite{snls3} with a sample of 472 SNe is used in this work.
For the BAO data, we consider the $r_s/D_V(z)$ measurements from 6dFGS ($z=0.1$), SDSS-DR7 ($z=0.35$), SDSS-DR9 ($z=0.57$), and
WiggleZ ($z=0.44$, 0.60, and 0.73) surveys, where the three data from the WiggleZ survey are correlated (for the data and their inverse covariance matrix,
see, e.g., Ref.~\cite{wmap9}).
For the CMB data, we adopt the Planck distance posterior given by Ref.~\cite{WW}.
It should be noted that dark energy only affects the CMB through the comoving angular diameter distance to the decoupling epoch (and the late-time ISW effect),
and so the distance information given by the CMB distance posterior is sufficient for the joint geometric constraint on dark energy.
We also use the direct measurement result of the Hubble constant in the light of the cosmic distance ladder from the HST, $H_0=73.8\pm 2.4$ km s$^{-1}$ Mpc$^{-1}$~\citep{Riess2011}.
Our procedure is as follows. Dark energy models are first constrained by using the current joint SN+BAO+CMB+$H_0$ data, and then
the best-fit dark energy models are chosen to be the fiducial models in producing the simulated mock SL test data.
The obtained simulated SL test data are thus well consistent with the current SN+BAO+CMB+$H_0$ data.
Therefore, it is rather appropriate to combine the mock SL test data with the current SN, BAO, CMB, and $H_0$ data
for further constraining dark energy models.
Next, we briefly review how to simulate the SL test data.
This method is just to directly measure the redshift variation of quasar Lyman-$\alpha$ absorption lines.
The redshift variation is defined as a spectroscopic velocity shift \cite{loeb},
\begin{equation}\label{eq6}
\ \Delta v \equiv \frac{\Delta z}{1+z}=H_0\Delta t_o\bigg[1-\frac{E(z)}{1+z}\bigg],
\end{equation}
where $\Delta t_o$ is the time interval of observation, and $E(z)=H(z)/H_0$ is given by specific dark energy models. In a flat universe, we have
\begin{equation}\label{Ez}
E(z)=\sqrt{\Omega_r(1+z)^4+\Omega_m(1+z)^3+(1-\Omega_r-\Omega_m){X}(z)},
\end{equation}
where $\Omega_r$ and $\Omega_m$ are the present-day density parameters of radiation and matter, respectively, and ${X}(z)\equiv\rho_{\rm de}(z)/\rho_{\rm de}(0)=\exp[3\int_0^z {1+w(z')\over 1+z'}dz']$.
According to the Monte Carlo simulations, the uncertainty of $\Delta v$ measurements expected by CODEX can be expressed as~\cite{Liske}
\begin{equation}\label{eq7}
\sigma_{\Delta v}=1.35
\bigg(\frac{S/N}{2370}\bigg)^{-1}\bigg(\frac{N_{\mathrm{QSO}}}{30}\bigg)^{-1/2}\bigg(
\frac{1+z_{\mathrm{QSO}}}{5}\bigg)^{f}~\mathrm{cm}~\mathrm{s}^{-1},
\end{equation}
where $S/N$ is the signal-to-noise ratio defined per 0.0125 ${\AA}$ pixel, $N_{\mathrm{QSO}}$ is the
number of observed quasars, $z_{\mathrm{QSO}}$ represents their
redshift, and the last exponent $f=-1.7$ for $2<z<4$ and $f=-0.9$ for $z>4$. We choose $N_{\mathrm{QSO}}=30$ mock SL data uniformly distributed among
six redshift bins of $z_{\rm QSO}\in [2,~5]$.
To simulate the SL test data, we first constrain the dark energy models by using the current SN+BAO+CMB+$H_0$ data.
We perform an MCMC likelihood analysis~\cite{cosmomc} to obtain ${\cal O}(10^6)$ samples for each model.
The obtained best-fit parameters are substituted into Eq.~(\ref{eq6}) to get the central values of the SL test data,
and we typically take $\Delta t_o=20$ and 30 yr, in our analysis.
The error bars are directly computed from Eq.~(\ref{eq7}) with $S/N=3000$.
\section{Results and discussion}
\begin{table
\caption{Fit results for the $\Lambda$CDM, $w$CDM, and $w_0w_a$CDM models using the
current CMB+BAO+SN+$H_0$ data.}
\label{table1}
\renewcommand{\arraystretch}{1.5}
\centering
\begin{tabular}{cccccccccc}
\\
\hline\hline
Parameter & $\Lambda$CDM & $w$CDM & $w_0w_a$CDM\\ \hline
$\Omega_bh^2$ & $0.02237^{+0.00025}_{-0.00024}$
& $0.02218^{+0.00025}_{-0.00029}$
& $0.02221^{+0.00028}_{-0.00028}$
\\
$\Omega_ch^2$ & $ 0.1174^{+0.0014}_{-0.0016}$
& $0.1201^{+0.0020}_{-0.0022}$
& $0.1194^{+0.0026}_{-0.0027}$
\\
$w_0$ & $-1~({\rm fixed})$
& $-1.103^{+0.058}_{-0.058}$
& $-1.158^{+0.133}_{-0.115}$
\\
$w_a$ & $0~({\rm fixed})$
& $0~({\rm fixed})$
& $0.271^{+0.425}_{-0.638}$
\\
$\Omega_{m}$ & $0.2953^{+0.0084}_{-0.0092}$
& $0.2844^{+0.0104}_{-0.0093}$
& $0.2847^{+0.0089}_{-0.0116}$
\\
$H_0$ & $68.81^{+0.75}_{-0.66}$
& $70.74^{+1.26}_{-1.30}$
& $70.52^{+1.40}_{-1.02}$
\\
\hline
\end{tabular}
\end{table}
In this work, we make a comparison for the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models in the cosmological parameter constraints
with the SL test. First, we constrain the three dark energy models by using the current CMB+BAO+SN+$H_0$ data combination.
Detailed fit results are given in Table~\ref{table1}.
Indeed, as indicated in Ref.~\cite{hdeLi}, when a dynamical dark energy model is considered, the value of $H_0$ will become larger, relieving the tension between
Planck data and $H_0$ direct measurement.
Using the best-fit parameters given in Table~\ref{table1}, the SL test data for constraining each model can be simulated and will be used in the analysis.
\begin{figure}
\begin{center}
\includegraphics[width=13cm]{fig1.pdf}
\end{center}
\caption{Reconstructed redshift drift using the current
SN+BAO+CMB+$H_0$ constraint results for 20-yr and 30-yr observations of SL test. In this plot, 1$\sigma$ and 2$\sigma$ uncertainties are shown. Error bars (1$\sigma$) of the simulated SL test mock data are also shown for a direct comparison.}
\label{fig1}
\end{figure}
To directly compare the accuracies of the current actual data with the future SL test data, we reconstruct the velocity shifts for the three dark energy models by using the
fit results given in Table~\ref{table1}, plotted in Fig.~\ref{fig1}, as colored bands.
These bands are obtained using the Monte Carlo method. Based on the parameter spaces constrained from the current data combination, the boundaries of $\Delta v$ could be determined by using Eq.~(\ref{eq6}).
Red and blue bands are for the 20-yr and 30-yr velocity-shift reconstructions, respectively.
We also plot the error bars in the SL test, given by Eq.~(\ref{eq7}), on the corresponding bands, in order to make a direct comparison with the reconstructed results from the current data.
The case of the $w$CDM model has been discussed in Ref.~\cite{msl1}. Now, one can direct compare the three dark energy models.
The conclusion is the same. The 20-yr SL observation would significantly improve the
accuracy; a 30-yr SL observation would be closer to the current combined observations in accuracy, implying that the SL test as a high-redshift supplement to other geometric measurements
will play a crucial role in the parameter estimation in the forthcoming decades.\footnote{To be more quantitative, we
take the velocity shift $\Delta v$ at $z=4.5$ as an example: the ratio of $1\sigma$ uncertainty from SL-20yr data to that from current data, $\sigma_{\rm SL}/\sigma_{\rm current}$, is $1.45$, $1.24$, and $1.24$ for the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM model, respectively, while the ratio of $1\sigma$ uncertainty from SL-30yr data to that from current data, $\sigma_{\rm SL}/\sigma_{\rm current}$, is $0.97$, $0.82$, and $0.83$ for the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM model, respectively.}
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{fig2.pdf}
\end{center}
\caption{Comparison of the $H(z)$ direct measurement data in SL test with the reconstructed $H(z)$ from the current SN+BAO+CMB+$H_0$ fit results. 1$\sigma$ and 2$\sigma$ uncertainties are shown.
Here $H(z)/(1+z)$ (instead of $H(z)$) is shown for a clearer display. }
\label{fig2}
\end{figure}
The SL test directly measures the redshift drifts in the range of $2\lesssim z\lesssim 5$; in other words, the SL test directly measures the Hubble expansion rate $H(z)$
at the high redshifts.
It is well known that the Hubble parameter $H(z)$ is related to the equation of state of dark energy through one integral [see Eq. (\ref{Ez}) for the integral $\int_0^z {1+w(z')\over 1+z'}dz'$ in $X(z)$], and the luminosity distance $d_L(z)$ (or the
angular diameter distance $d_A(z)$) is related to the equation of state of dark energy through two integrals
[both $d_L(z)$ and $d_A(z)$ are proportional to $D(z)=\int_0^z dz'/H(z')$ in a flat universe].
Thus, the direct measurements of $H(z)$ are of extreme
importance for constraining the property of dark energy.
If these high-redshift $H(z)$ data can be combined with some accurate low-redshift $H(z)$ data provided by other astrophysical methods, the capability of constraining dark energy
would be enormous.
Even though there are no such accurate low-redshift $H(z)$ data (there are indeed some low-redshift $H(z)$ data, but they are not sufficiently accurate \cite{Darling,Hz1,Hz2}), the high-redshift data
given by the SL test in combination with other low-redshift observations (such as SN and BAO) will also play a very significant role in constraining dark energy models.
Now, we wish to have a look at what accuracies the $H(z)$ measurements provided by SL test could achieve.
Hence, we plot the $H(z)$ evolutions for the three dark energy models in Fig.~\ref{fig2}.
In order to show the results more clearly, we actually plot the evolutions of $H(z)/(1+z)$ in this figure.
The red bands stand for the reconstructed $H(z)/(1+z)$ evolutions (with 1 and 2$\sigma$ uncertainties) for the three dark energy models from the fits to the current
SN+BAO+CMB+$H_0$ data, and the blue and black bars on the bands stand for the error bars of $H(z)/(1+z)$ measurements by 20-yr and 30-yr
observations of the SL test, respectively.
It can be seen from this figure that the accuracies of the SL high-redshift $H(z)$ data are not high compared to that of current data. However, in our following discussion, we will show that these SL high-redshift $H(z)$ data are extremely useful in breaking the significant degeneracies among the cosmological parameters in current data.
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{fig3.pdf}
\end{center}
\caption{Constraints (68.3\% and 95.4\% CL) in the $\Omega_m$--$H_0$ plane for $\Lambda$CDM, $w$CDM, and $w_0w_a$CDM models with current only, current+SL 20-yr and current+SL 30-yr data.}
\label{fig3}
\end{figure}
\begin{table*}\tiny
\caption{Errors of parameters in the $\Lambda$CDM ($\Lambda$), the $w$CDM ($w$), and the $w_0w_a$CDM ($w_0w_a$) models for the fits to
current only, current+SL 20-yr and current+SL 30-yr data.}
\label{table2}
\small
\setlength\tabcolsep{2.8pt}
\renewcommand{\arraystretch}{1.5}
\centering
\begin{tabular}{cccccccccccc}
\\
\hline\hline &\multicolumn{3}{c}{current only} &&\multicolumn{3}{c}{current + SL 20-yr} &&\multicolumn{3}{c}{current + SL 30-yr} \\
\cline{2-4}\cline{6-8}\cline{10-12
Error & $\Lambda$ & $w$ & $w_0w_a$ & & $\Lambda$ & $w$ & $w_0w_a$& & $\Lambda$ & $w$ & $w_0w_a$\\ \hline
$\sigma(w_0)$ & $-$
& $0.082$
& $0.204$&
& $-$
& $0.068$
& $0.173$&
& $-$
& $0.062$
& $0.165$\\
$\sigma(w_a)$ & $-$
& $-$
& 0.767&
& $-$
& $-$
& $0.690$&
& $-$
& $-$
& $0.648$
\\
$\sigma(\Omega_{m})$ & $0.0125$
& $0.0140$
& $0.0146$&
& $0.0040$
& $0.0037$
& $0.0064$&
& $0.0028$
& $0.0026$
& $0.0052$\\
$\sigma(H_0)$ & $1.00$
& $1.81$
& $1.73$&
& $0.38$
& $0.76$
& $0.76$&
& $0.30$
& $0.64$
& $0.65$\\
\hline
\end{tabular}
\end{table*}
In the existing data, in particular the Planck CMB data, the strong degeneracy between $\Omega_m$ and $H_0$ is well known.
We shall show that the SL test data can effectively break this degeneracy and thus help constrain the parameters $\Omega_m$ and $H_0$
to a high precision.
Figure~\ref{fig3} shows the joint constraints on the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models in the $\Omega_m$--$H_0$ plane.
The 68.3\% and 95.4\% CL posterior distribution contours are shown. The data combinations used are the current only, the current+SL 20-yr,
and the current+SL 30-yr combinations, and their constraint results are shown with white, red, and blue contours, respectively.
One can clearly see that the degeneracy between $\Omega_m$ and $H_0$ is well broken with the SL test data for all the three dark energy models.
The 1$\sigma$ errors of the parameters $w_0$, $w_a$, $\Omega_m$, and $H_0$ for the three models for the above three data combinations are given in Table~\ref{table2}.
From this table, one can directly figure out how the SL test data help improve the constraints.
With the 20-yr SL observation, the constraints on $\Omega_m$ and $H_0$ will be improved, respectively, by
68.0\% and 62.0\% for the $\Lambda$CDM model, by 73.6\% and 58.0\% for the $w$CDM model, and by $56.2\%$ and 56.1\% for the $w_0w_a$CDM model.
With the 30-yr SL observation, the constraints on $\Omega_m$ and $H_0$ will be improved, respectively, by
77.6\% and 70.0\% for the $\Lambda$CDM model, by 81.4\% and 64.6\% for the $w$CDM model, and by $64.4\%$ and 62.4\% for the $w_0w_a$CDM model.
Therefore, we can see that with a 30-yr observation of the SL test the geometric constraints on dark energy would be improved enormously.
For all the three dark energy models, the constraints on $\Omega_m$ and $H_0$ would be improved, relative to the current joint observations, by more than 60\%, with the SL 30-yr observation.
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{fig4.pdf}
\end{center}
\caption{ Constraints on the $\Lambda$CDM and the $w$CDM models in the $\Omega_m$--$H_0$ plane from the current only, SL 30-yr and current+SL 30-yr data. Owing to the fact that the orientations of the degeneracies in the two cases of current only and SL 30-yr constraints are very different, the strong degeneracy between $\Omega_m$ and $H_0$ in the current data is well broken by the SL test.}
\label{fig4}
\end{figure}
Figure~\ref{fig4} shows how the SL test breaks the strong degeneracy between $\Omega_m$ and $H_0$ in the current data constraint.
Here we take the $\Lambda$CDM model and the $w$CDM model as examples.
The white contours are for the constraints from the current combined geometric observations, and the gray contours are for the constraints from the SL 30-yr only observation.
One can see clearly that the strong degeneracy between $\Omega_m$ and $H_0$ appears in both cases of the current only constraint and the SL 30-yr only constraint,
but the degeneracy orientations in the two cases are very different, and the strong degeneracy in the current data is thus well broken by the SL test.
The blue contours are for the results of the joint current+SL 30-yr data constraints, from which one can easily see that once the high-redshift SL test data are combined with the
current geometric observations the capability of constraining dark energy would be enhanced enormously.
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{fig5.pdf}
\end{center}
\caption{The one-dimensional posterior distributions of $w$ for the $w$CDM model (left) and the two-dimensional posterior distributions of $w_0$ and $w_a$ for the $w_0w_a$CDM model (right), from the current only, current+SL 20-yr, and current+SL 30-yr constraints.}
\label{fig5}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=11cm]{fig6.pdf}
\end{center}
\caption{Reconstructed $w(z)$ evolutions in the $w_0w_a$CDM model by using the constraint results of the current only and the current+SL 30-yr data. In this plot, 1$\sigma$ and 2$\sigma$ uncertainties are shown. }
\label{fig6}
\end{figure}
We also discuss the impact of the SL test data on constraining the dark energy equation of state. The case for the $w$CDM model has been discussed in Ref.~\cite{msl1}. In this paper,
we will analyze the case for the $w_0w_a$CDM model, and make a comparison for the two cases.
In Fig.~\ref{fig5} we show the one-dimensional posterior distributions of $w$ for the $w$CDM model and the two-dimensional posterior distributions of $w_0$ and $w_a$ for the
$w_0w_a$CDM model, from the current only, current+SL 20-yr, and current+SL 30-yr constraints.
The corresponding errors of $w_0$ and $w_a$ are given in Table~\ref{table2}.
For the $w$CDM model, the constraints on $w$ can be improved by 17.1\%, and 24.4\%, with 20-yr, and 30-yr observations, respectively.
For the $w_0w_a$CDM model, the SL 20-yr observation helps improve the constraints on $w_0$ and $w_a$ by 15.2\% and 10.0\%, respectively;
the SL 30-yr observation helps improve the constraints on $w_0$ and $w_a$ by 19.1\% and 15.5\%, respectively.
Therefore, we conclude that a 30-yr observation of the SL test can help improve the constraint on constant $w$ by about 25\%, and
improve the constraints on $w_0$ and $w_a$ by about 20\% and 15\%, respectively.
We also see that the SL test data cannot break the degeneracy between $w_0$ and $w_a$.
Furthermore, in Fig.~\ref{fig6} we reconstruct the $w(z)$ evolutions in the $w_0w_a$CDM model by using the constraint results of the current only and the current+SL 30-yr data.
From the comparison, we find that the SL test cannot greatly improve the reconstruction of $w(z)$.
In fact, the conclusion that the SL test could not break the current degeneracy between $w_0$ and $w_a$ has also been drawn in Ref. \cite{sl7}, and the possible reason has
been discussed in the same paper (for more details, see Sec. V of Ref. \cite{sl7}). However, we will see in the next section that the future high-precision SN and BAO observations could break the
degeneracy between $w_0$ and $w_a$ and measure both of them to a high precision, and in this case the SL test would further improve the measurement precision of $w_a$ by
more than 50\%.
\begin{figure}
\begin{center}
\includegraphics[width=13cm]{fig7.pdf}
\end{center}
\caption{Reconstructed $H(z)$ evolutions in the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models, by using the
fit results from the current only and the current+SL 30-yr data. Only 1$\sigma$ uncertainties are shown in this plot.}
\label{fig7}
\end{figure}
We are also interested in the reconstruction of $H(z)$ with the SL test data.
In Fig.~\ref{fig7} we show the reconstructed $H(z)$ evolutions in the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models, by using the
fit results from the current only and the current+SL 30-yr data.
From Fig.~\ref{fig2} we have learned that the accuracy of the high-redshift $H(z)$ direct measurements with the SL test is worse than that of the reconstructed $H(z)$ results from the
current combined data constraint. However, owing to the fact that the SL test data can break the parameter degeneracies in the current low-redshift geometric measurements,
the reconstructed $H(z)$ results are improved in the low redshifts with the help of the SL test data.
In the current constraints on dark energy, there are important parameter degeneracies, and we have shown that the future redshift-drift observations could play a crucial role in breaking these degeneracies. The next step is to test what role the redshift-drift measurements would play in the future combined geometric constraints.
\section{Extended discussion concerning future geometric measurements}
In the above discussion, we showed how the future redshift-drift measurements would break the parameter degeneracies in the current geometric observations. However, when the CODEX experiment is ready to deliver its redshift-drift data in the future, other future geometric measurements data will also be available. Therefore, a further issue is to ask what role the SL test will play in improving the dark energy constraints in the future geometric measurements.
According to the report of the dark energy task force~\cite{DETF}, the most important future geometric measurements include the SN and the BAO observations (by the way, the most important structure growth measurements include the weak lensing and the galaxy clusters observations~\cite{DETF}). So in what follows we will only discuss the future long-term SN and BAO observations.
We simulate the future geometric measurements data using the method described in Ref.~\cite{DETF}.
As a concrete example, we simulate the future data based on the long-term space-based project called JDEM (Joint Dark Energy Mission) and simply describe the method in the following. For the details, we refer the reader to Ref.~\cite{DETF}.
For future SN data, aside from 2000 SNe distributed in 16 bins over the range from $z=0.1$ to $z=1.7$, a near sample of 500 SNe within $0.03<z<0.08$ is also considered.
The observables for SN data are apparent magnitudes $m_i=M+\mu(z_i)$, where $M$ represents the absolute magnitude, and $\mu(z_i)$ is the distance modulus given by
$\mu(z_i)= 5 \log_{10} d_L(z_i)+25$.
The luminosity distance $d_L(z)=(1+z)\int_0^z dz'/H(z')$ for a flat universe.
The uncertainty of apparent magnitude $m_i$ due solely to variation in the properties of SN is denoted as $\sigma_D$. Besides, we use two additional nuisance parameters $\mu^L$ and $\mu^Q$ to give a quadratic $z$-dependent effect of the peak luminosity of SNe: $\mu(z_i)\rightarrow\mu(z_i)+\mu^Lz_i+\mu^Qz_i^2$. For the near sample, an additional nuisance parameter $\mu^S$ is included to represent an offset between the photometric systems of the near and far samples: $\mu(z_i)\rightarrow\mu(z_i)+\mu^S$. The uncertainties of these parameters used to simulate mock future data are $\sigma_D=0.10$, $\sigma_{\mu^L}=\sigma_{\mu^D}=0.01/\sqrt2$, and $\sigma_{\mu^S}=0.01$.
We simulate 10000 mock BAO data uniformly distributed among 10 redshift bins of $z\in [0.5,~2]$, with each $\Delta z_i$ centered on the grid $z_i$. The observables are expansion rate $H(z)$ and comoving angular diameter distance $d_A^{co}(z)=d_L(z)/(1+z)$. The uncertainty of $\ln H(z_i)$ and $\ln d_A^{co}(z_i)$ can be expressed as
\begin{equation}
\sigma_H^i=x_0^H\frac{4}{3}\sqrt{\frac{V_0}{V_i}}f_{nl}(z_i),
\end{equation}
\begin{equation}
\sigma_d^i=x_0^d\frac{4}{3}\sqrt{\frac{V_0}{V_i}}f_{nl}(z_i),
\end{equation}
where the comoving survey volume in the redshift bin of $z_i$ is $V_i=1500(d_A^{co}(z_i))^2/H(z_i)$, and the erasure of the baryon features by non-linear evolution is factored in using $f_{nl}(z_i)=1$ for $z_i>1.4$ and $f_{nl}(z_i)=(1.4/z_i)^{1/2}$ for $z_i<1.4$. The parameters used in our simulation are $x_0^H=0.0148$, $x_0^d=0.0085$ and $V_0=\frac{2.16}{h^3}$ Gpc$^3$.
We also consider systematic errors in the BAO observation, which are modeled for both types of observable as independent uncertainties in the log of the distance measures in each redshift bin: $\sigma_s^i=0.01\sqrt{\frac{0.5}{\Delta{z_i}}}$, with $\Delta{z_i}=0.15$.
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{fig8.pdf}
\end{center}
\caption{Constraints (68.3\% and 95.4\% CL) in the $\Omega_m$--$H_0$ plane for $\Lambda$CDM, $w$CDM, and $w_0w_a$CDM models with future only and future + SL 30-yr data.}
\label{fig8}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{fig9.pdf}
\end{center}
\caption{The one-dimensional posterior distributions of $w$ for the $w$CDM model (left) and the two-dimensional posterior distributions of $w_0$ and $w_a$ for the $w_0w_a$CDM model (right), from the future only and the future + SL 30-yr constraints.}
\label{fig9}
\end{figure}
\begin{table*}\tiny
\caption{Errors of parameters in the $\Lambda$CDM ($\Lambda$), the $w$CDM ($w$), and the $w_0w_a$CDM ($w_0w_a$) models for the fits to
future only and future + SL 30-yr data.}
\label{table3}
\small
\renewcommand{\arraystretch}{1.5}
\centering
\begin{tabular}{ccccccccc}
\\
\hline\hline &\multicolumn{3}{c}{future only} &&\multicolumn{3}{c}{future + SL 30-yr} \\
\cline{2-4}\cline{6-8}
Error & $\Lambda$ & $w$ & $w_0w_a$ & & $\Lambda$ & $w$ & $w_0w_a$\\ \hline
$\sigma(w_0)$ & $-$
& $0.0083$
& $0.0091$&
& $-$
& $0.0067$
& $0.0090$\\
$\sigma(w_a)$ & $-$
& $-$
& $0.208$&
& $-$
& $-$
& $0.079$
\\
$\sigma(\Omega_{m})$ & $0.00078$
& $0.0021$
& $0.0108$&
& $0.00075$
& $0.0016$
& $0.0031$\\
$\sigma(H_0)$ & $0.0223$
& $0.0296$
& $0.0449$&
& $0.0218$
& $0.0286$
& $0.0349$\\
\hline
\end{tabular}
\end{table*}
Figure~\ref{fig8} shows the joint constraints on the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models in the $\Omega_m$--$H_0$ plane.
The 68.3\% and 95.4\% CL posterior distribution contours are shown. The data combinations used are the future only and the future + SL 30-yr combinations, and their constraint results are shown with white and blue contours, respectively.
The 1$\sigma$ errors of the parameters $w_0$, $w_a$, $\Omega_m$, and $H_0$ for the three models for the above data combinations are given in Table~\ref{table3}.
Note that here we use ``future'' to denote the data combination of future SN and BAO for convenience.
It is shown that with the 30-yr SL observation, the constraints on $\Omega_m$ and $H_0$ will be improved by
3.8\% and 2.2\% for the $\Lambda$CDM model, by 23.8\% and 3.4\% for the $w$CDM model, and by $71.3\%$ and 22.3\% for the $w_0w_a$CDM model.
In Fig.~\ref{fig9}, we show the one-dimensional posterior distributions of $w$ for the $w$CDM model (left panel) and the two-dimensional posterior distributions of $w_0$ and $w_a$ for the
$w_0w_a$CDM model (right panel), from the future only and the future + SL 30-yr constraints.
The corresponding 1$\sigma$ errors of $w_0$ and $w_a$ are given in Table~\ref{table3}.
For the $w$CDM model, the constraints on $w$ can be improved by 19.3\%, with the SL 30-yr observation.
For the $w_0w_a$CDM model, the SL 30-yr observation helps improve the constraints on $w_0$ and $w_a$ by 1.1\% and 62.0\%, respectively.
Therefore, we find that the redshift-drift observation could also play an important role in improving the dark energy constraints from the future geometric measurements,
especially for the $w_0w_a$CDM model. In the future geometric constraints for the $w_0w_a$CDM model, the SL 30-yr observation would help improve the measurement precisions of
$\Omega_m$, $H_0$, and $w_a$ by more than 70\%, 20\%, and 60\%, respectively.
\section{Summary}
In this paper, we have discussed the parameter estimation for the $\Lambda$CDM, the $w$CDM, and the $w_0w_a$CDM models with the future SL test data.
The SL test directly measures the expansion rate of the universe in the redshift range of $2\lesssim z \lesssim 5$ by detecting redshift drift in the spectra of Lyman-$\alpha$ forest
of distant QSOs, thus as a purely geometric measurement it provides an important supplement to other dark energy probes.
Following our previous work \cite{msl1}, in order to guarantee that the simulated SL test data are consistent with the other geometric measurement data,
we used the best-fitting dark energy models constrained by the current combined geometric measurement data as the fiducial models to produce the mock SL test data and
then used these simulated data to do the analyses.
We showed that the SL test data are extremely helpful in breaking the existing parameter degeneracies.
The strong degeneracy between $\Omega_m$ and $H_0$ in the current SN + BAO + CMB + $H_0$ constraint results for all the three models can be well broken by the SL test.
By analyzing and comparing the 20-yr and 30-yr observations of SL test, we found that the 30-yr observation could provide sufficiently important supplement to the other
observations. Compared to the current SN + BAO + CMB + $H_0$ constraint results, the 30-yr observation of SL test could improve the constraints on
$\Omega_m$ and $H_0$ by more than 60\% for all the three models.
But the SL test can only moderately improve the constraint on the equation of state of dark energy.
We showed that a 30-yr observation of SL test could help improve the constraint on constant $w$ by about 25\%, and improve the constraints on $w_0$ and $w_a$ by
about 20\% and 15\%, respectively.
We also analyzed how the SL test would impact on the dark energy constraints from the future geometric measurements. To do this analysis, we simulated the future SN and BAO
data based on the long-term space-based project JDEM. We found that the SL test could play a crucial role in the future joint geometric constraints. For example, the 30-yr observation of
SL test would help improve the measurement precision of $\Omega_m$, $H_0$, and $w_a$ by more than 70\%, 20\%, and 60\%, respectively, for the $w_0w_a$CDM model.
As a purely geometric measurement, the SL test has been proven to be a very important supplement to the other geometric measurement observations.
Actually, in order to differentiate the noninteracting dark energy, interacting dark energy, and modified gravity models, the geometric measurements should be compared to
the measurements of the growth of large-scale structure. A consistency test of the geometric and structural measurements might provide a diagnostic to the cause of the
acceleration of the universe in the future. Of course, the SL test will definitely play a significant role in doing such an analysis.
For the interacting dark energy models, the longstanding problem of large-scale instability was recently resolved by establishing a parameterized post-Friedmann framework
for interacting dark energy~\cite{PPF,PPF2}. Thus, the interacting dark energy models with the background interaction forms of both $Q\propto \rho_c$ and $Q\propto \rho_{de}$ are now
proven to be well behaved. It is, undoubtedly, worthy to study the interacting dark energy models with the SL test. In Ref.~\cite{msl1}, a preliminary SL test analysis has been
made for the constant $w$ model with $Q=\gamma H\rho_c$ and $Q=\gamma H\rho_{de}$. However, an analysis for the models with $Q=\Gamma \rho_c$ and $Q=\Gamma \rho_{de}$
(here $\Gamma$ is a constant) is still absent. We will leave the complete analysis for interacting dark energy models and modified gravity models in future work.
\acknowledgments
We acknowledge the use of {\tt CosmoMC}.We thank Yun-He Li for helpful discussion.
JFZ is supported by the Provincial Department of Education of
Liaoning under Grant No.~L2012087.
XZ is supported by the National Natural Science Foundation of
China under Grant No.~11175042 and the Fundamental Research Funds for the
Central Universities under Grant No.~N120505003.
|
\section{Introduction}\label{sec:introduction}
Crisis situations such as disasters brought on by natural hazards present unique challenges to those who study them, creating conditions that call for particular research methods~\cite{stallings_2012_unique}.
In this paper, we survey methods for studying disasters from the perspective of information processing and management, specifically methods for processing social media content.
Crisis situations---particularly those with little to no warning (known as ``sudden onset crises")---generate a situation that is rife with questions, uncertainties, and the need to make quick decisions, often with minimal information. When it comes to information scarcity, research in recent years has uncovered the increasingly important role of social media communications in disaster situations, and shown that information broadcast via social media can enhance situational awareness during a crisis situation~\cite{vieweg_2012_thesis}.
However, social media communications during disasters are now so abundant that it is necessary to sift through hundreds of thousands, and even millions, of data points to find information that is most useful during a given event.
The goal of this survey is to provide computer science researchers and software developers with computational methods they can use to create tools for formal response agencies, humanitarian organizations, and other end users with a way to successfully identify, filter, and organize the overwhelming amount of social media data that are produced during any given crisis. Such tools can help stakeholders make time-critical---and potentially life-saving---decisions.
\subsection{Social Media During Crisis Situations}\label{subsec:motivation}
\spara{Brief History.}
The use of Internet technologies to gather and disperse information in disaster situations, as well as to communicate among stakeholders, dates back to the late 1990s. Internet historians point to online newsgroups and email clients that were used to coordinate protests in Indonesia in 1998~\cite{poole2005internet}. In addition, there are cases of websites being set up in response to crises in 2003~\cite{palen2007citizen}.
To the best of our knowledge, 2004 is the first year in which a user-generated content website was used in response to a crisis; after the Indian Ocean Tsunami of December 26 that year, an electronic bulletin board was set up and moderated for 10 days.\footnote{\url{http://www.thefreelibrary.com/www.p-h-u-k-e-t.com+Has+Served+Its+Purpose+After+the+Tsunami\%3B+Site...-a0126803919}}
In addition, in the aftermath of Hurricane Katrina, which struck the city of New Orleans in the United States in 2005, significant emergency response activity took place on MySpace~\cite{shklovski_2010_katrina}.
One of the earliest known cases of people using microblogging service Twitter in an emergency was during severe wildfires that took place near San Diego, California (in the United States) in 2007.\footnote{Eric Frost, personal communication.} Since then, it has become common practice for affected populations and concerned others to use Twitter to communicate, ask questions, collect and spread information, and organize response efforts (among other tasks)~\cite{starbird2010chatter,Vieweg:2010,sarcevic2012beacons,starbird2013delivering,imran2014aidr,cobb2014designing}.
\spara{Today.}
The growing adoption of social media during disasters has created opportunities for information propagation that would not exist otherwise. Emergency response agencies routinely post information such as emergency alerts and advice through these channels,\footnote{See e.g. \url{https://blog.twitter.com/2013/twitter-alerts-critical-information-when-you-need-it-most}} but social media enables much more than ``top-down'' communications.
People post situation-sensitive information on social media related to what they experience, witness, and/or hear from other sources~\cite{hughes2009twitter}. This practice allows both affected populations and those outside the impact zone to learn about the situation \textit{first hand} and in near real-time.
We know that information posted to social media platforms in time- and safety-critical circumstances can be of great value to those tasked with making decisions in these fraught situations. Previous research has shown that information which contributes to situational awareness is reported via Twitter (and other social media platforms) during mass emergencies~\cite{Vieweg:2010,vieweg_2012_thesis,imran2014aidr}. Now, those tasked with formal response efforts---from local fire departments\footnote{\url{http://edition.cnn.com/2012/11/01/tech/social-media/twitter-fdny/}} to international aid agencies---are working to incorporate information broadcast on social media platforms into their processes and procedures. Many emergency responders and humanitarian officials recognize the value of the information posted on social media platforms by members of the public (and others), and are interested in finding ways to quickly and easily locate and organize that information that is of most use to them~\cite{hughes_2012_phd}.\footnote{Andrej Verity, personal communication}
Some agencies have even begun to formally incorporate social media monitoring and communication during mass emergency situations. The American Red Cross (ARC), in a survey, reported the effectiveness of social media and mobile apps.\footnote{\url{http://www.redcross.org/news/press-release/More-Americans-Using-Mobile-Apps-in-Emergencies}} ARC recently opened their Social Media Digital Operations Center for Humanitarian Relief. The goals of the center are to ``source additional information from affected areas during emergencies to better serve those who need help; spot trends and better anticipate the public's needs; and connect people with the resources they need, like food, water, shelter or even emotional support."\footnote{http://www.redcross.org/news/press-release/The-American-Red-Cross-and-Dell-Launch-First-Of-Its-Kind-Social-Media-Digital-Operations-Center-for-Humanitarian-Relief}
Though the ARC is currently one of the few (possibly the only) formal agencies to support such a center, it is likely that similar operations will begin within other organizations. Among other similar examples, the Australia Crisis Tracker,\footnote{\url{http://www.crisistracker.com.au/}} which is a machine-learning based tool to filter spam and categorize data into different event types, is also deployed with the Australian Red Cross.\footnote{\url{http://www.research.ibm.com/articles/crisis-tracker.shtml}}
Though formal response agencies express interest in incorporating social media into their processes, obstacles exist. For example, a recent survey by the US Congressional Research Service cites administrative cost as a significant barrier to adopting social media during emergencies:
``The number of personnel required to monitor multiple social media sources, verify the accuracy of incoming information, and respond to and redirect incoming messages is also uncertain ... Responding to each message in a timely manner could be time consuming and might require an increase in the number of employees ...''~\cite{lindsay_2011_social_media_disasters}
Others have expressed concerns including issues related to roles, responsibilities, and liabilities; difficulties evaluating the veracity, trustworthiness, and reliability of information; and information overload in general~\cite{vieweg_2014_integrating,hughes_2014_smem}.
Computational methods can help overcome some of these obstacles, by reducing the amount of information to be examined by humans. Automatic methods are necessary when human computation is limited, and in the following sections, we detail what those methods entail.
\subsection{Background Reading}\label{subsec:background-readings}
Sociologists began researching human behavior in mass emergency situations long before the Internet, or even modern computing.
The purpose of this section is not to provide the reader with an exhaustive list of sociology of disaster literature; we highlight a few foundational readings that are helpful for the computer science, information science, technology, and social media scholars to gain quick insight into the rich and varied field of sociology of disaster.
E.L. Quarantelli's 2002 chapter ``The Disaster Research Center (DRC) Field Studies of Organized Behavior in the Crisis Time Period of Disasters'' (in \textit{Methods of Disaster Research} edited by R.A. Stallings~\shortcite{quarantelli_2002_drc}) provides a brief history of one of the foremost disaster research institutes in the United States. Quarantelli gives background on the Disaster Research Center, and explains the strategic as well as academically-oriented decisions that were made in order to highlight the importance of studying the social science aspects of disaster.
In his edited volume ``Disasters by Design,'' Dennis S. Mileti~\shortcite{mileti_1999_disasters} and the contributing authors aim to reach a general (i.e. non-academic) audience and provide background on disasters caused by natural hazards. The volume is comprised of ``synthesized statements of what is known, collectively, about hazards and human coping strategies.'' Mileti and colleagues point to causes of disaster, which happen when three major systems---the physical, social, and built environments---interact in complex ways. The authors' goal is to give the reader a way to understand how to study disaster situations, with a final goal of helping members of the public create more resilient communities.
When it comes to combining studies of disaster with the use of information technology, including social media, the Harvard Humanitarian Initiative~\shortcite{hhe_2011_disaster2} presents an in-depth analysis of the response to the earthquake in Haiti in January 2010.
With a stronger focus in social media, a recent survey by Hughes, Peterson and Palen considers the motivating factors of emergency responders regarding their use of social media data. The authors describe the challenges they face, best practices regarding the adoption of social media by formal response organizations, and also touch on instances of integrated, end-to-end systems that are currently being built to meet these needs~\cite{hughes_2014_smem}. In addition, an article by Palen and Liu~\shortcite{palen2007citizen} was one of the first to provide an early assessment regarding how information and communication technology can support the participation of the public during crisis situations. Since then, many articles that focus on the role of social media in disaster have been published, but the two we mention here provide a good ``first glance'' to readers who are new to the field.
Our brief overview of foundational reading would not be complete without mentioning the much-discussed issue of trust and the use of social media. A recent ACM Computing Survey looks at this very topic ~\cite{Sherchan:2013:STS:2501654.2501661}. The authors review the various definitions of ``trust'' from a variety of academic disciplines, discuss the factors that contribute to notions of trust, and combine the complex and much-scrutinized idea of trust with computing and social network research.
\subsection{Scope and Organization}
The overarching problem we aim to confront in this article is that of extracting time-critical information from social media that is useful for emergency responders, affected communities, and other concerned populations in disaster situations.
We note that social media analysis has been used for a number of applications in the domains of economics, politics (e.g.~\citeN{hong2011does}) and public health (e.g.~\citeN{aramaki2011twitter}); for a survey, see~\citeN{cikm_2013_twitterreal}. Even if we consider only applications to time-critical settings, we note that often the same methods described in this survey can be applied to the analysis of social media during mass converge events, such as large political conventions, concerts, or sports events; or for monitoring the performance of a media campaign or a televised debate, among similar applications. However, while many of the methods and algorithms that we describe can be used for other purposes, we explain them from the perspective of their applications to mass emergencies, a topic that has a specific scientific and technical community, and that targets a particular set of use-cases.
The following two sections briefly describe our target end-user audience, and their information needs (Section~\ref{sec:users}), and end-to-end integrated systems (Section~\ref{sec:systems}). The subsequent sections form the main technical part of this survey and present a systematic analysis of the computational methods we cover
\begin{itemize}
\item Section~\ref{sec:data} starts with a general characterization of social media messages broadcast during disasters. Next, it introduces methods for programmatically acquiring data from social media, and for pre-processing the data.
\item Section~\ref{sec:event-detection} covers methods for the detection of new events, which involves detecting the first message on a given topic or sub-topic. The section also covers how to track these events, i.e. how to collect further messages belonging to the same topic.
\item Section~\ref{sec:finding_information} outlines methods to mine and aggregate information. These methods include unsupervised classification (and clustering), supervised classification, information extraction, and summarization.
\item Section~\ref{sec:semantics} presents how semantic technologies can be applied in this domain. This corresponds, first, to enrich the content with semantic information, and second, to use an ontology for disaster management to describe the content of the messages.
\end{itemize}
The final section concludes the survey, and outlines current research directions.
\section{Users and Information Needs}\label{sec:users}
Much of the research we present here focuses on the computational aspects of processing social media messages in time- and safety-critical situations. It is additionally important to consider the end users of these technological solutions; those who benefit from having curated information that describes a disaster or crisis and enhances situational awareness, including formal response agencies, humanitarian organizations, and members of the public.
\subsection{Public Participation in Crises}
Ideally, to understand how the public participates in social media during crises, we should start by asking how the public reacts in general to crises.
Contrary to Hollywood renditions of disaster situations, human response to crises is not one of panic and mayhem~\cite{mitchell2000catastrophe}. Victims of disaster do not lose control, run amok, nor flee the area in fear. Instead, they make quick decisions based on the information available to them at the time, which often allow them to save their own lives as well as help those around them~\cite{mitchell2000catastrophe}.
Neighbors, friends, and other members of the public are the first to respond when a disaster strikes. They rush to the scene to perform search and rescue operations, administer first aid, and perform critical tasks necessary in the first moments of response. Often, these ``first responders'' are victims of the disaster themselves~\cite{dynes1970organized}.
The role of the public in disaster response efforts is critical, and with the growing use of social media to gather and disperse information, organize relief efforts, and communicate, those members of the public who can play a valuable role in these situations is no longer limited to those in the area of impact.
As~\citeN{dynes1994community} explains, emergencies do not render victims incapable of helping themselves and others, nor create a situation in which they are unable to make intelligent, personally meaningful decisions. What emergencies \textit{do} create is an environment in which new and perhaps unexpected problems are presented, which members of the public are called upon to solve.
Research in recent years on the use of social media in disasters shows how members of the public, formal response agencies, and other stakeholders have taken to online outlets to perform tasks such as
communicating about hospital availability~\cite{starbird2013delivering},
coordinating medical responses~\cite{sarcevic2012beacons},
and communicating with the public during various crises~\cite{cobb2014designing}, among many others.
These users interact in complex ways including producing, distributing and organizing content~\cite{starbird2010chatter}.
\subsection{Differences in Information Needs}
The recognition that social media communications are a valid and useful source of information throughout the disaster lifecycle (preparation, impact, response, and recovery) is increasing among the many stakeholders who take action in disaster situations. In particular, members of the public, formal response agencies, and local, national and international aid organizations are all aware of the ability to use social media to gather and disperse timely information in the aftermath of disaster, but the specific information they seek---and their ability to put it to use---may differ~\cite{vieweg_2012_thesis}.
Depending on the circumstances of the disaster, and what roles and duties the various stakeholders are responsible for, their specific information needs will vary. For example, in a wildfire situation that affects a community, members of a formal response organization such as local police or area firefighters can benefit from information such as where people are smelling smoke, what precautions they are taking (e.g. clearing brush, watering yards), and what traffic patterns look like. In a large-scale, sudden-onset disaster such as a typhoon or earthquake, humanitarian agencies, such as the various branches of the United Nations,
benefit from information that details the current situation ``on the ground,'' such as where electricity has been disabled, or where people are without food and water.
In any disaster situation, members of the public play a variety of roles and take on many tasks; the information they find valuable may be very personal---i.e. hearing that a friend or loved one is safe, or it may be more broadly applicable, such as the status of a certain neighborhood or town.
Overall, the information any individual, group, or organization finds useful and seeks out in a disaster will depend upon their goals. Is it a group interested in providing food to children? Is it an organization that can set up a field hospital? Is it an individual living in a foreign country who is concerned about her or his family? The different types of information sought by these different stakeholders may be broadcast on Twitter, but to find it quickly, users rely on technological methods to sift through the millions of tweets broadcast at any given time to find useful information. Further information and a deeper perspective on users of social media in disaster can be found in \citeN{hughes_2014_smem} and \citeN{hughes2012evolving}.
\section{Systems for Crisis-Related Social Media Monitoring} \label{sec:systems}
Table~\ref{tbl:systems} provides examples of existing systems described in the literature that extract crisis-relevant information from social media.\footnote{The list is not extensive, and does not include tools such as Radian6 (\url{http://www.salesforcemarketingcloud.com/products/social-media-listening/}) that have not been described in the literature, but might be relevant for other reasons---e.g. in the case of Radian6, because it is used by the American Red Cross.}
The systems we include have varying degrees of maturity; some have been deployed in real-life situations, while others remain under development.
\begin{table}[t]
\caption{Example systems described in the academic literature that extract crisis-relevant information from social media.}
\label{tbl:systems}
\scriptsize\begin{tabular}{p{3.1in}l}\toprule
System name \\
~~ Data; example capabilities & Reference and URL \\\midrule
\textit{Twitris} & \cite{sheth2010understanding,purohit2013twitris} \\
~~ Twitter; semantic enrichment, classify automatically, geotag & \url{http://twitris.knoesis.org/} \\
\textit{SensePlace2} & \cite{maceachren2011senseplace2} \\
~~ Twitter; geotag, visualize heat-maps based on geotags & \url{http://www.geovista.psu.edu/SensePlace2/} \\[2pt]
\textit{EMERSE}: Enhanced Messaging for the Emergency Response Sector & \cite{caragea2011classifying} \\
~~ Twitter and SMS; machine-translate, classify automatically, alerts & \url{http://emerse.ist.psu.edu/}\\[2pt]
\textit{ESA}: Emergency Situation Awareness & \cite{yin2012using,power2014emergency}\\
~~ Twitter; detect bursts, classify, cluster, geotag & \url{https://esa.csiro.au/} \\[2pt]
\textit{Twitcident} & \cite{abel2012semantics}\\
~~ Twitter and TwitPic; semantic enrichment, classify & \url{http://wis.ewi.tudelft.nl/twitcident/} \\[2pt]
\textit{CrisisTracker} & \cite{rogstadius2013crisistracker}\\
~~ Twitter; cluster, annotate manually & \url{https://github.com/jakobrogstadius/crisistracker}\\[2pt]
\textit{Tweedr} & \cite{ashktorab2014tweedr}\\
~~ Twitter; classify automatically, extract information, geotag & \url{https://github.com/dssg/tweedr} \\[2pt]
\textit{AIDR}: Artificial Intelligence for Disaster Response & \cite{imran2014aidr}\\
~~ Twitter; annotate manually, classify automatically & \url{http://aidr.qcri.org/} \\[2pt]
\bottomrule
\end{tabular}
\end{table}
Most existing systems are built around the concept of a \textit{dashboard}, or a set of visual displays that provides a summary of social media during the crisis according to temporal, spatial, and thematic aspects. Common elements in these displays include:
\begin{itemize}
\item Lists/timelines showing recent or important messages, sometimes grouping the messages into clusters or categories.
\item Time series graphs representing the volume of a hashtag, word, phrase, or concept over time, and sometimes marking peaks of activity.
\item Maps including geotagged messages or interpolated regions, possibly layered according to different topics.
\item Pie charts or other visual summaries of the proportion of different messages.
\end{itemize}
These visual elements are powered by computational capabilities that include:
\begin{itemize}
\item Collections of social media messages matching a given criterion, from one or multiple social media and/or Short Message Service (SMS) streams, typically with a focus on Twitter (described in Section~\ref{sec:data}).
\item Natural Language Processing (NLP), including Named Entity Recognition (NER) and linking of named entities to concepts (described in Section~\ref{sec:data-preprocess}).
\item Extraction of information from the messages, including geotagging (described in Section~\ref{sec:geotag}).
\item Monitoring the volume of messages (or sets of messages) to detect or help detect sub-events within a crisis, and sometimes the crisis itself (described in Section~\ref{sec:event-detection}), possibly including the generation of user-defined alerts when certain conditions are met.
\item Clustering or automatic grouping of similar messages (described in Section~\ref{sec:unsupervised_classification}).
\item Classification of messages or groups of messages manually or automatically (see Section~\ref{sec:supervised_classification}).
\item Automatic translation of messages.
\end{itemize}
Though some of these systems are based on input or feedback provided by emergency responders and other officials, we note that to a large extent they are framed as a way to process social media data during crisis situations; their goal is not to address specific needs of emergency responders or other stakeholders.
This focus on \textit{processing} social media data possibly impacts the adoption of these systems by the practitioner community. Methodologies such as participatory design have been proposed to improve the matching between e.g. the needs of public information officers during a crisis, and the tools built by researchers and developers~\cite{hughes_2014_pio}.
When the goal of meeting the specific need of users \textit{is} stated explicitly in the design of systems, it often revolves around \textit{enhancing situational awareness}, defined in~\cite{endsley1995toward} as ``the perception of elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future.''
For instance, \textit{ESA}~\cite{yin2012using,power2014emergency} aims at enhancing situational awareness with respect to crises induced by natural hazards, particularly earthquakes. This is done by presenting information in time (frequency series) and space (maps), which is achieved by performing event detection, text classification, online clustering and geotagging.
Similarly, \textit{SensePlace2}~\cite{maceachren2011senseplace2} is presented as a \textit{geovisual analytics} system, which filters and extracts geographical, temporal and thematic information from tweets in order to present them in a layered map.
Data from social media can also be presented along with data from physical sensors, for instance to overlay earthquakes detected by seismic sensors on a map presenting social media data~\cite{avvenuti2014ears,musaevlitmus}.
In parallel to approaches that use Natural Language Processing (NLP) techniques to enhance situation awareness, \emph{Crisis Mapping} emerged as an alternative type of system, by employing digital volunteers to collect, classify, and geotag messages~\cite{okolloh2009ushahidi,meier_2015_dh}, and eventually by using the input from those volunteers to train machines to perform these tasks automatically~\cite{imran2014aidr}.
\section{Data Characterization, Acquisition, and Preparation}\label{sec:data}
Both academics and practitioners gather social media data during crisis events. In this section, we describe the common practices used to collect, represent, and process these data.
\iffalse
\subsection{Social Media}\label{sec:social-media}
Social media is a general term that encompasses a variety of platforms on which user-generated content can be disseminated and consumed, and where users can grow online connections with others. This definition currently includes blogging and micro-blogging, online social networking sites, social media sharing platforms, and {\em wikis}~\cite{vandijck2013social}.
The user-generated content posted on the social media platforms include, images, videos, text, and links to other sources of information.
In addition to sharing content, actions include commenting or appraising the content posted by others, and passing this information along to one's own network.
Based on a ``global social media census'' by Business Insider on October 2013, the ten largest online social networks in terms of monthly active users are:
Facebook (a social networking site, 1,200M active users per month),
YouTube (a video hosting site 1,000M),
QZone China (social networking site, 721M),
Sina Weibo (microblogging site, 500M),
WhatsApp (mobile chat application, 350M),
Google Plus (social networking site, 327M),
Tumblr (microblogging site, 300M),
LINE (mobile chat/talk application, 275M),
Twitter (microblogging site, 240M) and
WeChat (mobile chat application, 236M).\footnote{\url{http://www.businessinsider.com/a-global-social-media-census-2013-10}}
Of these, three are mobile-only platforms (WhatsApp, LINE, and WeChat), and all the others can be accessed through mobile applications.
\fi
\subsection{Characteristics of Messages Broadcast on Social Media in Disaster}
Social media is a general term that encompasses a variety of platforms on which user-generated content can be disseminated and consumed, and where users can connect with others. This definition currently includes blogging and micro-blogging, social networking sites, social media sharing platforms, and {\em wikis}~\cite{vandijck2013social}.
Activities such as staying in touch with friends and family, and connecting with others, have driven the growth of social media platforms.\footnote{\url{http://www.pewinternet.org/2011/11/15/why-americans-use-social-media/}}
Currently, different social networking sites are used for different purposes, but commonalities do exist. For instance, the top 3 activities on Twitter are to (1) post about daily activities, (2) upload and share photos, and (3) comment on posts of others; while on Facebook they are to (1) upload and share photos, (2) message with friends on a one-on-one basis, and (3) comment on posts of friends~\cite{globalwebindex2013q2}.
In any of these platforms, an increase in social media communications can be triggered by a variety of causes, which can be divided into \textit{endogenous} and \textit{exogenous}~\cite{crane2008robust}. Endogenous causes refer to phenomena in which an idea or ``meme'' gains popularity by a process of {\em viral contagion} or {\em information cascade}, where content spreads rapidly through a network, potentially reaching a significant fraction of all the users~\cite{chen_2014_propagation}.
Exogenous causes refer to large-scale events, usually happening in the physical world, of wide interest to social media users.
Emergencies and mass convergence events are examples of an exogenous cause, and during such events, we know that communications activity increases.
For instance, it has been observed that mobile network usage---both in terms of phone calls and SMS---increases in emergency situations~\cite{gao_2014_quantifying}.
The same is true for social media usage, which ``rises during disasters as people seek immediate and in-depth information''~\cite{fraustino_2012_start}
\footnote{For instance, activity on Facebook was observed to increase significantly in the areas most affected by the August 2014 earthquake in California, USA \url{https://www.facebook.com/notes/facebook-data-science/on-facebook-when-the-earth-shakes/10152488877538859}}
To illustrate the types of information that affected populations broadcast specifically on the popular microblogging platform Twitter, we turn to some example messages that have been highlighted in previous literature:
\begin{itemize}
\item ``OMG! The fire seems out of control: It's running down the hills!'' (bush fire near Marseilles, France, in 2009, quoted from Twitter in \citeN{longueville2009omg})
\item ``Red River at East Grand Forks is 48.70 feet, +20.7 feet of flood stage, -5.65 feet of 1997 crest. \#flood09'' (automatically-generated tweet during Red River Valley floods in 2009, quoted from Twitter in~\citeN{starbird2010chatter})
\item ``Anyone know of volunteer opportunities for hurricane Sandy? Would like to try and help in any way possible'' (Hurricane Sandy 2013, quoted from Twitter in~\citeN{purohit2013emergency})
\item ``My moms backyard in Hatteras. That dock is usually about 3 feet above water [photo]'' (Hurricane Sandy 2013, quoted from Reddit in~\citeN{leavitt2014upvoting})
\item ``Sirens going off now!! Take cover...be safe!'' (Moore Tornado 2013, quoted from Twitter in~\citeN{blanford2014tweeting})
\end{itemize}
Though the above are only a few examples, they convey a sense of the types of information posted during an event, and show that it is varied.
\citeN{vieweg_2012_thesis} points to this variation in her research that is based on a detailed study of four crisis events, in which she identifies in Twitter 35 types of information in three broad categories defined by~\citeN{mileti_1999_disasters}: social environment, built environment, and physical environment. She points out that social environment messages describe anything having to do with people and their reactions to the crisis, built environment messages correspond to information and updates about property and infrastructure, and physical environment messages include updates about the hazard agent, weather, and other environmental factors (see Section~\ref{subsec:content-categories} for details on different ways of categorizing this information).
Quantifying the amount of information found in social media based on type is even more difficult than locating that information in the first place. Important variations have been observed across crises (even for similar events) and at different stages of a crisis~\cite{blanford2014tweeting,imran2013practical}.
\citeN{olteanu2014crisislex} looked at the prevalence of three broad categories of information in tweets related to six crisis events. The results show large variabilities in the number of tweets reporting the negative consequences of an event (20\%-60\%), those offering or asking for donations (15\%-70\%) and those warning about risks or providing advice (5\%-20\%).
Using tweets as an indication of spatial zones during a disaster is also possible. For instance, Acar and Muraki~\citeyear{acar2011twitter} examine the use of Twitter during an earthquake in Japan and observed that tweets from affected areas include more requests for help and more warnings, while tweets from other areas which are far from the disaster epicenter tend to mostly include other types of information, such as concern and condolences.
\subsection{Data Acquisition}\label{sec:data-acquisition}
Most large social media platforms provide programmatic access to their content through an Application Programming Interface (API). However, the details of these APIs vary substantially from one platform to another, and also change over time.
APIs to access social media data typically belong to one of two types: those allowing to query an archive of past messages (also known as {\em search} APIs), and those allowing data collectors to subscribe to a real-time data feed (also known as {\em streaming} or {\em filtering} APIs).
Both types of APIs typically allow data collectors to express an information need, including one or several of the following constraints: (i) a time period, (ii) a geographical region for messages that have GPS coordinates (which are currently the minority), (iii) a set of keywords that must be present in the messages, which requires the use of a query language whose expressiveness varies across platforms.
In the case of archive/search APIs, messages are returned sorted by relevance (a combination of several factors, including recency), or just by recency. In the case of real-time/streaming/filtering APIs, messages are returned in order of their posting time.
Data collection strategies impact the data obtained and analytic results. For instance, selecting messages in the geographical region affected by a disaster vs. selecting messages based on a keyword-based query may return datasets having different characteristics~\cite{olteanu2014crisislex}.
\spara{Data availability.} \urldef{\weiboapi}\url{http://open.weibo.com/wiki/AP
In addition to issues regarding network connectivity during disasters~\cite{jennex2012social}, access to social media data is in general limited, which is a serious obstacle to research and development in this space~\cite{reuter2014technical}.
First, historical, archived social media can typically only be queried through {\em search APIs}, which are limited in terms of number of queries or data requested per unit of time. The exceptions are (i) a 1\% data sample collected from Twitter by the Internet Archive,\footnote{Available at \url{https://archive.org/details/twitterstream}} and (ii) collections created by researchers using automatic data collection methods (e.g. web crawling), which are against the terms of services of most social media platforms.
Second, access to real-time data is quite limited. Twitter offers a public streaming API providing a random sample of 1\% of all postings, plus the possibility of filtering all public postings by keywords.\footnote{Available at \url{https://dev.twitter.com/docs/api/streaming}}
This is in contrast with most large social media platforms, which do not offer this level of data access publicly.\footnote{As of November 2014, we found no publicly-available equivalent of Twitter's {\em streaming API} in Sina Weibo, Facebook, YouTube, Google Plus, or Tumblr.}
The consequence of these limitations is that most work on crisis-related messages is done using Twitter data (with few exceptions e.g.~Facebook~\cite{bird2012flooding} and Reddit~\cite{leavitt2014upvoting}), which provides an incomplete view of a crisis situation, as there are many different online social media sites that might be used for different purposes. Limitations in the amount of data that can be collected and in general a dependency on a small set of data providers or APIs, further reduces the efficiency and effectiveness of tools for handling crisis-related social media messages.
\iffalse
Sina Weibo allows users to programmatically poll for the latest 20 public postings, but not to obtain them in a streaming fashion,\footnote{As of April 2014, \weiboapi}
Facebook provides a stream of recent public postings through an API, but only to selected partners,\footnote{As of April 2014, \url{https://developers.facebook.com/docs/public_feed/}}
YouTube provides an API call for just a ``most popular by country'' stream,\footnote{As of April 2014, \url{https://developers.google.com/youtube/2.0/developers_guide_protocol_video_feeds#Standard_feeds}}
while Google Plus and Tumblr do not provide this functionality at all, allowing only to query the recent posts of a specific user (in the case of Google Plus)\footnote{As of April 2014, \url{https://developers.google.com/+/api/latest/}} or those having a specific tag (in the case of Tumblr).\footnote{As of April 2014, \url{http://www.tumblr.com/docs/en/api/v2}}
\fi
\subsection{Data Pre-Processing}\label{sec:data-preprocess}
Most researchers and practitioners prepare social media data by {\em pre-processing} it, using some of the methods outlined below, before performing the actual analysis.
Many pre-processing techniques are available, and the choice depends on the type of data at hand and the goals of the analysis.
\spara{Natural Language Processing (NLP).}
The text of the messages can be pre-processed by using an NLP toolkit. Typical operations include tokenization, part-of-speech tagging (POS), semantic role labeling, dependency parsing, named entity recognition and entity linking. A number of off-the-shelf implementations of these operations are available online, e.g. the Stanford NLP\footnote{\url{http://www-nlp.stanford.edu/software/}} or NLTK for Python.\footnote{\url{http://www.nltk.org/}}
Social media-specific NLP toolkits can also be used. For instance, ArkNLP~\cite{owoputi2013improved}, which is trained on Twitter data, is able to recognize Internet idioms such as ``ikr'' (\emph{I know, right?}) and assign them the correct Part Of Speech (POS) tag (interjection, in this case).
Additionally, higher-level operations can be applied, including applying sentiment analysis methods to infer aspects of the emotion conveyed by a piece of text~\cite{pang2008opinion}.
\spara{Feature extraction.}\label{subsec:feature-extraction}
For many automatic information processing algorithms (e.g. machine learning), each data item must be represented accurately as an information record. The representation of choice for text is typically a numerical vector in which each position corresponds to a word or phrase---this is known as the vector space model in information retrieval. The value in each position can be a binary variable, indicating the presence or absence of that word or phrase in the message, or a number following some weighting scheme, such as TF-IDF, which favors terms that are infrequent in the collection~\cite{mir2ed}.
To avoid having too many variables, textual features can be discarded by removing stopwords and functional words, or by normalizing words using stemming or lemmatization (e.g. considering ``damaging'' and ``damage'' as equivalent), or other means.
Additionally, other text-based features can be added, such as the length of the text in words or characters, and the number of question or exclamation marks. If some NLP pre-processing is performed on the text---such as part-of-speech-tagging---a feature such as \emph{noun:fire} (instead of \emph{verb:fire}) can be used to distinguish that the word ``fire'' is being used in a message as a noun (``I heard a fire alarm'') instead of a verb (``They should fire him'').
In the case of tweets, characteristics such as the presence of user mentions (``@user''), URLs, or hashtags can be included as features.
In the case of images or video, content-based features such as colors, textures and shapes can be included~(see e.g. \citeN{liu2007survey} for a survey). Additionally, features such as the date of a message, tags associated with it, the number of views/comments it has received, or information about its author, are often available in a platform-dependent manner.
Obviously, one can spend a great deal of time constructing features by hand. In order to guide this exploration, both researchers and practitioners should prioritize the development and understanding of features likely to be correlated with the target variable (e.g. tweet classification). By the same token, features valuable to one target variable may not be important at all for a different target variable. Although often under-appreciated, feature engineering is perhaps the most important part of a modeling exercise.
\spara{De-duplication.}
Further reduction of the amount of data to be processed can be achieved by removing near-duplicate messages. Given that in many social media platforms the number of people re-posting a message can be interpreted as a measure of its importance, whenever removing near-duplicates it is advisable to save the number of near-duplicates that have been found (e.g. as done in~\citeN{rogstadius2013crisistracker} to prioritize highly-reposted stories). De-duplication can be done by applying a clustering method (see Section~\ref{sec:unsupervised_classification}).
\spara{Filtering.}
A fraction of the messages collected will not be relevant for a given crisis. This fraction depends on the specific collection method used, as discussed by~\citeN{olteanu2014crisislex}, and on other factors, such as the presence of off-topic messages using the same tags or keywords as the on-topic ones~\cite{qu2011microblogging}. These messages can be post-filtered using human labeling or crowdsourcing, keyword-based heuristics, or automatic classification.
Additionally, many messages are posted automatically on social media for financial gain, exploiting the attention that a certain hashtag has received. These unsolicited commercial messages are known as {\em spam}~\cite{gupta2012credibility,uddinunderstanding} and there are well-studied methods that can remove a substantial portion of them~\cite{benevenuto2010detecting}.
Finally, in some cases we might want to also remove messages posted by automatic agents or social media {\em bots}. Their identification is similar to that of spammers.
\subsection{Geotagging and geocoding}\label{sec:geotag}
Attaching geographical coordinates to a message (a process known as {\em geotagging}) is useful for a number of tasks in disaster response~\cite{graham:geolocation,ikawa_2013_location,lingad:www2013}.
Geotagging allows the retrieval of information about a {\em local event}, by filtering the messages corresponding to a particular geographical region.
It allows the visualization of information about an event on a map, possibly making it more actionable for emergency responders.
Geotagging can also be used for higher-level tasks, such as helping predict epidemic transmission of diseases based on geographical proximity~\cite{brennan2013towards}.
The availability of machine-readable location information in social media messages, in the form of metadata, depends on the user's device having the capacity to know its location (e.g. via Global Positioning System (GPS)), on the specific client software having the capability to read this from the device, and most importantly, on the user enabling this feature explicitly (opting-in). In practice, only 2\% of crisis-related messages include machine-readable location information \cite{burton:twitter-geolocation}.
However, while explicit metadata about locations may be absent, many messages in social media do contain implicit references to names of places (e.g. ``The Christchurch hospital is operational'').
{\em Geocoding} refers to finding these geographical references in the text, and linking them to geographical coordinates.
This can be done by using a named entity extractor to extract potential candidates, and then comparing those candidates with a list of place names. This is the approach used by e.g.~\citeN{maceachren2011senseplace2} which uses Gate\footnote{\url{https://gate.ac.uk/}} for the first task and Geonames\footnote{\url{http://geonames.org/}} for the second.
While building a comprehensive database of geospatial information, including place names, is an important component of geotagging (see e.g.~\citeN{middleton2014real}), geotagging is not merely a dictionary look-up process because of ambiguities. These ambiguities are known as ``{\em geo/non-geo}'' and ``{\em geo/geo}.''
A geo/non-geo ambiguity occurs for instance in the message ``Let's play Texas Hold 'em,'' that does not refer to the state of Texas in the USA. A geo/geo ambiguity is found in the message ``There is a fire in Paris,'' which may refer to the capital of France, or to any of more than a dozen places on Earth sharing the same name. For instance, in~\citeN{sultanik2012rapid}, authors proposed an unsupervised approach to extract and disambiguate location mentions in Twitter messages during crisis situations.
In general, geotagging is done through probabilistic methods (see e.g.~\citeN{cheng_2010_where}), often exploiting contextual clues. These clues may include the general location of a crisis, information about nearby places, and location information indicated by users in their profiles~\cite{gelernter2011geo}.
\subsection{Archived versus Live Data Processing}\label{sec:retro-vs-online}
Depending on the urgency with which the output of an analysis is required, data may be provided to an algorithm either as an archive, for \textit{retrospective} analysis, or as a live data feed, for \textit{real-time} analysis. These correspond to two standard concepts in computer science: off-line processing and on-line processing.
Retrospective data analysis (\textit{off-line processing}) starts with a batch of data relevant to an event, usually containing messages over the entire time range of interest. For example, we might consider re-constructing a timeline of events in the aftermath of an earthquake, by looking at all of the tweets from the moment of the earthquake up to two weeks after. In deciding how to build the timeline, we have the complete context of events during this two week window.
Live data analysis (\textit{on-line processing}) is done over a stream of data relevant to the event, usually provided in real-time or with a short delay. For example, we might consider constructing a timeline of the events in the aftermath of an earthquake \textit{as we observe new tweets}; in deciding how to build the timeline, we have an incomplete context of the events and their future repercussions.
It is possible for algorithms to lie in between, operating on small batches of data at regular intervals (e.g. hourly, daily).
The trade-off between retrospective and live data is a matter of accuracy versus latency.
Retrospective data analysis maximizes our context and, as a result, gives us an accurate picture of the data. However, because we have to wait for the data to accumulate, we incur latency between when an event happens and when it is processed.
Live data analysis, on the other hand, minimizes the latency but, because we have partial information, we may incur lower accuracy.
The choice of collection methodology depends on the use case. Crisis responders may want lower latency in order to better respond to a developing situation; forensic analysts may want higher accuracy and have the benefit of waiting for data to be collected.
While developing an algorithm, we can use retrospective data to \textit{simulate} live data. This is a standard experimental methodology that has been used in the past for information filtering tasks~\cite{voorhees:trecbook} and, more recently, for crisis informatics~\cite{qi:temporal-summarization,trec2013:temporal-summarization-overview}.
\subsection{Challenges} \label{sec:data-challenges}
There are a number of challenges associated with the processing of social media messages. In this section, we group them into two high-level categories: scalability and content. We defer specific challenges (e.g. challenges to event detection) to their respective technical sections.
\spara{Scalability issues.}
Large crises often generate an explosion of social media activity.
Data size may be an issue, as for crises that last several days, millions of messages may be recorded. While the text of each message can be sorted, a data record for e.g. a Twitter message (140 characters of text) is around 4KB when we consider the metadata attached to each message. Thus, a Twitter collection for a crisis is then on the order of several hundred megabytes to a few gigabytes. In addition, multimedia objects such as images and videos may significantly increase the storage space requirements.
Data velocity may be a more challenging issue, especially considering that data does not flow at a constant rate but experiences drastic variations. The largest documented peak of tweets per minute during a natural hazard that we are aware of is 16,000 tweets per minute.\footnote{During Hurricane Sandy in 2012: http://www.cbsnews.com/news/social-media-a-news-source-and-tool-during-superstorm-sandy/}
Finally, redundancy, which is commonly cited as a scalability challenge, to some extent cannot be avoided in this setting. Repeated (re-posted/re-tweeted) messages are common in time-sensitive social media, even encouraged, as in some platforms messages that gain more notoriety are those that are simply repeated more.
\spara{Content issues.}\label{subsec:content-issues}
Microblog messages are brief and informal. In addition, this type of messaging is often seen by users to be more akin to speech, as opposed to a form of writing, which---compounded with technological, cross-lingual and cross-cultural factors---implies that on the Internet ``we find language that is fragmentary, laden with typographical errors, often bereft of punctuation, and sometimes downright incoherent''~\cite{baron2003language}.
This poses significant challenges to computational methods, and can lead to poor and misleading results by what is known as the ``garbage in, garbage out'' principle.
Messages are also highly heterogeneous, with multiple sources (e.g. traditional media sources, eyewitness accounts, etc.), and varying levels of quality. Quality itself is an important and complex question for crisis managers, and encompasses a number of attributes including objectivity, clarity, timeliness and conciseness, among others~\cite{friberg2011information}.
Additionally, different languages can be present in the same crisis and sometimes in the same message---a phenomena known as ``borrowing,'' and ``code switching.'' This makes it difficult for both machines and humans (e.g. content annotators) to understand or classify messages.
Finally, brief messages sent during a crisis often assume a shared context from which only a minor part is sometimes made explicit. The area of study in linguistics known as {\em pragmatics} focuses on ``communication in context,'' and explains how people are able to infer the meaning of the communications because humans are very adept at understanding context. So, in the case of Twitter communications, a reader can understand the tweet author's intent because she or he knows the context within which that tweet is being broadcast. Current computational methods are not able to make the same inferences humans do, and thus cannot achieve the same level of understanding~\cite{vieweg2014pragmatics}.
\spara{Privacy issues.}\label{subsec:privacy-issues}
Social media, since it is, by definition, content directly created by end users, might carry personally identifiable information (PII). Researchers and practitioners should be mindful of any explicit or inferred PII in the data. For example, an individual may not explicitly reveal PII such as location for privacy reasons; inferring this user's location may be in conflict with her expectation of privacy. As a result, academic work may have to be approved by a human subjects review process (ethics approval).
Practitioners, on the other hand, may be subject to a response organization's own standards and policies (see e.g.~\citeN{gilman:ocha-note} and \citeN{icrc:professional-standards}).
\section{Event Detection and Tracking}\label{sec:event-detection}
Most systems for social media processing during crises start with \textit{event detection}.
An event is the occurrence of something significant which is associated with a specific time and location~\cite{brants2003system}. However, due to the online nature of social media communications, events as they play out in social media may or may not be necessarily associated with a physical location.
In the context of social media, Dou et al.~\citeyear{dou2012leadline} define an event as: ``An occurrence causing changes in the volume of text data that discusses the associated topic at a specific time. This occurrence is characterized by topic and time, and often associated with entities such as people and location''.
Crisis and emergency situations typically fall into two broad categories: \emph{predicted} (or \emph{forewarned}) and \emph{unexpected}.
Some disaster events can be predicted to a certain level of accuracy based on meteorological or other data (e.g. this is the case with most large storms and tornadoes), and information about them is usually broadcast publicly before the actual event happens. Sometimes an event may not be explicitly anticipated, for example as with a mass protest, but still may be forecast from social media and other data \cite{ramakrishnan:EMBERS}.
Other events cannot be predicted (e.g. earthquakes), and in this case an automatic detection method is useful to find out about them as quickly as possible once they happen.
In this section, we study techniques available for the automatic detection of both predicted, and unexpected events.
Historically, methods for event detection and tracking in social media in the context of crises and emergencies, are adaptations of methods to perform these tasks with a more general scope: that of detecting news. These methods, in turn, are adaptations of methods to find new topics in general document collections.
\subsection{Background on Event Detection and Discovery}
A well-studied problem in Information Retrieval is detecting events in a stream of documents (see e.g. \citeN{allan2002topic}).
These documents can be news articles from traditional media sources, or posts on social media (e.g., tweets, Facebook posts, Flickr images).
Traditionally, the {\em Topic Detection and Tracking} (TDT) research community uses newswires as source data streams for event detection.
Various techniques are employed in TDT including story segmentation, topic detection, new event detection, link detection, and topic tracking.
\emph{Story segmentation} focuses on determining story boundaries from streaming speech recognition output, usually from radio or television broadcasts.
\emph{Topic detection} groups related documents together into cohesive topics.
\emph{New event detection} processes each new document to decide if it describes a previously unseen story.
\emph{Link detection} detects that if two documents are similar or not.
Finally, \emph{event tracking} follows the evolution of an event/topic to describe how it unfolds.
Event detection on social media is different from the traditional event detection approaches that are suitable for other document streams. Social media data emerge more quickly, and in larger volumes than traditional document streams. In addition, social media data are composed of short, noisy, and unstructured content that often require a different approach than what is used with traditional news articles.
Considering the unique characteristics of social media streams, we focus the remainder of this section on new event detection and event tracking.
Nevertheless, techniques and evaluation metrics from the TDT community provide insight into methods that might work for the Twitter domain.
\subsection{New Event Detection (NED)}
In the context of mass emergencies, New Event Detection (NED) refers to the task of discovering the first message related to an event by continuously monitoring a stream of messages.
NED decides whether a message is about something new that has not been reported in previous messages, or not~\cite{yang2009discovering}.
``New'' is normally operationalized as sufficiently different according to a similarity metric. Hellinger similarity, Kullback-Leibler divergence, and cosine similarity are among the metrics commonly used in NED~\cite{kumaran2004classification}.
As observed in Section~\ref{sec:retro-vs-online}, we can do retrospective (off-line) or on-line analysis~\cite{yang1998study}. While the most useful NED systems for emergencies are those who perform this analysis on-line, they are often adaptations or improvements of off-line methods, which we discuss next.
\subsubsection{Retrospective New Event Detection}
Retrospective NED refers to the process of identifying events using messages or documents that have arrived in the past. Methods for retrospective NED involve the creation of clusters of documents or messages based on a suitable definition of similarity between them, which may involve more than one dimension of them, e.g. using similar words, involving similar groups of people, occurring close to each other in time and/or space, etc.
For instance,
\citeN{zhao2007temporal} introduce a retrospective NED method that uses textual, social and temporal characteristics of the documents to detect events on social streams such as weblogs, message boards, and mailing lists.
They build multi-graphs using social textual streams, where nodes represent social actors, and edges represent the flow of information between actors. Clustering techniques and graph analysis are combined to detect an event.
\citeN{sayyadi2009event} introduce a retrospective NED approach that overlays a graph over the documents, based on word co-occurrences. They assume that keywords co-occur between documents when there is some topical relationship between them. Next, a community detection method over the graph is used to detect and describe events.
\citeN{pohl2012automatic} describe a two-phase clustering approach to identify crisis-related sub-events in photo-hosting site Flickr and video-hosting site YouTube.
During the first phase, which is to identify sub-events, clusters are formed by using only items that contain geographical coordinates. These coordinates are added automatically by the device used to capture the photo or video, or are added later by its author/uploader. Next, they calculate term-based centroids of the identified clusters using cosine distance to further describe the identified sub-events.
Another retrospective NED approach is presented in~\citeN{chen2009event}, with experiments run on Flickr. It uses photos, user-defined tags, and other meta-data including time and location to detect events. This approach simultaneously analyzes the temporal and geographical distribution of tags, and determines the event type (e.g. whether it is recurring) to form clusters. Finally, for each tag cluster, the corresponding photos are retrieved.
\citeN{ritter2012open} extract significant events from Twitter by focusing on certain types of words and phrases. In their system, called {\em TwiCal}, they extract event phrases, named entities, and calendar dates. To extract named entities, they use a named entity tagger trained on 800 randomly selected tweets. To extract event mentions they use a Twitter-tuned part-of-speech tagger~\cite{ritter2011named}. The extracted events are classified retrospectively into event types using a latent variable model that first identifies event types using the given data, and then performs classification.
Li et al.~\citeyear{li2012twevent} introduce \textit{Twevent}, a system that uses message segments instead of individual words to detect events. The authors claim that a tweet segment, which represents one or more consecutive words in tweets, contains more meaningful information than unigrams. The \textit{Twevent} approach works in phases. First, the individual tweet is segmented, and bursty segments are identified using the segments' frequency in a particular time window. Next, identified segments are retrospectively clustered using K-Nearest Neighbors (KNN) clustering. Finally, a post-filtering step uses Wikipedia concepts to filter the detected events.
\subsubsection{Online New Event Detection}
Online new event detection does not use previously seen messages or any prior knowledge about the events to be identified.
Online NED is typically performed with low latency (in real-time), in the sense that the time between seeing a document corresponding to a new event, and reporting that a new event has been detected, is relatively short.
\spara{Methods based on keyword burst.}
A straightforward approach is to assume that words that show sharp frequency increases over time are related to a new event.
For instance, Robinson et al.~\citeyear{robinson2013sensitive} introduce a system to detect earthquakes using Twitter. The earthquake detector, which is based on the Emergency Situation Awareness (\textit{ESA}) platform~\cite{power2014emergency}, checks for the keywords ``\texttt{earthquake}'' and ``\texttt{\#eqnz}'' in the real-time Twitter stream, and applies a burst detection method to analyze word frequencies in fixed-width time-windows and compare them to historical word frequencies. Unusual events are identified if the observed frequencies are much higher than those recorded in the past.
Earle et al.~\citeyear{earle2012twitter} have compared simple keyword-based approaches with data from seismological sensors, finding that while many earthquakes are not detected by Twitter users, detections are fast, ``considerably faster than seismographic detections in poorly instrumented regions of the world.''
Marcus et al.~\citeyear{marcus2011twitinfo} introduced \textit{TwitInfo}, a system for detecting, summarizing and visualizing events on Twitter. \textit{TwitInfo} collects tweets based on a user-defined query (e.g. keywords used to filter the Twitter stream).
It then detects events by identifying sharp increases in the frequency of tweets that contain the particular user-defined query as compared to the historical weighted running average of tweets that contain that same query.
Further, tweets are obtained from the identified events to identify and represent an aggregated sentiment (i.e., classifying tweets into positive and negative classes).
The authors evaluated the system on various events such as earthquakes and popular football games
A similar system, {\em TwitterMonitor}~\cite{mathioudakis2010twittermonitor}, also collects tweets from the Twitter stream and detects trends (e.g. emerging topics such as breaking news, or crises) in real-time. The trend detection approach proposed in their paper works in two phases. During the first phase, {\em TwitterMonitor} identifies bursty keywords which are then grouped based on their co-occurrences. Once a trend is identified, additional information from the tweets is extracted to analyze and describe the trend. For example, the system uses Grapevine's entity extractor~\cite{angel2009s} which identifies entities mentioned in the trends.
Another Twitter-specific event detection approach introduced by~\citeN{petrovic2010streaming} uses Locality Sensitive Hashing (LSH) for hashing a fixed number of recent documents in a bounded space, and processed in a bounded time, to increase the performance of nearest neighbors search.
With so many event detection systems, it is interesting to think about how they compare. McMinn et al.~\citeyear{McMinn:2013} describe a corpus to evaluate event detection methods, composed of 500 news events sampled over a four-week period, and including tweet-level relevance judgments for thousands of tweets referring to these events. While existing NED systems have not been evaluated against this corpus, we anticipate this calibration of systems to occur in the future.
\spara{Beyond keyword bursts.}
There are well-known problems of relying on increases in the frequency of a keyword (or a segment) to detect events.
For instance, consider popular hashtags such as ``\#musicmonday," which is used to suggest music on Mondays, or ``\#followfriday/\#ff,'' which are used to suggest people to follow on Fridays.
In these cases, there should be big pseudo-events detected every Monday and every Friday.
To address this problem, \citeN{becker2011beyond} present an approach to classify real-world events from non-events using Twitter. They use four types of features, which are temporal, social, topical, and Twitter-specific, to identify real events using the Twitter stream in real-time.
First, based on temporal features (i.e., volume of messages posted during an hour), they form initial clusters using the most frequent terms in the messages. Clusters are then refined using social features (i.e., users' interactions like re-tweets, replies, mentions). Next, they apply heuristics, for example, a high percentage of re-tweets and replies often indicates a non-event, whereas a high percentage of mentions indicates that there is an event.
Further, cluster coherence is estimated using a cosine similarity metric between messages and cluster centroid.
Finally, as the authors report that multi-word hashtags (e.g. ``\#musicmonday'' and ``\#followfriday'') are highly indicative of some sort of Twitter specific discussion and do not represent any real event, they check the frequency of such hashtags used in each cluster to further refine the results.
Weng et al.~\citeyear{weng2011event} present an algorithm for event detection from tweets using clustering of wavelet-based signals. Their approach involves three steps. First, they use wavelet transformation and auto correlation to find bursts in individual words, and keep only the words with high signal auto-correlations as event features. Then, the similarity for each pair of event-features is measured using cross correlation. Finally, they use a modularity-based graph partitioning algorithm to detect real-life events. One of the strong points of this approach over the traditional event detection approaches is the capability of differentiating real-life big events from trivial ones. This is achieved mainly by two factors: the number of words, and the cross-correlation among the words related to an event.
Unlike the approaches presented above, Corley et al.~\citeyear{corley2013social} present a method to detect and investigate events through meta-data analytics and topic clustering on Twitter.
Various features such as re-tweets, usage of different terms, and hashtags are analyzed for a certain time period to determine a baseline and a noise ratio. An event is detected once a particular feature value exceeds its noise boundaries and expected threshold.
Once an event has been detected, its related topics are identified using the topic clustering approach.
\spara{Domain-specific approaches.}
As in many natural language processing applications, approaches that are specific to a certain domain generally perform better than the approaches that are open-domain or generic.
For instance, Phuvipadawat and Murata~\citeyear{phuvipadawat2010breaking} describe a method for detecting breaking news from Twitter. First, tweets containing the hashtag ``\#breakingnews'' or the phrase ``breaking news'' are fetched from the Twitter streaming API. Grouping of the extracted tweets is then performed, based on content similarity, and using a variant of the TF-IDF technique. Specifically, the similarity variant assigns a high similarity score to hashtags and proper nouns, which they identify using the Stanford Named Entity Recognizer (NER) implementation.
The authors consider three types of features associated with tweets: statistical features (i.e., number of words in a tweet, position of the query word within a tweet), keyword-based features (i.e., the actual words in a tweet), and contextual features (e.g., words appearing nearby the query term, for instance if ``earthquake'' is a query, terms such as ``magnitude'' and ``rocks'' would be features in the tweet ``{\em 5.3 magnitude earthquake rocks parts of Papua New Guinea}'').
In order to determine if a tweet corresponds to one of these hazards or crises, they use Support Vector Machines (SVM)---a known supervised classification algorithm (more about supervised classification in Section~\ref{sec:supervised_classification}).
Data from traditional media sources can also be used for detecting newsworthy events. Not surprisingly, traditional media and social media have different editorial styles, perceived levels of credibility, and response delays to events.
\citeN{tanev2012enhancing} find news articles describing security-related events (such as gun fights), and use keywords in their title and first paragraph to create a query. This query is issued against Twitter to obtain tweets related to the event.
\citeN{dou2012leadline} describe {\em LeadLine}, an interactive visual analysis system for event identification and exploration.
{\em LeadLine} automatically identifies meaningful events in social media and news data using burst detection. Further, named entities and geo-locations are extracted from the filtered data to visualize them on a map through their interface.
Another domain-specific event-detection method is based on pre-specified rules and introduced in~\citeN{li2012tedas}. Their system, {\em TEDAS}, detects, analyzes, and identifies relevant crime and disaster related events on Twitter. First, tweets are collected based on iteratively refined rules (e.g., keywords, hashtags) from Twitter's streaming API. Next, tweets are classified via supervised learning based on content as well as Twitter-specific features (i.e., URLs, hashtags, mentions). Additionally, location information is extracted using both explicit geographical coordinates and implicit geographical references in the content. Finally, tweets are ranked according to their estimated level of importance.
\citeN{sakaki2010earthquake} detect hazards and crises such as earthquakes, typhoons, and large traffic jams using temporal and spatial information. \emph{LITMUS}~\cite{musaevlitmus} detects landslides using data collected from multiple sources. The system, which depends on the USGS seismic activity feed provider, the TRMM (NASA) rainfall feed, and social sensors (e.g. Twitter, YouTube, Instagram), detects landslides in real-time by integrating multi-sourced data using relevance ranking strategy (Bayesian model). Social media data is processed in a series of filtering steps (keyword-based filtering, removing stop-words, geotagging, supervised classification) and mapped based on geo-information either obtained from meta-data or from content. Events with high relevancy are identified as ``real events.''
\subsection{Event Tracking and Sub-Event Detection}\label{subsec:event-tracking-subevent-detection}
\spara{Event tracking.}
Event tracking refers to the task of studying how events evolve and unfold. For a general discussion on the subject, see~\citeN{allan2002topic} and \citeN{lee2013event}.
The way in which emergency response agencies deal with crisis events varies as a crisis unfolds.
Emergency situations typically consist of four phases: warning, impact, emergency, and recovery~\cite[page 51]{killian_2002_introduction}. During the warning phase, the focus is on monitoring the situation. Impact is when the disaster agent is actually at work, while the emergency phase is the immediate post-impact period during which rescue and other emergency activities take place. Recovery is the period in which longer-term activities such as reconstruction and returning to a ``normal'' state occur.
Various techniques have been proposed to identify event phases. For instance, Iyengar et al.~\citeyear{iyengar2011content} introduce an approach to automatically determine different phases of an event on Twitter. The approach, which is mainly based on content-based features of tweets, uses an SVM classifier and a hidden Markov model. Various content-specific features such as bag of words, POS (part-of-speech) tags, etc. are used to automatically classify tweets into three phases of an event: {\em before}, {\em during}, and {\em after.} A disaster-specific lexicon of discriminative words for each phase of the event can also be employed~\cite{chowdhurytweet4act}.
\spara{Sub-event detection.}
The detection of large-scale events has been studied with more attention than the detection of small-scale ``sub-events'' that happen as a crisis unfolds.
\citeN{pohl2012automatic} show the importance of sub-event detection during crisis situations. They use multimedia meta-data (tags and title) associated with content found on social media platforms such as YouTube and Flickr. Their framework uses a clustering approach based on self-organizing maps to detect sub-events.
First, a pre-selection of the data is performed based on user-identified keywords. The selected data is then passed to a sub-event detection module that performs clustering to further split the data into sub-events.
\citeN{khurdiya2012extraction} present a system for event and sub-event detection using Conditional Random Fields (CRF)~\cite{lafferty2001conditional}. Their system consists of four main modules: (1) a CRF-based event extractor to first extract actor, action, date, and location; event titles are also extracted using using CRFs; (2) an event resolution to find similar events; (3) an event compiler that characterizes events; and (4) an event reporting module which is the end-user interface used to browse events details.
\citeN{hua2013sted} introduce {\em STED}, a semi-supervised targeted-interest event and sub-event detection system for Twitter. To minimize the human effort required for labeling, they introduce an automatic label creation and expansion technique, which takes labels obtained from newspaper data and transfers them to tweets. They also propagate labels using mentions, hashtags, and re-tweets. Next, they build mini-clusters using a graph partitioning method to group words related to the event, and use supervised learning to classify other tweets using the examples provided by each mini-cluster. A final step on the classified output is to perform a location estimation using information from geo-coded tweets.
\begin{table}
\caption{Some of the event detection tools surveyed. The table includes the types of events for which the tool is built (open domain or specific), whether detection is performed in real-time, the type of query (open or ``kw''=keyword-based), and whether it has spatio/temporal or sub-event detection capabilities. Sorted by publication year.}
\label{tbl:event_detection_tools}
\centering\scriptsize
\begin{tabular}{ p{1.4cm} p{2.5cm} p{1.7cm} p{0.2cm} p{0.3cm} p{0.7cm} p{0.3cm} p{3.8cm} }
\toprule
System/tool & Approach & Event types & Real-time & Query type & Spatio-temporal & Sub-events & Reference \\
\midrule
\textit{TwitterMonitor} & burst \mbox{detection} & open domain & yes & open & no & no & \cite{mathioudakis2010twittermonitor} \\
\textit{TwitInfo} & burst \mbox{detection} & earthquakes+ & yes & kw & spatial & yes & \cite{marcus2011twitinfo} \\
\textit{Twevent} & \mbox{burst segment detection} & open domain & yes & open & no & no & \cite{li2012twevent} \\
\textit{TEDAS} & \mbox{supervised classification} & crime/disasters & no & kw & yes & no & \cite{li2012tedas} \\
\textit{LeadLine} & \mbox{burst detection} & open domain & no & kw & yes & no & \cite{dou2012leadline} \\
\textit{TwiCal} & \mbox{supervised classification} & \mbox{conflicts/politics} & no & open & temporal & no & \cite{ritter2012open} \\
\textit{Tweet4act} & \mbox{dictionaries} & disasters & yes & kw & no & no & \cite{chowdhurytweet4act} \\
\textit{ESA} & burst \mbox{detection} & open domain & yes & kw & spatial & no & \mbox{\cite{robinson2013sensitive}} \\
\bottomrule
\end{tabular}
\end{table}
Table~\ref{tbl:event_detection_tools} lists event detection systems. The majority of them are surveyed above; additional tools are covered in the following sections.
\subsection{Challenges}
In addition to the discussion of general data challenges in Section~\ref{sec:data-challenges}, the following are particularly relevant to event detection.
\spara{Inadequate spatial information.}
Spatial and temporal information are two integral components of an event. Most systems that rely on Twitter data for event detection face challenges to determine geographical information of tweets that lack GPS information. In this case, automatic text-based geotagging can be used, as described on Section~\ref{sec:geotag}.
\spara{Mundane events.}
People post mundane events on social media sites. These data points introduce noise, which creates further challenges for an event detection algorithm to overcome. For such cases, separation of real-life big events from trivial ones is required.
\spara{Describing the events.}
Creating a description or label for a detected event is in general a difficult task. Often the keywords that are more frequent during the event are presented as a description for the event in the form of a list of words (e.g. \{ \emph{sandy}, \emph{hurricane}, \emph{new york} \} ), but that list does not constitute a grammatically well-formed description (e.g. ``\emph{Hurricane Sandy hits New York}''). We will see one approach to address this in Section \ref{sec:summarization}.
\section{Clustering, Classification, Extraction and Summarization}
\label{sec:finding_information}
Once we have found social messages related to a crisis event or topic, there are several ways in which we might process them. In this section, we describe some approaches. Broadly, we separate techniques into those classifying the data item as a whole and those extracting useful information from the content of one or more data items.
\subsection{Classifying Social Media Items}
In many situations, we are interested in classifying social media items into one or more categories. In this section, we will describe three methods of classification with decreasing levels of manual human supervision. Text classification is a large field of research, so we will cover literature relevant to disasters. Interested readers should refer to other sources for a broader discussion \cite{Srivastava:2009:TMC:1571651}.
\subsubsection{Content Categories}\label{subsec:content-categories}
There is not a single standard or widely-accepted way of categorizing crisis-related social media messages. While some crisis-related ontologies have been proposed (see Section~\ref{sec:ontologies}), in general, different works use different approaches.
Table~\ref{tbl:event_dimensions_summary} summarizes various dimensions of social media content that different research articles have used to classify information:
\begin{compactenum}
\item By factual, subjective, or emotional content: to separate between facts (or combinations of facts and opinions), from opinions, or expressions of sympathy.
\item By information provided: to extract particular categories of information that are useful for various purposes.
\item By information source: to select messages coming from particular groups of users, e.g. eyewitness accounts or official government sources.
\item By credibility: to filter out messages that are unlikely to be considered credible.
\item By time: to filter messages that refer to different stages of an event, when temporal boundaries for the event are unclear.
\item By location: to select messages according to whether they originate from or near the place that was affected by an event, or from areas that were not affected.
\end{compactenum}
Other classification dimensions can be envisioned. In general, we observe that the selection of the set of categories used by researchers and practitioners is usually driven by two main factors: the data that is present in social media during crises, and the information needs of response agencies.
None of these factors is static, as both can change substantially from one crisis situation to another.
We also remark that different types of disaster elicit different distributions of messages~\cite{olteanu_2015_what}.
\begin{table}[p]
\caption{Classification of various dimensions of content posted on social media during high impact events with description and related work references.}
\label{tbl:event_dimensions_summary}
\centering \scriptsize
\begin{tabular}{ p{3.0cm} p{9.8cm} }
\toprule
Classification dimension & Description/examples \\ \midrule
\multicolumn{2}{l}{\bf By factual, subjective, or emotional content} \\
Factual information & \textit{(Examples under ``By information provided'')}\\
Opinions & opinions, criticism (e.g. of government response) \\
Sympathy & condolences, sympathy~\cite{kumar2013whom}; condolences~\cite{acar2011twitter}; support~\cite{Hughes:2014}; thanks, encouragement~\cite{bruns2014crisis}; prayers~\cite{olteanu2014crisislex} \\
Antipathy & \textit{schadenfreude}, animosity against victims (e.g. because of long-standing conflict)\\
Jokes & jokes, trolling~\cite{metaxas2013rise}\\[4pt]
\multicolumn{2}{l}{{\bf By information provided}} \\
Caution and advice & caution and advice~\cite{imran2013extracting}; warnings~\cite{acar2011twitter}; hazard, preparation~\cite{olteanu2014crisislex}; tips~\cite{leavitt2014upvoting}; advice~\cite{bruns2014crisis}; status, protocol~\cite{Hughes:2014} \\[2pt]
Affected people & people trapped, news~\cite{caragea2011classifying}; casualties, people missing, found or seen~\cite{imran2013extracting}; self reports~\cite{acar2011twitter}; injured, missing, killed~\cite{Vieweg:2010}; looking for missing people~\cite{qu2011microblogging} \\[2pt]
Infrastructure/utilities & infrastructure damage~\cite{imran2013extracting}; collapsed structure~\cite{caragea2011classifying}; built environment~\cite{Vieweg:2010}; closure and services~\cite{Hughes:2014} \\[2pt]
Needs and donations & donation of money, goods, services~\cite{imran2013extracting}; food/water shortage~\cite{caragea2011classifying}; donations or volunteering~\cite{olteanu2014crisislex}; help requests, relief coordination~\cite{qu2011microblogging}; relief, donations, resources~\cite{Hughes:2014}; help and fundraising~\cite{bruns2014crisis} \\[2pt]
Other useful information & hospital/clinic service, water sanitation~\cite{caragea2011classifying}; help requests, reports about environment~\cite{acar2011twitter}; consequences~\cite{olteanu2014crisislex} \\ \\
\multicolumn{2}{l}{\bf By information source} \\
Eyewitnesses/Bystanders & members of public~\cite{metaxas2013rise}, victims, citizen reporters, eyewitnesses~\cite{diakopoulos2012finding,olteanu2014crisislex,bruns2014crisis} \\
Government & administration/government~\cite{olteanu2014crisislex}; police and fire services~\cite{Hughes:2014}; government~\cite{bruns2014crisis}; news organization and authorities~\cite{metaxas2013rise} \\
NGOs & non-government organizations~\cite{DeChoudhury:2012}\\
News Media & news organizations and authorities, blogs~\cite{metaxas2013rise}, journalists, media, bloggers~\cite{DeChoudhury:2012}; news organizations~\cite{olteanu2014crisislex}; professional news reports~\cite{leavitt2014upvoting}; media~\cite{bruns2014crisis}
\\[4pt]
\multicolumn{2}{l}{\bf By credibility} \\
Credible information & newsworthy topics, credibility~\cite{castillo2013predicting}; credible topics~\cite{canini2011finding}; content credibility~\cite{gupta2012credibility}; users and content credibility~\cite{gupta2014tweetcred}; source credibility~\cite{thomson2012trusting}; fake photos~\cite{gupta2013faking} \\
Rumors & rumor~\cite{Hughes:2014,castillo2013predicting}\\\\
\multicolumn{2}{l}{\bf By time} \\
Pre-phase/preparedness & posted before an actual event occurs, helpful for the preparedness phase of emergency management~\cite{petak1985emergency}; pre-disaster, early information~\cite{iyengar2011content,chowdhurytweet4act}\\
Impact-phase/response & posted during the impact phase of an event, helpful for the response phase of emergency management~\cite{petak1985emergency}; during-disaster~\cite{iyengar2011content,chowdhurytweet4act}\\
Post-phase/recovery & posted after the impact of an event, helpful during the recovery phase~\cite{petak1985emergency}; post-disaster information~\cite{chowdhurytweet4act,iyengar2011content} \\ \\
\multicolumn{2}{l}{\bf By location} \\
Ground-zero & information from ground zero (victims reports, bystanders)~\cite{longueville2009omg,ao2014estimating} \\
Near-by areas & information originating close to the affected areas~\cite{longueville2009omg} \\
Outsiders & information coming from other parts of world, sympathizers~\cite{kumar2013whom}; distant witness (in the sense of~\cite{carvin_2013_distant}); not on the ground~\cite{starbird2012learning}; location inference~\cite{ikawa2012location}\\
\bottomrule
\end{tabular}
\end{table}
\subsubsection{Supervised Classification} \label{sec:supervised_classification}
\iffalse
\begin{figure}[t]
\centering
\includegraphics[width=0.89\columnwidth]{figs/supervised-approach-v2.pdf}
\caption{Supervised classification}
\label{fig:supervised}
\end{figure}
\fi
When a set of example items in each category is provided, a \emph{supervised classification} algorithm can be used for automatic classification.
This type of algorithm `learns' a predictive function or model from features of these examples (see Section \ref{subsec:feature-extraction}) in order to label new, unseen data items. This set of examples is referred to as the \textit{training set}.
After a model has been learned from the training data, it is evaluated using a different, hold-out set of labeled items, not used during the training. This second set of examples is referred to as the \emph{testing set}.
Depending on the nature of data in hand, different pre-processing techniques can be used.
In any case, the input items are transformed into feature vectors, following the methods described in Section~\ref{subsec:feature-extraction}.
\spara{Training examples.}
The number of training examples required to achieve good accuracy depends on many factors, including the number of categories into which messages have to be classified, and the variability of messages inside each category. Typical sizes of training sets range from a few hundred~\cite{yin2012using} to a few thousand~\cite{imran2014aidr}. More examples yield better results in general, with diminishing results after a certain point.
In general, the accuracy of models created using training data from one crisis decreases when applied to a different crisis, or when applied to the same crisis but at a different point in time~\cite{imran2014coordinating}.
\spara{Feature selection.}
Even if messages are brief, the feature space in which they are represented is typically high dimensional (e.g. one dimension for every possible term). This introduces a number of problems including the amount of computational resources required for the data analysis, and it also increases the chances of over-fitting the training data. In this case, a \emph{feature selection method} (e.g. mutual information) should be employed as a first step to discard input features that have little or no correlation with the given training labels. Feature selection is an active area of machine learning research and state of the art techniques can be found in modern textbooks or journals (e.g.~\citeN{Guyon:JMLR-feature-selection-special-issue}).
\spara{Learning algorithms.}
After features have been extracted and selected, a machine learning algorithm can be applied. Supervised classification algorithms include, among others, na{\"i}ve Bayes, Support Vector Machines (SVM), logistic regression, decision trees, and random forests.
The choice of a method is largely dependent on the specific problem setting. For instance, \emph{ESA}~\cite{yin2012using,power2014emergency} uses na{\"i}ve Bayes and SVM, \emph{EMERSE}~\cite{caragea2011classifying} uses SVM, \emph{AIDR}~\cite{imran2014aidr} uses random forests, and \emph{Tweedr}~\cite{ashktorab2014tweedr} uses logistic regression.
While in most cases algorithms are used to predict a single label for each element, adaptations of these algorithms that generate multiple labels for each element are sometimes employed (e.g.~\citeN{caragea2011classifying}).
\spara{Ensemble/stacked classification.}
In some cases an explicit model of a certain factor is desired, as exemplified by the work of \citeN{verma2011natural}. They observe that messages that contribute the most to situational awareness are also those that are expressed using objective (as opposed to subjective) language. In this case, one can create a \emph{stacked classifier} in which at one level certain characteristics of the message are modeled (e.g. by having a classifier that classifies messages as objective or subjective) and at the next level these characteristics are combined with other characteristics also modeled by specific classifiers (e.g. writing styles such as formal or informal), and with features from the message itself. \citeN{verma2011natural} find that this approach performs better than directly using the input features.
Comparison in terms of classification effectiveness/accuracy is a non-trivial task. This is due to the use of different datasets, different baselines, and different performance measures (e.g. AUC, F1, precision). However, classification accuracies reported by most of the systems we have reviewed for this survey range from 0.60 to 0.90.
Recent collections available for research may contribute to make systems easier to compare.\footnote{\url{http://crisislex.org/}}
\subsubsection{Unsupervised Classification} \label{sec:unsupervised_classification}
\iffalse
\begin{figure}[t]
\centering
\includegraphics[width=0.55\columnwidth]{figs/unsupervised-approach-v2.pdf}
\caption{Unsupervised classification}
\label{fig:unsupervised}
\end{figure}
\fi
Clustering is an \emph{unsupervised machine learning method}; a family of methods that seek to identify and explain important hidden patterns in unlabeled data. Unsupervised machine learning methods include clustering, dimensionality reduction (e.g. principal component analysis), and hidden Markov models, among others.
The process of performing clustering,
begins by ingesting a set of items (e.g., documents, tweets, images) which are then processed with the objective of grouping similar items together.
In general, the goal is to form clusters in such a way that elements within a cluster are more similar to each other than to the elements that belong to other clusters. Many clustering algorithms have been developed based on different approaches, examples include \emph{K-means} (centroid-based), \emph{hierarchical clustering} (connectivity-based), \emph{DBSCAN} (density-based), among many others. For an overview of clustering methods, see~\citeN[Part III]{zaki_meira_2014_datamin}.
In the context of dealing with social media data during crises, clustering can help reduce the number of social media messages that need to be processed/examined by humans, for instance by displaying multiple equivalent messages as a single item instead of multiple ones.
This is the approach used by \emph{CrisisTracker}~\cite{rogstadius2013crisistracker}, which is a crowdsourced social media curation system for disaster awareness. The system, which collects data from Twitter based on predefined filters (i.e., keywords, bounding box), groups these tweets into \emph{stories}, which are clusters of tweets. These stories are then curated/classified by humans, whose effort can be greatly reduced if they are asked to classify entire clusters instead of single tweets.
The specific clustering method employed in this case is locality-sensitive hashing, an efficient probabilistic technique that uses hash functions to detect near-duplicates in data.
Another example is \textit{SaferCity}~\cite{berlingeriosafercity}, which identifies and analyzes incidents related to public safety in Twitter, adopting a spatio-temporal clustering approach based on the modularity maximization method presented in~\citeN{blondel2008fast} for event identification. Clusters are then classified using a semantic labeling approach using a controlled vocabulary
and based on a rank score provided by the Lucene library.\footnote{http://lucene.apache.org/}
In addition to clustering methods that partition the items into groups, there are \emph{soft clustering} methods that allow an item to simultaneously belong to several clusters with varying degrees. This is the case of topic modeling methods, out of which Latent Dirichlet Allocation (LDA) is one of the most popular. In the crisis domain, \citeN{kireyev2009applications} use topics extracted using LDA, including a weighting scheme that accounts for the document frequency (number of tweets containing a word) of the words in the tweet, as well as for the length of the tweet. When applied to data from an Earthquake in Indonesia in 2009, it detects topics that cover different aspects of the crisis such as \{ \emph{tsunami}, \emph{disaster}, \emph{relief}, \emph{earthquake} \}, \{ \emph{dead}, \emph{bodies}, \emph{missing}, \emph{victims} \}
and \{ \emph{aid}, \emph{help}, \emph{money}, \emph{relief} \}.
There is free software available for creating these topics models, such as MALLET~\cite{mccallum2002mallet}, which has been applied to crisis data by~\citeN{karandikar2010clustering}.
\subsubsection{Discussion}
In the context of social media data analysis, many factors influence the choice of a learning approach (i.e. supervised vs. unsupervised). For instance, approaches based on unsupervised learning certainly have advantages over supervised ones in cases where obtaining training examples would be prohibitively costly, would introduce an unacceptable delay or would simply not be possible. Unsupervised approaches are more useful in cases where information classes are completely unknown from a information seeker point of view.
However, in terms of usefulness, the output of an unsupervised approach might be considered less useful than an approach based on supervised learning. Emergency responders may be used to specific categorizations and methodologies (e.g. the MIRA framework used by the United Nations\footnote{\url{https://docs.unocha.org/sites/dms/Documents/mira_final_version2012.pdf} (accessed Nov. 2014).}) and may expect that software tools output the same categories.
\subsection{Sub-Document Analysis}
In contrast to text classification schemes which make predictions about data items, sub-document analysis techniques extract granular information from the \textit{content} of the data items. In this section we present two important sub-document analysis methods in the context of crisis situations: information extraction and text summarization.
\subsubsection{Information Extraction}\label{subsec:information_extraction}
The task of automatically extracting structured information from unstructured (e.g., plain text documents) or semi-structured (e.g., web pages) data is known as Information Extraction (IE).
The most common information extraction task is named entity extraction, which consists of detecting regions of a text referring to people, organizations, or locations~\cite{liu2011recognizing,ritter2011named}. This is the first step towards semantic enrichment (see Section~\ref{subsec:semantic-enrichment}).
In the context of crisis-related social media, information extraction can be used, for instance, to transform tweets reporting injured people in natural language (e.g. ``\emph{5 injured and 10 dead in Antofagasta}'') to normalized records such as \{$<$people-affected=5, report-type=injury, location=Antofagasta, Chile$>$, $<$people-affected=10, report-type=fatal-casualty, location=Antofagasta, Chile$>$\}.
These records are machine-readable, which means they can be easily filtered, sorted, or aggregated.
Information extraction from social media is a challenging task because of the informal writing style and the presence of many ungrammatical sentences, as was noted in Section~\ref{subsec:content-issues}.
State-of-the-art approaches to information extraction involve the use of probabilistic sequential models such as hidden Markov models, conditional Markov models, maximum-entropy Markov models, or conditional random fields. Heuristics based on regular expressions can also be applied to this problem, although these are in general less effective than probabilistic methods.
Varga et al.~\citeyear{varga2013aid} use linguistic patterns and supervised learning to find ``trouble expressions'' in social media messages. For instance, in a tweet such as ``\emph{My friend said infant formula is sold out. If somebody knows shops in Sendai-city where they still have it in stock, please let us know},'' the nucleus of the problem is the sentence ``\emph{infant formula is sold out}.'' This extraction is then used to match tweets describing problems to tweets describing solutions to those problems.
Imran et al.~\citeyear{imran2013practical,imran2013extracting} apply conditional random fields to the extraction of information from tweets. Their method proceeds in two steps. First, tweets are classified to consider categories such as ``infrastructure damage,'' ``donations,'' and ``caution and advice.''
Then, a category-dependent extraction is done, where for instance for ``infrastructure damage'' tweets, the specific infrastructure reported damaged is extracted, while for ``donations'' tweets, the item being offered in donation is extracted.
\subsubsection{Summarization}\label{sec:summarization}
Another approach to dealing with information overload involves presenting users with a text-based representation of the evolving event. Text summarization refers to the generation of a text summary of a document or set of documents \cite{nenkova:summarization-survey}. Text summarization systems are optimized to generate a text summary that contain only the core topics discussed in the set of documents. Most systems produce this summary by \textit{extracting} key sentences from the input document. This is in contrast to systems that produce a summary by \textit{abstracting} or generating new sentences.
During crisis events, text summarization must be done in an \textit{incremental} and \textit{temporal} manner.
Incremental text summarization, also known as \textit{update summarization}, refers to generating a summary given:
\begin{inparaenum
\item the set of documents to be summarized, and
\item a reference set of documents which the user has read.
\end{inparaenum}
The objective for the system is to produce a summary only of data the user has not already read~\cite{tac2008:update-summarization-overview}. Temporal text summarization refers to creating an extractive summary from a set of time-stamped documents, usually in retrospect~\cite{allan:temporal-summaries,nallapati:event-threading,feng:cikm2009}.
Drawing on the work from summarization research, the TREC Temporal Summarization track focuses on generating updates relating to unfolding crisis events immediately after their occurrence~\cite{trec2013:temporal-summarization-overview}. Table~\ref{tab:trec-ts-example} presents an example of this type of summary. The focus of this initiative is to first define standardized metrics for the task and then to encourage the development of systems optimizing them.
These metrics include time-sensitive versions of precision and recall, ensuring that systems are penalized for \textit{latency}: delivering information about an event long after it occurred. In addition, the metrics include a redundancy penalty to prevent systems from delivering repetitive information. In order to optimize these metrics, systems can use staged text analysis with standard information retrieval measures \cite{trec2013:terrier}.
Alternatively, systems can use regression-based combinations of features from classic text summarization literature \cite{qi:temporal-summarization,trec2013:hltcoe}.
Other methods are purely content-based, hierarchically clustering sentence text~\cite{wang:update-summarization}. In the context of social media, Shou et al.~\citeyear{Shou:sumblr} propose a system for online update summarization based on incremental clustering. The performance is evaluated under experimental conditions different from the TREC track, making it difficult to compare with other results.
Although algorithms for summarization exist for crisis events, their development is still preliminary and several challenges remain. Foremost, the relative importance of different features is not well understood. To date, research has primarily used features from batch text summarization. A second challenge is scale. Many algorithms require aggressive inter-sentence similarity computation, a procedure which scales poorly.
\begin{table}[t]
\caption{Example summary from the TREC 2013 Temporal Summarization Track. Updates reflect new or updated information as it is reported. }
\label{tab:trec-ts-example}
\centering
{\scriptsize
\begin{tabular}{p{0.9in}p{4.2in}}
\toprule
time & update \\
\midrule
Nov 21, 2012 10:52 & Tel Aviv bus bombing; 13 injuries; reported on bus line 142; occured on Shaul Hamelech street; No claims of responsibility; 3 badly hurt; occured in the heart of Tel Aviv near military hdqtrs \\
Nov 22, 2012 20:49 & occurred in an area with many office buildings; occured in area with heavy pedestrian traffic; first notable bombing in Tel Aviv since 2006; At least 28 people were wounded; Hamas' television featured people praising the attack; Khaled Mashal, leader of Hamas, categorically rejected any connection of the bombing to his group; ... \\
Nov 26, 2012 04:33 & an Israeli Arab man was arrested on charges of planting the explosive device on the bus; Suspect was reportedly connected to Hamas; Suspect was reportedly connected to the Islamic Jihad \\
Nov 26, 2012 14:49 & The Romanian Foreign Minister condemned the bombing, \\
Nov 29, 2012 04:55 & govt rep refers to attack as terrorist attack \\
Nov 30, 2012 05:22 & Fears about Bus Bomb Before the Cease-Fire: Could derail peace talks \\
Nov 30, 2012 06:47 & The suspect remotely detonated the explosive device; suspect hid device in advance on the bus; The explosive device contained a large quantity of metal shrapnel designed to cause maximum casualties; The suspect later on confessed to carrying out the bus attack; suspect prepared the explosive device; suspect chose the target of the attack; ... \\%suspect purchased a mobile phone; suspect received an Israeli citizenship as part of a family unification process \\
\bottomrule
\end{tabular}}
\end{table}
\subsection{Challenges}\label{sec:datamin-challenges}
In addition to the discussion on generic data-related challenges in Section~\ref{sec:data-challenges}, the following are particularly relevant to mining:
\spara{Combining manual and automatic labeling.}
In a supervised learning setting, human labels are necessary, but they may be costly to obtain. This is particularly problematic in crises that attract a multilingual population, or for tasks that require domain knowledge (e.g. people who know informal, local place names in Haiti, and who speak Haitian Creole).
Labeled data are not always reliable, and may not be available at the time of the disaster; in this case, a hybrid approach that mixes human labeling and automatic labeling can be employed~\cite{imran_2013_engineering}. The selection of items to be labeled by humans can be done using \emph{active learning}, a series of methods to maximize the improvement in classification accuracy as new labels are received.
\spara{Domain adaptation.}
Ideally, one would like to avoid having to re-train an automatic classifier every time a new crisis occurs. However, simply re-using an existing classifier trained on data from a previous crisis does not perform well in practice, as it yields a substantial loss in accuracy, even when the two crises have several elements in common~\cite{imran2013practical}.
In machine learning, \emph{domain adaptation} (or \emph{domain transfer}) is a series of methods designed to maximize the accuracy of a classifier trained on one dataset, adapting it to continue to perform well on a dataset with different characteristics. To the best of our knowledge, these methods have not been applied to crisis-related social media data.
\section{Semantic Technologies in Disaster Response}\label{sec:semantics}
One of the main goals of semantic technologies is to allow users to easily search through complex information spaces, and to find, navigate, and combine information. In the context of social media use during crises and mass convergence events, this is achieved by \emph{linking} data elements to concepts in a machine-readable way, enabling the representation of a situation as a complex and interrelated set of elements.
\subsection{Semantic Enrichment of Social Media Content}\label{subsec:semantic-enrichment}
Semantic technologies are particularly useful in social media, because they provide a powerful method for dealing with the variety of expressions that can be used to refer to the same concept, and with the many relationships that can exist between concepts.
For instance, suppose we are looking for messages related to infrastructure damage using a keyword-query search. We would think that searching for something such as ``damage AND (airport OR port OR bridge OR building ...)'' would be sufficient, until we notice that it is not only difficult to cover every particular type of infrastructure (airport, port, bridge, building, etc.) but also to cover every particular instance of that type (for instance, there are tens of thousands of airports in the world).
\emph{Named entity linking} is a widely-used semantic technology that deals with the above problem. It operates in two phases. First, a \emph{named entity recognizer} module detects entities---such as names of persons, places, and organizations. Second, for each named entity that is found, a \emph{concept} is located that more closely matches the meaning of that named entity in that context.
Concepts are generally operationalized as Uniform Resource Locators (URLs). For instance, \citeN{zhou2010resolving} link named entities to the URLs of articles on Wikipedia. In a phrase such as ``\emph{Terminal 2 of JFK was damaged},'' the named entity corresponding to the segment ``JFK'' would be linked to the URL \url{https://en.wikipedia.org/wiki/John_F._Kennedy_International_Airport}.
There are several free, and commercial services that can be used to perform named entity linking, including
Alchemy,\footnote{\url{http://www.alchemyapi.com/}}
Open Calais,\footnote{\url{http://www.opencalais.com/}} and
Zemanta.\footnote{\url{http://developer.zemanta.com/}}
Once an element is linked to a concept, further automatic annotation can be done by following links from the concept. Returning to our example of ``\emph{Terminal 2 of JFK was damaged},'' if we go to its Wikipedia page,\footnote{Or to its related semantic resource, DBPedia \url{http://dbpedia.org/}.} we can learn that ``JFK'' is an instance of the class ``airport.'' The airport concept is represented in this case by the URL of a Wikipedia category page to which the JFK Airport page belongs (\url{https://en.wikipedia.org/wiki/Category:Airports}), which in turn belongs to the category of transport building and structures (\url{https://en.wikipedia.org/wiki/Category:Transport_buildings_and_structures}).
After named entity linking, messages that have been semantically enriched can be used to provide \emph{faceted search}, a popular approach to interactively search through complex information spaces. In faceted search the information of interest can be found not only by specifying a related keyword, but also by specifying a concept or concepts associated with the items of interest. In the example, we could select ``airport'' from a list of buildings and structures and then find a series of social media messages that are relevant, but that do not necessarily include the specific word ``airport'' in them.
\citeN{abel2011leveraging}, present an adaptive faceted search framework to find relevant messages on Twitter. The framework enriches tweets with semantics by extracting entities (i.e. persons, locations, organization), and then finding and linking those entities with external resources to create facets. Each facet enables search and navigation of relevant semantically related content.
In follow-up work~\cite{abel2012semantics}, they introduce \emph{Twitcident}, a system that supports semantic filtering, faceted search and summarization of tweets.
The semantic-based approach is also implemented in \emph{Twitris 2.0}~\cite{jadhav2010twitris} and \emph{EDIT}~\cite{traverso_2014_edit}, which present event-related social media capturing semantics in terms of spatial, temporal, and thematic dimensions.
\subsection{Ontologies for Disaster Management}\label{sec:ontologies}
Information technologies to support disaster response often involve interactions between software operated by different agencies, and/or provided by different developers or vendors.
Allowing computer systems to communicate information in a unified way is a key challenge in general, but especially during crisis events where different agencies must address different dimensions of a problem in coordination with each other~\cite{hiltz2011introduction}.
Interoperability at the semantic level requires centralized specifications describing machine-understandable common vocabularies of concepts and linkages between them. An effective way of achieving this is to use machine-understandable ontologies that define, categorize and maintain relationships between different concepts to facilitate common understanding, and unified communication.
\begin{table}
\caption{Crisis ontologies, including some of the classes and attributes they cover, and the format in which they are specified (OWL: Ontology Web Language, RDF: Resource Description Framework).}
\label{tbl:crisis_ontologies}
\centering \scriptsize
\begin{tabular}{ p{1.2cm} p{6.2cm} p{.5cm} p{1.7in}}
\toprule
Ontology Name & Coverage & Format & Reference \\ \midrule
SOKNOS & resources, damage, disasters & OWL & \cite{babitski2011soknos} \\
HXL & damage, geography, organization, disasters & RDF & \url{http://hxl.humanitarianresponse.info/} \\
SIADEX & processes, resources, geography & RDF & \cite{de2005siadex} \\
OTN & specific to infrastructure & OWL & \cite{lorenz2005ontology} \\
MOAC & damage, disasters, processes, resources & RDF & \url{http://observedchange.com/moac/ns/} \\
FOAF & emergency management people & RDF & \url{http://www.foaf-project.org/} \\
AktiveSA & \mbox{transportation, meteorology, processes, resources, people} & OWL & \cite{smart2007aktivesa} \\
IntelLEO & response organizations & RDF & \url{http://www.intelleo.eu/} \\
ISyCri & damages, processes, disasters & OWL & \cite{benaben2008metamodel} \\
WB-OS & features, components and information to build crisis management web sites & XML & \cite{chou2011ontology}\\
EDXL-RM & \mbox{data exchange language for resource management} & XML & \url{https://www.oasis-open.org/} \\
\bottomrule
\end{tabular}
\end{table}
Table~\ref{tbl:crisis_ontologies} lists some of the ontologies that have been introduced in recent years.
Some examples from this table:
\begin{itemize}
\item The {\em Humanitarian eXchange Language} (HXL)\footnote{\url{http://hxl.humanitarianresponse.info/}} is an ontology created in 2011 and 2012 and is currently under review; it describes 49 classes and 37 properties. The focus of HXL is mainly on four areas: organization (i.e., formal response agencies like military, charities, NGOs), disaster (i.e., classification of disasters such as natural, man-made), geography (i.e., event location, geo-location of displaced people), and damage (i.e., damages related to humans, infrastructure).
\item \emph{Management of A Crisis} (MOAC)\footnote{\url{http://observedchange.com/moac/ns/}} is an ontology with 92 classes and 21 properties covering four areas: disaster, damage, processes (i.e., rescue, search, evacuation processes), and resources (i.e., services, vehicles, tents). HXL and MOAC have elements in common: in both cases the objective is to describe different aspects of a crisis, including its effects, the needs of those affected, and the response to the crisis.
\item \emph{Integrated Data for Events Analysis} (IDEA)\footnote{\url{http://vranet.com/IDEA.aspx}} is a framework for coding social, economic and political events. It is used in the {\em Global Database of Events, Language and Tone} (GDELT),\footnote{\url{http://gdeltproject.org/}} which is a machine-generated list of event data extracted from news reports.
\item \emph{Service-orientierte ArchiteKturen zur Unterst{\"u}tzung von Netzwerken im Rahmen Oeffentlicher Sicherheit} (``Service-oriented architectures supporting networks of public security,'' SOKNOS)\footnote{http://soknos.de/} is an ontology for information integration for resource planning, including damage and resource categorization during disasters.
\end{itemize}
To support communication among different ontology-based systems, the problem of ontology heterogeneity needs to be solved by performing an \emph{ontology mapping}, which is the process of mapping the concepts of two ontologies from the same or from overlapping domains.
Many approaches have been proposed to perform ontology mapping. For instance,~\citeN{tang2006using} treated this as a decision-making problem and proposed an approach based on Bayesian decision theory. For a survey on ontology mapping techniques, see~\citeN{noy2004semantic}.
The ontologies in Table~\ref{tbl:crisis_ontologies} are crisis-specific, but not social-media specific. However, they can be combined with ontologies describing social media concepts such as users, tagging, sharing, and linking.
For instance, the ``Semantically Interlinked Online Communities'' (SIOC)\footnote{http://sioc-project.org/} ontology, originally developed to model sites such as blogs and online forums, has recently been extended to support the modeling of microblogs by adding concepts such as \emph{follower} or \emph{follows}.
An ontology specific to Twitter appears in~\citeN{celino2011making} and includes user sentiments and locations, while~\citeN{passant2008meaning} enables semantic tagging of social media data through an ontology called Meaning-Of-A-Tag (MOAT).
For a survey on ontologies developed for social media, see~\citeN{bontcheva2012making}.
\section{Summary and Future Research Directions}\label{sec:conclusions}
In a relatively short time period---roughly 4 to 6 years---the research community working on the topics we have covered here has achieved a fairly high degree of maturity with respect to filtering, classifying, processing, and aggregating social media data during crises.
However, the underlying (although sometimes explicitly stated) claim behind this line of work, i.e. that this research is \emph{useful} for the public and/or formal response agencies, that it has the potential to save lives and/or property during an emergency, remains to be seen.
While there are notable exceptions including the American Red Cross, the US Federal Emergency Management Agency, UN Office for the Coordination of Humanitarian Affairs and the Filipino Government, among others, the use of social media is still experimental for many organizations and not yet part of their normal, day-to-day operations.
There are two main directions in which we see future research going. First, continue deepening the data processing capabilities that have been the main focus of computing research on this topic thus far. Second, engage more deeply with human-centered approaches toward making the computing research the foundation of viable systems that emergency responders can implement.
\subsection{Deepening Data Processing Capabilities}
\spara{From situational awareness to decision support.}
The systems we described in Section~\ref{sec:systems} for processing social media during disasters have a strong focus on situational awareness, which is an important first step but might not be enough for emergency management.
During an emergency, social media is used as an information source in order to make decisions. Therefore, next-generation systems should be designed and evaluated in terms of their decision-support capabilities. This might even include forecasting using signals from social media.
\spara{Extending to other types of media.}
Data from various sources should be processed and integrated: ``The strategies of emergency services organizations must also recognize the significant interweaving of social and other online media with conventional broadcast and print media.''~\cite{bruns2014crisis}.
There are some examples of the processing of other types of information items during crises, including short messages (SMS)~\cite{Melville:2013:AVY:2487575.2488216}, news articles in traditional news media and blogs~\cite{leetaru_2013_gdelt}, and images~\cite{abel2012semantics}.
\spara{Verifying information.}
Systems that receive user-generated content are always exposed to abuse, which can be countered by a mixture of automatic and manual methods. For instance, algorithms to detect false product reviews are deployed by most major online retailers~\cite[Chapter 10]{liu_2012_sentiment}.
In the crisis domain, determining the credibility of information posted on social media is a major concern for those who process information (e.g. computer scientists, and software engineers), and for the information consumers (e.g. the public and formal response organizations)~\cite{hiltz2011introduction,hughes_2014_smem}.
Automatic classification can be used to filter out content that is unlikely to be considered credible~\cite{gupta2012credibility,castillo2013predicting}, or to annotate messages seen by users with credibility scores automatically~\cite{gupta2014tweetcred}. Additionally, the public itself can be mobilized to confirm or discredit a claim through crowdsourcing~\cite{popoola2013information}.
\subsection{Beyond Data Processing}
\spara{Designing with the users.}
Considering the number of systems that have been designed and built so far, there is little research on how usable and useful are those systems (with some exceptions, e.g.~\cite{robinson2013understanding,tucker2012straight}).
How should information be presented to users? How should users interact with it?
The key to answering these question lies with the users themselves, who should be brought into the process of designing the systems, dashboards, and/or visualizations they require to serve their needs. A highly regarded methodology for achieving this is \emph{participatory design}~\cite{hughes_2014_pio}.
\spara{Helping governments and NGOs communicate with the public.}
Three days \emph{before} Typhoon Pablo made its landfall in the Philippines in 2012, government officials were already calling users to use the hashtag ``\#pabloph'' for updates about the typhoon.
An effective use of hashtags has also been encouraged by the United Nations Office for the Coordination of Humanitarian Affairs~\cite{ocha_2014_hashtags}. Computational methods can be used not only to help formal response agencies choose which hashtags to use, but more generally, to help them design and evaluate effective communication strategies in social media (see~\citeN{veil2011work} regarding best practices for crisis communications using social media).
The conversation between the public and formal organizations can also be conducted through platforms that, instead of passively waiting for people to post information, ask them directly to answer certain questions that are relevance for the emergency response or relief operations~\cite{ludwig_2015_crowdmonitor}.
\spara{Social media for coordinating actions.}
The final output of the processing of social media messages is not limited to the presentation of information in a given format. Computational methods can be applied to augment the information in a number of ways.
For instance, Varga et al.~\citeyear{varga2013aid} match problem-tweets (``infant formula is sold out in Sendai'') to solution-tweets (``Jusco supermarket is selling infant formula''), and Purohit et al.~\citeyear{purohit2013emergency} match tweets describing urgent need of resources (``we are coordinating a clothing drive for families affected'') with tweets describing the intention to donate them (``I've a bunch of clothes I want to donate'').
We regard these efforts as preliminary results towards the ability to use social media as a mechanism for coordination of action in future emergency situations.
A recent special issue of the Journal on Computer Supported Collaborative Work explores various ways in which computing can support collaboration and coordination during an emergency~\cite{pipek_2014_special}.
\spara{Acknowledgments.} We are thankful to Jakob Rogstadius and Per Aarvik who pointed us to historical information.
We would like to thank our collaborators and co-authors Patrick Meier, Alexandra Olteanu, and Hemant Purohit.
\footnotesize{\bibliographystyle{ACM-Reference-Format-Journals}
|
\section{Introduction}
Through centuries, the stability of different phenomena has been studied by means of the bifurcation theory. Mathematically, the observed states are modeled as solutions of nonlinear equations, a mathematical context where the intuitive notion of stability is made precise. Roughly speaking, one expects that a slight change of a parameter in a system should not have great influence, but rather that stable solutions change continuously in a unique way. That expectation is verified by the Implicit Function Theorem. Consequently, as long as a continuous branch of solutions preserves its stability, no dramatic change is observed when the parameter is varied. However, if that \emph{ground state} loses its stability when the parameter reaches a critical value, then the state is no longer observed, and the system itself organizes a new stable state that \emph{bifurcates} from the ground state.
In \cite{S-W}, Smoller and Wasserman present a simple criterion based on the Index Morse determining the existence of bifurcation points which, together with the Implicit Function Theorem, has enabled several authors to study different stability problems. For instance, Bettiol and Piccione \cite{BP} used this technique to obtain results about the existence and non-existence of metrics with constant scalar curvature and as close as desired to a given homogeneous metric on the sphere. Jin, Lin and Xu in \cite{JLX} obtained some multiplicity results for the Yamabe equation on the sphere $\mathbb{S}^n$. Recently, Ram\'\i rez-Ospina \cite{RO} proved the existence of uncountably many unit volume constant scalar curvature metrics in product manifolds $\mathbb{T}^k\times M$ where $\mathbb{T}^k$ is a flat $k$-torus and $M$ is a compact manifold. Among all these works, we point out one authored by Al\'\i as and Piccione \cite{AP}, where it is shown the existence of an infinite sequence of isometric embeddings of tori with constant mean curvature (CMC) in Euclidean spheres that are not isometrically congruent to the CMC Clifford torus.
Motivated by this work, we propose to study the rigidity of families of CMC hypersurfaces in different (and more general) contexts. Roughly speaking, a family $\mathcal F$ of CMC hypersurfaces is (locally) rigid if given any other CMC hypersurface $\Sigma$ which is sufficiently close to some element of $\mathcal F$, then $\Sigma$ must be \emph{congruent} to some element of $\mathcal F$ (see Section~\ref{sub:niftheory} for details). Constant mean curvature hypersurfaces have a well-known variational characterization as critical points of the area functional under variations preserving certain volume. Our aim is to show when a family of hypersurfaces is rigid or it has some bifurcation point. A natural occurrence of foliations by CMC hypersurfaces appears in certain warped product spaces, which are the main object of our study.
Finally, we will apply our results to the spatial fiber of the so-called de Sitter and Anti-de Sitter Schwarzschild spacetimes. These classical solutions of the vacuum Einstein equation for static, spherically symmetric spacetimes describe the simplest model of a black hole which is known. Recently, these spaces have received an increasing attention, especially with the so-called AdS/CFT correspondence (or Maldacena duality), a conjectured equivalence between a string theory on a {\em bulk} space (typically, the product of Anti-de Sitter spaces by round spheres) and a quantum field theory without gravity on the {\em boundary} of the initial space, which behaves as an hologram of lower dimension (see \cite{fisicos} for a detailed introduction on the topic).
The paper is organized as follows. In Section 2, we establish the basic notions of bifurcation and local rigidity; and we present the variational setting in which we will work.
The bifurcation phenomenon under volume-preserving variations of CMC hypersurfaces is analyzed in Section 3. As a first step, we consider the case of families of CMC hypersurfaces whose associated potential function of the Jacobi operator is constant (see \eqref{potential}). For this general setting, including the space of CMC hypersurfaces which are orbits for some isometric action of a Lie group (as it happens in the warped product case), we obtain simple sufficient conditions to guarantee the stability of an one-parameter family of hypersurfaces. Our main results are summarized in the following statements (see Theorem \ref{teo1}, Corollary \ref{cor1} and Remark \ref{hypopen} (i) for details).
\begin{theoremA}
Let $\{\partial\Omega_\gamma\}_{\gamma\in I}$ be a family of CMC hypersurfaces in an $n$-dimensional Riemannian manifold $(M,g)$. Assume that, for each $\gamma$ the map
\[Q(\gamma):=||\mathrm I\!\mathrm I_{\gamma}||^2+(n-1)\mathrm{Ric}(n_\gamma,n_\gamma) \]
is constant on $\partial \Omega_\gamma$, where $||\mathrm I\!\mathrm I_{\gamma}||^2$ is the squared norm of the second fundamental of $\partial \Omega_\gamma$
and $\mathrm{Ric}(n_\gamma,n_\gamma)$ is the normalized Ricci curvature of $M$ evaluated at $n_\gamma$ the unit normal to $\partial\Omega_\gamma$.
Then, the family $\{\partial \Omega_{\gamma}\}_{\gamma\in I}$ is rigid for all $\gamma$ if
\[\mu_1(\gamma)-Q(\gamma)>0\]
where $\mu_1(\gamma)$ represents the first nonzero eigenvalue of the Laplacian defined on $\partial \Omega_\gamma$ with the induced metric. In particular, such a family is rigid if one of the following statements hold:
\begin{itemize}
\item[$i)$] $\partial \Omega_\gamma$ is convex\footnote{We recall that $\partial \Omega_\gamma$ is convex if and only if its second fundamental form is positive definite with respect to the inward normal direction.} and has non-positive Ricci curvature,
\item[$ii)$] $\mathrm{Ric}(n_\gamma,n_\gamma)\leq-1/(n-1)||\mathrm I\!\mathrm I_\gamma||^2$,
\item[$iii)$] $\partial \Omega_\gamma$ is Ricci flat and $\mu_1(\gamma)\geq||\mathrm I\!\mathrm I_\gamma||^2$, where $\mu_1(\gamma)$ is the first nonzero eigenvalue of $\Delta$ on $\partial \Omega_\gamma$.
\end{itemize}
\end{theoremA}
\smallskip
For the particular case of warped product spaces, we give sufficient conditions for both the existence and non-existence of bifurcation points. The results for this case are summarized as follow (see Propositions \ref{prop2.3}, \ref{prop5.1} for more details).
\begin{theoremB}
Let $((r_1,r_2)\times P,dr^2+\alpha^2(r)g^P)$ be a warped product space where $P$ is an $(n-1)$-dimensional closed (compact and without boundary) manifold; and consider the one-parameter family of CMC hypersurfaces $\{\{r\}\times P\}_{r\in(r_1,r_2)}$.
\begin{itemize}
\item[(i)] If
\[\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)<\frac{\hat{\mu}_1}{n-1}\] for all $r\in (r_1,r_2)$ where $\hat{\mu}_1$ is the first non-zero eigenvalue of the Laplacian on $(P,g^P)$, then $\{\{r\}\times P\}_{r\in(r_1,r_2)}$ has no bifurcation point.\smallskip
\item[(ii)] If
\[\lim_{r\rightarrow r_2 }\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)=\infty
\]
then, there exists an infinite sequence of bifurcation points.
\end{itemize}
\end{theoremB}
Furthermore, we give some geometric conditions (on terms of the scalar and mean curvatures of warped product and
the scalar curvature of its fiber), which allow us to ensure the rigidity of an one-parameter family of
CMC hypersurfaces in a warped product, see Corollary \ref{corsc}. This type of situation occurs in General Relativity, and we will discuss in Section 4 an application of these results to the spatial fiber of the de Sitter and Anti-de Sitter Schwarszchild spacetimes. We show that the family of spheres around the blackhole are locally rigid for the former (including the classical Schwarszchild spacetime) while the collapsing spheres on the latter have infinitely many bifurcation points.
\section{Preliminaries}\label{section2}
Let us state some of the basic elements, results and notations that we are going to use in the rest of the paper. Along this paper, we will consider an $n$-dimensional Riemannian manifold $(M^n,g)$. In general, a local coordinate system will be denoted by $(x_{i})_{i=1}^{n}$ and, in the adapted case for the boundary, we will assume that $\partial/\partial x_n$ is the normal vector to $\partial M$ pointing inward. The volume and area elements of $M$ and $\partial M$ will be denoted by $dv$ and $d\sigma$ respectively. \\
\subsection{On bifurcation theory}
\label{sub:niftheory}
This section is devoted to introduce roughly the basic framework about bifurcation theory needed for the rest of this paper. For details, we refer the readers to \cite{S-W,AP} and the references therein.
Let us consider a Banach space ${\mathcal{M}}$ and a continuous path of $C^{k}$-functionals (with $k\geq 2$) $\biffun{\lambda}:{\mathcal{M}}\rightarrow \mathbb{R}$ where $\lambda$ varies on a prescribed interval $I$. Assume that we also have a continuous path of critical points $\{x_{\lambda}\}\subset {\mathcal{M}}$ of the corresponding functionals, that is, $d\biffun{\lambda}(x_{\lambda})=0$.
Fix $\lambda_*\in I$. We will say that $\lambda_*$ is \textit{a point of bifurcation} if there exists a sequence $\{\lambda_n\}_n\subset I$ and a sequence $\{x_n\}_n$ satisfying:
\begin{itemize}
\item[(1)] $\lim\limits_{n\rightarrow \infty}\lambda_n=\lambda_*$ and $\lim\limits_{n\rightarrow \infty}x_n=x_*$,
\item[(2)] $d\biffun{\lambda_n}(x_n)=0$ for all $n$,
\item[(3)] $x_n\neq x_{\lambda_n}$ for all $n$.
\end{itemize}
If $\lambda_*$ is not a bifurcation point, we will just say that the family $\{x_{\lambda}\}_{\lambda}$ is \textit{locally rigid} at $\lambda_*$.
One of the classical criterion to determine when a point is of bifurcation is related with the so-called \textit{Morse index}. Recall that the Morse index of a critical point $x_{0}$ of a smooth functional $\biffun{}$ is equal to the dimension of the maximal subspace of the tangent space $T_{x_{0}}{\mathcal{M}}$ where the second variation $d^2(\biffun{})_{x_{0}}$ is negative definite. Such an index will be denoted in general by $i(x_{0},\biffun{})$.
On the one hand, essentially, a variation of the Morse index $i(\biffun{\lambda},x_{\lambda})$ along the interval $I$ will indicate the existence of a bifurcation point (see for instance \cite{S-W} for details). More precisely, under suitable fredholmness assumptions, if there exist $\lambda_1,\lambda_2\in I$ with $\lambda_1<\lambda_2$ such that: $d^2(\biffun{\lambda_i})_{x_{\lambda_i}}$ are non-singular for $i=1,2$ and $i(x_{\lambda_1},\biffun{\lambda_1})\neq i(x_{\lambda_2},\biffun{\lambda_2})$, then there exists a bifurcation point $\lambda_{*}\in (\lambda_1,\lambda_2)$.
On the other hand, using the Implicit Function Theorem, one sees that if $d^2(\biffun{\lambda_*})_{x_{\lambda_*}}$ is non singular then the family $\{x_{\lambda}\}$ is \textit{locally rigid} at $\lambda_*$. In particular, if $i(\biffun{\lambda},x_{\lambda})=0$ for all $\lambda$ in $I$, then there is no bifurcation point.
In this paper, using the above conditions, we will study local rigidity and bifurcation by analyzing the spectrum of $d^2(\biffun{\lambda})_{x_{\lambda}}$ for all $\lambda$. Essentially, we will determine the number of negative eigenvalues for each $\lambda$ (counting its multiplicity) and we will study the evolution of such a number as $\lambda$ runs along $I$.
\subsection{Stating the variational problem}
Let us recall the classical variational setup of the constant mean curvature (CMC) hypersurfaces as critical points of the area functional under variations preserving the volume.
\smallskip
In order to use the bifurcation theory above mentioned, let us describe the elements involved in this approach. The Banach space ${\mathcal{M}}$ will be the space of open subsets $\Omega$ of an $n$-dimensional Riemannian manifold $(M,g)$ with compact closure and whose smooth boundary is a connected and orientable smooth manifold. The variations on ${\mathcal{M}}$ will be defined in the following way: a differentiable map $X:(-\epsilon,\epsilon)\times \partial \Omega \rightarrow M$ is called a {\em CMC hypersurface variation of $\partial \Omega$} if $(i)$ $X_{t}:\partial \Omega\rightarrow M$, defined by $X_{t}(x)=X(t,x)$ is an immersion whenever $|t|<\epsilon$, for $\epsilon>0$ small enough; and $(ii)$ $X(0,x)=i(x)$ where $i$ is the inclusion map.
Given an element $\Omega\in {\mathcal{M}}$ and a variation $X$ of $\partial \Omega$ denote by $\partial \Omega_{t}=X_t(\partial \Omega)$. For values of $t$ small enough, $\partial \Omega_{t}$ is also a connected and oriented smooth submanifold. Moreover, it contains an open subset $\Omega_{t}$ whose closure is also compact. In conclusion, the hypersurface variation defines in a natural way a variation of the open subset $\Omega$ denoted by $\Omega_{t}$, which is also an element of ${\mathcal{M}}$. Among all the hypersurface variations, we are going to be interested in the so-called \textit{normal} variations. A hypersurface variation $X$ is normal if the variation vector field $\frac{\partial X}{\partial t}\big|_{t=0}$ is parallel to the unit normal vector field $N$ with respect to $\Omega$ and the volume is preserved. In particular, if we denote by $T_{\Omega}^N {\mathcal{M}}$ the linear subspace of $T_{\Omega} {\mathcal{M}}$ only determined by normal variations, we can deduce that $T_{\Omega}^N {\mathcal{M}}$ is naturally identified with the set of smooth functions with compact support $\mathcal{C}^\infty_0(\partial \Omega)$, where such an identification is essentially obtained by defining $f=\left\langle \frac{\partial X}{\partial t}\big|_{t=0}, N\right\rangle\in\mathcal{C}_0^\infty(M)$ for a normal variation $X$. For the volume-preserving property, let us recall that the volume form varies as
\begin{equation}\label{forvol}
\mathrm{Vol}(\Omega_{t})= \int_{\Omega_{t}}dv=\int_{\Omega}dv+ \int_{[0,t]\times \partial \Omega}X^*(dv).
\end{equation}
\noindent and so, the normal variations has an associated function $f$ satisfying
\begin{align}\label{volpre}\int_{\partial \Omega} f d\sigma=0.
\end{align}
Finally, we will consider the area functional restricted to such a volume-preserving variations. The Lagrange multiplier method leads us then to the functional $\biffun{\lambda}$ defined as:
\begin{equation}
\begin{array}{cccc}
\biffun{\lambda}: & {\mathcal{M}} & \longrightarrow & \mathbb{R}\\
& \Omega & \longmapsto & \textrm{Area}(\partial \Omega) + \lambda \mathrm{Vol}(\Omega),\end{array}
\end{equation}
where $\lambda$ varies on a prescribed interval $I$ (to be determined later).
\smallskip
Now, we are ready to determine the first variation of our functional $\biffun{\lambda}$. For this, consider $X$ a normal variation and $f$ its associated function satisfying \eqref{volpre}. Denote by
\[
A(t)=\textrm{Area}(\partial \Omega_t),\quad \mathcal{V}(t)=\mathrm{Vol}(\Omega_{t}).
\]
Taking into account \eqref{forvol}, we deduce that
\[
\frac{\partial}{\partial t}\Big|_{t=0}A(t)=-(n-1)\int_{\partial \Omega}fH d\sigma\quad \hbox{\and} \quad \frac{\partial}{\partial t}\Big|_{t=0} \mathcal{V}(t)=\int_{\partial \Omega} f d\sigma
\] where $H$ is the mean curvature on $\partial \Omega$ (for more details see \cite[Lemma 2.1]{BCE}).
So, that the first variation of our functional takes the following form
\begin{align}
\delta\biffun{\lambda}_{\Omega}(f)=& \frac{\partial}{\partial t}\Big|_{t=0} A(t) + \lambda \frac{\partial}{\partial t}\Big|_{t=0}\mathcal{V}(t)\nonumber\\
= &\int_{\partial \Omega} \left(-(n-1)H + \lambda\right)f d\sigma.
\end{align}
Therefore, the critical points of $\biffun{\lambda}$ under normal variations are open subsets of $M$ such that its boundaries are CMC-hypersurfaces with $H=\lambda/(n-1)$.
\smallskip
\begin{remark}\label{hypopen}\label{hypboundos}(i) It is important to remark that, for these computations, there is no real dependence on the open set $\Omega$ but on the hypersurface $\partial\Omega$. In fact, in the literature, it is more common to work in terms of hypersurfaces. However, for simplicity on this case, we have preferred this approach of open sets.
(ii) Note also that this is also valid for a more general setting. Assume that ${\mathcal{M}}$ is the space of open subsets $\Omega\subset M$ whose boundary is union of two disjoint sets $\partial\Omega=\Sigma_1\cup\Sigma_2$. We will assume that one of them, $\Sigma_1$, is a fixed set and so that the hypersurface variations only affects $\Sigma_2$. Under this assumption, the critical points of the functional will be open subsets $\Omega$ such that their boundaries are union of a (fixed) set $\bord{1}$ and a CMC hypersurface $\bord{2}$.
\end{remark}
As we are interested in rigidity results, we need to study the second variation of the functional on the critical points or, more precisely, its spectrum. Assume that $\Omega$ is a critical point for the functional, that is, that the boundary $\partial \Omega$ is an $(n-1)$-dimensional hypersurface with constant mean curvature $H=\lambda/(n-1)$. The second variation of the functional on $\Omega$ has been already computed in \cite{BCE}, and it has the following form:
\begin{equation}\label{secondvariationsur}
\delta^2 \biffun{\lambda}_{\Omega}(f):=\Delta f - \left(||\mathrm I\!\mathrm I||^2+(n-1)\mathrm{Ric}(n_0,n_0)\right)f
\end{equation}
\noindent where $\Delta$ represent the non-negative Laplace-Beltrami operator and $||\mathrm I\!\mathrm I||^2$ the squared norm of the second fundamental form both in $\partial \Omega$ with the induced metric by $g$; $\mathrm{Ric}(n_0,n_0)$ is the normalized Ricci curvature of $M$ evaluated at $n_0$ the normal vector to $\Omega$ pointing outward and $f$ satisfies \eqref{volpre}.
\section{Bifurcation under cmc hypersurface variations}
\subsection{General bifurcation results}
Consider now an one-parameter family $\{\Omega_\gamma\}_\gamma$ of open subsets in $M$ such that the boundary of each $\Omega_\gamma$, denoted by $\partial \Omega_{\gamma}$, is a compact hypersurface with constant mean curvature $H(\gamma)$. In particular, each element $\Omega_{\gamma}$ is a critical point for the functional $\biffun{H(\gamma)(n-1)}$ where we can study its rigidity character. As our aim is to see if there is a change in the Morse index, we need to
study the eigenvalues of the linear map $\delta^2(\biffun{H(\gamma)(n-1)})_{\Omega_\gamma}(f)$.
At this point, in order to avoid confusion, let us introduce the following notation
\[
\lambda(\gamma):=H(\gamma)(n-1),\qquad {\mathcal A}_{\gamma}:=\delta^2(\biffun{H(\gamma)(n-1)})_{\Omega_\gamma}.
\]
Notice, if the potential function in \eqref{secondvariationsur} is constant then the eigenfunctions of such a map will coincide with the eigenfunctions of the Laplacian and, even more, if we denote such potential by
\begin{eqnarray}\label{potential}
Q(\gamma):=||\mathrm I\!\mathrm I_\gamma||^2+(n-1)\mathrm{Ric}(n_{\gamma},n_{\gamma})
\end{eqnarray}
and if $f$ is an eigenfunction of the Laplacian associated to an eigenvalue $\mu(\gamma)$, then $\overline{\mu}(\gamma)$ is an eigenvalue for the second variation of the functional, where
\[
\overline{\mu}(\gamma)=\mu(\gamma)-Q(\gamma)
\]
(here $\gamma$ represents its dependence respect $\Omega_{\gamma}$). Finally, by the spectral theorem \cite[Theorem A.I.4]{BGM} we know that all the eigenvalues of the (non-negative) Laplacian $\Delta$ on $\partial \Omega_{\gamma}$, are determined by a sequence of eigenvalues $\{\mu_i(\gamma)\}_{i\in \mathbb{N}_0}$ satisfying\footnote{The set $\mathbb{N}_0=\mathbb{N}\cup\left\{0\right\}=\left\{0,1,2,\cdots\right\}$.}:
\[\mu_0(\gamma)=0<\mu_1(\gamma)\leq\dots\leq\mu_{i}(\gamma)\leq \mu_{i+1}(\gamma)\leq\dots\quad \textrm{and}\quad \lim_{i}\mu_i(\gamma)=\infty,\] repeated according to their multiplicity.
\begin{remark}
Since the potential $Q(\gamma)$ is constant, if $f$ is an eigenfunction for $\Delta$ also it is for the second variation of the functional ${\mathcal A}_\gamma$ (see \eqref{secondvariationsur}). However, for our problem, an eigenfunction for the latter should additionally satisfy the integral condition \eqref{volpre}. In particular, the eigenfunctions associated to the eigenvalue zero of $\Delta$, which are non-zero constants, are not valid eigenfunctions for $\mathcal{A}_{\gamma}$, as they do not satisfy such an integral condition. In conclusion, the eigenvalues on (\ref{valfun}) have to be considered for $i\in\mathbb{N}$.
\end{remark}
In conclusion, assuming that the potential is constant, all the eigenvalues of ${\mathcal A}_{\gamma}$ have the following form:
\begin{align}\label{valfun}
\overline{\mu_i}(\gamma)=\mu_i(\gamma)-Q(\gamma),\quad\textrm{for every }i\in\mathbb{N}.
\end{align}
As we have determined the spectrum of the second derivative completely, we can characterize both, the rigidity and the existence of bifurcation points on the family $\{\Omega_{\gamma}\}_{\gamma}$. Let us start giving a simple sufficient condition for the rigidity:
\begin{theorem}\label{teo1}
Let $\{\Omega_{\gamma}\}_{\gamma\in I}$ be a family of open subsets on $M$ whose boundary $\partial \Omega_{\gamma}$ satisfies both: it is a hypersurface with constant mean curvature $H(\gamma)$ and its associated potential $Q(\gamma)$ is constant on $\partial \Omega_{\gamma}$. Then, such a family is rigid if for all $\gamma$,
\[\mu_1(\gamma)-Q(\gamma)>0\]
where $\mu_1(\gamma)$ represents the first nonzero eigenvalue of the Laplacian defined on $\partial \Omega_\gamma$.
\end{theorem}
{\it Proof: } As we have already observed, $\Omega_\gamma$ is a critical point for the functional $\biffun{\lambda(\gamma)}$. Moreover, we have also proved that all the eigenvalues of ${\mathcal A}_{\gamma}$ are determined by \eqref{valfun}. So, taking into account that, for all $i\in\mathbb{N}$
\[
\overline{\mu_i}(\gamma)=\mu_i(\gamma)-Q(\gamma)\geq \mu_1(\gamma)-Q(\gamma)>0,
\]
then the second derivative of the functional is non singular for all $\gamma$. Therefore, we obtain the desired result.
\ \rule{0.5em}{0.5em}
\smallskip
In particular, we obtain the following simple conditions to ensure rigidity:
\begin{corolario}\label{cor1}
Consider a family $\{\Omega_{\gamma}\}_{\gamma}$ of open subsets of $M$ whose boundary $\partial\Omega_\gamma$ is a CMC hypersurface. Such a family is rigid if one of the following conditions holds:
\begin{itemize}
\item[$i)$] $\partial \Omega_\gamma$ is convex and has non-positive Ricci curvature,
\item[$ii)$] $\mathrm{Ric}(n_0,n_0)\leq-1/(n-1)||\mathrm I\!\mathrm I_\gamma||^2$,
\item[$iii)$] $\partial \Omega_\gamma$ is Ricci flat and $\mu_1(\gamma)\geq||\mathrm I\!\mathrm I_\gamma||^2$, where $\mu_1(\gamma)$ is the first nonzero eigenvalue of $\Delta$ on $\partial \Omega_\gamma$.
\end{itemize}
\end{corolario}
\smallskip
Now, let us formalize the criterion that we will use to ensure the existence of bifurcation points (recall Section \ref{section2}).
\begin{theorem}\label{teo2}
Let $\{\Omega_{\gamma}\}_{\gamma\in I}$ be a family of open subsets on $M$ whose boundary $\partial \Omega_{\gamma}$ is a hypersurface with constant mean curvature $H(\gamma)$. If there exist two values $\gamma_1,\gamma_2\in I$ with $\gamma_1<\gamma_2$ satisfying:
\begin{itemize}
\item[(i)] $\overline{\mu_{i}}(\gamma_j)\neq 0$ for all $i$ and $j=1,2$;
\item[(ii)] there exists $i_0$ such that $\overline{\mu_{i_0}}(\gamma_1)\overline{\mu_{i_0}}(\gamma_2)<0$,
\end{itemize} then there exists a bifurcation point $\gamma_*\in (\gamma_1,\gamma_2)$.
\end{theorem}
{\it Proof: }
For this result we have to show both, that ${\mathcal A}_{\gamma}$ is non-singular for $j=1,2$; and that $i(\Omega_{\gamma_1},\biffun{\lambda(\gamma_1)})\neq i(\Omega_{\gamma_2},\biffun{\lambda(\gamma_2)})$. The former condition is a direct consequence of assumption (i), as all the eigenvalues of such a linear map are nonzero. For the latter, observe that hypothesis (ii) is ensuring that the eigenvalue $\overline{\mu_{i_0}}$ is changing the sign between $\gamma_1$ and $\gamma_2$. Moreover, as the eigenvalues $\overline{\mu_i}$ are ordered, we can ensure that the number of negative eigenvalues between $\gamma_1$ and $\gamma_2$ has changed, so $i(\Omega_{\gamma_1},\biffun{\lambda(\gamma_1)})\neq i(\Omega_{\gamma_2},\biffun{\lambda(\gamma_2)})$.
Then the result is obtained. \ \rule{0.5em}{0.5em}
\\
\smallskip
\subsection{CMC hypersurfaces bifurcation in warped product spaces}
\label{hybifwsection}
Let us assume now that $(M,g)$ is an $n$-dimensional warped product space, that is, $M=(r_1,r_2)\times P$ is a product manifold endowed with the following metric
\begin{equation}\label{eq1}
g=dr^2 +\alpha^2(r)g^P
\end{equation}
where $(P,g^P)$ is an $(n-1)$-dimensional closed Riemannian manifold (i.e. a compact manifold without boundary) and $\alpha$ is a smooth positive function on $(r_1,r_2)$.
These spaces naturally define a family of open subsets which can be realized as critical points of the area-volume functional for some special $\lambda$. Such a family, denoted by $\{\Omega_{r}\}_{r\in (r_1,r_2]}$, is formed by elements $\Omega_{r}=(r_1,r)\times P$ whose boundary of each $\Omega_r$ is composed by a fixed set $\bord{1}=\{r_1\}\times P$ and other set $\bord{2}=\{r\}\times P$.
So, recalling Remark \ref{hypboundos} (ii) we have that $\Omega_r$ is a critical point for the functional $\biffun{\lambda(r)}$ where $\Sigma_2$ is a hypersurface with constant mean curvature (with respect to the inward unit normal $-\partial r$) given by\footnote{By notation, the dot will denote derivative of a real function.} \[H(r)=-(n-1)\frac{\dot{\alpha}(r)}{\alpha(r)}.\]
Now our aim is to study at what extent such a family is rigid. The particular advantage of the warped product spaces is that we can explicitly determine both, $||\mathrm I\!\mathrm I||^2$ and $\mathrm{Ric}_{p_0}$ in terms of the warping function $\alpha$. In fact, by simple computations we obtain that
\[
||\mathrm I\!\mathrm I||^2=(n-1)\left(\frac{\dot{\alpha}}{\alpha}\right)^2\quad\textrm{and} \quad \mathrm{Ric}_{p_0}(\partial r, \partial r)=-\frac{\ddot{\alpha}\alpha}{\alpha^2}
\]
which shows that the potential $Q(r)$ is constant on the boundary of the open set and leads us to the following expression for the second variation (recall (\ref{secondvariationsur}) and \eqref{potential})
\begin{equation}\label{eqbif}\begin{array}{rl}
{\mathcal A}_{r}(f)= & \Delta f - (n-1)\left(\dfrac{\dot{\alpha}^2-\ddot{\alpha}\alpha}{\alpha^2}\right)f.
\end{array}\end{equation}
\smallskip
Moreover, the induced metric on $\bord{2}$ is just $g^P$ multiplied by the constant $\alpha^2(r)$, which allows us to relate easily the spectrum of $\Delta$ with the corresponding spectrum of the Laplacian on $(P,g^P)$. In fact, if $\hat{\mu}$ is an eigenvalue of the latter, $\mu=\hat{\mu}/\alpha^2(r)$ will be an eigenvalue for the former.
Therefore, as $P$ is a compact manifold, applying again the spectral theorem we deduce that the eigenvalues of the Laplacian on $(P,g^P)$ are determined by a sequence $\{\hat{\mu}_i\}_{i\in\mathbb{N}_0}$ satisfying:
\[\hat{\mu}_0=0<\hat{\mu}_1\leq\dots\leq\hat{\mu}_{i}\leq \hat{\mu}_{i+1}\leq\dots\quad \textrm{and}\quad \lim_{i}\hat{\mu}_i=\infty,\] repeated according to their
multiplicity.
Summarizing, we deduce that the eigenvalues of the linear map ${\mathcal A}_{r}$ have the form described on (\ref{valfun}) with
\begin{equation}\label{valfun1}\mu_i(r)=\frac{\hat{\mu}_i}{\alpha^2(r)}\quad \hbox{and}\quad Q(r)=\frac{(n-1)(\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r))}{\alpha^2(r)}
\end{equation}
\noindent where $\hat{\mu}_i$ are the eigenvalues of the Laplacian of $(P,g^P)$. \\
Then, we can just give simple conditions for both, rigidity and existence of bifurcation points in terms of the warping function $\alpha$ and the eigenvalues of the Laplacian on $(P,g^P)$. For rigidity, we will just give a simple translation of Theorem \ref{teo1} (the proof is trivial from such a theorem and \eqref{valfun1}):
\begin{proposition}\label{prop2.3}
Let $\left((r_1,r_2)\times P,dr^2+\alpha^2(r)g^P \right)$ be a warped product space where $P$ is an $(n-1)$-dimensional closed manifold and consider the one-parameter family $\{\Omega_r\}_r$ defined as above. If \[\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)<\frac{\hat{\mu}_1}{n-1}\] for all $r\in (r_1,r_2)$, then $\{\Omega_r\}_r$ has no bifurcation points.
\end{proposition}
\begin{example}
From here, we easily conclude that the pseudo hyperbolic space (for more details about this space, see \cite{T}) \[\left
(r_1,r_2)\times\mathbb{S}^{n-1}, dr^2+\,e^{2r}\,g^{\mathbb{S}^{n-1}}\!\right) \]
is foliated by a locally rigid family of CMC hypersurfaces (recall Remark \ref{hypopen} (i)). The slices $\{r\}\times\mathbb{S}^{n-1}$ determine an one-parameter family of CMC hypersurfaces without bifurcation points
because $\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)=0$ and $\hat{\mu}_1=n-1$ (see \cite[Proposition C.I.1]{BGM}).\\
\end{example}
\smallskip
Next, our objective is to give conditions for $\alpha$ which ensure the existence of bifurcation points of the family $\{\Omega_{r}\}_{r\in(r_1,r_2)}$. Fix a value $r_0\in (r_1,r_2)$ and observe that, even if we do not know the sign on $\overline{\mu_i}(r_0)$, we can ensure that there exists $i_0\in \mathbb{N}$ such that for all $i\geq i_0$, $\overline{\mu_i}(r_0)>0$ (recall that $\hat{\mu}_i\rightarrow \infty$, \eqref{valfun} and \eqref{valfun1}). So, we only need to impose conditions which ensure that, for some $r>r_0$, $\overline{\mu_{i_0}}(r)<0$. This idea is exploited to obtain the following result.
\begin{proposition}\label{prop5.1}
Let $\left((r_1,r_2)\times P,dr^2+\alpha^2(r)g^P \right)$ be a warped product space where $P$ is an $(n-1)$-dimensional closed manifold and consider the one-parameter family of open subsets $\Omega_{r}=(r_1,r)\times P$ whose boundary is composed by a fix set $\bord{1}=\{r_1\}\times P$ and a CMC hypersurface $\bord{2}=\{r\}\times P$ (i.e., critical points for the functional $\biffun{\lambda(r)}$).
If
\begin{equation}\label{cond1}
\lim_{r\rightarrow r_2}\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)=\infty
\end{equation}
then, there exist infinitely many bifurcation points.
\end{proposition}
{\it Proof: }
Let us denote by \[h(r):=(n-1)\left(\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)\right)\] so that we can write the eigenvalues $\overline{\mu_{i}}(r)$ in the following way (recall \eqref{valfun} and \eqref{valfun1}):
\[
\overline{\mu_i}(r)=\frac{1}{\alpha^2(r)}\left(\hat{\mu}_i-h(r)\right).
\]
Fix $r_0\in (r_1,r_2]$ and consider $i_0\in \mathbb{N}$ such that $\overline{\mu_{i_0}}(r_0)>0$, that is, with
$ \hat{\mu}_{i_0}>h(r_0)$
(which is possible because $\lim_{i}\hat{\mu}_{i}=\infty$). As $h$ tends to infinity when $r\rightarrow r_2$ (recall (\ref{cond1})), we can find some value $\widetilde{r}>r_0$ such that
$\hat{\mu}_{i_0}<h(\widetilde{r})$,
and so, with $\overline{\mu_{i_0}}(\widetilde{r})<0$. Now, as $h$ is a continuous function and satisfies
\[h(r_0)<\hat{\mu}_{i_0}<h(\widetilde{r})
\]
we can obtain two values $r'_0,r'_1$ with $r_0\leq r'_0<r'_1\leq \widetilde{r}$ such that
\[
\hat{\mu}_{i_{-1}}<h(r'_0)<\hat{\mu}_{i_0}<h(r'_1)<\hat{\mu}_{i_1}
\] where $\hat{\mu}_{i_{-1}}$ (resp. $\hat{\mu}_{i_{1}}$) denotes the largest (resp., smallest) eigenvalue of the Laplacian $\Delta$ less (resp., greater) than $\hat{\mu}_i$. Then we deduce both, that $\overline{\mu_{i_0}}(r'_0)\overline{\mu_{i_0}}(r'_1)<0$ and that $\overline{\mu_{i}}(r'_j)\neq 0$ for all $i\in\mathbb{N}_0$ and $j=1,2$. Therefore, Theorem \ref{teo2} ensures the existence of a bifurcation point in the interval $(r_0,\widetilde{r})$. Now, we can repeat the process with $\widetilde{r}$, obtaining another bifurcation point in the interval $(\widetilde{r},r_2)$ for some $r_2>\widetilde{r}$. In conclusion, by induction, we prove the existence of infinitely many points of bifurcation.
\ \rule{0.5em}{0.5em}
\begin{remark}
As we can see in previous proof, a point of bifurcation always appears when the function $h(r)=(n-1)\left(\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)\right)$ cross the value of an eigenvalue of the Laplacian $\Delta$ (see Figure \ref{fig1}), obtaining a weaker condition than (\ref{cond1}) to ensure the existence of bifurcation points.
\end{remark}
\begin{figure}
\centering
\ifpdf
\setlength{\unitlength}{1bp}%
\begin{picture}(235.19, 189.06)(0,0)
\put(0,0){\includegraphics{dibujo1.pdf}}
\put(30.88,57.36){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i}$}
\put(24.38,87.88){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i+1}$}
\put(24.38,130.22){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i+2}$}
\put(24.38,175.02){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i+3}$}
\put(205.51,33.73){\fontsize{8.54}{10.24}\selectfont $r$}
\put(51.95,34.22){\fontsize{8.54}{10.24}\selectfont $r_1$}
\put(212.51,153.73){\fontsize{11.38}{13.66}\selectfont $h$}
\put(81.82,34.76){\fontsize{5.69}{6.83}\selectfont $r_{A}$}
\put(118.06,34.76){\fontsize{5.69}{6.83}\selectfont $r_{B}$}
\put(133.59,34.76){\fontsize{5.69}{6.83}\selectfont $r_{C}$}
\put(162.30,34.76){\fontsize{5.69}{6.83}\selectfont $r_{D}$}
\end{picture}%
\else
\setlength{\unitlength}{1bp}%
\begin{picture}(235.19, 189.06)(0,0)
\put(0,0){\includegraphics{dibujo1}}
\put(30.88,57.36){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i}$}
\put(24.38,87.88){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i+1}$}
\put(24.38,130.22){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i+2}$}
\put(24.38,175.02){\fontsize{8.54}{10.24}\selectfont $\hat{\mu}_{i+3}$}
\put(205.51,33.73){\fontsize{8.54}{10.24}\selectfont $r$}
\put(51.95,34.22){\fontsize{8.54}{10.24}\selectfont $r_1$}
\put(212.51,153.73){\fontsize{11.38}{13.66}\selectfont $h$}
\put(81.82,34.76){\fontsize{5.69}{6.83}\selectfont $r_{A}$}
\put(118.06,34.76){\fontsize{5.69}{6.83}\selectfont $r_{B}$}
\put(133.59,34.76){\fontsize{5.69}{6.83}\selectfont $r_{C}$}
\put(162.30,34.76){\fontsize{5.69}{6.83}\selectfont $r_{D}$}
\end{picture}%
\fi
\caption{\label{fig1}The function $h$ let us deduce where we can find bifurcation points. In this figure, $r_{A}$ is a bifurcation point, and there is another bifurcation point on the interval $(r_{B},r_{C})$. However, on $r_{D}$ we cannot ensure the existence of a bifurcation point, as $h$ does not cross the eigenvalue $\hat{\mu}_{i+2}$.}
\end{figure}
\begin{corolario}
If $r_2=\infty$ and $\alpha(r)=Cr^k + O(r^{k-1})$ with $k>1$ and $C\in\mathbb{R}\slash\left\{0\right\}$, then there are infinitely many bifurcation points.
\end{corolario}
{\it Proof: } This is just a direct computation showing that \[\lim_{r\rightarrow \infty}\left( \dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)\right)=\lim_{r\rightarrow \infty}\left(kC^2r^{2(k-1)}+O(r^{2k-3})\right)=\infty.\]
\ \rule{0.5em}{0.5em}\\
\smallskip
Finally, we devote the rest of this section to give a more geometrical viewpoint for previous results. For this, let us recall the formulae for the mean and scalar curvatures on warped products. The mean curvature of $\bord{2}$ is given by
\begin{align}\label{Hwarped}
H(r)=-(n-1)\frac{\dot{\alpha}(r)}{\alpha(r)}
\end{align}
while the scalar curvature in a point $(r,x)\in (r_1,r_2)\times P$ is given by
\begin{align*}
R{(r,x)}=\frac{1}{\alpha^2(r)}\left(R(x)-(n-1)\left((n-2)\dot{\alpha}^2(r) + 2\ddot{\alpha}(r)\alpha(r)\right)\right)
\end{align*}
where $R(x)$ denote the scalar curvature of the manifold $P$.
Taking into account the expression of $H(r)$, we are able to rewrite previous expression in the following way:
\begin{align*}
\begin{array}{rl} R(r,x)= & \dfrac{R(x)}{\alpha^2(r)}+2(n-1)\dfrac{\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)}{\alpha^2(r)}-n(n-1)\dfrac{\dot{\alpha}^2(r)}{\alpha^2(r)}\\\\ =&
\dfrac{R(x)}{\alpha^2(r)}+2(n-1)\dfrac{\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)}{\alpha^2(r)} - \dfrac{n}{n-1}H(r)^2.
\end{array}
\end{align*}
In particular,
\[
\frac{\dot{\alpha}^2(r)-\ddot{\alpha}(r)\alpha(r)}{\alpha^2(r)}=\frac{1}{2(n-1)}\left(R(r,x)-\frac{R(x)}{\alpha^2(r)}+\frac{n}{n-1}H(r)^2 \right)
\]
and so, substituting such expression in (\ref{eqbif}), we obtain the following
\[
{\mathcal A}_{r}(f)=\Delta f -\frac{1}{2}\left(R(r,x)-\frac{R(x)}{\alpha^2(r)}+\frac{n}{n-1}H(r)^2\right) f.
\]
With this expression, we are also able to rewrite the eigenvalues of the operator ${\mathcal A}_{r}$ (compare with \eqref{valfun} and \eqref{valfun1})
\[
\overline{\mu_{i}}(r)=\frac{1}{\alpha^2(r)}\left(\hat{\mu}_{i}+\frac{R(x)}{2}-\frac{\alpha^2(r)}{2}\left(R(r,x) + \frac{n}{n-1} H(r)^2 \right) \right), \;\textrm{ with }i\in\mathbb{N}.
\]
Using these expressions, we can translate Propositions \ref{prop2.3} and \ref{prop5.1} in more geometrical terms. In particular, we deduce the following two corollaries.
\begin{corolario}\label{corsc}
Assume that $R(r,x), R(x) $ and $H(r)$ satisfy
\[
\alpha^2(r)\left(R(r,x) + \frac{n}{n-1}H(r)^2\right)<2\hat{\mu}_1+R(x)
\]
for all $r$ and all $x\in P$. Then, there is no bifurcation point associated to our variational problem.
\end{corolario}
\begin{example}
We consider the following warped product \[\left((0,r_2)\times \mathbb{S}^{n-1}, \,\!dr^2+\sinh^2r\,g^{\mathbb{S}^{n-1}}\!\right).\]
Fixing the set $\Sigma_1=\left\{0\right\}\times\mathbb{S}^{n-1}$ it is possible to visualize the CMC hypersurfaces $\left\{r\right\}\times\mathbb{S}^{n-1}$ as the other set $\Sigma_2$ forming the boundary of the elements of an one-parameter family $\left\{\Omega_r\right\}_r$ in this warped product. By previous corollary and the fact that $R(r,x)=-n(n-1)$, $R(x)=(n-1)(n-2)$, $H(r)^2=(n-1)^2\coth^2 r$ and $\hat{\mu}_1=n-1$ we conclude that this family is rigid.
\end{example}
\begin{example}
Another interesting example occurs in the spatial fiber of the Sitter cusp spacetime (see \cite{H} for more details about this spacetime), that is, the following warped product \[\left(
(r_1,r_2)\times \mathbb{T}^{n-1},\,dr^2+e^{2r}\,g^{\mathbb{T}^{n-1}}\!\right),\]
where $ \mathbb{T}^{n-1}$ is the $(n-1)$-dimensional flat torus.
Note that if we fix a set $\Sigma_1=\left\{r_1\right\}\times\mathbb{T}^{n-1}$ for some $r_1\in\mathbb{R}$ and if we consider the CMC hypersurfaces $\left\{r\right\}\times\mathbb{T}^{n-1}$ as the set $\bord{2}$ constituting together with $\Sigma_1$ the boundary of the elements of an one-parameter family $\left\{\Omega_r\right\}_r$ we have that $R(r,x)\!=-n(n-1)$, $R(x)=0$, $H(r)^2=(n-1)^2$ and as $\mathbb{T}^{n-1}$ is a compact manifold is clear that $\hat{\mu}_1>0$. Moreover, from \cite[Theorem B.I.2]{BGM} we have that its spectrum is given by $\left\{\hat{\mu}=4\pi^2\left|y\right|^2\right\}$ where $y$ is the closest element to the unity in $\Gamma^*$ (the dual lattice of $\Gamma$ that define the torus in $\mathbb{R}^{n-1}$). Thus, using Corollary \ref{corsc} we conclude that this family has no bifurcation point.
\end{example}
\begin{corolario}
If
\[
\lim_{r\rightarrow r_2}\alpha^2(r)\left(R(r,x) +\frac{n}{n-1}H(r)^2\right)=\infty,
\]
then there exist infinitely many bifurcation points for the family $\left\{\Omega_{r}\right\}_r$ associated to the functional $\biffun{\lambda(r)}$ and the warped product space $(r_1,r_2)\times P$ where $P$ is a $(n-1)$-dimensional closed manifold.
\end{corolario}
\smallskip
\section{\textbf{An application on the de Sitter and Anti-de Sitter Schwarzschild models.} }\label{App}
In this section, we will apply previous results to the following Schwarzschild models: consider a $3$-dimensional Riemannian manifold $M= I\times \mathbb{S}^2$ endowed with the metric
\begin{align}\label{metricschwgen}
g_{K,E}=\psi_{K,E}(r)^{-2}dr^2+r^2\left(d\theta^2+sin^2\theta d\phi^2 \right),
\end{align}
where $K\neq0$ and $E$ are constants, $\psi_{K,E}(r)=\sqrt{1-2K/r+E r^2}$ and
$I$ is a maximal connected and open interval where $\psi_{K,E}$ is well defined. For instance, for $K>0$ and $E\geq 0$, $I=(\widehat{r},\infty)$ (with $\widehat{r}>0$) while for $K>0$ and $E<0$ the interval $I=(\widehat{r},0)$ where, in both cases, $\widehat{r}$ is a zero of the function $\psi_{K,E}$.
The two parameters $K$ and $E$ represent \textit{the black hole mass} (and so, it is usually considered positive) and \textit{the cosmological constant} respectively. When $E=0$, we obtain the classical model of the spatial fiber of the \textit{Schwarzschild spacetime}. For $E$ negative (resp. positive) we have the Riemannian model of the so-called de \textit{Sitter} (resp. \textit{Anti-de Sitter}) \textit{Schwarzschild spacetime}.
\smallskip
It follows directly that the variable change
\begin{equation}\label{varchan}ds=-\psi_{K,E}(r)^{-1}dr
\end{equation}
transform the metric $g_{K,E}$ into the warped metric
\begin{equation}\label{varchanmet}
g_{K,E}=ds^2+r(s)^2\left(d\theta^2+sin^2\theta \,d\phi^2 \right)
\end{equation}
\noindent making the results of previous section applicable. In this case, $\alpha(s)=r(s)$ and the CMC surfaces $\left\{s\right\}\times \mathbb{S}^2$ are $2$-spheres of radius $r(s)$, which can be seen as the boundaries of open subsets $\Omega_s$ that belong to a one-parameter family of $(M^{3},g_{K,E})$. Moreover, the eigenvalues for the $2$-spheres are well known (recall \cite[Proposition C.I.1]{BGM}), and satisfy that $\hat{\mu}_i=i(i+1)$ with $i\in \mathbb{N}_0$.
\smallskip
Let us consider two relevant cases (the rest will follow analogously). Consider first that both parameters $K$ and $E$ are positive (i.e., we consider a Anti-de Sitter Schwarzschild model) and take the interval $I=(\widehat{r},\infty)$ where $\widehat{r}$ is the only positive zero of $\psi_{K,E}$. Although the expression of $g_{K,E}$ becomes singular on $\widehat{r}$, it can be proved that the metric $g_{K,E}$ extends to a smooth Riemannian metric on $M=[\widehat{r},\infty)\times \mathbb{S}^2$. Finally, by the variable change on \eqref{varchan}, $M=[\widehat{s},\infty)\times \mathbb{S}^2$ and the metric $g_{K,E}$ takes the form on \eqref{varchanmet}. Then, we can state the following result
\begin{corolario}
Consider the spatial fiber of Anti-de Sitter Schwarzschild spacetime $(M^3,g_{K,E})$ (with $K,E>0$) as a warped product space. The family $\left\{\Omega_s\right\}_{s\in (\widehat{s},\infty)}$ where $\Omega_s=(\widehat{s},s)\times \mathbb{S}^2$ is locally rigid on all $s>\widehat{s}$.
\end{corolario}
{\it Proof: }
In order to apply Proposition \ref{prop2.3}, let us estimate the following:
\begin{align}\label{rl}
\dot{r}^2(s)-\ddot{r}(s)r(s) =1-\dfrac{3K}{r(s)}
\end{align}
where we have used that $\dot{r}(s)=-\sqrt{1-\dfrac{2K}{r(s)}+E r(s)^2}$.\\\\
Now, bearing in mind that $\hat{\mu}_{1}=2$ and $n=3$, we obtain that
\[
\dot{r}^2(s)-\ddot{r}(s)r(s)=1-\frac{3K}{r(s)}
<\frac{\hat{\mu}_1}{n-1}=1
\]
for $r(s)>\widehat{r}>0$ and $K>0$. So, Proposition \ref{prop2.3} applies showing that there are no bifurcation points. \ \rule{0.5em}{0.5em}
\\
\smallskip
Another way to argue the rigidity of this family $\left\{\Omega_s\right\}_s$ is to use Corollary \ref{corsc} and the fact $R(r,x)=-6E$, $H(r)^2=4/r^2\psi_{K,E}(r)^2=4/r^2-8K/r^3+4E$ and $R(x)=2$.\\
Now, let us consider the case when the cosmological constant $E$ is negative, and let us take $I=[\widehat{r},0)$ (again, the Riemannian metric extends smoothly). Take the variable change \eqref{varchan} in such a way that $M=[\widehat{s},0)\times \mathbb{S}^2$. In this case, the family $\{\Omega_{s}\}_{s\in(\widehat{s},0)}$ is no longer locally rigid, as we can see on the following result
\begin{corolario}
Consider the spatial fiber of de Sitter Schwarzschild spacetime $(M^3,g_{K,E})$ (i.e., with $E<0$ and $K>0$) as a warped product space. Then the family $\left\{\Omega_s\right\}_{s\in (\widehat{s},s)}$ where $\Omega_s=(\widehat{s},s)\times \mathbb{S}^2$ has infinitely many bifurcation points.
\end{corolario}
{\it Proof: }
Observe that
\begin{equation}
\lim_{s\rightarrow 0^-}\dot{r}^2(s)-\ddot{r}(s)r(s)=\lim_{r\rightarrow 0^-}1-\frac{3K}{r}=\infty.
\end{equation}
Then, by Proposition \ref{prop5.1} we conclude that the family $\left\{\Omega_s\right\}_s$ has infinitely many bifurcation points.
\ \rule{0.5em}{0.5em}
\bibliographystyle{amsplain}
|
\section*{Result}
\subsection*{Common features of CSQBC}
By checking the existing CSQBC protocols \cit
{qi150,qbc50,qbc78,qi197,qbc52,qbc89}, we find that they all share the
following common features (note that the names Alice and Bob are used
reversely in \cite{qbc78,qbc52,qbc89}):
(1) During the holding phase, the receiver Bob owns a quantum system $\Psi
\ encoding Alice's committed bit $b$. ($\Psi$\ can either be prepared by the
sender Alice, or be prepared by Bob and sent to Alice, who returns it to Bob
after performing some certain operations according to her choice of $b$. It
also does not matter whether Alice prepared and kept another quantum system
entangling with $\Psi$.)
(2) Bob knows the definitions of $\rho_{0}^{B}$\ and $\rho_{1}^{B}$ directly
before the end of the commit phase. (That is, these definitions are either
clearly stated by the protocol, or announced to Bob\ by Alice classically.
Bob does not need to perform operations on any quantum system to gain
knowledge of these definitions.) Here $\rho_{0}^{B}$\ and $\rho_{1}^{B}$\
are the density matrices of Bob's $\Psi$ corresponding to $b=0$ and $b=1$,
respectively.
(3) To detect Bob's cheating, at the unveil phase Alice can check whether
the state of $\Psi$\ is intact. (It does not matter whether the entire $\Psi
\ or only a small part can be checked.)
(4) To detect Alice's cheating, at the unveil phase Bob can learn a
nontrivial amount of information on the value of $b$ from $\Psi$, even
without any help from Alice.
The last feature indicates that there exists at least one operation known to
Bob, which can output a bit $b^{\prime }$ when being applied on $\Psi $, and
$b^{\prime }=b$ should occur with a probability larger than $1/2$. As a
result, there must be $\rho _{0}^{B}\neq \rho _{1}^{B}$. This is a main
difference from the original QBC, where there is generally $\rho
_{0}^{B}\simeq \rho _{1}^{B}$ so that it can be unconditionally secure
against dishonest-Bob.
The original purpose of CSQBC having these features is as follows. Alice's
cheating strategy suggested in the MLC no-go theorem is based on the
Hughston-Jozsa-Wootters (HJW) theorem \cite{qi73}, which applies to the case
$\rho _{0}^{B}\simeq \rho _{1}^{B}$. Therefore with feature (4), i.e., $\rho
_{0}^{B}\neq \rho _{1}^{B}$, Alice's cheating becomes detectable so that\
the MLC no-go theorem can be evaded. On the other hand, if Bob takes
advantages of $\rho _{0}^{B}\neq \rho _{1}^{B}$\ and performs measurements
to discriminate the committed bit $b$, the quantum state will be disturbed.
In this case, with feature (3) Bob's cheating will be detected with a
certain probability when Alice asks\ him to return the quantum state and
checks wether it remains undisturbed, so that the goal of CSQBC can be met.
But with a rigorous quantitative analysis on the probability of detecting
Bob's cheating, we will find that it is always not sufficiently large when
Bob applies some specific measurements. Therefore any CSQBC protocol having
the above four features will be bounded by the security limit below.
\subsection*{Notations and Bob's cheating strategy}
According to Eq. (9.22) of \cite{qi366}, the trace distance $D(\rho
_{0}^{B},\rho _{1}^{B})\equiv tr\left\vert \rho _{0}^{B}-\rho
_{1}^{B}\right\vert /2$\ (where $\left\vert A\right\vert \equiv \sqrt
A^{\dagger }A}$) between $\rho _{0}^{B}$ and $\rho _{1}^{B}$\ satisfie
\begin{equation}
D(\rho _{0}^{B},\rho _{1}^{B})=\max_{P}tr(P(\rho _{0}^{B}-\rho _{1}^{B})),
\end{equation
where the maximization is taken over all positive operators $P\leq I$, with
I$ being the identity operator. The above feature (2) of CSQBC guarantees
that Bob knows how $\rho _{0}^{B}$\ and $\rho _{1}^{B}$ are defined. Thus he
can find the positive projectors $P=P_{m}$\ that maximizes $tr(P(\rho
_{0}^{B}-\rho _{1}^{B}))$. If $\rho _{0}^{B}$\ stands a higher probability
to be projected successfully than $\rho _{1}^{B}$\ when applying $P_{m}$,
then we takes $P_{0}\equiv P_{m}$\ and $P_{1}\equiv I-P_{m}$. Otherwise we
takes $P_{0}\equiv I-P_{m}$\ and $P_{1}\equiv P_{m}$. Feature (1) ensures
that Bob owns the system $\Psi $\ encoding Alice's committed bit $b$ during
the holding phase. Therefore, by applying the positive operator-valued
measure (POVM) $\{P_{0}^{\dag }P_{0},P_{1}^{\dag }P_{1}\}$\ on $\Psi $, Bob
can discriminate between $\rho _{0}^{B}$\ and $\rho _{1}^{B}$ and learn
Alice's committed $b$ with the maximal probability allowed by $D(\rho
_{0}^{B},\rho _{1}^{B})$.
To analyze rigorously the probability for Bob to escape Alice's detection
with this POVM, let $H$\ be the global Hilbert space constructed by all
possible states of $\Psi $ (either $b=0$ or $1$). Since $P_{0}$, $P_{1}$\
are positive projectors, there exists an orthonormal basis $\{\left\vert
e_{i}\right\rangle \}$ of $H$ (the following proof remains valid regardless
whether $\{\left\vert e_{i}\right\rangle \}$ is known to Alice or Bob), in
which $P_{0}$, $P_{1}$ can be expressed a
\begin{eqnarray}
P_{0} &=&\sum\limits_{i}\left\vert e_{i}^{(0)}\right\rangle \left\langle
e_{i}^{(0)}\right\vert , \nonumber \\
P_{1} &=&\sum\limits_{i}\left\vert e_{i}^{(1)}\right\rangle \left\langle
e_{i}^{(1)}\right\vert , \label{projector2}
\end{eqnarray
where $\{\left\vert e_{i}^{(0)}\right\rangle \}\cup \{\left\vert
e_{i}^{(1)}\right\rangle \}=\{\left\vert e_{i}\right\rangle \}$.
Meanwhile, before Bob applying any measurement, the general form of the
initial state of $\Psi $\ can always be written a
\begin{align}
\left\vert \Phi \otimes \Psi \right\rangle _{ini}& =\sqrt{\alpha
\sum\limits_{i}\lambda _{i}^{(0)}\left\vert f_{i}^{(0)}\right\rangle \otimes
\left\vert e_{i}^{(0)}\right\rangle \nonumber \\
& +\sqrt{\beta }\sum\limits_{i}\lambda _{i}^{(1)}\left\vert
f_{i}^{(1)}\right\rangle \otimes \left\vert e_{i}^{(1)}\right\rangle ,
\label{initial}
\end{align
where $0\leq \alpha \leq 1$, $\beta =1-\alpha $, and $\sum\limits_{i}\lef
\vert \lambda _{i}^{(0)}\right\vert ^{2}=\sum\limits_{i}\left\vert \lambda
_{i}^{(1)}\right\vert ^{2}=1$\ (sum over all possible $i$ within each
corresponding subspace). The values of $\alpha $, $\beta $, $\lambda
_{i}^{(0)}$'s and $\lambda _{i}^{(1)}$'s\ are chosen by Alice according to
the value of her committed bit $b$.\ Here $\Phi $ is a quantum system that
Alice may introduce and keep to herself, which entangles with Bob's $\Psi $.
All $\left\vert f_{i}^{(0)}\right\rangle $'s and $\left\vert
f_{i}^{(1)}\right\rangle $'s are the vectors describing the state of $\Phi
, which are not required to be orthogonal to each other. In the case where
Alice does not introduce such a system, we can simply set all $\left\vert
f_{i}^{(0)}\right\rangle $'s and $\left\vert f_{i}^{(1)}\right\rangle $'s to
be equal, so that Eq. (\ref{initial}) still applies.
\subsection*{The security bound on Bob's cheating}
As elaborated in the 1st subsection of Methods section, when dishonest-Bob
applies the above POVM $\{P_{0}^{\dag }P_{0},P_{1}^{\dag }P_{1}\}$ on $\Psi
, we find that the probability for Bob's cheating to pass Alice's detection
successfully i
\begin{equation}
P_{B}=\frac{1}{2}+\frac{1}{2}(2\alpha -1)^{2}, \label{pc}
\end{equation
and the amount of mutual information he obtained i
\begin{equation}
I_{m}=1-h(\alpha ). \label{Im}
\end{equation
Here $h(\alpha )\equiv -\alpha \log _{2}\alpha -(1-\alpha )\log
_{2}(1-\alpha )$\ is the binary entropy function.
With Eqs. (\ref{pc}) and (\ref{Im}), we plot $P_{B}$ and $I_{m}$\ as a
function of $\alpha $\ in FIG. 1. Since $0\leq \alpha \leq 1$, FIG. 1 and
Eq. (\ref{pc}) both give
\begin{equation}
P_{B}\geq 50\%. \label{bound}
\end{equation
The minimum $P_{B}=50\%$\ can be reached when Alice chooses $\alpha =0.5$.
Thus we come to the conclusion that Bob can always learn Alice's committed
b $ with the maximal probability allowed by the trace distance between $\rho
_{0}^{B}$ and $\rho _{1}^{B}$, while his cheating stands at least $50\%$
chance to escape Alice's detection.
\begin{figure}[tbp]
\includegraphics{fig1}
\caption{Bob's successful cheating probability $P_{B}$ (red line) and mutual
information $I_{m}$ (blue line) on Alice's committed bit $b$ as a function
of $\protect\alpha $ that Alice chooses for the initial state Eq. (3). The
dash lines mark the values for the protocol in Ref. [5].}
\label{fig:epsart}
\end{figure}
It may look weird that FIG. 1 seems to indicate that the more amount of
information that Bob obtains, the easier he can pass Alice's detection. But
we must note that the amount of Bob's information is not chosen by himself.
Instead, it is determined by the value of $\alpha $ that Alice chooses. That
is, once Alice determines which state is used for encoding her committed
bit, the maximum amount of information that Bob can obtain is also fixed.
On the other hand, the above result indicates that Alice should make $\alpha
$\ as close to $0.5$\ as possible, so that Bob's information and successful
cheating probability can be minimized. However, note that she has to choose
the initial state Eq. (\ref{initial}) within the range restricted by the
protocol. Due to the feature (4) of CSQBC, the trace distance $D(\rho
_{0}^{B},\rho _{1}^{B})$\ has to be nonzero, Therefore, generally $\alpha $\
cannot be made very close to $0.5$, as we will see in the examples below.
\subsection*{Examples}
In the CSQBC protocol in \cite{qi150}, Bob's system $\Psi $\ is a single
qubit, whose state is either $\left\vert 0\right\rangle $\ or $\left\vert
-\right\rangle $\ ($\left\vert 1\right\rangle $\ or $\left\vert
+\right\rangle $) when Alice commits $b=0$ ($b=1$). Here $\left\vert
0\right\rangle $\ and $\left\vert 1\right\rangle $\ are orthogonal to each
other, $\left\vert \pm \right\rangle \equiv (\left\vert 0\right\rangle \pm
\left\vert 1\right\rangle )/\sqrt{2}$. So we have $\rho _{0}^{B}=(\left\vert
0\right\rangle \left\langle 0\right\vert +$\ $\left\vert -\right\rangle
\left\langle -\right\vert )/2$ and $\rho _{1}^{B}=(\left\vert 1\right\rangle
\left\langle 1\right\vert +$\ $\left\vert +\right\rangle \left\langle
+\right\vert )/2$. Defin
\begin{eqnarray}
\left\vert e^{(0)}\right\rangle &\equiv &\cos (-\pi /8)\left\vert
0\right\rangle +\sin (-\pi /8)\left\vert 1\right\rangle , \nonumber \\
\left\vert e^{(1)}\right\rangle &\equiv &\cos (3\pi /8)\left\vert
0\right\rangle +\sin (3\pi /8)\left\vert 1\right\rangle .
\end{eqnarray
Then Bob's operation for maximally discriminating $\rho _{0}^{B}$\ and $\rho
_{1}^{B}$\ is to measure $\Psi $\ in the basis $\{\left\vert
e^{(0)}\right\rangle ,\left\vert e^{(1)}\right\rangle \}$, i.e., he applies
the projector $P_{0}=\left\vert e^{(0)}\right\rangle \left\langle
e^{(0)}\right\vert $. When the projection is successful (unsuccessful), he
takes $b^{\prime }=0$ ($b^{\prime }=1$) as the decoded result. With this
method, $b^{\prime }$\ will match Alice's actual committed bit $b$ with the
probability $\cos ^{2}(\pi /8)\simeq 85.36\%$. Meanwhile, Alice's four input
states can be expanded in the $\{\left\vert e^{(0)}\right\rangle ,\left\vert
e^{(1)}\right\rangle \}$\ basis a
\begin{align}
\left\vert 0\right\rangle & =\cos (\pi /8)\left\vert e^{(0)}\right\rangle
+\sin (\pi /8)\left\vert e^{(1)}\right\rangle , \nonumber \\
\left\vert -\right\rangle & =\cos (\pi /8)\left\vert e^{(0)}\right\rangle
-\sin (\pi /8)\left\vert e^{(1)}\right\rangle , \nonumber \\
\left\vert 1\right\rangle & =-\sin (\pi /8)\left\vert e^{(0)}\right\rangle
+\cos (\pi /8)\left\vert e^{(1)}\right\rangle , \nonumber \\
\left\vert +\right\rangle & =\sin (\pi /8)\left\vert e^{(0)}\right\rangle
+\cos (\pi /8)\left\vert e^{(1)}\right\rangle .
\end{align
Comparing with Eq. (\ref{initial}), we can see that there is either $\alpha
=\cos ^{2}(\pi /8)$\ or $\alpha =\sin ^{2}(\pi /8)$. Substitute them into
Eq. (\ref{pc}) will both yield $P_{B}=\sin ^{4}(\pi /8)+\cos ^{4}(\pi
/8)=75\%$. That is, in the CSQBC protocol in \cite{qi150}, Bob can learn
Alice's committed bit with reliability $85.36\%$\ (i.e., his mutual
information is $1-h(0.8536)\simeq 0.4$ bit) before the unveil phase, while
he can pass Alice's security check with probability $75\%$. This protocol is
corresponding to the dash lines in our FIG. 1.
Another example can be found in \cite{HeComment}, where we illustrated how
our above cheating strategy applies on the CSQBC protocol in \cite{qbc52}.
This protocol looks more complicated than the one in \cite{qi150}, as the
committed bit $b$ is encoded with many qubits, instead of a single one. The
authors of \cite{qbc52} merely analyzed the individual attack of the
receiver (note that they used the names Alice and Bob reversely) where the
qubits are measured one by one. Then it is concluded that the cheating can
be detected with a probability arbitrarily close to $1$. But as we shown
above, instead of individual measurements, the dishonest receiver can apply
a two-element POVM $\{P_{0}^{\dag }P_{0},P_{1}^{\dag }P_{1}\}$ on the entire
state encoding the committed bit. When this state consists of many qubits,
each basis vector $\left\vert e_{i}\right\rangle $ of the Hilbert space $H$
is a multi-level state describing all qubits. Thus the projectors $P_{0}$,
P_{1}$ in Eq. (\ref{projector2}) are actually collective measurements. The
detailed form of $P_{0}$, $P_{1}$ is given in Eq. (2) of \cite{HeComment}.
As a result, it was further elaborated there that this collective
measurement is as effective as individual measurements on learning the
committed bit, while it causes much less disturbance on the multi-qubit
state. Once again, the probability for the cheater to escape the detection
was shown \cite{HeComment} to be not less than $50\%$. With the increase of
the qubit number $n$, this probability can even be arbitrarily close to
100\%$.
\subsection*{Alice's cheating strategy}
Alice's cheating strategy used in the MLC no-go theorem requires the
condition $\rho _{0}^{B}\simeq \rho _{1}^{B}$, which no longer holds in
CSQBC. Nevertheless, she can still apply the same strategy in CSQBC and try
her luck. To give a detailed description of the strategy, first let us model
the coding method in CSQBC more precisely. For generality, consider that in
the protocol, besides Bob's system $\Psi $, there is another system $E$.
Alice's different committed values of $b$ is encoded with different states
of the combined system $E\otimes \Psi $. System $E$ is kept at Alice's side
during the commit and holding phases, and is required to be sent to Bob at
the unveil phase to justify Alice's commitment. Let $\rho _{0}^{EB}$\ and
$\rho _{1}^{EB}$\ denote the density matrices of $E\otimes \Psi $
corresponding to $b=0$ and $b=1$, respectively. Note that in all existing
CSQBC protocols \cite{qi150,qbc50,qbc78,qi197,qbc52,qbc89}, there is no such
a system $E$. But we include it here, so that the model can cover more
protocols that may be proposed in the future.
In this scenario, Alice's cheating strategy is as follows. At the beginning
of the protocol she introduces an ancillary system $\Phi $ which is a copy of
$E\otimes \Psi $. Since the fidelity $F(\rho _{0}^{EB},\rho _{1}^{EB})\equiv
tr\sqrt{(\rho _{0}^{EB})^{1/2}\rho _{1}^{EB}(\rho _{0}^{EB})^{1/2}}$ between
$\rho _{0}^{EB}$\ and $\rho _{1}^{EB}$\ satisfies \cite{qi366
\begin{equation}
F(\rho _{0}^{EB},\rho _{1}^{EB})=\max_{\left\vert \psi _{0}\right\rangle
,\left\vert \psi _{1}\right\rangle }\left\vert \left\langle \psi _{0}\right.
\left\vert \psi _{1}\right\rangle \right\vert ,
\end{equation
where the maximization is over all purifications $\left\vert \varphi
_{0}\right\rangle $ of $\rho _{0}^{EB}$ and $\left\vert \varphi
_{1}\right\rangle $\ of $\rho _{1}^{EB}$ into $\Phi \otimes E\otimes \Psi $,
Alice finds the real and positive $\left\vert \psi _{0}\right\rangle $,
\left\vert \psi _{1}\right\rangle $ that reach the maximum, i.e.
\begin{equation}
F(\rho _{0}^{EB},\rho _{1}^{EB})=\left\langle \psi _{0}\right. \left\vert
\psi _{1}\right\rangle =\left\langle \psi _{1}\right. \left\vert \psi
_{0}\right\rangle .
\end{equation
Then she prepares the initial state of $\Phi \otimes E\otimes \Psi $ a
\begin{equation}
\left\vert \psi _{c}\right\rangle =\frac{\left\vert \psi _{0}\right\rangle
+\left\vert \psi _{1}\right\rangle }{N}, \label{psaic}
\end{equation
where the normalization constan
\begin{equation}
N=\sqrt{2+\left\langle \psi _{0}\right. \left\vert \psi _{1}\right\rangle
+\left\langle \psi _{1}\right. \left\vert \psi _{0}\right\rangle }.
\end{equation
She uses this state to complete the rest of the commit protocol. With this
method, the value of $b$ is not determined during the commit phase.
In the unveil phase, Alice decides whether she wants to unveil $b=0$ or $b=1
. Then she simply uses $\left\vert \psi _{c}\right\rangle $\ as $\left\vert
\psi _{b}\right\rangle $ to complete the protocol. From the symmetry of
\left\vert \varphi _{0}\right\rangle $ and $\left\vert \varphi
_{1}\right\rangle $\ in Eq.(\ref{psaic}), we can see that her successful
cheating probabilities for $b=0$ and $b=1$ are bot
\begin{eqnarray}
P_{A} &=&\left\vert \left\langle \psi _{0}\right. \left\vert \psi
_{c}\right\rangle \right\vert ^{2}=\frac{(1+\left\langle \psi _{0}\right.
\left\vert \psi _{1}\right\rangle )(1+\left\langle \psi _{1}\right.
\left\vert \psi _{0}\right\rangle )}{2+\left\langle \psi _{0}\right.
\left\vert \psi _{1}\right\rangle +\left\langle \psi _{1}\right. \left\vert
\psi _{0}\right\rangle } \nonumber \\
&=&\frac{1+F(\rho _{0}^{EB},\rho _{1}^{EB})}{2}. \label{pa}
\end{eqnarray
Therefore, in any specific CSQBC protocol, the Alice's exact cheating
probability can be calculated once the definition of $\rho _{0}^{EB}$, $\rho
_{1}^{EB}$ is known.
\subsection*{The optimal protocols are trivial}
Now we will try to find the CSQBC protocols which can optimally detect the
cheating of both parties, i.e., minimizing the sum of Alice's and Bob's
cheating probabilities.
Note that Eq. (\ref{pc}) depends on the specific value of $\alpha $ in the
state Eq. (\ref{initial}) that Alice chooses in a single run of the
protocol, while $F(\rho _{0}^{EB},\rho _{1}^{EB})$ in Eq. (\ref{pa}) is the
statistical result of all the legitimate states allowed by the protocol. Thus
it is hard to compare Eq. (\ref{pa}) and Eq. (\ref{pc}) directly and give a
general result without knowing the details on the composition of $\rho
_{b}^{EB}$ in a specific protocol.
Fortunately, in all existing CSQBC protocols \cit
{qi150,qbc50,qbc78,qi197,qbc52,qbc89}, there is no system $E$. The form of
the states of Bob's system $\Psi $ alone carries all the information of $b$.
Thus the trace distance $D(\rho _{0}^{EB},\rho _{1}^{EB})=D(\rho
_{0}^{B},\rho _{1}^{B})$. For any protocol of this kind (as well as
protocols having system $E$ but still satisfying $D(\rho _{0}^{EB},\rho
_{1}^{EB})=D(\rho _{0}^{B},\rho _{1}^{B})$), we can replace both $\alpha $\
and $F(\rho _{0}^{EB},\rho _{1}^{EB})$\ with $D(\rho _{0}^{B},\rho _{1}^{B})
, as elaborated in the 2nd subsection of Method, where we obtain
\begin{equation}
P_{A}\geq 1-\frac{D(\rho _{0}^{B},\rho _{1}^{B})}{2}, \label{pa lower}
\end{equation
an
\begin{equation}
P_{B}\geq \frac{1+D(\rho _{0}^{B},\rho _{1}^{B})^{2}}{2}. \label{pb lower}
\end{equation}
These two equations suggest that $P_{A}$\ and\ $P_{B}$\ cannot be minimized
simultaneously in the same protocol, because reducing $P_{A}$\ requires a
higher $D(\rho _{0}^{B},\rho _{1}^{B})$, while it will result in a higher
P_{B}$ at the same time.
Moreover, we must note that the above $P_{A}$\ and\ $P_{B}$\ are obtained
assuming that the actions of both parties in the protocol will always be
checked. But this is impossible, because they share the same system $\Phi
\otimes E\otimes \Psi $. In the unveil phase, either Bob will measure
E\otimes \Psi $ to check Alice's action, or he is required to return
\Psi $ to Alice who checks his action. These cannot be done simultaneously.
Suppose that in a CSQBC protocol, Bob's action is checked with probability
\zeta $ ($0\leq \zeta \leq 1$), and Alice's action is checked with
probability $1-\zeta $. When one's action is not checked, he/she can cheat
successfully with probability $1$. Thus the cheating probabilities $P_{A}$\
and\ $P_{B}$\ should be replaced b
\begin{equation}
P_{A}^{\ast }=\zeta +(1-\zeta )P_{A} \label{pas}
\end{equation
an
\begin{equation}
P_{B}^{\ast }=(1-\zeta )+\zeta P_{B}, \label{pbs}
\end{equation
respectively. Combining them with Eqs. (\ref{pa lower}) and (\ref{pb lower
), we find
\begin{eqnarray}
P_{A}^{\ast }+P_{B}^{\ast } &\geq &2-\frac{\zeta +D(\rho _{0}^{B},\rho
_{1}^{B})}{2} \nonumber \\
&&+\zeta D(\rho _{0}^{B},\rho _{1}^{B})\frac{1+D(\rho _{0}^{B},\rho _{1}^{B}
}{2}. \label{psum}
\end{eqnarray
Since $0\leq \zeta \leq 1$\ and $0\leq D(\rho _{0}^{B},\rho _{1}^{B})\leq 1
, we find another security lower bound of CSQB
\begin{equation}
P_{A}^{\ast }+P_{B}^{\ast }\geq \frac{3}{2}. \label{bound2}
\end{equation}
\begin{figure}[tbp]
\includegraphics{fig2}
\caption{The lower bound of the sum of the cheating probabilities $P_{A}^{\ast }+P_{B}^{\ast }$\
as a function of the trace distance $D(\rho _{0}^{B},\rho _{1}^{B})$ and the
probability $\zeta $ with which Bob's action is checked. $A$ and $B$ denote
the points that reach the minimum $P_{A}^{\ast }+P_{B}^{\ast }=3/2$.}
\label{fig:epsart}
\end{figure}
To find the optimal protocol that can reach this bound, we plot the lower
bound of $P_{A}^{\ast
}+P_{B}^{\ast }$\ as a function of $D(\rho _{0}^{B},\rho _{1}^{B})$ and
\zeta $\ in FIG. 2 according to Eq. (\ref{psum}). It shows that there are two
types of protocols that can both reach the minimum $P_{A}^{\ast
}+P_{B}^{\ast }=3/2$, denoted as points $A$ and $B$ in FIG. 2, respectively,
with the parameters (A) $D(\rho _{0}^{B},\rho _{1}^{B})=1$, $\zeta =0$,
and (B) $D(\rho _{0}^{B},\rho _{1}^{B})=0$, $\zeta =1$. Type (A) protocols
mean that $\rho _{0}^{B}$ and $\rho _{1}^{B}$ are orthogonal so that
P_{A}^{\ast }$ reaches its minimum $1/2$.
However, $\rho _{0}^{B}$ and $\rho _{1}^{B}$ can be distinguished
perfectly and Bob's action is never checked. Thus $P_{B}^{\ast }=1$,
i.e., he can always learn Alice's committed $b$ with reliability $1$ and
never get caught. In type (B) protocols, $\rho _{0}^{B}=\rho _{1}^{B}$ so
that Bob learns nothing about $b$. But Alice's action is never checked so
that she can unveil $b$ as whatever she wants, with a successful cheating
probability $P_{A}^{\ast }=1$. Therefore, we can see that these optimal
protocols are all trivial as they are completely insecure against one of the
parties. Thus they do not seem to have any practical usage.
\subsection*{The fair protocol}
Since the protocols that can minimize $P_{A}^{\ast }+P_{B}^{\ast }$\ all
look useless, let us consider the protocol satisfying $P_{A}^{\ast
}=P_{B}^{\ast }$ so that it is fair for both parties, and try to minimize
P_{A}^{\ast }$, $P_{B}^{\ast }$ in this case. From Eq. (\ref{37}) we can see
that the inequality Eq. (\ref{pb
lower}) can become equality when $\bar{\alpha}^{2}=\overline{\alpha^{2}}$, i.e.,
all the states allowed to be chosen in the protocol for committing the same
b$ value should have the same $\alpha $ value. Also, note that the lowest
bounds in Eqs. (\ref{pa lower}) and (\ref{psum}) cannot be reached by most
D(\rho _{0}^{B},\rho _{1}^{B})$, because these inequalities can become
equalities if and only if $F(\rho _{0}^{B},\rho _{1}^{B})=1-D(\rho
_{0}^{B},\rho _{1}^{B})$, which requires $\rho _{0}^{B}=\rho _{1}^{B}$.
Therefore, only the above optimal protocols can reach these bound. For this
reason, to calculate $P_{A}^{\ast }$ precisely in\ other protocols, we
should use Eq. (\ref{pa}) instead of Eq. (\ref{pa lower}). To compute
F(\rho _{0}^{EB},\rho _{1}^{EB})$\ in Eq. (\ref{pa}), for simplicity we consider
only the protocols in which there ar
\begin{eqnarray}
\rho _{0}^{EB} &=&\rho _{0}^{B}=\left[
\begin{array}{cc}
\alpha & 0 \\
0 & 1-\alph
\end{array
\right] , \nonumber \\
\rho _{1}^{EB} &=&\rho _{1}^{B}=\left[
\begin{array}{cc}
1-\alpha & 0 \\
0 & \alph
\end{array
\right] ,
\end{eqnarray
the
\begin{eqnarray}
F(\rho _{0}^{B},\rho _{1}^{B}) &=&2\sqrt{\alpha (1-\alpha )}, \nonumber \\
D(\rho _{0}^{B},\rho _{1}^{B}) &=&2\alpha -1.
\end{eqnarray
Combining them with Eqs. (\ref{pa}), (\ref{pas}), (\ref{pbs}) and (\ref{pb
lower}) (the latter becomes equality once we choose $\bar{\alpha}^{2}
\overline{\alpha^{2}}$), then by solving $P_{A}^{\ast }=P_{B}^{\ast }$ we yiel
\begin{equation}
\zeta =\frac{2\sqrt{\alpha (1-\alpha )}-1}{(2\alpha -1)^{2}+2\sqrt{\alpha
(1-\alpha )}-2}.
\end{equation
Any protocol satisfying this equation is fair for both parties. Now let us
find the minimal value of $P_{A}^{\ast }=P_{B}^{\ast }$. Substituting this
\zeta $ into Eq. (\ref{pbs}), we obtai
\begin{equation}
P_{A}^{\ast }=P_{B}^{\ast }=\frac{(2\sqrt{\alpha (1-\alpha )}+1)(2\alpha
^{2}-2\alpha +1)-2}{4\alpha ^{2}-4\alpha +2\sqrt{\alpha (1-\alpha )}-1}.
\label{pasfair3}
\end{equation
By solving $dP_{A}^{\ast }/d\alpha =0$, we find that the minimal cheating
probabilities in such protocols are $P_{A}^{\ast }=P_{B}^{\ast }=0.904$,
which can be obtained when $\alpha \simeq 0.885$, i.e., $\sqrt{\alpha
\simeq 0.941\simeq \cos (19.85^{\circ })$. In this case $\zeta \simeq 0.469$.
A simple protocol having these parameters is: Alice sends Bob the state
\cos (19.85^{\circ })\left\vert 0\right\rangle \pm \sin (19.85^{\circ
})\left\vert 1\right\rangle $\ ($\sin (19.85^{\circ })\left\vert
0\right\rangle \pm \cos (19.85^{\circ })\left\vert 1\right\rangle $) if she
wants to commit $b=0$\ ($b=1$). In the unveil phase, with probability $\zeta
\simeq 0.469$ Bob returns the state and Alice checks whether it remains
undisturbed, with probability $1-\zeta \simeq 0.531$ Bob measures the state
and checks whether it agrees with the value of Alice unveiled $b$.
Nevertheless, there is the difficulty in finding a method for deciding which
party will be checked in a single run of the protocol. Dishonest Alice (Bob)
would like to decrease $1-\zeta $\ ($\zeta $) so that $P_{A}^{\ast }$ (
P_{B}^{\ast }$)\ can be raised. Thus they do not trust each other and may
not collaborate. The CSQBC protocol in \cite{qi150} adopts a process called
\textquotedblleft the game\textquotedblright\ to handle this problem, which
is very similar to quantum coin flipping (QCF) protocols \cite{qi365}.
However, Ishizaka \cite{qi442v3} showed that this process provides extra
security loophole to Bob, so that there is a cheating strategy for him to
learn $b$ with reliability $61.79\%$\ (which is lower than what can be
obtained with our cheating strategy, as calculated in the Examples section)
while passing Alice's check with probability $100\%$ (which is higher than
that of our strategy). It was further shown in \cite{qi442} that due to the
inexistence of ideal black-boxed QCF, any CSQBC protocol based on biased QCF
cannot be secure. Therefore, it remains unclear how to build a fair CSQBC
protocol with $P_{A}^{\ast }=P_{B}^{\ast }$ while minimizing $P_{A}^{\ast }$
and $P_{B}^{\ast }$.
\section*{Discussion}
In summary, we showed that any CSQBC protocol having the above four features
is subjected to the security bound Eq. (\ref{bound}). Protocols satisfying
D(\rho _{0}^{EB},\rho _{1}^{EB})=D(\rho _{0}^{B},\rho _{1}^{B})$ is further
bounded by Eq. (\ref{bound2}). Note that the insecurity of QCF-based CSQBC
protocols (e.g., \cite{qi150,qbc50}) was already pinpointed out in \cit
{qi442v3,qi442}. But our proof also applies to the non-QCF-based ones.
Our result should not be simply considered as a generalization of the MLC
no-go proof. Instead, it is a complement. This is because the MLC no-go
proof applies to QBC protocol with $\rho _{0}^{B}\simeq \rho _{1}^{B}$. But
as pointed out in \cite{qbc52}, CSQBC does not need to satisfy this
requirement so that it may evade the MLC theorem. On the contrary, our proof
works for the case $\rho _{0}^{B}\neq \rho _{1}^{B}$, thus it fills the gap
where the MLC proof left. Meanwhile, the MLC theorem concentrates on the
cheating of Alice. It does not exclude the existence of protocols which is
unconditionally secure against dishonest Bob only. On the other hand, our
result shows that Bob can always cheat in CSQBC regardless Alice is honest
or not.
It will be interesting to study whether there can be CSQBC protocols without
the above four features. It seems that Kent's relativistic QBC \cit
{qi44,qbc24,qbc51} and our recent proposals \cite{HeJPA,HeQIP} do not
satisfy feature (1), while the protocol in \cite{HePRA} does not have
feature (2), as elaborated in \cite{HeProof}. However, these works are aimed
to achieve the original QBC, instead of CSQBC. Also, \cit
{HeJPA,HeQIP,HePRA,HeProof} have not gained wide recognition yet. Thus it is
still an open question whether it is possible to build non-relativistic
CSQBC protocols which are not limited by the above security bounds, without
relying on computational and experimental constraints.
\section*{Methods}
\subsection*{Calculating Bob's cheating probability}
Consider the POVM $\{P_{0}^{\dag }P_{0},P_{1}^{\dag }P_{1}\}$ defined in Eq.
(\ref{projector2}). After Bob applies it on $\Psi $, there can be two
outcomes.
(I) The projection outcome is $P_{0}$. Then Bob takes $b^{\prime }=0$ as his
decoded result of Alice's committed bit $b$. With Eqs.
(\ref{projector2}) and (\ref{initial}) we yiel
\begin{equation}
P_{0}\left\vert \Phi \otimes \Psi \right\rangle _{ini}=\sqrt{\alpha
\sum\limits_{i}\lambda _{i}^{(0)}\left\vert f_{i}^{(0)}\right\rangle \otimes
\left\vert e_{i}^{(0)}\right\rangle .
\end{equation
Thus this case will occurs with the probabilit
\begin{equation}
p_{I}=\alpha , \label{pI}
\end{equation
while the resultant state of $\Phi \otimes \Psi $ i
\begin{equation}
\left\vert \Phi \otimes \Psi \right\rangle _{I}=\frac{1}{\sqrt{p_{I}}
P_{0}\left\vert \Phi \otimes \Psi \right\rangle _{ini}.
\end{equation}
As described in feature (3) of CSQBC, at the unveil phase Alice may require
Bob to return $\Psi $ and check whether it remains intact in its initial
state. The maximal probability for Alice to find out that Bob has already
projected $\left\vert \Phi \otimes \Psi \right\rangle _{ini}$\ into
\left\vert \Phi \otimes \Psi \right\rangle _{I}$\ is bounded b
\begin{align}
\tilde{p}_{I}& =1-\left\vert _{I}\left\langle \Phi \otimes \Psi \right.
\left\vert \Phi \otimes \Psi \right\rangle _{ini}\right\vert ^{2} \nonumber
\\
& =1-\frac{1}{p_{I}}\alpha ^{2}.
\end{align
Thus the total probability for (case (I) occurred) $AND$ (Alice failed to
detect Bob's cheating) i
\begin{equation}
p_{I}(1-\tilde{p}_{I})=\alpha ^{2}.
\end{equation}
(II) The projection outcome is $P_{1}$. Then Bob takes $b^{\prime }=1$ as
his decoded result of Alice's $b$. No
\begin{equation}
P_{1}\left\vert \Phi \otimes \Psi \right\rangle _{ini}=\sqrt{\beta
\sum\limits_{i}\lambda _{i}^{(1)}\left\vert f_{i}^{(1)}\right\rangle \otimes
\left\vert e_{i}^{(1)}\right\rangle .
\end{equation
Obviously, this case will occurs with the probabilit
\begin{equation}
p_{II}=1-p_{I}.
\end{equation
Meanwhile, the resultant state of $\Phi \otimes \Psi $ in this case i
\begin{equation}
\left\vert \Phi \otimes \Psi \right\rangle _{II}=\frac{1}{\sqrt{p_{II}}
P_{1}\left\vert \Phi \otimes \Psi \right\rangle _{ini}.
\end{equation}
The maximal probability for Alice to find out that Bob has already projected
$\left\vert \Phi \otimes \Psi \right\rangle _{ini}$\ into $\left\vert \Phi
\otimes \Psi \right\rangle _{II}$\ is bounded b
\begin{align}
\tilde{p}_{II}& =1-\left\vert _{II}\left\langle \Phi \otimes \Psi \right.
\left\vert \Phi \otimes \Psi \right\rangle _{ini}\right\vert ^{2} \nonumber
\\
& =1-\frac{1}{p_{II}}\beta ^{2}.
\end{align
Thus the total probability for (case (II) occurred) $AND$ (Alice failed to
detect Bob's cheating) i
\begin{equation}
p_{II}(1-\tilde{p}_{II})=\beta ^{2}.
\end{equation}
Taking both cases (I) and (II) into consideration, the overall probability
for Bob's cheating to pass Alice's detection successfully i
\begin{eqnarray}
P_{B} &=&p_{I}(1-\tilde{p}_{I})+p_{II}(1-\tilde{p}_{II})=\alpha ^{2}+\beta
^{2} \nonumber \\
&=&\frac{1}{2}+\frac{1}{2}(2\alpha -1)^{2}. \label{pc*}
\end{eqnarray}
Meanwhile, since the projection outcome will either be $P_{0}$ or $P_{1}$\
with the probabilities $p_{I}$\ and $p_{II}=1-p_{I}$, respectively, Bob's
b^{\prime }$ will match Alice's $b$ with the probability $p_{I}$\ or
1-p_{I} $ too. Note that $h(1-p_{I})=h(p_{I})$.
Thus the amount of mutual information that Bob obtains with this POVM i
\begin{equation}
I_{m}=1-h(p_{I})=1-h(\alpha ). \label{Im*}
\end{equation}
\subsection*{Bounding the cheating probabilities with trace distance}
Suppose that there are many states allowed to be chosen randomly for
committing $b=0$ in the protocol, each of which takes the form of Eq. (\re
{initial}),
but with different values of the coefficients $\alpha $, $\beta $, $\lambda
_{i}^{(0)}$'s and $\lambda _{i}^{(1)}$'s. Bob applies the optimal POVM
to decode $b$. Then Eq. (\ref{initial}) indicates that he can learn $b$
correctly with probability $\bar{\alpha}$, i.e., the average of $\alpha $.
Meanwhile, it is well-known that the maximal probability for discriminating
two density matrices $\rho _{0}^{B}$, $\rho _{1}^{B}$ is $(1+D(\rho
_{0}^{B},\rho _{1}^{B}))/2$.\ Therefor
\begin{equation}
D(\rho _{0}^{B},\rho _{1}^{B})=2\bar{\alpha}-1.
\end{equation
Since Eq. (\ref{pc}) shows that Bob's average cheating probability for these
states i
\begin{equation}
P_{B}=\frac{1+\overline{(2\alpha -1)^{2}}}{2}\geq \frac{1+(2\bar{\alpha
-1)^{2}}{2}, \label{37}
\end{equation
we hav
\begin{equation}
P_{B}\geq \frac{1+D(\rho _{0}^{B},\rho _{1}^{B})^{2}}{2}. \label{pb lower M}
\end{equation
Similar discussion is also valid for the states for committing $b=1$, except
that $\alpha $ should be replace by $\beta =1-\alpha $. But Eq. (\ref{pb
lower M}) remains the same because Eq. (\ref{pc}) satisfies $P_{B}(1-\alpha
)=P_{B}(\alpha )$.
On the other hand, since \cite{qi366
\begin{equation}
F(\rho _{0}^{B},\rho _{1}^{B})\geq 1-D(\rho _{0}^{B},\rho _{1}^{B}),
\end{equation
from Eq. (\ref{pa}) we yield
\begin{equation}
P_{A}\geq 1-\frac{D(\rho _{0}^{B},\rho _{1}^{B})}{2}. \label{pa lower E}
\end{equation
|
\section{Introduction}
Khovanov-Lauda-Rouquier algebras (henceforth KLR algebras), also known as Quiver Hecke algebras, are a family of $\mathbb{Z}$-graded associative algebras introduced by Khovanov and Lauda \cite{khovanovlauda} and Rouquier \cite{rouquier} for the purposes of categorifying quantum groups. More specifically they categorify the upper-triangular part $\mathbf{f}=U_q(\mathfrak{g})^+$ of the quantised enveloping algebra of a symmetrisable Kac-Moody Lie algebra $\mathfrak{g}$ - see \S \ref{klr} for a precise statement. Let $I$ be the set of simple roots of $\mathfrak{g}$ and $\mathbb{N} I$ the monoid of formal sums of elements of $I$. For each $\nu\in\mathbb{N} I$ there is an associated KLR algebra $R(\nu)$.
In this paper we will assume that $\mathfrak{g}$ is of symmetric affine type. For now however, we will describe the theory developed in \cite{klr1,bkm} where $\mathfrak{g}$ is finite dimensional. The results of this paper generalise these results to the symmetric affine case.
One begins with choosing a convex order $\prec$ on the set of positive roots satisfying a convexity property - see Definition \ref{convexdefn}. It is this convex order which determines a PBW basis of $\mathbf{f}$. The representation theory of KLR algebras is built via induction functors from the theory of cuspidal representations. Write $\{\alpha_1\succ \cdots \succ\alpha_N\}$ for the set of positive roots, remembering that we are temporarily discussing the finite type case.
To each root $\alpha$ there is a subcategory of $R(\alpha)$-modules which are cuspidal defined in Definition \ref{cuspidaldefn}. There is a unique irreducible cuspidal module $L(\alpha)$. Let $\Delta(\alpha)$ be the projective cover of $L(\alpha)$ in the category of cuspidal $R(\alpha)$-modules.
Given any sequence $\p=(\p_1,\ldots,\p_N)$ of natural numbers, the proper standard and standard modules are defined respectively by
\begin{align*}
\overline{\Delta}(\p)&=L(\alpha_1)^{\circ \p_1} \circ \cdots \circ L(\alpha_N)^{\circ \p_N} \\
\Delta(\p)&= \Delta(\alpha_1)^{(\p_1)}\circ \cdots\circ \Delta(\alpha_N)^{(\p_N)}
\end{align*}
where $\circ$ denotes the induction of a tensor product and $(\p_i)$ is a divided power construction. Then in \cite{klr1} it is proved that the modules $\overline{\Delta}(\p)$ categorify the dual PBW basis, have a unique irreducible quotient and that these quotients give a classification of all irreducible modules. In \cite{bkm} it is proved that the modules $\Delta(\p)$ categorify the PBW basis and their homological properties are studied, justifying the use of the term standard.
Now let us turn our attention to the results of this paper where $\mathfrak{g}$ is of symmetric affine type. Again the starting point is the choice of a convex order $\prec$ on the set of positive roots. The theory of PBW bases for affine quantised enveloping algebras dates back to the work of Beck \cite{beck} and is considerably more complicated than the theory in finite type. It is a feature of the literature that the theory of PBW bases is only developed for convex orders of a particular form. We rectify this problem by presenting a construction of PBW bases in full generality.
For $\alpha$ a real root, the category of cuspidal $R(\alpha)$-modules is again equivalent to the category of $k[z]$-modules while the category of semicuspidal $R(n\alpha)$-modules is again equivalent to modules over a polynomial algebra. Whereas in finite type the proofs of these results currently rest on some case by case computations, here we give a uniform proof, the cornerstone of which is the growth estimates in \S \ref{growth}.
For the imaginary roots, the category of semicuspidal representations is qualitatively very different. The key observation here is that the R-matrices constructed by Kang, Kashiwara and Kim \cite{kkk} enable us to determine an isomorphism
\[
\operatorname{End}(M^{\circ n})\cong \mathbb{Q}}\def\partition{\mathcal{P}[S_n]
\]
where $M$ is either an irreducible cuspidal $R(\delta)$-module or an indecomposable projective in the category of cuspidal $R(\delta)$-modules (here $\delta$ is the minimal imaginary root). We are then able to use the representation theory of the symmetric group to decompose these modules $M^{\circ n}$. This presence of the symmetric group as an endomorphism algebra can also be seen to explain the appearance of Schur functions in the definition of a PBW basis in affine type.
With the semicuspidal modules understood we are able to prove our main theorems which are analogous to those discussed above in finite type. Namely families of proper standard and standard modules are constructed which categorify the dual PBW and PBW bases respectively. See Theorem \ref{pbwcat} and the following paragraph for this result. The proper standard modules have a unique irreducible quotient which gives a classification of all irreducibles and the standard modules satisfy homological properties befitting their name, leading to a BGG reciprocity theorem.
As a consequence we obtain a new positivity result, Theorem \ref{positif}, which states that when an element of the canonical basis of $\mathbf{f}$ is expanded in a PBW basis, the coefficients that appear are polynomials in $q$ and $q^{-1}$ with non-negative coefficients (and the transition matrix is unitriangular).
We thank A. Kleshchev, P. Tingley and B. Webster for useful conversations.
\section{Preliminaries
The purpose of this section is to collect standard notation about root systems and other objects which we will be making use of in this paper.
Let $(I,\cdot)$ be a Cartan Datum of symmetric affine type.
Following the approach of Lusztig \cite{lusztigbook}, this comprises a finite set $I$ and a symmetric pairing $\cdot\map{I\times I}{\mathbb{Z}}$ such that $i\cdot i=2$ for all $i\in I$, $i\cdot j\leq 0$ if $i\neq j$ and the matrix $(i\cdot j)_{i,j\in I}$ is of corank 1.
Such Cartan data are completely classified and correspond to the extended Dynkin diagrams of type A, D and E. We extend $\cdot\map{I\times I}{\mathbb{Z}}$ to a bilinear pairing $\mathbb{N} I\times \mathbb{N} I\to \mathbb{Z}$.
Let $\Phi^+$ be the set of positive roots in the corresponding root system.
We identify $I$ with the set of simple roots of $\Phi^+$. In this way we are able to meaningfully talk about elements of $\mathbb{N} I$ as being roots.
The set of real roots of $\Phi^+$ is denoted $\Phi^+_{re}$.
For $\nu=\sum_{i\in I}\nu_i\cdot i\in I$, define $|\nu|=\sum_{i\in I }\nu_i$. If $\nu$ happens to be a root, we also call this the height of the root and denote it $\operatorname{ht}(\nu)$.
Let $\Phi_f$ be the underlying finite type root system. A chamber coweight is a fundamental coweight for some choice of positive system on $\Phi_f$. If a positive system is given, let $\Omega$ denote the set of chamber coweights with respect to this system.
Let $p\map{\Phi}{\Phi_{f}}$ denote the projection from the affine root system to the finite root system whose kernel is spanned by the minimal imaginary root $\delta$. For $\alpha\in\Phi_f$, let $\tilde{\alpha}$ denote the minimal positive root in $p^{-1}(\alpha)$.
Let $W=\langle s_i\mid i\in I\rangle$ be the Weyl group of $\Phi$, generated by the simple reflection $s_i$ which is the reflection in the hyperplane perpendicular to $\alpha_i$.
Let $\Delta_f$ be the standard set of simple roots in $\Phi_f$. Let $W_f$ be the finite Weyl group.
Let $\P$ denote the set of partitions. A multipartition $\ula=\{\lambda_\w\}_{\w\in\Omega}$ is a sequence of partitions indexed by $\Omega$. We write $\ula\vdash n$ if $\sum_\w |\lambda_\w|=n$.
The symmetric group on $n$ letters is denoted $S_n$. If $\mu,\nu\in I$, the element $w[\mu,\nu]\in S_{|\mu+\nu|}$ is defined by
\[
w[\mu,\nu](i)=\begin{cases}
i+|\nu| & \text{if $i\leq \mu$} \\
i-|\mu| & \text{otherwise}.
\end{cases}
\]
\section{Convex Orders on Root Systems
\begin{definition}\label{convexdefn}
A convex order on $\Phi^+$ is a total preorder $\preceq$ on $\Phi^+$ such that
\begin{itemize}
\item If $\alpha\preceq \beta$ and $\alpha+\beta$ is a root, then $\alpha\preceq \alpha+\beta\preceq \beta$.
\item If $\alpha\preceq\beta$ and $\beta\preceq\alpha$ then $\alpha$ and $\beta$ are imaginary roots.
\end{itemize}
\end{definition}
\begin{theorem}
A convex order $\prec$ on $\Phi^+$ satisfies the following condition:
\begin{itemize}
\item Suppose $A$ and $B$ are disjoint subsets of $\Phi^+$ such that $\alpha\prec\beta$ for any $\alpha\in A$ and $\beta\in B$. Then the cones formed by the $\mathbb{R}_{\geq 0}$ spans of $A$ and $B$ meet only at the origin.
\end{itemize}\end{theorem}
\begin{remark}
In \cite{tingleywebster}, this condition replaces our first condition in their definition of a convex order. This theorem shows that their definition and our definition agree.
\end{remark}
\begin{remark}
The following proof requires being in finite or affine type since it depends on the positive semidefiniteness of the natural bilinear form.
We do not know if a similar statement is possible for more general root systems.
\end{remark}
\begin{proof}
We will write $(\cdot,\cdot)$ for the natural bilinear form on the root lattice.
Let $\{\alpha_i\}$ be a finite set of roots in $A$ and let $\{b_j\}$ be a finite set of roots in $B$.
For want of a contradiction, suppose that for some positive real numbers $c_i,d_j$ we have
\begin{equation}\label{cadb}\sum_i c_i\alpha_i= \sum_j d_j \beta_j\end{equation}
Any linear dependence between roots arises from linear dependencies over $\mathbb{Q}}\def\partition{\mathcal{P}$. Since $\mathbb{Q}}\def\partition{\mathcal{P}$ is dense in $\mathbb{R}$, without loss of generality we may assume that $c_i$ and $d_j$ are rational numbers. Clearing denominators, we may assume they are integers.
Now suppose we have a solution to (\ref{cadb}) where the $c_i$ and $d_j$ are positive integers with $\sum_i c_i + \sum_j d_j$ as small as possible.
For any $i\neq j$, if $\alpha_i+\alpha_j$ were a root, we could replace one occurrence of $\alpha_i$ and $\alpha_j$ by the single root $\alpha_i+\alpha_j$ to get a smaller solution, contradicting our minimality assumption. Therefore $\alpha_i+\alpha_j$ is not a root for any $i\neq j$. This implies that $(\alpha_i,\alpha_j)\geq 0$ for $i\neq j$.
If all $\alpha_i$ and $\beta_j$ are imaginary, this easily leads to a contradiction. So there exists at least one real root in the equation we are studying, without loss of generality say it is $\alpha_k$.
Applying $(\cdot,\alpha_k)$ for some $k$ leaves us with the inequality
\[
\sum_j d_j (\beta_j,\alpha_k)\geq c_k(\alpha_k,\alpha_k)>0.
\]
Therefore there exists $j$ such that $(\beta_j,\alpha_k)>0$, which implies that $\beta_j-\alpha_k$ is a positive root. By convexity this root must be greater than $\beta_j$. So now we may subtract $\alpha_k$ from both sides of (\ref{cadb}) to obtain a smaller solution, again contradicting minimality.
Therefore no solution to (\ref{cadb}) can exist, as required.
\end{proof}
The imaginary roots in any root system are all multiples of a fundamental imaginary root, which we
will denote $\delta$. In any convex order, these imaginary roots must all be equal to each other.
Let $\prec$ be a convex order. The set of positive real roots is divided into two disjoint subsets, namely
\[
\Phi_{\prec \delta} = \{ \alpha\in \Phi^+ \mid \alpha\prec \delta\},
\]and
\[
\Phi_{\succ \delta} = \{ \alpha \in \Phi^+ \mid \alpha\succ \delta\}.
\]
If we can write $\Phi_{\prec \delta} = \{ \alpha_1 \prec \alpha_2\prec \cdots \}$ and $\Phi_{\succ \delta} = \{ \beta_1 \succ \beta_2\succ \cdots \}$ for some sequences of roots $\{\alpha_i\}_{i=1}^\infty$ and $\{\beta_j\}_{j=1}^\infty$, then we say that $\prec$ is of \emph{word type}.
\begin{example}
Let $(V,\leq)$ be a totally ordered $\mathbb{Q}}\def\partition{\mathcal{P}$-vector space. Let $h\map{\mathbb{Q}}\def\partition{\mathcal{P}\Phi}{V}$ be an injective linear transformation. For two positive roots $\alpha$ and $\beta$, say that $\alpha\prec \beta$ if $h(\alpha)/|\alpha|<h(\beta)/|\beta|$ and $\alpha\preceq \beta$ if $h(\alpha)/|\alpha|\leq h(\beta)/|\beta|$.
This defines a convex order on $\Phi$.
\end{example}
In the above example, we can take $V=\mathbb{R}$ with the standard ordering to get the existence of many convex orders of word type.
An example of a convex order not of word type which we will make use of later on is the following:
\begin{example}\label{tworow}
Let $V=\mathbb{R}^2$ where $(x,y)\leq (x',y')$ if $x<x'$ or $x=x'$ and $y\leq y'$. Let $\Delta_f$ be a simple system in $\Phi_f$ and pick $\alpha\in\Delta_f$. Define $h\map{\mathbb{Q}}\def\partition{\mathcal{P}\Phi}{V}$ by $h(\tilde{\alpha})=(0,1)$, $h(\tilde{\beta})=(x_\beta,0)$ for $\beta\in\Delta_f\setminus\{\alpha\}$ where the $x_\beta$ are generically chosen positive real numbers, and $h(\delta)=0$. We extend by linearity, noting that $\{\delta\}\cup \{\tilde{\beta}\mid \beta\in\Delta_f\}$ is a basis of $\mathbb{Q}}\def\partition{\mathcal{P}\Phi$.
\end{example}
In this example, we have
\[
\widetilde{\delta-\alpha}\prec\widetilde{\delta-\alpha}+\delta\prec\widetilde{\delta-\alpha}+2\delta\cdots\prec\mathbb{Z}_{>0}\delta \prec \cdots \prec \tilde{\alpha}+2\delta\prec\tilde{\alpha}+\delta\prec\tilde{\alpha}
\] and all other positive roots are either greater than $\tilde{\alpha}$ or less than $\widetilde{\delta-\alpha}$.
Recall that $p$ is the projection from the affine root system to the finite root system.
\begin{lemma}
There exists $w\in W_f$ such that $p(\Phi_{\prec \delta})=w \Phi^+_f$ and $p(\Phi_{\succ \delta}) = w \Phi^-_f$.
\end{lemma}
\begin{proof}
First suppose that $\alpha\in p(\Phi_{\prec \delta})$ and $-\alpha\in p(\Phi_{\prec \delta})$. Then there are integers $m$ and $n$ such that the affine roots $-\alpha+m\delta$ and $\alpha+n\delta$ are both less than $\delta$ in the convex order $\prec$. By convexity, their sum $(m+n)\delta$ is also less that $\delta$, a contradiction.
Since a similar argument holds for $p(\Phi_{\succ \delta})$, we see that for each finite root $\alpha$, exactly one of $\alpha$ and $-\alpha$ lies in $p(\Phi_{\prec \delta})$.
Now suppose that $\alpha,\beta\in p(\Phi_{\prec \delta})$ and $\alpha+\beta$ is a root. Then for some integers $m$ and $n$, the affine roots $\alpha+m\delta$ and $\beta+n\delta$ are both less than $\delta$. By convexity, their sum $(\alpha+\beta)+(m+n)\delta$, which is also an affine root, is also less than $\delta$.
Therefore $\alpha+\beta\in p(\Phi_{\prec \delta})$.
We have shown that $p(\Phi_{\prec \delta})$ is a positive system in the finite root system $\Phi_f$. This suffices to prove the lemma.
\end{proof}
Define a finite initial segment to be a finite set of roots $\alpha_1\prec \alpha_2\prec\cdots\prec \alpha_N$ such that for all positive roots $\beta$, either $\beta\succ \alpha_i$ for all $i=1,\ldots,N$ or $\beta=\alpha_i$ for some $i$.
For any $w\in W$ define $\Phi(w)=\{\alpha\in\Phi^+\mid w^{-1}\alpha\in\Phi^-\}$.
\begin{lemma}\label{finiteinitial}
Let $\alpha_1\prec \alpha_2 \prec\cdots \prec \alpha_N$ be a finite initial segment. Then there exists $w\in W $ such that $\{ \alpha_1,\ldots \alpha_N \} =\Phi(w)$.
Furthermore there exists a reduced expression $w=s_{i_1}\cdots s_{i_N}$ such that $\alpha_k=s_{i_1}\cdots s_{i_{k-1}} \alpha_{i_k}$ for $k=1,\ldots N$.
\end{lemma}
\begin{proof}
The proof proceeds by induction on $N$. For the base case where $N=1$, any root $\alpha$ which is not simple is the sum of two roots $\alpha=\beta+\gamma$. By convexity of $\prec$, either $\beta\prec\alpha\prec\gamma$ or $\gamma\prec\alpha\prec\beta$. Either way, $\alpha\neq\alpha_1$ so $\alpha_1$ is simple, $\alpha_1=\alpha_i$ for some $i\in I$ and we take $w=s_i$.
Now assume that the result is known for initial segments with fewer than $N$ roots.
Let $v=s_{i_1}\ldots s_{i_{N-1}}$. Consider $v^{-1}\alpha_N$. By inductive hypothesis, it is a positive root. Suppose for want of a contradiction that $v^{-1}\alpha_N$ is not simple.
Then we can find positive roots $\beta$ and $\gamma$ such that $v^{-1}\alpha_N=\beta+\gamma$.
We can't have $v\beta=\alpha_N$ as this would force $\gamma=0$. If $v\beta=\alpha_j$ for some $j<N$ then $\beta=v^{-1}\alpha_j$ which by inductive hypothesis is in $\Phi^-$, a contradiction. Therefore either $v\beta$ is a positive root satisfying $v\beta\succ \alpha_N$ or $v\beta\in \Phi^-$. A similar statement holds for $v\gamma$.
To have both $v\beta$ and $v\gamma$ greater than $\alpha_N$ contradicts the convexity of $\prec$. Therefore, without loss of generality, we may assume $v\beta\in\Phi^-$. Then $-v\beta$ is a positive root with $v^{-1}(-v\beta)=-\beta$ which is a negative root, so by inductive hypothesis, $-v\beta=\alpha_j$ for some $j<N$. Now consider the equation $\alpha_N + (-v\beta)=v\gamma$. The convexity of $\prec$ implies that $v\gamma=\alpha_{j'}$ for some $j'<N$. This option is shown to be impossible in the previous paragraph, creating a contradiction. Therefore $v^{-1}\alpha_N$ must be a simple root.
Define $i_N\in I$ by $\alpha_{i_N}=v^{-1}\alpha_N$ and let $w=s_{i_1}\cdots s_{i_N}$. It remains to show that $\{ \alpha_1,\ldots \alpha_N \} = \{\alpha\in \Phi^+ \mid w^{-1}\alpha\in \Phi^-\}$.
If $\beta$ is a positive root that is not equal to $\alpha_j$ for some $j\leq N$, then by inductive hypothesis $v^{-1}\beta\in \Phi^+$. Then $w^{-1}\beta=s_{i_N}(v^{-1}\beta)\in\Phi^-$ if and only if $v^{-1}\beta=\alpha_{i_N}$ which isn't the case since this is equivalent to $\beta=\alpha_N$.
If $\beta=\alpha_j$ for some $j<N$ then $v^{-1}\beta\in\Phi^-$. So $w^{-1}\beta=s_{i_N}(v^{-1}\beta)\in\Phi^-$ unless $v^{-1}\beta=-\alpha_{i_N}$. This isn't the case since it is equivalent to $\beta=-\alpha_{i_N}$.
The above two paragraphs show that for a positive root $\beta$, if $\beta\in\{\alpha_1,\ldots ,\alpha_{N-1}\}$ then $w^{-1}\beta\in\Phi^+$ while if $\beta\notin \{\alpha_1,\ldots,\alpha_N\}$, then $w^{-1}\beta\in\Phi^+$. Since $w^{-1}\alpha_N=-\alpha_{i_N}\in\Phi^-$, this completes the proof.
\end{proof}
\begin{lemma}\cite{ito}
The restriction of a convex order to $\Phi_{\prec \delta}$ is of $n$-row type for some $n$, i.e. it is isomorphic to the ordinal $\omega\cdot n$.
\end{lemma}
For a convex order $\prec$, define
\[
I(\prec) = \{\alpha\in\Phi^+\mid \{\beta\in \Phi^+\mid \beta\prec \alpha\} \mbox{ is finite}\}
\]
\begin{lemma}\label{3.10}
Let $\prec$ be a convex order not of word type. Let $\beta$ be the smallest root that is not in any initial segment of $\Phi^+$. Let $S$ be a finite set of roots containing $\beta$. Then there exists a convex order $\prec'$ such that $I(\prec')=I(\prec)\cup \{\beta\}$ and the restrictions of $\prec $ and $\prec'$ to $S$ are the same.
\end{lemma}
\begin{proof}
Let $L$ be the set of roots in $\Phi^+$ less than or equal to $\beta$ under $\prec$. Then by \cite[Theorem 3.12]{cellinipapi}, there exists $v,t\in W$ with $t$ a translation
and $L=\cup_{n=1}^\infty \Phi(vt^n)$.
Let $w$ be such that $S\subset \{\alpha_1\prec\cdots\prec\alpha_N\}=\Phi_w$. There exists an integer $n$ such that $\Phi(w)\cup \{\beta\}\subset \Phi(vt^n)$. Let $v'=vt^n$. Since $\Phi(v')\supset \Phi(w)$, for any reduced expression of $w$, there exists a reduced expression of $v'$ beginning with that of $w$.
We choose the reduced decomposition of $w$ to be compatible with $\prec$. Then extend the reduced decomposition as per the above to get a new ordering $\prec'$ on $L$. This has the desired properties.
\end{proof}
\begin{theorem}\label{3.11}
Let $S$ be a finite subset of $\Phi^+$ and let $\prec$ be a convex order on $\Phi^+$. Then there exists a convex order $\prec'$ of word type such that the restrictions of $\prec$ and $\prec'$ to $S$ are equivalent.
\end{theorem}
\begin{proof}
Suppose our convex order begins
\[
\alpha_1\prec \alpha_2 \prec \cdots \prec \beta_1 \prec \beta_2 \prec \cdots
\]
and that $S\cap \{\alpha_i\mid i\in \mathbb{Z}^+\} \subset \{\alpha_1,\ldots,\alpha_n\}$. We now define inductively a sequence of convex orders $\prec_i$ with $I(\prec_i)=\{\alpha_i\mid i\in\mathbb{Z}^+\}\cup \{\beta_1,\ldots,\beta_i\}$ as follows:
Set $\prec_0=\prec$. Assume that $\prec_{i}$ is constructed. To construct $\prec_{i+1}$, apply Lemma \ref{3.10} with $S=\{\alpha_1,\ldots,\alpha_{n+i},\beta_1,\ldots,\beta_{i}\}$. We will take the convex order denoted $\prec'$ whose existence is given to us by Lemma \ref{3.10} as $\prec_{i+1}$.
Now let $\prec''=\lim_{i\to \infty} \prec_i$. If $\prec$ is of $n$-row type, then $\prec''$ will be of $(n-1)$-row type and the restrictions of $\prec$ and $\prec''$ to $S$ are the same. After iterating this process we reach a new convex order $\prec'$ whose restriction to $S$ is the same as $\prec$ and is of word type on $\Phi_{\prec\delta}$. Repeating this construction on the set of roots greater than $\delta$ completes the proof of this theorem.
\end{proof}
\begin{remark}
Using this theorem it will often be possible to assume without loss of generality that the convex order $\prec$ is of word type.
\end{remark}
\section{The Algebra $\mathbf{f}$}
The algebra $\mathbf{f}_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$ is the $\mathbb{Q}}\def\partition{\mathcal{P}(q)$ algebra as defined in \cite{lusztigbook} generated by elements $\{\theta_i\mid i\in I\}$. Lusztig defines it as the quotient of a free algebra by the radical of a bilinear form. By the quantum Gabber-Kac theorem, it can also be defined in terms of the Serre relations.
Morally, $\mathbf{f}_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$ should be thought of as the positive part of the quantised enveloping algebra $U_q(\mathfrak{g})$.
There is only a slight difference in the coproduct, necessary as the coproduct in $U_q(\mathfrak{g})$ does not map $U_q(\mathfrak{g})^+$ into $U_q(\mathfrak{g})^+\otimes U_q(\mathfrak{g})^+$.
There is a $\mathbb{Z}[q,q^{-1}]$-form of $\mathbf{f}_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$, which we denote simply by $\mathbf{f}$. It is the $\mathbb{Z}[q,q^{-1}]$-subalgebra of $\mathbf{f}_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$ generated by the divided powers $\theta_i^{(n)}=\theta_i^n/[n]_q!$, where $[n]_q!=\prod_{i=1}^n(q^i-q^{-i})/(q-q^{-1})$ is the $q$-factorial. If $\mathcal{A}$ is any $\mathbb{Z}[q,q^{-1}]$-algebra, we use the notation $\mathbf{f}_\mathcal{A}$ for $\mathcal{A}\otimes_{\mathbb{Z}[q,q^{-1}]}\mathbf{f}$.
The algebra $\mathbf{f}$ is graded by $\mathbb{N} I$ where $\theta_i$ has degree $i$ for all $i\in I$.
We write $\mathbf{f}=\oplus_{\nu\in\mathbb{N} I} \mathbf{f}_\nu$ for its decomposition into graded components.
Of significant importance for us is the dimension formula
\begin{equation}\label{dimfnu}
\sum_{\nu\in \mathbb{N} I}\dim \mathbf{f}_\nu t^\nu = \prod_{\alpha\in\Phi^+} (1-t^{\alpha})^{-\operatorname{mult}(\alpha)}
\end{equation}
The tensor product $\mathbf{f}\otimes \mathbf{f}$ has an algebra structure given by
\[
(x_1\otimes y_1)(x_2\otimes y_2)=q^{\beta_1\cdot\alpha_2}x_1 x_2 \otimes y_1 y_2
\] where $y_1$ and $x_2$ are homogeneous of degree $\beta_1$ and $\alpha_2$ respectively.
Given a bilinear form $(\cdot,\cdot)$ on $\mathbf{f}$, we obtain a bilinear form $(\cdot,\cdot)$ on $\mathbf{f}\otimes \mathbf{f}$ by
\[
(x_1 \otimes x_2,y_1\otimes y_2)=(x_1,x_2)(y_1,y_2).
\]
There is a unique algebra homomorphism $r\map{\mathbf{f}}{\mathbf{f}\otimes \mathbf{f}}$ such that $r(\theta_i)=\theta_i\otimes 1 + 1 \otimes \theta_i$ for all $i\in I$.
The algebra $\mathbf{f}$ has a symmetric bilinear form $\langle\cdot,\cdot\rangle$ satisfying
\begin{eqnarray*}
\langle\theta_i,\theta_i\rightarrow&=&(1-q^2)^{-1} \\
\langle x,yz\rightarrow &=& \langle r(x),y\otimes z\rightarrow.
\end{eqnarray*}
The form $\langle \cdot,\cdot\rightarrow$ is nondegenerate. Indeed, in the definition of $\mathbf{f}$ in \cite{lusztigbook}, $\mathbf{f}$ is defined to be the quotient of a free algebra by the radical of this bilinear form. It is known that $\mathbf{f}$ is a free $\mathbb{Z}[q,q^{-1}]$-module. Let $\mathbf{f}^*$ be the graded dual of $\mathbf{f}$ with respect to $\langle \cdot,\cdot\rightarrow$. By definition, $\mathbf{f}^*=\oplus_{\nu\in\mathbb{N} I} \mathbf{f}_\nu^*$. As twisted bialgebras over $\mathbb{Q}}\def\partition{\mathcal{P}(q)$, $\mathbf{f}_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$ and $\mathbf{f}^*_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$ are isomorphic, though there is no such isomorphism between their integral forms.
\section{KLR Algebras}\label{klr
A good introduction to the basic theory of KLR algebras appears in \cite[\S 4]{kleschevram}.
Although it is not customary, we will first give the geometric construction of KLR algebras, then discuss the standard presentation in terms of generators and relations.
Define a graph with vertex set $I$ and with $-i\cdot j$ edges between $i$ and $j$ for all $i\neq j$. Let $Q$ be the quiver obtained by placing an orientation on this graph.
For $\nu\in\mathbb{N} I$, define $E_\nu$ and $G_\nu$ by $$E_\nu=\prod_{i\to j}\operatorname{Hom}_{\mathbb{C}} (\mathbb{C}^{\nu_i},\mathbb{C}^{\nu_j}),$$
$$G_\nu=\prod_i GL_{\nu_i}(\mathbb{C}).$$
With the obvious action of $G_\nu$ on $E_\nu$, $E_\nu/G_\nu$ is the moduli stack of representations of $Q$ with dimension vector $\nu$.
Let $\flags_\nu$ be the complex variety whose points consist of a point of $E_\nu$, together with a full flag of subrepresentations of the corresponding representation of $Q$.
The variety $\flags_\nu$ is a disjoint union of smooth connected varieties. Let $\flags_\nu=\sqcup_{\bf i} \flags_\nu^{\bf i}$ be its decomposition into connected components and $\pi^{\bf i}\map{\flags_\nu^{\bf i}}{E_\nu}$ be the natural $G_\nu$-equivariant morphism. Define
\[
\mathcal{L}=\bigoplus_{\bf i} (\pi^{\bf i})_! \underline{\mathbb{Q}}\def\partition{\mathcal{P}}_{\flags_\nu^{\bf i}}[\dim \flags_\nu^{\bf i}] \in D^b_{G_\nu}(E_\nu).
\]
For each $\nu\in\mathbb{N} I$ we define the KLR algebra $R(\nu)$ by
\[
R(\nu)=\bigoplus_{d\in\mathbb{Z}} \operatorname{Hom}_{D^b_{G_\nu}(E_\nu)}(\mathcal{L},\mathcal{L}[d]).
\]
We now introduce the more customary approach via generators and relations. This presentation is due to \cite{vv} and \cite{rouquier}, and more recently over $\mathbb{Z}$ in \cite{maksimau}.
To introduce this presentation, we first need to define, for any $\nu\in\mathbb{N} I$,
\[
\,\mbox{Seq}\,(\nu)=\{{\bf i}=({\bf i}_1,\ldots,{\bf i}_{|\nu|})\in I^{|\nu|}\mid\sum_{j=1}^{|\nu|} {\bf i}_j=\nu\}.
\]
This is acted upon by the symmetric group $S_{|\nu|}$ in which the adjacent transposition $(i,i+1)$ is denoted $s_i$.
Define the polynomials $Q_{i,j}(u,v)$ for $i,j\in I$ by
\[
Q_{i,j}(u,v)=\begin{cases}
\prod_{i\to j}(u-v) \prod_{j\to i}(v-u)
&\text{if } i\neq j \\
0 &\text{if } i=j
\end{cases}
\] where the products are over the sets of edges from $i$ to $j$ and from $j$ to $i$, respectively.
\begin{theorem}
The KLR algebra $R(\nu)$ is the associative $\mathbb{Q}}\def\partition{\mathcal{P}$-algebra generated by elements $e_{\bf i}$, $y_j$, $\tau_k$ with ${\bf i}\in \,\mbox{Seq}\,(\nu)$, $1\leq j\leq |\nu|$ and $1\leq k< |\nu|$ subject to the relations
\begin{equation} \label{eq:KLR}
\begin{aligned}
& e_{\bf i} e_{\bf j} = \delta_{{\bf i}, \bf j} e_{\bf i}, \ \
\sum_{{\bf i} \in \,\mbox{Seq}\,(\nu)} e_{\bf i} = 1, \\
& y_{k} y_{l} = y_{l} y_{k}, \ \ y_{k} e_{\bf i} = e_{\bf i} y_{k}, \\
& \tau_{l} e_{\bf i} = e_{s_{l}{\bf i}} \tau_{l}, \ \ \tau_{k} \phi_{l} =
\tau_{l} \phi_{k} \ \ \text{if} \ |k-l|>1, \\
& \tau_{k}^2 e_{\bf i} = Q_{{\bf i}_{k}, {\bf i}_{k+1}} (y_{k}, y_{k+1})e_{\bf i}, \\
& (\tau_{k} y_{l} - y_{s_k(l)} \tau_{k}) e_{\bf i} = \begin{cases}
-e_{\bf i} \ \ & \text{if} \ l=k, {\bf i}_{k} = {\bf i}_{k+1}, \\
e_{\bf i} \ \ & \text{if} \ l=k+1, {\bf i}_{k}={\bf i}_{k+1}, \\
0 \ \ & \text{otherwise},
\end{cases} \\[.5ex]
& (\tau_{k+1} \tau_{k} \tau_{k+1}-\tau_{k} \tau_{k+1} \tau_{k}) e_{\bf i}\\
& =\begin{cases} \dfrac{Q_{{\bf i}_{k}, {\bf i}_{k+1}}(y_{k},
y_{k+1}) - Q_{{\bf i}_{k}, {\bf i}_{k+1}}(y_{k+2}, y_{k+1})}
{y_{k} - y_{k+2}}e_{\bf i}\ \ & \text{if} \
{\bf i}_{k} = {\bf i}_{k+2}, \\
0 \ \ & \text{otherwise}.
\end{cases}
\end{aligned}
\end{equation}
\end{theorem}
\begin{remark}
Although the polynomials $Q_{i,j}(u,v)$ are not exactly as they appear in \cite{khovanovlauda}, the reader should not be concerned when we quote results from \cite{khovanovlauda} as all of the arguments go through without change.
The discussion in \cite{kl2} shows that changing the ordering of the quiver $Q$ does not change the isomorphism type of $R(\nu)$.
\end{remark}
The KLR algebras $R(\nu)$ are $\mathbb{Z}$-graded, where $e_{{\bf i}}$ is of degree zero, $y_j e_{\bf i}$ is of degree ${\bf i}_j \cdot {\bf i}_j$ and $\phi_k e_{\bf i}$ is of degree $-{\bf i}_k \cdot {\bf i}_{k+1}$.
They satisfy the property that $R(\nu)_d=0$ for $d$ sufficiently negative (depending on $\nu$) and $R(\nu)_d$ is finite dimensional for all $d$. Relevant implications of these properties are that there are a finite number of isomorphism classes of simple modules and that projective covers exist.
All representations of KLR algebras that we consider will be finitely generated $\mathbb{Z}$-graded representations. If needed, we write $M=\oplus_d M_d$ for the decomposition of a module $M$ into graded pieces. A submodule of a finitely generated $R(\nu)$-module is finitely generated by \cite[Corollary 2.11]{khovanovlauda}.
For a module $M$, we denote its grading shift by $i$ by $q^iM$, this is the module with $(q^i M)_d = M_{d- i}$.
Given two modules $M$ and $N$, we consider $\operatorname{Hom}(M,N)$, and more generally $\operatorname{Ext}^i(M,N)$ as graded vector spaces. All Ext groups which appear in the paper will be taken in the category of $R(\nu)$-modules.
Let $\tau$ be the antiautomorphism of $R(\nu)$ which is the identity on all generators $e_{\bf i},y_i,\phi_j$. For any $R(\nu)$-module $M$, there is a dual module $M^\circledast = \operatorname{Hom}_\mathbb{Q}}\def\partition{\mathcal{P} (M,\mathbb{Q}}\def\partition{\mathcal{P})$, where the $R(\nu)$ action is given by $(x\lambda)(m)=\lambda(\tau(x)m)$ for all $x\in R(\nu)$, $\lambda\in M^\circledast$ and $m\in M$.
For every irreducible $R(\nu)$-module $L$, there is a unique choice of grading shift such that $L^\circledast\cong L$. \cite{khovanovlauda}
Let $\lambda,\mu\in\mathbb{N} I$.
Then there is a natural inclusion $\iota_{\lambda,\mu}:R(\lambda)\otimes R(\mu)\to R(\lambda+\mu)$, defined by $\iota_{\lambda,\mu}(e_{\bf i}\otimes e_{\bf j})=e_{{\bf i}{\bf j}}$, $\iota_{\lambda,\mu}(y_i\otimes 1)=y_i$, $\iota_{\lambda,\mu}(1\otimes y_i)=y_{i+|\lambda|}$, $\iota_{\lambda,\mu}(\phi_i\otimes 1)=\phi_i$, $\iota_{\lambda,\mu}(1\otimes \phi_i)=\phi_{i+|\lambda|}$.
Define the induction functor $\operatorname{Ind}_{\lambda,\mu}:R(\lambda)\otimes R(\mu)\mbox{-mod} \to R(\lambda+\mu)\mbox{-mod}$ by
\[
\operatorname{Ind}_{\lambda,\mu} (M) = R(\lambda+\mu)\bigotimes_{R(\lambda)\otimes R(\mu)} M.
\]
Define the restriction functor $\operatorname{Res}_{\lambda,\mu}:R(\lambda+\mu)\mbox{-mod} \to R(\lambda)\otimes R(\mu)\mbox{-mod}$ by
\[
\operatorname{Res}_{\lambda,\mu}(M) = \iota_{\lambda,\mu}(1_{R(\lambda)\otimes R(\mu)})M.
\]
The induction and restriction functors are both exact.
For a $R(\lambda)$-module $A$ and a $R(\mu)$-module $B$, we write $A\circ B$ for $\operatorname{Ind}_{\lambda,\mu}(A\otimes B)$. Under duality, the behaviour is
\begin{equation}\label{dualofaproduct}
(A\circ B)^\circledast \cong q^{\lambda\cdot\mu} B^\circledast \circ A^\circledast.
\end{equation}
Khovanov and Lauda \cite{khovanovlauda} prove the existence of a dual pair of isomorphisms
\begin{equation}\label{pmod}
\bigoplus_{\nu\in \mathbb{N} I} G_0(R(\nu)\mbox{-pmod}) \cong \mathbf{f}
\end{equation}
and
\begin{equation}\label{fmod}
\bigoplus_{\nu\in \mathbb{N} I} K_0(R(\nu)\mbox{-fmod}) \cong \mathbf{f}^*.
\end{equation}
The category $R(\nu)\mbox{-pmod}$ is the category of finitely generated projective $R(\nu)$-modules and $G_0$ means to take the split Grothendieck group. The category $R(\nu)\mbox{-fmod}$ is the category of finite dimensional $R(\nu)$-modules and $K_0$ means to take the Grothendieck group. We denote the class of a module $M$, identified with its image under the above isomorphisms, by $[M]$.
The action of $q\in \mathcal{A}$ is by grading shift.
The functors of induction and restriction decategorify to a product and coproduct. The isomorphisms above are then isomorphisms of twisted bialgebras.
If $M$ is a general finitely generated $R(\nu)$-module, then it has a well-defined composition series, where each composition factor appears with a multiplicity that is an element of $\mathbb{Z}((q))$. Thus we can consider $[M]$ to be an element of $\mathbf{f}^*_{\mathbb{Z}((q))}$.
Of great importance will be the following Mackey theorem. The general case stated below has the same proof as the special case presented in \cite{khovanovlauda}.
\begin{theorem}\cite[Proposition 2.18]{khovanovlauda}\label{mackey}
Let $\lambda_1,\ldots,\lambda_k,\mu_1\ldots,\mu_l\in\mathbb{N} I$ be such that $\sum_i \lambda_i=\sum_j \mu_j$ and let $M$ be a $R(\lambda_1)\otimes \cdots \otimes R(\lambda_k)$-module. Then the module
$ \operatorname{Res}_{\mu_1,\ldots,\mu_l}\circ\operatorname{Ind}_{\lambda_1,\ldots,\lambda_k}(M)
$ has a filtration indexed by tuples $\nu_{ij}$ satisfying $\lambda_i=\sum_j \nu_{ij}$ and $\mu_j=\sum_i\nu_{ij}$.
The subquotients of this filtration are isomorphic, up to a grading shift, to the composition $\operatorname{Ind}_\nu^\mu\circ \tau\circ \operatorname{Res}_\nu^\lambda(M)$. Here
$\operatorname{Res}_{\nu}^\lambda\map{\otimes_i R(\lambda_i)\mbox{-mod}}{\otimes_i(\otimes_jR(\nu_{ij}))\mbox{-mod}}$ is the tensor product of the $\operatorname{Res}_{\nu_{i\bullet}}$, $\tau\map{\otimes_i(\otimes_jR(\nu_{ij}))\mbox{-mod}}{\otimes_j(\otimes_iR(\nu_{ij}))\mbox{-mod}}$ is given by permuting the tensor factors and $\operatorname{Ind}_\nu^\mu\map{\otimes_j(\otimes_iR(\nu_{ij}))\mbox{-mod}}{\otimes_j R(\mu_j)\mbox{-mod}}$ is the tensor product of the $\operatorname{Ind}_{\nu_{\bullet i}}$.
\end{theorem}
We refer to the filtration appearing in the above theorem as the \emph{Mackey filtration}. It will be very common for us to make arguments using vanishing properties of modules under restriction to greatly restrict the number of these subquotients which can be nonzero.
For each $w\in S_{|\nu|}$, make a choice of a reduced decomposition $w=s_1\ldots s_n$ as a product of simple reflections. Define $\tau_w=\tau_{1}\cdots \tau_{n}$. In general $\tau_w$ depends on the choice of reduced decomposition though this is not the case for permutations of the form $w[\beta,\gamma]$.
\begin{theorem}\cite[Theorem 2.5]{khovanovlauda}\cite[Theorem 3.7]{rouquier}\label{basisofrnu}
The set of elements of the form $y_1^{a_1}\cdots y_{|\nu|}^{a_{|\nu|}}\tau_w e_{\bf i}$ with $a_1,\ldots,a_{|\nu|}\in\mathbb{N}$, $w\in S_{|\nu|}$ and ${\bf i}\in\,\mbox{Seq}\,(\nu)$ is a basis of $R(\nu)$.
\end{theorem}
\begin{theorem}\cite[Corollary 2.11]{khovanovlauda}
The KLR algebra is Noetherian.
\end{theorem}
We work over the ground field $\mathbb{Q}}\def\partition{\mathcal{P}$. It is proved in \cite{khovanovlauda} that any irreducible module is absolutely irreducible, so there is no change to the theory in passing to a field extension.
This also means that any irreducible module for a tensor product of KLR algebras is a tensor product of irreducibles, a fact we use without comment.
\section{Adjunctions
In addition to the induction and restriction functor defined in the previous section, there is also a coinduction functor $\operatorname{CoInd}_{\lambda,\mu}R(\lambda)\otimes R(\mu)\mbox{-mod} \to R(\lambda+\mu)\mbox{-mod} $, defined by
\[
\operatorname{CoInd}_{\lambda,\mu}(M) = \operatorname{Hom}_{R(\lambda)\otimes R(\mu)} (R(\lambda+\mu),M)
\]
where the $R(\lambda+\mu)$ module structure on $\operatorname{CoInd}_{\lambda,\mu}(M)$ is given by $(rf)(t)=f(tr)$ for $f\in\operatorname{CoInd}_{\lambda,\mu}(M)$ and $r,t\in R(\lambda+\mu)$.
The following adjunctions are standard:
\begin{proposition}
The functor $\operatorname{Ind}_{\lambda,\mu}$ is left adjoint to $\operatorname{Res}_{\lambda,\mu}$, while the functor $\operatorname{CoInd}_{\lambda,\mu}$ is right adjoint to $\operatorname{Res}_{\lambda,\mu}$.
\end{proposition}
As a $R(\lambda)\otimes R(\mu)$ module, $R(\lambda+\mu)$ is free of finite rank. This implies that the induction, restriction and coinduction functors all send projective modules to projective modules. As a consequence, there are natural isomorphisms of higher Ext groups
\begin{equation}\label{extind}
\operatorname{Ext}^i(A\circ B, C)\cong \operatorname{Ext}^i(A\otimes B,\operatorname{Res}_{\lambda,\mu} C)
\end{equation}
for all $A\in R(\lambda)\mbox{-mod}$, $B\in R(\mu)\mbox{-mod}$ and $C\in R(\lambda+\mu)\mbox{-mod}$.
Let $\sigma_\nu\map{R(\nu)}{R(\nu)}$ be the involutive isomorphism of $R(\nu)$ with $\sigma_\nu(e_{\bf i})=e_{w_0{\bf i}}$, $\sigma_\nu(y_i)=y_{|\nu|+1-i}$ and $\sigma_\nu(\tau_j e_{\bf i})=(1-2\delta_{{\bf i}_j,{\bf i}_{j+1}})\tau_{|\nu|-j}e_{w_0{\bf i}}$.
\begin{theorem}\cite[Theorem 2.2]{laudavazirani}
There is a natural equivalence of functors
\[
\sigma_{\lambda+\mu}^* \circ \operatorname{Ind}_{\lambda,\mu} \cong q^{(\lambda\cdot\mu)} \operatorname{CoInd}_{\lambda,\mu} \circ (\sigma_\lambda^*\otimes\sigma_\mu^*).
\]
\end{theorem}
\begin{proof}
The statement of this theorem in \cite{laudavazirani} includes a hypothesis that the modules in question are all finite dimensional.
Exactly the same proof works for graded modules all of whose pieces are finite dimensional, which covers all the modules we will ever cover across. The general case follows by writing a module as the direct limit of its finitely generated submodules (noting that $R(\lambda+\mu)$ is finite over $R(\lambda)\otimes R(\mu)$).
\end{proof}
\begin{remark}
Most importantly, applied to a module of the form $A\otimes B$ yields an isomorphism
\begin{equation*}
\operatorname{Ind}_{\lambda,\mu}(A\otimes B) \cong q^{(\lambda\cdot\mu)} \operatorname{CoInd}_{\mu,\lambda} (B\otimes A).
\end{equation*}
In particular, there is an isomorphism
\begin{equation}\label{oppositeadjunction}
\operatorname{Ext}^i(A,B\circ C)\cong q^{-(\lambda\cdot\mu)}\operatorname{Ext}^i( \operatorname{Res}_{\lambda,\mu} A, C\otimes B)
\end{equation}
for any $R(\lambda)$-module $C$, $R(\mu)$-module $B$ and $R(\lambda+\mu)$-module $A$.
\end{remark}
There are parabolic analogues of all the functors and results discussed in this section.
\section{The Ext Bilinear Form
By the decomposition theorem \cite{bbd}, we have
\[
\mathcal{L}\cong\bigoplus_{b\in\mathcal{B}_\nu} L_b\otimes \P_b
\]
where each $L_b$ is a nonzero finite dimensional graded vector space and $\P_b$ is an
irreducible $G_\nu$-equivariant perverse sheaf on $E_\nu$. The indexing set $\mathcal{B}_\nu$ can be taken to be the set of elements of weight $\nu$ in the crystal $\mathcal{\mathcal{B}}(\infty)$, though for our purposes it is not necessary to know this fact.
The maximal semisimple quotient of $R(\nu)$ is $\oplus_{b\in \mathcal{B}_\nu} \operatorname{End}(L_b)$ and hence the simple representations of $R(\nu)$ are the multiplicity spaces $L_b$.
The projective cover of $L_b$ is the module $\oplus_{d\in \mathbb{Z}}\operatorname{Hom}_{D^b_{G_\nu}(E_\nu)}(\mathcal{L},\P_b[d])$.
In this way we get a bijection between simple perverse summands of
$\pi_!\underline{\mathbb{Q}}\def\partition{\mathcal{P}}$ and irreducible representations of $R(\nu)$. By Lusztig's geometric construction of canonical bases, the class of a simple representation under the isomorphism (\ref{fmod}) lies in the dual canonical basis while the class of its projective cover under (\ref{pmod}) lies in the canonical basis.
As has been noted by Kato \cite{kato}, each algebra $R(\nu)$ is graded Morita equivalent to the algebra
\[
A(\nu)=\bigoplus_{d\in\mathbb{Z}}\operatorname{Hom}_{D^b_{G_\nu}(E_\nu)}\left(\bigoplus_{b\in\mathcal{B}_\nu} \P_b,\bigoplus_{b\in\mathcal{B}_\nu} \P_b[d]\right).
\]The algebra $A(\nu)$ is a $\mathbb{N}$-graded algebra with $A(\nu)_0$ semisimple. Under this Morita equivalence the self-dual irreducible module $L_b$ gets sent to a one-dimensional representation of $A(\nu)$ concentrated in degree zero.
\begin{lemma}\label{extwelldefined}
Let $M$ be a finitely generated representation of $R(\nu)$ and let $N$ be a finite dimensional representation of $R(\nu)$. Fix an integer $d$.
Then there exists $i_0$ such that $\operatorname{Ext}^i(M,N)_d=0$ for all $i>i_0$.
\end{lemma}
\begin{proof}
Replace $R(\nu)$ with the Morita equivalent algebra $A(\nu)$ and assume that $M$ and $N$ are $A(\nu)$-modules. Let $\cdots \to P^1\to P^0 \to M\to 0$ be a minimal projective resolution of $M$. As $M$ is finitely generated, there exists $d_0$ such that $M_j=0$ for $j<d_0$. Since $A(\nu)$ is nonnegatively graded with $A(\nu)_0$ semisimple, $P^i_j=0$ for $j<d_0+i$. The vector space $\operatorname{Ext}^i(M,N)$ is a subquotient of $\operatorname{Hom}(P^i,N)$ and for sufficiently large $i$, $\operatorname{Hom}(P^i,N)_d=0$ by degree considerations.
\end{proof}
By the above lemma, if $M$ is a finitely generated $R(\nu)$-module and $N$ is a finite dimensional $R(\nu)$-module, then the infinite sum
\[
( M,N ) = \sum_{i=0}^\infty (-1)^i \dim_q \operatorname{Ext}^i(M,N).
\]
is a well-defined element of $\mathbb{Z}((q))$. We thus get a pairing on Grothendieck groups
\[
( \cdot,\cdot )\map{\mathbf{f}^*_{\mathbb{Z}((q))}\times\mathbf{f}^*}{\mathbb{Z}((q))}.
\]
\begin{lemma}\label{extprops}
The pairing $(\cdot,\cdot)$ satisfies the following properties
\begin{align*}
( f(q)x,g(q) y) &= f(q)g(q^{-1})( x,y ) \\
( \theta_i,\theta_i^* ) &= 1 \\
( xy,z) &=( x\otimes y,r(z)) \\
( x,yz) &= q^{\beta\cdot\gamma}( r(x),z\otimes y)
\end{align*}
for all $x,y,z\in\mathbf{f}$, $f(q)\in\mathbb{Z}((q))$ ,$\mathfrak{g}(q)\in\mathbb{Z}[q,q^{-1}]$,
where $y$ and $z$ are homogeneous of degree $\beta$ and $\gamma$.
\end{lemma}
\begin{proof}
The first formula is obvious. The second is a simple computation in $R(i)\cong k[z]$. The third follows from (\ref{extind}) and the fourth follows from (\ref{oppositeadjunction}).
\end{proof}
Let $\langle x,y\rightarrow=( x,\bar{y})$. The pairing $\langle \cdot,\cdot\rightarrow$ can be extended by $\mathbb{Z}((q))$-linearity to give a bilinear pairing on $\mathbf{f}^*_{\mathbb{Z}((q))}$.
\begin{lemma}\label{extprops2}
The pairing $\langle \cdot,\cdot\rightarrow$ satisfies the following properties
\begin{align*}
\langle f(q)x,g(q)y\rightarrow &=f(q)g(q)\langle x,y\rightarrow \\
\langle\theta_i,\theta_i\rightarrow &=(1-q^2)^{-1} \\
\langle xy,z\rightarrow &=\langle x\otimes y,r(z)\rightarrow \\
\langle x,yz\rightarrow &=\langle r(x),y\otimes z\rightarrow
\end{align*}
\end{lemma}
\begin{proof}
These follow from the analogous formulae in Lemma \ref{extprops}. To derive the third we need to know that $r$ commutes with the bar involution while to derive the fourth we need to know that $\overline{yz}=q^{\beta\cdot\gamma}\bar{z}\bar{y}$ for homogeneous elements $y$ and $z$ of degree $\beta$ and $\gamma$.
\end{proof}
\begin{corollary}
The pairing $\langle \cdot,\cdot\rightarrow $ defined using the Ext-pairing is equal to the usual pairing on $\mathbf{f}$.
\end{corollary}
\begin{proof}
It is immediate that there is a unique pairing satisfying the properties of Lemma \ref{extprops2} and these properties define the pairing in \cite{lusztigbook}.
\end{proof}
\begin{lemma}\label{geometry}
Let $M$ be a finite dimensional $R(\nu)$-module with
\[
[M]=\sum_{i=m}^n \sum_L a_{i,L} q^i [L]
\]
where the second sum is over all self-dual simple modules $L$.
If $a_{n,L}\neq 0$ then $q^n L$ is a submodule of $M$ while if $a_{m,L}\neq 0$ then $q^m L$ is a quotient of $M$.
\end{lemma}
\begin{proof}
If this lemma is false, then there exist self-dual irreducible representations $L_1$ and $L_2$ of $R(\nu)$, and an integer $d\leq 0$ such that $\operatorname{Ext}^1(L_1,L_2)_d\neq 0$. Now replace $R(\nu)$ by the Morita equivalent $A(\nu)$. We compute $\operatorname{Ext}^1(L_1,L_2)$ by computing a minimal projective resolution of $L_1$. Since $A(\nu)$ is non-negatively graded with $A(\nu)_0$ semisimple, we see from this computation that $\operatorname{Ext}^1(L_1,L_2)$ is concentrated in degrees greater than zero.
\end{proof}
\section{Proper Standard Modules}\label{properstandard
\begin{definition}
Let $\alpha$ be a positive root and $n$ be an integer. A representation $L$ of $R(n\alpha)$ is called semicuspidal if
$\operatorname{Res}_{\lambda,\mu}L\neq 0$ implies that $\lambda$ is a sum of roots less than or equal to $\alpha$ and $\mu$ is a sum of roots greater than or equal to $\alpha$.
\end{definition}
\begin{lemma}\label{10.2}
Let $\alpha$ be a positive root, $m_1,\ldots,m_n\in\mathbb{Z}^+$ and $L_i$ be a semicuspidal representation of $R(m_i\alpha)$ for each $i=1,2,\ldots,n$. Then the module $L_1\circ\cdots\circ L_n$ is semicuspidal.
\end{lemma}
\begin{proof}
This immediate from Theorem \ref{mackey} and the definition of semicuspidality.
\end{proof}
\begin{definition}\label{cuspidaldefn}
Let $\alpha$ be a positive root. A representation $L$ of $R(\alpha)$ is called cuspidal if whenever $\operatorname{Res}_{\lambda,\mu}L\neq 0$ and $\lambda,\mu\neq 0$, we have that $\lambda$ is a sum of roots less than $\alpha$ and $\mu$ is a sum of roots greater than $\alpha$.
\end{definition}
\begin{remark}It is clear that if $\alpha$ is an indivisible root then any semicuspidal representation of $R(\alpha)$ is cuspidal. If $\alpha=n\delta$ for $n\geq 2$ we will see in Theorem \ref{19.7} that there are no cuspidal representations of $R(\alpha)$.
\end{remark}
\begin{definition}
A sequence of modules $L_1,\ldots,L_n$ is called \emph{admissible} if each $L_i$ is an irreducible semicuspidal representation of $R(m_i\alpha_i)$ with $m_i\in\mathbb{Z}^+$ and the positive roots $\alpha_i$ satisfying $\alpha_1\succ\alpha_2\succ\cdots\succ\alpha_n$.
\end{definition}
\begin{lemma}\label{restrictinduct}
Let $\alpha_1\succ\alpha_2\succ\cdots \succ \alpha_k$ and $\beta_1\succ\beta_2\succ\cdots\succ \beta_l$ be positive roots and $m_1,\ldots,m_k$, $n_1,\ldots,n_l$ be positive integers. Let $L_1,\ldots,L_k$ be semicuspidal representations of $R(m_1\alpha_1),\cdots,R(m_k\alpha_k)$ respectively. Then
\[
\operatorname{Res}_{n_1\beta_1,\ldots,n_l\beta_l} L_1\circ \cdots \circ L_k = \begin{cases}
0 &\mbox{unless } \underline{\beta}\leq\underline{\alpha} \\
L_1\otimes \cdots\otimes L_k &\mbox{if } \underline{\beta}=\underline{\alpha},
\end{cases}
\] where we are considering bilexicographical ordering on the multisets $\underline{\alpha}$ and $\underline{\beta}$.
\end{lemma}
\begin{proof}
Consider a nonzero layer of the Mackey filtration for $\operatorname{Res}_{n_1\beta_1,\ldots,n_l\beta_l} L_1\circ\cdots\circ L_n $. It is indexed by a set of elements $\nu_{ij}\in \mathbb{N} I$ such that $m_i\alpha_i=\sum_j \nu_{ij}$ and $n_j\beta_j = \sum_{i}\nu_{ij}$. For the piece of the filtration to be nonzero, it must be that $\operatorname{Res}_{\nu_{i,1},\ldots,\nu_{i,n}}L_i \neq 0$ for each $i$.
Suppose that $t$ is an index such that $m_i\alpha_i=n_i\beta_i$ for $i<t$. We will prove that in order for us to have a nonzero piece of the filtration, it must be that either $\beta_t\prec \alpha_t$ or $m_t\alpha_t=n_t\beta_t=\nu_{t,t}$.
By induction on $t$, we may assume that $\nu_{ii}=m_i\alpha_i=n_i\beta_i$ for $i<t$. Therefore $\nu_{i,j}=0$ for all $i$ and $j$ with $i\geq t$ and $j<t$.
Suppose $i\geq t$. Since the module $L_i$ is cuspidal, this implies that $\nu_{i,t}$ is a sum of roots less than or equal to $\alpha_i$, which are all less than or equal to $\alpha_t$.
Now $n_t\beta_t=\sum_{i\geq t}\nu_{i,t}$ is written as a sum of positive roots all less than or equal to $\alpha_t$. Therefore, by convexity of the ordering, either $\beta_t\prec \alpha_t$ or $n_t\beta_t=m_t\alpha_t$. In this latter case, equality in our inequalities must hold everywhere, hence $\nu_{t,t}=n_t\beta_t$ as required.
This is enough to conclude that $\underline{\alpha}\geq \underline{\beta}$ under lexicographical ordering.
\end{proof}
\begin{lemma}\label{irreduciblehead}
Let $\alpha_1\succ \alpha_2\succ\cdots\succ \alpha_n$ be roots, $m_1,\ldots,m_n$ be positive integers and $L_1,\ldots,L_n$ be irreducible semicuspidal representations of $R(m_1\alpha_1),\ldots,R(m_n\alpha_n)$ respectively.
Then \begin{enumerate}
\item the module $L_1\circ \cdots \circ L_n$ has a unique irreducible quotient $L$, and
\item $\operatorname{Res}_{m_1\alpha_1,\ldots,m_n\alpha_n} L_1\circ\cdots\circ L_n = \operatorname{Res}_{m_1\alpha_1,\ldots,m_n\alpha_n}L=L_1\otimes\cdots\otimes L_n$.
\end{enumerate}
\end{lemma}
\begin{proof}
Suppose that $Q$ is a nonzero quotient of $L_1\circ\cdots\circ L_n$. Then by adjunction there is a nonzero map from $L_1\otimes\cdots\otimes L_n$ to $\operatorname{Res} Q$. As $L_1\otimes \cdots\otimes L_n$ is irreducible, this map is injective.
The restriction functor is exact and by Lemma \ref{restrictinduct}, $\operatorname{Res} (L_1\circ\cdots\circ L_n)$ is simple. Therefore the head of $L_1\circ \cdots \circ L_n$ must be simple.
\end{proof}
If $L_1,\ldots,L_n$ is a sequence of representations, we define $A(L_1,\ldots,L_n)=\operatorname{cosoc}(L_1\circ\cdots L_n)$.
\begin{theorem}\label{fdmain}
Every irreducible module for $R(\nu)$ is of the form $A(L_1,\ldots,L_n)$ for exactly one set of irreducible semicuspidal representations $L_1,\ldots,L_n$ of $R(m_1\alpha_1),\ldots,R(m_n\alpha_n)$ respectively, where $\alpha_1\succ\cdots\succ\alpha_n$ are positive roots.
\end{theorem}
\begin{theorem}\label{numberofimaginarysemicuspidals}
If $\alpha$ is a positive real root and $n$ is a positive integer, there is one simple semicuspidal module for $R(n\alpha)$. For the imaginary roots, let $f(n)$ be the number of simple semicuspidal representations of $R(n\delta)$ (and set $f(0)=1$). Then
\[
\sum_{n=0}^{\infty} f(n) t^n = \prod_{i=1}^\infty (1-t^i)^{1-|I|}.
\]
\end{theorem}
\begin{proof}
We prove the above two theorems by a simultaneous induction on $\nu$.
First let us consider the case where $\nu$ is not of the form $n\alpha$ for some root $\alpha$. The number of irreducible representations of $R(\nu)$ is equal to $\dim{\mathbf{f}_\nu}$, which is the coefficient of $t^\nu$ in the power series (\ref{dimfnu}).
By inductive hypothesis applied to Theorem \ref{numberofimaginarysemicuspidals}, the number of admissible sequences of semicuspidal modules $(L_1,\ldots,L_n)$ is equal to $\dim\mathbf{f}_\nu$. By Lemma \ref{irreduciblehead}, each of the modules $A(L_1,\ldots,L_n)$ are irreducible, and by applying various restriction functors, we see via Lemma \ref{restrictinduct} that these modules are all distinct. Therefore we have identified all of the irreducible $R(\nu)$-modules in this case, proving Theorem \ref{fdmain}.
Now we turn our attention to the case where $\nu=k\alpha$ for some root $\alpha$. By the same arguments as in the previous case, the modules of the form $A(L_1,\ldots,L_n)$ where $n\geq 2$ yield all the irreducible modules for $R(k\alpha)$ except one, unless $\nu=n\delta$, when the construction yields all irreducible modules except $f(n)$. It suffices to prove that if $L$ is an irreducible representation of $R(\nu)$ with $L$ not of the form $A(L_1,\ldots,L_n)$ with $n\geq 2$, then $L$ is semicuspidal.
Suppose that $\lambda$ and $\mu$ are such that $\operatorname{Res}_{\lambda\mu}L\neq 0$. We need to prove that $\lambda$ is a sum of roots less than or equal to $\alpha$ (the result for $\mu$ is similar) and we may suppose that neither of $\lambda$ and $\mu$ is zero. Let $L_\lambda\otimes L_\mu$ be an irreducible submodule of $\operatorname{Res}_{\lambda\mu}L$. By inductive hypothesis $L_\lambda=A(L_1,\ldots,L_k)$ for some admissible sequence of semicuspidal representations. Suppose that $L_1$ is a $R(m\beta)$ module where $\beta$ is a root. Then $\operatorname{Res}_{m\beta,\nu-m\beta}L\neq 0$.
If $\beta\preceq\alpha$, then $\lambda$ is a sum of roots less than or equal to $\beta$ and hence a sum of roots less than or equal to $\alpha$.
Therefore without loss of generality we may assume that $\lambda=m\beta$ and that $L_\lambda$ is semicuspidal. For want of a contradiction, assume $\beta\succ \alpha$. We may further assume without loss of generality that $\beta$ is the maximal root for which $\operatorname{Res}_{m\beta,\nu-m\beta}L\neq 0$ for some positive integer $m$. We may further assume that $m$ is as large as possible.
By inductive hypothesis, write $L_\mu=A(M_1,\ldots,M_n)$ where $M_1$ is a $R(k\gamma)$-module for some root $\gamma$ and positive integer $k$.
Therefore $\operatorname{Res}_{\lambda+k\gamma,\mu-k\gamma}L\neq 0$. If $k\gamma\neq \mu$, then by maximality of $\beta$, $\lambda+k\gamma$ is a sum of roots less than or equal to $\beta$. By maximality of $m$, $\gamma \prec \beta$. By adjunction this implies that $L$ is a quotient of $L_\lambda\circ M_1\circ\cdots \circ M_n$. As $(L_\lambda,M_1,\ldots,M_n)$ is an admissible sequence of semicuspidal modules, this is a contradiction.
Therefore $L_\mu$ is semicuspidal, with $\mu=k\gamma$. By convexity $\gamma\prec\alpha\prec\beta$.
By adjunction there is a nonzero map from $L_\lambda\circ L_\mu$ to $L$. As $L$ is irreducible, this exhibits $L$ as $A(L_\lambda,L_\mu)$, a contradiction. This completes the proof.
\end{proof}
\section{Real Cuspidals
For $i\in I$, there is an automorphism $T_i$ of the entire quantum group $U_q(\mathfrak{g})$ satisfying
\[
T_i \theta_j = \sum_{k=0}^{-i\cdot j} (-q)^k \theta_i^{(k)} \theta_j \theta_i^{(-i\cdot j-k)}
\]
for all $i\neq j$.
In the notation of \cite{lusztigbook}, $T_i$ is the automorphism $T'_{i,+}$.
Now we will define the PBW root vectors for the real roots. Let $\alpha$ be a positive real root and suppose that $\alpha\prec\delta$. Let $S_\alpha=\{\beta\in \Phi^+\mid
\alpha-\beta\in \mathbb{N} I \}$. Then $S_\alpha$ is a finite set of roots. By Theorem \ref{3.11}, we can find a word convex order $\prec'$ whose restriction to $S_\alpha$ agrees with the restriction of $\prec$ to $S_\alpha$.
By Lemma \ref{finiteinitial} there exists $w\in W$ such that $\Phi(w)=\{\beta\in\Phi^+\mid \beta\preceq'\alpha\}$ and a reduced expression $w=s_{i_1}\ldots s_{i_N}$ such that $\alpha=s_{i_1}s_{i_2}\cdots s_{i_{N-1}} \alpha_{i_N}$. We define the root vector $E_\alpha\in\mathbf{f}$ by
\[
E_\alpha=T_{i_1}T_{i_2}\cdots T_{i_{N-1}} \theta_{i_N}
\]
If $\alpha$ happens to be greater than $\delta$, then in a similar vein we get a reduced expression but now define $E_\alpha\in\mathbf{f}$ by
\[
E_\alpha=T_{i_1}^{-1}T_{i_2}^{-1}\cdots T_{i_{N-1}}^{-1} \theta_{i_N}
\]
In all cases, we then define the dual root vector $E_\alpha^*=(1-q_\alpha^2)E_\alpha\in\mathbf{f}^*$.
A proof that the elements $E_\alpha$ and $E_\alpha^*$ are well defined based on \cite[Proposition 40.2.1]{lusztigbook} is possible. Alternatively, this result will follow from Theorem \ref{realdualpbw}.
For $\alpha\in\Phi^+_{\re}$, let $L(\alpha)$ be the unique self-dual cuspidal irreducible representation of $R(\alpha)$. The existence of a cuspidal irreducible module is Theorem \ref{numberofimaginarysemicuspidals} above while the fact that it can be chosen to be self-dual is in \cite[\S 3.2]{khovanovlauda}.
\begin{theorem}\label{realdualpbw}
Let $\alpha$ be a positive real root. Then $[L(\alpha)]=E_\alpha^*$.
\end{theorem}
\begin{proof}
Let $i_1,\ldots,i_N$ be as in the construction of $E_\alpha$ above. For $1\leq k\leq N$ let $\alpha_k=s_{i_1}\cdots s_{i_{k-1}} \alpha_{i_k}$. Then $\alpha_1\prec\cdots\prec\alpha_N=\alpha$.
First we will prove by induction on $n$ for $1\leq n\leq N$ that there exists $x_n\in \mathbf{f}^*_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$ such that $[L(\alpha)]=T_{i_1}\cdots T_{i_{n-1}}(x_{n})$.
For the case $n=1$, let $x_1=[L(\alpha)]$.
Now assume that the result is known for $n=k$ and consider the case $n=k+1$.
By \cite[Ch 38]{lusztigbook} we can write $x_k=\theta_{i_k}^* y + T_{i_k}(z)$ where $y,z\in \mathbf{f}^*_{\mathbb{Q}}\def\partition{\mathcal{P}(q)}$. Then
\[
[L(\alpha)]=(T_{i_1}\cdots T_{i_{k-1}}(\theta_{i_k}^*)) (T_{i_1}\cdots T_{i_{k-1}}(y)) + T_{i_1}\cdots T_{i_k}(z).
\]
Since $L(\alpha)$ is cuspidal and $\alpha_k\prec \alpha$, $\operatorname{Res}_{\alpha_k,\alpha_N-\alpha_k}L(\alpha)=0$. Therefore $[L(\alpha)]$ is orthogonal to the product $(T_{i_1}\cdots T_{i_{k-1}}(\theta_{i_k}^*)) (T_{i_1}\cdots T_{i_{k-1}}(y))$. Since this product is orthogonal to $T_{i_1}\cdots T_{i_k}(z)$, it must be that $(y,y)=0$.
If $\mathbb{Q}}\def\partition{\mathcal{P}(q)$ is embedded into $\mathbb{R}$ by sending $q$ to a sufficiently small real number, then the form $(\cdot,\cdot)$ on $\mathbf{f}_\nu$ is positive definite. Therefore $y=0$. We let $x_{k+1}=z$.
We have now proved the desired preliminary result by the principle of mathematical induction. Applying this when $k=N$, we see that $[L(\alpha)]=T_{i_1}\cdots T_{i_{n-1}}(x_{N})$ for some $x_N\in (\mathbf{f}^*_{\mathbb{Q}}\def\partition{\mathcal{P}(q)})_{\alpha_{i_N}}$. Therefore $x_N$ is a scalar multiple of $\theta_{i_N}$.
Since $[L(\alpha)]$ is an element of a $\mathbb{Z}[q,q^{-1}]$-basis of $\mathbf{f}^*$, the scalar must be a unit in $\mathbb{Z}[q,q^{-1}]$, thus of the form $\pm q^i$ for some $i\in \mathbb{Z}$.
Since $[L(\alpha)]$ is
invariant under the bar involution, $i=0$. The argument of \cite[Lemma 2.35]{kleshchev} shows that the sign is the positive one.
\end{proof}
\begin{proposition}
Let $\alpha$ be a real root and $n$ be a positive integer. The module $L(\alpha)^{\circ n}$ is the unique simple semicuspidal representation of $R(n\delta)$.
\end{proposition}
\begin{proof}
By Lemma \ref{10.2}, $L(\alpha)^{\circ n}$ is semicuspidal. Therefore $[L(\alpha)^{\circ n}]=f(q)[L]$ where $L$ is the unique semicuspidal representation of $R(n\alpha)$ and $f(q)\in\mathbb{N}[q,q^{-1}]$. By Theorem \ref{realdualpbw}, $[L(\alpha)^{\circ n}]=T_{i_1}T_{i_2}\cdots T_{i_{N-1}} (\theta_{i_N}^*)^n$ which is indivisible in $\mathbf{f}^*$, hence $L(\alpha)^{\circ n}$ is irreducible.
\end{proof}
\begin{remark}
This gives the existence of many modules called \emph{real} in the nomenclature of \cite{kkko}.
\end{remark}
\section{Root Partitions
Let $S$ be an indexing set for the set of self-dual irreducible representations of $R(n\delta)$, for all $n$. It will not be until Theorem \ref{19.7} that we exhibit a bijection between $S$ and $\partition^\Omega$. We write $L(s)$ for the representation indexed by $s\in S$.
Define a root partition $\pi$ to be an admissible sequence of self-dual irreducible semicuspidal representations.
To each root partition $\pi$ we define a function $f_\p\map{\Phi^+_{nd}}{\mathbb{N}}$ where if $f_\p(\alpha)$ is nonzero then there is a representation of $R(f_\p(\alpha)\alpha)$ in $\p$. Given two root partitions $\pi$ and $\sigma$ we say that $\pi<\sigma$ if there exist indivisible roots $\alpha$ and $\alpha'$ such that $f_\p(\alpha)<f_\sigma(\alpha)$, $f_\p(\alpha')<f_\sigma(\alpha')$ and $f_\p(\beta)=f_\sigma(\beta)$ for all roots $\beta$ satisfying either $\beta\prec\alpha$ or $\beta\succ\alpha'$. If $f_\pi=f_\sigma$ we say $\p\sim\sigma$.
Since there is exactly one irreducible semicuspidal representation of $R(n\alpha)$ for each $n$ and each real root $\alpha$, we can write the datum of a root partition in a more combinatorial manner.
Concretely we write a root partition in the form $\pi=(\beta_1^{m_1},\ldots,\beta_k^{m_k},s,\gamma_l^{n_l},\ldots,\gamma_1^{n_1})$. Here $k$ and $l$ are natural numbers, $s\in S$, $\beta_1,\ldots,\beta_k,\gamma_1,\ldots,\gamma_l$ are the set of real roots on which $f_\p$ is nonzero, $f_\pi(\beta_i)=m_i$, $f_\pi(\gamma_i)=n_i$ and
\[
\beta_1\succ\cdots\succ\beta_k \succ\delta\succ \gamma_l\succ\cdots\succ\gamma_1
\]
When we do have a bijection between $S$ and $\partition^\Omega$ then we will have a purely combinatorial description of a root partition.
Let $\pi=(\beta_1^{m_1},\ldots,\beta_k^{m_k},s,\gamma_l^{n_l},\ldots,\gamma_1^{n_1})$ be a root partition. Let $s_\lambda=\sum_{i=1}^k{\binom{m_i}{2}}+\sum_{j=1}^l\binom{n_j}{2}$. Define the proper standard module $\overline{\Delta}(\pi)$ to be
\[
\overline{\Delta}(\pi)=q^{s_\lambda}L(\beta_1)^{\circ m_1}\circ \cdots\circ L(\beta_k)^{\circ m_k} \circ L(s)\circ L(\gamma_l)^{\circ n_l} \circ \cdots \circ (\gamma_1)^{\circ n_1}.
\]
Let $L(\pi)$ be the head of $\overline{\Delta}(\pi)$. This is an irreducible module by Lemma \ref{irreduciblehead}.
\begin{theorem}\label{main}\cite{kleshchev}
The proper standard modules have the following property.
\begin{enumerate}
\item Up to isomorphism and grading shift, the set $\{L(\pi)\}$ as $\pi$ runs over all root partitions of $\nu$ is a complete and irredundant set of irreducible $R(\nu)$-modules.
\item The module $L(\pi)$ is self dual, i.e. $L(\pi)^\circledast \cong L(\pi)$.
\item If the multiplicity $[\overline{\Delta}(\pi):L(\sigma)]$ is nonzero, then $\sigma\leq \pi$. Furthermore $[\overline{\Delta}(\pi):L(\pi)]=1$.
\end{enumerate}
\end{theorem}
\begin{proof}
Part (1) is Theorem \ref{fdmain}. For part (2) note that since $L(\p)$ is irreducible, by \cite[\S 3.2]{khovanovlauda} $L(\pi)^\circledast\cong q^i L(\pi)$ for some $i$. By Lemma \ref{irreduciblehead}(2) and the fact that restriction commutes with duality, $i=0$. Part (3) follows from Lemma \ref{restrictinduct}.
\end{proof}
\section{Levendorskii-Soibelman Formula}\label{lsformula
By Theorem \ref{main} the classes $[\overline{\Delta}(\pi)]$ of the proper standard modules is a basis of $\mathbf{f}^*$. We call this the categorical dual PBW basis. Let $\{E_\pi\}$ be the basis of $\mathbf{f}$ dual to this with respect to $\langle\cdot,\cdot\rightarrow$. We shall call this basis the categorical PBW basis. Later we will identify the categorical PBW basis both with a basis coming from a family of standard modules, as well as an algebraically defined basis which generalises the approach of \cite{beck}.
The results in this section are an affine type analogue of the Levendorskii-Soibelman formula \cite[Proposition 5.5.2]{ls}. We refer to both Theorems \ref{ls} and \ref{lsf} as a Levendorskii-Soibelman formula.
\begin{theorem}\label{ls}
Let $\theta,\psi\in\Phi^+_{\re} \cup S$ with $\theta \succ\psi$. Expand $[L(\theta)][L(\psi)]-q^{(\theta\cdot\psi)}[L(\psi)][L(\theta)]$ in the standard basis
\[
[L(\theta)][L(\psi)]-q^{(\theta\cdot\psi)}[L(\psi)][L(\theta)]=\sum_\pi c_\pi [\overline{\Delta}(\pi)].
\]
If $c_\pi\neq 0$ for some root partition $\pi$ then $\pi< (\theta,\psi)$.
\end{theorem}
\begin{proof}
By Theorem \ref{main},
\[
[L(\theta)][L(\psi)]-[L(\theta,\psi)]\in\sum_{\pi<(\theta,\psi)}\mathbb{Z}[q,q^{-1}] [L(\pi)].
\]
Applying the bar involution on $\mathbf{f}^*$ yields
\[
q^{(\theta\cdot\psi)}[L(\psi)][L(\theta)]-[L(\theta,\psi)]\in \sum_{\pi<(\theta,\psi)}\mathbb{Z}[q,q^{-1}] [L(\pi)].
\]
Theorem \ref{main} also shows that
\[
\sum_{\pi<(\theta,\psi)}\mathbb{Z}[q,q^{-1}] [L(\pi)]=\sum_{\pi<(\theta,\psi)}\mathbb{Z}[q,q^{-1}] [\bar\Delta(\pi)]
\]
so upon subtraction we obtain the desired result.
\end{proof}
\begin{lemma}\label{orthog}
Let $\sigma$ and $\pi$ be two root partitions. Then $\langle [\overline{\Delta}(\sigma)],[\overline{\Delta}(\pi)]\rangle = 0$ unless $\sigma\sim\pi$.
\end{lemma}
\begin{proof}
We have $\langle [\overline{\Delta}(\sigma)],[\overline{\Delta}(\pi)]\rangle = \langle [L(\sigma_1)\otimes \cdots \otimes L(\sigma_l)],[\operatorname{Res}_\sigma\overline{\Delta}(\pi)]\rangle$ which, by Lemma \ref{restrictinduct}, is zero unless $\sigma\leq \pi$.
Also $\langle [\overline{\Delta}(\sigma)],[\overline{\Delta}(\pi)]\rangle=\langle [\operatorname{Res}_\pi(\overline{\Delta}(\sigma)],[L(\pi_1)\circ \cdots \circ L(\pi_k)]\rangle$ which again by Lemma \ref{restrictinduct} is zero unless $\pi\leq \sigma$.
\end{proof}
\begin{lemma}\label{thsimpsi}
If $\theta\in\Phi^+_\re\cup S$ then
\[
E_\theta \in \operatorname{span}_{\psi\sim \theta} [L(\psi)].
\]
\end{lemma}
\begin{proof}
By definition of $E_\theta$, we have $\langle E_\theta,[\overline{\Delta}(\pi)]\rangle = 0$ unless $\pi=\theta$. By Lemma \ref{orthog} and the fact that the classes $[\overline{\Delta}(\sigma)]$ are a basis of $\mathbf{f}^*$, this forces
\[
E_\theta \in \operatorname{span}_{\psi\sim \theta} [\overline{\Delta}(\psi)]= \operatorname{span}_{\psi\sim \theta} [L(\psi)].
\]\end{proof}
\begin{corollary}\label{improd}
If $\sigma,\pi\in S$ then $E_\sigma E_\pi$ is a linear combination of $E_\tau$ for $\tau\in S$.
\end{corollary}
\begin{proof}
This follows from Lemmas \ref{10.2} and \ref{thsimpsi}.
\end{proof}
\begin{lemma}
Let $\pi=(\pi_1,\ldots,\pi_k)$ be a root partition. Then $E_\pi=E_{\pi_1}\cdots E_{\pi_k}$.
\end{lemma}
\begin{proof}
By Lemma \ref{thsimpsi}, the element $E_{\pi_1}\cdots E_{\p_k}$ is a linear combination of elements of the form $[\overline{\Delta}(\sigma)]$ where $\sigma\sim \pi$. Therefore by Lemma \ref{orthog}, $E_{\pi_1}\cdots E_{\p_k}$ is orthogonal to all elements of the from $[\overline{\Delta}(\eta)]$ where $\eta\not\sim\pi$. For $\eta\sim\pi$, we compute
\[
\langle E_{\p_1}\cdots E_{\p_k} , [\overline{\Delta}(\eta)]\rangle = \prod_{i=1}^k \langle E_{\p_i},[L(\eta_i)]\rangle
\]
which is zero unless $\eta=\pi$ in which case it is equal to one. We have shown that the product $E_{\p_1}\cdots E_{\p_k}$ has all the properties which define $E_\pi$, hence is equal to $E_\p$.
\end{proof}
\begin{theorem}\label{lsf}
Let $\theta,\psi\in\Phi^+_{\re} \cup S$ with $\theta \succ\psi$. Then
\[
E_\theta E_\psi - q^{(\theta\cdot\psi)} E_\psi E_\theta \in \sum_{\pi<(\theta,\psi)} \mathbb{Z}[q,q^{-1}]E_\p.
\]
\end{theorem}
\begin{proof}
This is immediate from Theorem \ref{ls} and Lemma \ref{thsimpsi}.
\end{proof}
This yields an algorithm for expanding any monomial in the $E_\theta$ in the PBW basis. Namely given a monomial $E_{\kappa_1}E_{\kappa_2}\cdots E_{\kappa_k}$, repeatedly apply the following types of moves:
\begin{itemize}
\item
If $\kappa_l \prec \kappa_{l+1}$, replace $E_{\kappa_l}E_{\kappa_{l+1}}$ with $q^{\kappa_l\cdot\kappa_{l+1}}E_{\kappa_{l+1}}E_{\kappa_{l}}$ plus the correction term from Theorem \ref{lsf}.
\item If $\kappa_l$ and $\kappa_{l+1}$ are both imaginary replace the product $E_{\kappa_l}E_{\kappa_{l+1}}$ with a sum of terms $E_\theta$, which is possible by Corollary \ref{improd}.
\end{itemize}
\section{Minimal Pairs
Let $\alpha$ be a positive root. Define $S(\alpha)$ to be the quotient of $R(\alpha)$ by the two-sided ideal generated by the set of $e_{\bf i}$ such that $e_{\bf i} L=0$ for all semicuspidal modules $L$.
\begin{lemma}
There is an equivalence of categories between the category of $S(\alpha)$ modules and the full subcategory of semicuspidal $R(\alpha)$-modules.
\end{lemma}
\begin{proof}
It is clear from the definition that any semicuspidal $R(\alpha)$-module is a $S(\alpha)$-module.
Conversely suppose that $M$ is a $S(\alpha)$-module. Suppose that $\lambda=(\lambda_1,\ldots,\lambda_l)$ is a root partition such that $q^n L(\lambda)$ appears as a subquotient of $M$. Then $e_{\bf i} M\neq 0$ for some ${\bf i}$ which is the concatenation of ${\bf i}_1,\ldots,{\bf i}_l$ in $\,\mbox{Seq}\,(\lambda_1),\ldots,\,\mbox{Seq}\,(\lambda_l)$ respectively. If $\lambda\neq \alpha$, then $e_{\bf i} L=0$ for all semicuspidal $R(\alpha)$-modules $L$. Therefore $e_{\bf i}$ has zero image in $S(\alpha)$, contradicting $e_{\bf i} M\neq 0$. Hence all composition factors of $M$ are semicuspidal, so $M$ is semicuspidal.
\end{proof}
\begin{definition}
Let $\alpha$ be a positive root. A minimal pair for $\alpha$ is an ordered pair of roots $(\beta,\gamma)$ satisfying $\alpha=\beta+\gamma$, $\gamma\prec\beta$ and there is no pair of roots $(\beta',\gamma')$ satisfying $\alpha=\beta'+\gamma'$ and $\gamma\prec\gamma'\prec\beta'\prec\beta$.
\end{definition}
\begin{lemma}\label{oppositerestrict}
Let $\alpha$ be a positive root and let $(\beta,\gamma)$ be a minimal pair for $\alpha$. Let $L$ be a cuspidal representation of $R(\alpha)$. Then $\operatorname{Res}_{\gamma,\beta}L$ is a $S(\gamma)\otimes S(\beta)$-module.
\end{lemma}
\begin{proof}
Expand $[\operatorname{Res}_{\gamma,\beta}L]$ in the categorical dual PBW basis
\[
[\operatorname{Res}_{\gamma,\beta}L]= \sum_{\pi,\sigma} c_{\pi\sigma}E_\pi^* E_\sigma^*.
\]
Then
\[
c_{\pi\sigma}=\langle E_\pi\otimes E_\sigma,[\operatorname{Res}_{\gamma,\beta}L]\rangle =\langle E_\pi E_\sigma , [L]\rangle.
\]
In the previous section we showed how the Levendorskii-Soibelman formula gave an algorithm for expanding the product $E_\pi E_\sigma$ into the PBW basis. Each term $E_{\kappa_1}\cdots E_{\kappa_n}$ which appears at some point in this expansion has $\kappa_1\succeq\pi_1\succeq \beta$ and $\kappa_n\preceq \sigma_l\preceq \gamma$.
The only PBW basis elements which fail to be orthogonal to $[L]$ are those of the form $E_\alpha$ if $\alpha$ is real and $E_s$ with $s\in S$ if $\alpha$ is imaginary. For such a term to appear, it must arise as a result of applying Theorem \ref{lsf} to a term $E_{\kappa_1}E_{\kappa_2}$ with $\kappa_1+\kappa_2=\alpha$.
We have already showed that $\kappa_1\succeq \beta$ and $\gamma_2\preceq \gamma$. To apply the Levendorskii-Soibelman formula we need $\kappa_1\prec \kappa_2$ and we also know $\kappa_1+\kappa_2=\alpha$. Since $(\beta,\gamma)$ is a minimal pair, this forces $\kappa_1=\beta$ and $\kappa_2=\gamma$. Therefore the coefficient $c_{\pi\sigma}$ can only be nonzero if $\pi=\kappa_1$ and $\sigma=\kappa_2$. Hence $[\operatorname{Res}_{\gamma\beta}L]$ is a linear combination of elements of the form $[L_\gamma]\otimes [L_\beta]$ where $L_\gamma$ and $L_\beta$ are cuspidal representations of $R(\gamma)$ and $R(\beta)$. This implies that $\operatorname{Res}_{\gamma\beta}L$ is a $S(\gamma)\otimes S(\beta)$-module, as required.
\end{proof}
A chamber coweight $\w$ is said to be \emph{adapted} to the convex order $\prec$ if it is a fundamental coweight for the positive system $p(\Phi_{\succ\delta})$ in $\Phi_f$. Let $\w$ be such a chamber coweight. Then there exists a root $\alpha\in p(\Phi_{\succ\delta})$ such that $\langle\w,\alpha\rangle=1$ and $\langle\w,\beta\rangle=0$ for all $\beta\in p(\Phi_{\succ\delta})\setminus\{\alpha\}$. Let $\w_+=\tilde{\alpha}$ and $\w_-=\widetilde{-\alpha}$.
We will always assume that all chamber coweights are adapted to the given convex order.
\begin{lemma}
Let $\alpha$ be a positive real root which is not simple that does not have a real minimal pair. Then there exists a chamber coweight $\om$ adapted to $\prec$ such that $\alpha=\om_++n\delta$ or $\alpha=\om_-+n\delta$ for some $n\in\mathbb{N}$.
\end{lemma}
\begin{proof}
Since every root which is not simple has a minimal pair, if $\alpha$ has no real minimal pair it must be that $\alpha-\delta$ is also a root.
Without loss of generality suppose $\alpha\succ \delta$. The $p(\alpha)$ is a positive root in $\Phi_f$. We have to prove that $p(\alpha)$ is simple. Suppose for want of a contradiction that $p(\alpha)=\beta+\gamma$ for two positive roots $\beta$ and $\gamma$. Then $\alpha=\tilde{\beta}+\tilde{\gamma}+n\delta$ for some $n$. If $n\geq 0$ then we can use this expression to write $\alpha$ as a sum of two roots both greater than $\delta$, which proves that $\alpha$ has a real minimal pair.
Therefore the only case left to consider is if $\tilde{\beta}+\tilde{\gamma}-2\delta$ is a positive root. When writing $\tilde{\beta}+\tilde{\gamma}$ in the form $n\delta+x$ with $x\in\Phi_f$, $n\geq 2$ with equality if and only if $\beta$ and $\gamma$ are negative under the usual positive system on $\Phi_f$. Therefore it is impossible for $\tilde{\beta}+\tilde{\gamma}-2\delta$ to be a root.
\end{proof}
\section{Independence of Convex Order
In this section, we prove some results detailing how some modules which a priori depend on the entire convex order $\prec$, only depend on the positive system $p(\Phi_{\prec \delta})$.
\begin{theorem}\label{independence}
Let $\w$ be a chamber coweight.
The algebras $S(\omega_+)$ and $S(\om_-)$ only depend on the set $p(\Phi_{\prec \delta})$.
\end{theorem}
\begin{remark}
For a balanced convex order, this is \cite[Lemma 5.2]{kleshchev}.
\end{remark}
\begin{proof}
It suffices to prove that the simple modules $L(\w_-)$ and $L(\w_+)$ depend only on $p(\Phi_{\prec \delta})$.
We write $E_\alpha^\prec$ for the root vector defined using the convex order $\prec$. Let $\prec$ and $\prec'$ be two convex orders with $p(\Phi_{\prec \delta})=p(\Phi_{\prec' \delta})$. Without loss of generality we may assume that $\prec$ and $\prec'$ are of word type.
Label the roots smaller than $\delta$ as $\alpha_1\prec\alpha_2\prec\cdots$ and $\alpha_1'\prec'\alpha_2'\prec'\cdots$. Let $n$ and $N$ be such that
\[
\w_-\in\{\alpha_1,\ldots,\alpha_n\}\subset \{\alpha_1',\ldots,\alpha_N'\}
\]
Let $w$ be the element of $W$ such that $\Phi(w)=\{\alpha_1,\ldots,\alpha_n\}$ and let $u\in W$ be such that $\Phi(u)=\{\alpha_1',\ldots,\alpha_N'\}$. Then $\Phi(w)\subset \Phi(u)$.
Hence if we fix a reduced expression for $w$ (in particular the one used to define $E_{\w_-}^\prec$) then there exists a reduced expression for $u$ beginning with this fixed reduced expression for $w$.
By \cite[Prop 40.2.1]{lusztigbook} there exists a subspace $U^+(u)$ of $\mathbf{f}$ which contains $E_\alpha^\prec$ and $E_\alpha^{\prec'}$. The dimension of $U^+(u)_{\w_-}$ is equal to the number of ways of writing $\w_-$ as a $\mathbb{N}$-linear combination of roots in $\Phi^+(u)$. Any nontrivial expression contradicts the simplicity of $p(\w_-)$, hence this space is one-dimensional, so $E_\alpha^\prec$ and $E_\alpha^{\prec'}$ are scalar multiples of one another.
By Theorem \ref{realdualpbw}, $(1-q^2)E_{\w_-}$ is the character of the irreducible module $L(\w_-)$, hence this scalar must be one and the module $L(\w_-)$ is the same for the convex orders $\prec$ and $\prec'$. This completes the proof for $\w_-$ and the proof for $\w_+$ is similar.
\end{proof}
\begin{lemma}\label{strongminimalpair}
Let $(\beta,\gamma)$ be a minimal pair for $\delta$. Let $L_\beta$ and $L_\gamma$ be cuspidal $R(\beta)$ and $R(\gamma)$-modules respectively. Then $\operatorname{Res}_{\gamma\beta}(L_\gamma\circ L_\beta) \cong L_\gamma\otimes L_\beta$ and $\operatorname{Res}_{\gamma\beta}(L_\beta\circ L_\gamma) \cong q^{-\beta\cdot \gamma} L_\gamma\otimes L_\beta$.
\end{lemma}
\begin{proof}
By Lemma \ref{independence}, without loss of generality, assume our convex order $\prec$ is as in Example \ref{tworow}. Thus the only roots between $\gamma$ and $\beta$ are of the form $\gamma+n\delta$, $\beta+n\delta$ or $n\delta$.
Consider a nonzero quotient in the Mackey filtration of $\operatorname{Res}_{\gamma,\beta}(L_\gamma\circ L_\beta)$. Then we have $\lambda,\mu,\nu\in\mathbb{N} I$ such that $\lambda+\mu=\gamma$, $\mu+\nu=\beta$, $\lambda$ is a sum of roots less than or equal to $\gamma$, $\nu$ is a sum of roots greater than or equal to $\beta$, while $\mu$ is both a sum of roots greater than or equal to $\gamma$ and a (possibly different) sum of roots less than or equal to $\beta$.
Consider $\gamma=\lambda+\mu$ which has been written as a sum of roots less than or equal to $\beta$. No roots between $\gamma$ and $\beta$ can appear in this sum. By convexity of the convex order, the only options are $\mu=0$, $\mu=\gamma$ and $\mu=\beta$. We will have to show that the last two options are not possible.
So suppose for want of a contradiction that $\mu=\gamma$. Then $\nu=\beta-\gamma=\sum_i \nu_i$ with each $\nu_i$ larger than $\beta$. Note that there is at least two terms in this sum as $\beta-\gamma$ is not a root.
Since $(\gamma,\beta-\gamma)=-4$, there exists an index $j$ such that $(\gamma,\nu_j)<0$. Therefore $\gamma+\nu_j$ is a root. Now consider
\begin{equation}\label{bgn}
\beta = (\gamma+\nu_j) + \sum_{i\neq j} \nu_i.
\end{equation}
By convexity this implies $\gamma+\nu_j\prec\beta$ and as $\nu_j\succ\beta\succ\gamma$ it must be that $\gamma+\nu_j\succ\gamma$.
The equation (\ref{bgn}) implies $|\gamma+\nu_j|<|\beta|$. But on the other hand we've classified all roots $\alpha$ between $\beta$ and $\gamma$ and none of them satisfy $|\alpha|<\beta$, a contradiction.
The case $\mu=\beta$ is handled similarly.
Therefore there is only one term in the Mackey filtration, which is the one where $\mu=0$, whence we obtain the lemma.
\end{proof}
\section{Simple Imaginary Modules}\label{sim
We start by following \cite{kkk} and defining the $R$-matrices for KLR algebras. First we need to introduce some useful elements of $R(\nu)$.
For $1\leq a< n=|\nu|$ we define elements $\varphi_a\in R(\nu)$ by
\[
\varphi_a e_{\bf i} =
\begin{cases}
(\tau_ay_a-y_{a}\tau_a)e_{\bf i} & \text{if ${\bf i}_a={\bf i}_{a+1}$,} \\
\tau_ae_{\bf i} & \text{otherwise.}
\end{cases}
\]
These elements satisfy the following properties
\begin{lemma}\cite[Lemma 1.3.1]{kkk}\label{klem}
\hfill
\begin{enumerate}
\item $\varphi_a^2e_{\bf i}=(
Q_{\nu_a,\nu_{a+1}}(x_a,x_{a+1})+\delta_{\nu_a,\nu_{a+1}}) e_{\bf i}.$
\item
$\{\varphi_k\}_{1\le k<n}$ satisfies the braid relations.
\item
For $w\in S_n$, let $w=s_{a_1}\cdots s_{a_\ell}$ be a reduced expression of $w$ and
set $\varphi_w=\varphi_{a_1}\cdots\varphi_{a_\ell}$.
Then $\varphi_w$ does not depend on the choice of reduced expressions of $w$.
\item For $w\in S_n$ and $1\le k\le n$, we have $\varphi_w x_k=x_{w(k)}\varphi_w$.
\item
For $w\in S_n$ and $1\le k<n$,
if $w(k+1)=w(k)+1$, then $\varphi_w\tau_k=\tau_{w(k)}\varphi_w$.\label{ga5}
\item
$\varphi_{w^{-1}}\varphi_we_{\bf i}=\prod\limits_{ \substack{a<b,\\ w(a)>w(b)} }
(Q_{{\bf i}_a,{\bf i}_b}(x_a,x_b)+\delta_{{\bf i}_a,{\bf i}_b})e_{\bf i}$. \label{ga6}
\end{enumerate}
\end{lemma}
Let $M$ and $N$ be modules for $R(\lambda)$ and $R(\mu)$ respectively. Let $(\lambda,\mu)_n$ be the degree of $\phi_{w[\lambda,\mu]}$. Define the morphism $R_{M,N}\map{M\circ N}{q^{-(\lambda,\mu)_n}N\circ M}$ by
\[
R_{M,N}(u\otimes v)=\varphi_{w[\lambda,\mu]}v\otimes u.
\]
In \cite{kkk} an algebra homomorphism $\psi_z\map{R(\nu)}{\mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes R(\nu)}$ is constructed where $\psi_z(e_{\bf i})=e_{\bf i}$, $\psi_z(y_j)=y_j+z$ and $\psi_z(\tau_k)=\tau_k$. If $M$ is an $R(\nu)$-module we define the $R(\nu)$-module $M_z=\psi_z^*(\mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes M)$. The morphism $r_{M,N}\map{M\circ N}{q^{2s-(\lambda,\mu)_n}N\circ M}$ is now defined by
\[
r_{M,N} = \left( (z-w)^{-s} R_{M_z,N_w} \right) |_{z=w=0}.
\]
where $s$ is the largest possible integer for which this definition is possible. In \cite{kkk} it is shown that $r_{M,N}$ is a nonzero morphism and that these collections of morphisms satisfy the braid relation.
\begin{lemma}\label{risq}
Let $L_1$ and $L_2$ be two irreducible cuspidal representations of $R(\delta)$. Then the morphisms $r_{L_1,L_2}$ and $r_{L_2,L_1}$ are inverse to one another.
\end{lemma}
\begin{proof}
By adjunction
\[
\operatorname{Hom} (L_1\circ L_2,L_2\circ L_1) \cong \operatorname{Hom}(L_1\otimes L_2,\operatorname{Res}_{\delta,\delta} L_2\circ L_1).
\]
As $L_1$ and $L_2$ are cuspidal, the Mackey filtration of $\operatorname{Res}_{\delta,\delta} (L_2\circ L_1)$ has two nonzero pieces, namely $L_2\otimes L_1$ and $L_1\otimes L_2$. In particular this implies that $\operatorname{Hom} (L_1\circ L_2,L_2\circ L_1)$ is concentrated in degree zero. Since $r_{L_1,L_2}\neq 0$, the integer $s$ in the construction of $r_{L_1,L_2}$ must be equal to $(\delta,\delta)_n/2$.
For $j=1,2$,
pick a nonzero vector $v_j\in L_j$ such that $y_i v_j=0$ for all $i$.
The morphism $r_{L_2,L_1}r_{L_1,L_2}$ maps $v_1\otimes v_2$ to $\left((z'-z)^{-2s}\varphi_{w[\delta,\delta]}^2 v_1\otimes v_2\right)|_{z=z'=0}$ where the computation is taking place in $(L_1)_{z}\circ (L_2)_{z'}$ (by abuse of notation, we write $v$ for $1\otimes v\in L_z$). We can compute this using Lemma \ref{klem}(vi). Since $y_i v_j=0$ in $L_j$, we have $y_i v_j=zv_j$ in $(L_j)z$. Then the product on the right hand side of \ref{klem}(vi) acts by the scalar $(z'-z)^{(\delta,\delta)_n}$ on the vector $v_1\otimes v_1\in (L_1)_{z}\circ (L_2)_{z'}$. We've already computed $(\delta,\delta)_n=2s$ and hence $r_{L_2,L_1}r_{L_1,L_2} v_1\otimes v_2=v_1\otimes v_2$.
Since $L_1$ and $L_2$ are irreducible, $v_1\otimes v_2$ generates $L_1\circ L_2$. Therefore $r_{L_2,L_1}r_{L_1,L_2}$ is the identity.
\end{proof}
From the evident maps from $\operatorname{End}(L\circ L)$ to $\operatorname{End}(L^{\circ n})$,
the morphisms $r_{L,L}$ define $n-1$ elements, denoted $r_1,r_2,\ldots , r_{n-1} \in \operatorname{End}(L^{\circ n})$.
The following result was first noticed in a special case in
\cite[Theorem 4.13]{kmr}, and is fundamental to the paper
\cite{imaginaryschurweyl}.
\begin{theorem}\label{14.3}
Let $L$ be a cuspidal representation of $R(\delta)$.
There is an isomorphism $\operatorname{End}(L^{\circ{n}})\cong \mathbb{Q}}\def\partition{\mathcal{P}[S_n]$ sending $r_i$ to the transposition $(i,i+1)$.
\end{theorem}
\begin{proof}
By adjunction $\operatorname{End}(L^{\circ n})=\operatorname{Hom}(L^{\otimes n},\operatorname{Res}_{\delta,\ldots,\delta}L^{\circ n})$. Since $L$ is cuspidal, the Mackey filtration of $\operatorname{Res}_{\delta,\ldots,\delta}L^{\circ n}$ has exactly $n!$ nonzero subquotients, each isomorphic to $L^{\otimes n}$. Therefore $\dim\operatorname{End}(L^{\circ n})\leq n!$.
By Lemma \ref{risq}, $r_i^2=1$. The identity $r_ir_j=r_jr_i$ for $|j-i|>1$ is trivial and the braid relation $r_i r_{i+1} r_i= r_{i+1} r_i r_{i+1}$ is a general fact about the morphisms $r_{M,N}$ constructed in \cite{kkk}. This allows us to define $r_w$ for each $w\in S_n$.
Recall that in the proof of Lemma \ref{risq}, we showed that $s=(\delta,\delta)_n/2$, where $s$ is the integer appearing in the definition of $r_{L,L}$.
Therefore by induction on the length of $w$, using \cite[Proposition 1.4.4(iii)]{kkk}, we obtain
\[
r_w v\otimes\cdots \otimes v - \tau_{\iota(w)}v\otimes \cdots \otimes v \in \sum_{\ell(w')<\ell(\iota(w))}\tau_{w'}L\otimes \cdots \otimes L
\]
where $\iota\map{S_n}{S_{n|\delta|}}$ is the obvious embedding.
Therefore the endomorphisms $r_w$ are linearly independent.
Since the $r_i$ satisfy the Coxeter relations there is a homomorphism from $\mathbb{Q}}\def\partition{\mathcal{P}[S_n]$ to $\operatorname{End}(L^{\circ n})$. We have just shown it is injective. Surjectivity follows from the dimension estimate in the first paragraph of this proof.
\end{proof}
Let $\w$ be a chamber coweight. Let $L(\w)$ be the head of the module $L(\w_-)\circ L(\w_+)$.
\begin{lemma}\label{reslw}
The module $L(\w)$ is an irreducible module with $L(\w)^\circledast\cong L(\w)$. Furthermore $\operatorname{Res}_{\w_-,\w_+}L(\w)\cong L(\w_-)\otimes L(\w_+)$.
\end{lemma}
\begin{remark}
The irreducibility of $L(\w)$ is in \cite{tingleywebster} and can also be derived from \cite[Theorem 3.2]{kkko}. Our preference for giving this proof is that we wish to make use of the extra properties of $L(\w)$ that we establish.
\end{remark}
\begin{proof}
Using Theorem \ref{independence} and the convex order from Example \ref{tworow}, we may assume without loss of generality that $(\w_+,\w_-)$ is a minimal pair for $\delta$.
For any quotient $Q$ of $L(\w_-)\circ L(\w_+)$ there is, by adjunction, a nonzero morphism from $L(\w_-)\otimes L(\om_+)$ to $\operatorname{Res}_{\w_-,\w_+}Q$ which is injective as the source is irreducible.
Lemma \ref{strongminimalpair} implies that $\operatorname{Res}_{\w_-,\w_+} L(\w_-)\circ L(\w_+)\cong L(\w_-)\otimes L(\w_+)$. By exactness of the restriction functor, this forces the head of $L(\w_-)\circ L(\w_+)$ to be irreducible and furthermore $\operatorname{Res}_{\w_-,\w_+}L(\w)\cong L(\w_-)\otimes L(\w_+)$. The self-duality of $L(\w)$ follows since every simple module is self-dual up to a grading shift, duality commutes with restriction and the modules $L(\w_\pm)$ are self-dual.
\end{proof}
\section{The Growth of a Quotient}\label{growth
Let $z$ be the element $y_1+\cdots +y_{|\nu|}\in R(\nu)$. It is straightforward to check that $z$ is central.
The following lemma and proof appeared in an early version of \cite{bkm}.
\begin{lemma}\label{centretrick}
Let $R'(\nu)$ be the subalgebra of $R(\nu)$ generated by $e_{\bf i}$, ${\bf i}\in\,\mbox{Seq}\,(\nu)$, $\tau_i$ and $y_i-y_{i+1}$, $1\leq i<|\nu|$. Then multiplication induces an algebra isomorphism $\mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes R'(\nu)\to R(\nu)$.
\end{lemma}
\begin{proof}
An inspection of the presentation (\ref{eq:KLR}) of $R(\nu)$ shows that the set of elements of the form
\[
(y_1-y_2)^{a_1}(y_2-y_3)^{a_2}\cdots (y_{n-1}-y_n)^{a_{n-1}}\tau_w e_{\bf i}
\]
with $a_1,\ldots,a_{n-1}\in \mathbb{N}$, $w\in S_n$ and ${\bf i}\in\,\mbox{Seq}\,(\nu)$ is a spanning set for $R'(\nu)$. Since Theorem \ref{basisofrnu} provides us with a basis of $R(\nu)$, we can see that the collection of elements above forms a linearly independent set, hence is a basis for $R'(\nu)$. We compute
$$ny_n= z + \sum_{i=1}^{n-1}i(y_i- y_{i+1})$$
and thus $y_n$ is in the image of $\mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes R'(\nu)$. Therefore the multiplication map from $\mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes R'(\nu)$ to $R(\nu)$ is surjective. A dimension count using Lemma \ref{basisofrnu} shows that it must be an isomorphism.
\end{proof}
\begin{lemma}\label{scentretrick}
Let $\alpha$ be a positive root. There is an injection from $\mathbb{Q}}\def\partition{\mathcal{P}[z]$ into the centre of $S(\alpha)$.
\end{lemma}
\begin{proof}
Let $S'(\alpha)$ be the quotient of $R'(\alpha)$ by the two sided ideal generated by all $e_{\bf i}$ such that $e_{\bf i} L=0$ for all cuspidal representations $L$ of $R(\alpha)$.
Lemma \ref{centretrick} implies that $S(\alpha)\cong \mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes S'(\alpha)$. The image of $\mathbb{Q}}\def\partition{\mathcal{P}[z]\otimes \mathbb{Q}}\def\partition{\mathcal{P}$ provides us with our desired central subalgebra.
\end{proof}
Let $\alpha$ be an indivisible root, $L$ a cuspidal representation of $R(\alpha)$ and let $(\beta,\gamma)$ be a minimal pair for $\alpha$. Let $L''\otimes L'$ be an irreducible subquotient of $\operatorname{Res}_{\gamma,\beta}L$. By Lemma \ref{oppositerestrict}, $L''$ and $L'$ are cuspidal modules for $R(\gamma)$ and $R(\beta)$. We will call $(L',L'')$ a minimal pair for $L$.
We inductively define a word ${\bf i}_L\in\,\mbox{Seq}\,(\alpha)$ as the concatenation ${\bf i}_{L''}{\bf i}_{L'}$.
Let $T(L)$ be the subalgebra of $e_{{\bf i}_L} S(\alpha) e_{{\bf i}_L}$ generated by $y_1 e_{{\bf i}_L},\ldots,y_{|\alpha|}e_{{\bf i}_L}$.
\begin{lemma}\label{surjbydefn}
Let $L$ be a cuspidal representation and $(L',L'')$ be a minimal pair for $L$. The inclusion $R(\gamma)\otimes R(\beta)\to R(\alpha)$ induces a homomorphism from $T(L'')\otimes T(L')$ to $T(L)$.
\end{lemma}
\begin{proof}
Suppose $x\in \ker(R(\gamma)\to S(\gamma))$. Consider $x\otimes 1\in R(\gamma)\otimes R(\beta)\hookrightarrow R(\alpha)$. On $M\in S(\alpha)\mbox{-mod}$, $x\otimes 1$ acts in the way it does on
$\operatorname{Res}_{\gamma\beta}M$, which is a $S(\gamma)\otimes S(\beta)$-module. Therefore $x\otimes 1$ acts by zero and hence is in the kernel of $R(\alpha)\to S(\alpha)$.
\end{proof}
\begin{lemma}\label{zerodim}
Let $L$ be a cuspidal representation of $\alpha$.
The scheme $\operatorname{Proj} T(L)$ has a unique $\overline{\mathbb{Q}}$-point $[x_1:\cdots :x_{|\alpha|}]$, namely $x_1=\cdots=x_{|\alpha|}$.
\end{lemma}
\begin{proof}
We prove this by induction on the height of $\alpha$. Choose a minimal pair $(\beta,\gamma)$ for $\alpha$ and $(L',L'')$ for $L$.
Suppose that $[x_1:\cdots : x_{|\alpha|}]$ is a $\overline{\mathbb{Q}}$-point of $\operatorname{Proj} T(L)$. Then by Lemma \ref{surjbydefn} $[x_1:\cdots : x_{|\gamma|}]$ and $[x_{|\gamma|+1}:\cdots:x_{|\alpha|}]$ are points in $\operatorname{Proj} T(L'')$ and $\operatorname{Proj} T(L')$ respectively.
By inductive assumption, $x_1=\cdots = x_{|\gamma|}$ and $x_{|\gamma|+1}=\cdots = x_{|\alpha|}$.
Let $w=w[|\gamma|,|\beta|]$ and consider the element $\varphi_w^2 e_{{\bf i}_L}$. By Lemma \ref{klem}(vi) it lives in $T({\bf i}_L)$ and since $\varphi_w^2 e_{{\bf i}_L}=\varphi_w e_{{\bf i}_{L'}{\bf i}_{L''}}\varphi_w$, it lives in the kernel of the map from $R(\alpha)$ to $S(\alpha)$. Therefore $\varphi_w^2 e_{{\bf i}_L}$ is zero in $T({\bf i}_L)$.
Lemma \ref{klem}(vi) writes $\varphi_w^2 e_{{\bf i}_L}$ as a product of elements of the form $x_i-x_j$ where $i\leq |\gamma|$ and $j>|\gamma|$. Therefore any $\overline{\mathbb{Q}}$-point of $\operatorname{Proj} T({\bf i}_L)$ has $x_1=\cdots=x_{|\alpha|}$ as required.
\end{proof}
\begin{theorem}\label{oofone}
Let $\alpha$ be an indivisible root. Then $\dim S(\alpha)_d$ is bounded as a function of $d$.
\end{theorem}
\begin{proof}
Consider a composition series for $S(\alpha)$ as a $S(\alpha)$-module. Every composition factor must be cuspidal, so
\begin{equation}\label{sl}
[S(\alpha)]= \sum_L f_L(q) [L]
\end{equation}
where $f_L(q)\in \mathbb{N}((q))$ and the sum is over irreducible cuspidal representations $L$. For any ${\bf i}\in Seq(\nu)$, we therefore get the equality
\begin{equation}\label{fl}
\dim (e_{\bf i} S(\alpha))=\sum_L f_L(q) \dim e_{\bf i} L.
\end{equation}
Pick an irreducible cuspidal representation $L$ and let ${\bf i}_L$ be the corresponding word in $\,\mbox{Seq}\,(\nu)$. By Lemma \ref{zerodim} and the theory of the Hilbert polynomial, $\dim T({\bf i}_L)_d$ is a bounded function of $d$. From Theorem \ref{basisofrnu} we see that $e_{{\bf i}_L} S(\alpha)$ is finite over $T({\bf i}_L)$ and hence $\dim (e_{{\bf i}_L} S(\alpha))_d$ is a bounded function of $d$.
We take ${\bf i}={\bf i}_L$ in (\ref{fl}) and since $e_{{\bf i}_L}L\neq 0$, the Laurent series $f_L(q)=\sum_d f_L^{(d)} q^d$ has $f_L^{(d)}$ a bounded function of $d$. Equation (\ref{sl}) completes the proof.
\end{proof}
\section{An Important Short Exact Sequence}
Let $\alpha$ be a real root. Define $\Delta(\alpha)$ to be the projective cover of $L(\alpha)$ in the category of $S(\alpha)$-modules. Let $\om$ be a chamber coweight. Define $\Delta(\om)$ to be the projective cover of $L(\om)$ in the category of $S(\delta)$-modules.
\begin{lemma}\label{mpses}
Let $\alpha$ be an indivisible root.
Suppose that $(\beta,\gamma)$ is a minimal pair for $\alpha$. Let $\Delta_\beta$ and $\Delta_\gamma$ be finitely generated projective $S(\beta)$ and $S(\gamma)$-modules.
Then there is a short exact sequence
\[
0 \to q^{-\beta\cdot \gamma}\Delta_\beta\circ \Delta_\gamma\to \Delta_\gamma\circ \Delta_\beta\to C\to 0
\] for some projective $S(\alpha)$-module $C$.
\end{lemma}
\begin{proof}
By adjunction,
\[
\operatorname{Hom}(q^{-\beta\cdot \gamma}\Delta_\beta\circ \Delta_\gamma, \Delta_\gamma\circ \Delta_\beta)\cong \operatorname{Hom}(q^{-\beta\cdot\gamma}\Delta_\beta\otimes \Delta_\gamma,\operatorname{Res}_{\beta\gamma}\Delta_\gamma\circ\Delta_\beta).
\]
Since the modules $\Delta_\gamma$ and $\Delta_\beta$ are cuspidal, the Mackey filtration of $\operatorname{Res}_{\beta\gamma}\Delta_\gamma\circ\Delta_\beta$ has only one nonzero term, yielding an isomorphism
$$\operatorname{Res}_{\beta\gamma}\Delta_\gamma\circ\Delta_\beta\cong q^{-\beta\cdot \gamma}\Delta_\beta\otimes \Delta_\gamma.$$
Let $\phi\map{q^{-\beta\cdot\gamma}\Delta_\beta\circ \Delta_\gamma}{\Delta_\gamma\circ \Delta_\beta}$ be the image of the identity map on $q^{-\beta\cdot\gamma}\Delta_\beta\otimes\Delta_\gamma$ under the isomorphisms discussed above.
This map $\phi$ satisfies
\begin{equation}\label{explicitphi}
\phi(1\otimes (v_\beta\otimes v_\gamma)) = \tau_{w[\beta,\gamma]} 1\otimes (v_\gamma\otimes v_\beta)
\end{equation}
for all $v_\beta\in \Delta_\beta$ and $v_\gamma\in \Delta_\gamma$.
There are filtrations of $\Delta_\beta$ and $\Delta_\gamma$ where each successive subquotient is an irreducible cuspidal module for $R(\beta)$ or $R(\gamma)$ respectively. This induces a pair of filtrations on $\Delta_\beta\otimes \Delta_\gamma$ and $\Delta_\gamma\otimes \Delta_\beta$ where the successive subquotients are of the form $L_\beta \circ L_\gamma$ or $L_\gamma\circ L_\beta$ for cuspidal irreducible representations $L_\beta$ and $L_\gamma$ of $R(\beta)$ and $R(\gamma)$.
From the explicit formula (\ref{explicitphi}), we see that $\phi$ induces a morphism $\bar\phi$ on each subquotient $\bar\phi\map{q^{-\beta\cdot \gamma}L_\beta\circ L_\gamma}{L_\gamma \circ L_\beta}$ satisfying
\[
\bar\phi(1\otimes (v_\beta\otimes v_\gamma)) = \tau_{w[\beta,\gamma]} 1\otimes (v_\gamma\otimes v_\beta).
\]
By Theorem \ref{main}, the module $L_\beta\circ L_\gamma$ has an irreducible head $A(L_\beta,L_\gamma)$. Since $(\beta,\gamma)$ is a minimal pair, all other composition factors are cuspidal. Taking duals, $q^{\beta\cdot\gamma}L_\gamma\circ L_\beta$ has $A(L_\beta,L_\gamma)$ as its socle with all other composition factors cuspidal.
The morphism $\bar\phi$ therefore sends the head of $q^{-\beta\cdot \gamma}L_\beta\circ L_\gamma$ onto the socle of $L_\gamma\circ L_\beta$. Hence $\phi$ induces a bijection between all occurrences of non-cuspidal subquotients as sections of filtrations of $q^{-\beta\cdot \gamma}\Delta_\beta\circ \Delta_\gamma$ and $\Delta_\gamma\circ \Delta_\beta$. This shows that $\ker\phi$ and $\operatorname{coker}\phi$ are both cuspidal $R(\alpha)$-modules.
Suppose for want of a contradiction that $\ker\phi$ is nonzero. It is a submodule of
the finitely generated module $q^{-\beta\cdot\gamma}\Delta_\beta\circ \Delta_\gamma$. By \cite[Corollary 2.11]{khovanovlauda}, $R(\alpha)$ is Noetherian and hence $\ker\phi$ is finitely generated.
As $\ker\phi$ is cuspidal it is a $S(\alpha)$-module, so by Theorem \ref{oofone} we deduce that $\dim(\ker\phi)_d$ is bounded as a function of $d$.
The adjunction (\ref{oppositeadjunction}) yields a canonical nonzero map from $\operatorname{Res}_{\gamma,\beta}\ker\phi$ to $\Delta_\gamma\otimes \Delta_\beta$. If $X$ is the image of this map then we have $\dim X_d$ is a bounded function of $d$.
The modules $\Delta_\beta$ and $\Delta_\gamma$ are free over the central subalgebra $\mathbb{Q}}\def\partition{\mathcal{P}[z]$ of $S(\beta)$ and $S(\gamma)$. Therefore $\Delta_\beta\otimes \Delta_\gamma$ is a free $\mathbb{Q}}\def\partition{\mathcal{P}[z_1,z_2]$-module. Hence there are no nonzero submodules $M$ of $\Delta_\beta\otimes \Delta_\gamma$ for which $\dim M_d$ is a bounded function of $d$. This is a contradiction, implying $\phi$ is injective.
Now let $L$ be a cuspidal $R(\alpha)$-module. We apply $\operatorname{Hom}(-,L)$ to the short exact sequence
\[
0\to q\Delta_\beta\circ \Delta_\gamma \xrightarrow{\phi} \Delta_\gamma\circ \Delta_\beta \to \operatorname{coker}\phi\to 0.
\]
and obtain a long exact sequence. As $\operatorname{Res}_{\beta,\gamma}L=0$, we have
\[
\operatorname{Ext}^i(\Delta_\beta\circ \Delta_\gamma,L) = \operatorname{Ext}^i(\Delta_\beta\otimes \Delta_\gamma,\operatorname{Res}_{\beta,\gamma}L)=0.
\]
Therefore our long exact sequence degenerates into a sequence of isomorphisms
\begin{equation}\label{uno}
\operatorname{Ext}^i(\operatorname{coker}\phi,L)\cong \operatorname{Ext}^i(\Delta_\gamma\circ \Delta_\beta,L)
\end{equation}
and by adjunction we have
\begin{equation}\label{duo}
\operatorname{Ext}^i(\Delta_\gamma\circ \Delta_\beta,L)\cong \operatorname{Ext}^i(\Delta_\gamma\otimes \Delta_\beta,\operatorname{Res}_{\gamma,\beta}L).
\end{equation}
Lemma \ref{oppositerestrict} shows that $\operatorname{Res}_{\gamma,\beta}L$ is a $S(\gamma)\otimes S(\beta)$-module.
Since $\Delta_\gamma\otimes \Delta_\beta$ is a projective $S(\gamma)\otimes S(\beta)$-module, we derive that $\operatorname{Ext}^1(\Delta_\gamma\otimes \Delta_\beta,\operatorname{Res}_{\gamma,\beta}L)=0$. Tracing through the above isomorphisms yields $\operatorname{Ext}^1(\operatorname{coker}\phi,L)=0$ and therefore $\operatorname{coker}\phi$ is a projective $S(\alpha)$-module.
\end{proof}
\section{Cuspidal Representations of $R(\delta)$
We first explain the intertwined logical structure of this section and the following one. Each statement in Section 18 involves a positive root $\alpha$. We prove all the results in this section under an assumption that the results in Section 18 are known for all roots $\alpha$ of height less than the height of $\delta$. The reader wil not be worried about the forward references once the logical structure of Section 18 is known.
The results of Section 18 will be proved by a simultaneous induction on the height of the root $\alpha$. In particular, when Theorem 18.1 is proved for a root $\alpha$, it will be safe to assume that Theorem 18.2 is known for all roots of smaller height. There are references to the results of this section in Section 18. However they only appear when the root $\alpha$ under question is of height at least that of $\delta$. Thus there is no circularity and the argument is valid.
Let $\omega$ be a chamber coweight. Recall from \S \ref{sim} that $L(\omega)$ is the head of the module $L(\om_-)\circ L(\om_+)$ and is irreducible. Let $\Delta(\omega)$ be the projective cover of $L(\omega)$ in the category of $S(\delta)$-modules. We caution the reader that while $L(\w)$ will depend only on the chamber coweight $\w$ (as in \cite{tingleywebster}), the module $\Delta(\w)$ will depend not just on $\w$ but also on the positive system $p(\Phi_{\succ\delta})$.
\begin{theorem}\label{omegases} Let $\w$ be a chamber coweight.
There is a short exact sequence
\[
0 \to q^2 \Delta(\om_+)\circ \Delta(\om_-)\to \Delta(\om_-)\circ \Delta(\om_+)\to \Delta(\om)\to 0.
\]
\end{theorem}
\begin{proof}
As in the proof of Lemma \ref{reslw}, we may assume without loss of generality that $(\w_+,\w_-)$ is a minimal pair for $\delta$.
By Lemma \ref{mpses}, there is
a short exact sequence
\begin{equation}\label{cses}
0 \to q^2 \Delta(\om_+)\circ \Delta(\om_-)\to \Delta(\om_-)\circ \Delta(\om_+)\to C\to 0
\end{equation}
for some projective $S(\delta)$-module $C$.
As $C$ is cuspidal, $\operatorname{Res}_{\w_+\w_-}C=0$. By adjunction, this implies that $\operatorname{Ext}^i(q^2\Delta(\w_+)\circ \Delta(\w_-),C)=0$. From the long exact sequence obtained by applying $\operatorname{Hom}(-,C)$ to (\ref{cses}), we therefore get an isomorphism
\begin{equation}\label{1}
\operatorname{End}(C) \cong \operatorname{Hom} ( \Delta(\w_-)\circ\Delta(\w_+),C).
\end{equation}
By Lemma \ref{strongminimalpair} and adjunction,
\[
\operatorname{Ext}^1(\Delta(\w_-)\circ \Delta(\w_+),q^2 \Delta(\w_+)\circ\Delta(\w_-) )=
\operatorname{Ext}^1(\Delta(\w_-)\otimes \Delta(\w_+),q^4 \Delta(\w_-)\otimes \Delta(\w_+))
\]
which is zero since $\Delta(\w_+)\otimes \Delta(\w_-)$ is a projective $S(\w_+)\otimes S(\w_-)$-module. From the long exact sequence obtained by applying $\operatorname{Hom}(\Delta(\w_-)\circ\Delta(\w_+),-)$ to (\ref{cses}), we therefore have a surjection
from $\operatorname{End}(\Delta(\w_-)\circ \Delta(\w_+))$ onto $\operatorname{Hom}(\Delta(\w_-)\circ\Delta(\w_+),C)$.
Again we apply Lemma \ref{strongminimalpair} and adjunction to obtain
\[
\operatorname{End}(\Delta(\w_-)\circ\Delta(\w_+))\cong \operatorname{End}(\Delta(\w_-)\otimes \Delta(\w_+))
\]
By Theorem \ref{premorita} this is isomorphic to $\mathbb{Q}}\def\partition{\mathcal{P}[x,y]$ with $x$ and $y$ in degree 2.
Concentrating our attention to degree zero, we obtain $\operatorname{End}(C)_0\cong \mathbb{Q}}\def\partition{\mathcal{P}$. Therefore $C$ is indecomposable.
The module $L(\omega)$ is by construction a quotient of $\Delta(\w_-)\circ \Delta(\w_+)$. Since it is cuspidal, the same argument that produced the isomorphism (\ref{1}) yields an isomorphism
\[
\operatorname{Hom}(C,L(\w))\cong \operatorname{Hom}(\Delta(\w_-)\circ \Delta(\om_+),L(\w)).
\]
Therefore
$L(\om)$ is a quotient of $C$. Since $C$ is an indecomposable projective $S(\delta)$-module, it must be that $C$ is the projective projective cover of $L(\om)$.
\end{proof}
\begin{corollary}\label{stdim}
Let $\om$ be a chamber coweight. Then $[\Delta(\om)]\in\mathbf{f}$ and when specialised to $q=1$ is equal to $h_\w \otimes t$.
\end{corollary}
\begin{proof}
This is immediate from Theorems \ref{realpbw} and \ref{omegases}.
\end{proof}
As a consequence we also obtain the following theorem, which also appears in \cite{tingleywebster}.
\begin{theorem}\label{imaginaryclassification}
The set of all modules $L(\omega)$, as $\omega$ runs over the chamber coweights adapted to the convex order $\prec$, is a complete list of the cuspidal irreducible representations of $R(\delta)$.
\end{theorem}
\begin{proof}
Corollary \ref{stdim} shows that the modules $\Delta(\om)$ are a complete set of indecomposable projective modules for $S(\delta)$. In the last paragraph of the proof of Theorem \ref{omegases}, we showed that the module $L(\w)$ is a quotient of $\Delta(\om)$. We also proved that $L(\om)$ is simple in Lemma \ref{reslw}. Therefore the set of such $L(\w)$ is a complete set of irreducible cuspidal representations of $R(\delta)$.
\end{proof}
Let $\{n_\w\}_{\w\in\Om}$ be a sequence of natural numbers. Lemma \ref{risq} shows that the induced product
\[
\dct{w\in \Omega} L(\om)^{\circ n_\w}
\] is independent of the order of the factors.
Now we know the modules $L(\w)$ are pairwise nonisomorphic, we can use
the same argument as in Theorem \ref{14.3} to obtain a natural isomorphism
\begin{equation}\label{endsn}
\operatorname{End} \left( \dct{w\in \Om} L(\om)^{n_\w} \right) \cong \bigotimes_{w\in \Om} \mathbb{Q}}\def\partition{\mathcal{P}[S_{n_\w}].
\end{equation}
If $\{m_\w\}_{\w\in\Om}$ and $\{n_\w\}_{\w\in\Om}$ are two sequences of natural numbers then there is a natural inclusion
\begin{equation}\label{endinclusion}
\operatorname{End} \left( \dct{w\in \Om} L(\om)^{m_\w} \right) \otimes \operatorname{End} \left( \dct{w\in \Om} L(\om)^{n_\w} \right) \hookrightarrow \operatorname{End} \left( \dct{w\in \Om} L(\om)^{m_\w+n_\w} \right)
\end{equation}
which, under the isomorphism (\ref{endsn}) is the tensor product of the natural inclusions
\begin{equation}\label{symmetricinclusion}
\mathbb{Q}}\def\partition{\mathcal{P}[S_{m_\w}]\otimes \mathbb{Q}}\def\partition{\mathcal{P}[S_{n_\w}] \hookrightarrow \mathbb{Q}}\def\partition{\mathcal{P}[S_{m_\w +n_\w}].
\end{equation}
If $\om$ is a chamber coweight and $\lambda$ is a partition of $n$, we define
\[
L_\w(\lambda) = \operatorname{Hom}_{\mathbb{Q}}\def\partition{\mathcal{P}[S_n]} ( S^\lambda, L(\w)^{\circ n} )
\]
where $S^\lambda$ is the Specht module for $S_n$.
Let $\uline{\lambda}=\{\lambda_\w\}_{\w\in\Omega}$ be a multipartition. Then we define
\[
L(\uline{\lambda})=\dct{\w\in \Omega} L_\w(\lambda_\w)=\operatorname{Hom}_{\otimes\mathbb{Q}}\def\partition{\mathcal{P}[S_{n_\w}]} \left(\bigotimes_{\w\in\Omega} S^{\lambda_\w},\dct{\w\in \Om}L(\w)^{\circ n_\w}\right).
\]
Define the multi-Littlewood-Richardson coefficients by $$c^{\unu}_{\ula\umu}=\prod_{\w\in\Omega} c_{\lambda_\w\mu_\w}^{\nu_\w}$$ where $c_{\lambda\mu}^\nu$ is the ordinary Littlewood-Richardson coefficient, which we take to be zero if $|\nu|\neq |\lambda|+|\mu|$.
\begin{theorem}\label{lrl}
The family of modules $L(\ula)$ enjoy the following properties under induction and restriction:
\[
L(\ula)\circ L(\umu) = \bigoplus_{\unu} L(\unu)^{ \oplus c^{\unu}_{\ula\umu} }
\]
\[
\operatorname{Res}_{k\delta,(n-k)\delta} L(\unu) = \bigoplus_{\ula\vdash k,\umu\vdash n-k} L(\ula)\otimes L(\umu)^{\oplus c^{\unu}_{\ula\umu} }
\]
\end{theorem}
\begin{proof}
This follows from the observation above that the inclusions (\ref{endinclusion}) and (\ref{symmetricinclusion}) are equivalent under the isomorphism (\ref{endsn}), together with the known formulae for the induction and restriction of Specht modules for the inclusions $S_m\times S_n\to S_{m+n}$.
\end{proof}
As a particular case of Theorem \ref{lrl}, we have
\begin{equation}\label{lrr}
\operatorname{Res}_{k\delta,(n-k)\delta} L_\w (1^n) \cong L_\w(1^k) \otimes L_\w(1^{n-k}).
\end{equation}
\section{Homological Modules
See the beginning of the previous section for a discussion of the inductive structure of the arguments in this section.
\begin{theorem}\label{14.1}
Let $\alpha$ be an indivisible positive root. Let $\Delta$ and $L$ be $S(\alpha)$-modules with $\Delta$ projective. Then for all $i>0$,
\[
\operatorname{Ext}^i(\Delta,L) =0.\]
\end{theorem}
We remind readers that these Ext groups are taken in the category of $R(\alpha)$-modules which makes this result nontrivial.
\begin{proof}
Let $(\beta,\gamma)$ be a minimal pair for $\alpha$. If $\alpha$ is a real root, then by the inductive hypothesis applied to Theorem \ref{realpbw} and Corollary \ref{stdim}, there exist projective $S(\beta)$ and $S(\gamma)$-modules, $\Delta_\beta$ and $\Delta_\gamma$ such that $[\Delta_\beta][\Delta_\gamma]\neq q^{\beta\cdot \gamma}[\Delta_\gamma][\Delta_\beta]$. Therefore in the short exact sequence of Lemma \ref{mpses}, $C$ is a nonzero direct sum of copies of $\Delta(\alpha)$.
If $\alpha$ is imaginary, then without loss of generality assume that $\Delta$ is indecomposable projective, hence isomorphic to $\Delta(\om)$ for some $\om$. Then we use the short exact sequence of Theorem \ref{omegases} and so in all cases we have a short exact sequence
\begin{equation}\label{a}
0 \to q^{-\beta\cdot \gamma}\Delta_\beta\circ \Delta_\gamma\to \Delta_\gamma\circ \Delta_\beta\to C\to 0
\end{equation}
and it suffices to prove that $\operatorname{Ext}^i(C,L)=0$ for all cuspidal $R(\alpha)$-modules $L$.
By adjunction there is an isomorphism
\[
\operatorname{Ext}^i(\Delta_\gamma\circ \Delta_\beta,L)\cong \operatorname{Ext}^i(\Delta_\gamma\otimes \Delta_\beta,\operatorname{Res}_{\gamma\beta}L).
\]
Lemma \ref{oppositerestrict} shows that $\operatorname{Res}_{\gamma\beta}L$ is a $S(\gamma)\otimes S(\beta)$-module. Thus by inductive hypothesis we know that this Ext group is zero.
On the other hand, the group $\operatorname{Ext}^{i-1}(q^{-\beta\cdot \gamma}\Delta_\beta\circ \Delta_\gamma,L)$ is zero by adjunction and the cuspidality of $L$.
Now consider the short exact sequence (\ref{a}) and apply $\operatorname{Hom}(-,L)$ to get a long exact sequence of Ext groups.
In the long exact sequence the group $\operatorname{Ext}^i(C,L)$ is sandwiched between two groups which we have shown to be zero, hence must be zero itself.
\end{proof}
\begin{theorem}\label{realpbw}
Let $\alpha$ be a real root.
Inside $\mathbf{f}^*_{\mathbb{Z}((q))}$ we have $[\Delta(\alpha)]=E_\alpha$.
\end{theorem}
\begin{proof}
By Lemma \ref{14.1}, $\langle[\Delta(\alpha)],[L(\alpha)]\rangle=1$. We know that $\Delta(\alpha)$ only has $L(\alpha)$ appearing as a composition factor, and by Theorem \ref{realdualpbw}, $[L(\alpha)]=E_\alpha^*$. Therefore $\Delta(\alpha)$ is a scalar multiple of $E_\alpha$.
By \cite[Proposition 38.2.1]{lusztigbook}, the automorphisms $T_i$ preserve $(\cdot,\cdot)$, hence $\langle E_\alpha, E_\alpha^* \rangle =1$ and the scalar is 1.
\end{proof}
\begin{theorem}\label{premorita}
Let $\alpha$ be a real root.
The endomorphism algebra of $\Delta(\alpha)$
is isomorphic to $\mathbb{Q}}\def\partition{\mathcal{P}[z]$, where $z$ is in degree two.
\end{theorem}
\begin{proof}
As $\Delta(\alpha)$ is the projective cover of $L(\alpha)$ which is the unique simple $S(\alpha)$-module, the dimension of $\operatorname{End}(\Delta(\alpha))$ is equal to the multiplicity of $L(\alpha)$ in $\Delta(\alpha)$. Theorems \ref{realdualpbw} and \ref{realpbw} tell us that $[\Delta(\alpha)]=E_\alpha$ and $[L(\alpha)]=E_\alpha^*$. Since $E_\alpha^*=(1-q^2)E_\alpha$, we have $\dim\operatorname{End}(\Delta(\alpha))=(1-q^2)^{-1}$.
There is an injection from the centre of $S(\alpha)$ into $\operatorname{End}(\Delta(\alpha))$. By Lemma \ref{scentretrick}, there is an injection from $\mathbb{Q}}\def\partition{\mathcal{P}[z]$ into $\operatorname{End}(\Delta(\alpha))$. A dimension count shows that this injection must be a bijection, as required.
\end{proof}
\begin{corollary}\label{morita}
Let $\alpha$ be a positive real root. Then the algebras $S(\alpha)$ and $\mathbb{Q}}\def\partition{\mathcal{P}[z]$ are graded Morita equivalent.
\end{corollary}
\begin{proof}
The module $\Delta(\alpha)$ is a projective generator for the category of $S(\alpha)$-modules and its endomorphism algebra is $\mathbb{Q}}\def\partition{\mathcal{P}[z]$.
\end{proof}
\section{Standard Imaginary Modules}
\begin{lemma}\label{negativehoms}
Let $d\leq 0$ be an integer and let $\om$ and $\om'$ be two chamber coweights. Then \[
\dim\operatorname{Hom}(\Delta(\w),\Delta(\w'))_d = \begin{cases} \mathbb{Q}}\def\partition{\mathcal{P} & \text{if $d=0$ and $\w=\w'$,} \\
0 & \text{otherwise}.
\end{cases}
\]
\end{lemma}
\begin{proof}
Since $\Delta(\om)$ is a projective $S(\delta)$-module, the dimension of $\operatorname{Hom}(\Delta(\om),\Delta(\om'))$ is equal to the multiplicity of $L(\w)$ in $\Delta(\om')$.
We have
\[
\langle [L(\om)],[L(\om')] \rangle \in \delta_{\om\w'} + q \mathbb{Z}[[q]]
\]
and by Lemma \ref{14.1}, the bases $\{\Delta(\om)\}$ and $\{L(\om)\}$ are dual bases for the subspace of $\mathbf{f}_\delta$ spanned by the cuspidal modules. Therefore
\[
[\Delta(\w)]\in [L(\w)] + \sum_{x\in \Omega} q\mathbb{Z}[[q]]\cdot [L(x)]
\]
which shows the desired properties of the multiplicities.
\end{proof}
\begin{lemma}\label{impr}
The module $\dct{\om\in\Om}\Delta(\w)^{\circ n_\w}$ is a projective object in the category of $S(n\delta)$-modules.
\end{lemma}
\begin{remark}
We choose an arbitrary ordering of the factors in $\circ_{\om\in\Om}\Delta(\w)^{\circ n_\w}$. Lemma \ref{deltascommute} below shows that this choice of ordering is immaterial.
\end{remark}
\begin{proof}
Let $L$ be a semicuspidal $R(n\delta)$-module.
Therefore $\operatorname{Res}_{\delta,\ldots,\delta}L$ is a $S(\delta)\otimes\cdots\otimes S(\delta)$-module. By adjunction
\[
\operatorname{Ext}^1 (\dct{\om\in\Om}\Delta(\w)^{\circ n_\w},L)=\operatorname{Ext}^1(\bigotimes_{w\in \Om} \Delta(\om)^{\otimes n_\w},\operatorname{Res}_{\delta,\ldots,\delta}L)
\]
and since each $\Delta(\om)$ is a projective $S(\delta)$-module, this $\operatorname{Ext}^1$ group is trivial, as required.
\end{proof}
\begin{lemma}\label{deltascommute}
Let $\om$ and $\om'$ be two chamber coweights. Then $\Delta(\om)\circ \Delta(\om')\cong \Delta(\om')\circ \Delta(\om)$.
\end{lemma}
\begin{proof}
We assume that $\w\neq \w'$ as otherwise the result is trivial. By Lemma \ref{negativehoms} and a computation using adjunction and the Mackey filtration, we compute $\operatorname{End}(\Delta(\w)\circ\Delta(\w'))_0\cong \mathbb{Q}}\def\partition{\mathcal{P}$. Hence $\Delta(\w)\circ\Delta(\w')$ is indecomposable.
By Lemma \ref{impr} the module $\Delta(\om)\circ \Delta(\om')$ is a projective $S(2\delta)$-module which surjects onto $L(\om)\circ L(\om')$, hence is the projective cover of $L(\w)\circ L(\w')$ in the category of $S(2\delta)$-modules. By Lemma \ref{risq}, $L(\w)\circ L(\w')\cong L(\w')\circ L(\w)$, hence their projective covers are isomorphic.
\end{proof}
\begin{theorem}
Let $\{m_\w\}_{\w\in\Omega}$ and $\{n_\w\}_{\w\in\Omega}$ be two collections of natural numbers with $\sum_\w m_\w=\sum_\w n_\w$ and let $d\leq 0$ be an integer. Then
\[
\operatorname{Hom} ( \dct{\w\in\Omega} \Delta(\om)^{\circ m_\w},\dct{\w\in\Omega}\Delta(\w)^{\circ n_\w})_d\cong
\begin{cases}
\bigotimes_{\w\in\Omega} \mathbb{Q}}\def\partition{\mathcal{P}[S_{n_\w}] &\text{if $m_\w=n_\w$ for all $\w$ and $d=0$} \\
0 & \text{otherwise}
\end{cases}
\]
\end{theorem}
\begin{proof}
The Mackey filtration for $\operatorname{Res}_{\delta,\ldots,\delta}(\ic{\w\in\Omega}\Delta(\w)^{\circ n_\w})$ has $(\sum_{\w}n_\w)!$ nonzero subquotients, each a tensor product of projective $S(\delta)$-modules where the factor $\Delta(\om)$ appears $n_\w$ times.
Therefore the filtration splits, and by Lemma \ref{negativehoms} and adjunction, the Hom space under question is zero unless $m_\w=n_\w$ for all $\w$ and $d=0$. Furthermore in this case its dimension is $\prod_w n_\w!$.
Since $\circ_{\w\in\Omega}\Delta(\w)^{\circ n_\w}$ is a projective $S(n\delta)$-module and $\circ_{\w\in\Omega}L(\w)^{\circ n_\w}$ is a quotient of $\circ_{\w\in\Omega}\Delta(\w)^{\circ n_\w}$, every endomorphism of $\circ_{\w\in\Omega}L(\w)^{\circ n_\w}$ lifts to an endomorphism of $\circ_{\w\in\Omega}\Delta(\w)^{\circ n_\w}$. From the dimension counts in the previous paragraph and (\ref{endsn}), this lift is unique in degree zero and hence we get an algebra isomorphism
\[
\operatorname{End}(\dct{\w\in\Omega}\Delta(\w)^{\circ n_\w})_0 \cong \operatorname{End}(\dct{\w\in\Omega}L(\w)^{\circ n_\w}).
\]
So the result follows from (\ref{endsn}).
\end{proof}
For a multipartition $\uline{\lambda}=\{\lambda_\w\}_{\w\in\Omega}$ where each $\lambda_w$ is a partition of $n_\w$, we define
\[
\Delta(\uline{\lambda}) = \operatorname{Hom}_{\otimes \mathbb{Q}}\def\partition{\mathcal{P}[S_{n_\w}]} (\otimes S^{\lambda_w},\dct{\w\in\Omega}\Delta(\w)^{\circ n_\w})
\]
\begin{theorem}
The modules $\Delta(\lambda)$ behave in the following way under induction and restriction.
\[
\Delta(\ula)\circ \Delta(\umu)\cong \bigoplus_{\unu}\Delta(\unu)^{\oplus c^{\unu}_{\ula\umu}}
\]
\[
\operatorname{Res}_{k\delta,(n-k)\delta} \Delta(\unu) = \bigoplus_{\ula\vdash k,\umu\vdash n-k} \Delta(\ula)\otimes \Delta(\umu)^{\oplus c^{\unu}_{\ula\umu} }
\]
\end{theorem}
\begin{proof}
The proof is the same as that of Theorem \ref{lrl}
\end{proof}
Let $f_\lambda$ be the dimension of the Specht module $S^\lambda$ and for a multipartition $\ula=\{\lambda_\w\}_{\w\in\Omega}$, let $f_\ula=\prod_{\w} f_{\lambda_w}$.
As a $\mathbb{Q}}\def\partition{\mathcal{P}[S_n]$-module, $\mathbb{Q}}\def\partition{\mathcal{P}[S_n]$ decomposes as $\mathbb{Q}}\def\partition{\mathcal{P}[S_n]=\oplus_{\lambda} (S^\lambda) ^{\oplus f_\lambda}$. Therefore we obtain the decomposition
\begin{equation}\label{decom}
\dct{w\in \Omega}\Delta(\w)^{\circ n_w} \cong \bigoplus_{\uline{\lambda}\vdash n} \Delta(\uline{\lambda}) ^{\oplus f_{\uline{\lambda}}}.
\end{equation}
\begin{lemma}\label{19.4}
Let $\uline{\lambda}$ be a multipartition of $n$. The module $\Delta(\uline{\lambda})$ is indecomposable.
\end{lemma}
\begin{proof}
From the decomposition (\ref{decom}) we obtain inclusions
\begin{equation}\label{inclusion}
\bigoplus_\ula \,\mbox{Mat}\,_{f_\ula}(\mathbb{Q}}\def\partition{\mathcal{P}) \subset \bigoplus_\ula\,\mbox{Mat}\,_{f_\ula}(\operatorname{End}(\Delta(\uline{\lambda}))\subset\operatorname{End}(\dct{\w\in\Omega}\Delta(\w)^{\circ n_\w}).
\end{equation}
Comparing dimensions shows that these inclusions are isomorphisms in degree zero. Therefore $\operatorname{End}(\Delta(\uline{\lambda}))_0$ is isomorphic to $\mathbb{Q}}\def\partition{\mathcal{P}$, hence $\Delta(\ula)$ is indecomposable.
\end{proof}
\begin{lemma}\label{19.5}
If $\ula\neq \uline{\mu}$, then $\Delta(\ula)$ is not isomorphic to any grading shift of $\Delta(\uline{\mu})$.
\end{lemma}
\begin{proof}
Let $i\leq 0$ be an integer. The inclusions in (\ref{inclusion}) are all isomorphisms in degrees less than or equal to zero. Therefore $\operatorname{Hom}(\Delta(\lambda),\Delta(\mu))_i=0$ and thus $\Delta(\lambda)$ is not isomorphic to $q^i \Delta(\mu)$. Similarly $\Delta(\mu)$ is not isomorphic to $q^i \Delta(\lambda)$.
\end{proof}
\begin{theorem}
The set $\Delta(\underline{\lambda})$ is a complete set of indecomposable projective $S(n\delta)$-modules.
\end{theorem}
\begin{proof}
The module $\Delta(\uline{\lambda})$ is a direct summand of $\circ_{\w\in\Omega} \Delta(\w)^{\circ n_\w}$ which is projective by Lemma \ref{impr}, hence $\Delta(\ula)$ is projective. Lemmas \ref{19.4} and \ref{19.5} ensure that the set $\{\Delta(\ula)\}$ is an irredundant set of indecomposable projective $S(n\delta)$-modules, up to a grading shift. The number of indecomposable projective $S(n\delta)$-modules is equal to the number of semicuspidal $R(n\delta)$-modules. This number is known by Theorem \ref{numberofimaginarysemicuspidals}, hence we have found all of the indecomposable projectives.
\end{proof}
\begin{theorem}\label{19.7}
The set $\{L(\underline{\lambda})\}_{\ula\vdash n}$ is a complete set of self-dual irreducible $S(n\delta)$-modules.
\end{theorem}
\begin{proof}
The set $\Delta(\uline{\lambda})$ is a complete set of indecomposable projectives, so the set $\operatorname{hd}\Delta(\uline\lambda)$ is a complete set of irreducible $S(n\delta)$-modules. Since $\Delta(\uline\lambda)$ surjects onto $L(\uline\lambda)$, the set $\hd(L(\uline\lambda))$ is a complete set of irreducible $S(n\delta)$-modules.
So it suffices to prove that $L(\uline\lambda)$ is irreducible.
Let $X$ be a simple submodule of $L(\uline\lambda)$. Then $X$ is semicuspidal so is of the form $\hd(L(\uline\mu))$ for some multipartition $\uline\mu$. Therefore we get a nonzero morphism from $L(\uline\mu)$ to $L(\uline{\lambda})$. From the decomposition $\dct{\w\in\Omega}L(\w)^{\circ n_\w}\cong\oplus_\ula \Delta(\ula)^{\oplus f_{\ula}}$ we obtain inclusions
\begin{equation}\label{inclusion2}
\bigoplus_{\ula\vdash n} \,\mbox{Mat}\,_{f_\ula}(\mathbb{Q}}\def\partition{\mathcal{P}) \subset \bigoplus_{\ula\vdash n}\,\mbox{Mat}\,_{f_\ula}(\operatorname{End}(L(\ula))\subset\operatorname{End}(\dct{\w\in\Omega}L(\w)^{\circ n_\w}).
\end{equation}
Comparing dimensions shows that these inclusions are equalities and hence all morphisms from $L(\uline\mu)$ to $L(\uline\lambda)$ are either zero or isomorphisms. Hence $L(\uline{\lambda})$ must be irreducible, as required. The self-duality of $L(\ula)$ is immediate from the self-duality of $L(\om)$ and (\ref{dualofaproduct}).
\end{proof}
\begin{theorem}\label{14.1nd}
Let $\ula$ and $\umu$ be two multipartitions. Then
\[
\operatorname{Ext}^i(\Delta(\uline{\lambda}),L(\uline{\mu})) = \begin{cases}
\mathbb{Q}}\def\partition{\mathcal{P} &\text{if $\uline{\lambda}=\uline{\mu}$ and $i=0$}, \\
0 &\text{otherwise}.
\end{cases}
\]
\end{theorem}
\begin{proof}
In the course of proving Theorem \ref{19.7}, the module $\Delta(\uline{\lambda})$ was shown to be the projective cover of the irreducible module $L(\uline{\lambda})$ in the category of $S(n\delta)$-modules. This takes care of the $i=0$ case.
Now suppose that $i>0$. Since $\Delta(\uline{\lambda})$ is a direct summand of $\circ_{\w\in\Omega}\Delta(\w)^{\circ n_\w}$, it suffices to show that
\[
\operatorname{Ext}^i(\dct{\w\in\Omega}\Delta(\w)^{\circ n_\w},L(\uline{\mu}))=0.
\]
The module $\operatorname{Res}_{\delta,\ldots,\delta}L(\uline{\mu})$ has all composition factors a tensor product of cuspidal $R(\delta)$-modules. The result now follows from adjunction and Theorem \ref{14.1}.
\end{proof}
\begin{corollary}\label{19.10}
Let $\ula$ and $\uline{\mu}$ be two multipartitions. Then $\langle [\Delta(\ula)],[L(\umu)]\rangle=\delta_{\ula,\umu}$.
\end{corollary}
\section{The Imaginary Part of the PBW Basis
We now follow \cite{beckcharipressley} and define the imaginary root vectors. For comparison with their paper, we note that our $q$ is their $q^{-1}$. We will not be able to cite results from \cite{beckcharipressley} since they only work with convex orders of a particular type.
The aim of this section is to describe a purely algebraic construction of the PBW basis. We will prove that this algebraic construction agrees with the one coming from KLR algebras in Theorem \ref{pbwcat}.
Let $\om$ be a chamber coweight adapted to $\prec$. We first define elements $\psi_n^\om$ by
\[
\psi_n^\om= E_{n\delta-\om_+}E_{\om_+}-q^2 E_{\om_+} E_{n\delta-\om_+}.
\]
Before we continue, we show that the $\psi_n^\om$ lie in a commutative subalgebra of $\mathbf{f}$.
\begin{theorem}
If $L_1$ and $L_2$ are irreducible semicuspidal representations of $R(n_1\delta)$ and $R(n_2\delta)$ respectively, then $L_1\circ L_2\cong L_2\circ L_1$.
\end{theorem}
\begin{proof}
The modules $L_1$ and $L_2$ are both direct summands of modules of the form $\circ_{\om} L(\om)^{\circ n_\om}$. The space of homomorphisms between two modules of this form has already been computed to be concentrated in degree zero. Therefore $\operatorname{Hom}(L_1,L_2)$ is concentrated in degree zero. By the same argument as in the proof of Lemma \ref{risq}, the $R$-matrices $r_{L_1,L_2}$ and $r_{L_2,L_1}$ are inverse isomorphisms.
\end{proof}
\begin{corollary}
The subalgebra of $\mathbf{f}$ spanned by all semicuspidal representations of $R(n\delta)$ is commutative.
\end{corollary}
\begin{lemma}\label{20.3}
Let $\w\in \Om$ and $n\in\mathbb{N}$.
There exist semicuspidal representations $X$ and $Y$ of $R(n\delta)$ with $[X]-[Y]=\psi_n^\om$.
\end{lemma}
\begin{proof}
The same argument as in the proof of Lemma \ref{mpses} shows that we can take $X$ and $Y$ to be the cokernel and kernel of a map from $q^2\Delta(\om_+)\circ \Delta(n\delta-\om_+)$ to $\Delta(n\delta-\om_+)\circ \Delta(\om_+)$.
\end{proof}
\begin{corollary}\label{commute}
The elements $\psi^\w_n$ commute with each other.
\end{corollary}
Now we return to defining the imaginary part of the PBW basis and recursively define elements $P_n^\om$ by $P_0^\om=1$ and
\[
P^\om_n=\frac{1}{[n]}\sum_{s=1}^n q^{n-s} \psi^\om_s P^\om_{n-s}.
\]
Let $\lambda=(\lambda_1\geq\lambda_2\geq\cdots)$ be a partition and let $t\geq \ell(\lambda)$ be an integer. We define
\[
S_\lambda^\om = \det (P^\om_{\lambda_i-i+j})_{1\leq i,j\leq t}.
\]
By Corollary \ref{commute} the entries in this matrix all commute with each other so there is no ambiguity in the definition of the determinant.
The elements $P^\om_n$ here should be thought of as playing the role of the complete symmetric functions in the ring of all symmetric functions. This determinental definition shows that the elements $S^\om_\lambda$ are playing the role of the Schur functions. This point of view makes it clear that the definition of $S^\om_\lambda$ does not depend on $t$.
Let $\pi=(\beta_1^{m_1},\ldots,\underline{\lambda},\ldots,\gamma_1^{n_1})$ be a root partition. The PBW basis element $E_\pi$ is defined to be
\begin{equation}\label{pbwalg}
E_\pi = E_{\beta_1}^{(m_1)}\cdots E_{\beta_k}^{(m_k)} \left( \prod_{\w\in \Omega}S_{\lambda_\w}^\w \right)E_{\gamma_l}^{(m_l)} \cdots E_{\gamma_1}^{(n_1)}.
\end{equation}
\section{MV Polytopes
\begin{definition}
Let $M$ be an $R(\nu)$-module. The MV polytope of $M$, denoted $P(M)$, is the convex hull of the set
\[
\{ \mu \mid \operatorname{Res}_{\mu,\nu-\mu}M\neq 0\}
\]
\end{definition}
Let $\om$ be a chamber coweight. The $\om$-face of a polytope $P$ is defined to be the intersection of $P$ with the plane spanned by $\om_+$ and $\om_-$. The term face is justified by the following result.
\begin{proposition}
Suppose that $\om$ is adapted to the convex order $\prec$. If $L(\pi)$ is a simple module for some root partition $\pi$ such that the support of $\pi$ is contained in the span of $\om_-$ and $\om_+$, then the $\om$ face of $P(L(\pi))$ is a (possibly degenerate) 2-face of $P(L(\pi))$.
\end{proposition}
\begin{definition}
Let $\lambda$ be the functional on the span of $\om_-$ and $\om_+$ such that $\lambda(\om_+)=1$ and $\lambda(\om_-)=-1$.
The width of the $\om$-face of a polytope $P$ is equal to the maximum value of $\lambda(p)-\lambda(q)$ where $p$ and $q$ are two points in the $\om$-face of $P$.
\end{definition}
\begin{example}
The width of the $\om$-face of $P(L(n\delta-\om_+))$ is $n$.
\end{example}
We know this because a MV polytope is completely determined by its 2-faces, which are MV polytopes for rank two root systems.
In the rest of this section, we fix a choice of chamber coweight $\w$ adapted to $\prec$. Without loss of generality, we may assume that our convex order is of the form of Example \ref{tworow}.
Note that in our labelling of the irreducible semicuspidal modules for $R(\delta)$ by multipartitions, there are choices involved. Namely replacing
$r_{L(\om),L(\om)}$ by
its negative results in replacing the partition $\lambda_\w$ by its transpose. We make a choice of sign in $r_{L(\om),L(\om)}$ such that $L_\w((2))$ has $\om$-width 2.
The reason that such a choice is always possible is that the module $L(\om')\circ L(\om'')$ will only have $\w$-width at least two if $\w=\w'=\w''$ and by the Tingley-Webster classification, there exists a unique MV polytope for $2\delta$ of $\w$-width 2. It must thus come from one of the summands of $L(\om)\circ L(\om)$ and we may replace our $R$-matrix with its negative if necessary to ensure that this summand is the one indexed by the partition $(2)$.
This means that the $\omega$-face of the MV polytope for $L(\om,(1^2))$ is
\[
\begin{tikzpicture}
\node at (3,0){
\tikz[yscale=0.4, xscale=1.2] {
\node at (0,0) {$\bullet$};
\node at (0,4) {$\bullet$};
\node at (-1,1) {$\bullet$};
\node at (-1,3) {$\bullet$};
\draw [line width = 0.04cm] (0,0) -- (-1,1);
\draw [line width = 0.04cm] (-1,3) -- node[midway,left] {$(1)$}(-1,1);
\draw [line width = 0.04cm] (0,4) -- (-1,3);
\draw [line width = 0.04cm] (0,0) -- node[right,midway]{$(1,1)$}(0,4);
\draw [line width = 0.01cm, color=gray] (-1,1) -- (0, 2);
\draw [line width = 0.01cm, color=gray] (-1,3) -- (0, 2);
}
};
\end{tikzpicture}
\]
\begin{proposition}
Let $\lambda$ be a partition and $\w$ be a chamber coweight.
The module $L_\om(\lambda)$ has $\w$-width 1 if and only if $\lambda=(1^n)$.
\end{proposition}
\begin{proof}
We prove this proposition by an induction on $n$. The case $n=1$ is trivial and the case $n=2$ is true by the choice of normalisation of the $R$-matrix.
Note that for $\w'\neq\w$, the module $L(\w')$ has $\w$-width zero. Therefore the $\w$-width of $\dct{x\in \Omega} L(\lambda_x)$ is equal to the $\w$-width of $L(\lambda_w)$.
Therefore by induction we know exactly how many $\w$-faces of modules of the form $\dct{x\in\Omega} L(\lambda_x)$ with $|\lambda_w|<n$ have width less than or equal to one. By \cite{tingleywebster} this comprises all MV polytopes of $\w$-width less than or equal to one except for one polytope of $\w$-width one. Therefore there exists some partition $\mu\vdash n$ for which $L_\w(\mu)$ has $\w$-width one.
The restriction $\operatorname{Res}_{k\delta,(n-k)\delta}L(\mu)$ can only have composition factors $L_1\otimes L_2$ where $L_1$ and $L_2$ have $\w$-width at most one. These restrictions are given by the Littlewood-Richardson rule (\ref{lrl}). So by induction the only option is $\operatorname{Res}_{k\delta,(n-k)\delta}L(\mu)\cong L_\w(1^k)\otimes L_\w(1^{n-k})$ which for $n>2$ forces $\mu=(1^n)$, completing the proof.
\end{proof}
\begin{theorem}\label{22.6}
\[
\operatorname{Res}_{n\delta-\om_+,\om_+} L_\om(1^n)\cong A(L_\om(1^n),L(\om_-))\otimes L(\om_+).
\]
\end{theorem}
\begin{proof}
We perform an expansion in the dual PBW basis
\[
[\operatorname{Res}_{n\delta-\om_+,\om_+} L_\om(1^n)] =\sum_{\sigma,\pi} c_{\sigma,\pi} E_\sigma^* \otimes E_\pi^*.
\]
Then
\[
c_{\sigma,\pi} =\langle E_\sigma\otimes E_\pi,[\operatorname{Res}_{n\delta-\om_+,\om_+} L_\om(1^n)]\rangle = \langle E_\sigma E_\pi, [L_\om(1^n)]\rangle.
\]
We consider the algorithm of \S \ref{lsformula} which teaches us how to write the product $E_\sigma E_\pi$ in terms of the PBW basis.
Let $\pi_k$ be the smallest root appearing in $\pi$. If $\pi\neq \om_+$ then as $\w_+-\pi_k\in \mathbb{N} I$, it must be that $\pi_k \preceq \om_-$. Therefore at all stages in applying the algorithm for writing $E_\sigma E_\pi$ in terms of the PBW basis, any term $E_{\gamma_1}\cdots E_{\gamma_l}$ which appears has $\gamma_l\preceq \om_- \prec \delta$. Therefore no purely imaginary terms in the PBW basis can appear, and as $[L_\om(1^n)]$ is orthogonal to all PBW elements which are not purely imaginary, $c_{\sigma,\pi}=0$ for such $\pi$.
So we may assume $\pi=\om_+$.
Let $\sigma_k$ be the smallest root appearing in $\sigma$. Suppose that $\sigma_k$ is not of the form $m\delta+\om_-$. Then $\sigma_k\prec \om_-$. At the first stage of applying our algorithm, up to two terms $E_{\gamma_1}\cdots E_{\gamma_l}$ appear. One term has $\gamma_l=\sigma_k\prec w_-$ while the other term, if it exists, has $\gamma_l=\sigma_l+\om_+$ which is also less than $\om_-$, since by convexity it is less than $\om_+$ and we know all roots between $\om_+$ and $\om_-$. By the same argument as in the previous paragraph, $c_{\sigma,\pi}=0$ in this case too.
Therefore, when $c_{\sigma,\pi}\neq 0$, all roots that appear in $\sigma$ are all in the span of $\om_-$ and $\om_+$. This implies that every irreducible subquotient of $\operatorname{Res}_{n\delta-\om_+,\om_+} L_\om(1^n)$ is of the form $L(\sigma)\otimes L(\om_+)$ for some such root partition $\sigma$.
The largest root appearing in $\sigma$ is at most $\delta$ as $L_\om(1^n)$ is cuspidal. Therefore $\sigma=(\underline{\lambda},m\delta-\om_+)$ for some multipartition $\underline{\lambda}$ and positive integer $m$.
The $\om$-face of $P(L(m\delta-\om_+))$ has width $m$. Therefore the $\om$-face of $P(L(\sigma))$ has width at least $m$. As the $\om$-face of $P(L(\sigma))$ is a subset of the $\om$-face of $P(L_\om(1^n))$ which as width one, $m=1$.
Now by Theorem \ref{lrl},
\[
\operatorname{Res}_{(n-1)\delta,\delta} L_\om(1^n) \cong L_\om(1^{n-1})\otimes L(\om).
\]
Therefore the only option for $\underline{\lambda}$ is $1^n$ at $\omega$ and zero elsewhere, and furthermore $L(1^n_\om,\w_-)\otimes L(\om_+)$ must appear with multiplicity one, completing the proof.
\end{proof}
\begin{lemma}\label{zerocase}
Let $\w$ be a chamber coweight and $\alpha=\w_-$. There is a short exact sequence
\[
0\to q L(\alpha+\delta)\to L(\w)\circ L(\alpha)\to L(\w,\alpha)\to 0.
\]
\end{lemma}
\begin{proof}
Theorem \ref{main} tells us that $L(\w,\alpha)$ is the head of the module $L(\w)\circ L(\alpha)$ and that every other subquotient of $L(\w)\circ L(\alpha)$ is cuspidal. Therefore there is a short exact sequence
\[
0 \to X \to L(\w)\circ L(\alpha)\to L(\w,\alpha)\to 0
\]
for some cuspidal $R(\alpha+\delta)$-module $X$. Since the head of $L(\w)\circ L(\alpha)$ is known, Lemma \ref{geometry} implies that $[X]\in q\mathbb{N}[q]E_{\alpha+\delta}^*.$ Taking duals there is a short exact sequence
\[
0 \to L(\w,\alpha)\to L(\alpha)\circ L(\w)\to X^\circledast \to 0.
\]
We now consider
\[
\operatorname{Hom}(L(\alpha)\circ L(\w),L(\w)\circ L(\alpha))\cong \operatorname{Hom}(L(\alpha)\otimes L(\w), \operatorname{Res}_{\alpha,\delta}L(\w)\circ L(\alpha)).
\]
The restriction has two nonzero pieces in its Mackey filtration. The module $L(\alpha)\otimes L(\w)$ appears as a quotient and we use Lemma \ref{reslw} to identify the submodule as $L(\alpha)\otimes (L(\delta-\alpha)\circ L(\alpha))$.
Now we consider
\[
\operatorname{Hom}(L(\w),L(\delta-\alpha)\circ L(\alpha))\cong \operatorname{Hom}(q^2 L(\alpha)\otimes L(\delta-\alpha),L(\alpha)\otimes L(\delta-\alpha))
\]
where we have used the adjunction (\ref{oppositeadjunction}) and Lemma \ref{reslw} to reach this isomorphism. Therefore there is a unique (up to scalar) morphism from $L(\alpha)\circ L(\w)$ to $L(\w)\circ L(\alpha)$ in degree 2, and the only other possible morphisms are in degree zero from the other term in the Mackey filtration. When comparing this with $[X]\in q\mathbb{N}[q]E_{\alpha+\delta}^*$, the only option is that $X\cong q L(\alpha+\delta)$, as required.
\end{proof}
\begin{lemma}\label{zerocase2}
Let $\w$ be a chamber coweight and $\alpha=\w_-+n\delta$ for some natural number $n$. Then there are short exact sequences
\[
0\to qL(\alpha+\delta) \to L(\w)\circ L(\alpha) \to L(\w,\alpha)\to 0 \]\[
0 \to \Delta(\w)\circ\Delta(\alpha) \to \Delta(\alpha) \circ \Delta(\w) \to q\Delta(\alpha+\delta)\oplus q^{-1} \Delta(\alpha+\delta)\to 0.
\]
\end{lemma}
\begin{proof}
We prove the existence of these short exact sequences by an induction on $n$. The case $n=0$ for the first sequence is Lemma \ref{zerocase}. First we prove the existence of the first sequence for some $n>0$, assuming that both sequences are known for lesser values of $n$.
As in the proof of Lemma \ref{zerocase}, we have a short exact sequence\[
0 \to X \to L(\w)\circ L(\alpha)\to L(\w,\alpha)\to 0
\]
where $[X]\in q\mathbb{N}[q]E_{\alpha+\delta}^*$, and we wish to study
\[
\operatorname{Hom}(L(\alpha)\circ L(\w),L(\w)\circ L(\alpha))\cong \operatorname{Hom}(L(\alpha)\otimes L(\w), \operatorname{Res}_{\alpha,\delta}L(\w)\circ L(\alpha)).
\]
The Mackey filtration of $\operatorname{Res}_{\alpha,\delta}(L(\w)\circ L(\alpha))$ has two nonzero pieces. The module $L(\alpha)\otimes L(\w)$ appears as a quotient, and to understand the submodule, we need to first understand $\operatorname{Res}_{\alpha-\delta,\delta}L(\alpha)$.
By Lemma \ref{oppositerestrict}, we can write
\[
[\operatorname{Res}_{\alpha-\delta,\delta}L(\alpha)] = \sum_{x\in \Om} g_x(q) [L(\alpha-\delta)]\otimes [L(x)]
\]
for some polynomials $g_x(q)\in \mathbb{N}[q,q^{-1}]$ which satisfy $g_x(q)=g_x(q^{-1})$ since restriction commutes with duality.
For $x\in \Omega$, let $C_x$ be the projective $S(\alpha-\delta)$-module which appears in the short exact sequence of Lemma \ref{mpses}:
\[
0\to \Delta(x)\circ \Delta(\alpha-\delta) \to \Delta(\alpha-\delta)\circ \Delta(x)\to C_x\to 0.
\]
We compute
\begin{align*}
g_x(q)&=\langle E_{\alpha-\delta}\otimes E_x, [\operatorname{Res}_{\alpha-\delta}L(\alpha)] \rangle \\
&= \langle E_{\alpha-\delta}E_x,[L(\alpha)] \rangle \\
&= \langle E_{\alpha-\delta}E_x-E_x E_{\alpha-\delta},[L(\alpha)] \rangle \\
&= \langle [C_x],[L(\alpha)]\rangle.
\end{align*}
Therefore $C_x\cong g_x(q)\Delta(\alpha)$.
If $x\neq \w$ then we can compute the value of $[C_x]$ after specialising $q=1$ in $\mathbf{f}$ to obtain $g_x(1)$ is 0 or 1, which forces $g_x(q)$ to be 0 or 1.
For $x=\w$, we use the inductive hypothesis applied to the second short exact sequence to conclude that $g_\w(q)=q+q^{-1}$. Therefore $\operatorname{Res}_{\alpha,\delta} L(\w)\circ L(\alpha-\delta)$ has a submodule isomorphic to $q(L(\w)\circ L(\alpha-\delta) )\otimes L(\w)$.
By the inductive hypothesis this module receives a map from $q^2 L(\alpha)\otimes L(\w)$ and hence there exists a morphism from $L(\alpha)\circ L(\w)$ to $L(\w)\circ L(\alpha)$ of degree two. In fact this argument shows us we know even more, namely that all other morphisms between these modules are of degree zero. So the same argument as in Lemma \ref{zerocase} allows us to conclude $X\cong qL(\alpha+\delta)$, as required.
Now we deduce the second short exact sequence from the first. By Lemma \ref{mpses}, there exists a short exact sequence
\[
0\to \Delta(\w)\circ \Delta(\alpha) \to \Delta(\alpha)\circ \Delta(\w)\to C\to 0
\]
where $C$ is a projective $S(\alpha+\delta)$-module, hence isomorphic to $f(q)$ copies of $\Delta(\alpha+\delta)$ for some $f(q)\in\mathbb{N}[q,q^{-1}]$. The same argument computing pairings as above shows that $f(q)$ is equal to the multiplicity of $L(\alpha)\otimes L(\om)$ in $\operatorname{Res}_{\alpha,\delta}L(\alpha+\delta)$. The computation in $\mathbf{f}$ specialised at $q=1$ shows $f(1)=2$, and since $f(q)=f(q^{-1})$, we have $f(q)=q^i+q^{-i}$ for some $i\in\mathbb{Z}$.
The first exact sequence gives us a morphism from $qL(\alpha+\delta)$ to $L(\om)\circ L(\alpha)$ which by adjunction induces a nonzero morphism $\operatorname{Res}_{\alpha,\delta}L(\alpha+\delta)\to L(\alpha)\otimes L(\w)$. Therefore $i=1$, as required.
\end{proof}
\begin{proposition}\label{resolution}
Let $k$ and $l$ be positive integers. There is a short exact sequence
\[
0\to q A(L_\om(1^k),L((l+1)\delta-\om_+))\to L_\om(1^{k+1})\circ L(l\delta-\om_+)\to A(
L_\om(1^{k+1}),L(l\delta-\om_+))\to 0.
\]
\end{proposition}
\begin{proof}
This proof proceeds by an induction.
By Theorem \ref{main}, the module $L_\om(1^{k+1})\circ L(l\delta-\om_+)$ surjects onto $A(
L_\om(1^{k+1}),L(l\delta-\om_+))$ and all other subquotients are of the form $X_{\uline{\lambda},m}^i=q^iA(L(\underline{\lambda}),L((l+m)\delta-\alpha))$ for some $m>0$ and $\underline{\lambda}$ a multipartition of $k+1-m$.
Setting $n=k+1-m$, the following computation is straightforward as there is only one nonzero piece in the Mackey filtration.
\[
\operatorname{Res}_{n\delta,(l+m)\delta-\om_+} (L_\om(1^{k+1})\circ L(l\delta-\om_+) )\cong L_\om(1^n)\otimes
\left( L_\om(1^m))\circ L(l\delta-\om_+) \right)
\]
Note that if $X_{\uline{\lambda},m}^i$ is a subquotient of $L_\om(1^{k+1})\circ L(l\delta-\om_+)$ then $L(\uline{\lambda})\otimes L((l+m)\delta-\om_+)$ must appear as a subquotient of this restriction.
Immediately we see that $\lambda_\w=(1^n)$ and $\lambda_x=0$ for all other chamber coweights $x$.
Consider a subquotient of the form $X_{\uline{\lambda},m}^i$ with $\uline{\lambda}\neq 0$. Then by inductive hypothesis we know all that there is only a cuspidal subquotient of $ L_\om(1^m)\circ L(l\delta-\om_+)$ when $m=1$. Furthermore this cuspidal subquotient appears with multiplicity $q$, which completes the proof in this case.
So now turn our attention to the remaining case when $n=0$. The module $L_\om(1^{k+1})\circ L(l\delta-\om_+)$ has $\om$-width $l+1$ and the module $L((l+m)\delta-\w_+)$ has $\w$-width $l+m$. Therefore $m=1$. The result now follows from Lemma \ref{zerocase2}. \end{proof}
\section{Inner Product Computations}
For any natural number $n$ and chamber coweight $\w$, define $e_n^\om=[L_\om(1^n)]$.
\begin{lemma}\label{psinen}
Let $\om$ be a chamber coweight and $\{n_x\}_{x\in \Om}$ a collection of natural numbers with sum $n$. Then
\[
\langle \psi_n^\om ,\prod_{x\in\Om}e_{n_x}^x \rangle = \begin{cases}
(-q)^{n-1} &\text{if $n_x=0$ for all $x\neq \w$,} \\
0 & \text{otherwise}.
\end{cases}
\]
\end{lemma}
\begin{proof}
By Theorem \ref{independence}, we assume without loss of generality that our convex order is as in Example \ref{tworow}.
By definition $\psi_n^\om=E_{n\delta-\om_+} E_{\om_+}-q^2 E_{\om_+} E_{n\delta-\om_+}$. Since $\operatorname{Res}_{\om_+,n\delta-\om_+}L=0$ for any semicuspidal representation $L$, we have
\[
\langle \psi_n^\om ,\prod_{x\in\Om}e_{n_x}^x \rangle =\langle E_{n\delta-\om_+}\otimes E_{\om_+}, \prod_{x\in\Om}r(e_{n_x}^x) \rangle.
\]
The terms in the product all commute so without loss of generality we may assume that the $r(e_{n_\om}^\w)$ term is last.
Each term appearing in the product of the $r(e_{n_x}^x)$'s is a product of terms $y\otimes z$ with $y$ of degree at most $\delta$ and $z$ of degree at least $\delta$. Since we need a term of degree $(n\delta-\om_+,\om_+)$, the only option is that exactly one of the terms does not have degree $(n_x\delta,0)$.
That particular term will have degree $(n_x-\om_+,\om_+)$. Now for $r_{n_x-\om_+,\om_+}(e_{n_x}^x)$ to not be zero, it must be that $\operatorname{Res}_{n_x\delta-\om_+,\om_+}L_x(1^{n_x})\neq 0$ and hence the restriction $\operatorname{Res}_{n_x\delta-\om_+,\om_+}L(x)^{\circ n_x}$ is also not zero. By a Mackey argument this implies that $\operatorname{Res}_{\om_-,\om_+}L(x)\neq 0$.
By Lemma \ref{oppositerestrict} there is an injection $L(\om_-)\otimes L(\w_+)\to \operatorname{Res}_{\om_-,\om_+}L(x)$ and so by adjunction there is a nonzero map from $L(\om_-)\circ L(\w_+)$ to $L(x)$. By Theorem \ref{imaginaryclassification} $x=\w$.
Now Theorem \ref{22.6} and Proposition \ref{resolution} tell us that
\[
r_{n_\om\delta-w_+,\om_+}(e_{n_\om}^\om) = \sum_{j=1}^{n_\om} (-q)^{j-1} e_{n-j}^\om E_{j\delta-\om_+}^*\otimes E_{\w_+}^*.
\]
Therefore
\[
\langle \psi_n^\om ,\prod_{x\in\Om}e_{n_x}^x \rangle =\langle E_{n\delta-\om_+},
(\prod_{ \substack{x\in\Om, \\ x\neq \w}} e_{n_x}^x )(\sum_{j=1}^{n_\w}(-q)^{j-1} e_{n-j}^\om E_{j\delta-\om_+}^*)
\rangle.
\]
Since $\operatorname{Res}_{\delta,(n-1)\delta-\w_+}\Delta(n\delta-\w_+)=0$, there is only one possible term which can be nonzero, it only occurs when $n_x=0$ for all $x\neq \w$ and $j=n_\w$. The resulting inner product is easily evaluated to $(-q)^{n-1}$.
\end{proof}
\begin{lemma}\label{pnen}
For $n\geq 0$, we have
\[
\langle P_n^\om ,\prod_{x\in\Om}e_{n_x}^x \rangle
=\begin{cases}
1 &\text{if $n_x=0$ for all $x\neq \w$ and $n_\w\leq 1$,}\\
0 &\text{otherwise}.
\end{cases}
\]
\end{lemma}
\begin{proof}
From the definition of $P_n^\w$,
\begin{align*}
\langle P_n^\om ,\prod _{x\in\Om}e_{n_x}^x \rangle &= \frac{1}{n}\sum_{s=1}^n q^{n-s} \langle
\psi_s^\w P_{n-s}^\w, \prod _{x\in\Om}e_{n_x}^x \rangle. \\
&=\frac{1}{n}\sum_{s=1}^n q^{n-s} \langle
\psi_s^\w \otimes P_{n-s}^\w, \prod _{x\in\Om}r(e_{n_x}^x) \rangle.
\end{align*}
Since each $L_x(1^{n_x})$ is semicuspidal, the only relevant terms in $r(e_{n_x}^x)$ are of bidegree $(k\delta,l\delta)$ for some $k,l\in \mathbb{N}$, and all these terms are known by (\ref{lrr}). Therefore
\[
\langle P_n^\om ,\prod _{x\in\Om}e_{n_x}^x \rangle = \frac{1}{n}\sum_{s=1}^n q^{n-s} \langle
\psi_s^\w \otimes P_{n-s}^\w, \prod _{x\in\Om}\sum_{k=0}^{n_x} e_k^x \otimes e_{n_x-k}^x \rangle.
\]
This can easily be computed by an induction on $n$ together with Lemma \ref{psinen}.
\end{proof}
\section{Symmetric Functions
Let $\Lambda$ be the Hopf algebra of symmetric functions. We consider it over the ground ring $\mathbb{Z}[q,q^{-1}]$. It is isomorphic to $\mathbb{Z}[q,q^{-1}][h_1,h_2,\ldots]$ where $h_n$ is the complete symmetric function. Let $s_\lambda$ be the Schur function indexed by the partition $\lambda$. Let $(\cdot,\cdot)$ denote the usual inner product on $\Lambda$. We denote the coproduct on $\Lambda$ by $\Delta$.
Let $B$ be the subalgebra of $\mathbf{f}^*$ generated by the elements $e_{n}^\w$. For $x\in B$ we define $r_\delta(x)\in B\otimes B$ to be the sum of all terms in $r(x)$ of bidegree $(a\delta,b\delta)$.
\begin{lemma}\label{hopfiso}
There is an isomorphism of Hopf algebras $\psi\map{\Lambda^{\otimes \Omega}}{B}$ with
\[
\psi(\otimes_\w s_{\lambda_w}) = [L(\ula)]
\] where the coproduct on $B$ is $r_\delta$.
\end{lemma}
\begin{proof}
This is immediate from Theorem \ref{lrl}.
\end{proof}
Define an algebra homomorphism $\varphi\map{\Lambda^{\otimes \Omega}}{\mathbf{f}}$ by
\[
\varphi (\otimes_\w h_{n_\w}) = \prod_{\w\in\Omega} P^\w_{n_\w}.
\]
That such a homomorphism exists is because the $h_{n_\w}$ freely generate $\Lambda^{\otimes \Omega}$ as a commutative algebra and Corollary \ref{commute} which implies that the $P_{n_\w}^\w$ lie in a commutative subalgebra of $\mathbf{f}$.
\begin{lemma}\label{comp}
For all $x,y\in\Lambda^{\otimes \Omega}$ we have
\[
\langle \varphi(x),\psi(y) \rangle = (x,y).
\]
\end{lemma}
\begin{proof}
Lemma \ref{pnen} establishes this formula in the special case when $x=P_n^\w$. To deduce the general case from this particular case, we use $(xy,z)=(x\otimes y,z)$ and
$$\langle \varphi(xy),\psi(z)\rangle = \langle \varphi(x) \varphi(y),\psi(z)\rangle = \langle \varphi(x) \otimes \varphi(y), r_\delta (\psi(z)) \rangle = \langle \varphi(x) \otimes \varphi(y),\psi(\Delta(z))\rangle $$
where in the last step we used Lemma \ref{hopfiso}.
\end{proof}
\begin{corollary}\label{pairsandl}
Let $\om$ and $\om'$ be two chamber coweights and let $\lambda$ and $\mu$ be partitions. Then $\langle S_\lambda^\om,[L_{\om'}(\mu)] \rangle = \delta_{\om\om'}\delta_{\lambda\mu}$.
\end{corollary}
\begin{proof}
The Schur functions are orthonormal.
\end{proof}
\begin{theorem}\label{impbw}
Let $\ula=\{\lambda_\w\}_{\w\in\Omega}$ be a purely imaginary root partition. Then
$$[\Delta(\ula)]=\prod_{w\in\Omega}S^\w_{\lambda_\w}$$
\end{theorem}
\begin{proof}
The nondegeneracy of $(\cdot,\cdot)$ together with Lemma \ref{comp} implies that $\varphi$ is injective. By Lemmas \ref{20.3} and \ref{10.2}, the image of $\varphi$ lies in the subspace of $\mathbf{f}_{\mathbb{Z}((q))}^*$ spanned by the semicuspidal modules. A dimension count shows that the image is precisely the span of the semicuspidal modules.
Therefore $\Delta_\w(\ula)$ is a linear combination of the elements $S_{\uline{\mu}}$.
The pairings in Corollary \ref{19.10} and \ref{pairsandl} force $\Delta(\uline{\lambda})=S_{\uline{\lambda}}$.
\end{proof}
\section{Standard Modules
The nil Hecke algebra $NH_n$ is the algebra $R(ni)$ for any $i\in I$.
It is well known that the nil Hecke algebra is a matrix algebra over its centre, see for example \cite[Proposition 2.21]{rouquier2}. In particular, there is an isomorphism
\[
NH_n \cong \,\mbox{Mat}\,_{[n]!}(\mathbb{Q}}\def\partition{\mathcal{P}[x_1,\ldots,x_n]^{S_n})
\]
where each $x_i$ is in degree two.
Let $e_n$ be a primitive idempotent in $NH_n$.
\begin{theorem}
Let $\alpha$ be a real root. There is an isomorphism $\operatorname{End}(\Delta(\alpha)^{\circ n})\cong NH_n$.
\end{theorem}
\begin{proof}
The proof of \cite[\S 3]{bkm} works in this generality without any change.
\end{proof}
For any positive real root $\alpha$ and any positive integer $n$, we define the divided power standard module $\Delta(\alpha)^{(n)}$ to be
\[
\Delta(\alpha)^{(n)} =q^{n(n-1)/2} e_n (\Delta(\alpha)^{\circ n})
\]
\begin{lemma}\label{24.3}
Let $\alpha$ be a real root and $n$ a positive integer. Then
\[
\operatorname{Ext}^i(\Delta(\alpha)^{(n)},L(\alpha)^{\circ n}) \cong \begin{cases}
\mathbb{Q}}\def\partition{\mathcal{P} \text{ if $i=0$} \\
0 \text{ otherwise}
\end{cases}
\]
\end{lemma}
\begin{proof}
We compute by adjunction
\[
\operatorname{Ext}^i(\Delta(\alpha)^{\circ n},L(\alpha)^{\circ n})\cong \operatorname{Ext}^i(\Delta(\alpha)^{\otimes n},\operatorname{Res}_{\delta,\ldots,\delta}L(\alpha)^{\circ n}).
\]
The module $\operatorname{Res}_{\delta,\ldots,\delta}L(\alpha)^{\circ n}$ has a composition series with $n!$ subquotients, each isomorphic to some $q^jL(\alpha)^{\otimes n}$ and $[\operatorname{Res}_{\delta,\ldots,\delta}L(\alpha)^{\circ n}]=[n]!E_\alpha^*\otimes\cdots\otimes E_\alpha^*$. So by Theorem \ref{14.1}, for $i>0$ we have
\[
\operatorname{Ext}^i(\Delta(\alpha)^{\circ n},L(\alpha)^{\circ n})=0
\]
while for $i=0$ we also use the fact that $\Delta(\alpha)$ is the projective cover of $L(\alpha)$ in the category of $S(\alpha)$-modules to obtain
\[
\operatorname{Hom}(\Delta(\alpha)^{\circ n},L(\alpha)^{\circ n})\cong q^{{n \choose 2}} [n]!\mathbb{Q}}\def\partition{\mathcal{P}.
\]
Since $\Delta(\alpha)^{\circ n}\cong q^{{n \choose 2}} [n]!\Delta(\alpha)^{(n)}$, we obtain the desired result.
\end{proof}
Let $\pi=(\beta_1^{m_1},\ldots,\beta_k^{m_k},\underline{\lambda},\gamma_l^{n_l},\ldots,\gamma_1^{n_1})$ be a root partition. We define the corresponding standard module to be
\[
\Delta(\pi)=\Delta(\beta_1)^{(m_1)}\circ\cdots\circ\Delta(\beta_k)^{(m_k)}\circ \Delta(\ula)\circ\Delta(\gamma_l)^{(n_l)}\circ\cdots\circ
\Delta(\gamma_1)^{(n_1)}.
\]
Also define
\[
\nabla(\pi)=\overline{\Delta}(\pi)^\circledast.
\]
\begin{proposition}\label{homological}
Let $\pi$ and $\sigma$ be two root partitions. Then
\[
\operatorname{Ext}^i(\Delta(\pi),\nabla(\sigma)) =\begin{cases}
\mathbb{Q}}\def\partition{\mathcal{P} &\text{if } i=0 \text{ and } \pi=\sigma \\
0 &\text{otherwise}
\end{cases}
\]
\end{proposition}
\begin{proof}
Let $\pi=(\beta_1^{m_1},\cdots,\gamma_1^{n_1})$. Then by adjunction,
\[
\operatorname{Ext}^i(\Delta(\pi),\co(\sigma))=\operatorname{Ext}^i(\Delta(\beta_1)^{(m_1)}\otimes\cdots \otimes \Delta(\gamma_1)^{(n_1)},\operatorname{Res}_\pi\co(\sigma))
\]
and by Theorem \ref{main}, $\operatorname{Res}_\pi\co(\sigma)=0$ unless $\pi\leq\sigma$.
On the other hand, by the adjunction (\ref{oppositeadjunction}),
\[
\operatorname{Ext}^i(\Delta(\pi),\co(\sigma)) = \operatorname{Ext}^i ( \operatorname{Res}_\sigma \Delta(\pi),L(\gamma_1)^{\otimes n_1}\otimes \cdots \otimes L(\beta_1)^{\otimes m_1})
\]
and so again using Theorem \ref{main}, $\operatorname{Res}_\sigma \Delta(\pi)=0$ unless $\sigma\leq \pi$.
Thus the only case to consider is when $\sigma\sim\pi$. Remember that this means that $\sigma$ and $\pi$ agree except for the multipartition they contain. Let $\ula$ be the multipartition in $\pi$ and $\umu$ be the multipartition in $\sigma$.
By Theorem \ref{main}, $$\operatorname{Res}_\pi\co(\pi)\cong \operatorname{Res}_\p(\overline{\Delta}(\p)^\circledast)\cong (\operatorname{Res}_\p \overline{\Delta}(\p))^\circledast \cong L(\beta_1)^{\circ m_1}\otimes \cdots \otimes L(\gamma_1)^{\circ n_1}.$$ Therefore
\[
\operatorname{Ext}^*(\Delta(\pi),\nabla(\sigma))\cong \bigotimes_\alpha \operatorname{Ext}^*(\Delta(\alpha)^{(f_\p(\alpha))},L(\alpha)^{\circ f_\sigma(\alpha)}) \otimes \operatorname{Ext}^*(\Delta(\ula),L(\uline{\mu}))
\] where the tensor product is over all real roots $\alpha$.
The result now follows from Lemma \ref{24.3} and Theorem \ref{14.1nd}.
\end{proof}
\begin{theorem}\label{pbwcat}
Let $\p$ be a root partition.
The class of the standard module $\Delta(\p)$ is the PBW monomial $E_\p$, defined algebraically in (\ref{pbwalg}).
\end{theorem}
\begin{proof}
This follows from Theorems \ref{realpbw} and \ref{impbw}.
\end{proof}
Proposition \ref{homological} proves that the classes of the standard modules $\Delta(\pi)$ and the proper standard modules $\overline{\Delta}(\sigma)$ are orthogonal under $(\cdot,\cdot)$. Therefore the class of each proper standard module is an element of the dual PBW basis. So we have categorified both the PBW and dual PBW basis.
A module $M$ is said to have a $\Delta$-flag if it has a sequence of submodules $0=M_0\subseteq M_1\subseteq \cdots
\subseteq M_{n-1}\subseteq M_n=M$ such that each subquotient $M_{i+1}/M_i$ is isomorphic to $q^m\Delta(\pi)$ for some integer $m$ and some root partition $\pi$.
\begin{theorem}\label{deltaflag}
Let $M$ be a finitely generated $R(\nu)$-module such that $\operatorname{Ext}^1(M,\nabla(\pi))=0$ for all root partitions $\pi$. Then $M$ has a $\Delta$-flag. Furthermore $[M:\Delta(\pi)]=\dim\operatorname{Hom}(M,\nabla(\pi))$.
\end{theorem}
\begin{proof}
This is a standard argument, for example see \cite[Theorem 3.13]{bkm}.
\end{proof}
As a consequence we obtain the following BGG reciprocity for KLR algebras.
\begin{theorem}\label{15.4}
Let $\pi$ be a root partition and let $\P(\pi)$ be the projective cover of $L(\pi)$. Then $\P(\pi)$ has a $\Delta$-flag. For any root partition $\sigma$ the multiplicity $[\P(\pi):\Delta(\sigma)]$ is equal to the multiplicity $[\overline{\Delta}(\sigma):L(\pi)]$.
\end{theorem}
\begin{proof}
Since $\P(\pi)$ is finitely generated and projective it satisfies the hypotheses of Theorem \ref{deltaflag} and hence has a $\Delta$-flag. Furthermore the multiplicity of the module $\Delta(\sigma)$ in the flag is
\[
[\P(\pi):\Delta(\sigma)]=\dim\operatorname{Hom}(P(\pi),\co(\sigma)).
\]
As $\P(\pi)$ is the projective cover of $L(\pi)$, the dimension of this homomorphism space is equal to the multiplicity $[\nabla(\sigma)^\circledast:L(\pi)]$. By duality
$[\nabla(\sigma):L(\pi)]=[\overline{\Delta}(\sigma):L(\pi)^\circledast]$ and since $L(\pi)\cong L(\pi)^\circledast$ we are done.
\end{proof}
\begin{theorem}
The PBW basis (\ref{pbwalg}) is a basis of $\mathbf{f}$ as a $\mathbb{Z}[q,q^{-1}]$-module.
\end{theorem}
\begin{proof}
By Theorem \ref{15.4} and Theorem \ref{main}(3), the matrix expressing the set $\{[\Delta(\pi)]\}$ in terms of the basis $\{[\P(\pi)]\}$ is upper-triangular, with ones along the diagonal.
Therefore the set $\{[\Delta(\pi)]\}$ is a basis of $\mathbf{f}$ as a $\mathbb{Z}[q,q^{-1}]$-module.
\end{proof}
\begin{remark}
This is a generalisation, with a different proof, of a result of \cite{becknakajima}.
\end{remark}
\begin{proposition}
With respect to the PBW basis, the bar involution is unitriangular.
\end{proposition}
\begin{proof}
By Proposition \ref{homological}, the PBW basis is dual to the basis $[\nabla(\pi)]$ under the pairing $\langle,\cdot,\cdot\rangle$. It suffices to prove that the bar involution on $\mathbf{f}^*$ is unitriangular with respect to this basis. Since each $\nabla(\pi)$ is an induction product of self-dual simples up to an overall grading shift, it is easy to see that the bar involution is unitriangular by Theorem \ref{main}(3).
\end{proof}
Once we have that the bar-involution is unitriangular, it is straightforward to show that there exists a unique basis $b_\p$ of $\mathbf{f}$ which is bar-invariant and for which
\[
b_\p=E_\p + \sum_{\sigma<\pi} c_{\p\sigma} E_\sigma
\]
where $c_{\p\sigma}\in q\mathbb{Z}[q]$. Theorem \ref{positif} below shows that the basis $\{b_\p\}$ is the canonical basis, providing an algebraic characterisation of the canonical basis.
Thus from Theorem \ref{15.4} and the fact that the indecomposable projective modules categorify the canonical basis, we obtain the following positivity result.
\begin{theorem}\label{positif}
The change of basis matrix from the canonical basis to a PBW basis is unitriangular with off diagonal entries lying in $q\mathbb{N}[q]$.
\end{theorem}
\begin{proof}
The fact that the coefficients are all nonnegative is from Theorem \ref{15.4} and the fact that the indecomposable projective modules categorify the canonical basis. That the coefficients lie in $q\mathbb{Z}[q]$ follows from Lemma \ref{geometry}.
\end{proof}
This positivity result is new in affine type. In finite type this result is \cite[Corollary 10.7]{lusztigoneofthem} for particular convex orders and for all convex orders is due to Kato and the author independently in \cite{kato,klr1}.
\def$'${$'$}
|
\section{Introduction}
\label{Introduction}
Nonnegative data are pervasive. Consider the following four important applications, each of which give rise to nonnegative data matrices.
\begin{itemize}
\item In document collections, documents are stored as vectors. Each element of a document vector is a count (possibly weighted) of the number of times a corresponding term appears in that document. Stacking document vectors one after the other creates a nonnegative term-by-document matrix that represents the entire document collection numerically.
\item Similarly, in image collections, each image is represented by a vector, and each element of the vector corresponds to a pixel. The intensity and color of the pixel is given by a nonnegative number, thereby creating a nonnegative pixel-by-image matrix.
\item For item sets or recommendation systems, the information for a purchase history of customers or ratings on a subset of items is stored in a non-negative sparse matrix.
\item In gene expression analysis, gene-by-experiment matrices are formed from observing the gene sequences produced under various experimental conditions.
\end{itemize}
These are but four of the many interesting applications that create nonnegative data matrices (and tensors)
\cite{Langville2007:nmfsurvey}.
Three common goals in mining information from such matrices are: (1) to
automatically cluster similar items into groups, (2) to retrieve
items most similar to a user's query, and (3) identify interpretable critical dimensions within the collection. For the past decade, a technique called Latent
Semantic Indexing (LSI) \cite{BerryUSEbook}, originally conceived for the information retrieval problem and later extended to more general text mining problems, was a popular
means of achieving these goals. LSI uses a well-known factorization
of the term-by-document matrix, thereby creating a low rank approximation of the original matrix. This factorization, the singular value decomposition (SVD)
\cite{GolubVanLoanbook, MeyerSIAMbook}, is a classic technique in numerical
linear algebra.
The SVD is easy to compute and works well for points (1) and (2) above, but not (3).
The SVD does not provide users with any interpretation of its
mathematical factors or why it works so well. A common complaint from users is:
\emph{do the SVD factors reveal anything about the data collection?}
Unfortunately, for the SVD, the answer to this question is no, as explained in the next section. However, an
alternative and much newer matrix factorization, known as \emph{the
nonnegative matrix factorization (NMF), allows the question to be answered affirmatively.} As a result, it can be shown that the NMF works nearly as well as the SVD on points (1) and (2), and further, can also achieve goal (3).
Most examples and applications of the NMF in this paper refer to text mining because this is the area with which we are most familiar. However, the phrase ``term-by-document matrix" which we will use frequently throughout this paper can just as easily be replaced with gene-by-observation matrix, purchase-by-user matrix, etc., depending on the application area.
\section{Low Rank Approximations}
Applications, such as text processing, data mining, and image
processing, store pertinent information in a huge matrix.
This matrix $\b A$ is large, sparse, and often times nonnegative. In
the last few decades, researchers realized that the data matrix
could be replaced with a related matrix, of much lower rank. The
low rank approximation to the data matrix $\b A$ brought several
advantages. The rank-$k$ approximation, denoted $\b A_k$, sometimes
required less storage than $\b A$. But most importantly, the low
rank matrix seemed to give a much cleaner, more efficient
representation of the relationship between data elements. The low
rank approximation identified the most essential components of the
data by ignoring inessential components attributed to noise,
pollution, or inconsistencies. Several low rank
approximations are available for a given matrix: QR, URV, SVD, SDD, PCA, ICA, NMF, CUR, etc. \cite{OLeary1998, MeyerSIAMbook, Smildebook, Drineas2006}. In this section, we
focus on two such approximations, the SVD and the NMF, that have been applied to data mining
problems.
\subsection{The Singular Value Decomposition}
In 1991, Susan Dumais \cite{Dumais1991} used the
singular value decomposition (SVD) to build a low rank
approximation to the term-by-document matrix of information retrieval. In
fact, to build a rank-$k$ approximation
$\b A_k$ to the rank $r$ term-by-document matrix $\b A$, simply use the
$k$ most significant singular components, where $k < r$. That is,
\begin{eqnarray*}
\b A_k = \sum_{i=1}^k \sigma_i \b u_i \b v_i^T = \b
U_k \b \Sigma_k \b V_k^T,
\end{eqnarray*}
where $\sigma_i$ is the $i^{th}$ singular value of $\b A$, and $\b
u_i$ and $\b v_i^T$ are the corresponding singular vectors
\cite{GolubVanLoanbook}. The technique of replacing $\b A$ with the truncated
$\b A_k$ is called Latent Semantic Indexing (LSI) because the low rank
approximation reveals meanings and connections between documents that
were hidden, or latent, in the original noisy data matrix $\b A$.
Mathematically, the
truncated SVD has one particularly appealing property: of all
possible rank-$k$ approximations, $\b A_k$ is the best approximation
in the sense that $\| \b A- \b A_k \|_F$ is as small as possible
\cite{BerryUSEbook, BerryJessup1999}. Thus, the truncated SVD provides a nice baseline against which all other low-rank approximations can be judged for quantitative accuracy.
This optimality property is
also nice in practice. Algorithms for computing the $k$ most
significant singular components are fast, accurate, well-defined, and robust
\cite{Templatesbook, BerryUSEbook,
GolubVanLoanbook}. Two different algorithms will produce the same results up to
roundoff error. Such uniqueness and computational robustness are
comforting. Another advantage of the truncated SVD concerns building successive
low rank approximations. Once $\b A_{100}$ has been computed,
no further computation is required if, for example, for sensitivity
analysis or comparison purposes, other \emph{lower} rank
approximations are needed. That is, once
$\b A_{100}$ is available, then $\b A_k$ is available for any $k \le
100$.
LSI and the truncated SVD dominated text mining research in the 1990s
\cite{Baeza-Yatesbook, BerryCIRbook, BerryUSEbook,
BerryFierro1996, BerryJessup1999, BerryOBrien1998,
Blomthesis, BlomRuhe2001, Kumar2005, Dumais1991, HugheyBerry2000,
Littman2000, JiangBerry2000, LetscheBerry1997, WittenBerry1998, Zha1998:kvalue,
Berry2001:levelsets}. However, LSI is not perfect.
For instance, while it first appeared that the low rank approximation
$\b A_k$ would save storage over the original matrix $\b A$,
experiments showed that this was not the case. $\b A$ is generally
very sparse for text mining problems because only a small subset of
the terms in the collection are used in any particular document.
No matter how sparse the original term-by-document
matrix is, the truncated SVD produces singular components that are almost
always completely dense. In many cases, $\b A_k$ can require more
(sometimes much more) storage than $\b A$.
Furthermore, $\b A$ is
always a nonnegative matrix, yet the singular components are mixed in
sign.
The SVD's loss of the nonnegative structure of the
term-by-document matrix means that the factors of the truncated SVD provide no
interpretability.
To understand this statement, consider
a particular document vector, say, column 1 of $\b A$. The truncated
SVD represents document 1, $\b A_1$, as
\begin{eqnarray*}
\b A_1 \approx \sigma_1 v_{11} \pmatrix{ \vdots \cr \b u_1 \cr \vdots} +
\sigma_2 v_{12} \pmatrix{\vdots \cr \b u_2 \cr \vdots} +
\cdots +
\sigma_k v_{1k} \pmatrix{\vdots \cr \b u_k \cr \vdots },
\end{eqnarray*}
which reveals that document 1 is a linear combination of the singular
vectors
$\b u_i$, also called the basis vectors. The scalar weight
$\sigma_i v_{1i}$ represents the contribution of basis
vector
$i$ in document 1. Unfortunately, the mixed signs in $\b u_i$ and
$\b v_i$ preclude interpretation.
Clearly, the interpretability issues with the SVD's
basis vectors are caused by the mixed signs in the singular vectors. Thus,
researchers proposed an alternative low rank approximation that maintained the
nonnegative structure of the original term-by-document matrix.
As a result, the nonnegative matrix factorization (NMF) was created
\cite{LeeSeung1999, PaateroTapper1994}. The NMF replaces the role played by
the singular value decomposition (SVD). Rather than factoring $\b A$ as $\b U_k
\b \Sigma_k \b V_k^T$, the NMF factors $\b A$ as $\b W_k \b H_k$, where $\b W_k$
and $\b H_k$ are nonnegative.
\subsection{The Nonnegative Matrix Factorization}
Recently, the nonnegative matrix factorization (NMF) has been used to create a low rank approximation to $\b A$ that contains nonnegative factors called $\b W$ and $\b H$.
The NMF of a data matrix $\b A$ is created by solving the
following nonlinear optimization problem.
\vskip -15pt
\begin{eqnarray}
\min \| \b A_{m \times n} &-& \b W_{m \times k} \b H_{k \times n} \|^2_F, \\
s.t. && \b W \ge \b 0,\nonumber \\ && \b H \ge \b 0 \nonumber.
\label{NMFopt}
\end{eqnarray}
The Frobenius norm is often used to measure the error between the original matrix $\b A$ and its low rank approximation $\b W \b H$, but there are other possibilities \cite{Dhillon2005, LeeSeung1999, Paatero1997}.
The rank of the approximation, $k$, is a parameter that must be set by the user.
The NMF is used in place of other low rank factorizations, such as the singular value decomposition (SVD) \cite{MeyerSIAMbook}, because of its two primary advantages: storage and interpretability. Due to the nonnegativity constraints, the NMF produces a so-called ``additive parts-based" representation \cite{LeeSeung1999} of the data. One consequence of this is that the factors $\b W$ and $\b H$ are generally naturally sparse, thereby saving a great deal of storage when compared with the SVD's dense factors.
The NMF also has impressive benefits in terms of interpretation of its factors, which is, again, a consequence of the nonnegativity constraints. For example, consider a text processing application that requires the factorization of a term-by-document matrix $\b A_{m \times n}$.
In this case, $k$ can be considered the number of (hidden) topics
present in the document collection. In this case, $\b W_{m
\times k}$ becomes a term-by-topic matrix whose columns are the NMF
basis vectors. The nonzero elements of column 1 of $\b W$ (denoted $\b W_1$),
which is sparse and nonnegative,
correspond to particular terms. By considering the highest weighted terms in
this vector, one can assign a label or topic to basis vector 1. Figure
\ref{figure:interpretdim} shows four basis vectors for one
particular term-by-document matrix, the \texttt{medlars} dataset of medical abstracts, available at \verb"http://www.cs.utk.edu/~lsi/".
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=.5]{interpretdim.eps}
\caption{Interpretation of NMF basis vectors on \texttt{medlars} dataset}
\label{figure:interpretdim}
\end{center}
\end{figure}
For those familiar with the domain of this dataset, the NMF allows users the ability to
interpret the basis vectors. For instance, a user might attach
the label ``heart" to basis vector $\b W_1$ of Figure
\ref{figure:interpretdim}.
Similar interpretation holds for the
other factor $\b H$. $\b H_{k \times n}$ becomes a topic-by-document
matrix with sparse nonnegative columns. Element $j$
of column 1 of $\b H$ measures the strength to which topic $j$ appears
in document 1.
Another fascinating application of the NMF is image processing. Figure \ref{figure:LSnmffigures1} clearly demonstrates two advantages of the NMF over the SVD. First, notice that the NMF basis vectors, represented as individual blocks in the $\b W$ matrix, are very sparse (i.e., there is much white space). Similarly, the weights, represented as individual blocks in the $\b H_i$ vector, are also sparse. On the other hand, the SVD factors are nearly completely dense. Second, the basis vectors of the NMF, in the $\b W$ matrix, have a nice interpretation, as individual components of the structure of the face---ears, noses, mouths, hairlines. The SVD basis vectors do not create an additive parts-based representation.
In addition, the gains in storage and interpretability do not come at a loss in performance. The NMF and the SVD perform equally well in reconstructing an approximation to the original image.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=1]{LSnmffigures1.eps}
\caption{Interpretation of NMF and SVD basis vectors on face dataset, from \cite{LeeSeung1999}}
\label{figure:LSnmffigures1}
\end{center}
\end{figure}
Of course, the NMF has its disadvantages too. Other popular factorizations, especially the SVD, have strengths concerning uniqueness and robust
computation. Yet these become problems for the NMF. There is no unique global
minimum for the NMF. The optimization problem of Equation
\nref{NMFopt} is convex in either $\b W$ or $\b H$, but not in
both $\b W$ and
$\b H$, which means that the algorithms can only, if at all,
guarantee convergence to a local minimum. In practice, NMF users
often compare the local minima from several different starting
points, using the results of the best local minimum found. However,
this is prohibitive on large, realistically-sized problems. Not
only will different NMF algorithms (and there are many now \cite{Langville2007:nmfsurvey})
produce different NMF factors, the
same NMF algorithm, run with slightly different parameters, can
produce different NMF factors.
\subsection{Summary of SVD vs. NMF}
We pause to briefly summarize the advantages of these two competing low rank approximations.
The properties and advantages of the SVD include: (1) an optimality property; the truncated SVD produces the best rank-$k$ approximation (in terms of squared distances), (2) speedy and robust computation, (3) unique factorization; initialization does not affect SVD algorithms, and ( 4) orthogonality; resulting basis vectors are orthogonal and allow conceptualization of original data as vectors in space. On the other hand, the advantages of NMF are: (1) sparsity and nonnegativity; the factorization maintains these properties of the original matrix, (2) reduction in storage; the factors are sparse, which also results in easier application to new data, and (3) interpretability; the basis vectors naturally correspond to conceptual properties of the data.
The strengths of one approximation become the weaknesses of another. The most severe weakness of the NMF are its convergence issues. Unlike the SVD and its unique factorization, there is no unique NMF factorization. Because different NMF algorithms can converge to different local minima (and even this convergence to local minima is not guaranteed), initialization of the algorithm becomes critical.
In practice, knowledge of the application area can help guide initialization choices. We will return to such initialization issues in Section \ref{Initializations}.
\section{ALS Algorithms for the NMF}
\label{ALSAlg}
Several currently popular NMF algorithms
\cite{ChuPlemmons2004, LeeSeung1999, LeeSeung2000, PaateroTapper1994,
Paatero1997, XuLiu2003} do not create sparse factors, which are desired for
storage, accuracy, and interpretability reasons. Even with adjustments to create sparse factors,
the improved algorithms
\cite{Hoyer2002, LeeSeung2000, PiperPlemmons2005, BerryPlemmons2004} exhibit an
undesirable
\emph{locking} phenomenon, as explained below.
Thus, in this section, we propose two new NMF algorithms \cite{SASSlides4}, called ACLS and AHCLS, that
produce sparse factors and avoid the so-called locking problem.
Both algorithms are modifications to the
simple Alternating Least Squares (ALS) algorithm
\cite{PaateroTapper1994}, wherein $\b W$ is fixed and
$\b H$ is computed using least squares, then $\b H$ is fixed and
$\b W$ is computed using least squares, and so on, in alternating
fashion. The method of alternating variables is a well-known
technique in optimization
\cite{Nocedal1999}.
One problem with the first ALS algorithm applied to the NMF
problem (done by Paatero and Tapper in 1994 \cite{PaateroTapper1994}) was the lack of sparsity
restrictions. To address this, the ACLS algorithm adds a reward for sparse
factors of the NMF. The user sets the two parameters $\lambda_H$ and
$\lambda_W$ to positive values. Increasing these values increases
the sparsity of the two NMF factors. However, because there are no
upperbounds on these parameters, a user must resort to trial and error to find
the best values for $\lambda_H$ and
$\lambda_W$. The more advanced AHCLS \cite{SASSlides4}, presented in Section \ref{AHCLS}, provides better
sparsity parameters with more intuitive bounds.
\subsection{The ACLS Algorithm}
\label{ACLS}
The ACLS (Alternating Constrained Least Squares) algorithm is implemented
differently than the original ALS algorithm \cite{PaateroTapper1994} because
issues arise at each alternating step, where a constrained least squares problem
of the following form
\begin{eqnarray}
\min_{\b h_{j}}
\|
\b a_{j} -
\b W \b h_{j}
\|_2^2 + \lambda_H \| \b h_{j} \|_2^2 \quad s.t. \quad \lambda_H \geq 0, \b
h_{j}
\geq
\b 0
\label{vectorACLS}
\end{eqnarray}
must be
solved. The vectors $\b a_j$ and $\b h_j$ are columns of $\b A$ and $\b H$, respectively. Notice that the decision variable $\b h_j$ must be nonnegative. There are
algorithms specifically designed for this nonnegative constrained least squares
problem. In fact, the NNLS algorithm of Lawson and Hanson
\cite{Bjorck1996, LawsonHansonLSbook} is so common that it appears as a
built-in function in MATLAB. Unfortunately, the NNLS algorithm is
very slow, as it is an ``active set" method, meaning it can swap only one variable from the basis at a
time. Even the faster version of the NNLS algorithm by Bro and de Jong \cite{Bro1997} is still not fast enough, and the NNLS step remains the computational bottleneck.
As a result, in practice, compromises are made.
For example, a standard (unconstrained) least squares step is run \cite{Bjorck1996} and all
negative elements in the solution vector are set to 0. This ad-hoc
enforcement of nonnegativity, while not theoretically appealing,
works quite well in practice. The practical ACLS algorithm is shown
below.
\bigskip
\noindent
\framebox{
\parbox[t]{6.3in}{
\centerline {\sc Practical ACLS Algorithm for
NMF}
\smallskip
\noindent
input $\lambda_W$, $\lambda_H$ \\
$\b W$ = rand(m,k); \quad {\small \% initialize $\b W$ as
random dense matrix or use another initialization from Section
\ref{Initializations}}\\
for i = 1 : maxiter\\
\hspace*{.3in} {\sc (cls) } \quad \quad \ \ Solve for $\b H$ in matrix equation
$(\b W^T
\b W + \lambda_H \b I)\ \b H = \b W^T
\b A$. \quad \ {\small \% for $\b W$ fixed, find $\b
H$}\\
\hspace*{.3in} {\sc (nonneg) } \, Set all negative elements in $\b H$ to 0.
\\
\hspace*{.3in} {\sc (cls) } \quad \quad \ \ Solve for $\b W$ in matrix equation
$(\b H
\b H^T + \lambda_W \b I)\ \b W^T = \b H\b A^T
$. \ \ {\small \% for $\b H$ fixed, find $\b W$}\\
\hspace*{.3in} {\sc (nonneg) } \, Set all negative elements in $\b W$ to 0.
\\ end
}
}
\bigskip
\subsection{The AHCLS Algorithm}
\label{AHCLS}
ACLS uses a crude measure $\| \b x \|_2^2$ to approximate the sparsity of a vector $\b x$.
The AHCLS replaces this with a more sophisticated measure, $spar(\b x)$, which was invented by Hoyer \cite{Hoyer2004}.
\begin{eqnarray*}
spar(\b x_{n \times 1})=\;
\matrix{\sqrt{n} -
\|\b x\|_1 /
\|
\b x\|_2\cr
\noalign{\vskip 5pt\hrule\vskip 5pt}
\noalign{\vskip -2pt} \sqrt{n}-1}
\end{eqnarray*}
In AHCLS (Alternating Hoyer-Constrained Least Squares), the user defines two scalars $\alpha_W$ and $\alpha_H$ in addition to $\lambda_H$ and $\lambda_W$ of ACLS. For AHCLS, the two additional scalars $0 \leq \alpha_W, \alpha_H \leq 1$ represent a user's desired sparsity in each column of the factors. These scalars, because they range from 0 to 1, match nicely with a user's notion of sparsity as a percentage. Recall that $0 \leq \lambda_W, \lambda_H \leq \infty$ are positive weights associated with the penalties assigned to the density of $\b W$ and $\b H$. Thus, in AHCLS, they measure
how important it is to the user that $spar(\b W_{j*})=\alpha_W$ and
$spar(\b H_{j*})=\alpha_H$. Our experiments show that AHCLS does a better job of enforcing sparsity than ACLS does. And the four AHCLS parameters are easier to set. For example, as a guideline, we recommend $0 \le
\lambda_W,\lambda_H
\le 1$, with of course, $0 \leq \alpha_W, \alpha_H \leq 1$.
The practical AHCLS algorithm, using matrix systems and ad-hoc enforcement of negativity, is below. $\b E$ is the matrix of all ones.
\bigskip
\noindent
\framebox{
\parbox[t]{6.3in}{
\centerline {\sc Practical AHCLS Algorithm for
NMF}
\smallskip
\noindent
input $\lambda_W$, $\lambda_H$, $\alpha_W$, $\alpha_H$
$\b W$ = rand(m,k); \quad {\small \% initialize $\b W$ as
random dense matrix or use another initialization from Section
\ref{Initializations}}\\
$\beta_H = ((1-\alpha_H)\sqrt{k}+\alpha_H)^2$ \\
$\beta_W = ((1-\alpha_W)\sqrt{k}+\alpha_W)^2$ \\
for i = 1 : maxiter\\
\hspace*{.3in} {\sc (hcls) } \quad \ \ \ Solve for $\b H$ in matrix equation
$(\b W^T
\b W + \lambda_H \beta_H \b I - \lambda_H \b E)\ \b H = \b W^T
\b A$. \\
\hspace*{.3in} {\sc (nonneg) } \, Set all negative elements in $\b H$ to 0.
\\
\hspace*{.3in} {\sc (hcls) } \quad \ \ \ Solve for $\b W$ in matrix equation
$(\b H
\b H^T + \lambda_W \beta_W \b I-\lambda_W \b E)\ \b W^T = \b H\b A^T
$. \\
\hspace*{.3in} {\sc (nonneg) } \, Set all negative elements in $\b W$ to 0.
\\ end
}
}
\bigskip
\subsection{Advantages and Disadvantages of ACLS and AHCLS}
\label{AdDisadALS}
\subsubsection{Speed}
These algorithms have many advantages. For
instance, rather than computing the vectors in $\b H$ column by column (as is done in \cite{BerryPlemmons2004}), thereby solving sequential least squares problems of the form of Equation (\ref{vectorACLS}), one matrix system solve can be executed. Further,
because each CLS step solves a \emph{small} $k \times k$ matrix system, ACLS
and AHCLS are the fastest NMF algorithms available (and faster than current
truncated SVD algorithms). See Section \ref{ExpALS} for comparative run times. They converge quickly and give
very accurate NMF factors.
\subsubsection{Sparsity}
Only $\b W$ must be initialized, and sparsity is
incorporated for both NMF factors. We believe that avoidance of the
so-called
\emph{locking} phenomenon is one reason why the class of ALS algorithms works
well in practice. Nearly all other NMF algorithms, especially those of the
multiplicative update class
\cite{Hoyer2002, Hoyer2004, LeeSeung1999, LeeSeung2000, BerryPlemmons2004:siam,
Shahnaz2004, BerryPlemmons2004}, lock elements when they become 0. That is, during the iterative
process, once an element in either $\b W$ or
$\b H$ becomes 0, it must remain 0. For the basis vectors in the text mining problem, which are
stored in $\b W$, this means that in order to improve the objective
function, the algorithm can only remove terms from, not add terms to,
topic basis vectors. As a result, once the algorithm starts
down a path toward a particular topic vector, it must continue in that
direction. On the other hand, ALS algorithms do not {\it lock} elements, and
thus provide greater flexibility, allowing them to escape from a path
heading towards a poor local minimum.
\subsubsection{Convergence}
It has been proven that ALS algorithms will converge to a fixed point, but this fixed point may be a local extrema or a saddle point \cite{{Finesso2004, Gonzalez2005, Lin2005}}. The ACLS and AHCLS algorithms with properly enforced nonnegativity, for example, by the NNLS algorithm, are known to converge to a local minimum \cite{Dhillon2005, Lin2005}. However, our ad-hoc enforcement of nonnegativity, which drastically speeds up the algorithm (and improves sparsity), means there are no proofs claiming convergence to a local minimum; saddle points are now possible. (Actually, this is not so damning for our two ALS algorithms because most NMF algorithms suffer this same problem. The few NMF algorithms believed to guarantee convergence to a local minimum have been proven otherwise \cite{{Finesso2004, Gonzalez2005, Lin2005}}.)
Our experiments
\cite{SASSlides4, SASSlides3} and others \cite{Paatero1996, Paatero1999, PaateroTapper1994,
Paatero1997, Plemmons2005} have shown that the ALS fixed points can be superior to the results of other NMF algorithms.
\subsubsection{Nonnegativity}
Clearly, ad-hoc enforcement of nonnegativity is theoretically unattractive.
There are some alternatives to this ad-hoc enforcement of nonnegativity. For instance, one could convert from an alternating least squares approach to an alternating linear programming approach, whereby nonnegativity of variables is enforced in a natural way by the simple constraints of the linear programming
formulation. Yet, this has the same problem as the NNLS algorithm, lengthy execution time.
A second alternative to ad-hoc enforcement of nonnegativity is to add negativity penalties in the form of logarithmic functions to the NMF objective function \cite{LuWu2004}. This is a focus of future work.
\subsection{Numerical Experiments}
\label{ExpALS}
Figure \ref{figure:NMFalgs} compares our ACLS and AHCLS algorithms with the popular Lee-Seung mean squared error algorithm \cite{LeeSeung1999} and the GDCLS algorithm \cite{BerryPlemmons2004}. We use our own implementation of GDCLS, which is much faster than the implementation presented in \cite{BerryPlemmons2004}. The speed improvement results from our use of one matrix system rather than serial vector systems to solve the CLS step. This implementation trick was described above for the ACLS and AHCLS algorithms.
To create Figure \ref{figure:NMFalgs}, we used the \texttt{medlars} dataset of medical abstracts and the \texttt{cisi} dataset of library science abstracts.
These figures clearly show how the ALS-type algorithms outperform the Lee-Seung multiplicative update algorithms in terms of accuracy and speed. While the ALS-type algorithms provide similar accuracy to the GDCLS algorithm, they are much faster. This speed advantage continues to hold for much larger collections like the \texttt{reuters10} collection, used in Section \ref{Initializations}. (Details on the \texttt{reuters10} dataset appear in Section \ref{reuters10}.) On average the ACLS and AHCLS algorithms require roughly 2/3 the time of the GDCLS algorithm. Figure \ref{figure:NMFalgs} also reports the error in the optimal rank-10 approximation required by the SVD. Notice how close all NMF algorithms come to the optimal factorization error. Also, notice that ACLS and AHCLS require less time than the SVD to produce such good, sparse, nonnegative factorizations.
\begin{figure}[htbp]
\hspace*{-.2in}
\includegraphics[scale=1]{NMFalgsMED.eps}
\hfill
\includegraphics[scale=1]{NMFalgsCISI.eps}
\caption{Accuracy and Run-times of NMF Algorithms on \texttt{medlars} (left) and \texttt{cisi} (right) datasets}
\label{figure:NMFalgs}
\end{figure}
\section{Initializations}
\label{Initializations}
All NMF algorithms are iterative and it is well-known that they are sensitive to the initialization of $\b W$ and $\b H$ \cite{Wild2003:thesis}.
Some algorithms require that both $\b W$ and $\b H$ be
initialized
\cite{Hoyer2002, Hoyer2004, LeeSeung1999,
LeeSeung2000, PiperPlemmons2005}, while others require
initialization of only $\b W$
\cite{PaateroTapper1994, Paatero1997, Shahnaz2004, BerryPlemmons2004}.
\emph{In all
cases, a good initialization can improve the speed and accuracy of the
algorithms, as it can produce faster convergence to an improved local minimum} \cite{Smildebook}. A good initialization
can sidestep some of the convergence problems mentioned above, which is precisely why they are so important.
In this section, we compare several initialization procedures (two old and four new) by testing them on the ALS algorithms presented in Section \ref{ALSAlg}. We choose to use the ACLS and AHCLS algorithms because they produce sparse accurate factors and require about the same time as the SVD. Most other NMF algorithms require much more time than the SVD, often times orders of magnitude more time.
\subsection{Two Existing Initializations}
Nearly all NMF algorithms use
simple \emph{random initialization}, i.e., $\b W$ and $\b H$ are initialized as
dense matrices of random numbers between 0 and 1. It is well-known that random initialization does not generally provide a good first estimate for NMF algorithms \cite{Smildebook}, especially those of the ALS-type of \cite{Burdick1990, Li1993, Sanchez1990, Sands1980}. Wild et
al.
\cite{Wild2003:thesis, Wild2003, Wild2004} have shown that the \emph{centroid initialization}, built from the centroid decomposition \cite{Dhillon2001} is a much better
alternative to random initialization. Unfortunately, this decomposition is expensive as a preprocessing
step for the NMF. Another advantage of ALS algorithms, such as our ACLS and AHCLS, is that they only require initialization of $\b W$. In ALS algorithms, once $\b W^{(0)}$ is known, $\b H^{(0)}$ is computed quickly by a least squares computation. As a result, we only discuss techniques for computing a good $\b W^{(0)}$.
\subsection{Four New Initializations}
Some text mining software produces the SVD factors for other text tasks. Thus, in the event that the SVD factor $\b V$ is available, we propose a
\emph{SVD-centroid initialization} \cite{SASSlides3}, which initializes $\b
W$ with a centroid decomposition of the low dimensional SVD factor
$\b V_{n
\times k}$ \cite{Skillicorn2003}. While the centroid decomposition of $\b A_{m \times n}$ can be too time-consuming, the centroid decomposition of $\b V$ is fast because $\b V_{n \times k}$ is much smaller than $\b A_{m \times n}$.
When the SVD factors are not available, we propose
a very inexpensive procedure called \emph{random Acol
initialization}. Random Acol forms an initialization of each column
of the basis matrix $\b W$ by averaging
$p$ random columns of $\b A$. It makes more sense to
build basis vectors from the given data, the sparse document vectors
themselves, than to form completely dense random basis vectors, as random
initialization does. Random Acol initialization is very inexpensive, and lies
between random initialization and centroid initialization in terms of performance
\cite{SASSlides4, SASSlides3}.
We also present two more initialization ideas, one inspired by the $\b C$ matrix of the CUR decomposition \cite{Drineas2006}, and another by the term co-occurrence matrix \cite{Sandler2005}. We refer to these last two methods as \emph{random $\b C$ initialization} and \emph{co-occurrence initialization}, respectively. The random $\b C$ initialization is similar to the random Acol method, except it chooses $p$ columns at random from the longest (in the 2-norm) columns of $\b A$, which generally means the densest columns since our text matrices are so sparse. The idea is that these might be more likely to be the centroid centers. The co-occurrence method first forms a term co-occurrence matrix $\b C = \b A \b A^T$. Next, the method for forming the columns of $\b W^{(0)}$ described as Algorithm 2 of \cite{Sandler2005} is applied to $\b C$. The co-occurrence method is very expensive for two reasons. First, for text mining datasets, such as \texttt{reuters10}, $m>>n$, which means $\b C=\b A \b A^T$ is very large and often very dense too. Second, the algorithm of \cite{Sandler2005} for finding $\b W^{(0)}$ is extremely expensive, making this method impractical.
All six initialization methods are summarized in Table \ref{table:InitMethods}. The two existing methods appear first, followed by our four suggested methods.
\begin{table*}
\centering
\caption{Initialization Methods for the NMF}
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
Name & Proposed by & Pros & Cons \cr
\hline
Random & Lee, Seung \cite{LeeSeung2000} & easy, cheap to compute & dense matrices, no intuitive basis \cr
Centroid & Wild et al. \cite{Wild2003:thesis} & reduces \# NMF iterations, & expensive, must run clustering \cr
& & $\;\;\;\;$ firm, intuitive foundation & $\;\;\;\;$ algorithm on cols of $\b A$ \cr
\hline
SVD-Centroid & Langville \cite{SASSlides3} & inexpensive, reduces \# NMF& SVD factor $\b V$ must be available \cr
& & $\;\;\;\;$ iterations & \cr
Random Acol & Langville \cite{SASSlides4} & cheap, sparse matrices built & only slight decrease in number of \cr
& & $\;\;\;\;$ from original data & $\;\;\;\;$ NMF iterations \cr
Random $\b C$ & Langville adapts & cheap, sparse & not very effective \cr
& $\;\;\;\;$ from Drineas \cite{Drineas2006} & & \cr
Co-occurrence & Langville adapts & uses term-term similarities & large, dense co-occurrence matrix,\cr
& $\;\;\;\;$ from Sandler \cite{Sandler2005} & $\;\;\;\;$ & $\;\;\;\;$ very expensive computation \cr
\hline
\end{tabular}
\end{center}
\label{table:InitMethods}
\end{table*}%
\subsection{Initialization Experiments with Reuters10 dataset} \label{reuters10}
The \texttt{reuters10} collection is our subset of the Reuters-21578 version of the Reuter's benchmark document collection of business newswire posts. The Reuters-21578 version contains over 20,000 documents categorized into 118 different categories, and is available online.\footnote{\texttt{\small http://www.daviddlewis.com/resources/testcollections/\\reuters21578/}} Our subset, the \texttt{reuters10} collection, is derived from the set of documents that have been classified into the top ten most frequently occurring categories. The collection contains 9248 documents from the training data of the ``ModApte split" (details of the split are also available at the website above).
The numbers reported in Table \ref{table:InitExps} were generated by applying the alternating constrained least squares (ACLS) algorithm of Section \ref{ALSAlg} with $\lambda_H=\lambda_W=.5$ to the \texttt{reuters10} dataset.
The error measure in this table is relative to the optimal rank-10 approximation given by the singular value decomposition. For this dataset, $\| \b A - \b U_{10} \b \Sigma_{10} \b V_{10}^T \|_F=22656$. Thus, for example, the error at iteration 10 is computed as
\begin{eqnarray*}
\mbox{Error(10)}= \frac{\| \b A - \b W^{(10)} \b H^{(10)} \|_F - 22656 }{22656}.
\end{eqnarray*}
\begin{table*}
\caption{Experiments with Initialization Methods for the NMF}
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|}
\hline
Method & Time $\b W^{(0)}$ & Storage $\b W^{(0)}$ & Error(0) & Error(10) & Error(20) & Error(30) \cr
\hline
Random & .09 sec & 726K & 4.28\% & .28\% & .15\% & .15\% \cr
Centroid& 27.72 & 46K & 2.02\% & .27\% & .18\% & .18\% \cr
\hline
SVD-Centroid & $.65^\dagger$ & 56K & 2.08\% & .06\% & .06\% & .06\% \cr
Random Acol$^*$ & .05 & 6K & 2.01\% & .21\% & .16\% & .15\% \cr
Random $\b C^{\circ}$ & .11 & 22K & 3.35\% & .29\% & .20\% & .19\% \cr
Co-occurrence & 3287 & 45K & 3.38\% & .37\% & .27\% & .25\% \cr
\hline
ACLS time & & & .37 sec & 3.42 & 6.78 & 10.29 \cr
\hline
\end{tabular}
\end{center}
\label{table:InitExps}
$\dagger$ {\footnotesize provided $\b V$ of the SVD is already available} \\
\indent $^*$ {\footnotesize each column of $\b W^{(0)}$ formed by averaging 20 random columns of $\b A$} \\
\indent $^{\circ}$ {\footnotesize each column of $\b W^{(0)}$ formed by averaging 20 of the longest columns of $\b A$}\\
\end{table*}%
We distinguish between quantitative accuracy, as reported in Table \ref{table:InitExps}, and qualitative accuracy as reported in Tables \ref{table:BestBasisVectors} through \ref{table:CooccurrenceBasisVectors}. For text mining applications, it is often not essential that the low rank approximation be terribly precise. Often suboptimal solutions are ``good enough." After reviewing Tables \ref{table:BestBasisVectors}--\ref{table:CooccurrenceBasisVectors}, it is easy to see why some initializations give better accuracy and converge more quickly. They start with basis vectors in $\b W^{(0)}$ that are much closer to the best basis vectors found, as reported in Table \ref{table:BestBasisVectors}, which was generated by using the basis vectors associated with the best global minimum for the \texttt{reuters10} dataset, found by using 500 random restarts. In fact, the relative error for this global minimum is .009\%, showing remarkable closeness to the optimal rank-10 approximation. By comparing each subsequent table with Table \ref{table:BestBasisVectors}, it's clear why one initialization method is better than another. The best method, SVD-centroid initialization, starts with basis vectors very close to the ``optimal" basis vectors of Table \ref{table:BestBasisVectors}. On the other hand, random and random Acol initialization are truly random. Nevertheless, random Acol does maintain one clear advantage over random initialization as it creates a very sparse $\b W^{(0)}$. The Random C and co-occurrence initializations suffer from lack of diversity. Many of the longest documents in the \texttt{reuters10} collection appear to be on similar topics, thus, not allowing $\b W^{(0)}$ to cover many of the reuters topics.
Because the algorithms did not produce the ``wheat" vector always in column one of $\b W$, we have reordered the resulting basis vectors in order to make comparisons easier. We also note that the nonnegative matrix factorization did produce basis vectors that cover 8 of the 10 ``correct" reuters classifications, which appear on the last line of Table \ref{table:BestBasisVectors}. The two missing reuters classifications are \texttt{corn} and \texttt{grain}, both of which are lumped into the first basis vector labeled \texttt{wheat}. This first basis vector does break into two separate vectors, one pertaining to \texttt{wheat and grain} and another to \texttt{corn} when the number of basis vectors is increased from $k=10$ to $k=12$. We note that these categories have been notoriously difficult to classify, as previously reported in \cite{Dumais1998}.
\begin{table*}
\caption{Basis vectors of $\b W^{(50)}$ from \emph{Best Global Minimum} found for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(50)}$ & $\b W_2^{(50)}$ & $\b W_3^{(50)}$ & $\b W_4^{(50)}$ & $\b W_5^{(50)}$ & $\b W_6^{(50)}$ & $\b W_7^{(50)}$ & $\b W_8^{(50)}$ & $\b W_9^{(50)}$ & $\b W_{10}^{(50)}$ \cr
\hline
tonne & billion & share & stg & mln-mln & gulf & dollar & oil & loss & trade \cr
wheat & year & offer & bank & cts & iran & rate & opec & profit & japan\cr
grain & earn & company & money & mln & attack & curr. & barrel & oper& japanese\cr
crop & qrtr & stock & bill & shr & iranian & bank & bpd & exclude & tariff\cr
corn & rise & sharehol. & market & net & ship & yen & crude & net & import\cr
agricul.& pct & common & england & avg & tanker & monetary & price & dlrs & reagan\cr
\hline
\texttt{wheat} & \texttt{earn} & \texttt{acquisition} & & \texttt{interest} & \texttt{ship} & \texttt{frgn-exch.} & \texttt{oil} & & \texttt{trade} \cr
\hline
\end{tabular}
\end{center}
\label{table:BestBasisVectors}
\end{table*}%
\begin{table*}
\caption{Basis vectors of $\b W^{(0)}$ created by \emph{Random Initialization} for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(0)}$ & $\b W_2^{(0)}$ & $\b W_3^{(0)}$ & $\b W_4^{(0)}$ & $\b W_5^{(0)}$ & $\b W_6^{(0)}$ & $\b W_7^{(0)}$ & $\b W_8^{(0)}$ & $\b W_9^{(0)}$ & $\b W_{10}^{(0)}$ \cr
\hline
announce & wpp & formality & bulletin & matthews & dramatic & squibb & wag & cochran & erik \cr
medtec & reflagging & simply & awfully & nyt & boca raton & kuwaiti & oils & mln & support \cr
pac & kwik & moonie & blair & barrel & clever & dacca & hears & barriers & sale oil \cr
purina & tilbury & tmg & fresno & purina & billion & democrat & bwtr & deluxe & direct \cr
mezzanine & capacitor & bushnell & farm & june & bkne & induce & nestle & mkc & wheat \cr
foreign & grain & country & leutwiler & trend & clever & rate & fed. & econ. & aid \cr
\hline
\end{tabular}
\end{center}
\label{table:RandomBasisVectors}
\end{table*}%
\begin{table*}
\caption{Basis vectors of $\b W^{(0)}$ created by \emph{Centroid Initialization} for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(0)}$ & $\b W_2^{(0)}$ & $\b W_3^{(0)}$ & $\b W_4^{(0)}$ & $\b W_5^{(0)}$ & $\b W_6^{(0)}$ & $\b W_7^{(0)}$ & $\b W_8^{(0)}$ & $\b W_9^{(0)}$ & $\b W_{10}^{(0)}$ \cr
\hline
tonne & bank & share & medar & cts & iran & rate & oil & stg & strike \cr
wheat & rate & company & mdxr & mmln & gulf & dollar & trade & bill & port \cr
grain & dollar & offer & mlx & loss & attack & bank & price & take-up & union \cr
corn & billion & pct & mlxx & net & iranian & currency & barrel & drain & seaman \cr
crop & pct & stock & mich & shr & missile & market & japan & mature & worker \cr
agriculture & trade & dlrs & troy & dlrs & tanker & monetary & opec & money & ship \cr
\hline
\end{tabular}
\end{center}
\label{table:CentroidBasisVectors}
\end{table*}%
\begin{table*}
\caption{Basis vectors of $\b W^{(0)}$ created by \emph{SVD-Centroid Initialization} for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(0)}$ & $\b W_2^{(0)}$ & $\b W_3^{(0)}$ & $\b W_4^{(0)}$ & $\b W_5^{(0)}$ & $\b W_6^{(0)}$ & $\b W_7^{(0)}$ & $\b W_8^{(0)}$ & $\b W_9^{(0)}$ & $\b W_{10}^{(0)}$ \cr
\hline
tonne & billion & share & bank & cts & iran & dollar & oil & loss & trade \cr
wheat & year & offer & money & shr & gulf & rate & barrel & oper & japan\cr
grain & earn & company & rate & mln & attack & curr. & opec & profit & japanese\cr
corn & qrtr & stock & stg & net & iranian & yen & crude & cts & tariff\cr
crop & rise & pct & market & mln-mln & missile & japan & bpd & mln & import\cr
agricul.& pct & common & pct & rev & ship & economic & price & net & country \cr
\hline
\end{tabular}
\end{center}
\label{table:SVDBasisVectors}
\end{table*}%
\begin{table*}
\caption{Basis vectors of $\b W^{(0)}$ created by \emph{Random Acol Initialization} for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(0)}$ & $\b W_2^{(0)}$ & $\b W_3^{(0)}$ & $\b W_4^{(0)}$ & $\b W_5^{(0)}$ & $\b W_6^{(0)}$ & $\b W_7^{(0)}$ & $\b W_8^{(0)}$ & $\b W_9^{(0)}$ & $\b W_{10}^{(0)}$ \cr
\hline
mln & fee & agl & mln & mark &loss & official & dlrs & bank & trade \cr
denman & mortg. & tmoc & dlrs & mannes. & mln & piedmont & oper & bancaire & viermetz \cr
dlrs & billion & bank &share & dividend & cts & dollar &billion & austral & mln \cr
ecuador & winley & pct & seipp & mln &maki & interest & loss & neworld & nwa \cr
venezuela & mln & company & billion & dieter &name & tokyo & texaco & datron & cts \cr
revenue & fed & maki & dome & gpu & kato & japanese & pennzoil & share & builder \cr
\hline
\end{tabular}
\end{center}
\label{table:RandomAcolBasisVectors}
\end{table*}%
\begin{table*}
\caption{Basis vectors of $\b W^{(0)}$ created by \emph{Random $\b C$ Initialization} for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(0)}$ & $\b W_2^{(0)}$ & $\b W_3^{(0)}$ & $\b W_4^{(0)}$ & $\b W_5^{(0)}$ & $\b W_6^{(0)}$ & $\b W_7^{(0)}$ & $\b W_8^{(0)}$ & $\b W_9^{(0)}$ & $\b W_{10}^{(0)}$ \cr
\hline
analyst & dollar & econ. & bank & market & analyst & analyst & analyst & trade & rate \cr
lawson & rate & policy & rate & bank & market & industry & bank & dollar & trade \cr
market & econ. & pct & market & analyst & trade & price & currency & japan & official \cr
trade & mark & cost & currency & price & pct & market & japan & price & bank \cr
sterling & bank & growth & dollar & mark & last & believe & billion & japanese & market \cr
dollar & rise & trade & trade & good & official & last & cut & pct & econ. \cr
\hline
\end{tabular}
\end{center}
\label{table:RandomCBasisVectors}
\end{table*}%
\begin{table*}
\caption{Basis vectors of $\b W^{(0)}$ created by \emph{Co-occurrence Initialization} for \texttt{reuters10}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
$\b W_1^{(0)}$ & $\b W_2^{(0)}$ & $\b W_3^{(0)}$ & $\b W_4^{(0)}$ & $\b W_5^{(0)}$ & $\b W_6^{(0)}$ & $\b W_7^{(0)}$ & $\b W_8^{(0)}$ & $\b W_9^{(0)}$ & $\b W_{10}^{(0)}$ \cr
\hline
dept. & average & agricul. & national & farmer & rate-x & aver price & plywood & wash. & trade \cr
wheat & pct & wheat & bank & rate-x & natl & average & aqtn & trade & japan\cr
agricul. & rate & tonne & rate & natl & avge & price & aequitron & japan & billion\cr
tonne & price & grain & pct & avge & farmer & yield & medical & official & market \cr
usda & billion & farm & oil & cwt & cwt & billion & enzon & reagan & japanese \cr
corn & oil & dept. & gov. & wheat & wheat & bill & enzon & pct & import \cr
\hline
\end{tabular}
\end{center}
\label{table:CooccurrenceBasisVectors}
\end{table*}%
\section{Convergence Criterion}
\label{ConvCriterion}
Nearly all NMF algorithms use the simplest possible convergence criterion, i.e., run for a fixed number of iterations, denoted \texttt{maxiter}. This criterion is used so often because the natural criterion, stop when $\| \b A - \b W \b H \| \le \epsilon$, requires more expense than most users are willing to expend, even occasionally.
Notice that \texttt{maxiter} was the convergence criterion used in the ACLS and AHCLS algorithms of Section \ref{ALSAlg}. However, a fixed number of iterations is not a mathematically appealing way to control the number of iterations executed because the most appropriate value for \texttt{maxiter} is problem-dependent.
In this section, we use the ACLS algorithm applied to the \texttt{cisi} dataset to compare two convergence criterion: the natural but more expensive Frobenius norm measure, and our proposed angular measure, which we describe in the next paragraph. We note that we used the most efficient implementation of the Frobenius measure \cite{Langville2007:nmfsurvey}, which exploits the trace form of the Frobenius norm.
$$
\| \b A - \b W \b H \|_F^2 = trace(\b A^T \b A) - 2 \, trace(\b H^T \b W^T \b A) + trace(\b H^T \b W^T \b W \b H).
$$
In this equation $trace(\b A^T \b A)$ is a constant that does not change throughout the iterations, and thus, is only computed once and stored. At each iteration $2 \, trace(\b H^T \b W^T \b A)$ and $trace(\b H^T \b W^T \b W \b H)$ must be computed. However, some calculations required by these traces, such as $\b W^T \b A$ and $\b W^T \b W$, are already available from the least squares steps, and hence, need not be recomputed.
Our angular convergence measure is moderately inexpensive in storage and computation, and is intuitively appealing. Simply measure the angle $\theta_i$ between successive topic vectors, i.e., $\b W_i^{(j+1)}$ and $\b W_i^{(j)}$ at iterations $j$ and $j+1$. Once $\theta_i \le \epsilon$ for $i=1, \ldots, k$, stop because the topic vectors have converged satisfactorily. Mitchell and Burdick \cite{Mitchell1994} have shown that, in a different context, a similar measure converges simultaneously with the expensive convergence criterion based on the objective function, $\| \b A - \b W \b H \|$ \cite{Smildebook}. However, Figure \ref{figure:convcrit} clearly shows one problem with the angular convergence measure---it does not maintain continual descent, because the basis vectors compared from one iteration to the next are not required to maintain any fixed column order. After several iterations, the column ordering of the basis vectors in $\b W$ is less likely to change, making the angular measure more useful in later iterations of the algorithm. The angular convergence measure is much less expensive to compute than the Frobenius measure, but does require additional storage of the $\b W$ matrix from the previous iteration.
We note in practice, that regardless of the chosen convergence criterion, it is wise to only compute the measure every five or so iterations after some burn-in period.
\begin{figure}[htbp]
\hspace*{-.2in}
\includegraphics[scale=1]{NMFconvcritFro.eps}
\hfill
\includegraphics[scale=1]{NMFconvcritAng.eps}
\caption{Frobenius convergence measure (left) and angular convergence measure (right) of ACLS Algorithm on \texttt{cisi} datasets}
\label{figure:convcrit}
\end{figure}
The recent 2005 reference by Lin \cite{Lin2005} mentioned the related convergence criterion issue of stationarity. The fact that $\| \b A - \b W \b H \|$ (or some similar objective function) levels off does not guarantee stationarity. Lin advocates a stationarity check once an algorithm has stopped. For instance, the stationarity checks of Chu and Plemmons \cite{ChuPlemmons2004} may be used. Lin \cite{Lin2005} proposes a convergence criterion, that simultaneously checks for stationarity, and fits nicely into his projected gradients algorithm. We agree that a stationarity check should be conducted on termination.
\section{Conclusion}
The two new NMF algorithms presented in this paper, ACLS and AHCLS, are some of the fastest available, even faster than truncated SVD algorithms. However, while the algorithms will converge to a stationary point, they cannot guarantee that this stationary point is a local minimum. If a local minimum must be achieved, then we recommend using the results from a fast ALS-type algorithm as the initialization for one of the slow algorithms \cite{Lin2005} that do guarantee convergence to a local minimum. In this paper we also presented several alternatives to the common, but poor, initialization technique of random initialization. Lastly, we proposed an alternative stopping criterion that practical implementations of NMF code should consider. The common stopping criterion of running for a fixed number of iterations should be replaced with a criterion that fits the context of the users and their data. For many applications, iterating until $\| \b A - \b W \b H \|$ reaches some small level is unnecessary, especially in cases where one is most interested in the qualitative results produced by the vectors in $\b W$. In such cases, our proposed angular convergence measure is more appropriate.
|
\section{Introduction}
In navigation-oriented application domains such as autonomous\comment{ vehicles,} mobile robots, wireless sensor networks, security, surveillance, mechanical inspection, and more, graph representations are commonly employed for formulating and analyzing the central\comment{ multi-agent} navigation or area inspection problems. Many approaches to coverage problems \cite{Choset01,Correll08,Easton05} are based on static graph representations, as are visitation problems \cite{MVP} or related combinatorial optimization problems such as the $k$-Chinese Postman Problem \cite{Ahr06,Blum94} and $k$-Traveling Repairman Problem \cite{Edmonds73,Rao07}. But static graph structures do not represent the\comment{ kinds of} dynamic environments that can occur in applications of autonomous robots or non-player characters in video games and virtual worlds. In this paper, we present the \emph{Dynamic Map Visitation Problem} (DMVP), applying recent formulations of \emph{highly dynamic graphs} (or \emph{time-varying graphs} (TVGs)) \cite{Santoro12,Kuhn11} to an essential graph navigation problem: In DMVP, a team of agents must inspect a collection of critical locations on a map (represented as a graph) as quickly as possible, but the agents' environment may change during navigation.
The application of TVG models is essential to DMVP. In applications such as planetary exploration \cite{Wagner99}, search and rescue in hazardous environments (e.g., natural disasters, areas of armed conflict), or even ad-hoc network inspection, many aspects of the structure of graph waypoints and edges governing navigation can change during agent navigation, and TVG models can capture variation in graph structure in ways that static graphs cannot. Our paper presents new results about DMVP complexity and demonstrates distinctions among classes of TVGs; details of our main results are summarized in Section~\ref{ourResults}.
When incorporating dynamics into a problem such as DMVP, there are many options for how to constrain/model the dynamics of the graph. Dynamics can be deterministic (e.g., \cite{Santoro10,Wade11,Wade13,Mans13,Xuan03}) or stochastic (e.g., \cite{Baumann11,Lotker08}). In this paper, to provide a foundation for future work, and exemplify the aspects of topologies and dynamics that make our problem easy or hard, we focus on the deterministic case. The deterministic approach is also particularly relevant for situations in which some prediction of changes is feasible. Quite a bit of this previous work has required that the graph be connected \emph{at all times} \cite{Lotker08,Wade13,Kuhn10}. Indeed, for complete map visitation to be possible, every critical location must be eventually reachable. However, in application environments such as those outlined above, at any given time the waypoint graph may be disconnected. Our model must be general enough to allow for this phenomenon.
We adopt three classes of TVGs, each of which places constraints on edge dynamics. In $\mathcal{R}$, edges must reappear eventually; in $\mathcal{B}$, edges must appear within some time bound; in $\mathcal{P}$ edge appearances are periodic. These classes have proven to be critical to the TVG taxonomy \cite{Santoro12}. They have been studied with respect to problems such as broadcast \cite{Santoro10} and exploration \cite{Santoro13,Wade11}, with results relating to feasibility of computation and bounds on broadcast and exploration time. $\mathcal{R}$, $\mathcal{B}$, and $\mathcal{P}$ place intuitive constraints on the nature of dynamic navigation domains. Even the assumption of periodicity of edges has applications to navigation of transportation networks \cite{Santoro13,Wade11}, as well as environments periodically patrolled by other agents, who can prohibit or guarantee safe traversal of an edge.
In this paper, we shed further light on the computational hierarchy of $\mathcal{R}$, $\mathcal{B}$, and $\mathcal{P}$ \cite{Santoro10}, by analyzing them in the context of DMVP, a natural but difficult problem in global navigation. We provide hardness results for all three classes. For several restricted topologies, we demonstrate separation between the classes by showing severe inapproximability in $\mathcal{R}$, limited approximability in $\mathcal{B}$, and tractability in $\mathcal{P}$. We also give topologies in which DMVP in $\mathcal{R}$ is tractable and fixed parameter tractable, which may serve as a first step towards fully characterizing the topological features that make DMVP difficult. Because our goal in this paper is to cleanly differentiate the classes of dynamics we are exploring, rather than explore the interactions between multiple agents, our results here focus on the case of a single agent.
\subsection{Definitions and TVG Concepts}
\label{definitions}
As a foundation for our work, we adopt the definitions below from Santoro et al. \cite{Santoro12}.
\begin{definition}
A TVG (time-varying graph, dynamic graph, or dynamic network) is a five-tuple $\mathcal{G} = (V, E, \mathcal{T}, \rho, \zeta)$, where $\mathcal{T} \subseteq \mathbb{T}$ is the \emph{lifetime} of the system, \emph{presence function} $\rho(e,t) = 1 \iff$ edge $e \in E$ is available at time $t \in \mathcal{T}$, and \emph{latency function} $\zeta(e,t)$ gives the time it takes to cross $e$ if starting at time $t$. The graph $G = (V,E)$ is called the \emph{underlying graph} of $\mathcal{G}$, with $\lvert V \rvert = n$.
\end{definition}
In the most general case, $\mathbb{T}$ can be $\mathbb{R}$, and edges can be directed. However, in our work we consider the discrete case in which $\mathbb{T} = \mathbb{N}$, edges are undirected, and all edges have uniform travel cost $\zeta(e,t) = 1$ at all times. If agent $a$ is at $u$, and edge $(u,v)$ is available at time $\tau$, then $a$ can take $(u,v)$ during this time step, visiting $v$ at time $\tau+1$. As $a$ traverses $\mathcal{G}$ we say $a$ both \emph{visits} and \emph{covers} the vertices in its traversal, and we will henceforth use these terms interchangeably. A \emph{temporal subgraph} of a TVG $\mathcal{G}$ results from restricting the lifetime $\mathcal{T}$ of $\mathcal{G}$ to some $\mathcal{T}' \subseteq \mathcal{T}$. A \emph{static snapshot} is a temporal subgraph throughout which the availability of each edge does not change, i.e., edges are static.
\begin{definition}
$\mathcal{J} = \{(e_1,t_1),...,(e_k,t_k)\}$ is a \emph{journey} $\iff \{e_1,...,e_k\}$ is a walk in $G$ (called the \emph{underlying walk} of $\mathcal{J}$), $\rho(e_i,t_i) = 1$ and $t_{i+1} \geq t_i + \zeta(e_i,t_i)$ for all $i < k$. The \emph{topological length} of $\mathcal{J}$ is $k$, the number of edges traversed. The \emph{temporal length} is the duration of the journey: $(arrival \ date) - (departure \ date)$.
\end{definition}
Given a date $t$, a journey from $u$ to $v$ departing on or after $t$ whose arrival date $t'$ is soonest is called \emph{foremost}; whose topological length is minimal is called \emph{shortest}; and whose temporal length is minimal is called \emph{fastest}.
In \cite{Santoro12}, a hierarchy of thirteen classes of TVG's is presented. In related work on exploration \cite{Santoro13} and broadcast \cite{Santoro10}, focus is primarily on the chain $\mathcal{R} \supset \mathcal{B} \supset \mathcal{P}$ defined below. We adopt these classes into our domain, which we believe enforce natural constraints in our application environments.
\begin{definition}{(Recurrent edges)} $\mathcal{R}$ is the class of all TVG's $\mathcal{G}$ such that $G$ is connected, and $ \forall e \in E, \forall t \in \mathcal{T}, \exists t' > t$ s.t. $ \rho(e,t') = 1$.
\end{definition}
\begin{definition}{(Time-bounded recurrent edges)} $\mathcal{B}$ is the class of all TVG's $\mathcal{G}$ such that $G$ is connected, and $ \forall e \in E, \forall t \in \mathcal{T}, \exists t' \in [t, t + \Delta)$ s.t. $\rho(e,t') = 1$, for some integer $\Delta$.
\end{definition}
\begin{definition}{(Periodic edges)} $\mathcal{P}$ is the class of all TVG's $\mathcal{G}$ such that $G$ is connected, and $ \forall e \in E, \forall t \in \mathcal{T}, \forall k \in \mathbb{N}, \rho(e,t) = \rho(e,t + kp)$ for some integer $p$. $p$ is called the \emph{period} of $\mathcal{G}$.
\end{definition}
As much as possible, we also take standard notation and terms from the graph theory literature. We rely on several underlying graph topologies. A \emph{star} is a tree in which at most one vertex has degree greater than one. The leaves of a star are called \emph{points}. A \emph{spider} is a tree in which at most one vertex has degree greater than two. In other words, a spider consists of a set of vertex-disjoint paths, called \emph{arms}, each of which has exactly one endpoint connected to the common central vertex $c$. A \emph{comb} is a max-degree 3 tree, in which there exists a simple path containing every vertex of degree 3. Such a path is called a \emph{backbone} of the comb. Paths edge-disjoint to the backbone are called \emph{arms}. A leaf distance 1 from the backbone is called a \emph{tooth}. An \emph{r-almost-tree} is a connected graph with $\lvert V \rvert + r - 1$ edges, that is, $r$ edges can be removed to produce a tree.
\begin{prob}
Given a TVG $\mathcal{G}$ and a set of starting locations $S$ for $k$ agents in $G$, the TVG foremost coverage or dynamic map visitation problem (\emph{DMVP}) is the task of finding journeys starting at time 0 for each of these $k$ agents such that every node in $V$ is in some journey, and the maximum temporal length among all $k$ journeys is minimized. The decision variant asks whether these journeys can be found such that no journey ends after time $t$.
\end{prob}
We think of the input $\mathcal{G}$ as a temporal subgraph of some TVG $\mathcal{G}_\infty$ with lifetime $\mathbb{N}$ and the same edge constraints as $\mathcal{G}$. Thus, the limited information provided in $\mathcal{G}$ is used to compute complete solutions for agents covering $\mathcal{G}_\infty$. When unspecified, assume that DMVP refers to DMVP for a single agent.
\subsection{Main Results}
\label{ourResults}
Our results are summarized in Table~\ref{resultsTable}. We show that DMVP in $\mathcal{R}$ is NP-hard to approximate within any factor, when the underlying graph $G$ is restricted to a star or tree of max degree 3. We show that in $\mathcal{B}$ this problem is NP-hard to approximate within any factor less than $\Delta$, when $G$ is restricted to a spider or tree of max degree 3. We show that in $\mathcal{P}$, DMVP is NP-complete when $p=1$, and that there is a nontrivial class of graphs for which $p=2$ is NP-hard, but $p=1$ is not.
We show that in $\mathcal{R}$, DMVP is solvable in $O(T)$ when $G$ is a path, $O(Tn)$ when $G$ is a cycle, and $O(Tn^3+n^22^n)$ for general graphs, where $T$ is the duration of $\mathcal{G}$, as defined in Section \ref{modelAndProblem}. Furthermore, in $\mathcal{R}$, DMVP is fixed parameter tractable when $G$ is an $m$-leaf $O(1)$-almost tree, and poly-time solvable when $m = O(\lg n)$. In $\mathcal{B}$, we demonstrate a tight $\Delta$-approximation for trees, and a $2\Delta$-approximation for general graphs. We demonstrate a class of problems which are NP-hard in $\mathcal{B}$, but solvable by an online algorithm in $\mathcal{P}$. We show that DMVP in $\mathcal{P}$ is solvable in polynomial time when $G$ is a spider, for fixed $p$, and we show that when $p=2$, DMVP is solvable in linear time for general trees.
\begin{table}[h]
\caption{DMVP separations and results by TVG class and graph class\label{resultsTable}}
\begin{center}
\resizebox{12.5cm}{!} {
\begin{tabular}{| c | c | c | c |}
\hline
\multicolumn{4}{ | c | }{DMVP separations} \\
\hline
TVG class & spiders & max-degree 3 trees & general trees\\\hline
$\mathcal{R}$ & no approx. & no approx. & no approx. \\\hline
$\mathcal{B}$ & tight $\Delta$-approx. & tight $\Delta$-approx. & tight $\Delta$-approx. \\\hline
$\mathcal{P}$ & in P, for fixed $p$ & $O(n)$ exact, for $p=2$ & $O(n)$ exact, for $p=2$ \\\hline
\multicolumn{4}{ | c | }{$\exists$ graph class over which DMVP NP-hard in $\mathcal{P}$ with $p=2$, easy with $p=1$.} \\\hline
\end{tabular}}
\resizebox{12.5cm}{!} {
\begin{tabular}{| c | c | c | c | c |}
\hline
\multicolumn{5}{ |c| }{Complexity of exact algorithms in $\mathcal{R}$} \\
\hline
path & cycle & general graphs & $m$-leaf $c$-almost trees & $O(\lg n)$-leaf $c$-almost trees \\\hline
$O(T)$ & $O(Tn)$ & $O(Tn^3+n^22^n)$ & in FPT & in P \\\hline
\end{tabular}}
\end{center}
\end{table}
The remainder of this paper is organized as follows: preliminaries (\ref{modelAndProblem}), lower bounds (\ref{lowerBounds}), upper bounds (\ref{upperBounds}), open problems and discussion (\ref{openProbs}).
\section{Preliminaries}
\label{modelAndProblem}
For the minimization problem $\mbox{\emph{DMVP}}(\mathcal{G},S)$ and the corresponding decision problem $\mbox{\emph{DMVP}}(\mathcal{G},S,t)$, input is viewed as a sequence of graphs $G_i$ each represented as an adjacency matrix, with an associated integer duration $t_i$, i.e. $\mathcal{G} = (G_1,t_1),(G_2,t_2),$ $...,(G_m,t_m)$, where $G_1$ appears initially at time zero. Let $T = \sum_{i=1}^{m} t_i$. Note that since each $t_i$ can be encoded in $O(\lg t_i)$ space, it is possible for $T$ to be exponential in the size of $\mathcal{G}$. The following observation is required to show that the number of time steps of $\mathcal{G}$ that need to be considered for DMVP is in fact polynomial in the size of $\mathcal{G}$.
\begin{obs}
\label{2n}
When computing DMVP over $\mathcal{G}$, it is not necessary to consider each static temporal subgraph $(G_i,t_i)$ for more than $2n-3$ time steps.
\end{obs}
\begin{proof}
Suppose $G_i$ is the available static subgraph of $\mathcal{G}$ from times $\tau$ to $\tau + t_i$, and $t_i > 2n-3$. Suppose agent $a$ starts at vertex $u$ at time $\tau$. There are two cases:
Case 1: If $a$ can complete its coverage of $\mathcal{G}$ by only traversing in $G_i$, then in the worst case $a$ can execute any complete spanning tree traversal of $G_i$, which takes no more than $2n-3$ steps. In this case, it does not matter at which vertex $a$ ends up, because the task has been completed.
Case 2: If there is a vertex $v$ such that $a$ has not covered $v$ by time $\tau$, and $u$ and $v$ are in different connected components in $G_i$, then $a$ cannot complete coverage of $\mathcal{G}$ when $G_i$ is the available static subgraph. In this case it may matter which vertex $a$ ends up at, depending on which future edges will be available. The size of the connected component of $u$ in $G_i$ is at most $n-1$, so a spanning tree traversal of this component ending up back at $u$ takes no more than $2n-4$ steps. If $a$ would rather end up at a different vertex $w \neq u$, it simply traverses $w$'s branch of the spanning tree last, and returns up only to $w$, in fewer than $2n-4$ steps.
\end{proof}
By Observation \ref{2n}, for any $t_i > 2n-3$, when computing DMVP, all time steps after the first $2n-3$ can be ignored (skipped). DMVP over $\mathcal{G}$ can be computed by computing DMVP over $\mathcal{G}' = (G_1,\min(t_1,2n-3)),...,(G_m,\min(t_m,2n-3))$, and adding back the cumulative time skipped before completion. $\mathcal{G}'$ can clearly be derived from $\mathcal{G}$ in $O(m)$ time. The total duration of $\mathcal{G}'$ is $T' = \sum_{i=1}^{m} \min(t_i,2n-3) < 2nm - 3m$, which is polynomial in $\lvert \mathcal{G} \rvert$. Let $\epsilon(\tau)$ be the time skipped through time $\tau \leq T'$. $\epsilon(\tau)$ can be simply calculated for all $\tau \leq T'$ in $O(T')$ time. A similar $O(T')$ preprocessing step can be run to associate each time $\tau \in T'$ with the corresponding available static graph $G_i$, enabling $O(1)$ edge presence lookups $\rho(e,\tau)$.
Since all of the algorithms we present run in $\Omega(T')$ time, we can run these preprocessing steps for every instance of DMVP and not affect the asymptotic running time. Therefore, for the sake of simplicity, for the rest of our results we assume that this preprocessing has taken place, i.e., we think of $\mathcal{G}$ as $\mathcal{G'}$ and $T$ as $T'$, thereby avoiding the exponential nature of $T$. Note also that for the case of $\mathcal{P}$, the constraint of periodicity implies that it is only necessary to look at $p$ consecutive time steps of the input.
\section{Lower Bounds}
\label{lowerBounds}
As motivation for many of the results in this paper, it is important to note that MVP for a single agent is solvable in linear time on trees \cite{MVP}. To characterize the difficulty of DMVP in $\mathcal{R}$, we first show inapproximability over stars. A similar theorem was independently discovered in \cite{Michail14}.
\begin{theorem}
\label{npostar}
DMVP for a single agent in $\mathcal{R}$ is NP-hard to approximate within any factor, even when the underlying graph is a star.
\end{theorem}
\begin{proof}
We reduce from the set cover problem (SCP). Given a universe $\mathcal{U} = \{1,2,...,m\}$, a family $\mathcal{S}$ of subsets $s_1,s_2,...,s_n$ of $\mathcal{U}$, and an integer $k$, it is NP-complete to decide whether $\mathcal{S}$ contains a cover of $\mathcal{U}$ of size $k$ or less \cite{Karp}. Given an instance of SCP, construct a star $G = (V,E)$ with central vertex $c$; points $v_1,...,v_m$, corresponding to elements in $\mathcal{U}$; $p_1,...,p_n$, corresponding to sets in $\mathcal{S}$; and a single check point $p_0$. We use the following static subgraphs to construct a TVG $\mathcal{G}$. For all $s_i$ in $\mathcal{S}$, let $pass(i) = (V, \{c,p_i\})$, and let $take(i) = (V, E_i)$, where $E_i = \{(c,v_j) : j \in s_i \}$. Let $check = (V,\{c,p_0\})$. Let $finish = (V, F)$, where $F = \{(c,p_i) : 1 \leq i \leq n\}$. Consider the TVG $\mathcal{G} = (pass(1),1)$,$(take(1),t_1)$,$(pass(1),1)$,...,$(pass(n),1)$,$(take(n),t_n)$,$(pass(n),1)$,$(check,2)$,$(finish,2k-1)$, where $t_i = 2\lvert s_i \rvert, \forall i \in \{1,...,n\}$. The total duration of $\mathcal{G}$ is $D = 2n + 2\sum_{i=1}^n \lvert s_i \rvert + 2k + 1 < 2n + 2mn + 2k + 1$.
Consider the problem of deciding if DMVP over $\mathcal{G}$ with a single agent $a$ starting at $c$ has a solution of length no more than $D$. This problem is in NP, since given a journey over $\mathcal{G}$, we can easily check that it hits every vertex, and that all of its edges are available at the correct times. Such a solution can be no longer than $D$, since $D$ is the total duration of $\mathcal{G}$.
Suppose $\mathcal{S}$ contains a cover $\mathcal{C}$ of $\mathcal{U}$ of size $k$ or less. Then for all $s_i \in \mathcal{C}$, $a$ $takes$ $s_i$, that is, $a$ waits at $c$ during both instances of $pass(i)$, and visits all $v_j \in s_i$ and returns to $c$ during $take(i)$, which is possible since the duration of $take(i)$ is $2\lvert s_i \rvert$. Since $\mathcal{C}$ is a cover of $\mathcal{U}$, $a$ visits all $v_i$. For all $s_i \notin \mathcal{C}$, $a$ $passes$ $s_i$, that is, $a$ moves from $c$ to $p_i$ during the first $pass(i)$, waits at $p_i$ during $take(i)$, and returns to $c$ during the second $pass(i)$. During $check$, $a$ moves from $c$ to $p_0$, then back to $c$. At this point, since $\lvert \mathcal{C} \rvert \leq k$, $a$ has passed at least $n-k$ $s_i$'s. So, there are no more than $k$ $p_i$'s left unvisited. $a$ visits these during $finish$, thus completing visitation of all vertices of $G$ in no more than $D$ steps (e.g., Figure~\ref{RstarExample}).
Suppose there exists a solution to this instance of DMVP of length no more than $D$. Prior to $finish$, $a$ must have visited at least $n-k$ $p_i$'s, since $finish$ only lasts for $2k-1$ steps. So $a$ must have passed at least $n-k$ $s_i$'s. Taking and passing for a single $s_i$ are mutually exclusive, because if $a$ moves to $p_i$ during the first $pass(i)$, $a$ must wait during $take(i)$, and if $a$ both takes $s_i$ and moves to $p_i$ during the second $pass(i)$, $a$ will be trapped at $p_i$ until $finish$, and will never be able to reach $p_0$, which must be visited during $check$, the only input time at which $p_0$ is available. Thus, $a$ could have taken no more than $k$ $s_i$'s. During these $k$ or fewer takes, $a$ must have covered all $v_1,...,v_m$. So, the union of these $k$ or fewer $take(i)$'s contains all edges $(c,v_j)$, which implies that the corresponding $s_i$'s form a cover of $\mathcal{U}$ or size $k$ or less. Hence, the decision problem is NP-complete.
Consider the minimization version of the problem with the same setup. Since it is NP-hard to decide if there is a solution of length $D$ or less, it is NP-hard to find such a solution. But after $D$ steps, $a$ may have to wait an arbitrarily long time for the next edge is a feasible solution to appear, so any feasible solution that takes longer than $D$ steps can be arbitrarily long. Therefore, the problem cannot be approximated within any factor.
\end{proof}
\begin{figure}
\includegraphics[scale=1,trim=4cm 4cm 4cm 4cm]{starSnaps}
\caption{Thm.\ref{npostar} static snapshots for $U = \{1,2,3,4,5\}$, $S = \{\{1,2,4\},\{2,4\},\{3,4\},\{3,5\}\}$, $k = 2$.\label{RstarExample}}
\end{figure}
This inapproximability also holds over the restriction of underlying graphs to trees of max-degree 3, in particular, combs.
\begin{theorem}
\label{npocomb}
DMVP for a single agent in $\mathcal{R}$ is NP-hard to approximate within any factor, even when the underlying graph is a comb.
\end{theorem}
\begin{proof}
Analogous to Theorem \ref{npostar}, we reduce from the set cover problem (SCP) \cite{Karp}. Given an instance of SCP, construct a comb $G = (V,E)$ with backbone $b_0b_1b_{m+n+1}$; teeth $v_1,...,v_m$, corresponding to elements in $\mathcal{U}$, with $(b_i,v_i) \in E \ \forall i = 1,...,m$; teeth $p_1,...,p_n$, corresponding to sets in $\mathcal{S}$, with $(b_{m+i},p_i) \in E \ \forall i = 1,...,n$; and two check teeth $p_0$ and $p_{n+1}$ with $(b_0,p_0),(b_{m+n+1},p_{n+1}) \in E$. Let $B = \{(b_0,b_1),...,(b_{m+n-1},b_{m+n})\}$ be the set of all edges in the backbone of $G$. We use the following static subgraphs of $G$ to construct a TVG $\mathcal{G}$. For all $s_i$ in $\mathcal{S}$, let $pass(i) = (V, B \cup \{(b_{m+i},p_i)\})$, and let $take(i) = (V, E_i)$, where $E_i = B \cup \{(b_j,v_j) : j \in s_i \}$. Let $check = (V, \{(b_0,p_0)\})$. Let $finish = (V, F)$, where $F = B \cup \{(b_{m+i},p_i) : 1 \leq i \leq n+1\} \cup \{(b_{m+n+1}, p_{n+1})$. Let $back = (V, B)$.
Define the TVG $\mathcal{G} = (back, m+n),(pass(1),1),(take(1),3m),(pass(1),1),...$,$(back, m+n)$,\\$(pass(n),1)$,$(take(n),3m)$,$(pass(n),1),(back, m+n),(check,2),$$(finish,m+n+2+2k)$. The total duration of $\mathcal{G}$ is $D = n(4m+n+2)+(m+n)+2+(m+n+1+2k) = n^2 + 4mn + 4n + 2m + 2k + 3$.
Consider the problem of deciding if DMVP over $\mathcal{G}$ with a single agent $a$ starting at $b_0$ has a solution of length no more than $D$. This problem is in NP, since given a journey over $\mathcal{G}$, we can easily check that it hits every vertex, and that all of its edges are available at the correct times. Such a solution can be no longer than $D$, since $D$ is the total duration of $\mathcal{G}$.
Suppose $\mathcal{S}$ contains a cover $\mathcal{C}$ of $\mathcal{U}$ of size $k$ or less. Then for all $s_i \in \mathcal{C}$, $a$ $takes$ $s_i$, that is, $a$ travels to $b_0$ during \emph{back}, and visits all $v_j \in s_i$, and returns to the backbone during $take(i)$, which is possible since the duration of $take(i)$ is $3m$, which allows $a$ to take all of the at most $m$ available elements while traveling up the backbone. Since $\mathcal{C}$ is a cover of $\mathcal{U}$, $a$ visits all $v_i$. For all $s_i \notin \mathcal{C}$, $a$ $passes$ $s_i$, that is, $a$ moves to $b_{i+m}$ during $back$, and to $p_i$ during the first $pass(i)$, waits at $p_i$ during $take(i)$, and returns to $b_{i+m}$ during the second $pass(i)$. During the final $back$, $a$ moves to $b_0$, and during $check$, $a$ moves from $b_0$ to $p_0$, then back to $b_0$. At this point, since $\lvert \mathcal{C} \rvert \leq k$, $a$ has passed at least $n-k$ $s_i$'s. So, there are no more than $k$ $p_i$'s left unvisited. $a$ visits these during $finish$, each at cost 2 off the path length $m+n+2$ path up the backbone to $p_{n+1}$, thus completing visitation of all vertices of $G$ in no more than $D$ steps (e.g., Table 1).
Suppose there exists a solution to this instance of DMVP of length no more than $D$. Prior to $finish$, $a$ must have visited at least $n-k$ $p_i$'s, since $finish$ only lasts for $2k-1$. So $a$ must have passed at least $n-k$ $s_i$'s. Taking and passing for a single $s_i$ are mutually exclusive, because if $a$ moves to $p_i$ during the first $pass(i)$, $a$ must wait during $take(i)$, and if $a$ both takes $s_i$ and moves to $p_i$ during the second $pass(i)$, $a$ will be trapped at $p_i$ until $finish$, and will never be able to reach $p_0$, which must be visited during $check$, the only input time at which $p_0$ is available. Thus, $a$ could have taken no more than $k$ $s_i$'s. During these $k$ or fewer takes, $a$ must have covered all $v_1,...,v_m$. So, the union of these $k$ or fewer $take(i)$'s contains all edges $(c,v_j)$, which implies that the corresponding $s_i$'s form a cover of $\mathcal{U}$ or size $k$ or less. Hence, the decision problem is NP-complete.
Consider the minimization version of the problem with the same setup. Since it is NP-hard to decide if there is a solution of length $D$ or less, it is NP-hard to find such a solution. But after $D$ steps, $a$ may have to wait an arbitrarily long time for the next edge is a feasible solution to appear, so any feasible solution that takes longer than $D$ steps can be arbitrarily long. Therefore, the problem cannot be approximated within any factor.
\end{proof}
\begin{figure}
\includegraphics[scale=1,trim=4cm 4cm 4cm 4cm]{combSnaps}
\caption{Thm. 2 sample static snapshots for $U = \{1,2,3,4,5\}, s_1 = \{1,2,4\} \in S$, with $\lvert S \rvert = 4$, and $k = 2$.}
\end{figure}
We have a similar set of lower bounds for the case of $\mathcal{B}$, but with \emph{some} ability to approximate. We later show (Theorem \ref{btree}) that these approximation bounds are indeed tight for all trees.
\begin{theorem}
\label{bapx}
DMVP for a single agent in $\mathcal{B}$ is NP-hard to approximate within any factor less than $\Delta$, even when the underlying graph is a spider, $\forall \Delta > 1$.
\end{theorem}
\begin{proof}
We reduce from 3-partition. Given a multiset of $3m$ positive integers $S = \{s_1,...,s_{3m}\}$, it is strongly NP-complete to decide if they can be partitioned into $m$ sets where all sets have the same sum \cite{Garey}. Let $\sum_{i=1}^{3m} s_i$ = M. Then $B = M/m$ is the required sum for each partition.
Starting with the common central vertex $c$, construct a spider in the following way. For each $s_i \in S$, add a corresponding arm of length $s_i$. Add $m$ arms of length one to be used as checkpoints, and add a single long arm $A_0$ of some length $k > 2M + 2m$ (e.g., Figure \ref{bSpiderExample}). For the TVG used in this proof, arms over any period of time are in one of three modes: $steady$, $flashing$, or $carrying$. When an arm $A$ is steady over a period of time from $\tau$ to $\tau'$, all of its edges are constantly available during that period. When $A$ is flashing, all of its edges synchronously alternate between being unavailable for $\Delta-1$ steps, and available for 1 step, satisfying the time-bounded recurrence constraint. Formally, if $A$ is flashing, then $\forall e \in A, t \in [\tau,\tau'],$
$$
\rho(e,t) =
\begin{cases}
1 & \text{if } t-\tau \equiv p-1 \bmod p, \\
0 & \text{otherwise.}
\end{cases}
$$
When $A$ is carrying, all of its edges act as if $A$ were flashing, with the exception that the edge distance $i$ from $c$ is always available at time $\tau + i$, so that starting at time $\tau$, an agent at $c$ can travel along $A$ for $\tau' - \tau$ steps without waiting. Formally, if $A = ca_0...a_l$ is carrying, then $\forall (u,v) \in E(A), t \in [\tau,\tau'],$
$$
\rho(e,t) =
\begin{cases}
1 & \text{if } t-\tau \equiv p-1 \bmod p, \\
1 & \text{if } v = a_{t-\tau},\\
0 & \text{otherwise.}
\end{cases}
$$
Let $take$ be the temporal subgraph of duration $2B$ in which the arms corresponding to $s_i$'s are steady, and all others are flashing. Let $check$ be the temporal subgraph of duration 2 in which all checkpoint arms are steady, and all others are flashing. Let $finish$ be the temporal subgraph of duration $k$ in which all arms are flashing except for $A_0$, which is carrying. Let $\mathcal{G}$ be the TVG formed by alternating between $take$ and $check$ $m$ times, before ending with $finish$. The total duration of $\mathcal{G}$ is $D = 2M + 2m + k$.
Consider the problem of deciding if DMVP over $\mathcal{G}$ with a single agent $a$ starting at $c$ has a solution of length no more than $D$. Since $k > 2M + 2m$, such a solution must take the long arm last, as traversing this arm twice would result in a solution of length greater than $2k > (2M + 2m) + k = D$. Furthermore, since every arm must be traversed twice except the long arm, the topological length of a solution journey can be no less than $2M + 2m + k = D$. So a solution of temporal length no more than $D$ cannot involve waiting.
Suppose there exists a 3-partition of $S$. During each $take$, $a$ can explore the arms of the spider corresponding to one partition, and return to $c$ in exactly $2B$ steps. During the subsequent $check$, $a$ visits one checkpoint arm, and returns to $c$ in the allotted 2 steps. Repeating this process for the remaining $takes$ and $checks$, $a$ will cover all the $s_i$ arms and checkpoint arms without ever waiting, and end up again at $c$. Finally, during $finish$, $a$ takes the long arm $A_0$, reaching its leaf without waiting, completing coverage in $D$ steps.
Now, suppose this instance of DMVP has a solution of length $D$. To avoid waiting, $a$ must take complete arms and return to $c$ during each $take$, so that it is not stalled by flashing. Similarly, $a$ must take a single checkpoint arm and return to $c$ during each $check$. Doing this $m$ times, $a$ has effectively partitioned the $s_i$ arms into sets each of total length $B$. So, the decision problem is NP-complete, since 3-partition remains NP-complete even when input integers are given in unary. $a$ completes the solution by following the long arm $A_0$, which can be traversed in $k$ steps immediately after $a$ returns to $c$ from the final checkpoint.
Consider the minimization version of the problem with the same setup. Note that if $a$ does not begin taking $A_0$ right as $finish$ begins, $A_0$ will take at least $\Delta(k-1)+1$ to traverse, this best case occurring when $a$ does not have to wait for the first edge. In particular, if $a$ takes $A_0$ last, but has not visited all other arms before $finish$ starts, visiting those arms must have taken at least $2M + 2m + (\Delta - 1) > 2M + 2m, \forall \Delta > 1$, since $a$ must wait at least once during its traversal. The total cost of the solution is then at least $D' = \Delta(k-1) + 1 + 2M + 2m + (\Delta - 1) = \Delta k + 2M + 2m$. If $a$ does not take $A_0$ last, it must traverse $A_0$ twice, taking at least $\Delta(k-1) + 1 + k$ steps (this best case occurring when $a$ starts taking the long arm out right as $finish$ starts), and once it returns must wait at least once while traversing the remaining arms, making the length of the total solution at least $D'' = \Delta(k-1) + 1 + k + 2M + 2m + (\Delta - 1) = \Delta k + 2M + 2m + k > D'$. Take any real constant $\delta < \Delta$. Choose the least integer $N$ s.t. $N > \frac{1}{\Delta-\delta}$. Let $k = N\delta(2M+2m)$. Then,
$$(\Delta-\delta)N\delta(2M+2m) > \delta(2M+2m),$$
$$(\Delta-\delta)k > \delta(2M+2m),$$
$$ \Delta k > \delta(2M + 2m + k),$$
$$ \Delta k + 2M + 2m = D' > \delta D, \ \forall \Delta > 1.$$
Therefore, any solution that contains waiting cannot be a $\delta$-approximation. So, finding a $\delta$-approximation is equivalent to finding a solution with no waiting, i.e., a minimal solution, and thus is NP-hard. Hence, the problem is NP-hard to approximate within any factor less than $\Delta$.
\end{proof}
\begin{figure}
\centering
\includegraphics[scale=1]{underlyingSpider}
\caption{Underlying spider for Thm. \ref{bapx} for 3-partition input: $S = \{2,3,4,4,5,8\}$. \label{bSpiderExample}}
\end{figure}
\begin{theorem}
\label{bcomb}
DMVP for a single agent in $\mathcal{B}$ is NP-hard to approximate within any factor less than $\Delta$, even when the underlying graph is a comb, $\forall \Delta > 1$.
\end{theorem}
\begin{proof}
We use a similar extension to that for $\mathcal{R}$ to extend this result from spider to a comb with \emph{long enough arms}. We again reduce from 3-partition. Given a multiset of $3m$ positive integers $S = \{s_1,...,s_{3m}\}$, it is strongly NP-complete to decide if they can be partitioned into $m$ sets where all sets have the same sum \cite{Garey}. This result clearly still holds even when $m$ is even. So suppose $m$ is even and, let $\sum_{i=1}^{3m} s_i$ = M. Then $B = M/m$ is the required sum for each partition.
Let $l = \frac{7m^2}{2} - \frac{3m}{2} + 4$. Starting with a backbone $\beta=b_1...b_{4m+1}$, construct a comb in the following way: For each $s_i \in S$, add a corresponding arm of length $ls_i$ attached to $\beta$ at $b_{\frac{m}{2} + i}$. Add $m$ arms $c_1,...,c_m$ of length one to be used as checkpoints, with $c_i$ attached at $b_{\frac{m}{2}-\frac{i-1}{2}}$ if $i$ is odd, and $b_{\frac{7m}{2} + \frac{i}{2}}$ if $i$ is even. Add a single long arm of some length $k > 2lBm + \frac{7m^2}{2} - \frac{2m}{2} + 3$, attached at $b_{4m+1}$. For the TVG used in this proof (e.g., Figure \ref{bCombExample}), arms over any period of time are in one of three modes: $steady$, $flashing$, or $carrying$ (see the proof of Theorem \ref{bapx}).
Assume all edges in $\beta$ be available at all times, unless stated otherwise. Let $take$ be the temporal subgraph of duration $2lB + 3m$ in which the arms corresponding to $s_i$'s are steady, and all others are flashing. Let $check(j)$ be the temporal subgraph of duration $i + (i \bmod 2) + 2$ in which checkpoint $c_j$'s arm is steady, and all others are flashing. Let $finalcheck$ be the temporal subgraph of duration $\frac{m}{2} + 2$ in which checkpoint $c_m$'s arm is steady, and all others are flashing. Let $finish$ be the temporal subgraph of duration $k$ in which $\beta$ is flashing, and all arms are flashing except for the long arm of length $k$, which is carrying. Let $\mathcal{G}$ be the TVG $take,check(1),take,check(2),...,take,check(m-1),take,finalcheck,finish$. The duration of $\mathcal{G}$ up until the start of $finish$ is $d = 2lBm + 3m^2 + \sum_{i=1}^m (i + (i \bmod 2) + 2) + \frac{m}{2} + 3 = 2lBm + \frac{7m^2}{2} - \frac{2m}{2} + 3$. The total duration of $\mathcal{G}$ is $D = d + k$.
Consider the problem of deciding whether DMVP over $\mathcal{G}$ with a single agent $a$ starting at $b_\frac{7m}{2}$ has a solution of length no more than $D$. Since $k > 2lBm + \frac{7m^2}{2} - \frac{2m}{2} + 3$, such a solution must take the long arm last, as traversing this arm twice would result in a solution of length greater than $2k > (2lBm + \frac{7m^2}{2} - \frac{2m}{2} + 3) + k = D$.
Suppose there exists a 3-partition of $S$. During the $j$th $take$, if $j$ is odd (even), starting at $b_\frac{7m}{2}$ ($b_{\frac{m}{2}+1}$), $a$ can explore the arms of the spider corresponding to one partition, and end up at $b_{\frac{m}{2}+1}$ ($b_\frac{7m}{2}$) in exactly $2Bl + 3m$ steps. During the subsequent $check(j)$, $a$ visits $c_j$, and returns to $b_{\frac{m}{2}+1}$ ($b_\frac{7m}{2}$) in exactly $i + (i \bmod 2) + 2$ steps. During $finalcheck$, $a$ travels from $b_\frac{7m}{2}$ to $c_m$ and then to $b_{4m+1}$ in the allotted $\frac{m}{2} + 2$ steps. Finally, during $finish$, $a$ takes the long arm, reaching its leaf without waiting, completing coverage in $D$ steps.
Now, suppose this instance of DMVP has a solution of length $D$. If there is no 3-partition of $S$, then $a$ must wait during traversal of at least one arm corresponding to an $s_i$. Since the length of this arm is $ls_i$, $a$ must in fact wait for at least $l$ edges of this arm, incurring a cost of $(\Delta-1)l = (\Delta-1)(\frac{7m^2}{2} - \frac{3m}{2} + 2)$. Since in the length $D$ solution described above $a$ never waits on an arm, this incurred cost must be made up for by minimizing traversal of edges in $\beta$. However, in the length $D$ solution described above, $a$ only traverses $\beta$ for a total of $\frac{7m^2}{2}-\frac{3m}{2}+1 < (\Delta-1)l$ steps, $\forall \Delta > 1$. Therefore, the solution must be in the form of the solution described above in which a 3-partition of $S$ does indeed exists. So, the decision problem is NP-complete, since 3-partition remains NP-complete even when input integers are given in unary.
Consider the minimization version of the problem with the same setup. Note that if $a$ does not begin taking the long arm right as $finish$ begins, the long arm will take at least $\Delta(k-1)+1$ to traverse, this best case occurring when $a$ does not have to wait for the first edge. In particular, if $a$ takes the long arm last, but has not visited all other arms before $finish$ starts, visiting those arms must have taken at least $d + (\Delta - 1)$, since $a$ must wait at least once during their traversal, and the total cost of the solution is then at least $D' = \Delta(k-1) + 1 + d + (\Delta - 1) = \Delta k-1 + d$. If $a$ does not take the long arm last, it must traverse the long arm twice, taking at least $\Delta(k-1) + 1 + k$ steps (this best case occurring when $a$ starts taking the long arm right as $finish$ starts), and once it returns must wait at least once while traversing the remaining arms, making the length of the total solution at least $D'' = \Delta(k-1) + 1 + k + d + (\Delta - 1) = \Delta (k+1) + d > D'$. Take any real constant $\delta < \Delta$. Choose the least integer $N$ s.t. $N > \frac{1}{\Delta-\delta}$. Let $k = N\delta d$. Then
$$(\Delta-\delta)N\delta d > \delta d,$$
$$(\Delta-\delta)k > \delta d,$$
$$ \Delta k > \delta d + \delta k,$$
$$ \Delta k + d = D' > \delta D, \forall \Delta > 1.$$
Therefore, any solution that contains waiting cannot be a $\delta$-approximation. So, finding a $\delta$-approximation is equivalent to finding a solution with no waiting, i.e., a minimal solution, and thus is NP-hard. Hence, the problem is NP-hard to approximate within any factor less than $\Delta$.
\end{proof}
\begin{figure}
\centering
\includegraphics[scale=1]{underlyingComb}
\caption{3-partition underlying comb for some $S$ with $\lvert S \rvert = 6$. Agent starts at $b_8$. \label{bCombExample}}
\end{figure}
As is shown in Section~\ref{upperBounds}, there is a much greater potential for tractability of DMVP in $\mathcal{P}$ than in $\mathcal{B}$ or $\mathcal{R}$. However, the next result follows immediately via reduction from hamiltonian path by simply restricting $t$ to $n-1$.
\begin{theorem}
\label{phard}
DMVP for a single agent in $\mathcal{P}$ is NP-complete, when $p=1$.
\end{theorem}
\begin{proof}
$p=1$ is simply the static case, so the theorem follows immediately from the result that MVP is NP-complete for a single agent on general graphs \cite{MVP}.
\end{proof}
DMVP in $\mathcal{P}$ for $p=1$ is then also NP-complete for all classes of graphs for which hamiltonian path is NP-complete, in particular, planar graphs of maximum degree 3, bridgeless undirected planar 3-regular bipartite graphs, and 3-connected 3-regular bipartite graphs \cite{Saito80}. To show that $\mathcal{P}$ is an interesting dynamics class for DMVP in its own right, it is important to show that DMVP yields different hardness results over $\mathcal{P}$ than over static graphs. Thus, we construct a class of graphs for the following result:
\begin{theorem}
\label{p2}
There is an infinite class of graphs $C$ such that DMVP for a single agent in $\mathcal{P}$ over graphs in $C$ is NP-complete when $p=2$, but trivial when $p=1$.
\end{theorem}
\begin{proof}
Given a graph $G$ with an even number of vertices arbitrarily ordered $v_0,...,v_{n-1}$, construct a corresponding graph $G' \in C$ by adding $n$ new vertices $c_0,...,c_{n-1}$, and adding the edges $(v_i,c_i), (v_i,c_{i+1}),$ and $(c_i,c_{i+1})$ for all $0 < i < n$, where indices are taken mod $n$.
To show the problem is NP-complete for a single agent in $\mathcal{P}$ over graphs in $C$, with $p=2$, we reduce from the hamiltonian path problem \cite{Karp}. Consider a graph $G$ with an even number of vertices $n$, and one of those vertices $v_0$, with the problem of deciding whether $G$ contains a hamiltonian path starting at $v_0$. Take the graph $G' \in C$ corresponding to $G$ as the underlying graph of $\mathcal{G}$. $\mathcal{G}$ begins at time 0. In $\mathcal{P}$ with $p=2$, traversable edges can only be one of three possible types: (01) available at odd times but not even times, (10) available at even times but not odd times, (11) available at all times. Let all original edges of $G$ be of type 11. Let $(v_i,c_i)$ be of type 01 when $i$ is even and type 10 when $i$ is odd. Let $(v_i,c_{i+1})$ be of type 10 when $i$ is even and 01 when $i$ is odd. Let $(c_i,c_{i+1})$ be of type 10 when $i$ is even and 01 when $i$ is odd (see Figure \ref{P2Example}).
Consider the problem of deciding if DMVP over $\mathcal{G}$ for a single agent $a$ starting at $v_0$ has a solution of length no more than $2n-1$, i.e., a solution with no waiting, and in which each vertex is visited exactly once. This problem is clearly in NP. If there is a hamiltonian path in $G$ from $v_0$ vertex $v_i$, then this path will be constantly available in $\mathcal{G}$. $a$ can take this path in $n-1$ steps, and, ending at an odd time, immediately follow $(v_{i},c_{i})$ if $i$ is even, or $(v_{i},c_{i+1})$ if $i$ is odd, then follow either the path $c_ic_{i+1}c_{i+2}...c_{i-1}$ or $c_{i+1}c_{i+2}c_{i+2}...c_i$, the edges for which are always available as $a$ reaches the incoming vertices, thus completing the overall traversal in exactly $2n-1$ steps. Suppose there is a solution to this problem of length $2n-1$. By construction, if $a$ moves to any $c_i$ before covering every $v_j$, $a$ must then wait at least once at some $c_k$ before visiting any further $v_l$. This is because for all $c_i$, once $c_i$ is reached via either $(v_{i-1},c_i)$, $(v_i,c_i)$, or $(c_{i-1},c_i)$ the only edge that can be immediately taken without waiting is $(c_i,c_{i+1})$. So $a$ must visit all $v_j$ exactly once without visiting any $c_i$, thus following a path corresponding to a hamiltonian path in $G$.
However, if we consider the same setup over $G'$ but with $p = 1$, $v_0c_1v_1c_2v_2...$\\$c_{n-1}v_{n-1}c_0$ is always an optimal solution.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[scale=1]{p2Hard}
\end{center}
\caption{$\mathcal{G}$ (from Thm. \ref{p2}) with underlying graph $G'$ constructed from some six-node graph $G$. Edges labeled 10 are available at even times; 01 at odd times. \label{P2Example}}
\end{figure}
\section{Upper Bounds}
\label{upperBounds}
In this section, we map out a class of graphs over which DMVP in $\mathcal{R}$ is solvable in polynomial time. We first start with a very useful lemma. Note that a related observation (about turning around on a ring) was made in \cite{Wade13}.
\begin{lemma}[Turning around lemma]
\label{turnAround}
There is always an optimal solution $J$ that never turns around at a degree 2 vertex of the edge-induced subgraph of $J$ in $G$.
\end{lemma}
\begin{proof}
Suppose $v$ is a degree 2 vertex with neighbors $u,w$ in the edge-induced subgraph of $J$ in $G$. Suppose agent $a$ takes edge $(u,v)$ at time $\tau$, then \emph{turns around}, taking $(u,v)$ at time $\tau'$ as the very next edge in its traversal. Since $(v,w)$ is in the edge-induced subgraph of $J$, $a$ must traverse $(v,w)$ at some other time, thereby reaching $v$ at that time. So, $a$ could have waited at $u$ from times $\tau$ to $\tau' + 1$, instead of taking $(u,v)$ at time $\tau$, and the solution would still be optimal (see Figure~\ref{waysToCover}).
\end{proof}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=1.8]{possibleWays}
\end{center}
\caption{The 7 ways, satisfying Lemma 1, of covering the vertices of a length 5 path with degree 2 intermediate nodes. \label{waysToCover}}
\end{figure}
We apply Lemma~\ref{turnAround} to get the following solvability results for restricted classes of underlying graphs.
\begin{theorem}
\label{line}
DMVP for a single agent in $\mathcal{R}$ on a path is solvable in $O(T)$ time.
\end{theorem}
\begin{proof}
Consider DMVP for a single agent $a$ with underlying graph $G$ the path $v_0v_1...v_n$, and $a$ starting at $v_k$. To reach $v_0$, $a$ must cover all $v_{k-1},...,v_1$ along the way. Similarly, to reach $v_n$, $a$ must cover all $v_{k+1},...,v_{n-1}$. By Lemma \ref{turnAround}, an optimal solution can be constructed by first taking a foremost journey to either $v_0$ or $v_n$, then taking the foremost journey to the remaining endpoint. One of these two topological journeys, called the $left$ and $right$ journeys, must embody an optimal solution, but in the worst case edge availability must be checked for all $t \in \mathcal{T}$, yielding an $O(T)$ runtime.
\end{proof}
\begin{algorithm}[t]
\caption{DMVP-Line($\mathcal{G}, \{v_k\})$ \label{linealg}}
\begin{algorithmic}
\State $lLoc = rLoc = k$ \Comment{Both possible journeys start at $v_k$}
\State $lTurned = rTurned = complete = False$
\State $t = 0$
\While{not $complete$}
\If{not $lTurned$} \Comment{Advance left journey if possible}
\If{$\rho((v_{lLoc},v_{lLoc-1}),t) = 1$}
\State $lLoc = lLoc - 1$
\If{lLoc = 0} \Comment{Left endpoint reached}
\State $lTurned = True$
\EndIf
\EndIf
\Else
\If{$\rho((v_{lLoc},v_{lLoc+1}),t) = 1$}
\State $lLoc = lLoc + 1$
\If{rLoc = n} \Comment{Right endpoint reached}
\State $complete = True$
\EndIf
\EndIf
\EndIf
\If{not $rTurned$} \Comment{Advance right journey if possible}
\If{$\rho((v_{rLoc},v_{rLoc+1}),t) = 1$}
\State $rLoc = rLoc + 1$
\If{rLoc = n} \Comment{Right endpoint reached}
\State $rTurned = True$
\EndIf
\EndIf
\Else
\If{$\rho((v_{rLoc},v_{rLoc-1}),t) = 1$}
\State $rLoc = rLoc - 1$
\If{rLoc = 0} \Comment{Left endpoint reached}
\State $complete = True$
\EndIf
\EndIf
\EndIf
\State $t = t + 1$ \Comment{Step}
\EndWhile
\State \Return $t$
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\label{cycle}
DMVP for a single agent in $\mathcal{R}$ on a cycle is solvable in $O(Tn)$ time.
\end{theorem}
\begin{proof}
A similar case to Theorem \ref{line} can be made for the cycle $C = v_0v_1...v_nv_0$. Suppose $a$ starts at $v_0$ at time $0$. Consider an optimal visitation of $C$ for $a$. In this optimal solution, there is some vertex $v_k \neq v_0$ that is visited last. The second to last vertex is then either $v_{k-1}$ or $v_{k+1}$. If it is $v_{k-1}$, then $a$ must have already visited $v_{k+1}$ without visiting $v_k$. So, the edge $(v_{k+1},v_k)$ is never traversed. Therefore, the solution reduces to an optimal solution over the path graph $v_kv_{k-1}...v_{k+1}$ starting at $v_0$. Similarly, if instead $v_{k+1}$ is the vertex visited second to last, then $a$ must have already visited $v_{k-1}$ without visiting $v_k$, and the solution reduces to an optimal solution over the path $v_{k-1}v_{k-2}...v_{k+1}v_k$. Since there are $n-1$ possible final vertices for an optimal solution, the cost of an optimal solution can be computed by for each of these vertices computing the minimal cost between optimal coverage of each of the two corresponding paths using Algorithm~\ref{linealg}, and taking the minimum over all $n - 1$ vertices possible final vertices (see Figure \ref{waysToCycle}). This yields an $O(Tn)$ runtime.
\end{proof}
\begin{figure}
\centering
\includegraphics[scale=2]{cycleWays}
\caption{The 8 possible underlying walks of solutions, satisfying Lemma 1, to the 5-cycle starting at $v_0$. \label{waysToCycle}}
\end{figure}
Now we show that despite the severe inapproximability of DMVP over $\mathcal{R}$, we can always compute an optimal solution in exponential time.
\begin{theorem}
\label{dynam}
DMVP for a single agent in $\mathcal{R}$ is solvable in $O(Tn^3+n^22^n)$ time.
\end{theorem}
\begin{proof}
The proposed algorithm first computes all-pairs-all-times-foremost-journey of input TVG $\mathcal{G}$, using a straightforward dynamic programming algorithm, then uses this information to run another dynamic programming algorithm, conceived along the lines of a standard method for TSP \cite{Bellman}.
Let $d^t_{uv}$ be the length of the foremost journey from $u$ to $v$, starting at time $t$. Algorithm \ref{apatfj1} computes $d^t_{uv}$ for all vertex pairs $(u,v)$, and times $t \in \mathcal{T}$ for a given TVG $\mathcal{G}$.
At all times $t$, for all vertices $u \in V$, $d^t_{uu}$ is clearly 0. At time $T$, the time limit has been reached, so an agent cannot move to another vertex in any guaranteed time, and thus we set $d^T_{uv} = \infty$ for all $u \neq v$. For all $T - 1 \geq t \geq 0$, in the worst case an agent can wait at $u$ for one step, and take the foremost journey to $v$ starting at time $t+1$. If there is a better journey than this, it must consist of not waiting, rather taking one of the edges available at time $t$ from $u$ to some vertex $k$. Subsequently taking the foremost journey from $k$ to $v$ starting at time $t+1$ results in an optimal journey through $k$. Algorithm \ref{apatfj1} clearly runs in $O(Tn^3)$ time, and uses $O(Tn^2)$ space.
\begin{algorithm}[h]
\caption{All-pairs-all-times-foremost-journey($\mathcal{G}$) \label{apatfj1}}
\begin{algorithmic}
\ForAll{$u,v \in V \times V$} \Comment{Initialize base case for $t = T$.}
\If{$u = v$}
\State $d^T_{uv} = 0$
\Else
\State $d^T_{uv} = \infty$ \Comment{Since input ends at $T$, agent cannot move.}
\EndIf
\EndFor
\\
\For{$t = T-1,...,0$} \Comment{Work backwards until start time $t = 0$.}
\ForAll{$u,v \in V \times V$}
\If{$u = v$}
\State $d^t_{uv} = 0$
\Else
\State $d^t_{uv} = d^{t+1}_{uv} + 1$ \Comment{In worst case, just wait at $u$.}
\ForAll{$k \in V$}
\If {$\rho((u,k),t) = 1$} \Comment{Check for better route.}
\State $d^t_{uv} = \min(d^t_{uv},d^{t+1}_{kv} + 1)$
\EndIf
\EndFor
\EndIf
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
Algorithm \ref{expo1} uses the $d^t_{uv}$ values computed by Algorithm \ref{apatfj1} to compute the cost of a minimal solution to DMVP for a single agent in $\mathcal{R}$. Let $V' \subseteq V$ and $c(V',v)$ be the minimal time it takes to visit all vertices in $V'$ starting at vertex $s$ at time $0$ and ending at vertex $v \in V'$.
After initializing the minimal costs for visiting subsets up to size 2, the algorithm repeatedly uses the minimal costs for size $i$ subsets to calculate $c(V',v)$ for each size $i+1$ subset $V'$ and $v \neq s \in V'$. Once computed up to size $n$, the algorithm returns the minimal cost among journeys that cover all vertices. This is an optimal solution to DMVP as it is the minimum cost of taking foremost journeys between vertices that results in a complete cover. There are $2^n$ subsets of $V$, and so $n2^n$ subset-vertex pairs of the form $(V',v)$. For each of these, the algorithm computes the minimum of $O(n)$ values. So, Algorithm \ref{expo1} has running time $O(n^22^n)$. Since it saves one cost for each subset-vertex pair, Algorithm \ref{expo1} also uses $O(n2^n)$ space. Sequentially running Algorithm \ref{apatfj1} followed by Algorithm \ref{expo1}, we have a complete algorithm for DMVP for a single agent in $\mathcal{R}$, with combined running time $O(Tn^3+n^22^n)$.
\end{proof}
\begin{algorithm}[h]
\caption{$\mbox{\emph{DMVP}}(\mathcal{G},\{s\})$ \label{expo1}}
\begin{algorithmic}
\State $c(\{s\},s) = 0$ \Comment{Initialize subset of size 1.}
\ForAll{$v \neq s \in V$} \Comment{Initialize subsets of size 2.}
\State $c(\{s,v\},v) = d^0_{sv}$
\EndFor
\For{i = 3,...,n} \Comment{Build up to subsets of size n.}
\ForAll{$S \subseteq V s.t. \lvert S \rvert = i$}
\ForAll{$v \neq s \in V$}
\State $c(V',v) = \min_{u \neq s \in V' \setminus \{v\}}(c(V' \setminus \{v\}, u) + d^{c(V' \setminus \{v\},u)}_{uv})$
\EndFor
\EndFor
\EndFor
\Return $\min_{v \neq s \in V}(c(V,v))$
\end{algorithmic}
\end{algorithm}
Almost-trees have been previously studied with respect to fixed parameter tractability (e.g., \cite{Fiala01}). We use Theorem \ref{dynam} to generalize Theorems \ref{line} and \ref{cycle} with the following:
\begin{theorem}
\label{fpt}
DMVP in $\mathcal{R}$ is fixed parameter tractable, when $G$ is an $m$-leaf $c$-almost-tree, for fixed parameter $m$, and $c$ constant.
\end{theorem}
\begin{proof}
First, consider the restricted case where $G$ is an $m$-leaf tree. Since every leaf must be visited, and visiting all leaves implies coverage of the entire tree, there is a minimal solution that can be thought of as an ordering of the set of leaves of $G$, and the foremost journeys between them. In this case, there is only one \emph{way} to visit any node, namely, on the way to a leaf. Using this observation and Algorithm \ref{expo1} from the proof of Theorem \ref{dynam}, we see that we only need to consider all orderings of leaves, instead of all orderings of vertices, yielding a run time of $O(Tn^3 + m^22^m)$, which is indeed fixed parameter tractable for parameter $m$.
Suppose the underlying graph $G$ of $\mathcal{G}$ is an $m$-leaf $c$-almost-tree. Consider all edges $e$ such that removing $e$ from $G$ results in a $(c-1)$-almost-tree. Each of these edges lies on some path $P$ such that removing any edge of $P$ will similarly result in a $(c-1)$-almost-tree, and every intermediate vertex on the path has degree 2. Suppose $P$ is the path $v_0...v_l$. Since $G$ is an $m$-leaf $c$-almost-tree, there are $O(m)$ paths of this type. The edge-induced subgraph $G'$ of the underlying walk of an optimal covering of $\mathcal{G}$ can be any $(c-c')$-almost-tree $\subseteq G$, for $0 \leq c' \leq c$. For each $c'$, a solution involves selecting $c'$ paths, each of $O(n)$ length, from which to remove an edge. So, there are $O(m^{c'}n^{c'})$ possible choices of $(c-c')$-almost-trees, and thus $O(\sum_{c'=0}^c (m^{c'}n^{c'})) = O(m^cn^c)$ choices for $G'$. Every $G'$ has no more than $m+2c$ leaves. Since every edge of $G'$ is covered, by Lemma~\ref{turnAround}, there are at most 2 ways to cover each of the remaining $O(m)$ paths $v_0...v_l$ of intermediate vertex degree 2, namely: entering at $v_0$ and exiting at $v_l$, or entering at $v_l$ and exiting at $v_0$. Augment the set of leaves to be ordered in a solution with the selected ways of covering these paths, that is, select one of the consecutive subsequences $v_0v_l$ or $v_lv_0$ to be in the ordering. With this augmentation, we still have a set of $O(m)$ elements to be ordered, the optimal ordering of which can be computed via Theorem \ref{dynam} in $O(Tn^3 + m^22^m)$ time. The minimum over all ways of covering $G'$ can then be computed in $O(2^m)O(Tn^3 + m^22^m) = O(Tn^32^m + m^22^{2m})$. The overall minimum cost for covering $\mathcal{G}$ can then be computed by taking the minimum cost over all $O(m^c n^c)$ edge-induced subgraphs in $O(m^{c'}n^c)O(Tn^32^m + m^22^{2m}) = O(Tn^{3+c}f(m))$ time.
\end{proof}
The following result follows immediately for the case when $m = O(\lg n)$.
\begin{corollary}
\label{poly}
DMVP in $\mathcal{R}$ is solvable in polynomial time, if $\mathcal{G}$ is an $O(\lg n)$-leaf $c$-almost-tree, for $c$ constant.
\end{corollary}
We conjecture (see Section \ref{openProbs}) that the maximal class of graphs over which DMVP in $\mathcal{R}$ is poly-time solvable is the class of all graphs with polynomially many spanning trees, all of which have $O(\lg n)$ leaves.
Since DMVP in $\mathcal{B}$ is bounded by $2\Delta n$, the running time of the algorithm in Theorem \ref{dynam} on TVGs over $\mathcal{B}$ reduces to $O(\Delta n^4+n^22^n)$. We also see that we are able to greatly improve on approximation from $\mathcal{R}$ to $\mathcal{B}$:
\begin{theorem}
\label{btree}
DMVP in $\mathcal{B}$ over a tree can be $\Delta$-approximated in $O(n)$ time. This approximation is tight.
\end{theorem}
\begin{proof}
In \cite{MVP}, it is shown that minimal MVP cost $C$ can be computed in $O(n)$ for static graphs. In the dynamic case, the journey corresponding to following exactly the edges in the static solution when they are available can be followed, waiting at most $\Delta-1$ steps for each edge to appear before it is traversed. Since no solution can be better than $C$, and the proposed journey takes at most $\Delta C$, it must be a $\Delta$-approximation. From Theorems \ref{bapx} and \ref{bcomb}, we know there can be no better approximation. Hence, this approximation is tight.
\end{proof}
\begin{theorem}
\label{bapprox}
DMVP in $\mathcal{B}$ can be $2\Delta$-approximated by any online spanning tree traversal of $G$.
\end{theorem}
\begin{proof}
The topological length of a spanning tree traversal is no more than $2n-3$. In the worst case, waiting $\Delta-1$ time steps for each subsequent edge to appear results in coverage of $\mathcal{G}$ in $2n\Delta - 3\Delta$ steps. The fastest possible coverage of $\mathcal{G}$ is via the traversal of a hamiltonian path in $G$ without waiting, which takes $n-1$ steps, and $2\Delta(n-1) > 2n\Delta - 3\Delta$.
\end{proof}
Theorems \ref{bapx} and \ref{bcomb} show the tightness of Theorem \ref{btree}. Here, $\mathcal{B}$ is starkly differentiated from $\mathcal{R}$ in that we have at least some ability to approximate in $\mathcal{B}$. See Section \ref{openProbs} for a further discussion of the relationship between these two classes.
Similar to the case for $\mathcal{B}$, DMVP in $\mathcal{P}$ is bounded by $2pn$, so the running time of the algorithm in Theorem \ref{dynam} reduces to $O(pn^4+n^22^n)$. To exemplify the differences between $\mathcal{P}$ and $\mathcal{B}$, and motivate interest in the tractability of DMVP over $\mathcal{P}$, we first give the following simple example:
\begin{theorem}
\label{pcomb}
For any $p$, there is a class of problems over combs, for which DMVP in $\mathcal{B}$ is NP-hard, but in $\mathcal{P}$ is solvable by the online algorithm: take arms when you get to them.
\end{theorem}
\begin{proof}
Suppose an agent $a$ starts at $b_0$. $a$ can either take $A_0$ immediately, or travel along $B$ to visit other arms and return to visit $A_0$ at a later time. Suppose the fastest an agent starting at $b_0$ can visit the leaf of $A_0$ and return to $b_0$ is $l$ steps. Then the longest this could possibly take $a$ starting at time 0 is $l + (p-1)$ steps, this worst case occurring when $a$ must wait $p-1$ steps for the fastest journey to become available. Suppose the fastest journey from $b_0$ to $b_k$ takes $k'$ steps. Then in the worst case, traveling from $b_0$ to $b_k$ takes $k' + (p-1)$ steps. Suppose the fastest coverage of the remaining induced subgraph $G' = G \setminus (A_0 \cup \{b_0,...,b_{k-1}\})$ takes $m$ steps.
If $G'$ has only one arm, the foremost journey from $b_k$ to the leaf of this arm is clearly optimal. Assume that if $G'$ has $\alpha$ arms, then the following online algorithm results in an optimal solution: if at the endpoint of an unvisited arm, take that arm, otherwise visit the next unvisited vertex of $B$. In the $\alpha+1$ arm case, our agent starting at $b_0$ using this algorithm will complete coverage in no more than $l + (p-1) + k' + (p-1) + m = l + k + (2p-2) + m$ steps. But any solution in which $a$ does not take $A_0$ first must cost at least $l + k' + k + m$, as $a$ must retraverse $b_k...b_0$ on its way back to cover $A_0$. Since $k \geq 2p-2$, the online solution must be minimal.
It is straightforward to modify Theorem \ref{bcomb} (i.e., by appropriately scaling up the underlying graph: elongating arms, extending the backbone, separating arms along the backbone, and adding an additional check tooth at $b_0$) to show that DMVP over the above class of combs is NP-hard in $\mathcal{B}$, for a single agent starting at $b_0$.
\end{proof}
\begin{figure}
\centering
\includegraphics[scale=1]{longComb}
\caption{Segment of an underlying comb for which DMVP in $\mathcal{P}$ for an agent starting at $b_0$ is solved by the simple online tree traversal algorithm, regardless of the lengths of each arm $A_i$ \label{imgLongComb}}
\end{figure}
The quality of $\mathcal{P}$ we take advantage of above is that if the fastest journey between two nodes takes $d$ steps, the foremost journey can take no longer than $d + (p-1)$, while in $\mathcal{B}$ it can be as bad as $d\Delta$. We again harness this effect in the following result, a stronger theorem in the context of our inapproximability results for $\mathcal{R}$ and $\mathcal{B}$ (Theorems \ref{npostar} and \ref{bapx}):
\begin{theorem}
\label{pspider}
DMVP in $\mathcal{P}$ over a spider is solvable in polynomial time, for fixed $p$.
\end{theorem}
\begin{proof}
Starting at the center $c$ of the spider, it is never useful for an agent to travel along any arm, unless it reaches a leaf. That is, an optimal solution is essentially an optimal visitation order of the leaves. We can set up a cost function $c(l,t)$ giving the minimal time it takes to travel from $c$ to leaf $l$ and back, starting at time $t$. Since $\mathcal{G}$ is periodic, $c(l,t) = c(l,t+kp) \ \forall k \in \mathbb{Z}$. Suppose the fastest journey from $c$ to $l$ and back has cost $m(l)$. Let \emph{extra time} $e(l,t) = c(l,t) - m(l)$, be the cost above minimum incurred by traveling to $l$ and back starting at time $t$. $0 \leq e(l,t) < p \ \forall l, t$, since $a$ can always simply wait at most $p-1$ steps for the periodically fastest journey to be available. For any $l$, there are only $p^2$ possible functions $e(l,t)$, since for all $0 \leq t\pmod{p} \leq p-1$, $0 \leq e(l,t) < p$. Let $r(l,t)$ be the \emph{return time} $\mod p$ of traveling to $l$ and back, that is, $c(l,t) = i \implies r(l,t) = t + i \pmod{p}$. Classify each $l$ by $e$ and $r$. Let $l_1 \equiv l_2 \iff e(l_1,t) = e(l_2,t)$ and $r(l_1,t) = r(l_2,t) \ \forall t$. Since for each $t$ there are $p$ choices for $e$ and $p$ choices for $r$, there are $p^3$ such equivalence classes. During a traversal of the spider, leaves with a common equivalence class are interchangeable, since at a given time, taking any of the same class will result in equivalent incurrence of cost above minimum as well as equivalent return time. Thus, a minimal traversal consists of traversing an ordering of arms corresponding to a sequence of equivalence classes $q_i$, such that the frequency of $q_i$ in the sequence is the number of arms classified as $q_i$.
Notice that for any length $p$ traversed sequence $q_1q_2...q_p$, by the pigeonhole principle, there must be $q_i, q_j$ with $i < j$ such that $r(q_i,t_i) = r(q_j,t_j)$, where $t_i$ is the start time for traversing the $q_i$ arm, and $t_j$ is the start time for traversing the $q_j$ arm. Let $Q^t$ be a \emph{pattern} if $Q$ is a sequence of equivalence classes with $0 < \lvert Q \rvert \leq p$, and starting at $c$ at time $t$, the traversal of $Q$ returns at some time $t' \equiv t \pmod{p}$. Furthermore, $Q^t$ is not a pattern if it contains any subpatterns, i.e., traversed subsequence with equivalent start and end time. Any length $p$ sequence must contain a pattern. An optimal solution can be decomposed into a sequence alternating between patterns and non-pattern subsequences between patterns. Any pattern can be removed from its location starting at time $t$ and inserted at any different location $t' \equiv t \pmod{p}$, since the fact that the pattern has the equivalent start and end time means adjacent journeys will be unaltered, due to the periodicity of $\mathcal{G}$. In particular, any pattern $Q_1^t$ can be removed from its current location and inserted after any $Q_2^t$, without changing the cost of the solution. So, given any optimal solution, the following reordering process does not change the cost of the solution:
\begin{enumerate}
\item Divide the sequence into patterns $Q^t$ and stray arms in classes $q$ not in patterns
\item Set $i = 0$, $S = \{0\}$
\item Sequence all $Q^i$ together starting at $i$
\item Identify new patterns created by this move
\item Repeat 2 and 3 until nothing changes
\item Consider earliest start time $j$ of an arm such that $j \pmod{p} \notin S$ after final $Q_i$
\item Let $i = j$, and add $j$ to $S$.
\item Repeat 3 through 8 until nothing changes
\end{enumerate}
After this process, all $Q_i$ are grouped together for all $i$, with fewer than $p$ stray arms separating each of these sequences of patterns, since in step 6, if there is no such $j$, then there must be fewer than $p$ arms left, otherwise there would be a pattern among these arms. Thus, the reordered sequence also ends with fewer than $p$ stray arms. Since we started with an optimal solution, and the above process did not change the cost of the solution, there must be an optimal solution of this form. Since every such reordering begins with the $Q_0$ patterns, there are $(p-1)!$ orders in which the $p$ grouped sequences of patterns can show up. For each of these, there are at most $p$ clusters of stray arms each of length less than $p$. There are $O(p^5)$ ways to fill up these $O(p^2)$ slots with arms from the $p^3$ classes. Since there are $O((p^3)^p)$ possible patterns, there are $O(n^{(p^3)^p})$ ways to partition the remaining $O(n)$ arms in each class into patterns, and therefore $O(n^{(p^3)^pp^3}) = O(n^{(p^3)^{p+1}})$ ways to partition all classes into patterns. This yields $O((p-1)!p^5n^{(p^3)^{p+1}}) = O(n^{(p^3)^{p+1}})$ possible solutions of the form reached by executing steps 1-8 above, at least one of which must be optimal. The cost of each of these possible solutions, of which there are polynomially many in $\lvert \mathcal{G} \rvert$, can be easily computed in time polynomial in $\lvert \mathcal{G} \rvert$. If $a$ does not start at $c$, but rather on some arm $A$, $a$ can either visit the leaf of $A$ before returning or $c$, or return to $c$ directly. Compute DMVP for the remaining arms in each of these two cases will yield an overall optimal solution still in time polynomial in $\lvert \mathcal{G} \rvert$.
\end{proof}
This polynomial runtime can be significantly improved for the case of $p=2$.
\begin{theorem}
\label{p2GeneralTree}
DMVP in $\mathcal{P}$ over a tree is solvable in $O(n)$ time, when $p = 2$.
\end{theorem}
\begin{proof}
Consider DMVP in $\mathcal{P}$, with $p = 2$, over a tree $T$, for an agent starting at root $o$ at time 0. The proof proceeds as follows: We first show by induction that there is always an optimal solution that never enters any of the subtrees of $o$'s children more than once. We then show that when covered in its entirety, each subtree is of one of three types: ($fw11$) fastest coverage with return to root is always available, ($fw10$) fastest coverage is available only at even times, and ($fw01$) fastest coverage is available only at odd times. Alternating between $fw10$ and $fw01$ subtrees, and then taking the remaining subtrees in any order, before ending at a \emph{furthest leaf} results in an optimal solution, as we maximize how many subtrees are traversed optimally. We can recursively compute the type and costs of covering the maximal subtree rooted at each node $v$, in $O(\deg(v))$ time for each.
Suppose $o$ has adjacent edges $e_1 = (o,u_1),...,e_d = (o,u_d)$. Let $T^{u_i}$ be the maximal subtree rooted at $u_i$. Suppose agent $a$ starts at $o$. Recall that since $T$ is a tree, an optimal solution can be characterized as the set of leaves ordered by when they are visited.
Suppose an optimal solution ends on some leaf of $T^{u_i}$. Assume that when $T^{u_i}$ can be entered at most $k$ times during a solution, the solution that enters only once is still optimal. Suppose an optimal solution $S_{k+1}$ enters $T^{u_i}$ $k+1$ times. Then there is some non-empty subgraph $T^{u_i}_k$ of $T^{u_i}$ that $a$ covers the upon the $k$th entry, and $T^{u_i}_{k+1}$ it covers upon the $(k+1)$st entry. Let $T'$ be the subgraph of $T$ covered between the $k$th and $(k+1)$st entries into $T^{u_i}$. When $a$ arrives at $o$ before entering $T^{u_i}$ for the $k$th time, consider an alternate completion resulting in an alternate solution $S_k$, in which $a$ instead immediately covers $T'$ and ends up back at $o$, at a cost of at most 1 plus the cost of this coverage in $S_{k+1}$ (which must occur in $S_{k+1}$ after the $kth$ entry into $T^{u_i}$). If $S_k$ does incur a cost over $S_{k+1}$ for this traversal, then both $S_k$ and $S_{k+1}$ reach $o$ at times $\tau$ and $\tau'$ with equal parity. In $S_k$, $a$ completes coverage of $T$ during its $k$th entry to $T^{u_i}$, by covering $T^{u_i}_k$ and returning to $v_i$, then immediately covering $T^{u_i}_{k+1}$. If $\tau$ and $\tau'$ have equal parity, then $T^{u_i}_{k+1}$ is traversed in both $S_k$ and $S_{k+1}$ in equal time. Otherwise, $S_k$ may incur a cost of 1 over $S_{k+1}$ for this traversal. Therefore, the combined coverage of $T'$ and $T^{u_i}_k$ costs at most 1 more in $S_k$ than in $S_{k+1}$. $S_k$ may incur an additional cost of 1 over $S_{k+1}$ for its completing coverage of $T^{u_i}_{k+1}$, for a total incurrence of at most 2 over $S_{k+1}$ for these traversals. However, $S_{k+1}$ contains two additional traversals of $e_i$, so the total cost of $S_{k+1}$ can be no better than that of $S_k$. Therefore, the solution that enters $T^{u_i}$ only once is optimal.
Recall (from the proof of Theorem \ref{p2}) that in $\mathcal{P}$ with $p=2$, there are only three relevant dynamic edge types: 01, which are available at odd times but not even; 10, which are available at even times but not odd; and 11, which are available at all times. Let $T^u$ be a maximal subtree of $T$ rooted at $u$. Let $c_w(T^u,i)$ be the cost of the foremost journey covering $T^u$, starting at time $t \equiv i \bmod 2$, with returning to end back at $u$. Let $c(T^u,i)$ be the cost of the foremost journey covering without necessarily returning to $u$. Let $m_w(T^u)$ be the fastest cost of covering $T^u$ with returning to end back at $u$, and $m(T^u)$ be the fastest cost of covering $T^u$ without necessarily returning to $u$. Notice that, since $p = 2$, the foremost cost of these coverings at any time can be no more than $c_w(T^u,i) = m_w(T^u) + 1$ and $c(T^u,i) = m(T^u) + 1$, respectively. Classify $T^u$ as $fw11$ if a fastest coverage with return is always available; $fw10$ if it is available at even times, but not odd; $fw01$ if at odd times, but not even.
To enable our characterization of an optimal ordering of subtrees, it is necessary that for any tree $T$, if fastest coverage of $T$ with return to $o$ is always available, then fastest coverage takes even time, otherwise, fastest coverage takes odd time. To see this, first consider height 1 trees (i.e., stars). Regardless of when $a$ starts, if the numbers of 01 and 10 edges are equal, then $a$ can alternate between them, taking each at cost 3, and taking all 11 edges at cost 2, the total cost of which will be even. If $a$ starts at time 0 and there are more 10's than 01's, then $a$ can take one more 10 than 01 at cost 3, and take the rest at cost 4, for an odd total. Starting at time 1 $a$ cannot save on this extra edge. A similar case applies when there are more 01's than 10's, but for opposite start times. For trees of greater depth, since each subtree need not be entered more than once, we can use similar reasoning to the height 1 case. If there are the same number of $fw01$ as $fw10$ subtrees, alternate between them to take them all fastest (each at odd cost), so fastest coverage is always available at even cost. Otherwise, one more subtree of the more plentiful type could be taken fastest, depending on start time.
Since each maximal subtree is covered independently, we can (as in the proof of Theorem \ref{pspider}) look for \emph{patterns} in a solution given as an optimal ordering of child subtrees. Notice that in this case, as a result of the parity of coverage with return to $o$, there are only three patterns: $(fw11)$, $(fw10,fw01)$, and $(fw01,fw10)$. Given the above information for all of a tree $T^v$'s maximal child subtrees, we compute these values for $T^v$:
Following the characterization of an optimal solution given in the proof of Theorem \ref{pspider}, we can construct $S_0$, an optimal solution with return that starts at time $t \equiv 0 \bmod 2$ by first taking as many copies of $(fw10,fw01)$ as possible, since each of these will result in fastest traversals of the covered subtrees (because fastest coverage of $fw10$ and $fw01$ arms always takes odd time), taking one more $fw10$ if possible, before taking the remaining subtrees in any order. We construct $S_1$, a similar solution with return that starts at time $t \equiv 1 \bmod 2$, by taking an $fw01$ subtree before constructing a time 0 solution in the same way as $S_0$. Since we know the form of an optimal solution, we can easily compute the cost of each of these two solutions in $O(\deg(v))$. Then, for each $i \in \{0,1\}$ and each subtree $T^u$ of $T^v$, we calculate the cost of covering $T_v$ without return starting at time $t \equiv i \bmod 2$, such that $T^u$ is the final subtree covered. We can do this by subtracting the cost of covering $T^u$ in $S_i$, adding 1 if the removal of $T^u$ from $S_i$ necessarily decreases the number of subtrees taken optimally in $S_i$, and adding the cost of taking $T^u$ without return at the end of this new solution. The foremost cost of covering $T^v$ is the cost of the minimum solution over all $T^u$. Given the classification of $T^u$, this computation takes constant time for each $T^u$, and thus $O(\deg(v))$ overall. Given these costs, it is trivial to classify $T^v$. So, we can recursively compute all required values, at a cost of $O(\deg(v)$ per node, and thus $O(n)$ overall. The optimal cost of a complete solution is then $c(T,0)$.
\end{proof}
We hypothesize that more efficient algorithms, such as the one for the $p=2$ case, exist for this type of problem for greater values of $p$, and even general $p$, via this method of piecing together fast patterns. We have similar high hopes for larger classes of underlying graphs.
\section{Open Problems and Discussion}
\label{openProbs}
This paper presents significant advances towards isolating the maximal class of graphs over which DMVP in $\mathcal{R}$ is solvable in polynomial time. We conjecture that this maximal class is the class of all graphs with polynomially many spanning trees, all of which have $O(\lg n)$ leaves. Furthermore, we conjecture that this class is equivalent for $\mathcal{R}$ and $\mathcal{B}$. But we are very interested in expanding this class with respect to $\mathcal{P}$, motivated by our solvability results for $\mathcal{P}$ over subclasses of trees. We have shown that for the case of $p=2$, DMVP for a single agent over general trees can be computed in linear time. This result relies on the fact that we know how to optimally piece together patterns with period 2. New methods for finding optimal pattern sequences could greatly reduce computation for cases of $p>2$. We are hopeful that DMVP in $\mathcal{P}$ will be shown to be poly-time solvable over arbitrary trees or at least bounded degree trees, for greater values $p$ both fixed and not fixed.
Considering $\mathcal{B}$ and $\mathcal{R}$, $\mathcal{B}$ is clearly differentiated from $\mathcal{R}$ in that we have at least some ability to approximate in $\mathcal{B}$. There remains, however, an important open question: Is there any class $C$ of underlying graphs such that DMVP is NP-hard over $C$ in $\mathcal{R}$, but not in $\mathcal{B}$? We are particularly interested in whether or not DMVP in $\mathcal{B}$ is NP-hard when the underlying graph is a star and $\Delta$ is fixed, in particular, when $\Delta = 2$. Note: The proof of Theorem \ref{npostar} implies it is hard when $\Delta$ is some relatively small function of the input. We conjecture that even for $\Delta = 2$ this problem is NP-hard, but the highly-restricted nature of the input makes an answer to this problem more elusive than some of the others we have results for. Towards an answer to this question, we give the following observation:
\begin{obs}
\label{noWaiting}
DMVP
in $\mathcal{R}$ over a spider with arms of uniform length $l$, e.g., a star (when $l = 1$), can be decided in polynomial time, when $t$ disallows waiting, i.e., $t = 2n - l - d$, where $d$ is topological distance from $s$ to $c$.
\end{obs}
\begin{proof}
Suppose $G$ is a spider with arms of uniform length $l$. Then, $G$ has $n/l$ arms. Suppose $a$ starts at some vertex $s$ distance $d$ from the central vertex $c$. If $t = 2n - l - d$, then the only solution can be a waiting-free spanning tree traversal of $G$ starting at $s$. If $d > 0$, i.e., $s \neq c$, then the first leaf visited must be the leaf of $s$'s arm. If $a$ starts at $c$, any arm can be traversed first. In either case, starting at the first time $\tau$ that $a$ finds itself at $c$, $a$ must traverse the $O(n/l) = \alpha$ remaining arms $a_1,...,a_\alpha$ each in time ${2l}$, except for the final traversed arm, whose leaf is reached in $l$ steps from $c$, completing the solution. Starting at $\tau$, break the remaining time into $\alpha-1$ length $2l$ time blocks $b_1,...,b_{\alpha-1}$, and a final length $l$ time block $b_\alpha$. For each $a_i$, for each $b_j$, we can straightforwardly compute, in $O(l)$ time, whether or not $a$ can traverse $a_j$ and return to $c$ during $b_j$ without waiting. Deciding whether or not there exists a complete traversal of all $\alpha$ arms without waiting then reduces to the problem of finding a perfect bipartite matching between arms and time blocks, for which there are many known efficient polynomial time algorithms, e.g., \cite{Hopcroft}.
\end{proof}
Overall, our results show some instances where DMVP is tractable as well as showing that DMVP faces difficult computational challenges for some natural classes of underlying topologies and dynamics. These challenges motivate research into online, multi-agent solutions to the problem, since in many cases having a complete global view of the present and future does not appear to be very helpful; moreover, in agent-oriented applications ranging from software agents to mobile robots, the information available to teams of agents can be bounded both temporally and geographically, and such online, multi-agent approaches could be well suited to agent dynamics without diminishing tractability. We have begun to take steps in this direction using edge markovian TVG models \cite{Baumann11}. In these types of stochastic environments, investigating interactive agent policies is an especially interesting direction to pursue.
|
\section{Introduction}\label{sec1}
Complementarity and entanglement are basic concepts of quantum
theory. Heisenberg's uncertainty principle \cite{heisenberg} is
one of the most known restrictions imposed in the quantum world. A
truly non-classical character of entanglement was emphasized in
the Schr\"{o}dinger ``cat paradox'' paper \cite{cat35}. Today,
quantum properties are considered as powerful resources for
potential usage in communication and computation tasks
\cite{nielsen}. Historically, uncertainty relations were focused
on pairs of canonically conjugate variables \cite{lahti,brud11}.
Recent researches have shown that uncertainty relations give a
useful tool for studying complementarity aspects \cite{ww10}.
Quantum entanglement is used as a basic tool in quantum
parallelism, quantum cryptography, quantum dense coding, and
quantum teleportation \cite{hhhh09}.
The notion of mutually unbiased bases has much many links with
recent studies of quantum information protocols (see, e.g., the
review \cite{bz10} and references therein). Mutually unbiased
bases are also an interesting mathematical subject \cite{bz10}.
For example, the problem of maximal set of mutually unbiased bases
is still open. In general, the maximal number of MUBs in $d$
dimensions is still an open question \cite{bz10}. When $d$ is a
prime power, the answer $d+1$ is known \cite{wf89}. For other $d$,
i.e., for composite numbers, we only know that the maximal number
of MUBs does not exceed $d+1$. The author of \cite{kag14} proposed
the concept of mutually unbiased measurements. The principal
result is that a complete set of $d+1$ mutually unbiased
measurements has been built explicitly for arbitrary finite $d$
\cite{kag14}. Hence, we have come across different questions
concerning a possible usage of such measurements in quantum
information science.
The aim of the present work is to study mutually unbiased
measurements in some important respects. The paper is organized as
follows. In section \ref{sec2}, preliminary facts are reviewed. We
also prove an interesting relation between R\'{e}nyi's entropies
of three different orders. In Section \ref{sec3}, we derive an
upper bound on the sum of the indices of coincidence for a set of
mutually unbiased measurements. In Section \ref{sec4}, uncertainty
relations for an arbitrary number of such measurements are derived
in terms of R\'{e}nyi's and Tsallis' entropies. Both the
state-dependent and state-independent formulations are given.
Applications of mutually unbiased measurements in entanglement
detection are examined in Section \ref{sec5}. We will see an
evidence for trade-off between a sensitivity of the scheme and
costs for its implementation. In Section \ref{sec5}, we conclude
the paper with a summary of results.
\section{Preliminaries}\label{sec2}
In this section, we review the required material. First, some
notation for spaces and operators is introduced. We then recall
the definition of mutually unbiased measurements proposed in
\cite{kag14}. Further, we discuss the R\'{e}nyi and Tsallis
entropies which will be used as measures of uncertainties. We also
prove a relation between R\'{e}nyi's entropies of three different
orders.
Due to sensitivity of quantum states, a measurement stage is one
of central questions in quantum protocols \cite{nielsen}. Hence,
some selected types of measurements are of special interest in
quantum information processing. Mutually unbiased bases are used
in quantum state reconstruction \cite{wf89}, quantum error
correction \cite{gott96,cald97}, detection of quantum entanglement
\cite{shbah12}, and the mean king's problem \cite{vaid87,ena01}.
For arbitrary $d$, however, constructing a maximal set of mutually
unbiased bases is an open problem. One may try to fit
``unbiasedness'' with weaker conditions \cite{kag14}. In this way,
we will deal with mutually unbiased measurements.
Let $\lnp(\hh)$ be the space of linear operators on
$d$-dimensional Hilbert space $\hh$. By $\lsp(\hh)$, we denote the
set of positive semi-definite operators on $\hh$. A density
operator $\bro\in\lsp(\hh)$ is normalized by $\Tr(\bro)=1$.
For operators $\ax,\ay\in{\mathcal{L}}(\hh)$, their
Hilbert--Schmidt inner product is written as \cite{watrous1}
\begin{equation}
\langle\ax{\,},\ay\rangle_{\rm{hs}}:=\Tr(\ax^{\dagger}\ay)
\ . \label{hsdef}
\end{equation}
Let $\cla=\{|\phi_{n}\rangle\}_{n=1}^{d}$ and
$\clb=\{|\varphi_{n}\rangle\}_{n=1}^{d}$ be orthonormal bases in
$d$ dimensions. They are said to be mutually unbiased, when
\begin{equation}
\bigl|\langle\phi_{m}|\varphi_{n}\rangle\bigr|=\frac{1}{\sqrt{d}}
\ , \label{twb}
\end{equation}
for all $m,n=1,\ldots,d$. The set
$\cmb=\{\clb^{(1)},\ldots,\clb^{(M)}\}$ is a set of mutually
unbiased bases (MUBs), when each two bases from this set are
mutually unbiased. When dimensionality $d$ is not a prime power,
we do not know the maximal number of MUBs that can be constructed.
The answer is not known even for $d=6$ \cite{bz10}. In some
respects, mutually unbiased bases are connected with symmetric
informationally complete measurements, shortly SIC-POVMs
\cite{bz10}. Such measurements are also not easy to construct.
The authors of \cite{kag14} proposed a concept of mutually
unbiased measurements (MUMs). Let $\ppc=\{\ppm_{n}\}$ and
$\qpc=\{\qpm_{n}\}$ be two POVM measurements, each with $d$
elements. We assume that POVM elements satisfy
\begin{align}
& \Tr(\ppm_{n})=\Tr(\qpm_{n})=1
\ , \label{tmn1}\\
& \Tr(\ppm_{m}\qpm_{n})=\frac{1}{d}
\ . \label{dmn1}
\end{align}
The following fact follows from the assumptions. The
Hilbert--Schmidt product of two elements from the same POVM can be
described in terms of a single parameter $\varkappa$ \cite{kag14}:
\begin{equation}
\Tr(\ppm_{m}\ppm_{n})=\delta_{mn}{\,}\varkappa
+(1-\delta_{mn}){\>}\frac{1-\varkappa}{d-1}
\ . \label{mjmk}
\end{equation}
General bounds on the parameter $\varkappa$ are written as
$1/d\leq\varkappa\leq1$ \cite{kag14}. The set
$\mpb=\{\ppc^{(1)},\ldots,\ppc^{(M)}\}$ is a set of MUMs of the
efficiency $\varkappa$, when each two measurements obey the above
properties. It turns out that we can reach the aim to build a
complete set of $d+1$ mutually unbiased measurements in $d$
dimensions \cite{kag14}.
Let us consider $d^{2}-1$ operators that form an orthogonal basis
in the space of traceless Hermitian operators. For instance, we
can start with the generators of ${\textup{SU}}(d)$ \cite{kag14}.
Using such operators, one can built a family of traceless
Hermitian operators $\fm_{n}^{(b)}$ with labels $b=1,\ldots,d+1$
and $n=1,\ldots,d$. The constructed operators obey the condition
\begin{equation}
\Tr{\bigl(\fm_{m}^{(a)}\fm_{n}^{(b)}\bigr)}=0
\qquad (a\neq{b})
\ . \label{anqb}
\end{equation}
For one and the same label $b$, they also satisfy \cite{kag14}
\begin{equation}
\Tr{\bigl(\fm_{m}^{(b)}\fm_{n}^{(b)}\bigr)}=\bigl(1+\sqrt{d}\bigr)^{2}
\bigl[\delta_{mn}(d-1)-(1-\delta_{mn})\bigr]
\ , \label{bnqb}
\end{equation}
where $\delta_{mn}$ is the Kronecker symbol.
An explicit construction of MUMs is written as follows
\cite{kag14}. For $b=1,\ldots,d+1$ and $n=1,\ldots,d$, we
introduce operators
\begin{equation}
\ppm_{n}^{(b)}=\frac{\pen}{d}+t{\,}\fm_{n}^{(b)}
\ , \label{pnvfn}
\end{equation}
where $t$ is some parameter that should be chosen. The least
quantities among eigenvalues of the operators $\fm_{n}^{(b)}$
determine an interval, in which $t$ can be varied \cite{kag14}.
This interval should be such that $\ppm_{n}^{(b)}\in\lsp(\hh)$ for
all values of the labels. With the given $t$, the efficiency
parameter is calculated as
\begin{equation}
\varkappa=\frac{1}{d}+t^{2}\bigl(1+\sqrt{d}\bigr)^{2}(d-1)
\ . \label{dvit}
\end{equation}
The range of $t$ leads to the corresponding range of $\varkappa$.
The measurements
$\ppc^{(b)}=\bigl\{\ppm_{n}^{(b)}\bigr\}_{n=1}^{d}$ then form a
complete set of MUMs of the efficiency (\ref{dvit}).
As measures of an uncertainty in quantum measurements, we will use
the R\'{e}nyi and Tsallis entropies. The concept of entropy is of
great importance in both information theory and statistical
physics. In addition to the Shannon entropy, other entropic
measures were found to be useful. The R\'{e}nyi and Tsallis
entropies are both very important \cite{bengtsson}. For the given
probability distribution, its R\'{e}nyi $\alpha$-entropy is
defined as \cite{renyi61}
\begin{equation}
R_{\alpha}(p):=\frac{1}{1-\alpha}{\ }\ln\left(\sum\nolimits_{n} p_{n}^{\alpha}\right)
\, , \label{rpdf}
\end{equation}
where $\alpha>0$ and $\alpha\neq1$. This quantity is a
non-increasing function of $\alpha$ \cite{renyi61}. Other
properties of parametric dependence of (\ref{rpdf}) are discussed
in \cite{zycz}. The Renyi entropy of order $\alpha=2$ is also
known as the collision entropy \cite{ww10,bengtsson}. It is
written as
\begin{equation}
R_{2}(p)={-\ln}{\left(\sum\nolimits_{n}p_{n}^{2}\right)}
\> . \label{clen}
\end{equation}
In the limit $\alpha\to\infty$, we have the so-called min-entropy
\begin{equation}
R_{\infty}(p)=-\ln(\max{p}_{n})
\ . \label{mnen}
\end{equation}
The min-entropy is of specific interest in cryptography
\cite{ngbw12}. It is also linked with the extrema of the discrete
Wigner function \cite{MWB10}. Uncertainty bounds on R\'{e}nyi's
entropies are significant in studying the connection between
complementarity and uncertainty principles \cite{bosyk13a}. Using
the R\'{e}nyi entropy, the writers of \cite{rprz12} obtained
trade-off relations for a trace-preserving quantum operation. An
extension of such trade-off relations in terms of the unified
entropies was given in \cite{rast13a}.
The notion of Tsallis entropy is widely used in non-extensive
statistical mechanics \cite{gmt}. The non-extensive entropy of
positive degree $\alpha\neq1$ is defined as \cite{tsallis}
\begin{equation}
H_{\alpha}(p):=\frac{1}{1-\alpha}{\,}
\left(\sum\nolimits_{n} p_{n}^{\alpha}-1\right)
{\,}. \label{tsent}
\end{equation}
With the factor $\left(2^{1-\alpha}-1\right)^{-1}$ instead of
$(1-\alpha)^{-1}$, this function was deduced by Havrda and
Charv\'{a}t \cite{havrda}. The entropy (\ref{tsent}) is concave
for all $\alpha>0$. In more detail, properties of the entropy
(\ref{tsent}) and related functionals are considered in
\cite{sf06,rastkyb}. It is convenient to rewrite (\ref{tsent}) as
\begin{equation}
H_{\alpha}(p)=-\sum\nolimits_{n}p_{n}^{\alpha}{\,}\ln_{\alpha}(p_{n})
=\sum\nolimits_{n}p_{n}{\>\,}{\ln_{\alpha}}{\left(\frac{1}{p_{n}}\right)}
\ . \label{tsaln}
\end{equation}
Here, we used the $\alpha$-logarithm defined for $\alpha>0\not=1$
and $x>0$ as
\begin{equation}
\ln_{\alpha}(x)=\frac{x^{1-\alpha}-1}{1-\alpha}
\ . \label{lnadf}
\end{equation}
In the limit $\alpha\to1$, we obtain $\ln_{\alpha}(x)\to\ln{x}$
and the standard Shannon entropy
\begin{equation}
H_{1}(p)=-\sum\nolimits_{n}p_{n}{\,}\ln{p}_{n}
\ . \label{shaln}
\end{equation}
Of course, the right-hand side of (\ref{rpdf}) also gives
(\ref{shaln}) in this limit. Various applications of the above
entropies and their quantum counterparts are discussed in the book
\cite{bengtsson}.
Analyzing the case of detection inefficiencies, we will use the
method of \cite{rastmub,rastpsic}. To the given value
$\eta\in[0;1]$ and probability distribution $\{p_{n}\}$, one
assigns a ``distorted'' distribution:
\begin{equation}
p_{n}^{(\eta)}=\eta{\,}p_{n}
\ , \qquad
p_{\varnothing}^{(\eta)}=1-\eta
\ . \label{dspd}
\end{equation}
Here, the parameter $\eta\in[0;1]$ describes a detector
efficiency. By $p_{\varnothing}^{(\eta)}$, we denote the
probability of the no-click event. The described model of
distorted probabilities was introduced in studying entropic Bell
inequalities with detector inefficiencies \cite{rchtf12}. Further
development of Bell inequalities with detection inefficiencies was
given in \cite{rastqqt}. As was shown in the paper \cite{rastqqt},
for all $\alpha>0$ we have
\begin{equation}
{H_{\alpha}}{\bigl(p^{(\eta)}\bigr)}=\eta^{\alpha}H_{\alpha}(p)+h_{\alpha}(\eta)
\ , \label{qtlm0}
\end{equation}
where ${H_{\alpha}}{\bigl(p^{(\eta)}\bigr)}$ denotes the entropy
of ``distorted'' distribution (\ref{dspd}). As usual, the binary
Tsallis entropy $h_{\alpha}(\eta)$ is written as
\begin{equation}
h_{\alpha}(\eta):={}-\eta^{\alpha}\ln_{\alpha}(\eta)-(1-\eta)^{\alpha}\ln_{\alpha}(1-\eta)
\ . \label{bnta}
\end{equation}
Entropic uncertainty relations with detection inefficiencies were
derived for mutually unbiased bases in \cite{rastmub} and for a
general SIC-POVM in \cite{rastpsic}. We will apply this method to
mutually unbiased measurements.
Deriving uncertainty relations in terms of the R\'{e}nyi
entropies, we will deal with the following situation. Both the
collision entropy and min-entropy can be calculated or estimated
from below. We wish to obtain a lower bound on the R\'{e}nyi
entropy of order $\alpha\geq2$. An answer to the question is
written as follows.
\newtheorem{prp21}{Proposition}
\begin{prp21}\label{pon21}
For $\alpha\in[2;\infty]$, the R\'{e}nyi $\alpha$-entropy is
bounded from below as
\begin{equation}
R_{\alpha}(p)\geq\frac{1}{\alpha-1}{\>}R_{2}(p)+\frac{\alpha-2}{\alpha-1}{\>}R_{\infty}(p)
\ . \label{al2inf}
\end{equation}
\end{prp21}
{\bf Proof.} We will deal with finite $\alpha\geq2$. We first
write the inequality
\begin{equation}
\sum\nolimits_{n} p_{n}^{\alpha}\leq
(\max{p}_{n})^{\alpha-2}{\,}\sum\nolimits_{n} p_{n}^{2}
\ . \label{ainf1}
\end{equation}
The function $x\mapsto(1-\alpha)^{-1}\ln{x}$ decreases for
$\alpha\geq2$. Combining this fact with the formulas (\ref{rpdf}),
(\ref{clen}), and (\ref{mnen}) completes the proof.
$\blacksquare$
It should be pointed out that, for $\alpha\in[2;\infty]$, the
R\'{e}nyi $\alpha$-entropy can be estimated in terms of only the
collision entropy. As was mentioned in \cite{rastmub}, for
$\alpha\geq2$ we have
\begin{equation}
R_{\alpha}(p)\geq\frac{\alpha}{2(\alpha-1)}{\>}R_{2}(p)
\ . \label{ainf2}
\end{equation}
This fact directly follows from theorem 19 of the book
\cite{hardy}. It is easy to check that the second bound
(\ref{ainf2}) cannot be stronger than (\ref{al2inf}). Subtracting
the right-hand side of (\ref{ainf2}) from the right-hand side of
(\ref{al2inf}), we obtain
\begin{equation}
\frac{\alpha-2}{2(\alpha-1)}
{\,}\Bigl(
2R_{\infty}(p)-R_{2}(p)
\Bigr)\geq0
\ . \label{2rrin}
\end{equation}
The latter is equivalent to
$(\max{p}_{n})^{2}\leq\sum_{n}p_{n}^{2}$. Taking (\ref{al2inf}), we
may generally obtain better bounds. In the following, the result
(\ref{al2inf}) will be used in deriving uncertainty bounds for MUMs
in terms of the R\'{e}nyi entropies.
\section{Indices of coincidence for MUMs}\label{sec3}
In this section, we will study indices of coincidence for
measurements considered. Let $\ppc=\{\ppm_{n}\}_{n=1}^{d}$ be a
mutually unbiased measurement in $d$-dimensional Hilbert space. If
the pre-measurement state is described by density matrix $\bro$,
then the probability on $n$-th outcome is written as
\begin{equation}
p_{n}(\ppc|\bro)=\Tr(\ppm_{n}\bro)
\ . \label{prbn}
\end{equation}
The index of coincidence is then defined as the sum of squared
probabilities, namely
\begin{equation}
C(\ppc|\bro):=\sum_{n=1}^{d} p_{n}(\ppc|\bro)^{2}
\ . \label{icdf}
\end{equation}
It seems to be natural that the sum (\ref{icdf}) can be linked to
the quantity $\Tr(\bro^{2})$. This quantity, called the purity of
$\bro$, is frequently used since it is easy to compute
\cite{bengtsson}. The purity of a quantum state gives a good
characterization of the degree of information about its
preparation \cite{adesso05}. We have the following general bounds
on (\ref{icdf}):
\begin{equation}
\frac{1}{d}\leq{C}(\ppc^{(b)}|\bro)\leq1
\ . \label{1dc1}
\end{equation}
Here, the lower bound follows from the convexity of the function
$x\mapsto{x}^{2}$. The index of coincidence has been exactly
calculated for a single SIC-POVM \cite{rastmub} and for a general
SIC-POVM \cite{rastpsic}. In both cases, resulting expression
involves purity of the measured state. To derive entropic
uncertainty relations, we wish to estimate from above the sum of
the indices of coincidence for several mutually unbiased
measurements. It is natural that our bound is formulated in terms
of the purity.
\newtheorem{prp31}[prp21]{Proposition}
\begin{prp31}\label{pon31}
Let $\mpb=\{\ppc^{(1)},\ldots,\ppc^{(M)}\}$ be a set of $M$
mutually unbiased measurements of the efficiency $\varkappa$ in
$d$ dimensions. For arbitrary $\bro$, the sum of the corresponding
indices of coincidence obey
\begin{equation}
\sum_{\ppc\in\mpb}C(\ppc|\bro)\leq\frac{M-1}{d}+\frac{1-\varkappa+(\varkappa{d}-1){\,}\Tr(\bro^{2})}{d-1}
\ . \label{ubp1}
\end{equation}
\end{prp31}
{\bf Proof.} It follows from the construction of MUMs that any
density matrix can be represented as \cite{kag14}
\begin{equation}
\bro=\frac{\pen}{d}+\sum_{b=1}^{d+1}\sum_{n=1}^{d} r_{n}^{(b)}{\,}\fm_{n}^{(b)}
\ . \label{bsrp}
\end{equation}
Since the operators $\fm_{n}^{(b)}$ are all traceless and obey
(\ref{anqb}), we have
\begin{equation}
\Tr(\bro^{2})=\frac{1}{d}+\sum_{b=1}^{d+1}\sum_{m,n=1}^{d}
r_{m}^{(b)}r_{n}^{(b)}{\,}\Tr{\bigl(\fm_{m}^{(b)}\fm_{n}^{(b)}\bigr)}
\ . \label{tbr2}
\end{equation}
By (\ref{bnqb}), we have
$\Tr{\bigl(\fm_{m}^{(b)}\fm_{n}^{(b)}\bigr)}=-\bigl(1+\sqrt{d}\bigr)^{2}$
for $m\neq{n}$ and
\begin{equation}
\Tr{\bigl(\fm_{n}^{(b)}\fm_{n}^{(b)}\bigr)}=\bigl(1+\sqrt{d}\bigr)^{2}(d-1)
\ . \label{fnfn}
\end{equation}
Using these formulas, we obtain
\begin{align}
&\sum_{m,n=1}^{d}
r_{m}^{(b)}r_{n}^{(b)}{\,}\Tr{\bigl(\fm_{m}^{(b)}\fm_{n}^{(b)}\bigr)}
\nonumber\\
&=\bigl(1+\sqrt{d}\bigr)^{2}(d-1)\sum_{n=1}^{d}r_{n}^{(b)}r_{n}^{(b)}
-\bigl(1+\sqrt{d}\bigr)^{2}\sum_{\substack{m,n=1 \\ m\neq{n}}}^{d}r_{m}^{(b)}r_{n}^{(b)}
\nonumber\\
&=\bigl(1+\sqrt{d}\bigr)^{2}
\left(d\sum_{n=1}^{d}{r_{n}^{(b)}r_{n}^{(b)}}
-R^{(b)}R^{(b)}
\right)
{\,}, \label{srmrn}
\end{align}
where $R^{(b)}:=\sum_{n=1}^{d}r_{n}^{(b)}$. Substituting
(\ref{srmrn}) into (\ref{tbr2}), we obtain an expression for
$\Tr(\bro^{2})$.
Combining (\ref{pnvfn}) with (\ref{bsrp}), the probability of
$n$-th outcome in $b$-th measurement reads
\begin{align}
\frac{1}{d}+t{\,}\Tr{\bigl(\bro{\,}\fm_{n}^{(b)}\bigr)}
&=\frac{1}{d}+t\sum_{m=1}^{d}
r_{m}^{(b)}{\,}\Tr{\bigl(\fm_{m}^{(b)}\fm_{n}^{(b)}\bigr)}
\nonumber\\
&=\frac{1}{d}+t\bigl(1+\sqrt{d}\bigr)^{2}
\bigl(d{\,}r_{n}^{(b)}-R^{(b)}
\bigr)
{\,}. \label{pnbp}
\end{align}
The last expression is obtained similarly to (\ref{srmrn}).
Squaring this probability and further summing with respect to
$n=1,\ldots,d$, one gets
\begin{equation}
C(\ppc^{(b)}|\bro)=\frac{1}{d}+t^{2}\bigl(1+\sqrt{d}\bigr)^{4}
\left(d^{2}\sum_{n=1}^{d}{r_{n}^{(b)}r_{n}^{(b)}}
-d{\,}R^{(b)}R^{(b)}
\right)
{\,}. \label{cbbp}
\end{equation}
Here, we used
$\sum_{n=1}^{d}{\bigl(d{\,}r_{n}^{(b)}-R^{(b)}\bigr)}=0$. By
(\ref{1dc1}), we see that, for all $b$, the second term in the
right-hand side of (\ref{cbbp}) is non-negative.
Combining (\ref{tbr2}), (\ref{srmrn}), and (\ref{cbbp}), we
further write
\begin{align}
\sum_{b=1}^{M}C(\ppc^{(b)}|\bro)
&=\frac{M}{d}+t^{2}d\bigl(1+\sqrt{d}\bigr)^{4}\sum_{b=1}^{M}\left(d\sum_{n=1}^{d}{r_{n}^{(b)}r_{n}^{(b)}}
-R^{(b)}R^{(b)}
\right)
\nonumber\\
&\leq\frac{M}{d}+t^{2}d\bigl(1+\sqrt{d}\bigr)^{2}
\left(\Tr(\bro^{2})-\frac{1}{d}\right)
\nonumber\\
&=\frac{M-1}{d}+\varkappa+\frac{\varkappa{d}-1}{d-1}
\left(\Tr(\bro^{2})-1\right)
\, . \label{csmp}
\end{align}
At the last step, we used (\ref{dvit}). The quantity (\ref{csmp})
is easily reduced to the right-hand side of (\ref{ubp1}).
$\blacksquare$
The statement of Proposition \ref{pon31} provides an upper bound
on the sum of the indices of coincidence for a set of MUMs. For
the complete set of $d+1$ MUMs, we actually have an exact result
instead of inequality:
\begin{equation}
\sum_{b=1}^{d+1}C(\ppc^{(b)}|\bro)
=1+\frac{1-\varkappa+(\varkappa{d}-1){\,}\Tr(\bro^{2})}{d-1}
\ . \label{ubpd1}
\end{equation}
Indeed, the inequality (\ref{csmp}) is saturated with $M=d+1$. For
pure states, the right-hand side of (\ref{ubpd1}) becomes
$1+\varkappa$. This result was presented in \cite{kag14} and then
used in the context of entanglement detection in \cite{fei14}. The
inequality (\ref{ubp1}) is tight in the sense that it is always
saturated with the completely mixed state $\bro_{*}=\pen/d$. Since
operators $\fm_{n}^{(b)}$ are all traceless, we see from
(\ref{pnbp}) that
\begin{equation}
p_{n}(\ppc^{(b)}|\bro_{*})=\frac{1}{d}
\ , \label{pnras0}
\end{equation}
irrespectively to $n$ and $b$. For each $b=1,\ldots,d+1$,
therefore, the index of coincidence reads
\begin{equation}
C(\ppc^{(b)}|\bro_{*})=\sum_{n=1}^{d}\frac{1}{d^{2}}=\frac{1}{d}
\ . \label{pnras1}
\end{equation}
Hence, the left-hand side of (\ref{ubp1}) is equal to $M/d$ for
the completely mixed state $\bro_{*}$. By substitution
$\Tr(\bro_{*}^{2})=1/d$, the right-hand side of (\ref{ubp1}) gives
$M/d$ as well. Thus, our result is almost precise for impure
states with the purity close to $1/d$. The bound (\ref{ubp1}) may
also be saturated with pure states. Below, we will shortly mention
an example for MUBs. Note that the purity $\Tr(\bro^{2})$ can be
expressed in terms of the Bloch vector of $\bro$ \cite{rastpsic}.
Hence, the formulas (\ref{ubp1}) and (\ref{ubpd1}) can be
rewritten via the Bloch vector as well. We refrain from presenting
the details here.
It should be noticed that the results (\ref{ubp1}) and
(\ref{ubpd1}) are calculated for the aforementioned construction
of mutually unbiased measurements. Setting $\varkappa=1$,
nevertheless, the inequality (\ref{ubp1}) leads to the correct
result for mutually unbiased bases. As was shown in \cite{molm09},
for a set $\cmb=\{\clb^{(1)},\ldots,\clb^{(M)}\}$ of $M$ mutually
unbiased bases we have
\begin{equation}
\sum_{\clb\in\cmb} C(\clb|\bro)\leq\frac{M-1}{d}+\Tr(\bro^{2})
\ . \label{bap1}
\end{equation}
We consider $\bro=|\psi\rangle\langle\psi|$ with
$|\psi\rangle$ taken from one of the bases
$\clb^{(1)},\ldots,\clb^{(M)}$. In this case, the inequality
(\ref{bap1}) is actually saturated. Indeed, one the indices
$C(\clb|\bro)$ is then equal to $1$ and other are all $1/d$. The
sum of these indices is equal to the right-hand side of
(\ref{bap1}) for $\Tr(\bro^{2})=1$. Using (\ref{bap1}),
uncertainty relations in terms of the Shannon entropies have been
obtained \cite{molm09}. Extensions with the use of generalized
entropies were derived in \cite{rastmub}. The authors of
\cite{shbah12} considered applications of the bound (\ref{bap1})
in the context of entanglement detection. In the following, we
will use (\ref{ubp1}) for obtaining entropic bounds for an
arbitrary set of MUMs. We will also revisit applications of MUMs
in entanglement detection.
\section{Entropic uncertainty relations for MUMs}\label{sec4}
In this section, we present uncertainty relations for an arbitrary
set of mutually unbiased measurements. First, we obtain
uncertainty relations in terms of the R\'{e}nyi entropies. Second,
we give formulations in terms of the Tsallis $\alpha$-entropies of
order $\alpha\in(0;2]$. Both the state-dependent and
state-independent formulations are considered. In the Tsallis
case, we also address uncertainty relations with detection
inefficiencies. Our first result is posed as follows.
\newtheorem{prp41}[prp21]{Proposition}
\begin{prp41}\label{pon41}
Let $\mpb=\{\ppc^{(1)},\ldots,\ppc^{(M)}\}$ be a set of $M$
mutually unbiased measurements of the efficiency $\varkappa$ in
$d$ dimensions. For $\alpha\in[2;\infty]$ and arbitrary density
matrix $\bro$ on $\hh$, the averaged sum of R\'{e}nyi's entropies satisfies
the state-dependent bound
\begin{align}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{\alpha}(\ppc|\bro)
&\geq\frac{1}{\alpha-1}{\>}
{\ln}{\left(\frac{Md(d-1)}{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}\right)}
\label{rum01}\\
&+\frac{\alpha-2}{\alpha-1}
{\,}\biggl\{
\ln{d}-{\ln}{\left(1+M^{-1/2}\sqrt{\varkappa{d}-1}{\,}\sqrt{\Tr(\bro^{2})d-1}\right)}
\biggr\}
{\>}. \nonumber
\end{align}
\end{prp41}
{\bf Proof.} First, we will obtain a lower bound on the averaged sum of
collision entropies. Since the function $x\mapsto-\ln{x}$ is
convex, we write
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{2}(\ppc|\bro)=
\sum_{\ppc\in\mpb}\frac{1}{M}{\,}\Bigl(-\ln{C}(\ppc|\bro)\Bigr)
\geq{-\ln}{\left(\frac{1}{M}{\,}\sum_{\ppc\in\mpb}C(\ppc|\bro)\right)}
{\>}. \label{cov2}
\end{equation}
It follows from (\ref{ubp1}) that
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb}C(\ppc|\bro)\leq
\frac{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}{Md(d-1)}
\ . \label{ubpm1}
\end{equation}
As the function $x\mapsto-\ln{x}$ decreases, combining
(\ref{cov2}) with (\ref{ubpm1}) leads to the result
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{2}(\ppc|\bro)\geq
{\ln}{\left(\frac{Md(d-1)}{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}\right)}
{\>}. \label{cov22}
\end{equation}
The second step is to get a lower bound on the averaged sum of
min-entropies. It follows from lemma 3 of \cite{rastmub} that
\begin{equation}
\max\bigl\{p_{n}:{\,}1\leq{n}\leq{d}\bigr\}
\leq\frac{1}{d}\left(1+\sqrt{d-1}{\,}\sqrt{C(p)d-1}\right)
{\,}, \label{aust}
\end{equation}
where $C(p)$ is the index of coincidence. For clarity, we
introduce the function
\begin{equation}
g_{d}(x):=\frac{1}{d}{\>}\Bigl(1+\sqrt{d-1}{\,}\sqrt{xd-1}\Bigr)
{\>}. \label{gdxf}
\end{equation}
This function is concave and increasing. Combining these facts
with (\ref{aust}) and (\ref{ubpm1}), we obtain
\begin{align}
&\frac{1}{M}{\,}\sum_{\ppc\in\mpb}
\underset{n}{\max}{\,}p_{n}(\ppc|\bro)
\leq
\sum_{\ppc\in\mpb}\frac{1}{M}{\>\,}{g_{d}}{\bigl(C(\ppc|\bro)\bigr)}
\leq
g_{d}{\left(\frac{1}{M}{\,}\sum_{\ppc\in\mpb} C(\ppc|\bro)\right)}
\nonumber\\
&\leq{g}_{d}{\left(
\frac{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}{Md(d-1)}
\right)}
{\>}. \label{grp2}
\end{align}
Calculating the term (\ref{grp2}) in line with the definition
(\ref{gdxf}) leads to
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb}
\underset{n}{\max}{\,}p_{n}(\ppc|\bro)
\leq\frac{1}{d}{\,}
\left(1+M^{-1/2}\sqrt{\varkappa{d}-1}{\,}\sqrt{\Tr(\bro^{2})d-1}\right)
\, . \label{res1}
\end{equation}
Due to convexity of the function $x\mapsto-\ln{x}$, we further
write
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{\infty}(\ppc|\bro)\geq
{-\ln}{\left(
\frac{1}{M}{\,}\sum_{\ppc\in\mpb}
\underset{n}{\max}{\,}p_{n}(\ppc|\bro)
\right)}
\> . \label{res2}
\end{equation}
Since this function decreases, combining (\ref{res1}) with
(\ref{res2}) finally gives
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{\infty}(\ppc|\bro)\geq
\ln{d}-{\ln}{\left(1+M^{-1/2}\sqrt{\varkappa{d}-1}{\,}\sqrt{\Tr(\bro^{2})d-1}\right)}
\> . \label{res3}
\end{equation}
Using both the bounds (\ref{cov22}) and (\ref{res3}), we complete
the proof of (\ref{rum01}) due to (\ref{al2inf}).
$\blacksquare$
The statement of Proposition \ref{pon41} gives a state-dependent
lower bound on the sum of corresponding R\'{e}nyi's entropies. In
the case of MUBs, we have $\varkappa=1$. Let
$\cmb=\{\clb^{(1)},\ldots,\clb^{(M)}\}$ be a set of $M$ mutually
unbiased bases. For $\alpha\in[2;\infty]$, there holds
\begin{align}
\frac{1}{M}{\,}\sum_{\clb\in\cmb} R_{\alpha}(\clb|\bro)
&\geq\frac{1}{\alpha-1}{\>}
{\ln}{\left(\frac{Md}{M+\Tr(\bro^{2})d-1}\right)}
\label{rum01b}\\
&+\frac{\alpha-2}{\alpha-1}
{\,}\biggl\{
\ln{d}-{\ln}{\left(1+M^{-1/2}\sqrt{d-1}{\,}\sqrt{\Tr(\bro^{2})d-1}\right)}
\biggr\}
{\>}. \nonumber
\end{align}
For internal points of the interval $\alpha\in[2;\infty]$, this
result gives an improvement of the uncertainty relations of
\cite{rastmub}.
The bound (\ref{rum01}) is tight in the following sense. For
arbitrary $\alpha\in[2;\infty]$, this inequality is certainly
saturated with the completely mixed state. Substituting
$\Tr(\bro_{*}^{2})=1/d$, for $\alpha\geq2$ the relation
(\ref{rum01}) finally gives
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{\alpha}(\ppc|\bro_{*})\geq
\frac{1}{\alpha-1}{\>}\ln{d}+\frac{\alpha-2}{\alpha-1}{\>}\ln{d}=\ln{d}
\ . \label{exrp}
\end{equation}
Since the distribution (\ref{pnras0}) is uniform, we have
$R_{\alpha}(\ppc|\bro_{*})=\ln{d}$ for all $\alpha>0$. Thus, the
lower bound is actually saturated.
In the case of pure states, we obtain a state-independent lower
bound. If $\bro=|\psi\rangle\langle\psi|$, then we have
$\Tr(\bro^{2})=1$. Substituting this into the right-hand side of
(\ref{rum01}), one gets
\begin{align}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} R_{\alpha}(\ppc|\psi)
&\geq\frac{1}{\alpha-1}{\>}
{\ln}{\left(\frac{Md}{M+\varkappa{d}-1}\right)}
\label{rum01p}\\
&+\frac{\alpha-2}{\alpha-1}
{\,}\biggl\{
\ln{d}-{\ln}{\left(1+M^{-1/2}\sqrt{\varkappa{d}-1}\sqrt{d-1}\right)}
\biggr\}
{\>}. \nonumber
\end{align}
Of course, this lower bound is also valid for all mixed states.
For impure $\bro$, the lower bound (\ref{rum01}) is stronger than
(\ref{rum01p}) due to $\Tr(\bro^{2})<1$. That is, the right-hand
side of (\ref{rum01}) increases with a deviation of the purity
from $1$. Dependence of such a kind seems to be natural.
The bound (\ref{rum01}) covers the interval $\alpha\in[2;\infty]$.
For the interval $\alpha\in(0;2]$, we have a lower bound
independent of $\alpha$. Recall that the R\'{e}nyi entropy cannot
increase with growth of $\alpha$. When $0<\alpha<2$, the formula
(\ref{cov22}) remains valid after replacing $R_{2}(\ppc|\bro)$
with $R_{\alpha}(\ppc|\bro)$. In particular, we obtain lower
bounds on the sum of Shannon entropies. For a set $\mpb$ of $M$
MUMs of the efficiency $\varkappa$, we have
\begin{align}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} H_{1}(\ppc|\bro)
&\geq
{\ln}{\left(\frac{Md(d-1)}{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}\right)}
\nonumber\\
&\geq{\ln}{\left(\frac{Md}{M+\varkappa{d}-1}\right)}
{\>}. \label{smum02}
\end{align}
Substituting $M=d+1$, the latter formula gives the
state-independent relation derived in \cite{kag14}. Thus, we have
extended this result in the following three directions: (i) our
bounds hold for any set of MUMs; (ii) they are written in terms of
generalized entropies; (iii) they are state-dependent. Let us
proceed to the Tsallis formulation.
\newtheorem{prp42}[prp21]{Proposition}
\begin{prp42}\label{pon42}
Let $\mpb=\{\ppc^{(1)},\ldots,\ppc^{(M)}\}$ be a set of $M$
mutually unbiased measurements of the efficiency $\varkappa$ in
$d$ dimensions. For $\alpha\in(0;2]$ and arbitrary density matrix
$\bro$ on $\hh$, the averaged sum of Tsallis' entropies satisfies the
state-dependent bound
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} H_{\alpha}(\ppc|\bro)\geq
{\ln_{\alpha}}{\left(\frac{Md(d-1)}{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}\right)}
{\>}. \label{mum02}
\end{equation}
\end{prp42}
{\bf Proof.} We will use the following fact. For $\alpha\in(0;2]$
and arbitrary probability distribution, the Tsallis
$\alpha$-entropy obeys \cite{rastmub}
\begin{equation}
H_{\alpha}(p)\geq{\ln_{\alpha}}{\left(\frac{1}{C(p)}\right)}
{\,}. \label{cnv1}
\end{equation}
The inequality (\ref{cnv1}) is based on the fact that the function
$x\mapsto\ln_{\alpha}(1/x)$ is convex for $\alpha\in(0;2]$.
Applying Jensen's inequality to this convex function, one gets
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} H_{\alpha}(\ppc|\bro)\geq
\sum_{\ppc\in\mpb}\frac{1}{M}{\>}{\ln_{\alpha}}{\left(\frac{1}{C(\ppc|\bro)}\right)}
\geq{\ln_{\alpha}}{\left\{\biggl(\frac{1}{M}{\,}\sum_{\ppc\in\mpb}C(\ppc|\bro)\biggr)^{{\!}-1}\right\}}
{\>}. \label{cnv2}
\end{equation}
From $x\leq{y}$ we have $\ln_{\alpha}(1/x)\geq\ln_{\alpha}(1/y)$,
since the function $x\mapsto\ln_{\alpha}(1/x)$ is decreasing.
Combining (\ref{cnv2}) with (\ref{ubpm1}) completes the proof.
$\blacksquare$
The statement of Proposition \ref{pon42} gives a state-dependent
lower bound on the average Tsallis entropy for a set of $M$ MUMs.
Setting $\varkappa=1$, the bound (\ref{mum02}) is reduced to the
uncertainty relation derived in \cite{rastmub} for mutually
unbiased bases. For a pure state $\bro=|\psi\rangle\langle\psi|$,
its purity is equal to $1$. Then the bound (\ref{mum02}) reads
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} H_{\alpha}(\ppc|\psi)\geq
{\ln_{\alpha}}{\left(\frac{Md}{M+\varkappa{d}-1}\right)}
{\>}. \label{mum02pur}
\end{equation}
It is easy to see that the state-independent lower bound
(\ref{mum02pur}) remain valid for all states. Similarly to
(\ref{rum01}), the right-hand side of (\ref{mum02}) increases as
the purity decreases. When $\alpha=1$, the formulas (\ref{mum02})
and (\ref{mum02pur}) also lead to (\ref{smum02}).
Let us consider uncertainty relations with detection
inefficiencies. By the parameter $\eta\in[0;1]$, we characterize
an efficiency of used detectors. The maximum $\eta=1$ corresponds
to the inefficiency-free case. We will assume that, for any MUM,
the inefficiency-free distribution is distorted according to
(\ref{dspd}). In other words, for all $\ppc\in\mpb$ we write
\begin{equation}
p_{n}^{(\eta)}(\ppc|\bro)=\eta{\,}p_{n}(\ppc|\bro)
\ , \qquad
p_{\varnothing}^{(\eta)}(\ppc|\bro)=1-\eta
\ . \label{dsmd1}
\end{equation}
By $H_{\alpha}^{(\eta)}(\ppc|\bro)$, we mean the $\alpha$-entropy
calculated for (\ref{dsmd1}). The theoretical value
$H_{\alpha}(\ppc|\bro)$ is related to the case of
inefficiency-free implementation of measurements. Using
(\ref{qtlm0}) and (\ref{mum02}), we obtain
\begin{equation}
\frac{1}{M}{\,}\sum_{\ppc\in\mpb} H_{\alpha}^{(\eta)}(\ppc|\bro)\geq
\eta^{\alpha}{\>}{\ln_{\alpha}}{\left(\frac{Md(d-1)}{M(d-1)+(\varkappa{d}-1)\bigl(\Tr(\bro^{2})d-1\bigr)}\right)}
+h_{\alpha}(\eta)
\ , \label{mum2et}
\end{equation}
where $\alpha\in(0;2]$ and $\mpb$ is a set of $M$ MUMs of the
efficiency $\varkappa$. The result (\ref{mum2et}) is an entropic
uncertainty relation in the model of detection inefficiencies. We
see that the inefficiency-free bound (\ref{mum02}) is multiplied
by the factor $\eta^{\alpha}$ and also added by the binary entropy
$h_{\alpha}(\eta)$. Thus, additional uncertainties are induced by
non-ideal detectors \cite{rastmub,rastpsic}. By setting
$\alpha=1$, the relation (\ref{mum2et}) gives a lower bound on the
sum of the Shannon entropies.
In their usual form, uncertainty relations are not directly
applicable to study cryptographic security. To fill such a gap,
entropic uncertainty relations in the presence of quantum memory
should be considered \cite{renes10,coles12,piani14}. In the case
of two measurements, entropic uncertainty relations can be based
on majorization techniques. This important approach has been
studied in recent works \cite{prz13,fgg13,rpz14}. In certain
cases, the majorization approach has allowed to improve previous
bounds. It would be interesting to apply majorization techniques
to mutually unbiased measurement and compare resulting bounds with
the above one. We hope to address this question in future
investigations.
\section{Entanglement detection with arbitrary set of MUMs}\label{sec5}
In this section, we address a problem of entanglement detection.
Applications of mutually unbiased bases in this question were
analyzed in \cite{shbah12}. The entanglement detection via
SIC-POVMs has shortly been discussed in \cite{rastmub}. The use
of a general SIC-POVM for such purposes was discussed in
\cite{lifei14}. The authors of \cite{fei14} extended some results
of \cite{shbah12} with the use of mutually unbiased measurements.
More separability criteria are discussed in \cite{zhsl98,zhao14}.
Note that the scheme of \cite{fei14} is very particular in the
sense that it is based on using only complete sets of MUMs. Hence,
this scheme needs $d(d+1)$ local POVM elements. At the same time,
the method based on a SIC-POVM uses $d^{2}$ POVM elements
\cite{lifei14}. Here, we deal with a number of measurement
operators to be performed. From the viewpoint of implementation,
this number may be treated as an experimental complexity of the
given scheme of entanglement detection. We will show that
implementation costs of entanglement detection with MUMs can be
reduced essentially.
We now consider a bipartite system of two $d$-dimensional
subsystems. Its Hilbert space is the product
$\hh_{AB}=\hh_{A}\otimes\hh_{B}$ of two isomorphic spaces
$\hh_{A}$ and $\hh_{B}$. Let us choose the orthonormal basis
$\bigl\{|i_{S}\rangle\bigr\}$, where $S=A,B$, for each of the two
spaces $\hh_{A}$ and $\hh_{B}$. A maximally entangled pure state
is then expressed as
\begin{equation}
|\Phi_{AB}^{+}\rangle=\frac{1}{\sqrt{d}}{\,}\sum_{i=1}^{d} |i_{A}\rangle\otimes|i_{B}\rangle
\ . \label{mest}
\end{equation}
Entangled states are a basic resource in quantum information
science. Hence, the problem of efficient detection of entanglement
is of great importance \cite{hhhh09,mhph1999}.
Let us recall shortly basic notions. A product state is any state
of the form $\bro_{A}\otimes\bro_{B}$ \cite{bengtsson}. When both
the matrices $\bro_{A}$ and $\bro_{B}$ are rank-one, we have a
pure product state. A bipartite mixed state is called separable,
when its density matrix $\wbro_{AB}$ can be represented as a
convex combination of product states \cite{zhsl98}. That is, there
exist a probability distribution $\{q_{k}\}$ and two sets
$\{\bro_{A}^{(k)}\}$ and $\{\bro_{B}^{(k)}\}$ such that
\begin{equation}
\wbro_{AB}=\sum\nolimits_{k} q_{k}{\>}\bro_{A}^{(k)}\otimes\bro_{B}^{(k)}
\ . \label{sepsdf}
\end{equation}
Note that each separable state can also be expressed as a convex
combination of only pure product states. This fact easily follows
from (\ref{sepsdf}) by substitution of the corresponding spectral
decompositions. When representations of the form (\ref{sepsdf}) is
not possible, the state is called entangled \cite{zhsl98}.
To detect entanglement, we aim to use a collection of local
measurements. Let
$\mpb_{A}=\{\ppc_{A}^{(1)},\ldots,\ppc_{A}^{(M)}\}$ and
$\mqb_{B}=\{\qpc_{B}^{(1)},\ldots,\qpc_{B}^{(M)}\}$ be two sets of
$M$ MUMs. In $b$-th joint measurement, the pair $(m,n)$ of local
outcomes occurs with the probability
\begin{equation}
P^{(b)}(m,n)=
{\Tr}{\left(\bigl(\ppm_{m}^{(b)}\otimes\qpm_{n}^{(b)}\bigr)\wbro_{AB}\right)}
\, . \label{plmn}
\end{equation}
Following the idea of \cite{shbah12}, we introduce a quantity
\begin{equation}
J_{M}(\wbro_{AB})=\sum_{b=1}^{M}\sum_{n=1}^{d}
P^{(b)}(n,n)
\ . \label{jmnbn}
\end{equation}
In the case of mutually unbiased bases, this correlation measure
was proposed in \cite{shbah12}. The definition (\ref{jmnbn}) is an
immediate extension to MUMs.
\newtheorem{prp51}[prp21]{Proposition}
\begin{prp51}\label{pon51}
Let $\mpb_{A}$ be a set of $M$ mutually unbiased measurements of
the efficiency $\varkappa_{A}$ in the $d$-dimensional space
$\hh_{A}$. Let $\mqb_{B}$ be a set of $M$ mutually unbiased
measurements of the efficiency $\varkappa_{B}$ in the
$d$-dimensional space $\hh_{B}$. For all density matrices
$\bro_{A}\in\lsp(\hh_{A})$ and $\bro_{B}\in\lsp(\hh_{B})$, the
correlation measure satisfies
\begin{equation}
J_{M}(\bro_{A}\otimes\bro_{B})\leq
\prod_{S=A,B}
{\left(\frac{M(d-1)+(\varkappa_{S}d-1)\bigl(\Tr(\bro_{S}^{2})d-1\bigr)}{d(d-1)}\right)}^{1/2}
\, . \label{jrrab}
\end{equation}
\end{prp51}
{\bf Proof.} For a product state
$\wbro_{AB}=\bro_{A}\otimes\bro_{B}$, we clearly have
\begin{equation}
P^{(b)}(n,n)=
{p_{n}}{\bigl(\ppc_{A}^{(b)}|\bro_{A}\bigr)}{\,}
{p_{n}}{\bigl(\qpc_{B}^{(b)}|\bro_{B}\bigr)}
\>. \label{pnnab}
\end{equation}
Using equation (\ref{pnnab}) and the Cauchy--Schwarz inequality,
we then obtain
\begin{align}
J_{M}(\bro_{A}\otimes\bro_{B})&=
\sum_{b=1}^{M}\sum_{n=1}^{d}
{p_{n}}{\bigl(\ppc_{A}^{(b)}|\bro_{A}\bigr)}{\,}
{p_{n}}{\bigl(\qpc_{B}^{(b)}|\bro_{B}\bigr)}
\nonumber\\
&\leq
{\left(
\sum_{b=1}^{M}{C}{\bigl(\ppc_{A}^{(b)}|\bro_{A}\bigr)}
\right)^{1/2}}
{\left(
\sum_{b=1}^{M}{C}{\bigl(\qpc_{B}^{(b)}|\bro_{B}\bigr)}
\right)^{1/2}}
\, . \label{cnab}
\end{align}
Combining the latter with (\ref{ubp1}) completes the proof.
$\blacksquare$
The statement of Proposition \ref{pon51} leads to a necessary
criterion that the given bipartite state is a product state.
Applying this criterion for the given input $\wbro_{AB}$, we
should treat $\bro_{A}$ and $\bro_{B}$ as the reduced density
matrices. As usual, they are obtained by the partial trace
operation:
\begin{equation}
\bro_{A}=\Tr_{B}(\wbro_{AB})
\ , \qquad
\bro_{B}=\Tr_{A}(\wbro_{AB})
\ . \label{rarb}
\end{equation}
Relations between some norms of operators before and after partial
trace with applications to quantum entropies were obtained in
\cite{rastjst12}. To use the formula (\ref{jrrab}), we need only
purities of these density matrices. Substituting the purities in
the right-hand side of (\ref{jrrab}), we should then compare the
result with the actual value $J_{M}(\wbro_{AB})$. This value is
calculated from the measurement statistics. If the condition
(\ref{jrrab}) is violated, then the input $\wbro_{AB}$ is
certainly not a product state.
It is of interest to adopt (\ref{jrrab}) for the case, when
purities of the reduced density matrices are unknown. If we keep
at least purity of the state $\wbro_{AB}$ {\it per se}, then a
state-dependent form of the criterion can still be given. If the
input $\wbro_{AB}$ is a product state then
\begin{equation}
\Tr(\bro_{A}^{2}){\,}\Tr(\bro_{B}^{2})=\Tr(\wbro_{AB}^{2})
\ , \qquad
\Tr(\bro_{A}^{2})+\Tr(\bro_{B}^{2})\leq1+\Tr(\wbro_{AB}^{2})
\ . \label{arrb}
\end{equation}
The second relation is proved as follows. If real numbers
$x,y\in[0;1]$ are connected as $xy=a$, then $x+y\leq1+a$, by
convexity of the function $x\mapsto{x}+a/x$. Combining
(\ref{jrrab}) with (\ref{arrb}) leads to the following statement.
If the given density matrix $\wbro_{AB}$ is a product then
\begin{equation}
J_{M}(\wbro_{AB})\leq
\frac{1}{d(d-1)}{\>}
\sqrt{\Gamma^{2}+(\varkappa{d}-1)\bigl[\Gamma{d}
+(M+\varkappa{d}-1)(d-1)d{\,}\Tr(\wbro_{AB}^{2})\bigr]}
\ , \label{jabr1}
\end{equation}
where $\Gamma=M(d-1)-(\varkappa{d}-1)$. The formula (\ref{jabr1})
gives a necessary criterion, which depends on purity of the tested
bipartite state. As the violation of (\ref{jabr1}) has been
observed, we truly conclude that the tested state is not a product
state. Finally, we present the following state-independent bound.
\newtheorem{prp52}[prp21]{Proposition}
\begin{prp52}\label{pon52}
Let $\mpb_{A}$ be a set of $M$ mutually unbiased measurements of
the efficiency $\varkappa_{A}$ in the $d$-dimensional space
$\hh_{A}$. Let $\mqb_{B}$ be a set of $M$ mutually unbiased
measurements of the efficiency $\varkappa_{B}$ in the
$d$-dimensional space $\hh_{B}$. If the given state
$\wbro_{AB}$ is separable then
\begin{equation}
J_{M}(\wbro_{AB})\leq
\frac{1}{d}{\,}\sqrt{M+\varkappa_{A}d-1}{\,}\sqrt{M+\varkappa_{B}d-1}
\ . \label{jabr52}
\end{equation}
\end{prp52}
{\bf Proof.} We first observe the following. For any product
state, the quantity $J_{M}(\bro_{A}\otimes\bro_{B})$ is bounded
from above by the right-hand side of (\ref{jabr52}). This claim
directly follows from (\ref{jrrab}) and $\Tr(\bro_{S}^{2})\leq1$.
As the function $x\mapsto{x}^{2}$ is convex, we further have
\begin{equation}
J_{M}(\wbro_{AB})\leq\sum\nolimits_{k} q_{k}{\>}{J_{M}}{\bigl(\bro_{A}^{(k)}\otimes\bro_{B}^{(k)}\bigr)}
\> . \label{lkkab}
\end{equation}
Combining (\ref{lkkab}) with the above observation completes the
proof due to the condition $\sum_{k}q_{k}=1$. $\blacksquare$
The statement of Proposition \ref{pon52} gives a necessary
criterion for separability of bipartite states. In the case
$\varkappa_{A}=\varkappa_{B}=\varkappa$, this criterion reads
\begin{equation}
J_{M}(\wbro_{AB})\leq
\frac{M+\varkappa{d}-1}{d}
\ . \label{jabr2}
\end{equation}
With $M=d+1$, we deal with the scheme
using two complete sets of MUMs. Then the formula (\ref{jabr2}) is
reduced to $J_{M}(\wbro_{AB})\leq1+\varkappa$. The
latter necessary criterion for the separability was discussed in \cite{fei14}. We have seen that
entanglement detection may be proceeded with a lesser number of
measurements. This approach could be easy for implementation. On
the other hand, a sensitivity of the scheme will probably
decrease. In the following, we give a reason for trade-off between
a sensitivity of the scheme and its experimental complexity.
In general, the above upper bounds give only a necessary criterion
for the separability of bipartite states. For some classes of
states, however, this criterion may be sufficient as well. We
shall now illustrate this fact with isotropic states. Recall that
isotropic states are states of the form
\begin{equation}
\wbro_{{iso}}
=\gamma{\,}|\Phi_{AB}^{+}\rangle\langle\Phi_{AB}^{+}|+(1-\gamma){\,}\wbro_{AB*}
\ . \label{isost}
\end{equation}
Here, $\gamma\in[0;1]$ and $\wbro_{AB*}$ is the completely mixed
state on $\hh_{A}\otimes\hh_{B}$, namely
\begin{equation}
\wbro_{AB*}=\frac{\pen_{A}\otimes\pen_{B}}{d^{2}}
\ . \label{cmsab}
\end{equation}
To the given MUM $\ppc_{A}^{(b)}=\{\ppm_{n}\}$ of the efficiency
$\varkappa$, we assign a set $\ppc_{B}^{(b)*}=\{\ppm_{n}^{*}\}$ of
operators that are conjugate in the following sense. For all
$i,j=1,\ldots,d$, matrix elements obey
\begin{equation}
\langle{i}_{B}|{\,}\ppm_{n}^{*}|j_{B}\rangle=
\langle{j}_{A}|{\,}\ppm_{n}|i_{A}\rangle
\ . \label{cjpas}
\end{equation}
It is easy to check that the set $\ppc_{B}^{(b)*}$ is also a MUM
of the efficiency $\varkappa$. Using the property (\ref{cjpas}),
we obtain
\begin{equation}
\langle\Phi_{AB}^{+}|{\,}\ppm_{n}\otimes\ppm_{n}^{*}|\Phi_{AB}^{+}\rangle
=\frac{1}{d}{\>}\Tr(\ppm_{n}\ppm_{n})=\frac{\varkappa}{d}
\ . \label{apbp}
\end{equation}
For the isotropic state (\ref{isost}), calculations then lead to
the result
\begin{equation}
J_{M}(\wbro_{{iso}})=M
{\left(
\gamma\varkappa+\frac{1-\gamma}{d}
\right)}
\, . \label{jmiso}
\end{equation}
For the value $M=d+1$, this result was presented in \cite{fei14}.
Using the scheme with $M$ MUMs, we can certainly detect
entanglement for those values $\gamma$ that satisfy
\begin{equation}
\frac{M+\varkappa{d}-1}{Md}<\gamma\varkappa+\frac{1-\gamma}{d}
\ , \label{cogam}
\end{equation}
or merely $\gamma>1/M$. Thus, schemes with two sets of $M$
mutually unbiased measurements allow to detect the entanglement of
all isotropic states with $\gamma\in(1/M;1]$. For schemes with
mutually unbiased bases, this result was discussed in
\cite{shbah12}. In this sense, there is a good reason to realize
entanglement detection with MUMs instead of MUBs. The interval
$\gamma\in(1/M;1]$ widens with growth of $M$. Here, we see
some trade-offs between a sensitivity of the considered scheme and
costs for its implementation. The maximal interval is obtained for
$M=d+1$. It is known that isotropic states are certainly entangled
for $\gamma>1/(d+1)$ \cite{mhph1999}. Hence, using two complete
sets of MUMs allows to detect all the entanglement of isotropic
states \cite{fei14}.
Note that the described scheme of entanglement detection can be
realized for arbitrary $d$. Indeed, the explicit construction for
$d+1$ mutually unbiased measurements has been presented in
\cite{kag14}. However, acceptable values of $\varkappa$ cannot be
chosen {\it a priori}. In this way, therefore, we cannot generally
obtain a set of $d+1$ mutually unbiased bases. The problem of
maximal set of mutually unbiased bases seems to be very hard. In
practical questions, however, we may try to adopt MUMs instead of
MUBs. This possibility was already discussed in
\cite{kag14,fei14}. The above results also support such an
approach to applications of mutual unbiasedness in quantum
information processing.
\section{Conclusions}\label{sec6}
We have studied some properties of recently introduced mutually
unbiased measurements. For a set of several MUMs, we estimated
from above the sum of corresponding indices of coincidence. The
presented results are essentially based on such calculations. Then
we have obtained entropic uncertainty relations for a set of
several MUMs in terms of the R\'{e}nyi and Tsallis entropies. The
derived relations include both the state-dependent and
state-independent forms. The state-dependent bound on the sum of
corresponding R\'{e}nyi's entropies turns to be tight in some
sense. Namely, for all orders this bound is saturated with the
completely mixed state. We also obtained some improvement of the
R\'{e}nyi entropic bounds for mutually unbiased bases. The Tsallis
formulation allowed to address uncertainty relations with
detection inefficiencies. Applications of mutually unbiased
measurements in entanglement detection were considered in more
details. In particular, we obtained results for an arbitrary
number of MUMs used in entanglement detection. It seems that there
is a certain trade-off between a sensitivity of the scheme and
costs for its experimental implementation. In the literature,
mutually unbiased bases were considered as a suitable tool in
quantum state reconstruction, quantum error correction, and the
mean king's problem. It would be interesting to study possible
applications of MUMs in these questions. In this regard, the
results presented here may also be of significance.
\medskip
{\it Note added.} After this work was completed I learned about very recent results of Chen and Fei \cite{chenfei14}. These authors independently studied uncertainty relations for MUMs in terms of the R\'{e}nyi and Tsallis entropies. The uncertainty relations presented in my work differ in the following two respects. First, an arbitrary number of MUMs was considered, whereas the authors of \cite{chenfei14} deal with $d+1$ MUMs. Second, for $M=d+1$ the lower bound (\ref{rum01}) is stronger than the corresponding bound of \cite{chenfei14}.
\medskip
The author acknowledges fruitful discussions with Zbigniew
Pucha{\l}a and Karol \.{Z}yczkowski in Institute of Physics,
Jagiellonian University, Krak\'{o}w.
|
\section{Experimental details}
\subsection{Images of the Josephson parametric phase-locked oscillator}
Figure \ref{FigS_device}a shows the device image of the parametric
phase-locked oscillator (PPLO).
The device consists of a quarter wavelength (cavity length 2.6~mm) coplanar waveguide
(CPW) resonator with a dc-SQUID termination and a
pump line inductively coupled to the SQUID loop (mutual inductance $M \sim 1.0$~pH).
The critical current of the Josephson junction of the SQUID
is estimated to be 3.1~$\mu$A for each junction from
the fitting shown in Fig.~\ref{FigS_fr_flux}.
The device was fabricated by the planarized niobium trilayer
process at MIT Lincoln laboratory.
The resonator and the pump line are made of 150-nm-thick niobium film
sputtered on a Si substrate covered by 500-nm-thick SiO${}_2$ layer.
Figure \ref{FigS_device}b shows the magnified image of
the coupling capacitance between the microwave feedline and the resonator $C_{\rm in}^{\rm PO}$,
which is designed to be 15~fF.
The internal and external quality factors are measured to be
5200 and 340, respectively at $\omega_0^{\rm PO}/2\pi=10.507$~GHz,
which give the loaded quality factor $Q_{\rm L}^{\rm PO}$ of 320.
Figures~\ref{FigS_device}c and d show
the magnified optical image and the scanning electron micrograph
of the dc-SQUID part, respectively.
\subsection{Error budget of the Rabi-oscillation contrast}
In the main article, we show Rabi oscillations with a contrast of 90.7\%.
Here, we present our analysis on the loss of the contrast.
Possible sources of the error are
(i) incomplete initialization of the qubit, (ii) insufficient power of the locking signal (LS),
and (iii) qubit energy relaxation (including the gate error in the qubit-control $\pi$ pulse).
The error from the first source (incomplete initialization)
is estimated from the direct measurement of the background qubit excitation
by operating the device as Josephson parametric amplifier (JPA)~\cite{Lin13},
and found to be 2.6\%.
Half of this is equal to the non-zero minimum of the blue curve in
Fig.~4c in the main article.
This indicates that the error from the second source (non-locking error)
is negligible at least when the qubit is in state $|0\rangle$.
Non-locking error is also estimated from the measurement result shown
in Fig.~\ref{FigS_nonlock_err},
which is similar to the measurement shown in Fig.~2e in the main article,
and found to be negligible for both of the qubit states.
Minimum probability of $1\pi$ state decreases as we increase
$P_{\rm s}^{\rm PO}$ and saturates at the level determined by the background qubit excitation
when $P_{\rm s}^{\rm PO}$ is larger than $\sim -124$~dBm.
The microwave powers injected into the PPLO when the qubit is in $|0\rangle$ state
and in $|1\rangle$ state are shown
in the figure as $P_{\rm s}^0$ and $P_{\rm s}^1$, respectively.
Because of the dispersive shift in $\omega_0^{\rm r}$ and
finite internal loss of the readout resonator,
$P_{\rm s}^1$ is slightly lower than $P_{\rm s}^0$ which is set at $-122.5$~dBm.
$P_{\rm s}^1$ is estimated by directly measuring the amplitude of
the reflected microwave when the qubit is excited to $|1\rangle$ state.
Figure \ref{FigS_P0P1} shows the histograms of the voltage of the
reflected readout pulse when the qubit-control $\pi$ pulse is not
applied (a) and is applied (b) to the qubit.
The reflected voltage is measured by a single-shot readout using JPA~\cite{Lin13}.
The peak in Fig.~\ref{FigS_P0P1}a corresponds to the qubit state $|0\rangle$,
while the right peak in Fig.~\ref{FigS_P0P1}b
corresponds to the qubit state $|1\rangle$.
By fitting the peaks with Gaussian functions, we extract the voltage of the
reflected readout pulse when the qubit is in $|0\rangle$ state and
in $|1\rangle$ state, which we call $V_0$ and $V_1$, respectively.
$V_i$ and $P_{\rm s}^i$ ($i$=0,1) satisfy the following relation
\begin{equation}~\label{V_P}
\frac{P_{\rm s}^{0}G(P_{\rm s}^{0})}{P_{\rm s}^{1}G(P_{\rm s}^{1})} =
\Bigl( \frac{V_0}{V_1} \Bigr)^2,
\end{equation}
where $G$ represents the power gain of the JPA.
Note that $G$ is not necessarily independent of the input power
when the input power is high,
and we measured it independently (data not shown).
From Eq.~\ref{V_P}, $P_{\rm s}^{1}$ is estimated to be $-123.0$~dBm.
The rest of the error is attributable to the third source (qubit relaxation).
The energy relaxation time $T_1$ of the qubit is measured to be 690~ns (see below).
Assuming $1-\exp(-t_{\rm w}/T_1)=0.067$,
$t_{\rm w}=48$~ns is obtained, which is close to the sum of
$t_{\rm c}$, $t_{\rm p}$ and the response time of PPLO measured
to be $\sim10$~ns (Fig.~\ref{FigS_tcon}).
\subsection{Qubit energy relaxation time}
Figure \ref{FigS_T1} shows the probability of 0$\pi$ state as a function of
$t_{\rm p}$ when the qubit-control $\pi$ pulse is turned off and on.
We fit the data ($\pi$ pulse on)
from $t_{\rm p}=40$ ns to 3 $\mu$s with an exponential function,
and obtain the time constant of 690 ns.
This agrees with the qubit energy relaxation time $T_1$ of $694$ ns
obtained from an independent ensemble-averaged measurement
using standard pulse sequence for $T_1$ measurement, namely,
$\pi$ pulse followed by the delayed readout.
Small decay of the data without $\pi$ pulse is possibly due to the
qubit excitation induced by the readout pulse.
\section{Theory and simulations}
\subsection{Hamiltonian and equations of motion}
The Hamiltonian of the PPLO including a signal port for the locking signal (LS) and
a fictitious loss port for internal loss of the resonator is given by~\cite{Wallsbook}
\begin{eqnarray} \label{hamiltonian}
\mathcal{H}(t) &=& \mathcal{H}_{\rm sys}(t) + \mathcal{H}_{\rm sig} + \mathcal{H}_{\rm loss}, \\
\mathcal{H}_{\rm sys}(t)/\hbar &=& \omega_0^{\rm PO} \bigl[ a^\dagger a +
\epsilon \cos(\omega_{\rm p} t) (a + a^\dagger)^2 \bigr] + \gamma (a + a^\dagger)^4, \\
\mathcal{H}_{\rm sig}/\hbar &=& \int dk \Bigl[v_b k b_k^\dagger b_k +
\mathrm{i} \sqrt{\frac{v_b \kappa_1}{2 \pi}}
\Bigl( a^\dagger b_k-b_k^\dagger a \Bigr) \Bigr], \\
\mathcal{H}_{\rm loss}/\hbar &=& \int dk \Bigl[v_c k c_k^\dagger c_k +
\mathrm{i} \sqrt{\frac{v_c \kappa_2}{2 \pi}}
\Bigl( a^\dagger c_k-c_k^\dagger a \Bigr) \Bigr],
\end{eqnarray}
where $\omega_0^{\rm PO}$ is the static resonant frequency of the PPLO,
$\omega_{\rm p}$ and $\epsilon$ represent the frequency and
magnitude of the parametric modulation, respectively,
$\gamma$ represents the nonlinearity of the Josephson junction (JJ),
$a$ is the annihilation operator for the resonator,
$b_k$ ($c_k$) is the annihilation operator for the photon in the signal (loss) port
with a wave number $k$ and a velocity $v_b$ ($v_c$), and
$\kappa_1$ ($\kappa_2$) represents the coupling strength between the resonator
and the signal (loss) port.
The operators satisfy the following commutation rules:
$\bigl[ a,a^\dagger \bigr] = 1$, $\bigl[ b_k,{b_k'}^\dagger \bigr] = \delta (k-k')$, and
$\bigl[ c_k,{c_k'}^\dagger \bigr] = \delta(k-k')$.
The coupling constants $\kappa_1$ and $\kappa_2$ are related to the
external and internal quality factors of the resonator as follows:
$\kappa_1 = \omega_0^{\rm PO}/Q_{\rm e}^{\rm PO}$ and
$\kappa_2 = \omega_0^{\rm PO}/Q_{\rm i}^{\rm PO}$.
Below we consider the case where $\omega_{\rm p} = 2\omega_0^{\rm PO}$.
From the Heisenberg equations of motion for $b_k$, we obtain
\begin{equation}\label{Heq_b}
\frac{db_k(t)}{dt} = -\mathrm{i} vk b_k(t) - \sqrt{\frac{v\kappa_1}{2 \pi}} a(t).
\end{equation}
By solving this differential equation formally,
we have
\begin{equation}\label{sol_Heq_b}
b_k(t) = e^{- \mathrm{i} vkt} b_k(0)
- \sqrt{\frac{v\kappa_1}{2 \pi}} \int_{0}^{t} e^{- \mathrm{i} vk(t - t')} a(t')dt'.
\end{equation}
We introduce the real-space representation of the waveguide field by
$\widetilde{b}_r=(2\pi)^{-1/2}\int dk e^{\mathrm{i}kr} b_k$.
In this representation,
the waveguide field interacts with the resonator at $r=0$ and
the $r<0$ ($r>0$) region corresponds to the incoming (outgoing) field.
From Eq.~\ref{sol_Heq_b}, we have
\begin{eqnarray}
\widetilde{b}_r(t)
&=& \widetilde{b}_{r-vt}(0)
- \sqrt{\frac{\kappa_1}{v}} \theta(r)\theta(t-r/v)a(t-r/v) \label{br_op},
\end{eqnarray}
where $\theta(r)$ is the Heviside step function.
We define the input and output operators by
\begin{eqnarray} \label{bin_op}
\widetilde{b}_{\rm in}(t) &\equiv& \widetilde{b}_{-0}(t)
= \widetilde{b}_{-vt}(0), \\
\widetilde{b}_{\rm out}(t) &\equiv& \widetilde{b}_{+0}(t)
= \widetilde{b}_{\rm in}(t) - \sqrt{\frac{\kappa_1}{v}} a(t).
\end{eqnarray}
Using Eqs.~\ref{br_op} and \ref{bin_op},
the field operator $\widetilde{b}_r(t)$ at the resonator position ($r=0$) is given by
\begin{equation} \label{b0_op}
\widetilde{b}_0(t) = \frac{1}{\sqrt{2\pi}}\int b_k(t)dk =
\widetilde{b}_{\rm in}(t) - \frac{1}{2}\sqrt{\frac{\kappa_1}{v}} a(t).
\end{equation}
From the Heisenberg equations of motion for $a$, we obtain
\begin{equation}\label{Heq_a}
\frac{da}{dt} =
-\mathrm{i}[\mathcal{H}_{\rm sys}(t),a]
+ \sqrt{v\kappa_1} \widetilde{b}_0
+ \sqrt{v\kappa_2} \widetilde{c}_0.
\end{equation}
Using Eq.~\ref{b0_op} and its counterpart for $\widetilde{c}_0$,
Eq.~\ref{Heq_a} is rewritten as
\begin{equation}\label{Heq_a3}
\frac{da}{dt} =
-\mathrm{i}[\mathcal{H}_{\rm sys}(t),a]-\frac{\kappa}{2}a
+ \sqrt{v\kappa_1} \widetilde{b}_{\rm in}(t) +
\sqrt{v\kappa_2} \widetilde{c}_{\rm in}(t),
\end{equation}
where $\kappa=\kappa_1 + \kappa_2$.
Now we switch to a frame rotating at $\omega_0^{\rm PO}$
[namely,
$a(t)e^{\mathrm{i}\omega_0^{\rm PO} t} \to a(t)$,
$b_{\rm in}(t)e^{\mathrm{i}\omega_0^{\rm PO} t} \to b_{\rm in}(t)$, and
$c_{\rm in}(t)e^{\mathrm{i}\omega_0^{\rm PO} t} \to c_{\rm in}(t)$]
and drop the rapidly rotating terms in $\mathcal{H}_{\rm sys}(t)$.
Then the static system Hamiltonian is given by
\begin{equation} \label{hamiltonian_simple}
\mathcal{H}_{\rm sys}/\hbar =
\frac{\epsilon \omega_0^{\rm PO}}{2}(a^2 + a^{\dagger 2})
+ 6\gamma a^\dagger a^\dagger a a.
\end{equation}
Here, we neglected the term $12\gamma a^\dagger a$
which can be regraded as a small renormalization to $\omega_0^{\rm PO}$.
From Eqs.~\ref{Heq_a3} and \ref{hamiltonian_simple}, we have
\begin{equation}\label{Heq_a4}
\frac{da}{dt} =
-\frac{\kappa}{2}a
-\mathrm{i} \omega_0^{\rm PO} \epsilon a^\dagger
-12\mathrm{i}\gamma a^\dagger aa
+ \sqrt{v\kappa_1} \widetilde{b}_{\rm in}(t) +
\sqrt{v\kappa_2} \widetilde{c}_{\rm in}(t).
\end{equation}
Now we consider the classical amplitude of the resonator field, namely,
$\langle a(t) \rangle$.
We denote the locking signal applied from the signal port by
$E_{\rm s}(r,t)=E_{\rm s}^* e^{\mathrm{i}\omega_0^{\rm PO}(r/v-t)}$.
Then, we rigorously have
$\langle \widetilde{b}_{\rm in}(t) \rangle = E_{\rm s}^*$
and $\langle \widetilde{c}_{\rm in}(t) \rangle = 0$.
Dividing $\langle a(t)\rangle$ into its quadratures as
$\langle a(t)\rangle = q_x(t) - \mathrm{i}q_y(t)$,
their equations of motion are given by
\begin{eqnarray}
\frac{dq_x}{dt} &=& -\frac{\kappa}{2}q_x + \omega_0^{\rm PO} \epsilon q_y
- 12\gamma(q_x^2+q_y^2)q_y + \sqrt{\kappa_1}
|E_{\rm s}|\cos\theta_{\rm s},
\label{eq_of_mot_7} \\
\frac{dq_y}{dt} &=& -\frac{\kappa}{2}q_y + \omega_0^{\rm PO} \epsilon q_x
+ 12\gamma(q_x^2+q_y^2)q_x + \sqrt{\kappa_1}
|E_{\rm s}|\sin\theta_{\rm s},
\label{eq_of_mot_8}
\end{eqnarray}
where $E_{\rm s} = |E_{\rm s}|e^{\mathrm{i}\theta_{\rm s}}$,
and we approximated $\langle a^\dagger aa \rangle$ to be
$\langle a^\dagger \rangle \langle a \rangle^2$.
Equations~\ref{eq_of_mot_7} and \ref{eq_of_mot_8} can be recast as
\begin{eqnarray}
\frac{dq_x}{dt} &=& -\frac{\kappa}{2}q_x + \frac{\partial g}{\partial q_y}, \\
\frac{dq_y}{dt} &=& -\frac{\kappa}{2}q_y - \frac{\partial g}{\partial q_x},
\end{eqnarray}
where
\begin{equation}
g(q_x,q_y) = \frac{\epsilon}{2}\omega_0^{\rm PO}(q_y^2 - q_x^2) - 3\gamma(q_x^2 + q_y^2)^2 +
\sqrt{\kappa_1} |E_{\rm s}| (q_y \cos\theta_{\rm s} - q_x \sin\theta_{\rm s}).
\end{equation}
Considering that $P_{\rm p}/P_{\rm p0}=(\epsilon/\epsilon_0)^2$,
where $\epsilon_0=\kappa/(2\omega_0^{\rm PO})$ is the threshold for $\epsilon$,
these are Eqs.~2 to 4 in the main article.
\subsection{Master equation}
We denote the resonator transition operator by $s_{mn}=|m\rangle \langle n|$,
where $|m\rangle$ and $|n\rangle$ are the Fock states.
Its Heisenberg equation is given,
in the rotating frame [$s_{mn}(t)e^{\mathrm{i}\omega_0^{\rm PO}(n-m)t} \to s_{mn}(t)$], by
\begin{eqnarray}
\frac{d}{dt}s_{mn} &=&
\frac{\mathrm{i}}{\hbar}[\mathcal{H}_{\rm sys},s_{mn}]
+ \frac{\kappa}{2}(2a^\dagger s_{mn}a - s_{mn}a^\dagger a - a^\dagger as_{mn}) \nonumber \\
&& + \sqrt{v\kappa_1}[s_{mn},a^\dagger]\widetilde{b}_{\rm in}(t)
- \sqrt{v\kappa_1}\widetilde{b}_{\rm in}^\dagger(t)[s_{mn},a] \nonumber \\
&& + \sqrt{v\kappa_2}[s_{mn},a^\dagger]\widetilde{c}_{\rm in}(t)
- \sqrt{v\kappa_2}\widetilde{c}_{\rm in}^\dagger(t)[s_{mn},a].
\end{eqnarray}
where the static system Hamiltonian $\mathcal{H}_{\rm sys}$
is given by Eq.~\ref{hamiltonian_simple}.
Using again that
$\langle \widetilde{b}_{\rm in}(t) \rangle = E_{\rm s}^*$
and $\langle \widetilde{c}_{\rm in}(t) \rangle = 0$,
$\langle s_{mn} \rangle$ evolves as
\begin{eqnarray}~\label{Heq_Bexp}
\frac{d}{dt}\langle s_{mn} \rangle &=&
\frac{\mathrm{i}\omega_0^{\rm PO}\epsilon}{2} \Bigl(
\sqrt{m(m-1)}\langle s_{m-2,n} \rangle + \sqrt{(m+1)(m+2)} \langle s_{m+2,n} \rangle \nonumber \\
&& - \sqrt{n(n-1)}\langle s_{m,n-2} \rangle - \sqrt{(n+1)(n+2)} \langle s_{m,n+2} \rangle
\Bigr) \nonumber \\
&& + 6\mathrm{i}\gamma[m(m-1)-n(n-1)] \langle s_{mn} \rangle
+ \frac{\kappa}{2} \Bigl[
2\sqrt{(m+1)(n+1)}\langle s_{m+1,n+1} \rangle - (m+n) \langle s_{mn} \rangle \Bigr] \nonumber \\
&& + \sqrt{\kappa_1} E_{\rm s} \Bigl(
\sqrt{n}\langle s_{m,n-1} \rangle - \sqrt{m+1}\langle s_{m+1,n} \rangle
\Bigr) \nonumber \\
&& - \sqrt{\kappa_1} E_{\rm s}^* \Bigl(
\sqrt{n+1}\langle s_{m,n+1} \rangle - \sqrt{m}\langle s_{m-1,n} \rangle
\Bigr),
\end{eqnarray}
where we have used $a^\dagger s_{mn}a = \sqrt{(m+1)(n+1)}s_{m+1,n+1}$ and similar equalities.
Since $\langle s_{mn} \rangle = {\rm Tr}[\rho s_{mn}] = \rho_{nm}$,
Eq.~\ref{Heq_Bexp} is equivalent to the following master equation,
\begin{equation}~\label{dim_master}
\frac{d\rho}{dt} = -\frac{\mathrm{i}}{\hbar}[\mathcal{H}_{\rm int}, \rho]
+ \frac{\kappa}{2}(2a\rho a^\dagger - a^\dagger a \rho - \rho a^\dagger a),
\end{equation}
where
\begin{equation}
\mathcal{H}_{\rm int}/\hbar =
\frac{\epsilon \omega_0^{\rm PO}}{2}(a^2 + a^{\dagger 2})
+ 6\gamma a^\dagger a^\dagger aa
+ \mathrm{i}\sqrt{\kappa_1}|E_{\rm s}|
(e^{\mathrm{i}\theta_{\rm s}}a^\dagger - e^{\mathrm{-i}\theta_{\rm s}}a).
\end{equation}
By introducing dimensionless time $\tau = t\kappa/2$,
Eq.~\ref{dim_master} becomes
\begin{equation}~\label{master_eq}
\frac{d\rho}{d\tau} = -\mathrm{i}[\mathcal{H}'_{\rm int},\rho] +
(2a\rho a^\dagger - a^\dagger a\rho - \rho a^\dagger a).
\end{equation}
Here the dimensionless Hamiltonian $\mathcal{H}'_{\rm int}$ is given by
\begin{equation}
\mathcal{H}'_{\rm int} = \mathcal{H}_{\rm int}/(\hbar\kappa/2)
= \gamma'a^\dagger a^\dagger aa +
\frac{1}{2}\sqrt{\frac{P_{\rm p}}{P_{\rm p0}}}(a^2 + a^{\dagger2}) +
\mathrm{i}\sqrt{N^{\rm PO}}
(e^{\mathrm{i}\theta_{\rm s}}a^\dagger - e^{\mathrm{-i}\theta_{\rm s}}a),
\end{equation}
where $\gamma'=12\gamma/\kappa$,
$\sqrt{N^{\rm PO}}=\sqrt{\kappa_1}|E_{\rm s}|/(\kappa/2)$.
We numerically solve Eq.~\ref{master_eq} by
expanding $\rho$ in the number state basis~\cite{Johansson13},
namely $\rho=\sum_{m,n=0}^N \rho_{mn}|m\rangle \langle n|$,
and calculate the $Q$ function $Q(z)=\langle z|\rho|z \rangle/\pi$,
where $z$ is a coherent state~\cite{Wallsbook}.
\subsection{Experimental parameters}
To simulate the experiments, we need to determine the parameters
such as $N^{\rm PO}$, $P_{\rm p0}$, and $\gamma$.
$N^{\rm PO}$ is determined from the above definition and
$|E_{\rm s}|=\sqrt{P_{\rm s}^{\rm PO}/\hbar \omega_0^{\rm PO}}$.
Since we do not precisely know the mutual inductance between
the pump line and the SQUID loop in the PPLO, we can only roughly
determine $P_{\rm p0}$, the threshold for the pump power,
from the measurement shown in Fig.~2a in the main article,
and leave it as a semi-adjustable parameter.
$\gamma$ is calculated based on the theory in Ref.~\onlinecite{Wallquist06}.
It is related with the wavenumber $k$ of the first mode of the CPW resonator
terminated by a SQUID,
\begin{equation}~\label{gamma_Bk}
\gamma = -\Bigl( \frac{2\pi}{\Phi_0} \Bigr)^2\frac{\hbar B_k}{8C_k},
\end{equation}
where
\begin{eqnarray}~\label{B_k_param}
B_k &=& \frac{(1/4)\cos^2(kd)}{1+2kd/\sin(2kd)}, \\ \label{equivC}
C_k &=& \frac{C_{\rm cav}}{2}\Bigl[ 1+\frac{\sin(2kd)}{2kd} \Bigr] + C_{\rm J}\cos^2(kd).
\end{eqnarray}
Here, $d$ and $C_{\rm cav}$ are the length and the total capacitance of the CPW resonator,
respectively, and $C_{\rm J}$ is the total junction capacitance of the SQUID.
The flux dependent resonant frequency is given by
\begin{equation}~\label{fdep_omega0}
\omega_0^{\rm PO} (\Phi_{\rm sq})
= \frac{1}{\sqrt{L_k(C_k+C_{\rm in}^{\rm PO})}},
\end{equation}
where
\begin{equation}
1/L_k = \frac{(kd)^2}{2L_{\rm cav}}\Bigl[ 1+\frac{\sin(2kd)}{2kd} + \frac{2C_{\rm J}}{C_{\rm cav}}
\cos^2(kd) \Bigr].
\end{equation}
Here, $L_{\rm cav}$ is the total inductance of the CPW resonator, and $kd$ is determined by
the equation~\cite{Wallquist06},
\begin{equation}~\label{kd_eq}
kd\tan(kd) = \Bigl( \frac{2\pi}{\Phi_0} \Bigr) L_{\rm cav}2I_0\Bigl|
\cos\Bigl( \pi \frac{\Phi_{\rm sq}}{\Phi_0} \Bigr) \Bigr| -
\frac{C_{\rm J}}{C_{\rm cav}}(kd)^2,
\end{equation}
where $I_0$ represents the critical current of each of the SQUID JJ.
We fit the data shown in Fig.~\ref{FigS_fr_flux} by
Eq.~\ref{fdep_omega0} with fitting parameters of $I_{\rm c}$ and $C_{\rm cav}$,
which are determined to be 3.1~$\mu$A and 410~fF, respectively.
Using Eqs.~\ref{gamma_Bk} to \ref{equivC}, and \ref{kd_eq},
we calculate $\gamma$ to be
$-2.3\times10^5$ Hz at $\omega_0^{\rm PO}/2\pi=10.51$ GHz.
\subsection{Simulation of non-locking error}
Figure~\ref{fig_th}a shows an example of the calculated $Q$ function.
The parameters used in the calculation are
$P_{\rm p}/P_{\rm p0}=1.660$ (2.2~dB), $N^{\rm PO}=0.09$, $\theta_{\rm s}=\pi/2$.
The density matrix is truncated at $N=80$, which
we confirmed large enough for these parameters.
Also, the calculation is stopped at $\tau=20$,
which we confirmed long enough for the system to become stationary.
We clearly observe two distribution peaks, which correspond to
$0\pi$ and $1\pi$ states of PPLO.
Based on this result, we get the probability of $0\pi$ state
by integrating $Q$ for $q_x > 0$.
Figure~\ref{fig_th}b shows the probability of $0\pi$ state as a function of
the LS phase for different $N^{\rm PO}$'s.
$N^{\rm PO}$'s are chosen from those in Fig.~2d in the main article.
The agreement is fairly good.
|
\section{Introduction}
This communication follows the Noma-13 conference in September 2013, an enjoyable and fruitful meeting where one of us (YP) had a chance to hear of the model considered below \cite{zareq}. This model describes the evolution of the distribution of wealth in a population of individuals doing business pairwise. After each exchange there is a redistribution of money between the two individuals, without total loss or gain. A feature of this model, the ``Z-model" (with Z for Zaragoza) is its simple equilibrium solution (written below). Under its law of evolution, this equilibrium solution is stable and so attract most, if not all initial conditions satisfying convergence conditions (finite total probability and finite total wealth) \cite{zareq.1}. Moreover, an $H$-theorem is valid for this model \cite{apenko2013}. We show below that the evolution of higher momenta (mean square value, mean cubic value, etc.) of the wealth can be computed exactly, obviously under the condition that those momenta exist. We consider also situations where the momenta do not converge beyond a given order. An anonymous referee pointed out that something called ``q-model" has equations similar to the Z-model. Those q-models aim at describing the distribution of stress in random set of solid grains in contact with neighbours in such a way that the downward push of the weight of a grain and of the grain above it is distributed more or less randomly between its neighbours underneath. In this theory the equivalent of the time of the Z-model is played by the vertical direction and the time-iteration amounts to move down the pile to find the distribution of stress on grains. Even though the equations of this q-model look like the ones of the Z-model, their physical meaning is quite different. The interested reader may get a list of papers on the subject in the reference list of the lecture notes published in \cite{bouchaud}. Moreover the q-model, in order to get a row to row equation of iteration like the one of the Z-model has to assume that the vertical force on beads on the same horizontal row are statistically independent, which is presumably needed to get at the end something like a hyperbolic system, although the Cauchy-Poisson equations for regular elasticity are elliptic.
Because of its simple mathematical structure it makes sense to extend the Z-model by keeping the possibility of an exact solution for the momenta. This can be done with a straightforward extension maintaining the basic properties of conservation of the total probability and the total wealth. Although this modified Z-model looks very much like the original and reduces to it continuously as a parameter changes, it has completely different properties. In particular it shows an increase of the fluctuations of wealth as time goes, a rather unexpected property, absent in the original model. This makes the matter of Section \ref{sec:genZ}. In this respect the inequality of wealth, as studied below makes only a small part of this big subject, but it is at least one that one can try to describe quantitatively.
Motivated by this consideration of momenta, we look in Section \ref{sec:divmom} at what happens in the Z-model when the momenta do not converge, specifically when the distribution of wealth decays algebraically for large values so that momenta do not exist, at least initially, beyond a certain power (This might be related to what is called Pareto law, Pareto \cite{pareto} having predicted that the natural distribution of wealth decays algebraically for large values, a property of the mZ-model studied below). An interesting result of this analysis is that, after a certain number of iterations (namely after a finite amount of time) higher momenta converge although they diverged initially. Somehow, without venturing into the area of political science, this looks like the exact opposite of what is predicted sometimes (without relying on objective modelisation as much we can tell): fewer and fewer individuals get richer and richer although the other ones get poorer and poorer as time goes. This could have other explanations of course, like what is called the redistribution of wealth by the tax system in modern economies.
We shall explain first how to solve ``exactly" the moment problem, for a probability distribution decaying fast enough at infinity and then look at what happens if, initially, this probability distribution decays algebraically for large values.
In section \ref{richest} we give the probability distribution of the wealth of the ``richest man", namely the largest wealth of a given finite number of agents with a given probability distribution of the wealth with agents taken at random in the population. An explicit expression of this probability distribution of the maximum of wealth, with its limit in the case of a large number of agents, is given.
The last Section is a Summary and Conclusion section.
\section{The Z-model}
\label{sec:Zmodel}
In this model one considers a positive variable, with various names, $x$, $u$, etc, is for the amount of money owned by an individual. This amount changes in the course of time because of random exchanges between the individuals taking place at discrete time, in a synchronous way in the system. The fundamental quantity is $p_t(x)$, the probability that an individual taken at random in the population has an amount $x$ at time $t$. At the next time step $(t +1)$, due to the binary exchanges, $p_t(x)$ has changed according to the law of iteration found of reference \cite{zareq}:
\begin{equation}
p_{t+1} (x) = \int \int_{S(x)} {\mathrm{d}}u {\mathrm{d}}v \frac{p_{t} (u) p_{t} (v)}{u + v}
\mathrm{,}
\label{eq:iter}
\end{equation}
The domain of integration in Equation (\ref{eq:iter}) is defined by
$$ S(x) = \{ (u, v), u, v>0, u + v > x \}\mathrm{.}$$
This integral equation is for a function of $x$, positive variable. Because $p(.)$ is a probability distribution, it has to be positive or zero. Moreover it is normalised in such a way that $ \int_0^{\infty} {\mathrm{d}}u p_t (u) = 1$, and $t$ is a discrete index representing time.
This law of evolution of the wealth is derived as follows. Suppose two individuals, each one with the same probability of wealth, say $p(u)$, put their money in the same basket. Then the probability distribution for what is in the basket (the amount $w$) is
$$q(w) = \int_0^{\infty} {\mathrm{d}}v p(v) p(w - v) H(w-v) \mathrm{,}$$
where $H(.)$ is Heaviside function, zero for a negative argument and one otherwise. Suppose we share between two individuals the amount $w$ by taking randomly a value in $[0, w]$, give it to the first individual and the rest to the other. The probability distribution of what is taken by anyone of those individuals is
$$ r(s) = \frac{\chi_{w}(s)}{w} \mathrm{,}$$
where $\chi_{w}(s)$ is the characteristic function of the interval $[0, w]$. By extending this simple formula to the probability distribution $q(w)$ of the values of $w$, as derived above, one obtains:
$$ r(s) = \int_0^{\infty} \frac{ {\mathrm{d}}w}{w} H(w - s) \int_0^{\infty} {\mathrm{d}}v p(v) p(w - v) H(w-v) \mathrm{,}$$
After rearranging the integrals one finds
$$ r(s) = \int_{0}^{\infty}{\mathrm{d}} v' \int_{s-v'> 0}^{\infty} p(v') p(u') \frac{{\mathrm{d}} u'}{u' + v'} \mathrm{,}$$
which is a form of the right-hand side of Equation (\ref{eq:iter}).
Equation (\ref{eq:iter}) can be integrated explicitly, at least in some sense. Let us define the moments of $p_t(x)$ as
\begin{equation}
m_k (t) = \int {\mathrm{d}}u u^k p_t (u)
\mathrm{.}
\label{eq:momendef}
\end{equation}
We consider first the case where all momenta converge. In Section \ref{sec:divmom} we discuss the situation where some momenta do not exist at a given time because the integral (\ref{eq:momendef}) diverges at $k$ large, which is well possible because the ``physical" constraints on $p(u)$ is to have well defined (not diverging) values of $m_0$ and $m_1$ only.
From Equation (\ref{eq:iter}) one derives the following equation for the momenta of $p_{t+1} (.)$ as a function of the momenta of $p_{t} (.)$:
\begin{equation}
m_k (t+1) = \frac{1}{k + 1} \Sigma_{0\leq l \leq k} C_k^l m_{k-l} (t) m_{l} (t)
\mathrm{,}
\label{eq:momenteq}
\end{equation}
where $C_k^l = \frac{ k!}{(k - l)! l!}$ are the binomial coefficients.
This shows that the momenta of order $k$ at time $(t + 1)$ can be found if the momenta of smaller power at time $t$ are known. The formula is also consistent with the fact that $m_0 = 1$ at any time and that $m_1$ is a conserved positive constant (called later $m_1$). Let us look at the equation for $m_2$. It reads:
\begin{equation}
m_2 (t+1) = \frac{2}{3} ( m_2(t) + m_1^2)
\mathrm{,}
\label{eq:momenteq2}
\end{equation}
Because this equation is linear with respect to $m_2$ it can be integrated at once with the result (supposing $m_2 (0)$ given):
\begin{equation}
m_2 (t) = \left(\frac{2}{3}\right)^t m_2(0) + 2 m_1^2 \left[ 1 - \left(\frac{2}{3}\right)^t \right] = \left(\frac{2}{3}\right)^t (m_2(0) - 2 m_1^2) + 2 m_1^2
\mathrm{,}
\label{eq:momenteq2sol}
\end{equation}
The higher momenta can be computed also explicitely as functions of the initial data for the lower order momenta, the result becoming more and more cumbersome as the order increases. At third order one has:
\begin{equation}
m_3 (t+1) = \frac{1}{2} ( m_3(t) + 3 m_2(t) m_1)
\mathrm{,}
\label{eq:momenteq3}
\end{equation}
Let
$$S_3(t) = \frac{3}{2} m_2(t) m_1 \mathrm{.} $$ Therefore $$m_3(t) = \left(\frac{1}{2}\right)^t \left[ m_3(0) + \Sigma_{0\leq \theta \leq t} 2^{\theta} S_3 (\theta - 1)\right] \mathrm{,} $$ is a solution for $m_3(t)$ as a function of $m_1$, $m_2(0)$ and $m_3(0)$. The sums can be done explicitly because they involve geometric series.
The method of integration just explained does not work if one takes momenta with noninteger exponents because there is no finite equivalent of the binomial formula for such noninteger power.
\section{Definition and solution of a generalized Z-model}
\label{sec:genZ}
The Z-model can be generalized in the following way. In the original formulation, each of the two partners in a transaction have a random amount $u$ and $v$. During the transaction they put first the whole amount $(u+v)$ in a basket and then share its content randomly. The Z-model describing this satisfies the constraint that the total probability is one and that the total money is also conserved. This model has also the property that the equilibrium solution (namely the distribution of wealth such that $p_t(u) = p_{t+1}(u)$) is known explicitly and is
$$p_{eq} (u) = \frac{1}{ m_1} e^{-\frac{u}{m_1}} \mathrm{,} $$
Below we suggest a modified recursion relation, analogous to the one given in Equation (\ref{eq:iter}) but such that no simple expression of the equilibrium distribution can be found, even though the mass and first momentum $m_1$ is conserved (we keep the same notation, $m_k(t)$ for the $k$-th moment in the mZ-model, defined below, as in the Z-model). This model reads:
\begin{equation}
P_{t+1} (x) = \int \int_{S_a(x)} {\mathrm{d}}u {\mathrm{d}}v \frac{P_{t} (u) P_{t} (v)}{a u + (2-a) v}
\mathrm{,}
\label{eq:iterm}
\end{equation}
In this equation, $a$ is a real parameter, between $0$ and $2$, and $S_a(x)$ is defined by the condition $x < a u + (2-a) v$. In this model at the time of the transaction between the two individuals, one of the individual puts $(au)$ in the basket (instead of $u$ in the Z-model) and the other puts $(2-a)v$ in the basket, instead of $v$. Although this model is apparently not conservative, this is not the case. If we consider the symmetrical interaction for the pair of agents $(v,u)$, in this case the first agent will put $(av)$ in the basket and the second one $(2-a)u$. For both trades, those of the pairs $(u,v)$ and $(v,u)$, the total money to share in the basket is $2(u+v)$, then the total wealth is conserved. It can be interpreted that the excess of money in one of the trades is injected to cover the lack of money in the other trade. This is just one of the functions done by the bank system. Therefore, perhaps this is not such an unrealistic model because, nowadays (and very likely before), banks and even States rent money they do not really have and do that within constraints based on multiplicative factors of their actual wealth.
Like the Z-model, the modified Z-model (or mZ-model) defined by the iteration (\ref{eq:iterm}) satisfies the constraints of conservation of $m_0$ and $m_1$ if $m_0 = 1$. From simple algebra, one finds:
$$ m_0(t+1) = m_0(t)^2 \mathrm{,} $$ and
$$ m_1(t+1) = m_0(t) m_1(t) \mathrm{.} $$
Therefore the first two momenta are constant if $m_0 = 1$ and if $m_1$ converge, as we assume it.
Contrary to the case of the Z-model, there is no simple equilibrium solution. However it is possible to derive many properties of this equilibrium from the equations for the moments. This is because the denominator in the iteration formula is a linear function of $u$ and $v$ like in the Z-model. The recursion relation for the second moment is:
\begin{equation}
m_{2} (t + 1) = \frac{1}{3} \left[ (4 - 4 a + 2 a^2) m_2(t) + 2(2-a) a m_1^2\right]
\mathrm{,}
\label{eq:itermomentm}
\end{equation}
As can be easily checked, this reduces to the formula valid for the Z-model, Equation (\ref{eq:momenteq2}), in the case $a = 1$. However a very interesting difference appears in this iteration law (again, an iteration derived from the iteration for the probability distribution with no other assumption than the existence of the second moment). Actually this iteration may lead to an exponentially growing second moment. This happens if the coefficient of $m_2(t) $ in Equation (\ref{eq:itermomentm}) is larger than one. This happens if $a$ is outside of the interval $[1 -\frac{1}{\sqrt{2}}, 1 + \frac{1}{\sqrt{2}}]$ which is compatible with the condition that $0<a<2$. Therefore there can be an instability of the second moment leading to an indefinite increase of the width of the distribution of wealth. Without overstating this, one can say that this makes a model of ever increasing inequality as predicted by some socio-economical theories.
Moreover, for any $a$ different of $1$, the iteration of higher momenta become unstable. To show this, let us define $b = 1 -a$. The iteration of the $k$-th moment reads:
\begin{equation}
m_{k} (t + 1) = \frac{1}{k + 1} \left[ \left((1-b)^k + (1+b)^k\right) m_k(t) + l.o.t(t) \right] \mathrm{,}
\label{iter-km}
\end{equation}
In this equation, $l.o.t(t)$ is for the lowest order terms, depending on momenta of order less than $k$. Let us consider the smallest $k$ such that, for a given $a$, there is an exponential growth of this moment. Therefore $l.o.t(t)$ remains bounded as a function of time and so, if there is an instability, it is dominated after a sufficient number of iterations by the exponentially growing $ \left((1-b)^k + (1+b)^k\right) m_k(t) $. A little algebra shows that the coefficient of $m_k(t)$ on the right-hand side of Equation (\ref{iter-km}) is larger than 1 and the moment grows exponentially if $$\ln(1 + |b|) > \frac{\ln(k +1)}{k} \mathrm{,}$$
If $|b|$ is small, this is equivalent to the condition
$$ k > \frac{\ln(1/|b| + 1)}{|b|} \mathrm{.}$$
It shows that, however $|b|$ is small but not zero, the large order momenta are unstable under the iteration. Recall that $|b|$ small is equivalent to have a mZ-model formally close to the original Z-model. This also shows that, however small (but non zero) $|b|$ is, the steady distribution, if it exists, given by the iteration law should decay with a power law at large values of its argument to make diverge momenta with a large power. It is planned to return to this mathematically interesting question in a future publication.
\section{Diverging moments at time zero}
\label{sec:divmom}
In this Section we return to the Z-model in its original form and consider the following question: what happens to the iterations if the initial momenta diverge beyond a certain power? Indeed, because the initial condition is in principle rather free, provided $m_0 = 1$ and $m_1$ converges, one can always imagine an initial condition with a distribution of wealth decreasing algebraically for large powers. In this case momenta do not exist beyond a certain power. We consider below what happens in this case. In particular we show that, after a finite number of iterations, one recovers a converging moment with a power less than a value increasing as the iterations go.
We shall limit ourselves to situations where $p_0(u)$, the initial distribution of wealth, behaves at large $u$ as a power law, like
\begin{equation}
p_0(u) \approx l_{\alpha_0} u^{-\alpha_0}
\mathrm{,}
\label{eq;as1}
\end{equation}
where $ l_{\alpha_0}$ is a positive constant and $\alpha_0$ a positive exponent. To have finite probability and first momentum (finite total wealth) on must have $\alpha_0>2$. By putting this power law in the right-hand side of the functional iteration (\ref{eq:iter}), one obtains that at time $t = 1$, the distribution of wealth $p_1(u)$ decays with the power law:
\begin{equation}
p_1(u) \approx l_{\alpha_1} u^{-\alpha_1}
\mathrm{,}
\label{eq;as2}
\end{equation}
where $\alpha_1 = 2 \alpha_0 - 1$ and where
$$ l_{\alpha_1} = l_{\alpha_0}^2 B(\alpha_0) \mathrm{,}$$
where
$$ B(\alpha) = \int \int_{S(1)} {\mathrm{d}}u' {\mathrm{d}}v' \frac{(u'v')^{-\alpha}}{u' + v'} $$ is a numerical function of the argument $\alpha$. As the iteration formula shows, $\alpha$ increases as the iteration goes and so, as soon as it becomes big enough, momenta of a given power begin to exist, and follow later the explicit recursion formulae given in Equation (\ref{eq:momenteq}). This is correct because momenta of higher and higher order begin to converge the later as their power increases. Therefore the right-hand side of the recursion equation becomes all well defined when the highest moment becomes well defined, all momenta of a smaller power being already finite at this time.
\section{Probability distribution of the wealth of the richest man}
\label{richest}
Looking at the economic magazines, one is struck by their insistence on various lists of rich, if not very rich people, lists ordered according to their supposed wealth. Therefore it is of some interest to consider the question of the distribution of biggest wealth that can be reached within the models outlined in this work. We begin with a basic question of probability: given a probability distribution $p(x)$, and a number $\nu$ of independent trials, what is the largest value reached among those trials? This interesting question can be answered quite simply as demonstrated below. Then we apply this result to the case of the Z and of the mZ model.
Consider first the following problem: given $x_0$ positive, let us draw a number $x$ with probability distribution $p(x)$. What is the distribution of the maximum of $x_0$ and $x$, a maximum denoted as $X$? If $x$ is less than $x_0$ this maximum is $x_0$, in the opposite case it is $x$. Define $$ N(x) = \int_0^x {\mathrm{d}}x' p(x') \mathrm{.}$$ The probability that $x$ is less than $x_0$ is $N(x_0)$. Therefore the probability distribution of $X$ is
\begin{equation}
\Pi(X, x_0) = N(X) \delta(X-x_0) + p(X) H(X - x_0)
\mathrm{,}
\label{eq:prob1}
\end{equation}
where $H(.)$ is Heaviside function equal to 1 if its argument is positive and zero otherwise. The probability distribution $\Pi(X, x_0)$ is normalised in such a way that
$$ \int_0^{\infty} {\mathrm{d}}X \Pi(X, x_0) = 1 \mathrm{,}$$ a consequence of the property $N(\infty) = 1$.
Suppose now that $x_0$, instead of being taken as a fixed number is drawn at random with a probability distribution $q(x_0)$. Therefore the probability distribution of the maximum of $x$ and $x_0$ has to be averaged over the choices of $x_0$. This yields
\begin{equation}
P(X) = \int_0^{\infty} {\mathrm{d}}x_0 q(x_0) \Pi(X, x_0) = N(X) q(X) + p(X)\int_0^X {\mathrm{d}}x' q(x')
\mathrm{,}
\label{eq:prob2}
\end{equation}
One can check by performing the integrals in the quadrant $x, x' >0$ that $$\int_0^{\infty} {\mathrm{d}}X P(X) = \int_0^{\infty} {\mathrm{d}}x p(x) \int_0^{\infty} {\mathrm{d}}x' q(x') = 1\mathrm{.}$$
From equation (\ref{eq:prob2}) one can derive the probability distribution of the largest value drawn after $\nu$ (integer) independent trials , each one with the probability distribution $p(x)$. Let $P_{\nu}(x)$ be the probability distribution of the maximum of $\nu$ trials. After one trial $P_{1}(X) = p(X)$. From equation (\ref{eq:prob2}) one derives the recursion formula between $P_{\nu}(X)$ and $P_{\nu+1}(X)$:
\begin{equation}
P_{\nu+1}(X) = N(X) P_{\nu}(X) + p(X)\int_0^X {\mathrm{d}}x' P_{\nu}(x')
\mathrm{,}
\label{eq:prob3}
\end{equation}
Define now $Q_{\nu}(X) = \int_0^X {\mathrm{d}}x' P_{\nu}(x')$. This allows to write equation (\ref{eq:prob3}) like:
\begin{equation}
\frac{{\mathrm{d}}Q_{\nu+1}(X)}{{\mathrm{d}}X}= N(X) \frac{{\mathrm{d}}Q_{\nu}(X)}{{\mathrm{d}}X} + \frac{{\mathrm{d}}N(X)}{{\mathrm{d}}X} Q_{\nu}(X)
\mathrm{,}
\label{eq:prob4}
\end{equation}
This can be obviously integrated as
$$Q_{\nu+1}(X) = N(X) Q_{\nu}(X) + S_{\nu} \mathrm{,}$$
where $S_{\nu}$ is a constant of integration, independent on $X$. Because $Q_{\nu}(0) = 0 $ for all $S_{\nu}$, $S_{\nu} = 0$ also for all ${\nu}$. Therefore
\begin{equation}
Q_{\nu} (X)= \left (\int_0^X {\mathrm{d}}x' p(x') \right)^{\nu}
\mathrm{,}
\label{eq:prob5}
\end{equation}
and
\begin{equation}
P_{\nu} (X)= \nu p(X) \left (\int_0^X {\mathrm{d}}x' p(x') \right)^{\nu -1}
\mathrm{.}
\label{eq:prob6}
\end{equation}
Suppose $p(x)$ is a smooth function decaying continuously to zero as $x$ tends to infinity. In this case it is possible to get the asymptotic form of $P_{\nu} (X)$ at $\nu$ very large. Let us write $P_{\nu} (X)$ as an exponential
$$ P_{\nu} (X)= e^{T(\nu, X)} \mathrm{.}$$
with
$$ T(\nu, X) = \ln(\nu) + \ln(p(X)) + (\nu -1) \ln\left(\int_0^X {\mathrm{d}}x' p(x') \right) \mathrm{.}$$
In the limit $\nu$ large, one expects that the distribution $ P_{\nu} (X)$ has more and more weight at larger and larger values of $X$, which is also what is found by looking numerically at the shape of $ P_{\nu} (X)$ in this limit for various possible $p(X)$. See Figs. \ref{fig1} and \ref{fig2}. Therefore, in this limit, $P_{\nu} (X)$ should become more and more concentrated around the value of $X$ such that the derivative $\frac{\partial T(\nu, X)}{\partial X} = 0$. This derivative vanishes when $X$ is the root $X_{\nu}$ of
$$ \nu = 1 - \frac{p'\cdot N}{p^2} \mathrm{,}$$ where $p' = \frac{dp}{dX}$.
When $X_{\nu}$ is large, then $N(X_{\nu})=\int_0^{X_{\nu}} {\mathrm{d}}x' p(x') \approx 1$.
At $\nu$ large, this root $X_{\nu}$ is unique and large. This can be seen by noticing that $ - \frac{p'}{p^2} = \frac{d(1/p)}{dX} $, and by assuming that $1/p$ is a smooth function increasing monotonically to infinity as $x$ tends to infinity. To make its first momentum $m_1$ convergent $p(X)$ must decay faster than $x^{-2}$ at infinity, so that the derivative $ \frac{d(1/p)}{dX} $ must grow faster than $X$ at $X$ large. Therefore the function $X_{\nu}$ grows slower than $\nu$ as $\nu$ tends to infinity but it grows to infinity for any function $p(x)$ tending smoothly to zero as $x$ tends to infinity. This growth will depend on the behaviour of $p(x)$ as $x$ tends to infinity.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.5]{rich1}
\end{center}
\caption{$P_{\nu}(X)$ for different $\nu$ when $p(x)=e^{-x}$. Observe the monotonic increasing of $X_{\nu}$ with $\nu$. For this case, when $\nu\gg 10$, observe that $P_{\nu}(X_{\nu})$ is constant and $P_{\nu}(X)$ presents a soliton-like waveform.}
\label{fig1}
\end{figure}
The function $X_{\nu}$ gives the order of magnitude of the maximum wealth after $\nu$ iterations. By continuing the expansion of $ T(\nu, x)$ near $X_{\nu}$ to the quadratic order with respect to the difference $\delta X = X - X_{\nu}$,
one finds that
$$ T(\nu, X) \approx T(\nu, X_{\nu}) + \frac{\delta X^2}{2} \frac{\partial^2 T(\nu, X)}{\partial X^2} +...\mathrm{.}$$
where the second derivative is computed at $X = X_{\nu}$.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.5]{rich2}
\end{center}
\caption{$P_{\nu}(X)$ for different $\nu$ when $p(x)=xe^{-x}$. Observe the monotonic increasing of $X_{\nu}$ with $\nu$. For this case, when $\nu\gg 10$, observe that $P_{\nu}(X_{\nu})$ is not constant and $P_{\nu}(X)$ presents an increase of the maximal wave amplitude.}
\label{fig2}
\end{figure}
After some algebra and by taking into account that $\nu$ is large and that, in the limit $X_{\nu}$ large, $N(X_{\nu}) \approx 1$, one finds
$$ \frac{\partial^2 T(\nu, X)}{\partial X^2} \approx - p( X_{\nu}) \frac{d^2 (1/p)}{dX^2}\mathrm{.}$$
To prove that the width of the maximum of the distribution is much less than $X_{\nu}$, one can do the following approximate scaling argument. We have $\frac{p'}{p^2} \approx - \nu$. Assuming that $p' \approx \frac{p(X_{\nu})}{X_{\nu}}$, which is certainly correct for a probability distribution $p(.)$ decaying like a power law at large arguments, one finds $\nu \sim \frac{1}{ X_{\nu} p( X_{\nu})}$. Using the same kind of scaling argument one finds that
$$ \frac{\partial^2 T(\nu, X)}{\partial X^2} \sim \frac{-\nu p(X_{\nu})}{X_{\nu}} \sim -\frac{1}{X_{\nu}^2} \mathrm{.}$$
This shows that, at least for distributions $p(x)$ decaying like power laws, the width of the probability distribution $ P_{\nu} (X)$ is of order $X_{\nu}$ for $\nu$ very large, although its center is at $X_{\nu}$. In this case the width of the probability distribution and its center are large and of the same order of magnitude. Therefore one may guess that it behaves like
$$ P_{\nu} (X) \approx \frac{1}{X_{\nu}} \hat{P}\left(\frac{X}{X_{\nu}}\right) \mathrm{.}$$ where $\hat{P}$ is a positive numerical function of order one when its argument is of order one. It is normalized in such a way that $\int_0^{\infty}
\hat{P}(z) {\mathrm{d}} z = 1$. From the derivation, this function depends on the way $p(x)$ behaves as $x$ tends to infinity.
\section{Conclusions and perspectives}
\label{sec:concper}
Thanks to its mathematical structure the Z-model can be solved and somehow extended to bring interesting results with, perhaps, a connection to the complicated phenomenology of real economics. Despite its strongly nonlinear character it can be solved without assuming too many things. A remarkable feature of this model is its convergence to an exponential distribution of wealth. Of course any difference between reality and this model may have many explanations. Among others, it has been suggested, such as one of us (YP) also suggested it during the Noma-13 conference, that this model lacks an important element present in economies of developed countries, the tax system, with a more or less explicit claim of redistributing the wealth. Such a tax system could be perhaps represented by adding a third partner in each binary transaction, taking its pound of flesh at the transaction and redistributing it randomly at the next step, more or less the way the VAT tax (added value tax) works. This paper introduces also a modified Z-model, where at each transaction money is exchanged which is not actually possessed by the economic agents, something occuring all the time in modern economies. Amazingly this induces an instability in the distribution of wealth and makes grow indefinitely the higher momenta of its distribution, even though the total amount remains the same. Although this happens in a very idealised model, it could be closer to reality than the original Z-model with its rather narrow distribution of wealth.
\date{\today }
|
\section{I. Introduction}
Orbital degrees of freedom play an important role in solid-state systems. A prominent example is
the intriguing physics of heavy-fermion compounds that emerges from the interplay between dispersive
conduction-band orbitals on the one hand and strongly localized orbitals, with a large effective mass
and strong Coulomb interactions, on the other \cite{Hewson97,Coleman07,SiSteglich10,Gulacsi04}. However,
in systems of ultracold atoms in optical lattices \cite{BlochDalibardZwerger08,LewensteinSanperaAhufinger}
orbital degrees of freedom, spanning Bloch bands above a large energy gap, are typically frozen out, at
least in the interesting deep-lattice tight-binding regime where interactions are strong. Here, we
investigate the possibility to coherently open on-site orbital degrees of freedom in a strongly
interacting optical lattice system by means of near-resonant lattice shaking. We consider spinless
bosons in one dimension (1D) and show how to realize a ``dressed-lattice'' system, where effectively at
every lattice site the strongly localized ground-band orbital is nearly degenerate and coupled to the
much more dispersive first-excited-band orbital. The tunneling matrix elements of the two orbitals
differ strongly in magnitude and also in sign, with the latter leading to kinetic frustration. We
predict an orbital-driven phase transition between a Mott insulator (MI) and a superfluid (SF) state
when the population of the light orbitals is adiabatically increased by lowering the interorbital
detuning. As a consequence of frustration and strong interorbital interactions, this transition is found
to be either continuous or first-order, depending on parameters such as filling or lattice depth.
In contrast to the present proposal, in previous experiments atoms were transferred non-adiabatically to
excited bands of optical lattices by different methods \cite{GemelkeEtAl05,MuellerEtAl07,
SiasEtAl07,WirthEtAl11, OelschlaegerEtAl11,BakrEtAl11}. Moreover, lattice shaking has recently been
employed for band-coupling in the weakly interacting regime, where condensation into two possible
momentum states led to domain formation \cite{ParkerEtAl13}. Such band-coupling has been studied
theoretically for non/weakly interacting particles and isolated sites \cite{ArlinghausHolthaus11,
ArlinghausHolthaus12,Sowinski12,ZhangZhou14,ZhengEtAl14,ChoudhuryMueller14,DiLibertoEtAl14}. Also
orbtial coupling via magnetic resonances has been proposed \cite{PietraszewiczEtAl12} and there has been
theoretical interest in the physics of excited orbitals not involving lower-lying states
\cite{IsacssonGirvin05,WuEtAl06,Wu08a,Wu08b,LiEtAl12a,LiEtAl12b,PinheiroEtAl13}. Finally, the
perturbative admixture of excited orbitals has been studied
in theory \cite{LiEtAl06,LuehmannEtAl08,SchneiderEtAl09, Buechler10,HazzardMueller10, DuttaEtAl11,
LackiZakrzewski13, DuttaEtAl14} and experiment \cite{CampbellEtAl06,BestEtAl09,WillEtAl10,MarkEtAl11,
HeinzeEtAl11,JuergensenEtAl14}.
\section{II. Realizing the two-orbital model}
Consider spinless bosonic atoms of mass $m$ in an optical lattice
$V({\bm r})=V_0\sin^2(k_L x)-V_1\sin^2(2k_Lx)+V_\perp[\sin^2(k_L y)+\sin^2(k_L z)]$. The $y$ and $z$
directions are frozen out by a deep lattice, we will assume $V_\perp=30E_R$, such that an array of
1D tubes with a dimerized lattice [Fig.~\ref{fig:lattice}(a)] is created. The recoil energy
$E_R=\hbar^2k_L^2/2m$ is needed to localize a particle on a lattice constant $\pi/k_L$. For Rb$^{87}$ a
typical wave length of $2\pi/k_L= 852$ nm gives $E_R=2\pi\hbar\cdot 3.16$ kHz.
Each 1D tube is described by the multi-band Bose-Hubbard Hamiltonian
$\hat{H}_0=\hat{H}_\text{kin}+\hat{H}_\text{os}$, where
\begin{eqnarray}\label{eq:H0}
\hat{H}_\text{kin} &=&
-\sum_\ell\sum_\alpha (-1)^\alpha J_\alpha (\hat{b}^{\dag}_{\alpha (\ell+1)}\hat{b}^{\phantom\dag}_{\alpha \ell} +\text{h.c.}),
\\
\hat{H}_\text{os}&=& \sum_\ell\bigg[\sum_\alpha \epsilon_\alpha \hat{n}_{\alpha \ell}
+\sum_{\{\alpha\}} \frac{U_{\{\alpha\}}}{2}
\hat{b}^{\dag}_{\alpha_1\ell}\hat{b}^{\dag}_{\alpha_2 \ell}\hat{b}^{\phantom\dag}_{\alpha_3\ell}\hat{b}^{\phantom\dag}_{\alpha_4\ell}\bigg].
\end{eqnarray}
Here $\hat{b}^{\dag}_{\alpha \ell}$ and $\hat{n}_{\alpha \ell}$ are the bosonic creation and number operator for a
Wannier orbital $w_\alpha(x-x_\ell)$ of Bloch band $\alpha=0,1,\cdots$, localized at
$x_\ell=\ell\pi/k_L$
\cite{*[{Since we consider a weak dimerization only, $V_1<V_0$, we do not need to employ
generalized Wannier orbitals, localized in the left and right minimum of
each double well [Fig.~\ref{fig:lattice}(a)], as in }] [] VaucherEtAl07}.
The band-center energies and tunnel parameters fulfill
$0\equiv\epsilon_0<\epsilon_1<\cdots$ and $0<J_0<J_1<\cdots$, respectively. The interaction strengths
$U_{\{\alpha\}}\equiv U_{\alpha_1 \alpha_2 \alpha_3 \alpha_4}
=(2\hbar^2a_sa_\perp^2/m)\int\!\mathrm{d} x\, w_{\alpha_1}(x)w_{\alpha_2}(x)w_{\alpha_3}(x)w_{\alpha_4}(x)$
vanish for odd $\sum_i\alpha_i$, since $w_\alpha(x)=(-)^\alpha w_\alpha(-x)$, and depend on the
transverse localization length $a_\perp\simeq(V_\perp/E_R)^{1/4}/k_L$
and the scattering length $a_s$ ($\approx5.6$ nm for Rb$^{87}$).
\begin{figure}[t]
\includegraphics[width=\linewidth]{figure1}
\caption{(color online)
(a) Dimerized lattice with $V_1/V_0=0.5$.
(b) Impact of $V_1/V_0$ on $J_2/J_1$, $\Delta_{10}$, and $\Delta_{21}$, for $V_0/E_R=10$.
(c) Most dominant loss channels, scattering into third band (zero-photon process) and
driving-induced coupling (single-photon process), are off-resonant due to dimerization.
(d) Lattice parameters versus $V_0/E_R$ for $V_1/V_0=0.5$ and Rb$^{87}$.}
\label{fig:lattice}
\end{figure}
In the tight-binding regime, $\epsilon_1$ is typically much larger than the temperature and the chemical
potential so that the orbital degree of freedom $\alpha$ is frozen out. We wish to coherently open this
freedom by means of time periodic forcing with near-resonant frequency $\hbar\omega\approx\epsilon_1$.
In particular, the lowest band ($\alpha=0$) shall be coupled to the more dispersive first excited band
($\alpha=1$), without creating coupling to even higher-lying bands ($\alpha\ge2$). In order to achieve
such a controlled situation---also in the regime where interactions are strong compared to tunneling---we
combine two strategies: First we choose a driving scheme, namely sinusoidally shaking the lattice back and
forth, that for weak forcing does not lead to multi-``photon'' interband transitions at resonances
$\Delta_{\alpha'\alpha}\equiv\epsilon_{\alpha'}-\epsilon_{\alpha}\approx m \hbar\omega$ with
integer $|m|\ge2$. Second, we engineer the band structure by varying $V_1/V_0$ such that transitions to
band 2 remain off resonant: With increasing $V_1/V_0$ the bands organize in pairs (0,1), (2,3), \ldots
such that $\Delta_{10}$ and $\Delta_{32}$ as well as $J_1/J_0$ and $J_3/J_2$ decrease, while
$\Delta_{21}$ increases. For $V_0=10E_R$, already a slight dimerization $V_1/V_0=0.5$ ensures that
$\Delta_{10}\approx3.7 E_R$ is noticeably smaller than $\Delta_{21}\approx 5.3 E_R$, rendering the
$\Delta_{21}$ transition off-resonant when $\hbar\omega \approx \Delta_{10}=\epsilon_1$. At the same
time $V_1/V_0=0.5$ is small enough to keep a relatively large ratio $J_1/J_0\approx 5.7$, retaining the
desired feature that $\alpha=0$ particles are much less dispersive than $\alpha=1$
particles [Fig.~\ref{fig:lattice}(b,d)].
By moving the lattice like $K k_L/(\pi m\omega^2)\cos(\omega t)$ in $x$ direction, an inertial
force is created, described by
\begin{equation}\label{eq:Hdr}
\hat{H}_\text{dr}(t) =K\cos(\omega t) \sum_\ell \bigg[\sum_{\alpha} \ell \hat{n}_{\alpha \ell}
+\sum_{\alpha'\alpha} \eta_{\alpha'\alpha}\hat{b}^{\dag}_{\alpha'\ell}\hat{b}^{\phantom\dag}_{\alpha \ell} \bigg].
\end{equation}
Here $\eta_{\alpha'\alpha}=(k_L/\pi)\int\!\mathrm{d} x\, w_{\alpha'}(x) x w_{\alpha_\alpha}(x)$ vanishes for
even $\alpha'+\alpha$. We employ a time-periodic unitary transformation
$\hat{U}(t)=\exp(-i\sum_{\ell,\alpha}\hat{n}_{\alpha_\ell}\nu_\alpha\omega t)$ with
integers $\nu_\alpha$, designed to shift all band energies $\varepsilon_\alpha$ to values
$\epsilon_\alpha'\equiv\epsilon_\alpha-\nu_\alpha\hbar\omega\in(-\hbar\omega/2, \hbar\omega/2]$ that are
as close as possible to $\varepsilon_0=\varepsilon_0'=0$. This gives
$|\epsilon_1'|=|\epsilon_1-\hbar\omega|\ll\hbar\omega$ by choice of $\omega$ and
$|\epsilon_2'|=|\epsilon_2-2\hbar\omega|\sim\hbar\omega/2$ by choice of $V_1/V_0$; all other
$\epsilon_\alpha'$ are scattered somehow between $-\hbar\omega/2$ and $\hbar\omega/2$.
The periodic time dependence of the transformed Hamiltonian
$\hat{H}(t)=\hat{U}^\dag(\hat{H}_0+\hat{H}_\text{dr})\hat{U}-i\hbar\hat{U}^\dag\mathrm{d}_t\hat{U}$ appears in the interband coupling
parameters $K\eta_{\alpha'\alpha}\cos(\omega t) \mathrm{e}^{i(\nu_{\alpha'}-\nu_\alpha)\omega t}$ and
$U_{\{\alpha\}}\mathrm{e}^{i(\nu_{\alpha_1}+\nu_{\alpha_2}-\nu_{\alpha_3}-\nu_{\alpha_4})\omega t}$.
For weak forcing $K\ll\hbar\omega$ the driving frequency $\hbar\omega\sim\Delta_{10}$ is large compared
to the intraband terms as well as to the band coupling [Fig.~\ref{fig:lattice}(d)]. This allows to
average the rapidly oscillating terms in the Hamiltonian over one driving period and to approximately
describe the system by the effective time-independent Hamiltonian
$\hat{H}_\text{eff}=\frac{1}{T}\int_0^T\!\mathrm{d} t\,\hat{H}(t)$, reading
\begin{eqnarray}\label{eq:Hmulti}
\hat{H}_\text{eff}&=& \hat{H}_\text{kin}
+\sum_\ell \bigg[\sum_{\alpha}\epsilon_\alpha'\hat{n}_{\alpha \ell}
+\, K \sum_{\alpha'\alpha} \eta'_{\alpha'\alpha}
\hat{b}^{\dag}_{\alpha'\ell}\hat{b}^{\phantom\dag}_{\alpha \ell}
\nonumber\\&&
+\, \sum_{\{\alpha\}} \frac{U'_{\{\alpha\}} }{2}
\hat{b}^{\dag}_{\alpha_1\ell}\hat{b}^{\dag}_{\alpha_2 \ell}\hat{b}^{\phantom\dag}_{\alpha_3\ell}\hat{b}^{\phantom\dag}_{\alpha_4\ell}\bigg],
\end{eqnarray}
with $\eta'_{\alpha'\alpha}=\eta_{\alpha'\alpha}(\delta_{\nu_{\alpha'},\nu_{\alpha}+1}
+ \delta_{\nu_{\alpha'},\nu_{\alpha}-1})/2$ and $U'_{\{\alpha\}}=U_{\{\alpha\}}
\delta_{\nu_{\alpha_1}+\nu_{\alpha_2},\nu_{\alpha_3}+\nu_{\alpha_4}} $. For a more systematic derivation
of Eq.~(\ref{eq:Hmulti}), $\hat{H}_\text{eff}$ is defined as the generator of the time evolution over one
period \cite{Shirley65} and computed using degenerate perturbation theory in the extended Floquet
Hilbert space \cite{Sambe73}, similar like in Refs.~\cite{EckardtHolthaus07,EckardtHolthaus08b}. In
leading order one recovers Eq.~(\ref{eq:Hmulti}). The leading correction contains tiny second-order
coupling to bands $\alpha\ge3$ of order $c^2/\hbar\omega$ to be neglected, where $c\lesssim 0.1E_R$ is
a typical interband coupling matrix element and $\hbar\omega\gtrsim3E_R$.
It is a crucial property of lattice shaking (\ref{eq:Hdr}) that in $\hat{H}_\text{eff}$ the interband
coupling $K\eta_{\alpha'\alpha}'$ is a single-photon process, with $\nu_{\alpha'}=\nu_\alpha\pm1$, and
that scattering $U_{\{\alpha\}}'$ is a zero-photon process, with
$\nu_{\alpha_1}+\nu_{\alpha_2} = \nu_{\alpha_3}+\nu_{\alpha_4}$. No multi-photon processes are found for weak
driving. Thus, in $\hat{H}_\text{eff}$ above the bands 0 and 1 are coupled to band 2 only,
via the processes sketched in Fig.~\ref{fig:lattice}(c). These processes are, however, off-resonant, since
$\epsilon_2'\sim\hbar\omega$. The bands 0 and 1 are, therefore, to good approximation isolated and
described by the two-band (2B) model
\begin{eqnarray}\label{eq:Heff}
\hat{H}_{2B} &=&\hat{H}_\text{kin} +\sum_{\ell}\Big[\delta \hat{n}_{1\ell}
-\gamma(\hat{b}^{\dag}_{1\ell}\hat{b}^{\phantom\dag}_{0\ell}+\text{h.c.})+ 2U_{10}\hat{n}_{0\ell}\hat{n}_{1\ell}
\nonumber\\&&
+\,\frac{U_{00}}{2}\hat{n}_{0\ell}(\hat{n}_{0\ell}-1)+\frac{U_{11}}{2}\hat{n}_{1\ell}(\hat{n}_{1\ell}-1)
\Big],
\end{eqnarray}
where $\gamma=-K\eta_{10}/2$, $U_{\alpha'\alpha}\equiv U_{\alpha\alpha'\alpha'\alpha}$ and
$\delta=\Delta_{10}-\omega$. For $V_0=10E_R$, $V_1/V_0=1/2$, and $K=0.5E_R$ we obtain
$J_0\approx0.030E_R$, $J_1\approx0.17E_R$, $\gamma\approx0.13E_R$, $U_{00}\approx0.58E_R$,
$U_{01}\approx0.36E_R$, and $U_{11}\approx0.50E_R$.
\begin{figure}[t]
\includegraphics[width=\linewidth]{figure2}
\caption{ (color online)
(a) Sketch of the effective model, grey (white) circles correspond to the $\alpha = 1$ (0) state at each
site $\ell$, with energy $\delta$ (0).
(b) Ground-state compressibility $\partial_\mu n E_R$ in the $\mu$-$\delta$ plane for $V_0/E_R=10$ and
$V_1/V_0=1/2$.
(c) Correlations $\chi^{\ell}_{\alpha'\alpha}\equiv \langle\hat{b}^{\dag}_{\alpha'\ell}\hat{b}_{\alpha 0} \rangle /
\sqrt{n_{\alpha'}n_\alpha}$ and imbalance $n_0-n_1$ for fixed filling $n=1$ versus $\delta$,
with $V_1/V_0 =0.5$ and $V_0/E_R=15, 10, 5$ (from top to bottom). Numerical data in (b) and (c) obtained
for $M=30$ rungs under periodic boundary conditions using TEBD in imaginary time \cite{TEBD,Vidal04},
with bond dimensions 14 (b) and 30 (c).}
\label{fig:phases}
\end{figure}
$\hat{H}_{2B}$ describes a highly tunable 1D ladder system [Fig.~\ref{fig:phases}(a)] with interesting
properties: The tunneling matrix elements along both legs
(i.e.\ in both bands) differ in sign and magnitude. The former leads to maximal kinetic frustration with
a flux of $\pi$ per plaquette \cite{Honerkamp03,HottaFurukawa06,EckardtEtAl10,StruckEtAl11,GrescherEtAl13,
TielemanEtAl13,DharEtAl13,YudinEtAl14}. The latter
makes leg 0 more prone to localization than leg 1. The hybridization of both legs is controlled by the
energy separation $\delta$ and the coupling $\gamma$, which can be tuned via the frequency and strength
of the driving, respectively. Finally, the system features strong interorbital interactions $U_{10}$,
with two-particle energies $2U_{01}>U_{00}>U_{11}$.
In order to investigate the 2B model (\ref{eq:Heff}), the following experimental protocol can be
pursued. After the system is prepared in (or close to) the undriven ground state, populating band 0, the
driving strength $K$ is ramped up smoothly to the desired value. During this step the detuning
$\delta$ is still large enough to suppress any significant occupation of band 1. Then, the orbital
freedom is opened by smoothly lowering $\delta$.
\section{III. Orbital-driven Mott transition}
We study the ground state of $\hat{H}_{2B}$ versus $\delta$. For large positive
(negative) $\delta/|\gamma|$ only leg 0 (leg 1) will be occupied; the system effectively reduces to a
1D Bose-Hubbard chain. For integer filling of $n$ particles per site, the ground state of such a chain
with tunneling $J$ and on-site repulsion $U$ is a gapped (i.e.\ incompressible) MI with localized
particles if $J/U<(J/U)^{(n)}_c$, where $(J/U)^{(1)}_c\approx0.26$ \cite{ElstnerMonien99}. Otherwise, it
is a gapless SF with quasilong-range order. Thus, for $n=1$ the system is a MI for
$\delta\gg|\gamma|$, since $J_0/U_{00}\approx 0. 051$, and a SF for $-\delta\gg|\gamma|$, since
$J_1/U_{11}\approx 0.34$.
It is instructive to control $n$ via the chemical potential $\mu$, introduced by adding
$-\mu\sum_\ell(\hat{n}_{0\ell}+\hat{n}_{1\ell})$ to $\hat{H}_{2B}$. In Fig.~\ref{fig:phases}(b), we plot the
ground-state compressibility $\partial_\mu n$ in the $\mu$-$\delta$ plane, computed by time-evolving
block decimation (TEBD) in imaginary time \cite{TEBD,Vidal04}. As expected \cite{ElstnerMonien99}, for
$\delta/|\gamma|\gg 1$ we find incompressible MI phases at integer filling $n$, interrupted by SF phases
where $n$ changes in $\mu$
direction, while for $-\delta/|\gamma|\gg 1$ the system is a compressible SF.
When $\delta$ is lowered, the filling $n_0$ ($n_1$) of leg 0 (1) decreases (increases). In response,
an orbital-driven transition occurs, either between a MI and a SF or, for fractional filling, between
different SFs. For the given parameters, these are first-order transitions, except at the tip of the $n=1$
Mott phase, where a continuous transition is found. The discontinuous SF-to-SF transition, where the
ground state changes abruptly, happens when near $\delta=2(J_0-J_1)\approx-0.28 E_R$ a boson suddenly
prefers to delocalize with quasimomentum $\pi$ in leg 1, rather than with quasimomentum 0 in leg 0. The
discontinuous MI-to-SF transition, to be explained below, is more subtle.
A strong-coupling argument explains the orbital-driven MI-to-SF transition. Within the MI state, $n_1$
increases smoothly when $\delta$ is lowered, and the larger $n_1$ the larger is the reduction of kinetic
energy $\approx 2J_1(n_1+1)$ (or $\approx 2J_1n_1$) that a particle (or a hole) acquires by delocalizing
along leg 1 on the MI background. When the kinetic energy reduction of a particle-hole excitation exceed
its interaction-energy cost $\approx (U_{11}+2U_{01}n_0\delta_{n,1})$, these excitations proliferate and
the ground state becomes a SF as seen in Fig.~\ref{fig:phases}(b). This transition can also be observed
in Fig.~\ref{fig:phases}(c,middle) where we plot $n_0-n_1$ and the ground-state correlations
$\chi^{\ell}_{\alpha'\alpha}\equiv\langle\hat{b}^{\dag}_{\alpha'\ell}\hat{b}^{\phantom\dag}_{\alpha0}\rangle/\sqrt{n_{\alpha'}n_\alpha}$
versus $\delta$, for the same parameters and sharp filling $n=1$
\footnote{Experimentally $n_0$ and $n_1$ and the Fourier transforms of $n_0\chi^{\ell}_{00}$ and
$n_1\chi^{\ell}_{11}$ can be measured via band mapping \cite{MuellerEtAl07}.}.
While $|\chi_{11}^{\ell}|$ decays exponentially with $\ell$ in the MI phase, in the SF regime the decay
is only algebraic. Therefore, the transition is indicated by a significant increase of correlations on
longer distances such as $\chi^8_{\alpha\alpha}$. It is found near $\delta= 0.2E_R$, in fair agreement
with the above estimate giving $n_0-n_1\approx(8J_1-U_{11}-2U_{01})/(4J_1+2U_{01})\approx 0.10$.
In Fig.~\ref{fig:phases}(c, middle) we can identify three different types of MI states, characterized by
respective signs $s_\alpha\equiv\mathrm{sgn}(\chi^1_{\alpha\alpha})$ of the short-range correlations along both
legs. This is a consequence of kinetic frustration; while the tunneling matrix elements $-J_0$ and $J_1$
favor $s_0=+1$ and $s_1=-1$, the rung coupling $-\gamma$ favors $s_0=s_1$. We use the label MI$_{01}$ if
both legs retain their favored correlations ($s_0=-s_1=+1$), and MI$_\alpha$ if leg $\alpha$
dominates the other one [$s_0=s_1=+1$ (-1) for $\alpha = 0$ (1)]; similar labels are used for SF
states. Due to the strong interleg interactions $2U_{01}>U_{00},U_{11}$ the system does not feature the
chiral time-reversal symmetry broken MI or SF ground states with complex $\chi^\ell_{\alpha'\alpha}$
predicted in Ref.~\cite{DharEtAl13}. Treating both $\gamma$ and the $J_\alpha$ as perturbation the
MI$_0$-to-MI$_{01}$ transition is predicted to occur at
$\delta=U_{00} J_1/(2J_0) \approx1.6$ \cite{Supplemental}.
Experimentally, this transition is hardly observable, since it occurs at tiny
$n_1\simeq|\gamma/\delta|^2\approx 0.007$. The MI$_{01}$-to-MI$_1$, happening when
$\delta$ is lowered further, is of greater importance. A perturbative treatment of the tunneling matrix
elements $J_\alpha$, neglecting $\gamma$ and $\delta$ on interaction-dominated doubly occupied sites,
predicts this transition to occur when
$n_0-n_1\approx[(J_1-J_0)U_{00}-4J_0U_{01}]/[(J_1-J_0)U_{00}+4J_0U_{01}]\approx0.31$ \cite{Supplemental},
in reasonable agreement with the numerics.
These transitions can be observed also in a deeper lattice, where the system remains a MI for small
$\delta$ [Fig.~\ref{fig:phases}(c,top)].
For a lower lattice depth of $V_0/E_R=5$, the MI-to-SF transition occurs earlier and already within the
MI$_{01}$ regime [Fig.~\ref{fig:phases}(c,bottom)]. This is explained by the above estimates that
predict the MI-to-SF transition to occur when $n_0-n_1\approx 0.94$, well before the estimated value
$n_0-n_1\approx 0.23$ for the MI$_{01}$-to-MI$_1$ transition is reached. As a consequence, the MI-to-SF
transition is rendered discontinuous. The discontinuity results from an abrupt change in the structure
of the short-range correlations along leg 0. Namely, the SF phase is of SF$_1$ type, with
$s_0=s_1=-1$, such that the short-range correlations along leg 0 have to undergo a finite jump at the
MI$_{01}$-to-SF$_1$ transition.
The same argument also explains the first-order nature of the orbital-driven MI-to-SF transitions for
$V_0/E_R=10$ at filling $n\ge2$, visible as a sharp jump of the compressibility in
Fig.~\ref{fig:phases}(b).
All in all, the fact that the orbital-driven MI-to-SF transition can be discontinuous results from the
combination of kinetic frustration, tunneling imbalance $J_1\gg J_0$, and strong interband interactions
$U_{01}$, all stemming from the spatial structure of the Wannier orbitals.
\section{IV. Preparation dynamics and heating}
\begin{figure}[t]
\includegraphics[width=\linewidth]{figure3}
\caption{ (color online)
Real-time (RT) evolution of the effective model (\ref{eq:Hmulti}) with three bands (3B) $\alpha=0,1,2$, obtained by TEBD \cite{TEBD,Vidal04} with bond dimensions 24 (a), 22 (b), and 22-40 (c).
(a) Occupations and overlaps with the [imaginary time-evolved (IT)] ground states of the 3B and
2B model. Starting in the ground state at $\delta=\epsilon_1-\omega=1E_R$, $\delta$ is lowered
linearly to $\delta=-0.5E_R$ within a time of $T_r=500 \hbar/E_R$; for $n=1$, $V_0/E_R=10$, $V_1/V_0=0.5$,
$K/E_R=0.5$ and $M=16$ rungs under periodic boundary condtions.
(b) After-ramp overlap for $M=10$, like in (a) but varying $T_r$ and $K$.
(c) After-ramp overlap for $M=10$, like in (a), but varying $V_1/V_0$.}
\label{fig:prep}
\end{figure}
When $\delta$ is lowered slow enough, the system is expected to approximately follow the ground state
of the 2B model (\ref{eq:Heff}), unless the first-order transition is crossed. This desired dynamics is
effectively adiabatic \cite{EckardtHolthaus08b}, i.e.\ adiabatic with respect to $\hat{H}_{2B}$, but
diabatic with respect to tiny coupling matrix elements neglected in $\hat{H}_{2B}$. We have simulated the
time evolution of the system [Fig.~\ref{fig:prep}(a)] using TEBD \cite{TEBD,Vidal04}. For parameters
like in Fig.~\ref{fig:phases}(c,middle), $\delta/E_R$ is ramped from $1$ to $-0.5$ within a time
$T_r=500\hbar/E_R\approx 25$ms. In order to probe the validity of the 2B model (\ref{eq:Heff}), we
include the major ``loss'' processes depicted in Fig.~\ref{fig:lattice}(c) by employing Hamiltonian
(\ref{eq:Hmulti}) with three bands ($\alpha=$0,1,2). In Fig.~\ref{fig:prep}(a), one can clearly see that
the $\alpha=2$ occupation $n_2$ remains very low and that the overlap with the
instantaneous 2B ground state stays close to 1. Both clearly shows that the driving does not cause
detrimental heating and justifies a description of the driven system in terms of the 2B model
(\ref{eq:Heff}). It, moreover, indicates that an effectively adiabatic time evolution is possible,
despite a noticeable dip of the overlap at the Mott transition (resembling the behavior of Landau-Zener
sweeps \cite{LimBerry91}). Thus, the protocol allows for the preparation of stable low-entropy states in
an excited Bloch band.
The overlap plotted in Fig.~\ref{fig:prep}(b) versus $T_r$ and $K$ measures the effective adiabaticity.
Too small $T_r$ and $\gamma\propto K$ spoil the adiabatic dynamics within the 2B model and for too large
$K$ the coupling to band 2 becomes relevant. Moreover, for too large $K$ and $T_r$
slow second-order loss processes (not included) can occur. Finally, Fig.~\ref{fig:prep}(c) shows that
for strong interactions a minimal dimerization of $V_1/V_0=0.5$ is crucial. Different from the weakly
interacting case \cite{ParkerEtAl13}, we find significant transfer to the second excited band 2 for the
simple cosine lattice ($V_1/V_0=0$).
\section{V. Conclusion and outlook}
We have shown that lattice shaking is a feasible tool to coherently open on-site orbital degrees of
freedom in a strongly interacting optical lattice system and that the interplay between Wannier orbits
of different structure gives rise to rich physics already for spinless bosons in 1D.
Extending the scheme to spinful fermions, the interplay between strongly localized and dispersive orbital
states should permit to mimic aspects of the intriguing heavy-fermion physics and to realize
periodic-Anderson-like models \cite{Hewson97,Coleman07,SiSteglich10,Gulacsi04}. The extension to higher
dimensional lattices should provide a feasible scheme for the preparation of low-entropy states in
excited bands as they have been discussed before \cite{IsacssonGirvin05,WuEtAl06,Wu08a,Wu08b,LiEtAl12a,
LiEtAl12b,PinheiroEtAl13} and, moreover, to couple them to strongly localized lowest-band orbits.
Finally, by employing sufficiently off resonant forcing, keeping $\delta$ large enough, one might
enhance and control the perturbative admixtures of excited bands \cite{LiEtAl06,LuehmannEtAl08,
SchneiderEtAl09, Buechler10,HazzardMueller10, DuttaEtAl11, LackiZakrzewski13, DuttaEtAl14,CampbellEtAl06,
BestEtAl09,WillEtAl10,MarkEtAl11,HeinzeEtAl11,JuergensenEtAl14}, e.g.\ in order to enhance superexchange
processes by engineering density-dependent tunneling.
\begin{acknowledgments}
We thank Mikl{\'o}s Gul{\'a}csi for discussion. CS is grateful for support by the Studienstiftung des
deutschen Volkes.
\end{acknowledgments}
\section{Appendix A: Transition between MI$_0$ and MI$_{01}$}
For large negative $\delta$, one can treat both $\gamma$ and the tunneling matrix elements $J_\alpha$ as
a perturbation. For unit filling $n=1$ the unperturbed ground state takes the simple form
\begin{equation}
|\psi_0\rangle = \prod_\ell \hat{b}^{\dag}_{0\ell}|\text{vac}\rangle
\end{equation}
with the vacuum state $|\text{vac}\rangle$. A finite correlation $\langle\hat{b}^{\dag}_{10}\hat{b}^{\dag}_{11}\rangle$ will appear in
third order. Namely one has
\begin{eqnarray}
\langle\psi|\hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{11}|\psi\rangle
&\simeq& \langle\psi_1|\hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{11}|\psi_2\rangle+\langle\psi_2|\hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{11}|\psi_1\rangle
\nonumber\\
&=&2\langle\psi_1|\hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{11}|\psi_2\rangle
\end{eqnarray}
with $|\psi_k\rangle$ denoting the state correction appearing in $k$th order perturbation theory.
Here the relevant term of $|\psi_1\rangle$ reads
\begin{equation}
\frac{-\gamma \hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{00}}{-\delta}|\psi_0\rangle
= \frac{\gamma}{\delta} \,\hat{b}^{\dag}_{10}\prod_{\ell\ne0}\hat{b}^{\dag}_{0\ell}|\text{vac}\rangle,
\end{equation}
and the relevant term of $|\psi_2\rangle$ takes the form
\begin{eqnarray}&&
\bigg(\frac{\gamma\hat{b}^{\dag}_{11}\hat{b}^{\phantom\dag}_{01} \,J_0\hat{b}^{\dag}_{01}\hat{b}^{\phantom\dag}_{00}}{(2U_{01}+\delta)U_{00}}
-\frac{J_1\hat{b}^{\dag}_{11}\hat{b}^{\phantom\dag}_{10}\,\gamma\hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{00}}{(2U_{01}+\delta)\delta}
\bigg)|\psi_0\rangle
\nonumber\\&&
=\Big(\frac{2J_0}{U_{00}}-\frac{J_1}{\delta}\Big)\frac{\gamma}{2U_{12}+\delta}\,
\hat{b}^{\dag}_{11}\prod_{\ell\ne0}\hat{b}^{\dag}_{0\ell}|\text{vac}\rangle.
\end{eqnarray}
With that we arrive at
\begin{equation}
\langle\hat{b}^{\dag}_{10}\hat{b}^{\dag}_{11}\rangle \simeq
\Big(\frac{2J_0}{U_{00}}-\frac{J_1}{\delta}\Big)\frac{2\gamma^2}{(2U_{12}+\delta)\delta}
\end{equation}
leading to a sign change when both terms in the round bracket cancel each other. The change from
positive to negative sign corresponds to the transition from MI$_0$ to MI$_{01}$ that is thus expected
to occur at
\begin{equation}
\delta=\frac{U_{00}J_1}{2J_0}.
\end{equation}
\section{Appendix B: Transition between MI$_{01}$ and MI$_{1}$}
We assume sharp filling $n=n_0+n_1=1$ and treat the tunnel terms as a perturbation.
The unperturbed on-site problem is then given by
\begin{eqnarray}
\hat{H}_0 &=& \delta \delta \hat{n}_1 -\gamma (\hat{b}^{\dag}_1\hat{b}^{\phantom\dag}_0+\hat{b}^{\dag}_0\hat{b}^{\phantom\dag}_1) + 2U_{01}\hat{n}_0\hat{n}_1
\nonumber\\&&
+\,\frac{U_{00}}{2}\hat{n}_0(\hat{n}_0-1)+\frac{U_{11}}{2}\hat{n}_1(\hat{n}_1-1)
\end{eqnarray}
where we dropped the site index $\ell$. In the subspace of one particle on a site the unperturbed
ground state reads
\begin{equation}
|\psi^{(0)}\rangle= (a_0\hat{b}^{\dag}_0+a_1\hat{b}^{\dag}_1) |\text{vac}\rangle,
\end{equation}
with energy $\varepsilon_0=\frac{\delta}{2}-\frac{1}{2}[\delta^2+4\gamma^2]^{1/2}$ per site and
$a_1/a_0 = -\varepsilon_0/\gamma$, with $a_0^2+a_1^2=1$, giving in leading order perturbation theory
\begin{equation}\label{eq:occupations}
n_0\simeq a_0^2 \quad\text{and}\quad n_1\simeq a_1^2.
\end{equation}
In the course of the perturbation calculation we also need
defect states with one particle less (a hole) and one extra particle. The hole state is simply given by
the vacuum
\begin{equation}
|\psi^{(h)}\rangle= |\text{vac}\rangle,
\end{equation}
with energy $\varepsilon_h=0$.
The subspace with two particles on a site contains three states. For simplicity, we neglect the
hybridization coupling $\gamma$ and approximate the eigenstates with an additional particle by states
with sharp occupations of the orbitals $\alpha$,
\begin{eqnarray}
|\psi^{(p20)}\rangle &=& \frac{1}{\sqrt{2}}(\hat{b}^{\dag}_0)^2|\text{vac}\rangle,
\\
|\psi^{(p11)}\rangle &=& \hat{b}^{\dag}_0\hat{b}^{\dag}_1|\text{vac}\rangle,
\\
|\psi^{(p02)}\rangle &=& \frac{1}{\sqrt{2}}(\hat{b}^{\dag}_1)^2|\text{vac}\rangle,
\end{eqnarray}
with unperturbed energies $\varepsilon_{p20}=U_{00}$, $\varepsilon_{p11}=2U_{01}+\delta$, and
$\varepsilon_{p02}=2U_{11}+2\delta$.
Re-introducing the site index $\ell$ the unperturbed ground state reads
\begin{equation}
|\psi_0\rangle =\prod_\ell |\psi^{(0)}_\ell\rangle = \prod_\ell (a_0\hat{b}^{\dag}_{0\ell}+a_1\hat{b}^{\dag}_{1\ell})|\text{vac}\rangle.
\end{equation}
The correlation function between the neighboring sites $0$ and $1$ obtains a finite value in the first
order of the perturbation expansion with respect to tunneling
\begin{eqnarray}\label{eq:0corr}
\langle\psi|\hat{b}^{\dag}_{\alpha 0}\hat{b}^{\phantom\dag}_{\alpha 1}|\psi\rangle
&\simeq& \langle\psi_0|\hat{b}^{\dag}_{\alpha0}\hat{b}^{\phantom\dag}_{\alpha1}|\psi_1\rangle+\langle\psi_1|\hat{b}^{\dag}_{\alpha 0}\hat{b}^{\phantom\dag}_{\alpha1}|\psi_0\rangle
\nonumber\\
&=&2\langle\psi_0|\hat{b}^{\dag}_{\alpha 0}\hat{b}^{\phantom\dag}_{\alpha1}|\psi_1\rangle.
\end{eqnarray}
Here the relevant terms of the first-order state correction $|\psi_1\rangle$ possess an extra
particle in one of the three possible states on site 1 and a hole on site 0. These terms are related to
the perturbation $-J_0\hat{b}^{\dag}_{01}\hat{b}^{\phantom\dag}_{00}+J_1\hat{b}^{\dag}_{11}\hat{b}^{\phantom\dag}_{10}$ and read
\begin{eqnarray}
&&\bigg[
\frac{a_0^2J_0}{U_{00}-2\varepsilon_0}(\hat{b}^{\dag}_{01})^2
- \frac{a_1^2J_1}{U_{11}+2\delta-2\varepsilon_0}(\hat{b}^{\dag}_{11})^2
\nonumber\\&&
-\,\frac{a_0a_1(J_1-J_0)}{U_{01}+\delta-2\varepsilon_0}\hat{b}^{\dag}_{11}\hat{b}^{\dag}_{01}
\bigg]
\prod_{\ell\ne0,1}(a_0\hat{b}^{\phantom\dag}_0+a_1\hat{b}^{\phantom\dag}_1)|\text{vac}\rangle.
\end{eqnarray}
With this expression we obtain from Eqs.~(\ref{eq:0corr}) and (\ref{eq:occupations}) that
\begin{eqnarray}\label{eq:leg0corr}
\langle\hat{b}^{\dag}_{00}\hat{b}^{\phantom\dag}_{01}\rangle
&\simeq& \frac{2n_0}{U_{00}(2U_{01}+\delta-2\varepsilon_0)}
\Big[2n_0 J_0 (2U_{01}+\delta-2\varepsilon_0)
\nonumber\\&&
-\,n_1(J_1-J_0)(U_{00}-2\varepsilon_0)\Big]
\end{eqnarray}
and
\begin{eqnarray}
\langle\hat{b}^{\dag}_{10}\hat{b}^{\phantom\dag}_{11}\rangle
&\simeq& -\frac{2n_1}{(U_{11}+2\delta-2\varepsilon_0)(2U_{01}+\delta-2\varepsilon_0)}
\nonumber\\&&\times
\Big[2n_1 J_1 (2U_{01}+\delta-2\varepsilon_0)
\nonumber\\&&\quad
+\,n_0(J_1-J_0)(U_{11}+2\delta-2\varepsilon_0)\Big].
\end{eqnarray}
The transition from MI$_{01}$ to MI$_1$ is related to $\langle\hat{b}^{\dag}_{00}\hat{b}^{\phantom\dag}_{01}\rangle$ becoming negative.
Approximating $2U_{01}+\delta-2\varepsilon_0\approx 2U_{01}$, which is consistent with our previous
approximation to neglect $\gamma$ on doubly occupied sites, the transition is expected to occur when
\begin{equation}
\frac{n_1}{n_0}\approx\frac{4J_0U_{01}}{(J_1-J_0)U_{00}}
\end{equation}
or, equivalently, when
\begin{equation}
n_0-n_1\approx\frac{(J_1-J_0)U_{00}-4J_0U_{01}}{(J_1-J_0)U_{00}+4J_0U_{01}}.
\end{equation}
|
\section{Introduction}
Photoproduction of mesons provides important information about the excitation spectrum of
the nucleon that, despite various long-lasting experimental and theoretical efforts, is
still not sufficiently understood. The number of predicted states (see Review of Quark Model
in Ref.~\cite{PDG_12} exceeds the experimentally observed number and the properties of
some identified states are markedly different from those expected.
The difficulty of identifying the relevant degrees of freedom of Quantum Chromodynamics (QCD)
in the nonperturbative region using effective models can perhaps be solved in the future
by lattice calculations. Nevertheless, further precise experimental input is needed since
the majority of the available data comes from pion scattering experiments,
which could leave states that couple only weakly to $\pi N$ undiscovered. This situation is
currently changing due to the world-wide effort to measure the photoproduction of mesons
off nucleons with tagged photon beams. Not only angular distributions but also many different
polarization observables from measurements with circularly and linearly polarized beams
and longitudinally and transversely polarized targets are becoming available, e.g., from
the CLAS experiment at the Thomas Jefferson National Accelerator Facility (JLab),
the CBELSA/TAPS experiment at ELSA, and the Crystal Ball/TAPS
experiment at MAMI. The first results of this program are summarized in the Baryon Particle Listings
in the Review of Particle Physics \cite{PDG_12} by the Particle Data Group (PDG). The new results
from photon-induced reactions are quite important, e.g., for the coupled-channel partial-wave
analysis of the Bonn-Gatchina group (BnGn) \cite{Anisovich_12} or the partial-wave analysis of
the SAID group \cite{Workmann_12}.
The majority of these recent measurements investigated photoproduction off free protons.
The complementary program for the neutron target is less advanced due to the complications
arising from the use of quasifree neutrons bound in light nuclei as targets. However, this
program is very important for the investigation of $N^{\star}$ resonances because the
$\gamma NN^{\star}$ helicity couplings are isospin dependent. In some cases it is even
possible that the excitation of states is forbidden for the proton (or at least strongly
suppressed) but allowed for the neutron (Moorehouse selection rules \cite{Moorehouse_66}).
Therefore, the isospin decomposition of the amplitudes requires measurements
of photoproduction of mesons off neutrons. Light nuclei such as $^3$He and, in particular,
the loosely bound deuteron are the best available targets. In comparison to measurements
with free protons some complications arise. The first is of a technical nature.
The classification of the final state requires the detection and identification of the
recoil nucleon. This is challenging for all-neutral final states (neutron, decay-photons
from neutral mesons like $\pi^0$, $\eta$), which are produced in some of the most interesting reactions.
At present, only almost $4\pi$ covering electromagnetic calorimeters with good particle
identification capabilities can efficiently measure such reactions. All excitation functions,
angular distributions, and other observables will be smeared by the Fermi motion of the
bound nucleons. This problem can in principle be overcome by a complete reconstruction of
the final-state nucleon-meson kinematics, which reveals the `true'
$W = \sqrt{s} = \sqrt{(p_{N'}+p_m)^2}$ ($p_{N'}$, $p_m$: recoil nucleon and meson four-momenta)
of the reaction. However, for tagged-photon experiments this
means that the resolution of $W$ is no longer defined by the energy resolution
of the incident tagged-photon beam but by the typical resolution of the reaction-product
detector (which is usually much worse). The last problem is the possible modification of the
experimental results due to nuclear effects, in particular final-state interactions (FSI)
between the nucleons or between mesons and nucleons.
Such effects can be investigated for proton photoreactions where the results for free protons
can be compared with quasifree measurements on protons bound in the deuteron. This gives some
indication of whether, for a specific reaction channel, FSI effects are important and can
test FSI models before they are applied to quasifree neutron measurements.
\subsection{Photoproduction of $\eta$ mesons}
Photoproduction of $\eta$ mesons attracted much interest when this reaction became
experimentally accessible with a precision comparable to pion photoproduction. Due to its
isoscalar nature only isospin 1/2 $N^{\star}$ resonances can decay to the nucleon ground state
via $\eta$ emission. Furthermore, due to its relatively large mass (compared to pions) the number
of relevant partial waves is still small at excitation energies where so far many predicted
low-spin $N^{\star}$ states are `missing'. This simplifies the interpretation of the data.
\begin{figure}[b]
\centering
\includegraphics[width=0.48\textwidth]{fig_01.eps}
\caption{(Color online) Total cross section for $\gamma p\rightarrow \eta p$ averaged over data from
\cite{Krusche_95,Renard_02,Dugger_02,Crede_05,Nakabayashi_06,Bartalini_07,Crede_09,Williams_09,McNicoll_10}.
Model curves are from MAID 1 \cite{Chiang_02}, MAID 2 \cite{Chiang_03}, SAID \cite{McNicoll_10}, and
BnGn \cite{Anisovich_05}.}
\label{fig:proton}
\end{figure}
Experimental progress for measurements of the $\gamma p\rightarrow p\eta$ reaction
with free protons was huge during the last decade, so that now photoproduction of
$\eta$ mesons is probably the best studied final state apart from pion
production. Total cross sections, angular distributions, and some polarization observables have been
investigated at all major tagged photon facilities sometimes even with repeated and improved
experiments \cite{Krusche_95,Ajaka_98,Bock_98,Renard_02,Dugger_02,Crede_05,Nakabayashi_06,Bartalini_07,Bartholomy_07,Elsner_07,Crede_09,Williams_09,Sumihama_09,McNicoll_10}.
For the discussion of the gross features of $\eta$ production off the proton, Fig.~\ref{fig:proton}
summarizes the results for the total cross section (averaged over all available data).
At threshold ($E_{\gamma}$= 708~MeV, $W=$ 1486~MeV) the reaction is completely dominated by the
excitation of the $S_{11}$(1535) resonance \cite{Krusche_97}. Contributions from the $P_{11}$(1440)
(Roper) resonance have never been directly identified. The Review of Particle Physics cites a
branching ratio of (0$\pm$1)\% \cite{PDG_12}. The $D_{13}$(1520) resonance makes a tiny contribution
(branching ratio (0.23$\pm$0.04)\% \cite{PDG_12}), which is negligible for the total cross section
but was identified via an interference term in the angular distributions \cite{Krusche_95,Weiss_03}
and, even more significantly, in the photon beam asymmetry \cite{Ajaka_98,Elsner_07}. At slightly
higher energy the $S_{11}$(1650) interferes (for the proton) destructively with the $S_{11}$(1535).
In the S$_{11}$ region (see \cite{Krusche_03} for a summary) contributions from
non-resonant backgrounds seem to be small.
At slightly higher energies contributions from the $D_{15}$(1675), $D_{13}$(1700), $P_{11}$(1710),
and $P_{13}$(1720) resonances can be expected. Branching ratio estimates given by PDG \cite{PDG_12}
are 10--30\% for the $P_{11}$, (4$\pm$1)\% for the $P_{13}$, and (0$\pm$1)\% for the two $d$-wave
states. The results differ between the available analyses. Total cross sections from some analyses
(MAID 1: $\eta$-MAID isobar model \cite{Chiang_02}, MAID 2: $\eta$-MAID reggeized isobar model
\cite{Chiang_03}, BnGn: Bonn-Gatchina coupled-channel analysis \cite{Anisovich_05},
SAID partial-wave analysis \cite{McNicoll_10}) are compared in Fig.~\ref{fig:proton} to the average
of all available data. In the region around photon energies of 1~GeV agreement between
the model fits is not very good and, as discussed, e.g., in \cite{Elsner_07}, the relative contributions
of different resonances differ quite a bit in the models.
For the neutron, the range from threshold throughout the $S_{11}$(1535) resonance has been studied
intensively using the deuteron or helium isotopes as targets
\cite{Krusche_95_2,Hoffmann_Rothe_97,Weiss_03,Weiss_01,Pfeiffer_04,Hejny_99,Hejny_02}.
The experiments found consistently a neutron-to-proton cross-section ratio for quasifree production
close to 2/3 and very small coherent contributions, which was interpreted as evidence for a dominant
isovector excitation of the $S_{11}$(1535) (the isoscalar admixture in the proton
amplitude is only $\approx$9\% \cite{Krusche_03}).
Above this energy range many open questions exist. Most analyses (MAID, SAID) find a negative
sign for the $A^n_{1/2}$ (in the following all values are in units of $10^{-3}$ GeV$^{-1/2}$)
neutron helicity coupling of the four star $S_{11}$(1650) resonance (PDG: $-$15$\pm$21) and thus
a destructive interference between the two $S_{11}$ states. A negative sign is also preferred by
quark models (e.g., Capstick \cite{Capstick_92}: $-35$; Burkert et al. \cite{Burkert_03}: $-$31$\pm$3).
However, the more recent analyses of the Bonn-Gatchina group \cite{Anisovich_13} (25$\pm$20)
and Shresta and Manley \cite{Shrestha_12} (11$\pm$2) found positive values corresponding
to a constructive interference. The $\eta$-MAID model \cite{Chiang_02} found a much larger
$\eta$-decay branching ratio than quoted in PDG for the $D_{15}$(1675) state (17\%). It thus
predicted a significant contribution of this state to $\gamma n\rightarrow n\eta$ because it
has much larger photon couplings for the neutron than for the proton. Furthermore, there were
predictions that the non-strange $P_{11}$-like member of the conjectured baryon antidecuplet
\cite{Diakonov_97} should be electromagnetically excited more strongly on the neutron, should
have a large decay branching ratio to $N\eta$, an invariant mass around 1.7 GeV, and a width
of a few tens of MeV \cite{Diakonov_97,Polyakov_03,Arndt_04}.
Motivated by these open problems several experiments have recently studied this reaction.
Exclusive measurements of $\gamma p\rightarrow \eta p$ and $\gamma n\rightarrow \eta n$ on the deuteron in quasifree
kinematics were performed by GRAAL \cite{Kuznetsov_07}, at the Laboratory of Nuclear Science
at Sendai (LNS-Sendai) \cite{Miyahara_07}
and by the CBELSA/TAPS collaboration \cite{Jaegle_08,Jaegle_11}. A prominent structure
in the total cross section of $\gamma n\rightarrow n\eta$ at incident photon energies
around 1~GeV was first found by the GRAAL experiment. This peak with a width of only $\approx$25~MeV
appeared also in the CBELSA/TAPS data at $W = 1.67$ GeV when the true center-of-mass energy
$W$ was reconstructed from the final-state $\eta$ meson and the recoil neutron. Using the
same analysis, the cross section for $\gamma p\rightarrow \eta p$ was determined and good agreement to
previous direct measurements on the free proton was found, demonstrating that nuclear
effects could be reliably controlled by this method. Moreover, it was found that around the
same value of $W$ the total proton cross section shows a dip-like structure, which was
confirmed by the latest high-statistics measurement on free protons at MAMI-C \cite{McNicoll_10}.
The origin of the dip in the proton cross section was recently discussed for various scenarios
(narrow resonances, threshold effects from the $\gamma p\rightarrow \omega p$ reaction)
in the framework of the BnGn model \cite{Anisovich_13a}.
At the moment the nature of the peak in the neutron cross section and the dip in the proton
cross section is not understood nor is it clear whether they are correlated. The only
scenario that is ruled out in the case of the neutron is that the peak originates from an
isolated conventional broad resonance. Various scenarios have been suggested in the literature.
They range from different coupled-channel effects of known nucleon resonances
\cite{Shklyar_07,Shyam_08}, interference effects in the $S_{11}$ partial wave
\cite{Anisovich_09, Anisovich_13}, effects from strangeness threshold openings
\cite{Doering_10}, to intrinsically narrow states
\cite{Arndt_04,Choi_06,Fix_07,Anisovich_09,Shrestha_12}.
The available data are insufficient for an unambiguous analysis.
In this work we present in detail the results from a new high statistics measurement of the
total cross section and angular distributions for the reactions $\gamma p\rightarrow \eta p$ and $\gamma n\rightarrow \eta n$
extracted from data taken with a liquid deuterium target at Mainz.
The main experimental results have been summarized in a preceding Letter \cite{Werthmueller_13}.
Both the $\eta\rightarrow 2\gamma$ and the $\eta\rightarrow 3\pi^{0}\rightarrow 6\gamma$ neutral decay modes were used for the reconstruction
of the $\eta$ meson, leading to an unprecedented statistical quality of the results and to
stringent limits for systematic uncertainties.
Measurements of further quantities, such as single and double polarization observables, are
of course highly desirable and are already under way at MAMI.
\section{Experimental setup}
The data were measured during three different beam times at the MAMI electron accelerator
facility in Mainz \cite{Herminghaus_76,Kaiser_08} using the standard A2 setup for photon
beam experiments. Details for the experimental parameters (targets, beams) are summarized
in \cite{Oberle_13}. The electron beam, having an energy of
$E_0 = 1508$~MeV (1558~MeV during part of the beam time) and a current ranging from 4.5
to 10 nA after the last accelerator stage of MAMI-C, was used to produce photons via
bremsstrahlung in a 10 $\mu$m copper radiator. The scattered electrons were momentum-analyzed
up to 95\% of the initial beam energy in the Glasgow photon tagger
\cite{Anthony_91,Hall_96,McGeorge_08}. The large magnetic dipole of the spectrometer and
the 353 half-overlapping plastic scintillators installed in the focal plane allow an energy
reconstruction of the bremsstrahlung photons via $E_\gamma = E_0 - E_{e^-}$ from the
measured electron energy $E_{e^-}$ with a resolution of 2--5 MeV. Electrons corresponding
to photon energies below 400 MeV were not recorded to increase the event rate and to prevent
damage to the phototubes of the low energy tagger detectors.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{fig_02.eps}
\caption{(Color online) Diagram of the experimental setup created by the Geant4 model. The Crystal
Ball detector was cut along the x-axis to show the inner detectors and the target.}
\label{fig:setup}
\end{figure}
The photon beam was collimated using a 4 mm diameter lead collimator and impinged on the 4.72 cm
long (3.0 cm for part of the beam time) Kapton target cylinder of 4 cm diameter. Outgoing
particles were detected using the two calorimeters Crystal Ball (CB) \cite{Starostin_01}
and TAPS \cite{Novotny_91,Gabler_94}. A schematic diagram of the detector setup is
shown in Fig.~\ref{fig:setup}. The CB consists of two hemispheres with in total 672
optically insulated NaI(Tl) crystals of 15.7 radiation length thickness, covering all
azimuthal angles for the polar angle range $20^\circ < \theta < 160^\circ$.
All crystals point towards the center of the sphere where the target is mounted.
The distance from the center to the detector modules is 25 cm.
The energy resolution for photons can be described as
$\Delta E / E = 2\%/(E[\mathrm{GeV}])^{0.36}$ while typical angular resolutions are
$\Delta\theta = 2^\circ$--$3^\circ$ and $\Delta\phi = 2^\circ$--$4^\circ$ \cite{Starostin_01}.
The forward hole of the CB is covered by the hexagonal TAPS wall, which is made of
384 hexagonally shaped BaF$_2$ crystals with a thickness of 12 radiation lengths. TAPS was
installed 1.46 m downstream from the target covering polar angles from
$5^\circ$ to $21^\circ$. The photon energy resolution is parametrized as
$\Delta E / E = 1.8\% + 0.8\%/(E[\mathrm{GeV}])^{0.5}$ \cite{Gabler_94}.
The fine granularity of the detector
elements leads to excellent resolution in the polar angle (better than $1^\circ$), while
$\Delta\phi = 1^\circ$--$6^\circ$.
Neutral and charged particles were discriminated by plastic scintillators in both detectors.
A 50 cm long barrel of 24 strips with a width of 4 mm surrounded the target and acted as
particle identification detector (PID) for the CB \cite{WATTS_05}. In TAPS charged particles
were identified with individual 5 mm thick plastic scintillators that were installed in front
of every detector element. The multi-wire proportional chamber (MWPC) surrounding the PID
in the CB was not active for the present experiment.
The experimental trigger was relatively open and consisted of a total energy sum threshold
in the CB and a minimal total `hit' multiplicity in the CB and TAPS. The energy sum was implemented
as the analog sum of all signals from the CB and its threshold was adjusted to correspond to
an energy of around 300 MeV, mainly to reject $\pi^0$ production events.
The 672 crystals of the CB were grouped into 45 sectors each containing up to 16
neighboring crystals and TAPS was divided into 6 triangular sectors. If the signal from at least one crystal
in a sector exceeded a threshold of about 30 MeV (35 MeV in TAPS) that sector contributed to the event
multiplicity.
A minimal total multiplicity in the CB and TAPS
of two was set to accept events from the $\eta\rightarrow 2\gamma$ decay. For a part of the beam time
a multiplicity of 3 was required to increase statistics for the $\eta\rightarrow 3\pi^{0}$ decay and other
multi-photon channels. This data set could not be used for the $\eta\rightarrow 2\gamma$ decay.
\section{Data analysis}
\subsection{Subtraction of tagger random coincidences}
All electron hits in the photon tagger were recorded for each event triggered by the
calorimeters. The tagger itself did not contribute to the trigger decision because
for every CB/TAPS trigger there was almost always a hit in the tagger.
The random coincidences were subtracted by a standard side-band analysis of the production-detector
tagger coincidence. Fig.~\ref{fig:rnd_subtr} shows the relative timing spectra between the
tagger and both the CB and TAPS detectors. The time resolution with respect to TAPS was around
0.9 ns compared to the 1.5 ns that could be achieved with the CB.
Hence, whenever possible, tagger coincidence time evaluation was performed with photons
detected in TAPS. The true coincidences located in the peak ($C_T$) (hatched blue area in
Fig.~\ref{fig:rnd_subtr}) were determined by a subtraction of the random coincidences ($C_R$),
which were analyzed with cuts on the random background in the time spectrum (regions $R_1$, $R_2$)
with proper normalization. This procedure was applied to all spectra. The random background
windows in Fig.~\ref{fig:rnd_subtr} are only a schematic representation. They were much wider
in the actual analysis ($2\times 200$ ns) so that statistical uncertainties from random
background were negligible.
\begin{figure}[th]
\centering
\includegraphics[width=0.48\textwidth]{fig_03.eps}
\caption{(Color online) Tagger-calorimeter coincidence time spectra (sum of all channels): Left side: timing between tagger
and the Crystal Ball. Right side: timing between tagger and TAPS.
Hatched red areas: random background ($R_1$ and $R_2$) and random coincidence ($C_R$)
windows. Hatched blue areas: true coincidence window ($C_T$).}
\label{fig:rnd_subtr}
\end{figure}
\begin{table}[b]
\centering
\begin{tabular}{|c|c|c|}
\hline $\eta$ decay & reaction & cluster selection criteria \\
\hline
& $\gamma N\rightarrow \eta (N)$ & (2n \& 0c) or (2n \& 1c) or (3n \& 0c) \\
$2\gamma$ & $\gamma p\rightarrow \eta p$ & 2n \& 1c \\
& $\gamma n\rightarrow \eta n$ & 3n \& 0c \\
\hline
& $\gamma N\rightarrow \eta (N)$ & (6n \& 0c) or (6n \& 1c) or (7n \& 0c) \\
$3\pi^0$ & $\gamma p\rightarrow \eta p$ & 6n \& 1c \\
& $\gamma n\rightarrow \eta n$ & 7n \& 0c \\ \hline
\end{tabular}
\caption{Summary of the basic event selection criteria based on the number and type
(n=neutral, c=charged) of detected clusters. See text for more details.}
\label{table:event_selection}
\end{table}
\subsection{Particle reconstruction and reaction identification}
\label{sec:iden}
Clusters in the calorimeters were built from adjacent crystals where the deposited energy
in each single crystal exceeded a typical threshold of 2 MeV in the CB and 3--5 MeV in TAPS.
If the total deposited energy in all crystals was below 20 MeV, the cluster was ignored in
the analysis.
The reconstructed clusters in the CB and TAPS were first classified as neutral or charged.
Clusters were marked as charged if a coincident hit in the corresponding PID (CB) or
veto element (TAPS) was found, otherwise they were marked as neutral. Depending on the
reaction to be measured, a condition on the number and type of clusters was set.
An overview of these conditions is given in Table \ref{table:event_selection}. To measure
the reactions $\gamma p\rightarrow \eta p$ and $\gamma n\rightarrow \eta n$ exclusive measurements were performed, i.e., the
detection of the recoil nucleons was required. In the inclusive measurement of $\gamma N\rightarrow \eta (N)$
the recoil nucleon could also be undetected. The $\eta$ meson was identified using the
neutral decays $\eta\rightarrow 2\gamma$ and $\eta\rightarrow 3\pi^{0}\rightarrow 6\gamma$ as described in the following.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{fig_04.eps}
\caption{(Color online) Typical invariant-mass spectra for two bins of incident photon energy $E_\gamma$
in the range of $\eta$ mesons for the $\eta\rightarrow 2\gamma$
and $\eta\rightarrow 3\pi^{0}$ decays in coincidence with recoil protons $p$, recoil neutrons $n$,
and without any condition for recoil nucleons $(N)$. The indicated cuts (red lines) have been
applied to the spectra discussed below.
}
\label{fig:inv_mass_raw}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_05.eps}
\caption{(Color online) $\eta$-$N$ coplanarity $\Delta\phi$ used for the reaction identification:
Top two rows: $\eta\rightarrow 2\gamma$ analyses. Bottom two rows: $\eta\rightarrow 3\pi^{0}$ analyses.
Columns: bins of incident photon energy $E_\gamma$. Black squares: experimental data.
Curves: simulations of pure signal (blue), all background contributions (dashed magenta,
see text), sum of signal and background (green). Red lines: cut markers.}
\label{fig:cop}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_06.eps}
\caption{(Color online) Missing mass $\Delta m_X$ used for the reaction identification:
Top three rows: $\eta\rightarrow 2\gamma$ analyses. Bottom three rows: $\eta\rightarrow 3\pi^{0}$ analyses.
Columns: bins of incident photon energy $E_\gamma$. Black squares: experimental data.
Curves: simulations of pure signal (blue), all background contributions (dashed magenta,
see text), sum of signal and background (green). Red lines: cut markers.}
\label{fig:mm}
\end{figure*}
The charged cluster was always assumed to originate from the recoil proton. In the
$\eta\rightarrow 2\gamma$ analysis with proton coincidence the $\eta$ meson four-momentum was then immediately
reconstructed from the remaining two neutral clusters. In the case of neutron coincidence,
a $\chi^2$ search was performed among the detected three neutral clusters finding the
minimal
\begin{equation}
\chi^2 = \left( \frac{m_\eta - m_{\gamma\gamma}}{\Delta m_{\gamma\gamma}}\right)^2
\end{equation}
of all three photon pair combinations. The invariant mass of the two photons and
the real $\eta$ mass are denoted as $m_{\gamma\gamma}$ and $m_\eta$, respectively.
The uncertainty of the invariant mass $\Delta m_{\gamma\gamma}$ was evaluated with respect to
individual photon cluster angular and energy resolutions of the detector system. The required
resolutions $\Delta\theta(\theta), \Delta\phi(\theta)$ and $\Delta E(E)$ were determined
from Monte Carlo (MC) simulations and the known energy resolutions of the CB and TAPS.
Once the best combination was found the remaining cluster was marked as the neutron candidate.
Events with wrong assignments of the neutrons are sufficiently rejected by the later
applied analysis cuts (see Sec.~\ref{sec:add_checks}).
In the $\eta\rightarrow 3\pi^{0}$ analyses the minimal
\begin{equation}
\chi^2 = \sum\limits_{i=1}^{3}\left( \frac{m_{\pi^0} - m_{\gamma\gamma,i}}{\Delta m_{\gamma\gamma,i}}\right)^2
\end{equation}
over all possible combinations to form three $\pi^0$ mesons with masses $m_{\pi^0}$ out of
six or seven neutral clusters was used to combine the best three $\pi^0$ mesons to
an $\eta$ meson. More details about the $\chi^2$ analysis are given in \cite{Witthauer_13}.
Since the contribution of the energy resolution to the two photon invariant mass is larger
than the contribution of the angular resolution, the
energy reconstruction of mesons can be optimized by applying the correction
\begin{equation}
E^{'} = E \frac{m_m}{m_{\gamma\gamma}}
\end{equation}
to the reconstructed energy $E$, where $m_{\gamma\gamma}$ and $m_m$ are the invariant mass
of the decay photons and the real meson mass, respectively, thus obtaining the energy
$E^{'}$, which has better resolution. This method correcting the detected photon energies
within their resolution was used for the final $\eta$ reconstruction as well as for the
intermediate state $\pi^0$ reconstruction in the $\eta\rightarrow 3\pi^{0}$ analyses.
Typical spectra of the $2\gamma$ and $3\pi^0$ invariant masses obtained after this event
selection are shown in Fig.~\ref{fig:inv_mass_raw} for two ranges of incident photon energy.
The resolution in the $\eta\rightarrow 3\pi^{0}$ channel is better due to the constraints posed by the
intermediate state $\pi^0$ mesons (cuts of $\pm3\sigma$ were applied on the $m_{\pi^0}(2\gamma)$
masses in the reconstruction of the three intermediate $\pi^0$ mesons (not shown)). For all
further spectra aiming at the reaction identification, rough cuts on the invariant
$\eta$-mass indicated in Fig.~\ref{fig:inv_mass_raw} were applied to suppress
backgrounds from single and double pion production.
The major part of the background from other reaction channels was removed by a
proper identification of the signal reaction using kinematic cuts. These cuts
were already applied before the use of further particle identification spectra, such as pulse-shape
analysis, time-of-flight versus energy, and $\Delta E -E$ analysis, because they can be much more
reliably modeled by MC simulations and are therefore less critical sources for
systematic uncertainties.
For the exclusive analyses (which are more important) a coplanarity cut involving the detected
recoil nucleons can be established.
Namely, it is required that the $\eta$ meson, the recoil nucleon, and the incoming photon lie
in one plane. This can be translated into a condition on the azimuthal
angle difference $\Delta\phi$ between $\eta$ meson and recoil nucleon $N$ using
\begin{equation}
\Delta\phi =
\begin{cases}
\phi_\eta - \phi_N &\mbox{if } \phi_\eta - \phi_N \geq 0 \\
2\pi - |\phi_\eta - \phi_N| &\mbox{if } \phi_\eta - \phi_N < 0
\end{cases}
,
\end{equation}
where $\phi_i$ are the corresponding azimuthal angles of the two reconstructed particles in
the lab frame. The resulting distributions are shown in Fig.~\ref{fig:cop} along with the applied
cuts at $\pm 2\sigma$. Because of the
Fermi motion of the initial-state nucleons, the distributions are broader than in the
analysis of free proton data. The distributions are very well reproduced by MC
simulations. Significant background is only visible for the $2\gamma n$ final state. It originates
from the $\pi^0 n$ final state when one of the decay photons is misidentified as a neutron and
the corresponding neutron assigned as a photon. Background reactions where all final-state
particles have been detected and correctly identified will of course pass this cut.
More sensitive is a cut on the missing mass $\Delta m_X$ of the reaction, which was calculated under
the assumption of quasifree production of $\eta$ mesons off nucleons at rest via
\begin{equation}
\Delta m_{X} = \sqrt{\left(E_\gamma + m_N - E_\eta\right)^2 - \left(\vec{p}_\gamma - \vec{p}_\eta\right)^2}-m_N,
\end{equation}
where $E_\gamma$, $\vec{p}_{\gamma}$ are respectively the energy and momentum of the incident photon,
$E_\eta$, $\vec{p}_\eta$ are respectively the reconstructed energy and momentum of the meson and $m_N$ is
the mass of the recoil nucleon. The latter (no matter if detected or not) was treated
as a missing particle. Effects from nuclear Fermi motion, which were ignored in this analysis,
broaden the experimentally observed structures. Typical spectra for $\eta\rightarrow 2\gamma$ and $\eta\rightarrow 3\pi^{0}$ decays in
inclusive mode and in coincidence with recoil nucleons are summarized in Fig.~\ref{fig:mm}.
The reaction identification cuts discussed above were applied to these spectra.
Events from quasifree $\eta$-production peak around zero missing mass, while background
from $\eta\pi$ final states, where the pion escaped detection, appears at positive missing
mass and increases strongly with incident-photon energy. Because of this energy dependence,
$E_\gamma$-dependent symmetric cuts around the signal maximum were used. They account also for the
small offsets (identical for data and MC simulations) of the peak positions from zero for higher
incident photon energies.
The background from $\eta\pi$ production (mainly charged pions) can pass all previous cuts when
the charged pion is emitted with low energy or at small polar angle and escapes detection.
In that case, the coplanarity cut does not help because either the energy of
the pion is so low or its polar angle is so small that it does not disturb the azimuthal
correlation between recoil nucleon and $\eta$ meson. Nevertheless, the coplanarity cut
removes a significant fraction of the background as can be seen (Fig.~\ref{fig:mm}) in the
comparison of the missing mass spectra for the exclusive reactions (with coplanarity cut)
to the inclusive reaction (without it). As a consequence, a very strict $\pm0.5\sigma$
missing-mass cut was applied to the inclusive reactions whereas in the exclusive cases a
broader $\pm1.5\sigma$ cut was applied. For photon energies below the $\eta\pi$ production threshold
no cut was applied because no background was visible. Sufficient rejection of the background
was checked using simulations of various $\eta\pi$ production channels (their relative
cross sections are known) that gave together with the simulated signal distributions
a good description of the measured distributions (see green curves).
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{fig_07.eps}
\caption{(Color online) Pulse-shape analysis (PSA) distributions for particles detected in TAPS:
Left column: $\eta\rightarrow 2\gamma$ analyses. Right column: $\eta\rightarrow 3\pi^{0}$ analyses.
Top three rows: photon candidates in the inclusive and the two exclusive analyses.
Bottom two rows: proton and neutron candidates in the exclusive analyses.
Black lines: cut markers. Counts increase from violet to red.}
\label{fig:psa}
\end{figure}
\subsection{Additional checks}
\label{sec:add_checks}
\begin{figure*}[t]
\centering
\includegraphics[width=0.47\textwidth]{fig_08a.eps}
\includegraphics[width=0.47\textwidth]{fig_08b.eps}
\caption{(Color online) Left side: Time-of-flight analysis (TOF) distributions for particles detected
in TAPS: Left column: $\eta\rightarrow 2\gamma$ analyses. Right column: $\eta\rightarrow 3\pi^{0}$ analyses.
Top three rows: photon candidates in the inclusive and the two exclusive analyses.
Bottom two rows: proton and neutron candidates in the exclusive analyses.
Counts increase from violet to red. \\
Right side: $\Delta E$--$E$ distributions for protons: Left column: $\eta\rightarrow 2\gamma$ analyses.
Right column: $\eta\rightarrow 3\pi^{0}$ analyses. First row: protons in the CB (all events).
Second row: protons in the CB (accepted events). Third row: protons in TAPS (all events).
Fourth row: protons in TAPS (accepted events). Counts increase from violet to red.}
\label{fig:tof_deltaE}
\end{figure*}
So far only the information from the charged particle identification detectors
(PID and TAPS-Veto) and the $\chi^2$ analysis have been used for particle identification.
Further constraints on particle types can be obtained for hits in TAPS from a pulse-shape
analysis (PSA) and a time-of-flight versus energy analysis, and for hits in TAPS and in
the CB from $\Delta E -E$ analyses and from cluster-size analyses.
The PSA uses the fact that the pulse shape produced by protons and neutrons in BaF$_2$ differs
strongly from the signals coming from photons. This is due to the different mechanisms of
energy deposition by these particles in matter. They result in different contributions
to the slow ($\tau=650$ ns) and the fast ($\tau=0.9$ ns) scintillation light components
of BaF$_2$. Therefore, in the TAPS data acquisition the signals are integrated over two ranges
(short gate: 40 ns, long gate: 2 $\mu$s) giving two signal integrals---one containing
mostly the short component and one containing the total signal. After an energy calibration
based on photon signals, the short-gate energy $E_s$ and the long-gate energy $E_l$ can be
compared to separate different particles. Convenient distributions are obtained by plotting
the PSA-radius $r_{\mathrm{PSA}}$ versus the PSA-angle $\phi_{\mathrm{PSA}}$ using the
transformations
\begin{equation}
r_{\mathrm{PSA}} = \sqrt{E_s^2 + E_l^2} \quad \mathrm{and} \quad \phi_{\mathrm{PSA}} = \arctan (E_s / E_l).
\end{equation}
The distributions for photon and nucleon candidates for all analyses are shown in
Fig.~\ref{fig:psa}. Since the calibration of $E_s$ was made by setting $E_s = E_l$ for
photons, the photons are located at $\phi_{\mathrm{PSA}} = 45^\circ$ for all PSA-radii.
The different mean energies of the decay photons originating from the $\eta\rightarrow 2\gamma$ and the
$\eta\rightarrow 3\pi^{0}$ decays are clearly represented by larger and smaller PSA-radii, respectively.
Protons and neutrons are located at lower angles in bands showing a characteristic energy
dependence. All analysis cuts discussed above were applied to the plotted distributions and
very little background contamination is visible, hence the influence of the applied PSA
cuts on the event selection is rather small. Some noticeable contamination was found in
the nucleon spectra (in their lower, right corners) where low energy photons or electrons,
which did not activate the veto detectors, are visible. However, no significant residue
of the photon band was observed in these spectra.
With respect to the photon mean positions $c_\gamma(r_{\mathrm{PSA}}) \approx 45^\circ$,
cuts were established by fitting $\phi_{\mathrm{PSA}}$-distributions for different bins of
PSA-radii. Photons were then only accepted within a $r_{\mathrm{PSA}}$-dependent $3\sigma$
band around $c_\gamma(r_{\mathrm{PSA}})$. Accepted nucleons had to be located at smaller
angles than the left photon cut position for
$r_{\mathrm{PSA}} < 85$ MeV and $r_{\mathrm{PSA}} > 380$ MeV.
For PSA-radii between these two values no cut was applied because high energy
punch-through nucleons were located in this area. The cuts were kept so conservative because
the background level already established by the other cuts was low and because
the PSA analysis could not be included in the MC simulations as
modeling of the two light components of BaF$_2$ is not available.
Additional information on the detected particles provided by the various detectors,
although not used for the application of cuts, was checked for signs of any deficiencies
in the event selection.
Because of the fast response of BaF$_2$ the distance from the target was sufficient for TAPS
to provide a useful time-of-flight (TOF) measurement.
The deposited energy plotted versus inverse velocity, $t_{\mathrm{TOF}}$ [ns/m], shows
distinct distributions for the different particle types (left side of Fig.~\ref{fig:tof_deltaE}).
Photons are located around 3.3 ns/m and the different energy of photons from the $2\gamma$ and
$3\pi^0$ decays is again clearly visible. For protons a fairly tight correlation between inverse
velocity and deposited energy can be seen.
Neutrons are located above 4 ns/m and do not show any correlation between
time-of-flight and deposited energy (because the latter is more or less random).
The neutron spectra at the bottom of Fig.~\ref{fig:tof_deltaE} show no residual trace of the
proton band indicating good separation between protons and neutrons in TAPS.
Actually, none of the spectra shows significant background structures from other
particle species, which demonstrates the good event selection by the previously discussed
analysis cuts.
Detected proton candidates could additionally be checked using the deposited energy in
the PID and Veto detectors. On the right side of Fig.~\ref{fig:tof_deltaE} the
corresponding distributions for candidates in the CB and TAPS are shown for both $\eta$
decay channel analyses. The spectra labeled `all', where no analysis cuts were applied,
can be compared to the spectra `acc' representing the accepted events after all cuts.
In the case of the CB, large background contaminations from charged pions and electrons are cleanly
removed in the analysis. Resolution in TAPS is worse due to inferior optical coupling of
the Veto scintillators; nevertheless, signatures of protons and electrons can be clearly
identified. The latter are also effectively suppressed by the analysis cuts.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{fig_09.eps}
\caption{(Color online) Cluster size distributions in the $\gamma n\rightarrow \eta n\rightarrow 2\gamma n$ (solid red histograms) and the
$\gamma n\rightarrow \eta n\rightarrow 3\pi^{0} n$ (dashed blue) analyses: Top row: particles in the CB. Bottom row: particles in
TAPS. Left column: photons. Right column: neutrons. }
\label{fig:clsize}
\end{figure}
There is no direct event-by-event discrimination of photons and neutrons in the CB
(no PSA and TOF has poor resolution due to the short flight path).
The separation is entirely based on the $\chi^2$ analysis of the invariant masses
of the `photon' pairs. However, on average there is a distinction between photon
and neutron hits by the size of the corresponding clusters, i.e., the number of
activated detector modules. Most neutron clusters consist of four or fewer detector
elements while high energy photons (from the $\eta\rightarrow2\gamma$ decay) produce
clusters of up to twelve crystals. The mean energy of the $\eta$ decay photons from
the $\eta\rightarrow 3\pi^{0}$ channel is smaller, therefore they produce smaller cluster sizes more
similar to neutrons. The measured cluster size distributions for the CB and TAPS are
shown in Fig.~\ref{fig:clsize}.
As expected the cluster size distributions for photons are quite different for the two
decay channels while they are very similar for neutrons.
This is a strong indication that a clean photon-neutron separation
was achieved, even for the CB where no PSA or TOF information could be used.
\subsection{Final yield extraction}
The final yields were extracted from invariant-mass spectra after the application of the
cuts discussed above. Typical examples for the different reaction types for some
energy ranges are shown in Fig.~\ref{fig:inv_mass} and compared to the distributions
obtained with MC simulations. The invariant-mass peaks from the
exclusive analyses are almost background free. Small background components
are most visible for the $\eta\rightarrow 2\gamma$ decay in the inclusive reaction.
The residual backgrounds were subtracted by a fit made using the peak shape from the
simulated distribution together with a polynomial background.
For the extraction of angular distributions, the entire analysis
procedure was of course done separately for each data point of the angular distribution.
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_10.eps}
\caption{(Color online) Invariant masses $m(2\gamma)$ and $m(3\pi^0)$ used for the identification of the
$\eta$ mesons: Top three rows: $\eta\rightarrow 2\gamma$ analyses. Bottom three rows: $\eta\rightarrow 3\pi^{0}$ analyses.
Columns: bins of incident photon energy $E_\gamma$. Black squares: experimental data.
Blue curves: simulation.}
\label{fig:inv_mass}
\end{figure*}
\subsection{$W$ reconstruction as final-state invariant mass}
\label{sec:w_rec}
A full reconstruction of the kinematics was performed for the exclusive analyses.
This method allowed a calculation of the Fermi momentum of the participant nucleon in the
initial state. A cut on the momentum rejecting events with momenta above 80 MeV was used
in a special version of the analysis (later referred to as type II) which attempted to optimize
the resolution of the $W$ reconstruction. In the standard analysis no such
cut was applied, partly due to the loss of statistics and because it introduces additional
systematic effects in the extraction of the cross section close to threshold, where
larger Fermi momenta play a crucial role.
Quasifree cross sections calculated as functions of the photon beam energy $E_\gamma$ are
affected by the Fermi motion of the initial-state nucleons, which are
bound inside the deuteron. This means that a fixed value of $E_\gamma$ corresponds to a
distribution of center-of-mass energies $W$ due to the convolution with the Fermi momentum
distribution. The resulting cross sections are then smeared compared to the fundamental
cross sections at fixed values of $W$. This loss of resolution is mainly a problem when
sudden changes occur in the latter, as in $\eta$ photoproduction at threshold and in
the region of interest around $W = 1680$ MeV. Therefore, in this work the `true' center-of-mass
energy was calculated by a full reconstruction of the reaction in impulse approximation
from the final state \cite{Jaegle_11}. For a limited angular region, $W$ could additionally
be reconstructed using a time-of-flight measurement of the recoil nucleons.
From the measurement of the energies and directions of the decay photons, the four-momentum
of the $\eta$ meson could be completely reconstructed. The same is in principle possible
for final-state protons although a special calibration would be needed to deduce the kinetic
energy from the energy deposition in the detector. As there is no correlation between the
kinetic energy of neutrons and their deposited energy, a reconstruction of the neutron
energies was not possible in this way. The energy of the recoil nucleon was thus treated
as unknown in the reconstruction with only the measured angles used to reconstruct
the direction of its momentum. This was done in the same way for protons and neutrons
in order to maintain identical systematic uncertainties. Additional unknown quantities
are the three components of the spectator nucleon momentum in the final state,
leaving in total four unknown quantities since $E_\gamma$, all masses of the involved
particles, and the momentum of the incident deuteron (at rest) are known.
These four variables were determined from the four equations following from energy and
momentum conservation. In this way, the three momenta of final-state participant (and spectator)
nucleon and the final-state invariant mass of the recoil-nucleon-meson pair could be
reconstructed in the plane-wave impulse approximation.
The kinetic energy of nucleons detected in TAPS could also be determined via a
time-of-flight measurement. Regarding the different time resolutions of the detectors, the
best measurement would be possible for events with a photon in TAPS along with the recoil
nucleon. Due to the reaction kinematics and the experimental trigger, there are practically
no such events in the $\eta\rightarrow 2\gamma$ analyses and only results with low statistics
could be extracted in the $\eta\rightarrow 3\pi^{0}$ analyses. Therefore, the main TOF results were deduced from time
measurements relative to the tagger, which are affected by an inferior time resolution.
TOF measurements were not possible for the CB because of the short flight path and the poor time
resolution in the NaI(Tl) crystals. Due to this restriction, $W$ could only be reconstructed
for $\eta$ polar angles $\theta_{\eta}^{*}$ in the center-of-mass frame with $-1 < \cos(\theta_{\eta}^{*}) < -0.5$.
Nevertheless, this independent method serves as a check for the $W$ reconstruction
discussed above.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{fig_11.eps}
\caption{(Color online) Resolution of the $W$ reconstructions illustrated by the response curves
of the kinematic reconstruction (solid curves) and the
reconstruction via time-of-flight (dotted curves) at several discrete values of $W$
(dashed lines). The response curves were obtained from the $\gamma n\rightarrow \eta n\rightarrow 2\gamma n$ simulation.}
\label{fig:wres}
\end{figure}
Cross sections as a function of $W$ reconstructed with the two discussed methods are no longer
affected by nuclear Fermi motion but depend on the experimental resolution for the reconstructed
$W$. The resolution was investigated with MC simulations of the experimental setup (see
section \ref{sec:cs_extr}). Phase-space events at several fixed values of $W$
($\delta$-functions) were generated and the responses of the detectors were modeled.
The same analysis as used for experimental data was applied to the simulated data and
the resulting distributions for $W$ from the $\gamma n\rightarrow \eta n\rightarrow 2\gamma n$ analysis are summarized in
Fig.~\ref{fig:wres}.
The effects from the energy and angular resolution for the
$\eta$ mesons enter into both analyses in the same way for the determination of
the $\eta$ four-momentum. The resolution for the angle of the recoil nucleon also enters
into both analyses via the definition of the direction of the nucleon momentum.
All these factors increase the width of the observed distributions with rising $W$.
The angular resolution for the recoil nucleon degrades significantly above $W \sim 1550$ MeV
where the majority of nucleons are detected in the CB for which the polar angle resolution
$\Delta\theta$ is worse than in TAPS ($\Delta\phi$ being similar).
The final ingredient, the kinetic energy of the recoil nucleon, is determined from
the kinematic reconstruction or from the TOF measurement. It is evident from
Fig.~\ref{fig:wres} that the first method results in better resolution, in particular for
larger values of $W$. For the TOF reconstruction for higher energies even the centroids
are shifted. This is due to the fact that the TOF resolution for neutrons is not very good
(additional uncertainty is introduced because they can interact at any depth in the
crystals and the TOF flight path is not very long) and that, at larger kinetic energies,
the TOF-energy relation becomes so flat that small effects in TOF result in large uncertainties
for the energies.
In the case of the kinematic reconstruction, the corresponding
resolution for the $\eta\rightarrow 3\pi^{0}$ analysis is slightly better---as it also is for the proton
analyses. When approximated with Gaussians they all show a nearly linear rise from
$\Delta W(\mathrm{FWHM}) \sim 10$ MeV at 1515 MeV to $\Delta W \sim 40$ MeV at 1820 MeV.
In summary, a FWHM resolution of 30 MeV in the region of interest around $W = 1680$ MeV
could be achieved with the kinematic reconstruction of the final-state invariant mass,
while the TOF reconstruction yields an inferior resolution of about 70 MeV.
\subsection{Extraction of cross sections}
\label{sec:cs_extr}
The obtained yields were normalized to differential cross sections by using the target
surface density of 0.2304 nuclei/barn (0.147 nuclei/barn for one of the beam times), the
flux of the incoming photon beam, the analysis-dependent detection efficiency and the
$\eta$ decay branching ratios $\Gamma_{2\gamma} = 39.41\%$ and $\Gamma_{3\pi^0} = 32.68\%$,
respectively \cite{PDG_12}.
\begin{figure}[b]
\centering
\includegraphics[width=0.48\textwidth]{fig_12.eps}
\caption{Photon flux on the target for the May 2009 dataset:
Left side: measured photon flux as a function of $E_\gamma$. \\
Right side: flux as a function of $W$ obtained from folding with the deuteron Fermi
momentum distribution.}
\label{fig:flux}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_13.eps}
\caption{(Color online) Detection efficiencies as a function of $\cos(\theta_{\eta}^{*})$ for the exclusive analyses for
different bins of $W_{\mathrm{kin}}$: Solid curves: $\eta\rightarrow 2\gamma$ proton (blue) and neutron (red) analyses.
Dashed curves: $\eta\rightarrow 3\pi^{0}$ proton (blue) and neutron (red) analyses.}
\label{fig:deteff}
\end{figure*}
Yields for all bins in $(E_\gamma,\cos(\theta_{\eta}^{*}))$, $(W_{\mathrm{kin}},\cos(\theta_{\eta}^{*}))$ and $(W_{\mathrm{TOF}},\cos(\theta_{\eta}^{*}))$ (with $\cos(\theta_{\eta}^{*})$ evaluated
in the corresponding center-of-mass frame) were individually extracted by integrating the
appropriate $2\gamma$ or $3\pi^0$ invariant-mass histograms. Having applied all analysis cuts
as discussed in Sec.~\ref{sec:iden}, these histograms were background-free for the
exclusive analyses. In the case of the inclusive analyses, the background was non-negligible
since no cuts on the recoil nucleon could be applied. Here, the signals were extracted
using a fit of the distributions consisting of the combined peak shape from the simulated distribution
and a second-order polynomial function for the background. Energy (but not $\cos(\theta_{\eta}^{*})$) dependent
contributions to the yields originating from the target windows were subtracted using data
that were measured while the target cell was empty. These contributions were about
5--7\% and showed a rather smooth energy dependence.
The photon flux on the target as a function of $E_\gamma$ was calculated via
\begin{equation}
N_{\gamma}(E_{\gamma}) = N_{e^-}(E_{\gamma}) \epsilon_{tg}(E_{\gamma}).
\end{equation}
The number of electrons $N_{e^-}(E_{\gamma})$ in the photon tagger was counted during the
whole experiment. The so-called tagging efficiency $\epsilon_{tg}(E_{\gamma})$, i.e.,
the fraction of correlated photons passing through the beam collimator, was determined in
frequent, dedicated measurements at low beam intensity. Running at these conditions
ensured that random electron coincidences were minimized, and that the photon detection
efficiency of the lead-glass detector, which was moved into the photon beam, was still
close to 100\%. Besides these absolute values of the tagging efficiency, relative values
were available at all times from the measured relative beam intensity using an
ionization chamber placed at the end of the photon beam line. By normalization of
the relative values to the absolute measurements, a time-dependent tagging efficiency
was calculated. The resulting flux
integrated over one of the beam times is shown as a function of $E_\gamma$ at
the left side of Fig.~\ref{fig:flux}. For the normalization of the cross section
obtained using either the kinematic or the TOF reconstruction, an effective photon flux as
a function of $W$ had to be calculated. The effective distribution of $W$ values was
calculated by folding the incoming photon beam energy distribution with the nucleon
momentum distribution inside the deuteron. For the latter the wave function of the Paris
$N$--$N$ potential was used \cite{Lacombe_81}. The resulting flux is shown at the right
side of Fig.~\ref{fig:flux}.
The detection efficiency was determined with a Geant4-based model \cite{Agostinelli_03}
of the experimental setup. Events covering the complete phase space of quasifree $\eta$
production were generated and tracked by the simulation. The resulting detector information
was analyzed using the same analysis as for real data. In addition, the experimental
trigger conditions had to be modeled realistically. In fact, an even more restrictive
implementation of the CB energy sum trigger was imposed (also on the experimental data),
namely that only the decay photons were allowed to contribute. The same restriction was
implemented for the hit multiplicity to avoid systematic differences in the proton and
neutron analyses due to the different interaction of these particles with the detectors.
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{fig_14.eps}
\caption{(Color online) Total cross sections as a function of the photon beam energy $E_\gamma$:
(a) Comparison of the two $\eta$ decay analyses of the proton
(blue diamonds and triangles), neutron (red stars and circles) and the inclusive analyses
(black crosses and squares).
(b) Comparison of the averaged proton (blue triangles) and neutron (red circles) results
and of their sum (black crosses) to the inclusive data (black squares).
Hatched areas: total systematic uncertainties of inclusive (black), proton (blue) and
neutron (red) data.}
\label{fig:tcs}
\end{figure*}
Subsequently, the detection efficiency was calculated as the ratio of detected and
generated events for each bin of the excitation functions. Some examples are shown in
Fig.~\ref{fig:deteff} for the exclusive analyses. Special attention was given to the
detection efficiencies of the recoil nucleons, as the systematic uncertainties of the hadronic
models in the energy range covered by this experiment were suspected to be rather large.
Especially the tracking of low energy neutrons through different materials requires
specific and accurate cross sections for the nuclear reactions involved. The proton
efficiencies are highly sensitive to the modeled detector geometries and material budgets
as well. A measurement on a hydrogen target was used to check and correct the
nucleon detection efficiencies obtained by simulation. For this purpose, relative
corrections of the simulated nucleon efficiencies for the free reactions $\gamma p\rightarrow \eta p$ and
$\gamma p\rightarrow \pi^0\pi^+ n$ were deduced from a comparison of hydrogen experimental data and the
corresponding simulation, and applied in the deuteron analyses. For nucleons detected in
TAPS corrections for recoil protons (neutrons) were on average around +7.3\% (+12.1\%)
(the detection efficiency was overestimated by the simulation), while for the CB
they were -1\% (-3.5\%) (underestimated on average by simulation). In the gap region between
the CB and TAPS the corrected nucleon efficiencies were found to be still inaccurate.
This corresponds to values of $\cos(\theta_{\eta}^{*})$ around $-0.6$ for $W > 1550$ MeV where the
sharp efficiency dependence on $\cos(\theta_{\eta}^{*})$, especially for the proton,
can be clearly seen in Fig.~\ref{fig:deteff}.
As a solution, differential cross sections as functions of the lab polar angle of the
nucleons were interpolated in the problematic regions and correction factors were calculated
that were finally applied on an event-by-event basis in the analysis.
\subsection{Systematic uncertainties}
Common to all results are the global systematic uncertainties of the photon flux (3\%),
the target surface density (4\%), the $\eta$ decay branching ratios ($<$ 1\%) and the
approximately constant uncertainty of the empty target subtraction (2.5\%). The systematic
uncertainty in the photon flux mainly comes from the absolute measurements of the tagging
efficiency and was estimated by the extreme values of the normalization of the relative
flux measurements to the absolute measurements. The target surface density depends on the length
of the target cell, which is subject to deformations when the target is cooled
down. The systematic uncertainties of the $\eta$ decay branching ratios are almost
negligible \cite{PDG_12}. Due to the low statistics of the empty target runs,
a conservative estimate of roughly half the relative yield contribution (2.5\%) was made.
Several systematic uncertainties were found to be of rather different importance for the
various analyses and showed a strong energy and $\cos(\theta_{\eta}^{*})$ dependence. Therefore, they were
calculated individually and for all bins of the obtained cross sections. First, the CB
energy sum trigger was found to be of great importance especially for the $\eta\rightarrow 3\pi^{0}$
analysis. Its uncertainty was estimated by slightly different applications of the software
model trigger in the analysis of simulated data. All analysis cuts were varied by $\pm3\%$
and systematic uncertainties were estimated from the differences between the results.
Uncertainties in the nucleon detection efficiencies were estimated taking into account effects of
different hadronic models in Geant4, trigger and cut effects in the analysis of the
hydrogen data used for the efficiency correction, and the influence of the correction applied for
the data corresponding to the TAPS-CB gap region.
\begin{figure*}[p]
\centering
\includegraphics[width=0.9\textwidth]{fig_15.eps}
\caption{(Color online) Differential cross sections as a function of the photon beam energy $E_\gamma$
and $\cos(\theta_{\eta}^{*})$: Blue triangles: exclusive proton. Red circles: exclusive neutron.
Black crosses: sum of proton and neutron data. Black squares: inclusive data.}
\label{fig:dcs}
\end{figure*}
The many small systematic uncertainties from the different analysis steps were added quadratically
and the result was added linearly to the uncertainties of the photon flux and the target
density.
The total uncertainty for the inclusive $\eta\rightarrow 2\gamma$ analysis shows almost no energy and
$\cos(\theta_{\eta}^{*})$ dependence and is about 7\%. The total uncertainty for the corresponding
$\eta\rightarrow 3\pi^{0}$ analysis falls from 11\% at threshold to 7\% at $E_\gamma = 1$ GeV above which
it is constant. The reason for these increased values are the higher uncertainties in the
backward region of $\cos(\theta_{\eta}^{*})$, which are related to the CB energy sum trigger. The systematic
uncertainties for the proton analyses are almost energy and $\cos(\theta_{\eta}^{*})$ independent (with the exception
of the most forward angular bin and the CB-TAPS gap region located at $\cos(\theta_{\eta}^{*})\sim-0.5$ for higher values of $W$)
and are about 6--7\%. The systematic uncertainties
for the neutron analyses show a more pronounced energy and $\cos(\theta_{\eta}^{*})$ dependence. The total values
for the $\eta\rightarrow 2\gamma$ ($\eta\rightarrow 3\pi^{0}$) analysis are around 12\% (15\%) at threshold, have a local
maximum of 13\% (14\%) near $W = 1580$ MeV, and fall more or less linearly to 9\% (10\%)
at the maximum energy.
\section{Results}
The results presented in this section were obtained by combining the datasets from all
three beam times. Furthermore, the data of the $\eta\rightarrow 2\gamma$ and $\eta\rightarrow 3\pi^{0}$
analyses were averaged according to their statistical weights to calculate the final cross
sections. Differential cross sections were extracted as functions of $(E_\gamma,\cos(\theta_{\eta}^{*}))$,
$(W_{\mathrm{kin}},\cos(\theta_{\eta}^{*}))$ and $(W_{\mathrm{TOF}},\cos(\theta_{\eta}^{*}))$ where $\cos(\theta_{\eta}^{*})$ was always evaluated in the corresponding
center-of-mass frame. Total cross sections were obtained by fitting the angular
distributions with Legendre polynomials.
\subsection{Cross sections as a function of $E_\gamma$}
The total cross-section results are shown in Fig.~\ref{fig:tcs}. On the left side,
the data from the two $\eta$ decay analyses are compared to each other. With the exception
of the neutron data in the threshold region, the cross sections extracted from the $\eta\rightarrow 3\pi^{0}$
analyses are slightly larger. This could be due to residual background from direct $3\pi^0$
photoproduction (the invariant mass spectrum of direct $3\pi^0$ production peaks in the
S$_{11}$(1535) region close to the $\eta$-mass due to trivial kinematic relations).
Other effects at lower
photon beam energies could be caused by the CB energy sum trigger, whereas at higher
energies cluster overlaps in the $\eta\rightarrow 3\pi^{0}$ analyses could lead to systematic effects.
At the right side of Fig.~\ref{fig:tcs}, the data averaged over the $\eta$ decays
are shown. The inclusive result and the sum of the proton and neutron cross sections
are compared. Since the coherent production of $\eta$ mesons off the deuteron is very
small \cite{Weiss_01}, the two exclusive
cross sections should add up to the inclusive data. Within a range of 10\%, which is
compatible with all of the involved systematic uncertainties, this is indeed the case.
The good agreement is also clearly visible in the corresponding angular distributions
that are shown in Fig.~\ref{fig:dcs}. In the region of the \mbox{$S_{11}$(1535)} resonance
($E_\gamma = $ 758--923 MeV) for example,
the proton and neutron distributions are curved in opposite directions due to an
interference with the \mbox{$D_{13}$(1520)} with different signs \cite{Weiss_03}. Their sum is flat
which is reproduced by the direct inclusive measurement. Also at higher energies, the angular
distributions of sum and direct measurement agree very well. This is a strong indication
that systematic uncertainties, although quite large in case of the neutron data, are
generally well under control.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{fig_16.eps}
\caption{(Color online) Total cross sections as a function of the final-state invariant mass
$W_{\mathrm{kin}}=m(\eta N)$: Blue triangles: proton data. Red circles: neutron data scaled by $3/2$.
Black stars: free proton data from MAMI-C \cite{McNicoll_10}.
Hatched areas: total systematic uncertainties of proton (blue) and neutron (red) data.}
\label{fig:tcsw}
\end{figure}
\begin{figure*}[p]
\centering
\includegraphics[width=0.99\textwidth]{fig_17.eps}
\caption{(Color online) Differential cross sections for $\gamma p\rightarrow \eta p$ as a function of the final state invariant mass
$W_{\mathrm{kin}}=m(\eta p)$ and $\cos(\theta_{\eta}^{*})$: Filled blue triangles: exclusive proton.
Open green triangles: quasifree data from CBELSA/TAPS \cite{Jaegle_11}.
Black stars: free proton data from MAMI-C \cite{McNicoll_10}.
Black curves: Legendre fits to the present results. Hatched blue areas: total systematic uncertainties
in the present work.}
\label{fig:dcsw_p}
\end{figure*}
\begin{figure*}[p]
\centering
\includegraphics[width=0.99\textwidth]{fig_18.eps}
\caption{(Color online) Differential cross sections for $\gamma n\rightarrow \eta n$ as a function of the final-state invariant mass
$W_{\mathrm{kin}}=m(\eta n)$ and $\cos(\theta_{\eta}^{*})$: Filled red circles: exclusive neutron.
Open green circles: quasifree data from CBELSA/TAPS \cite{Jaegle_11}.
Open black squares: quasifree data from GRAAL \cite{Kuznetsov_07}.
Black curves: Legendre fits to the present results. Hatched red areas: total systematic uncertainties
in the present work.}
\label{fig:dcsw_n}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_19.eps}
\caption{(Color online) Phenomenological fits of the total $\gamma n\rightarrow \eta n$ cross sections as functions of $W_{\mathrm{kin}}$ and
$W_{\mathrm{TOF}}$: Top row: $\eta\rightarrow 2\gamma$ analyses. Bottom row: $\eta\rightarrow 3\pi^{0}$ analyses.
Columns from left to right: Kinematic $W$ reconstruction with standard (Kin.~I)
and more strict (Kin.~II) cuts, time-of-flight $W$ reconstruction with standard (TOF I)
and more strict (TOF II) cuts. Points: Original data (filled red circles) and background
subtracted data (open black circles). The TOF data were scaled by a factor of 4.
Curves: Total fit (solid black), \mbox{$S_{11}$(1535)} contribution (dashed black), integrated background
Breit-Wigner (dotted black) and narrow BW (solid red).}
\label{fig:pos_width}
\end{figure*}
\subsection{Cross sections as a function of $W_{\mathrm{kin}}$}
Fig.~\ref{fig:tcsw} shows the total cross sections as a function of the final-state
invariant mass $W_{\mathrm{kin}}$ obtained using the kinematic reconstruction of the nucleon energies.
As discussed before, no effects from Fermi motion should be present in these data---it should
only be affected by the resolution of the $W$ reconstruction. Therefore, the proton data can be
directly compared to data measured on the free proton target as, for example, obtained at MAMI-C
\cite{McNicoll_10}. The main characteristic features of the latter data are reproduced.
There are some discrepancies at threshold, which are most probably due to
the much poorer (ca.~one order of magnitude) resolution in $W$ of the current analysis and
the complicated proton detection efficiency in the region of the \mbox{$S_{11}$(1535)} resonance.
Above $W = 1600$ MeV the two data sets deviate by up to 15\% for the highest measured
invariant masses. The differential cross sections for $\gamma p\rightarrow \eta p$ are shown in
Fig.~\ref{fig:dcsw_p} and also compared to the free proton data.
In addition, the quasifree data obtained by the CBELSA/TAPS collaboration \cite{Jaegle_11}
are plotted. The very precise angular distributions of \cite{McNicoll_10} are in general
well reproduced by this work. Some residual effects from the uncertain proton detection
efficiency in the CB-TAPS gap region are still visible. They are located in the energy
bins for 1518 MeV $< W <$ 1655 MeV, first around $\cos(\theta_{\eta}^{*})\approx 0.3$ and then slowly moving to
backward angles up to $\cos(\theta_{\eta}^{*})\approx -0.65$. The issues in the determination of these data
points are accounted for by increased systematic uncertainties. Altogether, the quasifree and free
proton data agree quite well (for most kinematics within systematic uncertainties), which indicates
that for this reaction channel nuclear effects from FSI are not important, so that the quasifree
neutron data can be regarded as close approximation of the free $\gamma n\rightarrow \eta n$
cross sections. This is by no means trivial. In a similar investigation of photoproduction of
$\pi^0$ mesons off nucleons bound in the deuteron \cite{Dieterle_14} substantial effects (on the
order of 25\%) were found and also $\eta$ production off nucleons bound in $^3$He nuclei
is strongly affected by FSI \cite{Witthauer_13}.
The total cross section for $\gamma n\rightarrow \eta n$ shown in Fig.~\ref{fig:tcsw} was scaled
by $3/2$ to compensate for the known ratio $\sigma_n/\sigma_p \approx 2/3$
\cite{Krusche_95_2,Hoffmann_Rothe_97, Hejny_99} in the maximum of the \mbox{$S_{11}$(1535)} resonance.
The shapes of the corresponding peaks in the proton and neutron cross-section data are very similar.
The small deviations are caused by the different systematic
effects in the proton and neutron detection efficiencies. Above $W = 1615$ MeV
the neutron cross section deviates strongly from the proton results and exhibits
a pronounced peak-like structure around $W = 1670$ MeV. This structure, already observed
by earlier measurements \cite{Kuznetsov_07,Miyahara_07,Jaegle_11}, is thus confirmed
by this work with much superior statistical significance.
The corresponding differential cross sections for $\gamma n\rightarrow \eta n$ are shown in Fig.~\ref{fig:dcsw_n}.
The data from GRAAL \cite{Kuznetsov_07} and CBELSA/TAPS \cite{Jaegle_11} are plotted for
comparison. There is reasonable agreement between all data in the region of the \mbox{$S_{11}$(1535)}
resonance. In the vicinity of the peak-like structure, some deviations between the different
measurements are visible. Above around $W = 1800$ MeV, the agreement with the GRAAL data
improves again. The much better statistical quality of the present data compared to the previous
measurements is obvious from the figure.
\subsection{Properties of the structure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_20.eps}
\caption{(Color online) Differential cross sections as a function of $W_{\mathrm{kin}}$ for different bins of $\cos(\theta_{\eta}^{*})$:
Points: Original data (filled red circles) and background-subtracted data
(open black circles). Curves: Total fit (solid black),
\mbox{$S_{11}$(1535)} contribution (dashed black), integrated background Breit-Wigner (dotted black),
total background (\mbox{$S_{11}$(1535)} + broad BW, solid blue) and narrow BW (solid red).}
\label{fig:ang}
\end{figure*}
\begin{table}[b]
\centering
\begin{tabular}{|l|c|c|c|}
\hline
& $W_R$ & $\Gamma_R$ & $\sqrt{b_{\eta}}A_{1/2}^n$ \\
& [MeV] & [MeV] & [$10^{-3}$GeV$^{-1/2}$] \\
\hline
Kin.~I $2\gamma$ & 1670$\pm$1 & 27$\pm$3 (50$\pm$3) & 12.1$\pm$0.8 \\
Kin.~II $2\gamma$ & 1669$\pm$1 & 25$\pm$6 (44$\pm$5) & 11.8$\pm$1.0 \\
Kin.~I $3\pi^0$ & 1669$\pm$1 & 30$\pm$5 (49$\pm$4) & 12.9$\pm$0.8 \\
Kin.~II $3\pi^0$ & 1665$\pm$3 & 53$\pm$17 (66$\pm$14) & 15.6$\pm$2.7 \\
\hline
Best estimate & 1670$\pm$5 & 28$\pm$5 (50$\pm$10) & 12.3$\pm$0.8 \\
\hline
TOF I $2\gamma$ & 1658$\pm$2 & (42$\pm$4) & 13.2$\pm$0.7 \\
TOF II $2\gamma$ & 1651$\pm$3 & (45$\pm$8) & 18.1$\pm$1.7 \\
TOF I $3\pi^0$ & 1658$\pm$3 & (41$\pm$9) & 13.9$\pm$1.5 \\
TOF II $3\pi^0$ & 1663$\pm$3 & (20$\pm$9) & 11.3$\pm$2.0 \\
\hline
Best estimate & 1658$\pm$7 & (42$\pm$10) & 13.3$\pm$2.0 \\
\hline
\end{tabular}
\caption{Overview of the extracted parameters from the phenomenological fits shown in
Fig.~\ref{fig:pos_width}: The values in parentheses correspond to the fits where the $W$
resolution was not taken into account via convolution with the signal parameterization.
Uncertainties are statistical only, except for the couplings of the kinematic reconstruction and the
`best estimates' which reflect also the scatter between the different fits and analyses, respectively.}
\label{table:fit_params}
\end{table}
The nature of the narrow structure observed for the $\gamma n\rightarrow n\eta$ reaction
around invariant masses of $W$ = 1670~MeV is not yet understood. The phenomenological
properties of this structure were analyzed with the same simplified ansatz as in
\cite{Jaegle_11}. It consists of a Breit-Wigner (BW) function with energy-dependent width
for the contribution of the \mbox{$S_{11}$(1535)} resonance, a narrow standard BW function for the
observed structure, and an additional broad BW function parameterizing the remaining
background contributions at higher energies. The data obtained from the kinematic $W$
reconstruction and the reconstruction via TOF (for $-1 < \cos(\theta_{\eta}^{*}) < -0.5$) were fitted separately.
It was found that the broad BW function was not needed to describe the data from the TOF
reconstruction (this is so because those data are restricted to $\eta$-backward angles
where the background is much different from forward angles). Individual fits for the
$\eta\rightarrow 2\gamma$ and $\eta\rightarrow 3\pi^{0}$ data were performed. The results are shown in Fig.~\ref{fig:pos_width}
and the extracted parameters for position $W_R$, width $\Gamma_R$, and
electromagnetic coupling $A^n_{1/2}$ (multiplied by the square root of the unknown $N\eta$
branching ratio $b_{\eta}$) assuming an $J=1/2$ state are summarized in
Table \ref{table:fit_params}. Kin.~I, Kin.~II
and TOF I, TOF II represent datasets obtained with different analysis cuts, where in the
sets II more strict cuts on the $\eta$ missing mass ($\pm0.5\sigma$), the $\eta$-$n$
coplanarity ($\pm0.5\sigma$) and the reconstructed Fermi momentum ($p_F<80$ MeV) were
applied (see Sec.~\ref{sec:iden}).
The data depending on $W_{\mathrm{kin}}$ were additionally analyzed with a fit taking into account the
resolution of the $W$ reconstruction, which was estimated via simulation
(see Sec.~\ref{sec:w_rec}). While the extracted parameters for position and coupling
did not vary much so that they could be simply averaged, the extracted width was
considerably reduced from around 50 MeV to 30 MeV. This indicates that a significant fraction
of the observed width is related to the experimental resolution and that the intrinsic width is
narrower. The width extracted this way can hence be seen as an approximation of the true width
while the width obtained with the standard fit corresponds to an upper limit only.
With the exception of the parameters extracted from the Kin.~II analysis of the $\eta\rightarrow 3\pi^{0}$
channel, which suffer from a large reduction in statistics, all parameters
corresponding to the kinematic $W$ reconstruction are in good agreement within statistical
uncertainties. A slight improvement in resolution can be seen for the Kin.~II analysis of the
$\eta\rightarrow 2\gamma$ channel leading to smaller parameters for the width. No such effect can be seen
in the $\eta\rightarrow 3\pi^{0}$ data where the reliability of the fit seems to be reduced by lower
statistics. A best estimate was calculated only taking into account the type I analyses,
as they have better statistics than the type II analyses, and the values are shown in
Table \ref{table:fit_params}.
The data from the TOF reconstruction cover only $1/4$ of the solid angle resulting in much
lower statistics. Nevertheless, they serve as an independent check for the presence and the
properties of the structure. Somewhat lower values for the position and width were obtained.
The coupling was estimated assuming an isotropic angular distribution, which
(see Figs.~\ref{fig:ang},\ref{fig:ang_int}) is only a rough approximation, but the results
are in fair agreement with the other analysis.
The smaller width is surprising since the $W$ resolution of the TOF
reconstruction was estimated to be twice that of the kinematic reconstruction, although
this is probably too pessimistic. The reason for the different parameter values will be
discussed below. The fits of the data from the analyses with narrower cuts
are less reliable due to poorer statistics.
Therefore, as for the kinematic reconstruction method, only the type I data were used to deduce
a total best estimate of the parameters.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{fig_21.eps}
\caption{(Color online) Contributions from $S_{11}$ BW (open, black squares), background function (open,
blue triangles), and narrow BW (red dots) for different angular bins integrated over
excitation energy. (The most extreme angles have been omitted because of unstable
fit results due to lack of statistics.)}
\label{fig:ang_int}
\end{figure}
In Fig.~\ref{fig:ang} the differential cross sections are presented as a function of
$W_{\mathrm{kin}}$ for different bins of $\cos(\theta_{\eta}^{*})$. The same phenomenological fits as discussed above
were performed to reveal the angular dependence of the structure. The position varies
between $W_R = 1665$ MeV at backward angles and $W_R = 1680$ MeV at forward angles.
Also the width is reduced at backward angles. This explains the lower values for position
and width also observed in the results of the TOF reconstruction where $-1 < \cos(\theta_{\eta}^{*}) < -0.5$.
On the one hand, the shifting position disfavors the scenario of a single resonance.
On the other hand, it could also be caused by the simplified
ansatz for the phenomenological fitting (which does of course not include any interference
effects). The angular dependence of the strength of the narrow structure is shown in
Fig.~\ref{fig:ang_int}. Due to the simplified ansatz this is only a qualitative indication
for the variation of the strength over the angular distribution. The figure shows
the contributions of the three fit components integrated over the excitation energy.
The `$S_{11}$' contribution shows the expected behavior (since only one BW function was used
this reflects effectively the contribution from the $S_{11}$ and the $S_{11}$-$D_{13}$
interference, which peaks at forward and backward angles and has a minimum around $\cos(\theta_{\eta}^{*}) = 0$).
The phenomenological background subsumes contributions from higher lying $P$- and $D$-states,
their interferences, and non-resonant background and has therefore no simple interpretation.
The angular dependence of the narrow structure does not agree with the most simple scenarios
for its nature, e.g., not with a narrow $P_{11}$ state interfering with the broad $S_{11}$
states. The angular distribution of a $P_{11}$ state is isotropic and the interference term
between $P_{11}$ and $S_{11}$ is proportional to $\cos(\theta_{\eta}^{*})$. The resulting angular distribution
would thus have a maximum at forward angles and a minimum at backward angles or vice versa
(depending on the sign of the interference).
However, Figs.~\ref{fig:ang},\ref{fig:ang_int} show that the structure almost vanishes at
extreme forward angles and is also small at extreme backward angles. Its largest contribution
lies between $\cos(\theta_{\eta}^{*})\pm 0.5$.
A recent fit of the Bonn-Gatchina partial wave analysis \cite{Anisovich_14} reproduced
the peak-like structure in the total cross section and also the angular distributions in the
corresponding energy range. In this solution the bump in the total cross section is caused
by interference effects in the $S_{11}$ partial wave. This interpretation requires a sign
change (relative to the value given by PDG \cite{PDG_12}) of the electromagnetic coupling of
the $S_{11}$(1650) for the neutron.
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\textwidth]{fig_22.eps}
\caption{(Color online) Comparison of Legendre coefficients $A_i$ extracted from fits of the differential
cross sections: quasifree proton from this work (filled blue triangles) and
CBELSA/TAPS \cite{Jaegle_11} (open black triangles), quasifree neutron from this work
(filled red circles) and CBELSA/TAPS (open black circles) and free proton data from
MAMI-C \cite{McNicoll_10} (black stars). Solid lines: MAID model predictions
\cite{Chiang_02}. Dashed vertical lines: markers at $W = 1680$ MeV.}
\label{fig:legendre}
\end{figure*}
Finally, a comparison of the Legendre coefficients $A_i$ is shown in Fig.~\ref{fig:legendre}.
The $A_i$ were obtained by fitting the angular distributions with a series of Legendre polynomials
$P_i$ up to fourth order
\begin{equation}
\frac{d\sigma}{d\Omega}(W, \cos(\theta_{\eta}^{*})) = \frac{q_{\eta}^{*}(W)}{k_{\gamma}^{*}(W)}
\sum_{i=0}^{4} A_i(W) P_i(\cos(\theta_{\eta}^{*}))\,,
\end{equation}
where $q_{\eta}^{*}$ and $k_{\gamma}^{*}$ are the $\eta$ and photon momenta in the
center-of-mass frame, respectively. The data from the exclusive proton and neutron analyses
are plotted along with those from CBELSA/TAPS \cite{Jaegle_11} and the free proton
measurement from MAMI-C \cite{McNicoll_10}. The energy dependences of the proton data of
this work are in general close to the latter, which indicates
that cross sections can be reliably extracted from measurements in quasifree kinematics.
Below $W = 1600$ MeV, some larger discrepancies especially for $A_1$ and $A_3$ are
observed. This is probably caused by the proton detection efficiency which is
problematic in this region, as discussed in Sec.~\ref{sec:cs_extr}. Nevertheless,
above this energy, there is better agreement and also the data from
CBELSA/TAPS are close to our results. The description of the proton data by the
MAID model \cite{Chiang_02} is, as expected, reasonable.
As already seen in the neutron differential cross section data, there are some
discrepancies between the current results and those from CBELSA/TAPS, although the
general trends are confirmed. The most significant discrepancy (note
the logarithmic scale) is in the $A_0$ coefficient, which (apart from the phase-space
factor $q_{\eta}^{*}/k_{\gamma}^{*}$) is proportional to the total
cross section. The sign changes in the vicinity of $W = 1680$ MeV of $A_1$ and $A_2$
are reproduced. In case of the latter, the different signs for proton and neutron
at low energies are due to interference between the \mbox{$S_{11}$(1535)} and the \mbox{$D_{13}$(1520)} resonances
\cite{Weiss_03}. $A_2$ is proportional (neglecting other contributions) to the helicity
couplings $A^N_{1/2}$ of these states, which have equal signs for protons and neutrons for
the \mbox{$D_{13}$(1520)} state, whereas they are opposite for the \mbox{$S_{11}$(1535)} resonance. This is more
or less reproduced by MAID, while the model fails in the description of $A_1$.
In the discussion of
the results from $\eta$ electroproduction the change of sign in $A_1$ was interpreted as
$s$-$p$ wave interference \cite{Denizli_07}. If only $S_{11}$ ($E_{0+}$ multipole) and
$P_{11}$ ($M_{1-}$ multipole) states are considered, $A_1$ would be directly proportional
to $\mathrm{Re}(E_{0+}^*M_{1-})$. A change of sign would then mean that the relative phase
between the two multipoles is changing rapidly due to one of them passing through a resonance. The rough
picture of $A_3$ given by the CBELSA/TAPS measurement is now clarified by the better statistical
quality of the present results. The coefficient seems to be negative throughout the entire energy region.
\section{Summary and conclusions}
Differential and total cross sections of $\eta$ photoproduction off the proton and
the neutron were simultaneously measured in quasifree kinematics on a deuteron target
from threshold up to $E_\gamma = 1.4$ GeV. The $\eta$ mesons were identified using the
two neutral decays $\eta\rightarrow 2\gamma$ and $\eta\rightarrow 3\pi^{0}$. Exclusive measurements were performed by
detecting the recoil nucleons. The total dataset included $4.29\times 10^6$
events of inclusive $\eta$ production as well as $1.86\times 10^6$ events with coincident
protons and $0.63\times 10^6$ events with coincident neutrons. With the inclusive
measurement, the systematic uncertainties of the nucleon detection efficiencies could be
checked via $\sigma_{(N)} = \sigma_{p} + \sigma_{n}$ knowing that coherent contributions
are very small. Effects from Fermi motion were avoided by
a reconstruction of the center-of-mass energy $W$ from the final state. The technical
procedure of a kinematic reconstruction of the kinetic energy of the recoil nucleons was
cross-checked by a time-of-flight measurement. Both methods are only affected by the
corresponding detector resolution, which for the kinematic reconstruction was determined
by MC simulations to be $\Delta W < 40$ MeV (FWHM). The results for $\gamma p\rightarrow \eta p$
are mostly in good agreement with data from inclusive hydrogen measurements taking
into account the poorer resolution in $W$ and effects from the complicated proton efficiency.
The results for $\gamma n\rightarrow \eta n$ are of unprecedented statistical quality and confirm the
existence of a peak in the total cross section at $W_R = (1670\pm 5)$ MeV with a width of
$\Gamma_R = (50\pm 10)$ MeV. Correcting for the finite experimental resolution gave an estimate of
the intrinsic width of $\Gamma_R = (28\pm 5)$ MeV. If the structure would be related to a
single $J=1/2$ state its strengths would correspond to
$\sqrt{b_{\eta}}A_{1/2}^n = (12.3\pm 0.8) \times 10^{-3}\,\mathrm{GeV}^{-1/2}$.
However, the precise differential cross sections revealed that the strength depends
on $\cos(\theta_{\eta}^{*})$; in particular it is suppressed at extreme backward and forward angles which
disfavors such a scenario. More sophisticated partial-wave analyses of the data are
under way. First results in the framework of the BnGn model \cite{Anisovich_14} describe
the data better with a scenario where the main effect is related to interferences in the
$S_{11}$ sector than with the introduction of a narrow $P_{11}$ state. However, also in
this approach contributions from other partial waves are needed to reproduce the non-trivial
angular distributions.
We wish to acknowledge the outstanding support of the accelerator group and operators of
MAMI. This work was supported by Schweizerischer Nationalfonds (200020-132799, 121781,
117601, 113511), Deutsche Forschungsgemeinschaft (SFB 443, SFB/TR 16),
DFG-RFBR (Grant No. 05-02-04014), UK Science and Technology Facilities Council,
(STFC 57071/1, 50727/1), European Community-Research Infrastructure Activity (FP6),
the U.S. DOE, U.S. NSF, and NSERC (Canada).
\bibliographystyle{apsrev}
|
\section{Introduction}
\label{sect:intro}
The Milky Way is known to be a barred spiral galaxy, as shown by
surveys of molecular gas \citep[][]{dht01}, neutral gas
\citep[][]{burt88}, and by infrared star counts \citep[][]{bcb05}. We
do not know its precise structure, e.g., the number and position of
spiral arms, the location and length of the Galactic bar, however. The
obstacle is that the Sun is located on the edge of the Galactic disk
close to the Galactic plane, so that we can only see the superposition
of various structure features (e.g., spiral arms, spurs, arm branches,
and Galactic bars) along the observed line of sight. It is difficult
to disentangle the features and determine the relative position of
spiral tracers accurately.
To outline the precise structure of the Galaxy, a large number of
spiral tracers should be detected over the disk with distances
determined as accurately as possible. Observationally, it has been
shown that the narrow and sharply defined spiral arms of a galaxy
could be well delineated by the distribution of massive star forming
regions or giant molecular clouds \citep[GMCs;
e.g.,][]{iga75,gg76,dwbw80,gcbt88,ggc93,rus03,bt08,hhs09}. HII
regions and 6.7 GHz methanol masers are situated in massive star
forming regions, and closely related to massive young stars. GMCs are
the birthplace of massive stars, and have been widely used to reveal
bright spiral arms and the gas arms of galaxies \citep[][]{bt08}. In
the Milky Way, they are particularly useful in mapping the spiral
pattern, because they are bright and can be widely detected throughout
the Galactic plane \citep[as far as $\sim$20 kpc away from the Sun,
e.g.,][]{abbr11,dt11}. The radio recombination lines from HII
regions, the methanol maser lines near 6.7 GHz, and the molecular
rotational transitions are almost unabsorbed by the interstellar
medium.
Much effort has been dedicated to search for Galactic HII regions,
GMCs, or masers, then to determine their distances to infer the spiral
structure of our Galaxy. A good amount of observational data exist in
literature
\citep[e.g.,][]{dwbw80,ch87,srby87,gcbt88,ahck02,rus03,pdd04,swa+04}. We
have cataloged from the literature 815 HII regions and 963 GMCs with
known trigonometric, photometric, or kinematic distances, and combined
all these data to delineate the spiral structure of the Milky Way
\citep{hhs09}. A polynomial-logarithmic model of spiral arms was
proposed to fit the data distribution and match the observed
tangential directions of spiral arms.
In the last few years, much progress toward understanding the spiral
structure of our Galaxy has been made. First of all, a large number of
new spiral tracers have been detected
\citep[e.g.,][]{rjj+09,abbr11,gm11,gbnd14}. Some newly discovered HII
regions or GMCs even as far as $\sim$20 kpc from the Sun
\citep[][]{abbr11,dt11} provide important data to reveal the spiral
structure in the outer Galaxy. The kinematic distance ambiguity for
many HII regions, GMCs, and 6.7 GHz methanol masers has been solved
\citep[e.g.,][]{rjh+09,lra+11,gm11,abbr12,uhl12,jd12}, which enables
us to estimate their kinematic distances. In addition, the more
reliable trigonometric or photometric distances of a number of HII
regions or masers have been measured
\citep[e.g.,][]{rbr+10,shr+10,mdf+11,xmr+11,xlr+13,rza+12,zms+14,fb14}. The
available spiral tracers have almost {\it doubled} as compared to
those listed in \citet{hhs09}.
Second, some fundamental parameters of the Galaxy have been
observationally modified. Based on the stellar kinematics in the solar
neighborhood, \citet{sbd10} revised the solar motions with respect to
the Local Standard of Rest (LSR) as $U_\odot=11.10$~km~s$^{-1}$
(radially inward), $V_\odot=12.24$~km~s$^{-1}$ (in the direction of
Galactic rotation), and $W_\odot=7.25$~km~s$^{-1}$ (vertically upward,
all in J2000), which are inherently incorporated in the LSR velocities
of spiral tracers, hence influence the estimation of their kinematic
distances. A consistent value of the distance of the Sun to the
Galactic center (GC), $R_0=8.3$~kpc, was proposed by \citet{brm+11},
who incorporated studies on stellar orbits in the GC
\citep[][]{gsw+08,get+09} and trigonometric parallax measurements
toward Sgr\,B2~\citep{rmz+09} and massive star forming regions
\citep{rmz+09b,brm+11}. The circular rotation of the LSR, $\Theta_0$,
was estimated to be 239~km~s$^{-1}$ \citep[][]{brm+11}. Some recent
papers \citep[e.g.,][]{scho12,rmb+14} confirmed the results of
\citet{brm+11}. The values of both $R_0$ and $\Theta_0$ are different
from the IAU recommended values, $R_0=$~8.5~kpc and
$\Theta_0=$~220~km~s$^{-1}$.
The exciting new progress mentioned above motivates us to update the
catalogs of the spiral tracers in \citet{hhs09}, namely, the HII
regions, GMCs, and also to add the 6.7 GHz methanol masers in this
paper, and then to delineate the spiral structure of the Milky
Way. This work is organized as follows. In Sect. 2, we update the
catalogs of spiral tracers and make them publicly available. In
Sect. 3, we discuss the distributions of different kinds of spiral
tracers and their combination. Then, we use the models of three- and
four-logarithmic spiral arms and a model of polynomial-logarithmic
spirals to connect the identified arm segments. Discussions and
conclusions are given in Sect. 4.
\section{Tracers for the Galactic spiral structure}
\label{sect:Obs}
\begin{figure}
\centering\includegraphics[width=0.47\textwidth]{dis_op.ps}
\caption{Distribution of HII regions with photometric (blue) or
trigonometric (red) distances. Two black stars indicate the location
of the Sun ($x=0.0$ kpc, $y=8.5$ kpc) and the GC ($x=0.0$ kpc,
$y=0.0$ kpc). Q1 $-$ Q4 indicate the four Galactic
quadrants. Position uncertainties are indicated by error bars in $X$
and $Y$ directions.}
\label{dis_op_all}
\end{figure}
The primary tracers for spiral arms are HII regions, GMCs, and 6.7 GHz
methanol masers. To outline a precise spiral pattern, the distances of
spiral tracers should be determined as accurately as possible. At
present, the most reliable measurements come from the trigonometric
parallax observations \citep[e.g.,][]{xrzm06,hna+12,rmb+14} and the
photometric method \citep[e.g.,][]{rag07,fb14}. The HII regions with
known photometric or trigonometric distances are shown in
Fig.~\ref{dis_op_all}, which cannot be used to infer the large-scale
spiral pattern of our Galaxy because data are sparse and limited
within about 6~kpc from the Sun.
To uncover the spiral pattern of the Milky Way, we have to involve the
large number of spiral tracers with kinematic distances at
present. The spiral tracers in the inner Galaxy (i.e., the distances
to the GC less than $R_0$) have two possible kinematic distances
corresponding to one $V_{\rm LSR}$, which is known as the kinematic
distance ambiguity. The ambiguity of a HII region can be solved by the
HI/H$_2$CO emission/absorption
observations~\citep[e.g.,][]{was+03,swa+04,ab09} and by the HI
self-absorption method~\citep[e.g.,][]{tl08,ab09}. The latter was also
applied to solve the distance ambiguity for
GMCs~\citep[e.g.,][]{rjh+09} and 6.7 GHz methanol
masers~\citep[e.g.,][]{gm11}.
\subsection{Tracer data}
\subsubsection{ HII regions:}
Based on the distribution of $\sim$100 complexes of H$\alpha$ and/or
H109$\alpha$ emission sources, i.e., bright HII regions,
\citet[][hereafter GG76]{gg76} first proposed a Galactic spiral
pattern with four arm segments. \citet{dwbw80} and \citet{ch87}
discovered more HII regions in the first and fourth Galactic
quadrants, which improved the model of GG76. \citet{ahck02},
\citet{was+03}, and \citet{swa+04} observed H110$\alpha$ and H$_2$CO
4.8 GHz lines simultaneously toward many ultracompact HII (UCHII)
regions, and resolved their kinematic distance ambiguity. The
distribution of UCHII regions traces three arm segments in the first
Galactic quadrant. \citet{rus03} compiled a large sample of $\sim$400
star forming complexes, and fitted the data with models of
two-,~three- and four-logarithmic spirals. She found that the four-arm
model is more appropriate to represent the data
distribution. \citet{pdd04} collected $\sim$180 HII regions with
known distances and found that the distribution is consistent with the
model of GG76.
In the past few years, there have been many HII region
measurements. Notably, Green Bank Telescope HII Region Discovery
Survey\footnote{http$://$www.cv.nrao.edu/hrds/}\citep[][]{babr10,abbr11}
and its extension \citep{abb+13} detected 603 discrete hydrogen radio
recombination line components toward 448 targets. About 60 of them
have negative velocities in the first Galactic quadrant or have
positive velocities in the fourth Galactic quadrant, implying that
they are located in the outer Milky Way or in the 3 kpc Arms
\citep[][]{dt08}. The kinematic distance ambiguity for the 149 newly
discovered HII regions was solved by \citet{abbr12}. The Arecibo HII
Region Discovery Survey \citep[][]{bab12} has detected 37 new HII
regions, kinematic distances were derived for 23 of
them. \citet{dze+11} and \citet{hze+11} observed H110$\alpha$ and
H$_2$CO 4.8 GHz lines simultaneously toward more than 200 candidates
of UCHII regions with the Urumqi 25~m telescope, and obtained
kinematic distances for about 20 HII regions. \citet{lra+11} solved
the kinematic distance ambiguity problem for more than 800 CS~(2$-$1)
emission sources~\citep[][]{bnm96}, which are associated with the
confident candidates of UCHII regions \citep[][]{wc89a}. By analyzing
the HI absorption data toward a sample of compact HII regions
identified from the Red MSX Source Survey \citep{hlo+05},
\citet{uhl12} solved the kinematic distance ambiguity for about 105
HII regions. Similar work was conducted by \citet{jd12} for 75 HII
regions. In addition, the relative accurate photometric or
trigonometric distances have been measured or revised for some HII
regions. \citet{pco10} identified the ionizing stars in nine HII
regions, and derived their photometric distances. \citet{mdf+11} used
the near-infrared color images to study the stellar content of 35 HII
regions, and estimated their spectrophotometric distances.
\citet{rza+12} analyzed the data of OB stars in two star forming
complexes \citep[see][]{rus03}, NGC\,6334 and NGC\,6357, and revised
their photometric distances. \citet{fb14} obtained the
spectrophotometric distances for 103 HII regions in the second and
third Galactic quadrants. By measuring the trigonometric parallax of
associated masers
\citep[e.g.,][]{rbr+10,okh+10,shr+10,nnh+11,shd+11,srd+12,irm+13,xlr+13,zrm+13,wsr+14},
the distances of about 30 HII regions have been determined with high
accuracy\footnote{The Bar and Spiral Structure Legacy (BeSSeL)
Survey. http$://$bessel.vlbi-astrometry.org}\fnmsep\footnote{The
Japanese VLBI Exploration of Radio Astrometry
(VERA). http$://$veraserver.mtk.nao.ac.jp}.
We collect the HII regions with line measurements from the
references. The catalog contains about 4500 line measurements toward
more than 2500 HII regions or confident candidates. More than 1800 of
them have their distances determined. The photometric or trigonometric
distance and the solutions of the kinematic distance ambiguity if
available are collected from literature and listed in
Table~\ref{tab_a1}.
\subsubsection{GMCs}
Giant molecular clouds have proven to be good tracers of the spiral
structure \citep[][]{cdt86,dect86,gcbt88}. The famous
Sagittarius-Carina Arm was clearly delineated by the distribution of
GMCs \citep{gcbt88}. \citet{srby87} identified 268 GMCs in the first
Galactic quadrant and outlined three arm segments. \citet{efre98}
collected the known GMCs and fitted the data with a four-arm
model. \citet{rjh+09} solved the kinematic distance ambiguity of 750
GMCs identified from the Galactic Ring Survey of $^{13}$CO (1$-$0)
\citep[][]{jrs+09,rjj+09}, and showed three arm segments in the
distribution of GMCs. \citet{dt11} identified ten $^{12}$CO (1$-$0)
emission components having negative $V_{\rm LSR}$ in the first
Galactic quadrant, which are probably located in a spiral arm lying
outside the Outer Arm. For the fourth Galactic quadrants, a
preliminary identification of GMCs in the inner Galaxy was made by
\citet{bnt89}. Recently, \citet{gbnd14} reanalyzed the Columbia $-$
U. de Chile Survey data of CO, and identified 92 GMCs. The kinematic
distance ambiguity for 87 of them has been solved.
The GMC catalog in \citet{hhs09} is now updated to include these new
results. Molecular clouds with a mass greater than 10$^4$ $M_\odot$
are regarded to be massive enough and considered as GMCs in this
work. We collect about 1300 GMCs as listed in Table~\ref{tab_a2}, and
more than 1200 of them have distances determined.
\begin{figure*}[t]
\centering\includegraphics[width=170mm]{lv_hii_maser_gmc.ps}
\caption{Longitude-velocity diagrams for the HII regions ({\it
upper}), GMCs ({\it middle}), and 6.7 GHz methanol masers ({\it
lower}), which are overlaid on the image of $^{12}$CO~(1$-$0) of
\citet[][the IAU standard solar motions are adopted]{dht01}. The
open squares indicate HII regions with the symbol area proportional
to the excitation parameters (see Eq.~\ref{eq_U}). The open circles
indicate the GMCs with the symbol size proportional to
log($M_{GMCs}$) (see Sect. 2.4). The black crosses indicate the
6.7 GHz methanol masers with a constant symbol size. A spiral arm
model (see the left panel of Fig.~\ref{model_pol}) is overlaid to
indicate the locations of the identified arm segments: arm-1: the
Norma Arm; arm-2: the Scutum-Centaurus Arm; arm-3: the
Sagittarius-Carina Arm; arm-4: the Perseus Arm; arm-5: the Outer
Arm; arm-6: the Outer+1 Arm; and the Local Arm is also given. }
\label{lvall}
\end{figure*}
\subsubsection{6.7 GHz methanol Masers}
The 6.7 GHz methanol masers \citep{ment91} have been widely detected
from massive star forming regions in our Galaxy
\citep[e.g.,][]{pmb05,elli06,gcf+10,gcf+12}. Early detections of 519
6.7 GHz methanol masers were summarized by \citet{pmb05}. Since then,
\citet{pgd07} detected 86 masers in a 18.2 deg$^2$ area with the
Arecibo, and 48 of them are new detections. \citet{elli07} observed
200 {\it Spitzer}/GLIMPSE sources with Mt Pleasant 26~m and Ceduna
30~m radio telescopes, and they detected methanol masers in 38 sites,
nine of which are new. \citet{xlh+08} searched for 6.7 GHz methanol
masers toward a sample of 22 GHz water maser sources with the
Effelsberg 100~m telescope, and got ten new detections. \citet{cbhc09}
measured 18 methanol masers toward a sample of young stellar objects
with the VLA. The most important progress of the 6.7 GHz methanol
maser discovery is the Methanol Multibeam (MMB) Survey of the entire
Galactic plane
\citep[e.g.,][]{gcf+09,gcf+10,gcf+12,cfg+10,cfg+11}. \citet{fcf10}
searched for the 6.7 GHz methanol masers toward 296 massive star
forming regions, and detected 55 masers, 12 of which are
new. \citet{oah+13} performed a sensitive search of methanol masers
toward 107 sources discovered in the {\it Herschel} Infrared Galactic Plane
Survey \citep{msb+10}, and detected 32 masers, 22 of which are
new. \citet{sxc+14} searched for 6.7 GHz methanol masers toward 318
dust clumps, and found 29 masers, 12 of which are new. In addition,
the distances for many 6.7 GHz methanol masers have been
determined. \citet{pmg09} solved the kinematic distance ambiguity for
86 masers with the HI self-absorption method. \citet{gm11} collected a
large sample of 6.7 GHz methanol masers from references and new
detections from the MMB survey. Their database covers the longitude
ranges of 270$^\circ \leq$ $l$ $\leq$ 358$^\circ$, 5$^\circ \leq$ $l$
$\leq$ 67$^\circ$, and the latitude range of $|b| \leq$
1.5$^\circ$. The kinematic distance ambiguity for more than 400
methanol masers was solved by the HI self-absorption method.
We collect more than 900 methanol masers from the references above
(see Table~\ref{tab_a3}), about 750 of which have distances
determined, mostly kinematic, though some were measured by the
trigonometric parallax or the photometric method. The observed spectra
of 6.7 GHz methanol masers toward massive star forming regions always
have multicomponents. The velocities of the maser spots toward a
source are in a range of typically a few km~s$^{-1}$, sometimes more
than 10 km~s$^{-1}$~\citep[e.g., see][]{gcf+10}. The median velocity
instead of the peak velocity was suggested to be a better proxy of the
systematic velocity of the source~\citep{gm11}. Therefore, in our work,
we use the median velocities to calculate the kinematic distances of
masers. The systematic velocities and hence the distances determined
from the observed maser spectra may be affected by noncircular motions
and therefore have large uncertainties.
\subsubsection{Longitude-velocity diagram of spiral tracers}
The gas component in the Milky Way has been systematically observed
with the CO rotational transitions or the HI 21cm line. The
longitude-velocity diagrams represent the kinematic and dynamic
features of interstellar gas, which offer observational constraints on
the spiral structure \citep[e.g.,][]{dht01}. The key step in {\it
mapping} the gaseous spiral arms is to reconstruct the gas
distribution from the observed longitude-velocity maps. In the inner
Galaxy, the distances of gas components are difficult to
determine. The spiral structure derived from diffuse HI and CO
distributions is far from clear at present \citep[e.g.,
see][]{ns03,ns06}.
In contrast, HII regions, GMCs, and masers serve as excellent tracers
of spiral arms. Their distances can be obtained by trigonometric,
photometric, or kinematic method. As shown in Fig.~\ref{lvall},
almost all features in the longitude-velocity map of $^{12}$CO~(1$-$0)
\citep{dht01}\footnote{http$://$www.cfa.harvard.edu/mmw/MilkyWayinMolClouds.html}
have counterparts in the distributions of collected HII regions and
masers. With distances, one can reveal the unknown spiral arms.
\subsection{Fundamental parameter $R_0$, $\Theta_0$, and solar motions}
For most of the spiral tracers without photometric or trigonometric
distances, we have to obtain their kinematic distances, which depend
on the adopted rotation curve and the fundamental parameters of the
Galaxy, including the distance to the GC, $R_0$, the circular orbital
speed at the Sun, $\Theta_0$, and the solar motions with respect to
the LSR.
According to the IAU standard, the distance to the GC is
$R_0=8.5$~kpc, and the circular orbital speed at the Sun is
$\Theta_0=220$~km~s$^{-1}$. By measuring the S-star orbits in the GC
regions, \citet{gsw+08} obtained $R_0=8.4\pm0.4$~kpc, and
\citet{get+09} estimated $R_0=8.33\pm0.35$~kpc. By measuring the
trigonometric parallax toward Sgr\,B2, \citet{rmz+09} obtained a value
of $R_0=7.9^{+0.8}_{-0.7}$~kpc. \citet{rmz+09b} and \citet{brm+11}
fitted the parallax measurements of massive star forming regions and
derived the best fitted value of $R_0=8.4\pm0.6$~kpc and
$\Theta_0=254\pm16$~km~s$^{-1}$. A weighted average of
$R_0=8.3\pm0.23$~kpc was derived from the above measurements by
\citet{brm+11}, and the fitted value of $\Theta_0$ is in the range of
223$-$280~km~s$^{-1}$ for different rotation curves. The ratio of
$\Theta_0/R_0$ has been well constrained to be
$29.4\pm0.9$~km~s$^{-1}$~kpc$^{-1}$. \citet{brm+11} suggested a value
of $\Theta_0=239\pm7$~km~s$^{-1}$. \citet{scho12} obtained
$R_0=8.27\pm0.29$~kpc and $\Theta_0=238\pm9$~km~s$^{-1}$ by analyzing
the stellar kinematics. \citet{hna+12} found $R_0=8.05\pm0.45$~kpc and
$\Theta_0=238\pm14$~km~s$^{-1}$, using the VLBI astrometry data of 52
Galactic masers. \citet{rmb+14} estimated that $R_0=8.34\pm0.16$~kpc
and $\Theta_0=240\pm8$~km~s$^{-1}$ by analyzing the new parallax
measurements toward massive star forming regions. \citet{cfg+14} found
$R_0=8.36\pm0.11$~kpc with a statistical study on star cluster
dynamics and S-star orbits.
\begin{figure}[!b]
\centering\includegraphics[width=0.47\textwidth]{te.ps}
\caption{Distribution of estimated electron temperature ($T_e$) of
HII regions from radio recombination line and radio continuum
observations in Table~\ref{tab_a1}.}
\label{dis_te}
\end{figure}
\begin{figure}[!t]
\centering\includegraphics[width=0.47\textwidth]{Uhist.ps}
\caption{Distributions of the excitation parameters of HII regions
({\it upper}) and the masses of GMCs ({\it lower}). We adopted the
IAU standard $R_0=8.5$~kpc and $\Theta_0=220$~km~s$^{-1}$ and the
standard solar motions together with a flat rotation curve in
deriving the kinematic distances from velocities if no photometric
or trigonometric distance is available.}
\label{weight}
\end{figure}
The IAU standards of solar motions with respect to the LSR are
$U_\odot=10.27$~km~s$^{-1}$, $V_\odot=15.32$~km~s$^{-1}$, and
$W_\odot=7.74$~km~s$^{-1}$ in J2000. The values of $U_\odot$ and
$W_\odot$ do not change considerably in the recent research. But the
solar motion in the direction of Galactic rotation, $V_\odot$, is
under debate. By analyzing the Hipparcos data, \citet{db98} obtained
the solar motions $U_\odot=10.00\pm0.36$~km~s$^{-1}$,
$V_\odot=5.25\pm0.62$~km~s$^{-1}$, and
$W_\odot=7.17\pm0.38$~km~s$^{-1}$. According to these parameters,
\citet{rmz+09b} analyzed the parallax measurements toward 18 massive
star forming regions, and they found a velocity difference of
$\sim$15~km~s$^{-1}$ in rotation, which may be induced by an
erroneous value of $V_\odot$. Higher values of
$V_\odot=12.24-14.6$~km~s$^{-1}$ are also suggested by some recent
studies \citep[e.g.,][]{fa09,sbd10,cab+11,rmb+14}. Based on the solar
motion parameters given by \citet[][]{sbd10}, \citet{brm+11}
reanalyzed the trigonometric parallax data of massive star forming
regions, and found the peculiar rotation velocity of
$\sim$$8\pm2$~km~s$^{-1}$.
Throughout this work, we adopt two sets of $R_0$, $\Theta_0$, and
solar motions: one is the IAU standard $R_0=8.5$~kpc and
$\Theta_0=220$~km~s$^{-1}$, and solar motions of
$U_\odot=10.27$~km~s$^{-1}$, $V_\odot=15.32$~km~s$^{-1}$, and
$W_\odot=7.74$~km~s$^{-1}$; the other is $R_0=8.3$~kpc and
$\Theta_0=239$~km~s$^{-1}$ \citep[][]{brm+11,rmb+14}, and solar
motions of $U_\odot=11.10\pm1.2$~km~s$^{-1}$,
$V_\odot=12.24\pm2.1$~km~s$^{-1}$ and $W_\odot=7.25\pm0.6$~km~s$^{-1}$
\citep{sbd10}. A flat rotation curve \citep[][]{rmz+09b,rmb+14,brm+11}
and the rotation curve of \citet[][hereafter BB93]{bb93} are adopted
in calculating the kinematic distances.
\subsection{Distances of the Galactic spiral tracers}
If the photometric or trigonometric distance is available for a
tracer, we adopt it directly. Otherwise, the kinematic distance is
estimated using the observed $V_{\rm LSR}$ and a rotation curve with
the adopted $R_0$, $\Theta_0$, and solar motions. For GMCs and 6.7 GHz
methanol masers, their kinematic distances are calculated in the same
way as that for HII regions.
Some HII regions have more than one measurement of $V_{\rm LSR}$from
the same or different emission lines (Table~\ref{tab_a1}). Here, the
mean velocity ($V_{\rm LSR}$=$\Sigma\frac{V_i}{N}$) is adopted. We
revise the mean $V_{\rm LSR}$ according to the adopted solar motions,
then calculate the kinematic distance with a flat rotation curve
\citep[][]{rmz+09b,rmb+14,brm+11} or the rotation curve of BB93. The
FORTRAN package kindly provided by \citet{rmz+09b} was modified and
used in our calculations. The sources in the range of Galactic
longitudes of $-$15$^\circ$ to 15$^\circ$ should be carefully
calculated for their kinematic distances \citep[][]{tw14} because of
the Near and Far 3 kpc Arms \citep[e.g.,][]{dt08,gcm+11,jdd+13} and
the Galactic bar(s) \citep[][]{hgm+00,cbm+09}. We first inspect their
possible associations with the Near and Far 3 kpc Arms according to
the arm parameters given by \citet[][]{dt08}, and then calculate their
distances.
Some HII regions, GMCs, and/or 6.7 GHz methanol masers are probably
associated with each other. To remove the redundancy, the associations
are identified following these criteria: (1) if an association was
identified in literature, we adopted it directly; (2) if a HII
region/maser is located within a GMC in projection considering its
measured angular size, and their velocity difference is
$\leq$~10~km~s$^{-1}$ \citep{fdt90}, we regard them as an association;
(3) if the coordinates of a HII region and a maser are closer than
2$^{\prime}$, and their velocity difference is $\leq$~10~km~s$^{-1}$
\citep{gm11}, they are regarded as association. We checked the
associations one by one, and adopted the distances of the HII regions
or GMCs in the associations preferentially for the distributions
because of the larger uncertainties of the systematic velocities derived
from the 6.7 GHz methanol maser spectra. For the associations of HII
regions and GMCs, we adopted the parameters of HII regions
(coordinates and weights) for the distributions.
\subsection{Weights for tracers of Galactic spiral structure}
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{dis_hii_8.5_err.ps}
\centering\includegraphics[width=0.47\textwidth]{dis_hii_8.5.ps}\\
\centering\includegraphics[width=0.47\textwidth]{dis_hii_8.5_err_1kpc.ps}
\centering\includegraphics[width=0.47\textwidth]{polyarm_hii_8.5_show.ps}
\caption{{\it Upper panels}: distributions of HII regions projected
into the Galactic plane with ({\it left}) and without ({\it right})
position error bars. {\it Lower left}: HII region distribution
for those only with position uncertainties better than 1~kpc. {\it
Lower right}: HII region distribution overlaid with a spiral
arm model (see the {\it left panel} of Fig.~\ref{model_pol}) to indicate
the identified arm segments. The area of open squares is
proportional to the excitation parameters (see Eq.~\ref{eq_U}). The
IAU standard $R_0=8.5$~kpc and $\Theta_0=220$~km~s$^{-1}$ and the
standard solar motions together with a flat rotation curve are
adopted in deriving the kinematic distances from velocities if no
photometric or trigonometric distance is available. The coordinates
originate from the GC, and the Sun is located at $x=$~0.0~kpc,
$y=$~8.5~kpc. The open red ellipse indicates the Galactic bar
\citep{cbm+09}, and the laurel-gray ellipse indicates the
best-fitted Near 3 kpc Arm and Far 3 kpc Arm \citep{gcm+11}.}
\label{dis_hii}
\end{figure*}
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{dis_gmc_8.5_err.ps}
\centering\includegraphics[width=0.47\textwidth]{dis_gmc_8.5.ps}\\
\centering\includegraphics[width=0.47\textwidth]{dis_maser_8.5_err.ps}
\centering\includegraphics[width=0.47\textwidth]{dis_maser_8.5.ps}\\
\caption{{\it Upper panels}: distributions of GMCs with ({\it left})
and without ({\it right}) position error bars. {\it Lower panels}:
distributions of 6.7 GHz methanol masers with ({\it left}) and
without ({\it right}) position error bars. The symbols for GMCs and
masers are the same as those in Fig.~\ref{lvall}. The open red
ellipse and the laurel-gray ellipse are the same as those in
Fig.~\ref{dis_hii}.}
\label{dis_gmc_mas}
\end{figure*}
In order to use available tracers to outline the spiral structure, an
appropriate weight factor should be used \citep{hhs09} to represent
the relative importance of a tracer for the spiral
structure. Reasonably, the brighter a HII region, or more massive a
GMC, the more important a tracer for spiral structure, and therefore a
larger weight is assigned.
For HII regions, we adopt the excitation parameter \citep{sm69} as the
weight. The excitation parameter $U$ (in pc~cm$^{-2}$) is defined as:
\begin{equation}
U=4.5526\{\alpha(\nuup,T_e)^{-1}\nuup^{0.1}T_e^{0.35}S_{\nuup}D^{2}\}^{1/3},
\label{eq_U}
\end{equation}
here, $T_e$ is the electron temperature in K; $S_{\nu}$ is the radio
flux density in Jy, collected from literature
\citep[e.g.,][]{dwbw80,ch87,gwbe95,kc97} if available; $\nu$ is the
observed frequency in GHz; $D$ is the distance to the tracer in kpc;
and $\alpha(\nu,T_e)$ is a parameter close to 1~\citep[][]{sm69}, and
for simplicity, we adopt $\alpha(\nu,T_e)=1$. For a HII region, if
$T_e$ is given in the literature, we adopt it directly. The mean value
is used if there is more than one measurement, see
Table~\ref{tab_a1}. Otherwise, we adopt the most probable value of
$T_e=$~6500~K (see Fig.~\ref{dis_te}). The excitation parameter $U$
is calculated via Eq.~\ref{eq_U}, and the distribution of $U$ is shown
in Fig.~\ref{weight}, which ranges from 1 pc~cm$^{-2}$ to $\sim$300
pc~cm$^{-2}$ with a peak near 50 pc~cm$^{-2}$. The ranges of size or
luminosity of the Galactic HII regions are comparable to those in
other spiral galaxies \citep[e.g., M31, M51,][]{amb11}. In the catalog
of Galactic HII regions in Table~\ref{tab_a1}, ultracompact HII
regions are included, though they have much smaller sizes and radio
flux density, but they are important to indicate spiral structure in
some regions.
The weight of a HII region is defined as $W_{HII} = \frac{U}{U_0}$,
here $U_0 = 100$ pc~cm$^{-2}$. Clearly, $W_{HII}$ has a value in the
range of 0.0$-$3.0. For those HII regions without measurements of
$S_{\nu}$, a weight factor $W_{HII}=0.1$ is assigned to exploit the
distribution of entire HII region dataset.
For GMCs, we use their masses as the weight factor
\citep[see][]{hhs09}. The mass of each molecular cloud from literature
was re-scaled with the adopted distances in Table~\ref{tab_a2}, and
the distribution of $M_{GMCs}$ in the ranges of 1$\times 10^4
M_{\odot}$ to $\sim$10$^7 M_{\odot}$ is shown in the lower panel of
Fig.~\ref{weight}. The range of GMC mass in our Galaxy is comparable
to nearby spiral galaxies, e.g., M31 \citep[][]{kgf+14}. Here, we
adopt $W_{GMCs}=\log(M_{GMCs}/10^4M_{\odot})$ as the weight, and their
values are in the ranges from 0.0$-$3.0. For the methanol masers, we
simply assign the weight factor $W_{m}=0.1$ for each of them.
We also consider the weight due to the distance uncertainty of each
tracer. A weight factor $W_x=0.5/\sigma_x$ and $W_y=0.5/\sigma_y$ is
assigned to each tracer \citep{hhs09}. If the uncertainty of the
photometric or trigonometric distance is available in reference, we
adopt it directly. If the kinematic distance is used, we assume a
systematic velocity uncertainty of $\pm$~7~km s$^{-1}$, and then the
distance uncertainty is calculated via the adopted rotation curve and
the Galaxy fundamental parameters. Furthermore, if a tracer has a
distance accuracy better than 0.5 kpc in $x$ or $y$ direction, the weight
is assigned to be 1.0, i.e., $W_x$ = 1.0 if $\sigma_x< 0.5$ kpc;
and/or $W_y$ = 1.0 if $\sigma_y < 0.5$ kpc.
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{dis_all_8.5_err.ps}
\centering\includegraphics[width=0.47\textwidth]{dis_all_8.5.ps}\\
\centering\includegraphics[width=0.47\textwidth]{Rdis_contour_8.5.ps}
\centering\includegraphics[width=0.47\textwidth]{hhs09_8.5_flat_RC.ps} \\
\caption{{\it Upper panels}: distributions of HII regions, GMCs, and
6.7 GHz methanol masers with ({\it left}) and without ({\it
right}) position error bars. The symbols for HII regions, GMCs,
and masers are the same as those in Fig.~\ref{lvall}. {\it
Lower left}: color map of all tracers brightened with a Gaussian
function with the amplitude of a weight factor, compared to the
color map in the {\it lower right panel} which is constructed using the
815 HII regions and 963 GMCs collected in our previous work
\citep{hhs09}. }
\label{dis_all}
\end{figure*}
\section{The spiral structure of the Milky Way}
\subsection{Spiral structure revealed by different tracers}
\subsubsection{HII regions}
The large dataset of HII regions gives a clear presentation of the
spiral pattern, see Fig.~\ref{dis_hii}. Some remarkable features are
discussed below.
The Local Arm, where the Sun is located, is a remarkable arm segment
as indicated by a black-solid line in the lower-right panel of
Fig.~\ref{dis_hii}. It starts near the Perseus Arm ($x \sim$5 kpc,
$y \sim$7.5 kpc), and seems to have two branch-like structures near
the Sun, one to $x \sim$ $-$1 kpc and $y \sim$9 kpc close to the
Perseus Arm, above the fitted position of the Local Arm (see
Fig.~\ref{dis_hii}), the other to $x \sim$ $-$1.5 kpc and $y \sim$8.5
kpc close to the Carina Arm and overlapped with the fitted Local
Arm. \citet{xlr+13} suggested that the Local Arm is closer to the
Perseus Arm than to the Sagittarius Arm, and very likely is an arm
branch from the Perseus Arm.
In the first Galactic quadrant (Q1), there are three obvious arm
segments. From the inside out, they are the Scutum Arm, the
Sagittarius Arm and the Perseus Arm. In the outer Galaxy regions ($x$
from $\sim$7 kpc to $14$ kpc, $y$ from $\sim$ $-$10 kpc to $-$1 kpc),
there are indications for the Outer Arm and even the Outer+1 Arm. In
the inner Galaxy ($x \sim$2~kpc, $y \sim$3~kpc), the assemblies of
HII regions may be related with the near ends of the Galactic bar
and/or the 3 kpc Arms.
In the second (Q2) and third (Q3) Galactic quadrants, i.e., the
anticentral regions, a remarkable arm segment ($x$ from $-$4 kpc to 4
kpc, $y\sim$10 kpc) is the Perseus Arm. Outside the Perseus Arm,
there are the extension of the Outer Arm or even that of the Outer+1
Arm. The kinematic distance anomalies \citep{robe72} in the second and
third Galactic quadrants do not significantly influence the structure
features shown in Fig.~\ref{dis_hii} because most of the bright HII
regions have photometric or trigonometric distances
(cf. Fig.~\ref{dis_op_all}).
In the fourth Galactic quadrant (Q4), three arm segments are
obvious. From the outside in, they are the Carina Arm, the Centaurus
Arm and the Norma Arm. The assemblies of HII regions in the inner
Galaxy ($x \sim$ $-$2~kpc, $y \sim$ $-$3~kpc) maybe related to the far
ends of the Galactic bar and/or the 3 kpc Arms.
The large uncertainties of kinematic distances of HII regions may
result in some fuzzy structure features. In the lower left panel of
Fig.~\ref{dis_hii}, only the HII regions with position uncertainties
better than 1 kpc are plotted. The separation between the adjacent arm
segments is clearer, typically larger than the position uncertainties
of spiral tracers, which confirms the existence of the arm segments
discussed above.
\begin{figure}
\centering\includegraphics[width=0.47\textwidth]{dis_all_HI_8.5.eps}
\caption{Spiral arm tracer distribution overlaid on the HI map
of \citet{lbh06}. The symbols for HII regions, GMCs, and masers
are the same as those in Fig.~\ref{lvall}.}
\label{dis_hi}
\end{figure}
\subsubsection{GMCs}
Arm features traced by GMCs are shown in
Fig.~\ref{dis_gmc_mas}. Most of the arm features recognized in the
distribution of HII regions have their counterparts in the
distribution of GMCs.
In the first Galactic quadrant, we find three arm segments, which are
the Scutum Arm, the Sagittarius Arm and the Perseus Arm from the
inside out. In the distant regions of the first quadrant
($D\gtrsim$~20~kpc, $l\sim10\degr-60\degr$), several CO emission
features were revealed by \citet{dt11}. One fully mapped CO emission
source has a mass of $\sim$5~$\times$~10$^4$ $M_\odot$. Other CO
emission features are not yet fully mapped, but are also probably from
GMCs. These GMCs may indicate the arm segment(s) beyond the Outer Arm.
Some GMCs ($x\sim$12~kpc, $y\sim$4~kpc) identified by \citet{dtb90}
may trace the Outer Arm or the Outer+1 Arm.
In the second and third Galactic quadrants many GMCs exist that were
identified by \citet{mab97}, \citet{hcs01}, and \citet{nomf05}. But
arm-like features are not obvious probably due to the anomalies of
kinematic distances \citep{robe72}, as most of their distances are
kinematic.
In the fourth quadrant, there are two obvious arm segments. From the
outside in, they are the Carina Arm and the Centaurus Arm. The
assemblies of GMCs near $x \sim$ $-$2~kpc and $y \sim$3~kpc may trace
part of the Norma Arm. The GMCs near $x \sim$ $-$3~kpc and $y \sim$
$-$2~kpc may be related to the extension of the Norma Arm and/or with
the Near 3 kpc Arm.
\subsubsection{6.7 GHz methanol masers}
The distribution of 6.7 GHz methanol masers (Fig.~\ref{dis_gmc_mas})
resembles that of HII regions or GMCs but with a much larger data
scatter. Most of their distances are kinematic, which depends on the
measured $V_{\rm LSR}$. As discussed in Sect. 2.1.3, the systematic
velocity derived from the observed maser lines may be significantly
affected by noncircular motions of the maser spots, and hence has
larger uncertainties of derived distances than those for HII regions
and GMCs.
\subsubsection{Spiral structure seen from tracers}
\begin{figure}
\centering\includegraphics[width=0.47\textwidth]{log_seta.ps}
\caption{ Distributions of HII regions ({\it upper}), GMCs ({\it
middle}), and 6.7 GHz methanol masers ({\it lower}) in the
log($r$) $-$ $\theta$ diagram. The symbols for HII regions, GMCs,
and masers are the same as those in Fig.~\ref{lvall}. Here, $r$
is the distance to the GC, $\theta$ starts from the positive
$x$-axis and increases counterclockwise. The arm segments are
identified and roughly separated by the dashed-lines. }
\label{log_seta}
\end{figure}
The distribution of three kinds of spiral tracers is shown in
Fig.~\ref{dis_all}, after removing the redundancy, as described in
Sect. 2.3.
To better indicate spiral arms, following \citet{hhs09},
we use a Gaussian function to brighten each tracer:
\begin{equation}\label{eq:LebsequeI}
L(x,y)=\sum_i {\frac{W(i)} {2\pi
\sigma^2}}exp(-{\frac{(x_{i}-x)^{2}+(y_{i}-y)^{2}} { 2
\sigma^2}}).
\end{equation}
Here, $W(i)=W_{HII}$ or $W_{GMCs}$ or $W_{m}$ is the weight factor
discussed in Sect. 2.4; $x_i$ and $y_i$ are the coordinates of the
$i$th spiral tracer. A color intensity map is shown in the lower left
panel of Fig.~\ref{dis_all}, in which $\sigma$ = 0.2 is adopted as
the Gaussian width. Different values of $\sigma$ in a suitable range
(e.g., 0.05$-$0.4) yield a similar result. Several distinct arm
segments are clearly shown. Comparing to the previous arm structure in
\citet[][see the lower right panel of Fig.~\ref{dis_all}]{hhs09}, we
found that the Outer Arm, part of the Outer+1 Arm, the Local Arm, the
Centaurus Arm and the Norma Arm are significantly better outlined by
more tracers.
Comparison of the tracer distribution with the distribution of HI gas
is shown in Fig.~\ref{dis_hi}, which gives us an intuitive feeling
about the extension of the identified arm segments. In the outer
regions of the first quadrant ($x \sim$12~kpc, $y \sim$ $-$8~kpc), tens
of newly discovered HII regions \citep{abbr11} and GMCs \citep{dt11}
are consistent with an HI arm-like feature, and then connected to a
cluster of GMCs ($x \sim$12~kpc, $y \sim$4~kpc). In the third and
fourth Galactic quadrants, from the outside in, the three remarkable
HI arms correspond to the extensions of the Outer Arm ($x$ from
$\sim$ $-$13 kpc to $-$7~kpc, $y\sim$12~kpc, $l\sim200\degr-270\degr$),
the Perseus Arm ($x\sim$ $-$7~kpc, $y\sim$9~kpc, $l\sim275\degr$), and
the Carina Arm ($x\sim$ $-$10 kpc). In particular, the Carina Arm traced
by HII regions and GMCs is well matched with distribution of HI gas.
In summary, the distribution of HII regions and GMCs do show some
obvious arm segments. Particularly, these arm segments are able to
match the HI arms in the outer Galaxy. However, the connections of arm
segments in different Galactic quadrants are still not intuitive.
\subsection{Fitting models to tracer distributions }
We can roughly identify the arm segments in the log($r$)$-$$\theta$
diagrams as separated by the dashed lines in
Fig.~\ref{log_seta}. From the bottom up, they are the Norma Arm
(arm-1), the Scutum-Centaurus Arm (arm-2), the Sagittarius-Carina Arm
(arm-3), the Local Arm, the Perseus Arm (arm-4), the Outer Arm (arm-5)
and the Outer+1 Arm (arm-6). Near the direction of $\theta \sim$90$^\circ$,
the spiral tracers are mixed, the Scutum-Centaurus Arm and
the Sagittarius-Carina Arm can barely be distinguished. Near the
direction of $\theta \sim$320$^\circ$, where the data points are also
mixed, the different arm segments cannot be separated with high
confidence. In addition, there is a lack of tracers at large $r$, the
Outer Arm and the Outer+1 Arm can only be identified marginally. The
Local Arm ($\theta \sim$$50^\circ-110^\circ$, log($r$) $\sim$2.1) is
obviously short and has a distinctive pitch angle as recognized in the
distribution of HII regions. To explore the number and position of
spiral arms in our Galaxy, fitting the tracer data with models is
desired.
Many models of spiral structure have been proposed
\citep[e.g.,][]{val08,val13}. The paradigmatic model is the four arm
segments first recommended by GG76, which is supported by observations
later~\citep[e.g.,][]{dwbw80,ch87,eg99,ds01,rus03,lbh06,bb14}, though
other spiral patterns have been proposed to describe the structure of
the Galaxy \citep[see, e.g.,][]{al97,lmd01,lra+11,peb08,hhs09}. Here,
we first fit the tracer data with the conventional logarithmic
spirals, and then present a model with polynomial-logarithmic spiral
arms to connect the tracers.
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{3arm_hii_8.5.ps}
\centering\includegraphics[width=0.47\textwidth]{4arm_hii_8.5.ps}\\
\centering\includegraphics[width=0.47\textwidth]{3arm_all_8.5.ps}
\centering\includegraphics[width=0.47\textwidth]{4arm_all_8.5.ps}\\
\caption{{\it Upper panels}: best-fitted three-arm model ({\it
left}) and four-arm model ({\it right}) for the distribution of
HII regions. {\it Lower panels}: best-fitted three-arm model
({\it left}) and four-arm model ({\it right}) for the distribution
of three kinds of spiral tracers. The symbols are the same as those
in Fig.~\ref{lvall}. The open red ellipse indicates the Galactic
bar \citep{cbm+09}, and the laurel-gray ellipse shows the
best-fitted Near 3 kpc Arm and Far 3 kpc Arm \citep{gcm+11}. The
dashed-lines indicate the observed tangential directions (see
Table~\ref{tan_obs}).}
\label{model_log}
\end{figure*}
\setlength{\tabcolsep}{0.8mm}
\begin{table*}[!t]
\caption{Arm parameters of the best-fitted models of logarithmic
spirals (see Eq.~\ref{eq_log}).}
\begin{center}
\begin{tabular}{lcccccccccccccccc}
\hline
\hline
Models$\setminus$Arm parameters & $R_{1}$&$\theta_{1}$& $\psiup_{1}$ & $R_{2}$ & $\theta_{2}$ &
$\psiup_{2}$ & $R_{3}$ & $\theta_{3}$ & $\psiup_{3}$ & $R_{4}$ &
$\theta_{4}$& $\psiup_{4}$ & $R_{5}$ & $\theta_{5}$ & $\psiup_{5}$& $Z$ \\
& (kpc)&$(^{\circ})$& $(^{\circ})$ & (kpc) & $(^{\circ})$ &
$(^{\circ})$ & (kpc) & $(^{\circ})$ & $(^{\circ})$ &
(kpc)& $(^{\circ})$& $(^{\circ})$& (kpc) &$(^{\circ})$ & $(^{\circ})$&\\
\hline
\hline
\multicolumn{14}{l}{Fitting of models to HII regions, $R_0=8.5$~kpc, $\Theta_0=220$~km~s$^{-1}$.} \\
3-arm model& {3.42} & {44.4} & { 8.85} & {3.72} & {192.3} & {9.29} & {3.01} & {219.4} & {8.62}& & & & {8.61}& {51.3}& {0.50} &{0.17} \\
4-arm model& {3.40} & {44.9} & {11.70} & {4.41} & {190.1} & {10.82} & {3.58} & {217.6} & {10.68} & {3.60} &
{320.9} & {11.06} & {8.64} & {50.4} & {0.99} & {0.16} \\
\hline
\multicolumn{14}{l}{Fitting of models to all three kinds of spiral tracers (HII regions, GMCs, and 6.7 GHz methanol masers).} \\
3-arm model& {3.40} & {45.8} & {9.12} & {3.78} & {194.7} & {9.29} & {3.47} & {230.4} & {7.61}& & & & {8.59}& {54.7}& {0.62} &{0.20} \\
4-arm model& {3.23} & {41.2} & {10.64} & {4.27} & {189.0} & {11.15} & {3.56} & {217.9} & {10.63} & {3.81} &
{309.8} & {8.85} & {8.64} & {53.8} & {0.59} & {0.17} \\
\hline\hline
\multicolumn{14}{l}{Fitting of models to HII regions, $R_0=8.3$~kpc, $\Theta_0=239$~km~s$^{-1}$.} \\
3-arm model& {3.19} & {36.8} & {9.00} & {3.27} & {191.9} & {10.36} & {3.13} & {230.7} & {8.16}& & & & {8.20}& {51.1}& {2.63} & {0.16} \\
4-arm model& { 3.35} & {44.4} & {11.43} & {4.61} & {192.4} & {9.84} & {3.56} & {218.6} & {10.38} & {3.67} &
{330.3} & {10.54} & {8.21} & {55.1} & {2.77} & {0.16} \\
\hline
\multicolumn{14}{l}{Fitting of models to all three kinds of spiral tracers.} \\
3-arm model& {3.22} & {44.1} & {9.25} & {3.43} & {184.5} & {9.50} & {3.10} & {210.3} & {7.80} & & & & {8.17} & {47.8} & {2.68} &{0.18} \\
4-arm model& {3.27} & {38.5} & {9.87} & {4.29} & {189.0} & {10.51} & {3.58} & {215.2} & {10.01} & {3.98} &
{320.1} & {8.14} & {8.16} & {50.6} & {2.71} & {0.16} \\
\hline\hline
\end{tabular}
\end{center}
\label{34para}
\tablefoot{{\bf Notes.} For the $i$th spiral arm, $R_{i}$
is the initial radius, $\theta_{i}$ is the start azimuth angle,
and $\psiup_{i}$ is the pitch angle. The parameters of the Local
Arm are denoted with subscript 5. The fitting factor $Z$ (see
Eq.~\ref{eq1}) for each model is shown in the last column.}
\end{table*}
\begin{table*}
\caption{Tangential directions of spiral arms derived from
observations (part of the references are taken from the collection
by Englmaier \& Gerhard 1999 as shown in their Table 1). }
\begin{center}
\begin{tabular}{lccccccc}
\hline
\hline
Tracers and References & Near 3 kpc & Scutum & Sagittarius & Carina & Centaurus & Norma & Far 3 kpc \\
& $(^{\circ})$ & $(^{\circ})$ & $(^{\circ})$ & $(^{\circ})$ & $(^{\circ})$& $(^{\circ})$ & $(^{\circ})$ \\
\hline
HI: wea70, bs70, hen77 & & & 50 & & 310 & 328 & \\
integrated $^{12}$CO: ccdt80, gcb+87 & 24 & 30.5 & 49.5 & & 310 & 330 & \\
$^{12}$CO clouds: dect86 & 25 & 32 & 51 & & & & \\
warm CO clouds: ssr85 & 25 & 30 & 49 & & & & \\
HII regions: loc89, dwbw80 & 24 & 30 & 47 & & 305 & 332 & \\
$^{26}$Al: cgdh96 & & & 46 & & 310 & 325 & \\
Radio 408 MHz: bkb85 & & & 48 & &310,302& 328 & 339 \\
2.4$\mu$m: hmm+81 & & & & & & 332 & 339 \\
60$\mu$m: bdt90 & & & & & 313 & 329 & 340 \\
$^{12}$CO: bro92 &25 & 33 & 55 & 282 & 309 & 328 & 337 \\
$^{12}$CO: bro08 & & & & & 308 & 328 & 336 \\
$^{12}$CO: ban80, cd76 &23.5 & & & & & & \\
$^{12}$CO: dt08 &23 & & & & & & 337 \\
\hline
Median & 24 & 30.5 & 49.3 & 282 & 310 & 328 & 338 \\
\hline\hline
\end{tabular}
\end{center}
\label{tan_obs}
\tablefoot{{\bf Notes.} The median values are adopted in this work if
there is more than one measurement. bs70: \citet{bs70}; bkb85:
\citet{bkb85}; bdt90: \citet{bdt90}; bro92: \citet{bro92}; bro08:
\citet{bron08}; ban80: \citet{bani80}; ccdt80: \citet{ccdt80};
cgdh96: \citet{cgdh96}; cd76: \citet{cd76}; dect86: \citet{dect86};
ssr85: \citet{ssr85}; dt08: \citet{dt08}; dwbw80: \citet{dwbw80};
hen77: \citet{hend77}; hmm+81: \citet{hmm+81}; gcb+87:
\citet{gcb+87}; loc89: \citet{lock89}; wea70: \citet{weav70}.}
\end{table*}
\subsubsection{The logarithmic spiral arm models}
\label{log_mod}
Spiral arms of galaxies have been conventionally approximated by
logarithmic spirals~\citep[e.g.,][and references therein]{rus03}. For
the Milky Way, the two-arm logarithmic spirals can neither fit the
distribution of massive star forming regions/GMCs nor match the
observed tangential directions of spiral arms \citep{rus03,hhs09}. The
models of three- and four-arm spirals can fit data equally well
\citep{rus03,hhs09}. In this work, we focus on the spiral structure
traced by massive star forming regions and GMCs. As shown in some
galaxies, the spiral structure traced by the stellar component may be
different from that traced by massive star forming regions or gas
\citep[e.g.,][]{gp98}. For the Milky Way, the stellar component is
commonly accepted to be dominated by a two-arm spiral
pattern~\citep[][]{ds01,bcb05,cbm+09,fa12}.
In polar coordinates ($r,\theta$), the $i$th arm can be given as
logarithmic form:
\begin{equation}
\ln\frac{r}{R_{i}}=(\theta-\theta_{i})\tan\psiup_{i},
\label{eq_log}
\end{equation}
where $r$, $\theta$ are the polar coordinates centered at the GC,
$\theta$ starts at the positive $x$-axis and increases counterclockwise;
$\theta_{i}$, $R_{i}$, and $\psiup_{i}$ are the start azimuth angle,
the initial radius, and the pitch angle for the $i$th spiral arm,
respectively. To search for the optimized values of each parameter, we
minimize the factor \citep[see][]{rus03,hhs09}:
\begin{equation}\label{eq1}
Z=\frac{1}{\sum{W_i}}\sum
W_{i}\sqrt{{(x_{i}-x_{t})^{2}} {W_{x_i}^2} +
{(y_{i}-y_{t})^{2}} {W_{y_i}^2}},
\end{equation}
here, $W_{i}$ is the weight discussed in Sect. 2.4; $x_{i}$ and
$y_{i}$ are the Cartesian coordinates of a spiral tracer, their
uncertainties are represented by $W_{x_i}$ and $W_{y_i}$,
respectively; $x_{t}$ and $y_{t}$ are the coordinates of the nearest
point from the fitted spiral arms to the tracer. The Minuit package,
with the $simplex$ and $migrad$ search routines (Nelder \& Mead 1965),
is adopted to minimize the factor $Z$.
The model is first fitted to only the HII regions, and then to the all
three kinds of spiral tracers (i.e., HII regions, GMCs, and 6.7 GHz
methanol masers). We realized that the distances of many HII regions
have been determined more credibly by photometric or trigonometric
method, especially for those in the second and third Galactic
quadrants where the kinematic anomalies dominate. The known HII
regions are widely distributed in a large region of the disk (see
Fig.~\ref{dis_hii} and \ref{dis_gmc_mas}).
\begin{table*}
\caption{Tangential directions of spiral arms derived from the
best-fitted models. }
\begin{center}
\begin{tabular}{lcccccccc}
\hline
\hline
Models$\setminus$Spiral arms & Near 3 kpc & Scutum & Sagittarius & Carina & Centaurus & Norma & Far 3 kpc & $\sqrt{\Sigma_{i}(\phi_{i}^{model}-\phi_{i}^{obs})^{2}}$/N\\
& 24$^{\circ}$ & 30.5$^{\circ}$ & 49.3$^{\circ}$ & 282$^{\circ}$ & 310$^{\circ}$ & 328$^{\circ}$ & 338$^{\circ}$ \\
\hline
\hline
\multicolumn{8}{l}{Best-fitted models to HII regions, $R_0=8.5$~kpc, $\Theta_0=220$~km~s$^{-1}$.} & \\
3-arm model & {22.6} [1.4] & {34.7} [4.2]& {57.5} [8.2]& {285.0} [3.0]& {312.5} [2.5]& {327.4} [0.6]& & {1.7}\\
4-arm model & {22.1} [1.9] & {33.6} [3.1]& {55.3} [6.0]& {284.7} [2.7]& {309.8} [0.2]& {324.2} [3.8]& & {1.4} \\
PL model & {24.3} [0.3] & {33.3} [2.8]& {51.1} [1.8]& {286.1} [4.1]& {310.2} [0.2]& {326.8} [1.2]& & {0.9}\\
\hline
\multicolumn{8}{l}{Best-fitted models to all three kinds of spiral tracers.} \\
3-arm model &{22.3} [1.7] & {37.5} [7.0]& {58.7} [9.4]& {283.7} [1.7]& {310.9} [0.9]& {327.4} [0.6]& & {2.0}\\
4-arm model &{21.2} [2.8] & {34.9} [4.4]& {54.1} [4.8]& {286.1} [4.1]& {311.9} [1.9]& {326.8} [1.2]& & {1.4} \\
PL model &{24.9} [0.9] & {33.6} [3.1]& {50.8} [1.5]& {286.5} [4.5]& {310.1} [0.1]& {326.6} [1.4]& & {1.0}\\
\hline\hline
\multicolumn{8}{l}{Best-fitted models to HII regions, $R_0=8.3$~kpc, $\Theta_0=239$~km~s$^{-1}$.} \\
3-arm model &{21.9} [2.1] & {35.1} [4.6]& {54.6} [5.3]& {286.1} [4.1]& {312.8} [2.8]& {328.0} [0.0]& & {1.5}\\
4-arm model &{22.4} [1.6] & {33.5} [3.0]& {54.7} [5.4]& {285.9} [3.9]& {310.8} [0.8]& {324.1} [3.9]& & {1.4} \\
PL model &{23.8} [0.2] & {32.5} [2.0]& {49.7} [0.4]& {288.9} [6.9]& {310.2} [0.2]& {327.3} [0.7]& &{1.2} \\
\hline
\multicolumn{8}{l}{Best-fitted models to all three kinds of spiral tracers.} \\
3-arm model &{21.6} [2.4] & {36.0} [5.5]& {55.3} [6.0]& {287.0} [5.0]& {312.3} [2.3]& {328.1} [0.1]& & {1.7}\\
4-arm model &{22.4} [1.6] & {35.8} [5.3]& {53.9} [4.6]& {287.6} [5.6]& {311.9} [1.9]& {326.3} [1.7]& & {1.6} \\
PL model &{22.9} [1.1] & {32.7} [2.2]& {51.7} [2.4]& {288.7} [6.7]& {310.6} [0.6]& {327.9} [0.1]& & {1.3} \\
\hline\hline
\end{tabular}
\end{center}
\label{tan_mod}
\tablefoot{{\bf Notes.} The deviations between the modeled and
observed tangential directions ($|\phi_{i}^{model}-\phi_{i}^{obs}|$,
see Table~\ref{tan_obs} for $\phi_{i}^{obs}$) are given in
brackets.}
\end{table*}
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{polyarm_hii_8.5.ps}
\centering\includegraphics[width=0.47\textwidth]{polyarm_all_8.5.ps}\\
\caption{ Best-fitted models of polynomial-logarithmic (PL) spirals
to only the HII regions ({\it left}) and all three kinds of spiral
tracers ({\it right}). The symbols are the same as those in
Fig.~\ref{lvall}. The dashed-lines indicate the observed
tangential directions (see Table~\ref{tan_obs}). In the far end of
the GC direction, the spiral tracers generally have large position
error bars (see e.g., Fig.~\ref{dis_hii}), resulting in the large
uncertainties of the fitted positions of arm segments which are
indicated as dashed-lines for Galaxy longitude range of 340$^\circ
\leq$ $l$ $\leq$20$^\circ$.}
\label{model_pol}
\end{figure*}
The best-fitted models to HII regions are displayed in the upper
panels of Fig.~\ref{model_log}, and the corresponding parameters of
spiral arms are listed in Table~\ref{34para}. Both the three-arm model
and the four-arm model can connect most spiral tracers. The fitting
factor $Z$ = 0.17 for the best-fitted three-arm model is slightly
larger than that for the four-arm model ($Z$ = 0.16). However, the
connections of the known arm segments differ in the two models. In the
four-arm model, the Norma Arm starts from the near end of the Galactic
bar, then extends to the Outer Arm. While in the three-arm model, the
Norma Arm is connected to the Perseus Arm, which is fitted as an
individual spiral arm in the four-arm model. The Scutum Arm in the
four-arm model starts from a cluster of HII regions ($x\sim$4 kpc,
$y\sim$0 kpc) connected to the Centaurus Arm, then extends to be the
Outer+1 Arm. In the three-arm model, however, the Scutum Arm starts
from the far end of the Galactic bar, then connects to the Centaurus
Arm, and extends to the Outer Arm. The Sagittarius-Carina Arm is well
fitted in both of the two models, but their extensions are
different. The Local Arm is fitted additionally in both of the two
models as a short arm segment, which starts near the Perseus Arm, and
then extends to the fourth quadrant until it approaches the Carina Arm.
The tangential directions of spiral arms obtained by the best-fitted
four-arm model are slightly more consistent with observations than
those from the three-arm model, as indicated by
$\sqrt{\Sigma_i(\phi_{i}^{model}-\phi_{i}^{obs})^{2}}$/N in
Table~\ref{tan_mod}, though the deviations can be as large as
4$^{\circ}$$-$6$^{\circ}$ for the Norma Arm and the Sagittarius
Arm. Most grand-design spiral galaxies approximate two-fold rotational
symmetry~\citep[e.g,][]{kbc+92,rkb98,bt08}. For our Galaxy, a two-fold
symmetry of the spiral structure has been proposed \citep{dam13},
which is based on analyzing the observational data of stellar
component~\citep[e.g.,][]{bcb05} and the discovery of the Far 3 kpc
Arm \citep[][]{dt08} and distant GMCs and HII regions
\citep[][]{dt11,abbr11}. The two-fold symmetry is more satisfied by
the best-fitted four-arm model than the three-arm model.
The best-fitted models to all three kinds of tracers are displayed in
the lower two panels of Fig.~\ref{model_log}. The modeled Norma Arm,
Scutum-Centaurus Arm, Sagittarius-Carina Arm, and Perseus Arm are
consistent with the models from the fitting to HII regions. The
disagreements are found for the outer Galaxy regions, in particular in
the second and third quadrants, where many GMCs exist. Their
distributions are somewhat messy, because of large uncertainties in
the kinematic distances. Exact distances should be measured
\citep[e.g., W3OH in the Perseus Arm,][]{xrzm06}, otherwise the
structure features traced by many GMCs in the second and third
quadrants are questionable. In contrast, most of the bright HII
regions in the second and third quadrants have photometric or
trigonometric distances (see Fig.~\ref{dis_op_all}). The arm features
traced by HII regions are reliable, which motivates us to recommend
the best-fitted four-arm model shown in the upper right panel of
Fig.~\ref{model_log}.
\begin{table}
\caption{Arm parameters for the best-fitted PL spiral arm models
(see Eq.~\ref{eq_pol}).}
\begin{center}
\begin{tabular}{lcccccc}
\hline
\hline
$i$th arm & $a_i$ & $b_i$ & $c_i$ & $d_i$ & $\theta_{start}$ & $\theta_{end}$ \\
& & & & & ($^\circ$) & ($^\circ$) \\
\hline
\hline
\multicolumn{6}{l}{HII regions, $R_0=8.5$~kpc, $\Theta_0=220$~km~s$^{-1}$, $Z$ = {0.16}.} \\
arm-1 & {1.1923} & { 0.1499} & {-0.007056} & {0.0} & 40 & 250 \\
arm-2 & {7.4144} & {-2.4119} & {0.3105} & {-0.01223} & 275 & 620 \\
arm-3 &{ 6.8191} & {-2.1630} & {0.2888} & {-0.01163} & 275 & 570 \\
arm-4 & {2.6019} & {-0.3315}& { 0.03829} & {0.0} & 275 & 500 \\
arm-5 & {1.7840} & {-0.04095} & {0.01956} & {0.0} & 280 & 475 \\
arm-6 & {3.1816} & {-0.5408} & {0.07024} & {0.0} & 280 & 355 \\
& $R_i$ (kpc) & $\theta_i$ ($^{\circ}$) & $\psiup_i$ ($^{\circ}$) & & & \\
Local Arm & {8.64} & {52.0} & {1.00} & & & \\
\hline
\multicolumn{6}{l}{All three kinds of spiral tracers, $Z$ = {0.18}.} \\
arm-1 & {1.2170} & { 0.1442} & {-0.007552} & {0.0} & 40 & 250 \\
arm-2 & {7.4413} & {-2.4138} & {0.3103} & {-0.01222} & 275 & 620 \\
arm-3 &{ 6.8185} & {-2.1632} & {0.2887} & {-0.01162} & 280 & 625 \\
arm-4 & {1.8419} & {-0.09367}& {0.02001} & {0.0} & 280 & 500 \\
arm-5 & {1.7020} & {-0.01485} & {0.01522} & {0.0} & 280 & 500 \\
arm-6 & {2.0497} & {-0.05168} & {0.01807} & {0.0} & 280 & 405 \\
& $R_i$ (kpc) & $\theta_i$ ($^{\circ}$) & $\psiup_i$ ($^{\circ}$) & & & \\
Local Arm & {8.60} & {51.9} & {0.76} & & & \\
\hline
\hline
\multicolumn{6}{l}{HII regions, $R_0=8.3$~kpc, $\Theta_0=239$~km~s$^{-1}$, $Z$ = {0.16}.} \\
arm-1 & {1.1668} & { 0.1198} & { 0.002557} & {0.0} & 40 & 250 \\
arm-2 & {5.8002} & {-1.8188} & {0.2352} & {-0.008999} & 275 & 620 \\
arm-3 &{ 4.2300} & {-1.1505} & {0.1561} & {-0.005898} & 275 & 570 \\
arm-4 & {0.9744} & { 0.1405}& {0.003995} & {0.0} & 280 & 500 \\
arm-5 & {0.9887} & { 0.1714} & {0.004358} & {0.0} & 280 & 475 \\
arm-6 & {3.3846} & {-0.6554} & {0.08170} & {0.0} & 280 & 355 \\
& $R_i$ (kpc) & $\theta_i$ ($^{\circ}$) & $\psiup_i$ ($^{\circ}$) & & & \\
Local Arm & {8.10} & {52.2} & {2.36} & & & \\
\hline
\multicolumn{6}{l}{All three kinds of spiral tracers, $Z$ = {0.17}.} \\
arm-1 & {1.1320} & { 0.1233} & {0.003488} & {0.0} & 40 & 250 \\
arm-2 & {5.8243} & {-1.8196} & {0.2350} & {-0.009011} & 275 & 620 \\
arm-3 &{ 4.2767} & {-1.1507} & {0.1570} & {-0.006078} & 275 & 575 \\
arm-4 & {1.1280} & { 0.1282}& {0.002617} & {0.0} & 280 & 500 \\
arm-5 & {1.7978} & {-0.04738} & {0.01684} & {0.0} & 280 & 500 \\
arm-6 & {2.4225} & {-0.1636} & {0.02494} & {0.0} & 280 & 405 \\
& $R_i$ (kpc) & $\theta_i$ ($^{\circ}$) & $\psiup_i$ ($^{\circ}$) & & & \\
Local Arm & {8.17} & {57.8} & {2.84} & & & \\
\hline\hline
\end{tabular}
\end{center}
\label{5arm}
\tablefoot{{\bf Notes.} The fitting is made to HII regions and all
three kinds of spiral tracers, respectively.}
\end{table}
\begin{figure}[!t]
\centering\includegraphics[width=0.47\textwidth]{pitch_hii_8.5.ps}
\caption{Pitch angle ($\psiup_{i}$) variation of arm segments in
the best-fitted polynomial-logarithmic model ({\it left panel} of
Fig.~\ref{model_pol}).}
\label{pitch_a}
\end{figure}
\subsubsection{The polynomial-logarithmic spiral arm model}
A single value of the pitch angle cannot well describe the spiral arms
for many galaxies \citep[e.g.,][]{sj98,sr13}, and the variation of
pitch angle can sometimes exceed 20 percent \citep[][]{sr13}. To fit
both the data distribution and also the observed tangential
directions, we proposed a polynomial-logarithmic (PL) spiral arm model
\citep{hhs09} to connect the identified arm segments.
In polar coordinates ($r$, $\theta$), a PL spiral arm is expressed as:
\begin{equation}
\ln r = a_i + b_i \theta+c_i\theta^2+d_i\theta^3.
\label{eq_pol}
\end{equation}
To search the best-fitted models, the Minuit package is adopted to
minimize the fitting factor $Z$ (see Eq~\ref{eq1}). We first fit the
six identified arm segments (see the log($r$)$-$$\theta$ diagrams in
Fig.~\ref{log_seta}) individually by a PL spiral to obtain the
initial values of arm parameters. Then, the six arm segments and the
Local Arm are fitted together to derive the best model. As a purely
logarithmic spiral has shown to be good enough to fit the Local Arm
(Fig.~\ref{model_log}), we adopt it here as well. The model is also
fitted to the HII regions and to all three kinds of spiral tracers.
The best-fitted model to HII regions is shown in the left panel of
Fig.~\ref{model_pol}, the corresponding arm parameters are listed in
Table~\ref{5arm}. This model has advantages in connecting most known
spiral tracers. From the inside out, the Norma Arm (arm-1), the
Scutum-Centaurus Arm (arm-2), the Sagittarius-Carina Arm (arm-3), the
Perseus Arm (arm-4), and the Local Arm both are well fitted. The Outer
Arm (arm-5) and the Outer+1 Arm (arm-6) are also delineated, although
less spiral tracers are found for the outer Galaxy. The modeled
tangential directions for the PL arm models are in better agreement
with observations than those of the best-fitted four-arm model, 0.9
cf. 1.4, see Table~\ref{tan_mod}.
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{4arm_hii_8.5.eps}
\centering\includegraphics[width=0.47\textwidth]{polyarm_hii_8.5.eps}
\caption{Distribution of HII regions was overlaid on the widely used
image of the Milky Way spiral structure
(NASA/JPL-Caltech/R.Hurt). The symbols are the same as those in
Fig.~\ref{lvall}. The outlines are the best-fitted four-arm model
({\it left}, also see the {\it upper right panel} of
Fig.~\ref{model_log}) and the best-fitted PL spiral arm model ({\it
right}, also see the {\it left panel} of Fig.~\ref{model_pol}).}
\label{overlay}
\end{figure*}
The best-fitted model to all three kinds of tracers is shown in the
right panel of Fig.~\ref{model_pol}. The best-fitted arm-1 to arm-4,
and also the Local Arm are consistent in the models. Large
discrepancies in positions are found for the best-fitted Outer Arm and
the Outer+1 Arm in this model compared with the model of HII regions
partially because of the lack of HII regions in the outer Galaxy and
the unreliability of the kinematic distances for many GMCs in the
second and third quadrants (Sect.~\ref{log_mod}). In the distant
regions of the first quadrant, the CO emission features revealed by
\citet{dt11} may trace a more distant arm segment even beyond the
Outer+1 Arm (also see Fig.~\ref{model_log}). These CO emission
features may be related with the extension of the Sagittarius-Carina
Arm (arm-3), but cannot be reasonably fitted by a single spiral
arm. More measurements of distant spiral tracers are necessary.
The pitch angle for a PL arm segment is expressed in the form of $
tan(\psiup_{i})= b_i + 2c_i\theta + 3d_i\theta^2$, which varies
significantly with the azimuthal angle as shown in
Fig.~\ref{pitch_a}. In some galaxies, e.g., NGC 628 and NGC 6946,
the variations of pitch angles are also obvious
\citep[][]{ccb+12,sr13}. The relative positions of spiral tracers in
our Galaxy have larger uncertainties than those in face-on
galaxies. Therefore, high quality data (e.g., the trigonometric
parallax) is necessary to uncover the properties of spiral arms in our
Galaxy.
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{4arm_hii_8.3.ps}
\centering\includegraphics[width=0.47\textwidth]{4arm_all_8.3.ps}\\
\centering\includegraphics[width=0.47\textwidth]{polyarm_hii_8.3.ps}
\centering\includegraphics[width=0.47\textwidth]{polyarm_all_8.3.ps}\\
\caption{{\it Upper panels}: distributions of HII regions ({\it
left}) and all three kinds of spiral tracers ({\it right}) with
the flat rotation curve with newly observed $R_{0}$ = 8.3 kpc and
$\Theta_0$ = 239 km~s$^{-1}$ \citep{brm+11}. The symbols are the
same as those in Fig.~\ref{lvall}. The best-fitted four-arm
models are also shown. {\it Lower panels}: the best-fitted PL
models for these two distributions are presented.}
\label{Ro}
\end{figure*}
\subsection{Comparison with the widely used image of the Galaxy spiral
structure}
With the updated catalogs of Galactic HII regions, GMCs, and 6.7 GHz
methanol masers, we explore the spiral pattern of the Galaxy. Some
revealed arm features are not consistent with the well-known picture
of the Galaxy spiral structure (Fig.~\ref{overlay}):
(1) The extension of the Local Arm. In the concept map of the Galaxy
structure (the background of Fig.~\ref{overlay}), the Local Arm
starts near the Perseus Arm, then extends as an independent arm
segment between the Perseus Arm and the Sagittarius-Carina Arm ($x
\sim$ $-$3 kpc, $y \sim$9 kpc). This feature is not confirmed by the
distribution of HII regions. As shown in Fig.~\ref{overlay}, the
Local Arm does not extend far from the Sun. It starts near the Perseus
Arm, then extends to the fourth quadrant, and closely approaches to
the Carina Arm at $x \sim$ $-$1.5 kpc, $y \sim$8.3 kpc, also see
Fig.~\ref{Ro}.
(2) The start point of the Scutum Arm. The Scutum Arm is commonly
believed to originate from the near end of the Galactic bar. This
contradicts the distribution of the cataloged spiral tracers. In the
best-fitted models, the Scutum Arm is related to many bright HII
regions in the distant parts of the first Galaxy quadrant around
$x\sim$4 kpc, $y\sim$0 kpc, which is consistent with the spiral arm
models of GG76 and \citet{ne2001}.
(3) The demarcation of the Sagittarius-Carina Arm and the
Scutum-Centaurus Arm. Around the direction of the GC ($|x|\lesssim$ 2
kpc, 4 kpc $\lesssim y \lesssim$ 7 kpc), no clear demarcation is found
in the data distribution between the best-fitted Sagittarius-Carina
Arm and the Scutum-Centaurus Arm, which is also shown in the
distribution of HII regions with photometric or trigonometric
distances (Fig.~\ref{dis_op_all}).
(4) The extensions of the Sagittarius-Carina Arm and the
Scutum-Centaurus Arm. As seen in Fig.~\ref{overlay}, large
discrepancies are found for the extension of the Sagittarius-Carina
Arm or the Scutum-Centaurus Arm (in the regions with $y <$ $-$6 kpc)
between our best-fitted models and that of the concept map.
\begin{figure*}
\centering\includegraphics[width=0.47\textwidth]{dis_bb93_8.5.ps}
\centering\includegraphics[width=0.47\textwidth]{bb93_contour_8.5.ps}\\
\caption{{\it Left}: distributions of HII regions, GMCs, and 6.7 GHz
methanol masers projected into the Galactic plane. The symbols are
the same as those in Fig.~\ref{lvall}. The kinematic distances
are estimated using the rotation curve of BB93. {\it Right}: color
intensity map of spiral tracers. The IAU standard $R_{0}$ = 8.5
kpc and $\Theta_0$ = 220 km~s$^{-1}$ and standard solar motions
are adopted in deriving the kinematic distances if no photometric
or trigonometric distance is available.}
\label{bb93}
\end{figure*}
\subsection{Influence of the fundamental parameters to the derived spiral
structure}
\label{sec_r0}
As discussed in Sect. 2.2, recent observations suggest that $R_{0}$,
$\Theta_0$, and solar motions are different from the IAU standards,
which should influence the derived spiral pattern of our Galaxy from
the cataloged tracers. To show the possible influence, we
re-calculated the distances of spiral tracers with a flat rotation
curve for $R_{0}$ = 8.3 kpc and $\Theta_0$ = 239 km~s$^{-1}$
\citep{brm+11} together with the new solar motions of \citet{sbd10}.
The results are given in Fig.~\ref{Ro}. The arm-like features are
very similar to that with the IAU standard, but slightly shrinked,
especially for the outer Galaxy regions. We also fit the data
distribution with the four-arm logarithmic spiral arm models and the
PL arm models. The best fits are shown in Fig.~\ref{Ro}, the
corresponding arm parameters are listed in Table~\ref{34para} and
Table~\ref{5arm}. The modeled tangential directions are compared with
observations in Table~\ref{tan_mod}.
\subsection{Influence of the rotation curve on the derived spiral
structure}
\label{sec_bb93}
Besides the flat rotation curve, the rotation curve of BB93 has also
been widely used in previous studies. As shown by \citet{hhs09}, the
BB93 rotation curve is reasonable for the whole Galaxy in comparison
with that of \citet{clem85} and \citet{fbs89}. We re-calculated the
kinematic distances of the collected spiral tracers with the BB93
rotation curve and the results are shown in Fig.~\ref{bb93}. The
arm-locations are very similar to those calculated with the flat
rotation curve. This is reasonable because the BB93 rotation curve is
almost flat in the galactocentric radii between $\sim$5 kpc and about
15 kpc.
\subsection{The Galactic warp}
The Galactic warp was first discovered in early surveys of HI gas
\citep{kerr57,burt88}, which is vertically distorted even to distances
larger than 3 kpc from the Galactic plane~\citep{ds91}. Evidence for
Galactic warp has recently been observed again from HI absorption
\citep{dsg+09}, red clump giants \citep{lch+10,boby10}, extended low
density warm ionized medium \citep{cmf+09}, 2MASS infrared stars
\citep{rmr+09}, young open clusters \citep{vmc+08}, molecular gas
\citep{nomf05}, Galactic HII regions \citep{pdd04}, and also star
forming complexes \citep{rus03}.
With the catalogs of Galactic HII regions, GMCs, and 6.7 GHz methanol
masers, we get significant evidence of warp in the outer Galaxy as
shown in Fig.~\ref{warp}, where the best-fitted four-arm model is
overlaid. In the first and second quadrants, the Outer Arm and the
Outer+1 Arm are above the Galactic plane (positive). In the third and
fourth quadrants, the Outer Arm, the Perseus Arm, and the Carina Arm
are below the Galactic plane (negative).
\begin{figure}[!b]
\centering\includegraphics[width=0.47\textwidth]{Gal_warp.ps}
\caption{Evidence of Galactic warp as shown by the distributions of
HII regions, GMCs, and 6.7 GHz methanol masers. Note that the
diamonds here indicate the tracers of $b < 0.0^\circ$, and the
blue crosses indicate the tracers of $b > 0.0^\circ$. The symbol
size is proportional to the offset from the Galactic plane. The
outlines are the best-fitted four-arm model (see the {\it upper
right panel} of Fig.~\ref{model_log}). }
\label{warp}
\end{figure}
\section{Discussions and conclusions}
We update the catalogs of Galactic HII regions, GMCs, and 6.7 GHz
methanol masers to study the large-scale spiral arm structure.
Compared with our previous results in \citet{hhs09}, the samples of
HII regions and GMCs are much enlarged. We include the maser data
though they are marginally helpful. If the photometric or
trigonometric distances of these tracers are not available in the
literature, kinematic distances are estimated using a Galaxy rotation
curve with two sets of $R_0$, $\Theta_0$, and solar motions, where one
set is the IAU standard and the other is from the new observational
results. We calculate the excitation parameters of HII regions and the
masses of GMCs and scale them as the weight factors, then obtain the
weighted distribution of these spiral tracers and also their
combinations.
We have shown that the distribution of HII regions delineates a clear
spiral structure. Obvious arm segments can be identified. From the
inside out, they are the Norma Arm, the Scutum-Centaurus Arm, the
Sagittarius-Carina Arm, the Perseus Arm, the Outer Arm, and probably
the Outer+1 Arm. The Local Arm is identified as a short arm
segment. The identified arm segments traced by massive star forming
regions and GMCs can match the HI arms in the outer Galaxy, suggesting
that the Milky Way is a grand-design spiral galaxy.
We fit the models of three-arm and four-arm logarithmic spirals to
only the HII regions and to all three kinds of spiral tracers,
respectively. Both models can connect most spiral tracers, but the
connections of known arm segments differ in the two models. A PL
spiral arm model is able to not only connect most spiral tracers, but
also match the observed tangential directions.
\begin{acknowledgements}
We thank the referee for helpful comments, and Dr. L{\'e}pine and
Dr. Bronfman for kindly providing us their data. The authors are
supported by the Strategic Priority Research Program ``The Emergence
of Cosmological Structures" of the Chinese Academy of Sciences,
Grant No. XDB09010200, and the National Natural Science Foundation (NNSF) of
China No. 11473034. L.G.H. is also supported by the Young Researcher
Grant of National Astronomical Observatories, Chinese Academy of
Sciences.
\end{acknowledgements}
\scriptsize
\bibliographystyle{aa}
|
\section{Introduction}
In the context of quantum information and quantum technology the idea of
reading classical data by means of quantum states arises quite naturally~\cite{Pirandola2011a,Pirandola2011b}. In general, the standard
implementations of reading are based on optical technologies: the task is
the readout of a digital optical memory, where information is stored by
means of the optical properties of the memory cells that are in turn probed
by shining light, e.g. a laser beam, on them. The probing light is usually
denoted as the \emph{transmitter}. Interesting features arise in the regime
in which the transmitter has to be treated quantum mechanically. The maximum
rate of reliable readout defines the quantum reading capacity~\cite{Pirandola2011b}. The latter
can overcome the classical reading capacity, obtained by probing with
classical light, in several relevant settings. The (possibly quantum)
transmitter that is needed to extract the encoded information is prepared in
some initial state. By scanning a particular cell the transmitter changes
its properties in a way depending on the cell. The task is to recognize
which cell occurs based on the output state of the transmitter after it has
been detected and measured. Therefore the problem of reading reduces to the
problem of distinguishing the output states of the transmitter.
In such optical settings one needs to consider two main coding protocols
depending on the trade off between energy and coherence of the transmitters
and the channels that are being used. The first protocol is the so-called
\emph{amplitude shift keying} (ASK) in which the changes in the state of the
transmitter are caused by the cell-dependent losses in the intensity of the
transmitted signal~\cite{Pirandola2011a,Pirandola2011b,Lupo2013,Spedalieri2012,Tej2013}. The second main
protocol is the so-called \emph{phase shift keying} (PSK)~\cite{Hirota2011,Dallarno2012,Nair2011, Guha2013, Bisio2011,Dallarno2012a}.
This is a type of coding which does not produce
energy dissipation. On the other hand, it requires a very high coherence of
the transmitter, a feature that might be realized in realistic implementations~\cite{Dallarno2012}.
If the transmitter is quantum, the cells play the role of effective quantum
channels. The ASK protocol then corresponds to a dissipative channel coding,
while the PSK protocol is a particular case of unitary coding corresponding to a unitary channel.
Within the ASK protocol, it can be shown that in the low-energy regime there is an energy
threshold above which the maximally entangled transmitter, i.e. a two-mode
squeezed state, yields a better reading efficiency than any of the classical states with the same
energy~\cite{Pirandola2011a,Pirandola2011b,Spedalieri2012}. The general result is still
valid in the presence of some noise-induced decoherence. Within the ASK
protocol, coding is then realized by local channels corresponding to cells
with different reflectivities.
In the PSK protocol, the coding is realized by means of local unitary
operations, specifically local phase shifts~\cite{Hirota2011,Dallarno2012}. In the ideal,
noise-free protocol the transmitter is taken to be in a pure Gaussian
quasi-Bell state, i.e. a Bell-type superposition of quasi-orthogonal
coherent states. In this scheme, the resulting quantum advantage is
absolute, in the sense that quantum reading of the classical information
encoded via a phase shift of $\pi$ is achieved with vanishing error, while
any classical state of the transmitter always yields a finite error
probability.
In both the ASK and PSK protocols the transmitter is assumed to be a
bipartite system such that only one part of it scans the memory cell. This
choice is motivated by the fact that it maximizes distinguishability at the
output when the state of the transmitter is quantum. As already mentioned,
the reading efficiency is characterized by the probability of error.
Information is encoded in binary memory cells with indices $0$ and $1$. It
is thus written using only two local channels that are assumed to occur with
equal \emph{a priori} probabilities. Given the bi-partite input transmitter
$\rho _{AB}$, the two possible output states will be denoted by
$\rho_{AB}^{(0)}$ and $\rho _{AB}^{(1)}$.
The probability of error in distinguishing the two output states when
reading a memory cell by means of the same input $\rho_{AB}$ is given by the
well-known Helstrom formula \cite{Helstrom1976}:
\begin{equation}
P_{err} = \frac{1}{2} - \frac{1}{4}d_{Tr}\left(\rho_{AB}^{(0)},\rho_{AB}^{(1)}\right) \, ,
\label{Helstrom}
\end{equation}
where $d_{Tr} \equiv \|\rho_{AB}^{(0)} - \rho_{AB}^{(1)}\|_{Tr}$ is the trace distance,
with $\|X\|_{Tr}=\tr\sqrt{XX^{\dagger}}$. With our normalization convention, the trace distance ranges from
$0$ to a maximum of $2$ for orthogonal pure states.
In the original reading protocols the goal is to minimize $P_{err}$ over the
set of possible transmitter states $\rho_{AB}$ at fixed encoding in the
memory cells~\cite{Pirandola2011b,Dallarno2012}. The problem is thus
dependent on the type of memory device being used.
Here instead we wish to
provide a device-independent characterization of a given transmitter by
considering the worst-case scenario that maximizes the probability of error
$P_{err}$ over all possible codings. Once the worst-case scenario is
identified, one can then compare different classes of transmitter states to
identify the ones that minimize the maximum probability of error $P_{err}^{(\max)}$.
We will show
that the maximum probability of error $P_{err}^{(\max)}$ is a monotonically decreasing function of
the amount of quantum correlations present in the transmitter state
$\rho_{AB}$, as quantified by a recently introduced measure of quantum correlations,
the so-called discord of response~\cite{Roga} in its Gaussian version~\cite{Buono2014}
(for general reviews on quantum correlations and distinguishability of quantum states and
on discord-like correlations see~\cite{Spehner2014,Modi2012}). As a
consequence, every state with non-vanishing discord of response is able to
read any type of memory device with maximal $P_{err}<1/2$. On the other
hand, for each classical transmitter, i.e. for transmitter with
vanishing discord of response, there will always exist at least one memory
device which is completely invisible, i.e. for which $P_{err}=1/2$.
In these considerations we exclude the situations when two channels are chosen to be
arbitrarily similar. In this case the probability of error always approaches $1/2$
independently on the chosen transmitter.
In Sec.~\ref{errorprobdiscord} we derive the exact analytical relation
between the maximum probability of error and the Gaussian discord of response. In Sec.~\ref{Gausstrans}
we discuss the properties of classical and quantum Gaussian transmitters, comparing squeezed-thermal, thermal-squeezed
and coherent-thermal states; in subsection \ref{GausstransA} we derive upper and lower bounds on
the maximum probability of error, and in subsection \ref{phaseshiftreading} we identify the unitary coding
that maximizes the quantum Chernoff bound, namely the upper bound on the maximum
probability of error, for the classes of quantum Gaussian transmitters considered.
The unitary channel which maximizes the bounds on the probability of error turns out to be a particular PSK coding,
namely the unique traceless one, which is realized for a $\pi/2$ phase shift.
In Sec.~\ref{comparisonclass} we compare the
performance of Gaussian quantum states of light with classical states (coherent thermal states).
We show that strongly discordant squeezed thermal states possess a higher
reading efficiency than the corresponding classical states of light, that is
non-discordant Gaussian coherent thermal transmitters with the same total number
of photons (fixed energy). This realizes an important instance of
noise-enhanced quantum advantage over the corresponding noisy classical resources.
In Sec.~\ref{comparisonquant} we compare different families of discordant
Gaussian states, the squeezed thermal and the thermal squeezed states at fixed total number of photons
or squeezing. While for both classes of states the entanglement obviously
decreases with increasing thermal noise, we show that for squeezed thermal
states the discord is an increasing function of the number of thermal
photons at fixed squeezing, while the opposite holds for thermal squeezed states. Moreover,
for squeezed thermal transmitters, the quantum Chernoff bound is independent
of thermal noise. As a consequence, this type of transmitter plays a privileged
role in the considered class of quantum Gaussian resources because the associated
quantum efficiency is either enhanced or unaffected by increasing the thermal
noise. Thus squeezed thermal transmitters realize an instance of noise-enhanced
or noise-independent quantum resources at fixed squeezing. Both in the classical-to-quantum
comparison and in the comparison of different quantum resources, the key enhancement of
quantum advantage is realized in the situation of strongest asymmetry of the distribution
of thermal noise among the field modes: {\em local} noise enhancement leads
to {\em global} enhancement of quantum correlations.
The main results are summarized and some outlook perspectives for future work are discussed
in Sec.~\ref{summary}. Detailed calculations and auxiliary reasonings are reported in four Appendixes.
\section{Probability of error, bounds, and discord of response}
\label{errorprobdiscord}
In protocols of quantum reading with unitary coding the
two local channels acting on the input probe state $\rho_{AB}$ are unitary,
and denoted as $U^{(0)}_A$ and $U^{(1)}_A$. Therefore, in this type of protocol
$\rho_{AB}^{(0)} = U_A^{(0)}\rho_{AB}U_A^{(0)\dagger}$ and
$\rho_{AB}^{(1)} = U_A^{(1)}\rho_{AB}U_A^{(1)\dagger}$, so that the
probability of error reads
\begin{equation}
P_{err} = \frac{1}{2} - \frac{1}{4}d_{Tr}\left(U_A^{(0)}\rho_{AB}U_A^{(0)\dagger},
U_A^{(1)}\rho_{AB}U_A^{(1)\dagger}\right) \, .
\end{equation}
Since the trace norm is invariant under local unitary transformations, one
has equivalently
\begin{equation}
P_{err} = \frac{1}{2} - \frac{1}{4}d_{Tr}\left(\rho_{AB}, \widetilde{\rho }_{AB} \right) \, ,
\label{proberr}
\end{equation}
where $\widetilde{\rho }_{AB} = W_{A} \rho _{AB} W_{A}^{\dagger }$ and $W_A = U_A^{(0)\dagger}U_A^{(1)}$ is still a local unitary transformation
acting on the transmitted subsystem $A$. The absolute upper bound for the
probability of error is thus $1/2$, corresponding to a situation in which there
is no way to distinguish the two output states and therefore the memory
device becomes completely invisible to the transmitter.
In general, computing the trace distance proves to be extremely challenging~\cite{Audenaert2007}, even more so for Gaussian states of infinite-dimensional continuous-variable systems~\cite{Pirandola2008}. Therefore one has to look for analytically computable {\em a priori} upper and lower bounds. A natural upper bound on the probability of error, Eq.~(\ref{Helstrom}), in distinguishing two states $\rho_1$ and $\rho_2$ occurring with the same probability is provided by the quantum Chernoff bound $QCB$~\cite{Calsamiglia2008}:
\begin{equation}
P_{err} \leq QCB \equiv \frac{1}{2} \left[ \inf_{t\in (0,1)} \tr (\rho_1^{t}\rho_2^{1-t}) \right] \; .
\label{QCB}
\end{equation}
If the states $\rho_1$ and $\rho_2$ are not arbitrary, but they are qubit-qudit states related by a local, single-qubit unitary transformation, as in the case of the quantum reading protocol with unitary coding, for which $\rho_1 = \rho_{AB}$ and $\rho_2 = \widetilde{\rho }_{AB}$, then the quantum Chernoff bound $QCB$ is achieved for $t=1/2$ in Eq.~(\ref{QCB}), as discussed in Appendix~\ref{chbm}:
\begin{equation}
QCB = \frac{1}{2}\left[ \tr \left( \sqrt{\rho _{AB}} \sqrt{\widetilde{\rho }_{AB}} \right) \right] \; ,
\label{QCBLU}
\end{equation}
and its expression coincides with the quantum Bhattacharyya coefficient \cite{Pirandola2008,Sihui2008}, which provides an upper bound on $QCB$ for
arbitrary quantum states. The same result, Eq.~(\ref{QCBLU}), holds for Gaussian states related by traceless local symplectic transformations, as discussed in the following sections and in Appendix~\ref{chbm}.
Next, considering the Uhlmann fidelity yields a complete hierarchy of lower and upper bounds~\cite{Pirandola2008}:
\begin{equation}
LBP_{err} \leq P_{err} \leq QCB \; ,
\label{errorineq}
\end{equation}
where the lower bound on the probability of error $LBP_{err} \equiv (1-\sqrt{1-{\cal{F}}})/2$, and
the Uhlmann fidelity ${\cal{F}}$ between two quantum states $\rho_1,\rho_2$ is defined as
$\c F(\rho_1,\rho_2) \equiv \left( \tr{\sqrt{\sqrt{\rho_1}\rho_2\sqrt{\rho_1}}} \right)^2$.
For the quantum reading protocol with unitary coding, let us consider the maximum probability of error
in distinguishing the output of a binary memory cell encoded using one identity
and one arbitrary unitary channel $W_{A}$ chosen in the set of local unitary
operations with non-degenerate harmonic spectrum. The latter is the spectrum
of the complex roots of the unity and its choice is motivated by observing
that it excludes unambiguously the identity from the set of possible
operations: indeed, unitary operations with harmonic spectrum are orthogonal
(in the Hilbert-Schmidt sense) to the identity. We further assume that the
coding is unbiased, that is the two channels are equiprobable.
The worst-case scenario is defined by the probability of error Eq.~(\ref{proberr}) being the largest possible:
\begin{equation}
P_{err}^{(\max )}\equiv \max_{\{W_{A}\}}P_{err}=\frac{1}{2}-\frac{1}{4} \min_{\{W_{A}\}}d_{Tr}\,\left( \rho _{AB},\widetilde{\rho }_{AB}\right) .
\label{PerrMax}
\end{equation}
Let us now consider a recently introduced measure of quantum correlations, the so-called
discord of response~\cite{Roga}:
\begin{equation}
{\mathcal{D}}_{R}^{x}(\rho _{AB})\equiv \min_{\{W_{A}\}}{\cal{N}}_x^{-1}d_{x}^{2}\left(\rho_{AB},\widetilde{\rho }_{AB}\right) \, ,
\label{DiscOfResp}
\end{equation}
where the index $x$ denotes the possible different types of well behaved, contractive metrics under completely positive and trace-preserving
(CPTP) maps. The normalization factor ${\cal{N}}_x$ depends on the given metrics and is chosen in such a way to assure that ${\mathcal{D}}_{R}^{x}$
varies in the interval $[0,1]$. Finally, the set of local unitary operations $\{W_A\}$ includes all and only those local unitaries with
harmonic spectrum.
In the following, we will need to consider both the probability of error and different types of upper and lower bounds on it. Therefore we will be concerned with three different discords of response corresponding to three types of contractive distances: trace, Hellinger, and Bures.
The trace distance $d_{Tr}$
between any two quantum states $\rho_1$ and $\rho_2$ is defined as:
\begin{equation}
d_{Tr}\left( \rho_1, \rho_2 \right) \equiv \tr \left[ \sqrt{ \left( \rho_1 - \rho_2 \right)^{2}} \right] \, .
\label{tracedistance}
\end{equation}
The Bures distance, directly related to the fidelity ${\cal{F}}$, is defined as:
\begin{equation}
d_{Bu}\left( \rho_1, \rho_2 \right) \equiv \sqrt{ 2 \left( 1 - \sqrt{ {\cal{F}}(\rho_1,\rho_2) } \right) } \; \, .
\label{Buresdistance}
\end{equation}
Finally, the Hellinger distance is defined as:
\begin{equation}
d_{Hell}\left( \rho_1, \rho_2 \right) \equiv \sqrt{ \tr \left[ \big( \sqrt{\rho_1} - \sqrt{\rho_2} \big)^{2} \right] } \; \, .
\label{helldistance}
\end{equation}
For each discord of response, trace, Hellinger, and Bures, the normalization factor in Eq.~(\ref{DiscOfResp}) is, respectively:
${\cal{N}}_{Tr}^{-1} = 1/4$, ${\cal{N}}_{Hell}^{-1} = {\cal{N}}_{Bu}^{-1} = 1/2$.
If the two states $\rho_1$ and $\rho_2$ are bipartite Gaussian states related by local traceless symplectic transformations or bipartite qubit-qudit states related by a local single-qubit unitary operation, that is $\rho_1 = \rho_{AB}$ and
$\rho_2 = \widetilde{\rho }_{AB} = W_{A} \rho _{AB} W_{A}^{\dagger }$, then, by exploiting Eq.~(\ref{QCBLU}),
it is straightforward to show that the quantum Chernoff bound is a simple, monotonically non-increasing simple
function of the Hellinger distance:
\begin{equation}
QCB = \frac{1}{4}\left( 2 - d_{Hell}^{2}(\rho_{AB},\widetilde{\rho }_{AB}) \right) \; .
\end{equation}
It is then immediate to show that the maximum of $QCB$ over the set of local unitary operations $\{W_{A}\}$ with completely
non-degenerate harmonic spectrum is a simple linear function of the Hellinger discord of response:
\begin{equation}
QCB^{max} = \frac{1}{2}\big( 1 - {\cal D}_{R}^{Hell}(\rho _{AB}) \big) \; .
\end{equation}
The discord of response quantifies the response of a quantum state to least-disturbing local unitary perturbations
and satisfies all the basic axioms that must be obeyed by a \emph{bona fide} measure of quantum correlations~\cite{Roga}: it vanishes if
and only if $\rho _{AB}$ is a classical-quantum state; it is invariant under local unitary operations; by fixing a well-behaved metrics
such as trace, Bures, or Hellinger, it is contractive under CPTP maps on subsystem $B$, i.e. the subsystem that is not perturbed by the
local unitary operation $W_A$; and reduces to an entanglement monotone for pure states, for one of which it also assumes the maximum possible value ($1$).
By comparing Eqs.~(\ref{PerrMax}) and (\ref{DiscOfResp}) with $x=Tr$, it is immediate to relate the maximum probability of error $P_{err}^{(\max )}$
to the trace discord of response ${\mathcal{D}}_{R}^{Tr}$:
\begin{equation}
P_{err}^{(\max )}=\frac{1}{2}-\frac{1}{2}\sqrt{{\mathcal{D}}_{R}^{Tr}(\rho_{AB})} \; \, .
\label{perrdiscord}
\end{equation}
From Eq.~(\ref{perrdiscord}) it follows that half of the square root of the
trace discord of response yields the difference between the absolute maximum of
the probability of error (i.e. $1/2$) and the maximum probability of error at
fixed transmitter state $\rho_{AB}$.
A vanishing trace discord of response implies that there exists at least one memory that
cannot be read by classical-quantum transmitters. A maximum trace discord of response
(${\mathcal{D}}_{R}^{Tr}=1$) implies that, irrespective of the coding, the maximally
entangled transmitter will read any memory without errors: indeed, any local
unitary operation with harmonic spectrum transforms a maximally entangled
state into another maximally entangled state orthogonal to it, and therefore yields perfect
distinguishability at the output.
\section{Quantum reading with squeezed thermal states} \label{Gausstrans}
In the following, in order to compare the efficiency of classical (non-discordant) and quantum noisy sources of light in reading protocols, we will consider
two-mode Gaussian states of the electromagnetic field. The states with vanishing first moments of the quadratures are fully described by their covariance matrix $\sigma$~\cite{Weedbrook2012,Salerno,Ferraro}:
\begin{equation}
\sigma = \frac{1}{2}\begin{bmatrix}
a & 0 & c_1 & 0 \\
0 & a & 0 & c_2 \\
c_1 & 0 & b & 0 \\
0 & c_2 & 0 & b
\end{bmatrix} \; ,
\label{sqthcorm}
\end{equation}
The range of values of $a, b, c_1$ and $c_2$ for which the corresponding states are physical (i.e. correspond to positive density matrices)
is determined by the Heisenberg uncertainty relation stated in symplectic form:
\begin{equation}
\sigma+\frac{i}{2}\omega\oplus\omega\geq 0 \; ,
\label{uncert}
\end{equation}
where $\omega=\begin{bmatrix}
0 & 1 \\
-1 & 0
\end{bmatrix}$ is the symplectic form. In all the paper, if we refer to the symmetric states we mean $a=b$.
In the following we will focus on two rather general classes of (undisplaced) Gaussian states, the squeezed thermal states (STS)
and the thermal squeezed states (TSS). The former are defined by two-mode squeezing
$S(r)=\exp {\{ra_{1}^{\dagger }a_{2}^{\dagger }-r^{\ast }a_{1}a_{2}\}}$ applied on, possibly non-symmetric, two-mode thermal
states. Notice that the denomination STS is sometimes used in the literature to denote any Gaussian states characterized by the
covariance matrix Eq.~(\ref{sqthcorm}) with $c_1=-c_2$.
In the present work we adopt the convention that STSs describe a physically rather frequent
situation in which the thermal noise acts possibly non-symmetrically on the two modes, that is $N_{th_1}\neq N_{th_2}$, and thus the total number of thermal photons is $N_{th_1}+N_{th_2}$. Here $r$ is the two-mode squeezing parameter and $a_{i}$ are the
annihilation operators in each of the two modes ($i=1,2$). The diagonal and off-diagonal covariance
matrix elements for these states, respectively $a=a_{sq-th}$, $b=b_{sq-th}$ and $c_1=-c_2=c_{sq-th}$, read:
where
\begin{eqnarray}
a_{sq-th}\! &\!=\!&\!\cosh (2r)\!+\!2N_{th_{1}}\cosh ^{2}(r)\!+\!2N_{th_{2}}\sinh ^{2}(r)\nonumber \, , \\
b_{sq-th}\! &\!=\!&\!\cosh (2r)\!+\!2N_{th_{2}}\cosh ^{2}(r)\!+\!2N_{th_{1}}\sinh ^{2}(r) \, ,\nonumber \\
c_{sq-th}\! &\!=\!&\!(1+N_{th_{1}}+N_{th_{2}})\sinh (2r)\label{ccvsqth} \, .
\end{eqnarray}
where $N_{s}=\sinh^{2}{(r)}$ is the number of squeezed photons.
Thermal squeezed states (TSSs) describe the reverse physical situation: an initially two-mode squeezed vacuum is allowed to evolve
at later times in a noisy channel and eventually thermalizes with an external environment characterized by a total number of thermal
photons $N_{th_1}+N_{th_2}$. The covariance matrix elements of TSSs, respectively $a=a_{th-sq}$, $b=b_{th-sq}$ and $c_1=-c_2=c_{th-sq}$, are:
\begin{eqnarray}
a_{th-sq} &=&2N_{s}+1+2N_{th_1}\, , \nonumber \\
b_{th-sq} &=&2N_{s}+1+2N_{th_2}\, , \nonumber \\
c_{th-sq} &=&2\sqrt{N_{s}(N_{s}+1)} \; . \label{cthsq}
\end{eqnarray}
The same covariance matrix, Eq.~(\ref{sqthcorm}), also describes classical uncorrelated
tensor product states, which we assume to be Gaussian. Thermal states are obtained letting
$c=c_{cl}=0$, $a=a_{cl}=1+2N_{th_1}$ and $b=b_{cl}=1+2N_{th_2}$. These Gaussian states are classical in the
sense that they can be written as convex combinations of coherent states
and, moreover, they are the only Gaussian states with vanishing discord~\cite{Adesso2010,Adesso2011}.
Notice that in the standard quantum optics terminology the wording classical states is used to denote any state with positive Glauber-Sudarshan $P$-representation. In the following, without loss of generality, we will identify party $A$ with mode $a_{1}$ and party $B$ with mode $a_{2}$.
\subsection{Probability of error: upper and lower bounds, and Gaussian discords of response}
\label{GausstransA}
For unitary-coding protocols with Gaussian transmitters, Gaussian local (single-mode) unitary operations acting on an infinite-dimensional Hilbert space are implemented by local (single-mode) symplectic transformations acting on the covariance matrix $\sigma$ of two-mode Gaussian input states $\rho_{AB}^{(\sigma)}$. In the following we will consider only traceless transformations. The traceless condition must be imposed in order to exclude trivial coding by two identical channels, for which the maximum probability of error is always $1/2$. Moreover, imposing tracelessness allows to investigate and determine the correspondence between reading efficiency and quantum correlations, as will be shown in the following.
Denoting by $F_A$ the local traceless symplectic transformations acting on mode $A$, the two local unitary operations implementing the encodings of the binary memory cells are the identity $\1_{A}\oplus\1_{B}$ and $F_A\oplus\1_B$.
In order to assess the performance of quantum and classical Gaussian resources in the unitary-coding quantum reading protocol we need to evaluate the upper and lower bounds, Eqs.~(\ref{QCB}) and (\ref{errorineq}), on the maximum probability of error $P_{err}^{(\max )}$, Eqs.~(\ref{PerrMax}) and (\ref{perrdiscord}), for Gaussian two-mode transmitters $\rho_{AB}^{(\sigma)}$. To this end, we introduce first the {\em Gaussian} discord of response~\cite{Buono2014}, i.e the discord of response obtained by minimizing over local unitaries restricted only to the subset of local symplectic, traceless, transformations $F_A$:
\begin{equation}
{\cal{GD}}_{R}^{x}(\rho_{AB}^{(\sigma)}) \equiv \min_{\{F_A\}} {\cal{N}}_x^{-1} d_{x}^2\left(\rho_{AB}^{(\sigma)},\widetilde{\rho }_{AB}^{(\sigma)}
\right) \, ,
\label{intdr}
\end{equation}
where the index $d_x$ stands for trace, Hellinger, or Bures distance with the same normalization factors ${\cal{N}}_x^{-1}$ as before, and $F_A^T$ is the transpose of the symplectic matrix $F_A$ and $\widetilde{\rho }_{AB}^{(\sigma)} \equiv \rho_{AB}^{(F_A \sigma F_A^T)}$.
The Gaussian discord of response provides an upper bound to the true discord of response of Gaussian states and vanishes on and only on Gaussian classical states (subset of separable states that are in product form). The main properties of the Gaussian discord of response are reported in Appendix~\ref{appdiscord}.
In complete analogy with Eq.~(\ref{perrdiscord}) the maximum probability of error in discriminating two Gaussian transmitters related by a local symplectic transformation can be expressed as a simple function of the trace Gaussian discord of response:
\begin{equation}
P_{err}^{(\max )} = \frac{1}{2} - \frac{1}{2}\sqrt{{\mathcal{GD}}_{R}^{Tr}\left(\rho_{AB}^{(\sigma)}\right)} \; \, .
\label{Gaussperrdiscord}
\end{equation}
Specializing the bounds given by Eq.~(\ref{errorineq}) to the maximum probability of error in distinguishing Gaussian states, one has:
\begin{equation}
LBP_{err}^{(\max )} \leq P_{err}^{(\max )} \leq QCB^{(\max )} \, ,
\label{Gaussbounds}
\end{equation}
where the lower bound $LBP$ is a simple monotonically non-increasing function of the Bures Gaussian discord of response:
\begin{eqnarray}
LBP_{err}^{(\max )} = \frac{1}{2}\left( 1-\sqrt{1 - \big( 1 - {\mathcal{GD}}_{R}^{Bu} \big)^{2}} \right)
\, , \label{LBPmax}
\end{eqnarray}
and the upper bound $QCB$ is a simple linear, monotonically non-increasing function of the Hellinger Gaussian discord of response:
\begin{eqnarray}
QCB^{(\max )} = \frac{1}{2}\left( 1 - {\mathcal{GD}}_{R}^{Hell} \right) \, .
\label{QCBmax}
\end{eqnarray}
Therefore, for increasing Gaussian discords of response the bounds on the probability of error decrease correspondingly. The explicit expressions of the quantum Chernoff bound $QCB$, the Hellinger Gaussian discord of response, the Uhlmann fidelity, and the Bures Gaussian
discord of response are derived in Appendices~\ref{UhlmannGaussian} and \ref{ChernoffGaussian}.
\subsection{Maximum probability of error: $\pi/2$ phase shift}
\label{phaseshiftreading}
The probability of error in distinguishing $\rho_{AB}^{(\sigma)}$ from $\widetilde{\rho }_{AB}^{(\sigma)} \equiv \rho_{AB}^{(F_A \sigma F_A^T)}$ is given by Eq.~(\ref{proberr}) with the local symplectic transformations $F_A$ replacing $W_{A}$. Among the local unitary operations $F_A$ which can implement the unitary-coding reading protocol, an important subset includes the single-mode phase shifts $P_{\phi }$ acting on mode $a_{1}$, parameterized by the
angle parameter $\phi$: $P_{\phi }=\exp {(-i\phi a_{1}^{\dagger }a_{1})}$.
Under a local phase shift the local mode $a_{1}$ is transformed as follows: $\widetilde{a}_1 = P_{\phi }a_{1}P_{\phi }^{\dagger } = \exp {(-i\phi )}a_{1}$, while the two-mode covariance matrix $\sigma $ transforms according to
$(F_{\phi }\oplus \mathbbm{1})\sigma (F_{\phi } \oplus \mathbbm{1})^{T}$, where the symplectic matrix $F_{\phi }$ reads
\begin{equation}
F_{\phi }=%
\begin{bmatrix}
\cos {\phi } & \sin {\phi } \\
-\sin {\phi } & \cos {\phi }
\end{bmatrix} \, .
\end{equation}
For the maximum probability of error, Eq.~(\ref{Gaussperrdiscord}), the upper bound is achieved, from
Eqs.~(\ref{Gaussbounds}) and (\ref{QCBmax}), in terms of a simple linear function of the Hellinger Gaussian
discord of response. The latter, in turn, is obtained by minimizing the Hellinger distance over the entire set of
local unitary operations implemented on the covariance matrix by local symplectic, traceless, transformations.
For squeezed thermal and thermal squeezed states one finds that this minimum is realized by the $\pi/2$ phase shift $F_{\pi /2}$,
that is the only possible traceless phase shift. Therefore the extremal unitary-coding protocol in the ensemble of local
traceless symplectic operations is realized by a particular PSK coding, the phase shift $\pi/2$, which is the only traceless
PSK coding. The details of the proof are reported in Appendix~\ref{chbm}.
On the other hand the quantity $LBP_{err}^{(\max )}$, Eq.~(\ref{LBPmax}), evaluated at $\pi/2$,
may not be optimal but certainly still provides a lower bound on the maximum probability of error:
\begin{equation}
LBP_{err}(F_{\pi/2})\leq LBP_{err}^{max}\leq P_{err}^{max} \; .
\end{equation}
Since for a $\pi/2$ phase shift the corresponding transformation is implemented by the traceless symplectic matrix $F_{\pi/2}=\begin{bmatrix}0&1\\-1&0\end{bmatrix}$, the expectation values of the canonical quadrature operators $x$ and $p$
transform as follows: $\<x\> \rightarrow -\<p\>$ and $\<p\> \rightarrow \<x\>$. Therefore, undisplaced thermal
Gaussian states ($\<x\> = \<p\> = 0$) are left invariant, and the worst-case PSK coding $(\mathbbm{1},F_{\pi /2})$ is
completely invisible to classical transmitters (thermal states) since the $\pi /2$ shift
does not change their covariance matrix. The probability of error $P_{err}$ for every such
classical transmitter always achieves the absolute maximum $1/2$. Viceversa, the very same coding can always be
read by any quantum Gaussian transmitter with nonvanishing Gaussian discord of response. As a consequence, quantum transmitters
always outperform undisplaced classical transmitters in device-independent, worst-case scenario quantum reading.
The situation changes when we consider displaced thermal states, as displacement unavoidably increases distinguishability.
Indeed the coherent and thermal coherent states are very efficient in detecting phase shift transformations. Nonetheless, in
the next section we will show that thermal coherent transmitters are outperformed by noisy quantum ones provided that the
distribution of the thermal noise among the modes in the quantum resource is strongly {\em non-symmetric}.
\section{Comparing classical and quantum resources: noise-enhanced quantum transmitters}
\label{comparisonclass}
We have seen that without displacement classical transmitters (thermal states) are completely
blind to reading. Introducing displacement enhances the distinguishability
of output states and turns classical states (thermal coherent states) in useful transmitters.
It is straightforward to show that distinguishability and the reading efficiency increase by implementing a single-mode
displacement rather than a two-mode one with equal single-mode amplitudes.
Let us then consider a scenario in which one compares
discordant quantum transmitters with displaced classical ones. We will show that in this
case, that is comparing noisy quantum resources with distinguishability-enhanced noisy classical ones,
discordant transmitters can outperform classical ones, and that the quantum advantage increases
with increasing (thermal) noise.
Stated precisely, given the same coding $(\1,F_{\pi /2})$ acting locally on the first mode, we want to identify the regimes in which the
probability of error associated to a quantum transmitter is smaller than the probability of error associated to a thermal coherent one.
From Eqs.~(\ref{QCB}) and (\ref{errorineq}) this is equivalent to identifying the regimes in which the upper bound $QCB$ on the
probability of error using squeezed thermal
transmitters, denoted by $QCB^{sq-th}$, is smaller than the lower bound $LBP_{err}$
using thermal coherent states, that will be denoted
by $LBP_{err}^{coh-th}$. Obviously, only a constrained comparison at given fixed
physical quantities is meaningful. We will thus compare squeezed
thermal states and displaced thermal states at
fixed purity and fixed total number of photons. We will observe that the quantum advantage is achieved provided
the covariance matrix is not symmetric with respect to exchange of the modes.
With these notations, the requirement for a \emph{bona fide} quantum advantage reads as follows:
\begin{equation}
QCB^{sq-th} \leq LBP_{err}^{coh-th} \; .
\label{uplow}
\end{equation}
Both the coherent thermal and squeezed thermal states are two extremal classes of the general family of states which can be described as squeezed displaced thermal states (SDTS), defined as:
\begin{equation}
\rho_{SDTS} = S(r)D(\alpha)\rho_{th}(N_{th_1},N_{th_2})D(\alpha)^{\dagger} S(r)^{\dagger} \; ,
\end{equation}
where $S(r)=\exp(r a_1^{\dagger}a_2^{\dagger}-r a_1a_2)$ is the two-mode squeezing operator and we assume that squeezing parameter $r$ is real. Here $D(\alpha)=\exp(\alpha a_1^{\dagger}-\bar{\alpha} a_1)$ is the single-mode displacement operator and $\rho_{th}(N_{th_1},N_{th_2})\equiv\rho_{th_1}\otimes \rho_{th_2}$ is the non-symmetric two-mode thermal state, where $\rho_{th_i}\equiv \frac{1}{1+N_{th_i}}\sum_{m=0}^\infty (\frac{N_{th_i}}{1+N_{th_i}})^m\ket{m_i}\bra{m_i}$.
The purity $\mu=1/(16\det\sigma)^{1/2}$ of the SDTS is a function of the covariance matrix $\sigma$ and
depends only on the number of thermal photons:
\begin{equation}
\mu=\frac{1}{(1+2N_{th_1})(1+2N_{th_2})} \; .
\end{equation}
The total number of photons, $N_T=<a_1^{\dagger}a_1+a_2^{\dagger} a_2>$, in the SDTS reads:
\begin{equation}
N_T = \left( N_{th_1} + N_{th_2}\right)\left( 1 + 2N_s \right) + 2N_s \left( 1 + |\alpha|^2 \right) + |\alpha|^2 \; ,
\end{equation}
where $N_s=\sinh(r)^2$ is the number of squeezed photons.
Putting $r=0$, SDTSs reduce to thermal coherent states $\rho(0,\alpha,N_{th_1},N_{th_2})$ with total number of photons $N_T=N_{th_1}+N_{th_2}+|\alpha|^2$. Decreasing the displacement amplitude $\alpha$ the distinguishability of coherent thermal transmitters is reduced. We want to investigate whether this loss of distinguishability can be compensated by the quantum contribution due to increase of $r$ keeping $N_T$ and the purity fixed. In the limiting situation when $\alpha=0$ the corresponding quantum state $\rho(r,0,N_{th_1},N_{th_2})$ is a squeezed thermal state (STS). In the following we will show that for STSs $\rho(r,0,N_{th_1},N_{th_2})$ and thermal coherent states $\rho(0,\alpha,N_{th_1},N_{th_2})$ with equal total number of photons $N_T$, Ineq.~(\ref{uplow}) is satisfied for some ranges of $N_{th_1}$ and $N_{th_2}$. The condition of equal total number of photons $N_T$
implies $|\alpha|^2 = 2\sinh(r)^2(1+N_{th_1}+N_{th_2})$.
In order to evaluate the Uhlmann fidelity ${\cal{F}}$ and the quantum Chernoff bound $QCB$ in Eq.~(\ref{uplow})
we need to know how the phase shift $F_{\pi/2}$ transforms the transmitters
that we wish to compare: the squeezed thermal states and the
thermal coherent states. The dependence of ${\cal{F}}$ and
$QCB$ on the displacement vector and on the covariance matrix of general
Gaussian states is reported in Appendices~\ref{UhlmannGaussian} and
\ref{ChernoffGaussian}. The Uhlmann fidelity providing the lower bound on
$P_{err}$ for thermal coherent states depends only on the displacement
vector, since the covariance matrix of thermal coherent states is unaffected by the
action of the symplectic transformation
$\left(F_{\pi/2}\oplus\mathbbm{1}_B\right)\sigma\left( F_{\pi/2}\oplus
\mathbbm{1}_B\right)^T$, where $F_{\pi/2}\oplus\mathbbm{1}_B=
\begin{bmatrix}
0 & -1 \\
1 & 0
\end{bmatrix}
\oplus\mathbbm{1}_B$.
Without loss of generality, the displacement vector of a thermal coherent state can be written as
$\<u\>_{coh-th} = [\sqrt{2}\abs{\alpha},0,0,0]^T$. Under a $\pi/2$ phase shift the difference $\delta$ between the final and
the initial displacement vectors reads as follows:
\begin{equation}
\delta=F_{\pi/2}%
\begin{bmatrix}
\abs{\sqrt{2}\alpha} \\
0 \\
0 \\
0%
\end{bmatrix}%
-%
\begin{bmatrix}
\abs{\sqrt{2}\alpha} \\
0 \\
0 \\
0%
\end{bmatrix}%
=%
\begin{bmatrix}
\abs{\sqrt{2}\alpha} \\
-\abs{\sqrt{2}\alpha} \\
0 \\
0%
\end{bmatrix}
\; .
\end{equation}
The Uhlmann fidelity of a thermal coherent state is then
\begin{equation}
\c F^{coh-th} = \exp{\left(-\frac{2\abs{\alpha}^2}{\c A}\right)} \; ,
\end{equation}
where $\c A=(1+2N_{th_1})$. The $QCB$ of non-symmetric, undisplaced squeezed thermal
state depends only on the covariance matrix, Eq.~(\ref{sqthcorm}), with
entries Eqs.~(\ref{ccvsqth}), and its explicit
expression is reported in Appendix~\ref{ChernoffGaussian}.
In Fig.~\ref{boundsenergy}, upper panel, we report the exact values of $P_{err}^{sq-vac}$ for a squeezed vacuum with squeezing $r$, $P_{err}^{coh}$ for a coherent state $\ket{\alpha}$, and $P_{err}^{sq-coh}$ for a squeezed displaced vacuum with squeezing $r'$ and displacement $\beta$, in the absence of noise, $N_{th_1}=N_{th_2} = 0$, and at fixed total photon number $N_T = \abs{\alpha}^2 = 2\sinh^2{(r)}=\abs{\beta}^2(1+2\sinh^2{(r')})+2\sinh^2{(r')}$.
The coherent states outperform the quantum resources given by the undisplaced squeezed vacuum. The coherent transmitters are then compared with squeezed displaced vacua of the same energy. Even if the latter include a classical contribution due to displacement and a quantum contribution due to squeezing they are still outperformed by the classical coherent states. The quantum efficiency converges to the classical one in the high-energy limit. For completeness, in Fig.~\ref{boundsenergy} we also report the quantum Chernoff bound $QCB^{sq-vac}$ for the squeezed-vacuum transmitters.
\begin{figure}[tbp]
\includegraphics[width=7.0cm]{gridnth1nth2.pdf}
\caption{Upper panel: behavior, as a function of the total photon number $N_T $, of the probability of error $P_{err}$ in the absence of thermal noise ($N_{th_1}=N_{th_2} = 0$). Blue dotted line: probability of error $P_{err}^{sq-vac}$ of
squeezed vacuum states. Black dashed line: $P_{err}^{coh}$ of
coherent states. Green dot-dashed line: $P_{err}^{sq-coh}$ of squeezed displaced vacuum states with displacement $\alpha=\frac{1}{2}N_{T}$. Red solid line: quantum Chernoff bound $QCB^{sq-vac}$ of squeezed vacuum states yielding the upper bound on $P_{err}$ with quantum transmitters.
No quantum gain is not observed in this regime.
Central panel: behavior of quantum and classical bounds on $P_{err}$ as
functions of $N_T$ at fixed asymmetric thermal noise: $N_{th_1} = 5$, $N_{th_2}=0$. Red
solid line: quantum upper bound $QCB^{sq-th}$ on $P_{err}$ with undisplaced
squeezed thermal states. Black dashed line: classical lower bound $LBP_{err}^{coh-th}$ on $P_{err}$
with thermal coherent states.
Lower panel: same as central panel, but with stronger thermal noise:
$N_{th_1} = 8$, $N_{th_2}=0$. With increasing $N_T$ the quantum upper bound goes below the
corresponding classical figures of merit and quantum transmitters certainly outperform classical ones.}
\label{boundsenergy}
\end{figure}
In the presence of symmetric thermal noise, $N_{th_1} = N_{th_2}$, there is no improvement in the quantum efficiency relative to the classical one. Introducing non-symmetric thermal noise, e.g. $N_{th_1}> N_{th_2}$, the Gaussian discords of response, that are intrinsically asymmetric quantities with respects to the subsystems in a given bipartition, increase dramatically, and so does the corresponding quantum reading efficiency. As a consequence, for sufficiently strong non-symmetric thermal noise the quantum resources outperform the classical ones. In the presence of non-symmetric noise exact expressions for $P_{err}$ are no
longer available. Therefore, in the central and lower panels of Fig.~\ref{boundsenergy} we report the exact lower and upper bounds on $P_{err}$ based
on the Uhlmann fidelity ${\cal{F}}$ and on the quantum Chernoff bound $QCB$. We observe that at intermediate
values of the total number of photons $N_T$ the quantum upper bound $QCB^{sq-th}$ on $P_{err}$ is strictly lower than the classical lower
bound $LBP_{err}^{coh-th}$, assuring that the quantum resources outperform the classical ones.
The classical transmitters (thermal coherent states) recover the quantum efficiency for large values of the total photon number.
Moreover, comparing the central and the lower panels in Fig.~\ref{boundsenergy}, we observe that as the number of thermal photons $N_{th_1}$ is
increased, the range of values of the total photon number $N_T$ for which
one has a quantum advantage increases.
In Fig.~\ref{rconstans} we provide a plot of the contour lines for the
differences $QCB^{sq-th} - LBP_{err}^{coh-th}$ for different asymmetries: $N_{th_2}=0$ (upper panel) and $N_{th_2}=0.5$ (lower panel) as functions of the total
photon number $N_T$ and of the purity (or, equivalently of the number of
thermal photons $N_{th_1}$). When these differences become negative, Ineq.~(\ref{uplow}) is satisfied and the quantum resources certainly outperform the
classical ones.
From the upper panel of Fig.~\ref{rconstans}, for $N_{th_2}=0$ comparing noisy quantum
transmitters with noisy coherent ones, one observes that $QCB^{sq-th} -
LBP_{err}^{coh-th} < 0$ in a large region of parameters. Fixing the
squeezing, so that the change in the total photon number $N_T$ is due only
to the change in the number of thermal photons $N_{th_1}$, corresponds to a
straight line in the plane (in the figure, drawn at $r=0.8$). Remarkably,
for these iso-squeezed states the quantum advantage increases with
increasing number of thermal photons. This is an instance of noise-enhanced
quantum efficiency that will be discussed further in Sec.~\ref{comparisonquant}.
In the lower panel of Fig.~\ref{rconstans} we decrease the asymmetry $(N_{th_2}=0.5)$.
We observe that the quantum gain is also achieved but in the range of much higher $N_{T}$.
Again, fixing the squeezing, e.g. at $r=0.8$, we notice that the quantum advantage increases with
thermal noise.
We remark that these results are obtained in a
scenario in which we compare the minimum quantum efficiency (upper bound on the
error probability using quantum transmitters) with the maximum classical
efficiency (lower bound on the error probability using coherent thermal
transmitters). Therefore the actual quantum advantage will be even larger.
\begin{figure}[!th]
\includegraphics[width=7.0cm]{gridcontour.pdf}
\caption{Contour plot providing the contour lines for the differences $
QCB^{sq-th} - LBP_{err}^{coh-th}$ for $N_{th_2}=0$ (upper panel) and for $N_{th_2}=0.5$
(lower panel) as functions of the total photon number $N_T $ and of the number of thermal photons $N_{th_1}$. The region in which
these quantities assume negative values corresponds to quantum transmitters
certainly outperforming coherent thermal ones. The red dashed curve identifies its
boundary. The straight solid yellow lines in both panels corresponds to a fixed degree
of squeezing $r=0.8$. Moving along this lines in the direction of increasing
number of thermal photons $N_{th_1}$ one observes that as noise grows there is
a growing advantage in using quantum transmitters over classical ones. This
behavior provides an instance of noise-enhanced quantum performance. The effect is reduced when the asymmetry $N_{th_1}-N_{th_2}$ decreases.}
\label{rconstans}
\end{figure}
The quantum advantage disappears in the symmetric situation $N_{th_1}=N_{th_2}$.
Indeed, the inequality $N_{th_1} > N_{th_2}$ expresses the condition that the mode which passes through the coding channels is more noisy. This condition is unfavorable for thermal coherent states and favorable for STSs. Namely, in STSs with fixed finite squeezing, increasing the number of thermal photons in the first mode certainly increases the discord of response and, as a consequence, increases also the reading efficiency for this type of transmitters. This phenomenon is further analyzed in the following sections on the comparison of different quantum transmitters. These two concatenated effects cause the advantage of quantum states over the classical transmitters in the protocol of quantum reading with noisy transmitters. The asymmetry between the local thermal noise terms is the crucial element for realizing the enhancement of the reading efficiency. As we will see in the next section, the behavior of STSs with increasing number of thermal photons in the symmetric situation $N_{th_1}=N_{th_2}$, although not sufficient to realize a quantum advantage over classical resources, favors STSs among other noisy quantum transmitters.
\section{Comparing noisy quantum resources}
\label{comparisonquant}
In the previous section we compared classical and quantum transmitters and
for worst-case scenario we identified the regimes in which noisy but discordant quantum resources
outperform classical thermal coherent ones. We also observed that the quantum advantage
can increase, at fixed squeezing, with increasing thermal noise. We will now
compare the behavior of squeezed-thermal and thermal-squeezed states
in order to investigate how thermal noise affects the quantum efficiency
of different classes of quantum transmitters. We shall compare symmetric squeezed thermal
and thermal squeezed transmitters either at fixed number of thermal photons or at fixed squeezing.
We will then consider how non-symmetric noise further enhances the quantum efficiency by suppressing
the upper bound on the probability of error. Finally, we will investigate how the quantum efficiency
of different quantum transmitters improves when multiple reading operations are implemented at
fixed thermal noise.
\subsection{Comparing symmetric squeezed thermal and thermal squeezed transmitters: fixed noise}
\label{quantumthermal}
Let us start by comparing quantum reading with symmetric squeezed thermal and symmetric thermal squeezed
transmitters at fixed number of thermal photons and its performance as a function of the total number of photons. This comparison
is motivated by the fact that the interplay between quantum and thermal fluctuations
is very different for these two classes of quantum states. Squeezed thermal states (STSs) are
obtained by applying on thermal states, namely states that have already
thermalized (e.g. at the output of a noisy channel) a purely quantum operation, two-mode squeezing, that can be interpreted
as a re-quantization of the thermal vacuum. Viceversa, thermal squeezed states (TSSs) are
realized by letting pure squeezed vacua evolve and eventually thermalize in a noisy channel.
Both squeezed thermal and thermal squeezed states are two extremal classes of the very general family of squeezed thermal squeezed displaced states (STSDSs) which are defined as follows:
\begin{eqnarray}
&& \rho(r,N_{th_1},N_{th_2},r',\alpha) = \\
&&S(r)\Phi_{N_{th_1},N_{th_2}}\Big[S(r')D(\alpha)\rho_{vac}D(\alpha)^{\dagger}S(r')^{\dagger}\Big] S(r)^{\dagger} \, .
\nonumber
\end{eqnarray}
Here $S(r)$ and $S(r')$ are two-mode squeezing operators with different squeezing parameters $r$ and $r'$,
$D(\alpha)$ is a single-mode displacement operator, $\Phi_{N_{th_1},N_{th_2}}$ is a noisy channel introducing $N_{th_1}$ and $N_{th_2}$
thermal photons respectively in the first and in the second mode. The channel acts on a given Gaussian state adding the number of thermal photons to the diagonal entries of its covariance matrix. Finally, $\rho_{vac}=\ket{00}\bra{00}$ denotes the two-mode vacuum state.
We study this family of states at constant fixed values of the parameters $N_{th_1}$ and $N_{th_2}$.
The total number of photons in a STSDS is:
\begin{eqnarray}
&&N_T=<a_1^{\dagger}a_1+a_2^{\dagger} a_2>=\\
&&(N_{th_1}\! +\! N_{th_2}) \cosh(2 r')\! +\! (1 \! + \! |\alpha|^2) \cosh(2 (r \! + \! r')) \! - \!1 \, .
\nonumber
\end{eqnarray}
Consider first the situation without displacement, $\alpha=0$, and with symmetric thermal noise $N_{th_1}=N_{th_2}=N_{th}$.
Putting $r=0$, STSDSs reduces to thermal squeezed states TSTs $\rho(0,N_{th},N_{th},r',0)$.
Decreasing $r'$ and correspondingly increasing $r$ while keeping $N_T$ and $N_{th}$ fixed, in the limit $r' \rightarrow 0$ one recovers
the squeezed thermal states STSs $\rho(r,N_{th},N_{th},0,0)$.
Let us compare these two extremal classes of quantum Gaussian transmitters that coincide for $N_{th}=0$ and differ for $N_{th} \neq 0$ or, in non symmetric situations, when either $N_{th_1} \neq 0$ and/or $N_{th_2} \neq 0$.
In Fig.~\ref{quantumquantum}, upper left panel, we observe that for nonvanishing but small number of thermal photons $N_{th}$ the upper bound on the probability of error $P_{err}$ using STSs still remains above the lower bound on $P_{err}$ using TSSs. By further increasing thermal noise, as shown in the upper right panel of Fig.~\ref{quantumquantum}, all bounds with STSs are below all bounds with TSSs and the STSs certainly outperform the TSSs. Due to the different effects of the noise in STSs and TSSs we observe the clear advantage of using STSs over TSSs in a quantum reading protocol.
The lower panels of Fig.~\ref{quantumquantum} show the comparison done for displaced thermal squeezed states $\rho(0,N_{th},N_{th},r',\alpha)$ and displaced squeezed thermal states $\rho(r,N_{th},N_{th},0,\alpha)$. At fixed total number of photons we observe that single-mode displacement always increases the reading efficiency for both classes of states, while decreasing the squeezing reduces and eventually wipes out the quantum advantage of STSs over TSSs. These two classes of states coincide in the limiting case $r=r'=0$ in which they both recover the classical thermal coherent states. As we have seen in the previous section, the advantage of STSs over classical states is recovered by considering non-symmetric thermal noise.
\begin{figure}[tbp]
\includegraphics[width=9.0cm]{gridquantquant1.pdf}
\caption{Upper left panel: behavior, as a function of the total number of photons $N_T $, of the upper and lower bounds on the probability of error $P_{err}$ using either undisplaced squeezed thermal states (STSs) or undisplaced thermal squeezed states (TSSs) with fixed, symmetric, thermal noise: $N_{th_1} = N_{th_2}=0.2$. Upper right panel: the same but with $N_{th_1}=N_{th_2}=2$. In this case one observes that beyond a threshold value of $N_{T}$ the STSs certainly outperform the TSSs. Lower panels similar to the upper panels but with nonvanishing displacement $|\alpha|^2=\frac{1}{2}(N_T-N_{th_1}-N_{th_2})$. The upper bounds on the probability of error are given here by the quantum Bhattacharyya coefficient $UBP_{err}=1/2\tr\sqrt{\rho_1}\sqrt{\rho_2}$, which for states that include displacement does not necessary coincide with the quantum Chernoff bound $QCB$. The displacement increases the efficiency of the reading, however it does not guarantee with certainty the noise-enhanced performance of STSs with respect to TSSs.}
\label{quantumquantum}
\end{figure}
\subsection{Comparing symmetric squeezed thermal and thermal squeezed transmitters: fixed squeezing}
When comparing the behavior of the Uhlmann fidelity, quantum Chernoff bound, and
the Gaussian discords of response under variations of the classical noise at a fixed level
of quantum fluctuations (squeezing), we expect a radically diverging behaviors of the STSs with
respect to the TSSs. On intuitive grounds, since fidelity, Chernoff bound, and discord are measures of
distinguishability between an input state and the corresponding output after
a local disturbance, if we compare STSs and TSSs we notice from the structure of
their covariance matrices, see Eqs.~(\ref{cthsq}), that as
$N_{th}$ increases the correlation part of the STSs increases, while it remains constant in TSSs.
Indeed, the quantum Chernoff bound $QCB$ and of the Uhlmann fidelity $\c F$ for any two Gaussian
states of the form Eq.~(\ref{sqthcorm}) with $a=b$ and $c=c_1=-c_2$, related by a $\pi/2$ phase shift, take the form:
\begin{equation}
QCB = \frac{a^2-c^2}{2a^2-c^2} \; ,
\label{derivchernoff}
\end{equation}
\begin{equation}
\c F = \frac{4}{\left[1+c^2-a^2+\sqrt{(c^2-a^2)^2+1+2a^2}\right]^{2}} \; .
\label{derivfidel}
\end{equation}
In Fig.~\ref{e18} we report the behavior of the upper bound on the probability of error $QCB$, Eq.~(\ref{derivchernoff}), and of
the lower bound $LBP_{err}$ (which is a monotonic increasing function of the Uhlmann fidelity $\c F$, Eq.~(\ref{derivfidel}))
for the squeezed thermal and thermal squeezed states as functions of the number of thermal photons at fixed squeezing.
We observe that for TSSs $\c F$ and $QCB$ both increase with increasing thermal noise, converging asymptotically to
the absolute maximum ($1/2$) of the probability of error. Therefore, the quantum efficiency of TSSs is
suppressed by increasing the thermal noise.
On the contrary, for STSs $QCB$ remains constant
and $LBP_{err}$ decreases. This behavior guarantees that the probability of error, at fixed squeezing, is bound to vary in a
restricted interval below $0.1$. In the given example, the squeezing amplitude $r$ has been fixed at a relatively low value
$r \simeq 0.9$. Increasing the level of squeezing will further reduce the maximum value achievable by the probability of error.
In conclusion, the quantum advantage associated to squeezed thermal states is paramount at fixed, even moderate, squeezing,
and increases monotonically with increasing thermal noise.
\begin{figure}[tbp]
\includegraphics[width=7.9cm]{sttsnew.pdf}
\caption{Behavior of the quantum Chernoff bound $QCB$ and of the lower bound on the probability of error $LBP_{err}$
as functions of the number of thermal photons $N_{th}$, at fixed number of squeezed photons $N_s = 1$, for thermal squeezed
and squeezed thermal states. Blue dashed line: $QCB$ for thermal squeezed states. Dotted red line: $LBP_{err}$ for thermal
squeezed states. Solid black line: $QCB$ for squeezed thermal states. Orange dot-dashed line: $LBP_{err}$ for squeezed thermal
states. The colored areas between the upper and lower bounds denote the admissible intervals of variation for the probability
of error $P_{err}$. Increasing thermal noise suppresses the efficiency of thermal squeezed transmitters and increases the
efficiency of squeezed thermal ones.}
\label{e18}
\end{figure}
A more detailed understanding of these opposite behaviors can be gained by looking at the variation of the measures of
distinguishability with respect to the variations of the thermal noise and of the parameters of the covariance matrix.
Consider a generic measure of distinguishability denoted by $f(\rho_1,\rho_2)$ where $f$, among others, includes the Uhlmann
fidelity $\c F$ and the quantum Chernoff bound $QCB$. Consider then the {\em total} derivative of $f$ with respect to $N_{th}$,
keeping $r$ constant:
\begin{equation}
\frac{df}{dN_{th}}=\frac{\partial f}{\partial a}\bigg|_{c}\frac{\partial a}{\partial N_{th}}+\frac{\partial f}{\partial c}\bigg|_{a}\frac{\partial c}{\partial N_{th}} \; .
\label{derivative}
\end{equation}
Specializing to either $\c F$ or $QCB$ we obtain the explicit expressions of their derivatives, as reported in Appendix \ref{Derivatives}. From these explicit
expressions it follows that it is always
$$
\frac{\partial QCB}{\partial a}\big|_{c} \geq 0 \; , \; \frac{\partial QCB}{\partial c}\big|_{a} \leq 0 \; , \; \frac{\partial \c F}{\partial a}\big|_{c} \geq 0 \; , \; \frac{\partial \c F}{\partial c}\big|_{a} \leq 0 \; ,
$$
irrespective of the type of quantum transmitter considered.
Hence, if $f$ represents either the Uhlmann fidelity or the quantum Chernoff bound, the derivative $\frac{\partial f}{\partial a}\big|_{c}\geq 0$.
This behavior agrees with the intuition that the operation of increasing the diagonal entries of the covariance matrix and keeping the off-diagonal entries constant acts like a thermal channel which makes the initial state and the final state after the phase shift less distinguishable.
The behavior $\frac{\partial f}{\partial c}\big|_{a}\leq 0$ for both $\c F$ and $QCB$ is also intuitively clear, since changes in $\sigma$ under $F_{\pi/2}$ are the greater the larger the off-diagonal entries when keeping the diagonal $a$ constant.
Let us now discuss the state-dependent derivatives: for STSs and TSSs the partial derivatives $\frac{\partial a}{\partial N_{th}}$
and $\frac{\partial c}{\partial N_{th}}$ are non-negative, therefore they cannot oppose the behavior of the state-independent part. For STSs they are given by $2\cosh{(2r)}$ and $2\sinh{(2r)}$ respectively, while for TSSs $\frac{\partial a}{\partial N_{th}}=2$ and $\frac{\partial c}{\partial N_{th}}=0$.
The behavior of $\cal F$ or $QCB$ with increasing $N_{th}$ depends then on the ratio of the positive and negative parts on the left hand side of Eq.~(\ref{derivative}).
For TSSs there is only a positive contribution in Eq.~(\ref{derivative}) and both $\c F$ and $QCB$ increase with increasing number of thermal photons. As a consequence, both the lower and the upper bounds on the probability of error must increase, as observed in Fig.~\ref{e18}.
On the other hand, for STSs the negative contribution always prevails when considering the Uhlmann fidelity, while the positive and negative contributions always cancel exactly when considering the quantum Chernoff bound, leading to a constant upper bound on the probability of error, as observed in Fig.~\ref{e18}.
The constant behavior of $QCB$ as a function of thermal noise for STSs can be also seen directly from Eq.~(\ref{derivchernoff}). This equation can be rewritten straightforwardly only in terms of $a/c$. Indeed, this ratio for STSs does not depend on $N_{th}$.
In this section we have considered reading protocols with binary coding given by the identity and the phase shift $\pi/2$, and transmitters implemented by symmetric STSs. This is actually a worst-case scenario in two respects. On the one hand, the phase shift $\pi/2$ provides the worst possible coding among all traceless local symplectic operations (maximum probability of error, device-independent reading). On the other hand, the {\em symmetric} STSs provide the worst possible transmitters among general STSs.
Indeed, in the next subsection we will show that non-symmetric STSs provide much larger quantum efficiencies and even effectively suppress the probability of error.
\subsection{Non-symmetric squeezed thermal states: noise-suppressed bounds on the probability of error}
\label{noiseenhancedQCB}
One might speculate that the increment of the Bures discord of response for increasing thermal noise and the corresponding decrement of the lower bound on the probability of error are due to the particular relation with the Bures metrics induced by the Uhlmann fidelity.
However, this is not the case. We will now show that if one considers non-symmetric two-mode STSs then also the Hellinger discord of response increases under increasing local thermal noise and therefore the corresponding upper bound on the probability of error decreases as well.
This is a strong indication that the true probability of error decreases as well with increasing thermal noise and thus that the use of discordant, non-symmetric STSs yields an absolute advantage, even over the use of entangled pure states, namely two-mode squeezed vacua with the same amount of squeezing as in the corresponding STSs.
The covariance matrix of non-symmetric two-mode STSs is given in Eq.~(\ref{sqthcorm}) with the parameters given in Eqs.~(\ref{ccvsqth}).
The corresponding $QCB$ achieves its maximum for the $\pi/2$ phase shift, as proven in Appendix~\ref{chbm}, and the exact expression of $QCB$ for non-symmetric STSs related by a $\pi/2$ phase shift is:
\begin{equation}
QCB=\frac{ab-c^2}{2ab-c^2}\;.
\label{derivchernoffas}
\end{equation}
Let us consider the variation $\frac{d QCB}{dN_{th_1}}$ of the quantum Chernoff bound, at constant squeezing $r$ and constant number of thermal photons
$N_{th_2}$ in the second mode, whose analytical expression is provided in Appendix~\ref{Derivatives}. From this expression it is clear that there is a range of values of $N_{th_1}$ and $N_{th_2}$, namely $N_{th_1} > N_{th_2}$, for which $\frac{d QCB}{dN_{th_1}} < 0$. Therefore, in this regime $QCB$ decreases with increasing $N_{th_1}$. On the other hand, $QCB$ increases with increasing $N_{th_1}$ if $N_{th_1}<N_{th_2}$. Henceforth, in the symmetric situation $N_{th_1}=N_{th_2}=N_{th}$ the quantum Chernoff bound is maximum and constant, independent of $N_{th}$, as discussed in the previous section.
\begin{figure}
\includegraphics[width=9cm]{qcbnth1.pdf}
\caption{Quantum Chernoff bound for non-symmetric STSs related by the phase shift $F_{\pi/2}$, as a function of $N_{th_1}$. In each panel the number of thermal photons $N_{th_2}$ in the second mode is fixed at a constant value. Upper left panel: $N_{th_2}=0.01$. Upper right panel: $N_{th_2}=0.1$. Lower left panel: $N_{th_2}=1$. Lower right panel: $N_{th_2}=1$ and extended range of values of $N_{th_1}$, in order to show the asymptotic vanishing of $QCB$ with increasing local thermal noise. For all panels the two-mode squeezing is fixed at $r=0.5$. The maximum of $QCB$ is achieved for symmetric STSs and provides the upper bound on the maximum probability of error $P_{err}^{\max}$ of the worst-case scenario.}
\label{qcbnth1}
\end{figure}
In Fig.~\ref{qcbnth1} we report the behavior of $QCB$ as a function of $N_{th_1}$ for different fixed values of $N_{th_2}$ and fixed squeezing $r$. In this physical situation the quantum Chernoff bound decreases with increasing {\em local} thermal noise and vanishes asymptotically for $N_{th_1} \rightarrow \infty$. Therefore the probability of error in a Gaussian quantum reading protocol can be made arbitrarily small by using non-symmetric STSs transmitters with very large {\em local} thermal noise.
This very remarkable result may look at first quite counter-intuitive. In fact, the crucial point is that this feature is obtained by the {\em global quantum} operation of two-mode squeezing applied to a two-mode thermal state with very strong asymmetry in the {\em local} thermal noises affecting the two field modes. It is therefore not entirely unexpected that the consequences can be dramatic. While entanglement certainly decreases, the operation of squeezing a larger amount of noise can increase quantum state distinguishability by {\em "orthogonalizing"} on a larger portion of Hilbert space with respect to the thermal states.
\subsection{Squeezed thermal and squeezed vacuum states}
\label{noisypure}
Collecting all the previous results we are finally in the position to compare the best resources of device-independent Gaussian quantum reading, namely the noisy and discordant non-symmetric STSs, to the best absolute resources of Gaussian quantum reading, namely pure entangled two-mode squeezed vacuum states (TMSVSs). In the limit of infinite squeezing the TMSVSs are maximally entangled pure Einstein-Podolsky-Rosen (EPR) states whose probability of error in a quantum reading protocol vanishes identically. In absolute terms, TMSVSs are certainly the best among classical and quantum resources in a reading protocol with continuous variables. Indeed, in Fig.~\ref{sqthsqvac} we report the behavior of the exact probability of error for TMSVSs and the lower bound on it for non-symmetric STSs as functions of the total number of photons at fixed thermal noise, that is for {\em arbitrarily increasing squeezing} as the total number of photons increases. One observes that the lower bound on the probability of error for non-symmetric STSs is always above the exact probability of error for TMSVSs, converging towards it only asymptotically.
\begin{figure}
\includegraphics[width=8cm]{sqthsqvac.pdf}
\caption{Behavior as a function of the total number of photons $N_T$ of the probability of error $P_{err}^{sq-vac}$ for two-mode squeezed vacuum transmitters (TMSVs) and of the lower bound on the probability of error $LBP_{err}^{sq-th}$ for two-mode squeezed thermal transmitters (STSs) with $N_{th_1}=1$ and $N_{th_2}=0$. TMSVs have better reading efficiency than STSs. The squeezing in TMSVs is larger than the one in STSs at each fixed value of $N_T$. The two efficiencies converge asymptotically with increasing total number of photons.}
\label{sqthsqvac}
\end{figure}
On the other hand, it is also important to compare TMSVSs and non-symmetric STSs in terms of the concrete use of resources in realistically feasible experimental scenarios. In Fig.~\ref{stsvstmsvs} and Fig.~\ref{stsvstmsvs1} we report the behavior of the exact probability of error $P_{err}^{sq-vac}$ associated to TMSVS transmitters compared to the upper and lower bounds $QCB^{sq-th}$ and $LBP_{err}^{sq-th}$ for non-symmetric STS transmitters as functions of the total number of photons. In the case of TMSVSs the total number of photons obviously depends only on the squeezing and the behavior of the probability of error is the same as the one reported in Fig.~\ref{sqthsqvac}. However, at variance with Fig.~\ref{sqthsqvac}, in Fig.~\ref{stsvstmsvs} and \ref{stsvstmsvs1} we compare it with the lower and upper bounds for ST transmitters at a {\em fixed} finite value, low and comparably easy to produce experimentally, of the squeezing. In this case, the total number of photons in STSs varies only with the amount of thermal photons. Fig.~\ref{stsvstmsvs} shows the behavior of the exact probability of error for TMSVS transmitters and the bounds on the probability of error for non-symmetric STSs, as functions of the total number of photons $N_T$ and constant squeezing parameter fixed at $r=0.5$. For comparison, in Fig.~\ref{stsvstmsvs1} we report the same quantities but for a larger fixed two-mode squeezing $r=1$. Indeed, the higher the squeezing, the better the bounds on the probability of error using STS transmitters approximate the exact probability of error for TMSVS transmitters.
The crucial difference is that in real-world experimental setups it is comparatively much easier and less resource-demanding to implement a scheme relying on non-symmetric STSs with enhanced thermal noise and quantum discord than to produce pure (noise-free) TMSVSs with enhanced squeezing and entanglement.
Therefore, at {\em fixed} squeezing, we can compare the two classes of transmitters for {\em different} values of the total number of photons $N_T$ and ask for the threshold value of $N_{th_1}$ above which the discordant STSs certainly perform better than the entangled TMSVSs at the same fixed level of squeezing (the noise on the second mode being also fixed at a given reference value, say e.g. $N_{th_2}=0$). This threshold is thus determined by the condition $QCB^{sq-th} = P_{err}^{sq-vac}$. We give here two numerical examples for two different realistic values of the two-mode squeezing achievable in the laboratory with current technologies. For $r=0.5$, we have that $QCB^{sq-th} \leq P_{err}^{sq-vac}$ as soon as $N_{th_1} \geq 3.6$. For $r=1$, we have that $QCB^{sq-th} \leq P_{err}^{sq-vac}$ as soon as $N_{th_1} \geq 2.6$. Therefore, the higher the fixed level of squeezing, the lower is the level of thermal noise and quantum discord required to match the performance of pure entangled TMSVSs. Alternatively, increasing $N_{th_1}$ further above the threshold, we can also look for the complementary information on the minimum threshold values of $r$ (more easily realizable in the laboratory) above which STSs match or surpass the performance of TMSVSs at higher values of the squeezing (harder to achieve experimentally). In other words, we can introduce the concept of {\em effective} squeezing $r_{eff}$ associated to the value $N_{th}^{eff}(r,r_{eff})$ such that for $N_{th} > N_{th}^{eff}(r,r_{eff})$ STSs perform better than TMSVSs with given squeezing $r > r_{eff}$.
In conclusion, device-independent quantum reading is a remarkable protocol of quantum technology with noisy resources for which the best transmitters are discordant non-symmetric squeezed thermal states whose performance is optimized by realizing a fine trade-off between increased local thermal noise and fixed global two-mode squeezing, yielding noise-enhanced quantum correlations and state distinguishability.
\begin{figure}
\includegraphics[width=8cm]{stsvstmsvs.pdf}
\caption{Behavior as functions of the total number of photons $N_T$ of the probability of error $P_{err}^{sq-vac}$ for two-mode squeezed vacuum transmitters, and of the lower and upper bounds on the probability of error $LBP_{err}^{sq-th}$ and $QCB^{sq-th}$ for non-symmetric two-mode squeezed thermal transmitters. The latter two quantities are plotted for variable $N_{th_1}$ at {\em fixed} squeezing $r=0.5$, as well as fixed reference thermal noise in the second field mode $N_{th_2} = 0$.}
\label{stsvstmsvs}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{figr1.pdf}
\caption{Behavior as functions of the total number of photons $N_T$ of the probability of error $P_{err}^{sq-vac}$ for two-mode squeezed vacuum transmitters, and of the lower and upper bounds on the probability of error $LBP_{err}^{sq-th}$ and $QCB^{sq-th}$ for non-symmetric two-mode squeezed thermal transmitters. The latter two quantities are plotted for variable $N_{th_1}$ at {\em fixed} squeezing $r=1$, as well as fixed reference thermal noise in the second field mode $N_{th_2} = 0$.}
\label{stsvstmsvs1}
\end{figure}
\subsection{Many copies}\label{manycop}
Let us now analyze the case in which the total number of photons can vary by considering many copies of the transmitter,
that is repeating the reading protocol many times independently. Using $n$ copies of the system the Uhlmann fidelity
and the quantum Chernoff bound decrease as powers of $n$. Therefore, the probability of error can decrease both in the
case of squeezed thermal and thermal squeezed states.
The interesting question which arises here is how many copies we need in both cases to achieve a given level of probability of error.
The number of copies defines for instance the time needed for reading one bit of information in the given coding. Therefore this process
is interesting from the point of view of assessing the reading time and the strength of the sources of squeezed light that one needs.
\begin{figure}[!th]
\includegraphics[width=8cm]{nnew.pdf}
\caption{Upper and lower bounds on the probability of error, using squeezed thermal (STSs) and thermal squeezed (TSSs) transmitters,
as a function of the number of copies of each transmitter, at fixed number of squeezed and thermal photons in each single
copy: $N_s=0.1$ and $N_{th}=1$. Blue dashed line: $QCB$ for thermal squeezed states. Dotted red line: $LBP_{err}$ for thermal
squeezed states. Solid black line: $QCB$ for squeezed thermal states. Orange dot-dashed line: $LBP_{err}$ for squeezed thermal
states. In order to achieve $P_{err}=1/8$ it is enough to take at most $n=7$ copies of STSs, while
the needed number of copies of TSSs is at least $n=20$.}
\label{en}
\end{figure}
Let us for instance assume that we require a value of the probability of error $1/8$, having for each copy of the squeezed thermal transmitter
the thermal noise fixed at $N_{th}=1$ and the weak squeezing fixed at $N_s=0.1$, see Fig.~\ref{en}. Looking at the upper bound (worst-case scenario),
the number of copies which are needed, in order to achieve the desired level of probability of error, is at most $n=7$.
Taking instead the thermal squeezed transmitter with the same squeezing and thermal noise in each copy, we see from Fig.~\ref{en}
that one needs, considering the lower bound (best-case scenario), at least $n=20$ copies.
These behaviors illustrate very clearly the advantage of using noise-enhanced quantum correlations. Indeed, comparing Figs.~\ref{en} and \ref{e18}, we see that by keeping a fixed level of squeezing and increasing the thermal noise, the number of copies of squeezed thermal transmitters needed to achieve a given level of precision stays constant, while the number of copies of thermal squeezed transmitters must increase.
\section{Conclusions and outlook}
\label{summary}
We have investigated Gaussian quantum reading protocols realized by weak optical sources in the worst-case scenario for quantum transmitters with respect to classical (thermal coherent) ones. For protocols that involve local unitary operations in the process of reading by continuous-variable Gaussian optical fields, we have showed that the maximum probability of error in reading binary memory cells is directly related to the amount of quantum correlations in a given transmitter, as quantified by the trace Gaussian discord of response. This relation allows to quantify the reading efficiency in terms of quantum correlations, providing a natural operational interpretation to the Gaussian discord of response.
Indeed, the latter is a well-defined measure of quantum state distinguishability under the action of local unitary operations. Therefore, the more discordant is the transmitter, the smaller is the maximum probability of error when using quantum resources. This relation then allows to determine the physical regimes of state purity and signal strength for which one has a net advantage in using quantum resources over classical thermal coherent ones.
Since the trace distance is in general uncomputable for Gaussian states, we have introduced exact upper and lower bounds on the maximum probability of error. We have showed that these bounds are expressed in terms of other type of quantum discords. In particular, the lower bound is expressed in terms of the Bures Gaussian discord of response, while the upper bound, provided by the quantum Chernoff bound maximized over the set of possible local unitary operations, is expressed in terms of the Hellinger Gaussian discord of response for squeezed thermal states and thermal squeezed states.
Both bounds decrease with an increasing amount of quantum correlations, providing a precise quantitative estimate of the quantum advantage obtained by using discordant resources over the corresponding thermal coherent ones. Moreover, the Bures and Hellinger discords of response are of further independent interest, as they play a central role in other quantum protocols studied recently, ranging from the assessment and use of local quantum uncertainty in optimal phase estimation~\cite{Girolami2013}, the efficiency of black-box quantum metrology~\cite{Interferometry2014,Spehner2013,GaussianMetrology2014}, and the quantum advantage of discordant resources in the protocol of quantum illumination~\cite{Farace2014}.
After comparing quantum and classical resources, we have discussed two fundamental classes of Gaussian quantum transmitters: symmetric squeezed thermal states (STSs) and symmetric thermal squeezed states (TSSs). We have shown that the actual beneficial or detrimental effects of environmental noise depend on the type of quantum state being considered. Considering STSs as quantum transmitters, the upper and lower bounds on the probability of error decrease with increasing thermal noise and therefore the quantum reading efficiency increases. The opposite behavior is observed when considering TSSs: in this case both the upper and the lower bounds on the maximum probability of error increase and therefore the quantum reading efficiency decreases with increasing thermal noise.
Finally, we went a step further and investigated the use of non-symmetric STSs. For such transmitters, also the quantum Chernoff bound decreases when the local thermal noise increases in one mode and remains fixed in the second mode. Indeed, the quantum Chernoff bound vanishes asymptotically with very large local thermal noise and therefore the probability of error must also vanish. In other words, non-symmetric two-mode STSs with imbalanced thermal noise between the two modes achieve an asymptotically vanishing probability of error for very large values of the noise imbalance. For such asymptotic states the Hellinger and Bures Gaussian discords of response attain their maximum value, and the quantum Chernoff bound and Uhlmann fidelity vanish. As a consequence, all upper and lower bounds on the probability of error vanish, the probability of error itself vanishes, and perfect reading is approached asymptotically.
Since the quantum reading efficiency of non-symmetric two-mode squeezed thermal states is a non-decreasing function of thermal noise, there is no evident advantage in using pure-state squeezed transmitters, the two-mode squeezed vacuum states, or low-noise ones over non-symmetric two-mode STSs with large noise imbalance between the field modes, {\em as long as} the squeezing is kept {\em fixed} at a realistic, finite constant value achievable in concrete experiments with currently available technology. Hence, noisy STSs transmitters {\em can} provide a better quantum efficiency at {\em fixed} two-mode squeezing, provided thermal noise (number of thermal photons) is enhanced beyond the threshold value above which the upper bound on the probability of error (quantum Chernoff bound) for STSs goes below the exact probability of error for TMSVSs with the same, fixed level of squeezing.
This remarkable phenomenon of noise-assisted quantum correlations and quantum efficiency is eventually due to the fact that quantum state distinguishability is intimately related to the concept of geometric quantum correlations, as measured by the discords of response, and the observation that the former can increase under increasing thermal noise. In particular, maximum {\em local} noise enhancement leads to maximum {\em global} enhancement of quantum correlations. In forthcoming studies we will provide a general characterization and quantification of noise-suppressed vs. noise-enhanced quantum correlations for different classes of quantum states~\cite{taming2014}, and we will investigate the relations between different types of quantum correlations according to states, metrics, and operations~\cite{RSI2014}.
\acknowledgments
F.I. acknowledges valuable discussions with Gerardo Adesso. The authors acknowledge financial support from the Italian
Ministry of Scientific and Technological Research under the PRIN 2010/2011 Research Fund, and from the EU FP7 STREP Projects iQIT,
G.A. No. 270843, and EQuaM, G.A. No. 323714.
|
\section{Introduction}
The world largest statistics of $\tau$ leptons collected
at $e^+ e^-$ $B$ factories (Belle \cite{belle,kekb} and
$BABAR$ \cite{Aubert:2001tu,Kozanecki:2000cm}) and
LHCb \cite{Alves:2008zz} opens new era in the precision tests of the Standard Model (SM).
Basic tau properties like lifetime, mass, couplings,
electric dipole moment (EDM), anomalous magnetic dipole moment
and other are introduced as free parameters in the theory,
or they can be calculated in the SM. Hence the former
parameters should be measured experimentally as precise as
possible, while the latter ones provide the unique
possibility to test SM and search for the effects
of New Physics (NP). An essential progress has been made
in the study of the main $\tau$ properties at Belle and $BABAR$,
namely lifetime \cite{Belous:2013dba,Lusiani:2005sy},
mass \cite{Abe:2006vf,Aubert:2009ra}, EDM \cite{Inami:2002ah},
coupling constant ratios \cite{Aubert:2009qj} have been
measured with the best or competitive to the world best
accuracies \cite{Beringer:1900zz}.
In the SM $\tau$ decays due to the charged weak interaction
described by the exchange of $W^{\pm}$ with a pure vector
coupling to only left-handed chirality fermions. There are
two main classes of tau decays: leptonic decays\footnote{Unless specified
otherwise, charge conjugate decays are implied throughout the paper.}
($\tau^-\to \ell^-\bar{\nu_{\ell}}\nu_{\tau}$,
$\tau^-\to \ell^-\bar{\nu_{\ell}}\nu_{\tau}\gamma$,
$\tau^-\to \ell^- \ell'^+\ell'^- \bar{\nu_{\ell}}\nu_{\tau}$;~
$\ell,\ell' = e,\mu$), and hadronic decays.
Leptonic decays provide very clean laboratory to probe
electroweak couplings \cite{Fetscher:1993ki}, which is
complementary or competitive to the precision studies of muon
in the experiments with muon beams \cite{Kuno:1999jp}. Plenty
of NP models can be tested and constrained in the precision
studies of the dynamics of $\tau$ decays with leptons \cite{Haber:1978jt}-\cite{McKeen:2011aa}.
Hadronic decays of $\tau$ offer unique tools for the precision
study of low energy QCD \cite{Pich:1997ym}. The hadronic system
is produced from the QCD vacuum via decay of the $W^-$ boson
into $\bar{u}$ and $d$ quarks (Cabibbo-allowed decays)
or $\bar{u}$ and $s$ quarks (Cabibbo-suppressed decays).
As a result the decay amplitude can be factorized into a
purely leptonic part including the $\tau$ and $\nu_{\tau}$
and a hadronic spectral function.
Various decay modes are interesting to study precisely the structure of the
hadronic spectral functions~\cite{Kuhn:1992nz} and measure precisely parameters
of the intermediate states, testing the Wess-Zumino-Witten anomaly~\cite{Wess:1971yu},
chiral theory~\cite{Pich:1987qq,Pich:1995bw}, and relations to $e^+e^-$ cross sections
following from the conservation of the vector current~\cite{Eidelman:1990pb}.
Measurement of the inclusive hadronic spectral function of Cabibbo-allowed
decays is important for the precision determination of $\alpha_s$ \cite{Braaten:1991qm},
while the inclusive strange hadronic spectral function is used to evaluate
$s$-quark mass and $V_{us}$ element of Cabibbo-Kobayashi-Maskawa (CKM)
quark flavor-mixing matrix \cite{Gamiz:2004ar}. Recently lots of important results in the hadronic
sector of $\tau$ physics have been obtained at Belle and $BABAR$ \cite{Bevan:2014iga}.
In the leptonic sector CP symmetry violation (CPV) is strongly suppressed
in the SM ($A_{\rm SM}^{\rm CP}\lesssim 10^{-12}$) leaving enough room to
search for the effects of NP~\cite{Delepine:2005tw}. Of particular interest are
strangeness changing Cabibbo-suppressed hadronic $\tau$ decays, in which
large CPV could appear from a charged scalar boson exchange~\cite{Weinberg:1976hu}-\cite{Kiers:2008mv}.
Probabililty of lepton flavor violating (LFV) decays of charged leptons is
extremely small in the Standard Model
(for example ${\cal B}(\tau\to\ell\nu) \sim \Delta m^4_\nu / m^4_W < 10^{-54}$ \cite{Deppisch:2012vj}).
Many models beyond the SM predict LFV decays with the branching fractions
up to $\lesssim 10^{-7}$ \cite{Ellis:2002fe}-\cite{CorderoCid:2005gr}.
As a result observation of LFV is a clear signature of New Physics.
$\tau$ lepton is an excellent laboratory to search for the LFV
decays due to the enhanced couplings to the new particles as well as large
number of LFV decay modes. Study of different $\tau$ LFV decay modes allows
one to test various NP models. Huge statistics collected by Belle and $BABAR$
was used in the searches for 48 LFV $\tau$ decays, the upper limits on the
branching fractions for the most of LFV modes approach $10^{-8}$ level, which
allows one to constrain the parameter spaces of many NP models \cite{Bevan:2014iga}.
Recently LHCb collaboration performed results of the search for LFV $\tau^-\to\mu^-\mu^+\mu^-$,
lepton number (LNV) and barion number violating (BNV) $\tau$ decays with proton
$\tau^-\to p\mu^-\mu^-,~\bar{p}\mu^+\mu^-$ at the Large Hadron Collider \cite{Aaij:2013fia}.
\section{Measurement of $\tau$ lifetime at Belle}
\subsection{Tests of lepton universality}
Lepton universality in the charged lepton sector of the SM
is the fundamental assumption about lepton flavor-independent
structure of the charged weak interaction. It is introduced in
the theory as an equality of the coupling constants for $e^-$,
$\mu^-$ and $\tau^-$: $g_e=g_\mu=g_\tau$.
This universality can be experimentally tested by comparing
the rates of the leptonic decays: $\tau^-\to e^-\bar{\nu}_e\nu_\tau$,
$\tau^-\to \mu^-\bar{\nu}_\mu\nu_\tau$ and $\mu^-\to e^-\bar{\nu}_e\nu_\mu$.
The total decay width with electroweak radiative corrections of
lepton $L^-$ ($L=\mu,~\tau$) reads \cite{Marciano:1988vm}:
\begin{equation}
\Gamma(L^-\to\ell^-\bar{\nu}_{\ell}\nu_L(\gamma))=\frac{{\cal B}(L^-\to\ell^-\bar{\nu}_{\ell}\nu_L(\gamma))}{\tau_L}=\frac{g^2_L g^2_\ell}{32M^4_W}\frac{m^5_L}{192\pi^3}F_{\rm corr}(m_L,m_\ell),
\end{equation}
\begin{equation}
F_{\rm corr}(m_L,m_\ell)=f(x)\biggl(1+\frac{3}{5}\frac{m^2_L}{M^2_W}\biggr)\biggl(1+\frac{\alpha(m_L)}{2\pi}\biggl(\frac{25}{4}-\pi^2\biggr)\biggr),
\end{equation}
\begin{equation}
f(x)=1-8x+8x^3-x^4-12x^2\ln{x},~x=m_\ell/m_L,
\end{equation}
where $m_L$($m_\ell$) and $g_L$($g_\ell$) are mass and coupling constant of initial (final) lepton,
$\tau_L$ is lifetime of initial lepton, $M_W$ and $\alpha(m_L)$ are $W^-$ boson mass and fine-structure
constant at the energy scale of $m_L$.
Taking into account that \cite{Beringer:1900zz}:
\[ {\cal B}(\mu^-\to e^- \bar{\nu}_e\nu_\mu(\gamma))={\cal B}(\mu^-\to e^- \bar{\nu}_e\nu_\mu)+\]
\begin{equation}
{\cal B}(\mu^-\to e^- \bar{\nu}_e\nu_\mu\gamma)+{\cal B}(\mu^-\to e^- \bar{\nu}_e\nu_\mu e^+ e^-)=1,
\end{equation}
the $g_\tau /g_e$ and $g_\tau /g_\mu$ ratios of the coupling constants can be extracted:
\begin{equation}
\frac{g_\tau}{g_e}=\sqrt{{\cal B}(\tau^-\to\mu^-\bar{\nu}_\mu\nu_\tau(\gamma))\frac{\tau_\mu}{\tau_\tau}\frac{m^5_\mu}{m^5_\tau}\frac{F_{\rm corr}(m_\mu,m_e)}{F_{\rm corr}(m_\tau,m_\mu)}},
\label{gtau:ge}
\end{equation}
\begin{equation}
\frac{g_\tau}{g_\mu}=\sqrt{{\cal B}(\tau^-\to e^-\bar{\nu}_\mu\nu_\tau(\gamma))\frac{\tau_\mu}{\tau_\tau}\frac{m^5_\mu}{m^5_\tau}\frac{F_{\rm corr}(m_\mu,m_e)}{F_{\rm corr}(m_\tau,m_e)}}.
\label{gtau:gmu}
\end{equation}
As it is seen from Eq.~\ref{gtau:ge} and \ref{gtau:gmu} precise measurement of the branching fraction
of leptonic $\tau$ decay, $\tau$ mass, and $\tau$ lifetime are necessary for the accurate tests of
lepton universality.
According to the last report from Heavy Flavor Averaging Group (HFAG) \cite{Amhis} lepton
universality was confirmed in the $g_\tau /g_e$ and $g_\tau /g_\mu$ ratios with the accuracy
of about 0.2\%:
\[ g_\tau/g_e=1.0024\pm 0.0021,~g_\tau/g_\mu=1.0006\pm 0.0021 \]
However, recently LEP Electroweak Collaboration published results on the test of
lepton universality in W-boson decays \cite{Schael:2013ita}, and the ratio of the
branching fraction of $W^-$-boson decay to $\tau^-\bar{\nu}_\tau$ to the average
branching fraction of $W^-$-boson decay to $\mu^-\bar{\nu}_\mu$ and $e^-\bar{\nu}_e$
was found to be 2.6 standard deviations away from unity:
\[\frac{2{\cal B}(W^-\to\tau^-\bar{\nu}_\tau)}{{\cal B}(W^-\to\mu^-\bar{\nu}_\mu) + {\cal B}(W^-\to e^-\bar{\nu}_e)}=1.066\pm 0.025.\]
So, the improvement of the accuracy of the test of lepton universality with
leptonic decays of $\tau$ is still rather actual task. The reduction of the
uncertainty of tau lifetime to the negligible level will allow one to test
lepton universality in $g_\tau /g_e$ and $g_\tau /g_\mu$ ratios with the
accuracy of about 0.1\% even with the current values of the $\tau$ mass
and branching fraction uncertainties.
\subsection{Measurement of $\tau$ lifetime at Belle}
This analysis \cite{Belous:2013dba} is based on the statistics with the luminosity integral of
$\int Ldt=711$~fb$^{-1}$ collected with the Belle detector
at the KEKB asymmetric-energy $e^+ e^-$ collider operating
at the $\Upsilon(4S)$ resonance and 60~MeV below. The data sample comprises
653$\times 10^6~\tau^+\tau^-$ pairs. Events where both taus decay to three
charged pions and neutrino were selected: $e^+e^-\to\tau^+\tau^-\to (\pi^+\pi^+\pi^-\bar{\nu}_{\tau},~\pi^+\pi^-\pi^-\nu_{\tau})$
(or shortly $(3\pi)^+ - (3\pi)^-$).
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.4\textwidth]{fig1.eps}
\hspace{5mm}
\includegraphics[width=0.4\textwidth]{fig2.eps}
\\
\parbox[t]{0.48\textwidth}{\caption{The scheme of $(3\pi)^+ - (3\pi)^-$ event in laboratory frame.}\label{ctaumeth}}
\parbox[t]{0.48\textwidth}{\caption{The scheme of $(3\pi)^+ - (3\pi)^-$ event
in c.m.s.~$\cos{\psi_{1,2}} = \frac{2E_{\tau}E_{h_{1,2}}-M^2_{\tau}-m^2_{h_{1,2}}}{2p_{\tau}p_{h_{1,2}}}$,
$h^-_1 = (3\pi)^-$, $h^+_2 = (3\pi)^+$.}\label{taudir}}
\end{figure}
At the asymmetric-energy $e^+ e^-$ collider the angle between
$\tau^+$ and $\tau^-$ in laboratory frame is smaller than $180^\circ$,
so the $\tau^+\tau^-$ production point can be calculated from the
intersection of two trajectories defined by the $\tau$-lepton
decay vertices and their momentum directions, see Fig.~\ref{ctaumeth}.
The position of beam interaction point is not needed at all in this method.
$\tau$ momentum direction is determined with the two-fold ambiguity in the
center-of-mass system (c.m.s.), for the analysis average axis is used,
see Fig.~\ref{taudir}.
In this method lifetimes of $\tau^-$ and $\tau^+$ can be measured
separately to test CPT symmetry conservation.
The following criteria were applied to select signal $(3\pi)^+ - (3\pi)^-$
events: six charged pions with zero net charge and no
other tracks are found; the thrust value in the c.m.s.~is
greater than 0.9; three pions (triplet) with $\pm 1$ net charge
in each hemisphere, separated by the plane perpendicular
to the thrust axis in the c.m.s.; there are no additional
$K^0_S$, $\Lambda$, and $\pi^0$ candidates; absolute value
of the transverse momentum of the $6\pi$ system is greater
than 0.5~GeV/$c$; the mass of $6\pi$ system should
satisfy the requirement $4$~GeV/$c^2<M(6\pi)<10.25$~GeV/$c^2$; the
pseudomass of each triplet of pions is $\sqrt{M^2_h+2(E_{\rm beam}-E_h)(E_h-P_h)}<1.8$~GeV/$c^2$,
$h=(3\pi)^-,~(3\pi)^+$; each triplet vertex-fit quality fulfill $\chi^2<20$;
the distance ($dl$) of the closest approach of $\tau^-$ and $\tau^+$
trajectories in laboratory frame satisfy $dl<0.03$~cm.
Finally 1148360 events were selected with the background contamination
of about 2\%, the dominant background comes from the continuum
$e^+ e^-\to q\bar{q}$ ($q=$u,~d,~s) events.
The probability density function (p.d.f.) for the measured
$\tau$ decay length distribution is written in the form:
\begin{equation}
{\cal P}(x)={\cal N}\int e^{-x'/\lambda_{\tau}}R(x-x';\vec{P})dx'+{\cal N}_{\rm uds}R(x;\vec{P})+B_{\rm cb}(x),
\end{equation}
where $x = \ell/(\beta_{\tau}\gamma_{\tau})$ is normalized $\tau$ decay length, ${\cal N}$ is normalisation
constant, $\lambda_{\tau}$ is estimator of $c\tau_{\tau}$ and $c\tau_{\tau}=\lambda_{\tau}+\Delta_{\rm corr}$
($\Delta_{\rm corr}$ is determined from MC), ${\cal N}_{\rm uds}$ is contribution of background
from $e^+e^-\to q\bar{q}~(q=u,~d,~s)$ (predicted by MC), $B_{\rm cb}(x)$ is p.d.f. contribution to describe background from
$e^+e^-\to q\bar{q}~(q=c,~b)$ (fixed from MC), $R(x;\vec{P})$ is detector resolution function (see Fig.~\ref{ctaures}),
parametrized by:
\[ R(x;\vec{P}) = (1-2.5x)\cdot\exp\biggl(-\frac{(x-P_1)^2}{2\sigma^2}\biggr),\]
\begin{equation}
\sigma=P_2+P_3|x-P_1|^{1/2}+P_4|x-P_1|+P_5|x-P_1|^{3/2}.
\label{resofun}
\end{equation}
$\lambda_{\tau}$, ${\cal N}$ and $\vec{P}=(P_1,...,P_5)$ are free parameters of the fit.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.4\textwidth]{fig3.eps}
\hspace{5mm}
\includegraphics[width=0.4\textwidth]{fig4.eps}
\\
\parbox[t]{0.48\textwidth}{\caption{Distribution of the difference between the reconstructed
and true $\tau$ decay length values, line shows result of the fit to Eq.~\ref{resofun}.
Distribution of the residuals for the fit is also shown.}\label{ctaures}}
\parbox[t]{0.48\textwidth}{\caption{Distribution of the experimentally measured decay length
(points with errors), result of the fit is shown by black solid line. Red histogram and solid
line show the MC prediction and parametrisation for the $e^+ e^-\to q\bar{q}$ ($q=$u,~d,~s)
and two-photon backgrounds, blue histogram and solid line show the MC prediction and
parametrisation for the $e^+ e^-\to q\bar{q}$ ($q=$c,~b) background. Distribution of the
residuals for the fit is also shown.}\label{ctaufit}}
\end{figure}
From the fit of experimental data $\tau$ lifetime estimator is obtained to be
$\lambda_{\tau}=86.53\pm 0.16$~$\mu$m, applying MC correction $\Delta_{\rm corr}=0.46$~$\mu$m
the $\tau$ lifetime (multiplied by speed of light) value is $c\tau_{\tau}=86.99\pm 0.16$~$\mu$m.,
where the error is statistical only. The result of the fit is demonstrated in Fig.~\ref{ctaufit}.
The main sources of systematic uncertainties are summarized in Table~\ref{tab:syst}.
\begin{table}[htbp]
\begin{center}
\begin{tabular}{l@{~}c}
\hline
Source & $\Delta c\tau_\tau$ ($\mu$m) \\
\hline
Silicon vertex & \multirow{2}*{0.090} \\
detector alignment & \\
Asymmetry fixing & 0.030 \\
Fit range & 0.020 \\
Beam energy, ISR, FSR & 0.024 \\
Background contribution & 0.010 \\
$\tau$-lepton mass & 0.009 \\
\hline
Total & 0.101 \\
\hline
\end{tabular}
\caption{Systematic uncertainties of $c\tau_\tau$.}\label{tab:syst}
\end{center}
\end{table}
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.5\textwidth]{fig5.eps}
\caption{Summary of $\tau$ lifetime measurements.}\label{ctauresul}
\end{figure}
The obtained results for the product of the lifetime and speed of light and for the lifetime are:
\[ c\tau_{\tau} = (86.99 \pm 0.16({\rm stat}) \pm 0.10({\rm syst}))~\mu{\rm m.} \]
\[ \tau_{\tau} = (290.17 \pm 0.53({\rm stat}) \pm 0.33({\rm syst}))~{\rm fs.} \]
Belle result on $\tau_{\tau}$ and previous measurements are shown in Fig.~\ref{ctauresul}.
For the first time the upper limit on the relative $\tau_\tau$
difference between $\tau^+$ and $\tau^-$ was measured to be:
\[ |\tau_{\tau^+}-\tau_{\tau^-}|/\tau_\mathrm{average}<7.0\times10^{-3}\mathrm{\, \, at\, \, 90\%\,\, CL.} \]
With the new $\tau_\tau$ Belle result the $g_\tau /g_e$ and $g_\tau /g_\mu$
ratios were recalculated to revise lepton universality:
\[ g_\tau/g_e=1.0031\pm 0.0016,~g_\tau/g_\mu=1.0013\pm 0.0016. \]
It is seen that the uncertainty of the ratios was improved by a factor
of about 1.3 in comparison with the last HFAG result \cite{Amhis},
and now the $g_\tau / g_e$ ratio is almost 2 standard deviations away from unity.
\section{Hadronic $\tau$ decays with $K^0_S$ at Belle}
\subsection{Measurement of the branching fractions}
The analysis \cite{Ryu:2014vpc} is based on the data sample with the
luminosity integral of ${\cal L}=669$~fb$^{-1}$ which comprises
615 million $\tau^+\tau^-$ events.
One inclusive decay mode $\tau^-\to K^0_S X^-\nu_{\tau}$
and 6 exclusive hadronic tau decay modes with $K^0_S$
($\tau^-\to\pi^-K^0_S\nu_\tau$, $\tau^-\to K^-K^0_S\nu_\tau$, $\tau^-\to\pi^-K^0_SK^0_S\nu_\tau$,
$\tau^-\to\pi^-K^0_S\pi^0\nu_\tau$, $\tau^-\to K^-K^0_S\pi^0\nu_\tau$, $\tau^-\to\pi^-K^0_SK^0_S\pi^0\nu_\tau$)
were studied.~After the standard Belle $\tau\tau$ preselection criteria \cite{Ryu:2014vpc}
the following selection criteria are applied: the thrust value in the c.m.s.
is greater than 0.9; event is separated into two hemispheres separated by the
plane perpendicular to the thrust axis; in the tag hemisphere one-prong tau decay
is required (decays with $e$, $\mu$ or $\pi/K(n\geq 0)\pi^0$);
in the signal hemisphere particular hadronic final state is required;
$K^0_S$ candidate is reconstructed from a pair of oppositely charged
pions with $\pi^+\pi^-$ invariant mass within the range $0.485$~GeV/$c^2<M_{\pi\pi}<0.511$~GeV/$c^2$,
decay length in the $r-\varphi$ plane satisfying $2$~cm$\leq L_{\rm{\perp}}\leq 20$~cm, and
$z$-distance between the two helices at the $\pi^+\pi^-$ vertex position before the fit
is required to be $\Delta Z_{1,2}<2.5$~cm; $\pi^0$ candidate is reconstructed from
a pair of $\gamma$'s with normalized $\gamma\gamma$ invariant mass satisfying
$-6<(m_{\gamma\gamma}-m_{\pi^{0}})/\sigma_{\gamma\gamma}<5$; charged kaon/pion
identification parameter $\mathcal{P}_{K/\pi}=\mathcal{L}_{K}/(\mathcal{L}_{\pi}+\mathcal{L}_K)$
is required to be $\mathcal{P}_{K/\pi}>0.7(<0.7)$ for kaon (pion); the total energy of extra
photons (not from the signal $\pi^0$ candidates) in laboratory frame should satisfy
$\sum E_{\rm{\gamma}}^{\rm{LAB}}< 0.2$~GeV.
For the inclusive mode $\tau^-\to K^0_S X^-\nu_{\tau}$ the number of
signal events is obtained from the fit of the $\pi^+\pi^-$ invariant
mass distribution of $K^0_S$ candidate.
\begin{table}[htbp]
\begin{center}
\small
\begin{tabular}{@{}c@{~}|@{~}c@{~}|@{~}c@{~}c@{~}c@{~~}c@{~~}c@{~~}c@{}}
\hline
Mode & $K^0_SX^-$ & $\pi^-K^0_S$ & $K^-K^0_S$ & $\pi^-K^0_S\pi^0$ & $K^-K^0_S\pi^0$ & $\pi^-K^0_SK^0_S$ & $\pi^-K^0_SK^0_S\pi^0$ \\
\hline
$N^{\rm data}$ & {\scriptsize $397806\pm 631$} & {\scriptsize $157836\pm 541$} & {\scriptsize $32701\pm 295$} & {\scriptsize $26605\pm 208$} & {\scriptsize $8267\pm 109$} & {\scriptsize $6684\pm 96$} & {\scriptsize $303\pm 33$} \\
$\frac{N^{\rm bg}}{N^{\rm data}}$ (\%) & {\scriptsize $4.20\pm 0.46$} & {\scriptsize $8.86\pm 0.05$} & {\scriptsize $3.55\pm 0.07$} & {\scriptsize $5.60\pm 0.10$} & {\scriptsize $2.43\pm 0.10$} & {\scriptsize $7.89\pm 0.24$} & {\scriptsize $11.6\pm 1.60$} \\
$N^{\rm sig}$,~$10^3$ & {\scriptsize $3947\pm 7$} & {\scriptsize $1793\pm 5$} & {\scriptsize $319.3\pm 1.8$} &
{\scriptsize $833.6\pm 7.0$} & {\scriptsize $322.6\pm 4.5$} & {\scriptsize $244.7\pm 3.3$} & {\scriptsize $21.05\pm 1.40$} \\
$\varepsilon_{\rm det}$ (\%) & 9.66 & 7.09 & 6.69 & 2.65 & 2.19 & 2.47 & 0.82 \\
$\bigl(\frac{\Delta{\cal B}}{{\cal B}}\bigr)_{\rm syst}$ (\%) & 2.4 & 2.5 & 4.0 & 3.9 & 5.2 & 4.4 & 8.1 \\
\hline
\end{tabular}
\normalsize
\caption{Results of the event selection, detection efficiency and
systematic uncertainties of the branching fractions.}\label{tab:ksdata}
\end{center}
\end{table}
Number of selected events ($N^{\rm data}$) as well as
background admixture ($N^{\rm bg}/N^{\rm data}$) for
each decay mode are shown in Table~\ref{tab:ksdata}.
The main non-$\tau\tau$ background comes from $e^+e^-\to q\bar{q}~(q=u,~d,~s,~c)$ events.
To take into account cross-feed background six decay modes are analysed simultaneously
and number of signal events of $i$-th mode is calculated according:
\[ N^{\rm sig}_{i} = \sum_{j} (\mathcal{E}^{-1})_{ij}(N^{\rm data}_{j} - N^{\rm bg}_{j}),~i=1\div 6,\]
where $\mathcal{E}_{ij}$ is $6\times 6$ efficiency matrix (diagonal elements of $\mathcal{E}_{ij}$ are shown
in Table~\ref{tab:ksdata}), $N^{\rm sig}$ numbers for all modes are also summarized in Table~\ref{tab:ksdata}.
For the $\tau^-\to\pi^-K^0_S\nu_\tau$, $\tau^-\to K^-K^0_S\nu_\tau$, $\tau^-\to\pi^-K^0_S\pi^0\nu_\tau$ and
$\tau^-\to K^-K^0_S\pi^0\nu_\tau$ modes with large statistics lepton tag is applied and normalisation
to the two-lepton events $(\tau^{\mp}\to e^{\mp}\nu\nu,~\tau^{\pm}\to\mu^{\pm}\nu\nu)$ (or shortly $e-\mu$)
is used to calculate branching fractions:
\[ {\cal B}_{i} = \frac{N^{\rm sig}_{i}}{N^{\rm sig}_{e-\mu}}\frac{{\cal B}_{e}{\cal B}_{\mu}}{{\cal B}_{e}+{\cal B}_{\mu}}, \]
where $N^{\rm sig}_{e \textrm{-}\mu}$ is the number of $e-\mu$ events, ${\cal B}_\ell$ is the branching fraction of
$\tau^-\to\ell^-\bar{\nu}_\ell\nu_\tau$, $\ell=e,~\mu$.
To increase statistics for the remaining $\tau^-\to\pi^-K^0_SK^0_S\nu_\tau$ and
$\tau^-\to\pi^-K^0_SK^0_S\pi^0\nu_\tau$ modes one-prong decay tag and luminosity normalisation method are used:
\[ {\cal B}_{i} = \frac{N^{\rm sig}_{i}}{2{\cal L}\sigma_{\tau\tau}{\cal B}_{1-\rm prong}},\]
where $\sigma_{\tau\tau}=(0.919\pm 0.003)$~nb \cite{Banerjee:2007is} is the cross section of $\tau^+\tau^-$
production, ${\cal B}_{1-\rm prong}=(85.35 \pm 0.07)\%$ \cite{Beringer:1900zz}
is the one-prong decay branching fraction. The branching fractions for one inclusive and six
exclusive $\tau$ hadronic decay modes with $K^0_S$ measured at Belle are summarized in Table~\ref{tab:brres}.
The branching fractions for $\tau^-\to K^-K^0_S \nu_\tau$ and $\tau^-\to K^-K^0_S\pi^0\nu_\tau$
are measured for the first time at the $B$ factories. The results on branching fractions of
all studied decay modes are consistent with the previous experiments and have better precision.
Upper limits on the branching fractions of two $\tau$ decay modes with two $K^0_S$ obtained
recently at $BABAR$ \cite{Lees:2012de} are shown in Table~\ref{tab:brres} as well.
\begin{table}[htbp]
\begin{center}
\begin{tabular}{@{~}lcl@{~}}
\hline
\hline
Mode & Branching fraction & Reference \\
\hline
$K^0_S X^- \nu_\tau$ & $(9.15 \pm 0.01 \pm 0.15) \times 10^{-3}$ & Belle \\
$\pi^- K^0_S \nu_\tau$ & $(4.16 \pm 0.01 \pm 0.08) \times 10^{-3}$ & Belle \\
$K^- K^0_S \nu_\tau$ & $(7.40 \pm 0.07 \pm 0.27) \times 10^{-4}$ & Belle \\
$\pi^- K^0_S \pi^0 \nu_\tau$ & $(1.93 \pm 0.02 \pm 0.07) \times 10^{-3}$ & Belle \\
$K^- K^0_S \pi^0 \nu_\tau$ & $(7.48 \pm 0.10 \pm 0.37) \times 10^{-4}$ & Belle \\
$\pi^- K^0_S K^0_S \nu_\tau$ & $(2.33 \pm 0.03 \pm 0.09) \times 10^{-4}$ & Belle \\
$\pi^-K^0_SK^0_S\pi^0\nu_\tau$ & $(2.00 \pm 0.22 \pm 0.20) \times 10^{-5}$ & Belle \\
$K^-K^0_SK^0_S\nu_\tau$ & $< 6.3\times 10^{-7}~{\rm at~90\%~CL}$ & $BABAR$ \\
$K^-K^0_SK^0_S\pi^0\nu_{\tau}$ & $< 4.0\times 10^{-7}~{\rm at~90\%~CL}$ & $BABAR$ \\
\hline
\end{tabular}
\caption{Summary of the branching fractions of the $\tau$ decay modes with
$K^0_S$ from Belle \cite{Ryu:2014vpc} and $BABAR$ \cite{Lees:2012de}. The first
uncertainty is statistical and the second one is systematic.}\label{tab:brres}
\end{center}
\end{table}
\subsection{Study of mass spectra in $\tau^-\to\pi^-K^0_SK^0_S\pi^0\nu_\tau$ sample}
The invariant masses of the $K^0_S K^0_S \pi^0$ and $\pi^- K^0_S$ subsystems
for $\tau^-\to\pi^-K^0_SK^0_S\pi^0\nu_\tau$ events are shown in Fig.~\ref{ksdynam}.
Clear peak at the mass of about $1280$~MeV/$c^2$ is seen in the $K^0_S K^0_S \pi^0$
mass distribution as well as some bump around $1420$~MeV/$c^2$, taking into account the
quantum numbers of the hadronic final state observed structures are associated with
the intermediate $f_1(1285)$($J^{PC}=1^{++}$) and $f_1(1420)$($J^{PC}=1^{++}$)
pseudovector mesons. The $\pi^- K^0_S$ distribution exhibits clear signal from the
intermediate $K^{*-}(892)$ vector meson. No additional resonancelike structures
are observed in the other mass spectra.
\begin{figure}[hbtp]
\centering
\includegraphics[width=0.49\textwidth]{fig6a.eps}
\includegraphics[width=0.49\textwidth]{fig6b.eps}
\\
\caption{Invariant mass of the $K^0_S K^0_S \pi^0$ (left) and $\pi^- K^0_S$ (right)
subsystems for the $\tau^-\to\pi^-K^0_SK^0_S\pi^0\nu_\tau$ events. Points with errors
are experimental data, the hatched histogram is background from
$\tau^-\to\pi^- K^0_S K^0_S \nu_\tau$, and the shaded (yellow) histogram is the
continuum $q\bar{q}$ background. Solid line is the result of the fit by the
$\tau^-\to (f_1(1285)\pi^-~+~f_1(1420)\pi^-~+~K^{*-}K^0_S)\,\nu_\tau$
model for signal events. The $(f_1(1285)\pi^-~+~f_1(1420)\pi^-)\,\nu_\tau$
and $K^{*-} K^0_S \nu_\tau$ contributions are shown by the dashed (red)
and dotted (green) lines respectively.}\label{ksdynam}
\end{figure}
To determine the observed contributions simultaneous fit of the $M(K^0_S K^0_S \pi^0)$
and $M(\pi^- K^0_S)$ mass spectra is performed. The result of the fit is also shown
in Fig.~\ref{ksdynam}.
The fractions (significances) of three contributions in $\tau^-\to\pi^-K^0_SK^0_S\pi^0\nu_\tau$
are extracted to be $(34\pm 5)\%$ ($12\sigma$), $(12\pm 3)\%$ ($4.8\sigma$) and $(54\pm 6)\%$ ($7.8\sigma$)
for the $f_1(1285)\pi^-\nu_\tau$, $f_1(1420)\pi^-\nu_\tau$ and $K^{*-}(892)K^0_S\pi^0\nu_\tau$ mechanisms,
respectively. With obtained fractions, products of the branching fractions for the subprocesses
are calculated (the first uncertainty is statistical and the second one is systematic):
\begin{eqnarray*}
{\cal B}(\tau^-\to f_1(1285)\pi^-\nu_\tau)\cdot{\cal B}(f_1(1285)\to K^0_S K^0_S\pi^0) & = & (0.68\pm 0.13\pm 0.07)\times 10^{-5}, \\
{\cal B}(\tau^-\to f_1(1420)\pi^-\nu_\tau)\cdot{\cal B}(f_1(1420)\to K^0_S K^0_S\pi^0) & = & (0.24\pm 0.05\pm 0.06)\times 10^{-5}, \\
{\cal B}(\tau^-\to K^*(892)^-K^0_S\pi^0\nu_\tau)\cdot{\cal B}(K^*(892)^-\to K^0_S\pi^-) & = & (1.08\pm 0.14\pm 0.15)\times 10^{-5}.
\end{eqnarray*}
\section{Search for lepton flavor violating $\tau$ decays at LHCb}
In the analysis the statistics of about 1~fb$^{-1}$ collected by LHCb
at the proton-proton c.m.s. energy $\sqrt{s}=7$~TeV in 2011 is used.
The inclusive $\tau^-$ production cross section at LHCb is
$\sigma_{\rm inc}(\tau^-)=80~\mu$b (80\% of $\tau^-$ come from $D^-_s\to\tau^-\bar{\nu}_\tau$ decay).
LFV decay $\tau^-\to\mu^-\mu^+\mu^-$ and the LNV and BNV ($|\Delta({\rm B-L})|=0$)
decays $\tau^-\to \bar{p}\mu^+\mu^-$ and $\tau^-\to p\mu^-\mu^-$ were studied.
The decay $D^-_s\to\phi\pi^-$ followed by $\phi\to\mu^+\mu^-$
(or shortly $D^-_s\to\phi(\mu^+\mu^-)\pi^-$), which has similar
experimental signature is used for the calibration and normalization.
The following selection criteria are applied:
good tracks with transversal momenta satisfying $p^{\rm 1trk}_{T}>0.3$~GeV/$c$ ;
the transversal momentum of the combined three-track system
is required to be $p^{\rm 3trk}_{T}>4$~GeV/$c$; decay length of
the $\tau$ candidate satisfy $\lambda_\tau >$100~$\mu$m; angle
between 3-track momentum and radius vector of $\tau$ decay
vertex should be small; invariant mass of $\mu^+\mu^-$
satisfy $M_{\mu^+\mu^-}>0.45$~GeV/$c^2$ to suppress background
from the $D^-_s\to\eta(\to\mu^+\mu^-\gamma)\mu^-\bar{\nu}_\mu$ decay.
For the further analysis the region around $\tau$ mass
$(m_{\tau}-20$~MeV/$c^2) <{\cal M}_{\rm inv}<(m_{\tau}+20$~MeV/$c^2)$ ($\pm 2\sigma_M$)
is blinded (so called blinded analysis).
After all selections each $\tau$ candidate is characterized
by a probability to be signal or background according to the
values of three likelihood parameters:
${\cal M}_{\rm 3body}$ includes geometrical properties to
identify displaced 3-body $\tau$ decays, ${\cal M}_{\rm PID}$ is particle
identification based on the information from the detector subsystems: RICH, ECAL, Muon \cite{Alves:2008zz},
${\cal M}_{\rm inv}$ is invariant mass of $\tau$ decay products.
The range for each classifier, ${\cal M}_{\rm 3body}$ (varies from -1 to 1) and ${\cal M}_{\rm PID}$ (varies from 0 to 1),
is subdivided into 6 and 5 unequally spaced bins respectively.
And ${\cal M}_{\rm inv}$ within $\pm 20$~MeV/$c^2$ mass window around $\tau$ mass is subdivided
into 8 equally spaced bins to extract number of signal events.
Result of the fit of the data sidebands for particular bins in ${\cal M}_{\rm 3body}$ and ${\cal M}_{\rm PID}$
for the three decay modes are shown in Fig.~\ref{lfv:fit}.
\begin{figure}[htbp]
\begin{center}
\begin{minipage}[b]{0.32\linewidth}
\centering
\begin{overpic}[width=\textwidth]{fig7a.eps}
\put (20,75) {\scriptsize{\bf(a)}}
\put (35,80) {\tiny{${\cal M}_{\rm 3body} \in [0.65, 1.0]$}}
\put (35,70) {\tiny{${\cal M}_{\rm PID} \in [0.725, 1.0]$}}
\end{overpic}
\end{minipage}
\begin{minipage}[b]{0.32\linewidth}
\centering
\begin{overpic}[width=\textwidth]{fig7b.eps}
\put (20,75) {\scriptsize{\bf(b)}}
\put (35,77) {\tiny{${\cal M}_{\rm 3body} \in [0.40, 1.0]$}}
\end{overpic}
\end{minipage}
\begin{minipage}[b]{0.32\linewidth}
\centering
\begin{overpic}[width=\textwidth]{fig7c.eps}
\put (20,75) {\scriptsize{\bf(c)}}
\put (33,77) {\tiny{${\cal M}_{\rm 3body} \in [0.40, 1.0]$}}
\end{overpic}
\end{minipage}
\caption{Invariant mass distributions and fits
of the mass sidebands in data for:
(a) $\tau^-\to\mu^-\mu^+\mu^-$ candidates,
(b) $\tau^-\to\bar{p}\mu^+\mu^-$ candidates, and
(c) $\tau^-\to p\mu^-\mu^-$ candidates.}\label{lfv:fit}
\end{center}
\end{figure}
The branching fraction of signal decay is calculated according:
\[ {\cal B}_{\rm sig}={\cal B}(D^-_s\to\phi(\mu^+\mu^-)\pi^-)\frac{f^{D_s}_\tau}{{\cal B}(D^-_s\to\tau^-\bar{\nu}_\tau)}\frac{\epsilon^{\rm det}_{\rm norm}\epsilon^{\rm trg}_{\rm norm}}{\epsilon^{\rm det}_{\rm sig}\epsilon^{\rm trg}_{\rm sig}}\frac{N_{\rm sig}}{N_{\rm norm}},\]
where $N_{\rm sig}$($N_{\rm norm}$) is number of selected signal (normalisation) events,
$\epsilon^{\rm trg}_{\rm sig},~\epsilon^{\rm det}_{\rm sig}$ ($\epsilon^{\rm trg}_{\rm norm},~\epsilon^{\rm det}_{\rm norm}$)
are trigger and detection efficiency for signal (normalisation) events, $f^{D_s}_\tau=0.78\pm 0.05$ is fraction of
$\tau$'s from $D^-_s$ decays.
The obtained upper limits on the branching fractions:
\begin{eqnarray*}
{\cal B}(\tau^-\to\mu^-\mu^+\mu^-) & < & 8.0\times 10^{-8}~{\rm at}~90\%~{\rm CL}, \\
{\cal B}(\tau^-\to\bar{p}\mu^+\mu^-) & < & 3.3\times 10^{-7}~{\rm at}~90\%~{\rm CL}, \\
{\cal B}(\tau^-\to p\mu^-\mu^-) & < & 4.4\times 10^{-7}~{\rm at}~90\%~{\rm CL}.
\end{eqnarray*}
Figure~\ref{lfvres} summarizes the upper limits on the branching fractions of 48
LFV modes obtained at $B$ factories. After the LHCb upgrade \cite{Lucchesi:2014gra} it will be possible
to improve the upper limit on ${\cal B}(\tau^-\to\mu^-\mu^+\mu^-)$ in comparison
with the current best limit obtained by Belle \cite{Hayasaka:2010np}.
For the first time the upper limits on the branching fractions of
the LNV and BNV $\tau^-\to \bar{p}\mu^+\mu^-$ and $\tau^-\to p\mu^-\mu^-$
decays were measured at LHCb.
\begin{figure}[hbtp]
\centering
\includegraphics[width=\textwidth]{fig8.eps}
\caption{Summary of the upper limits (90\%CL) on the branching fractions of 48 LFV $\tau$ decays
measured at $B$ factories.~Recent result on ${\cal B}(\tau^-\to\mu^-\mu^+\mu^-)$ from LHCb is shown
by black square.}\label{lfvres}
\end{figure}
\section{CP violation in $\tau$ decays at $B$ factories}
Recent studies of CPV in the $\tau^-\to\pi^- K_S(\geq\pi^0)\nu_{\tau}$ decays
at $BABAR$~\cite{BABAR:2011aa} as well as in the $\tau^-\to K_S\pi^-\nu_{\tau}$ decay
at Belle~\cite{Bischofberger:2011pw} provide complementary information about sources
of CPV in these hadronic decays.
The decay-rate asymmetry $A_{\rm CP}=\frac{\Gamma(\tau^+\to\pi^+ K_S(\geq\pi^0)\nu_{\tau})-\Gamma(\tau^-\to\pi^- K_S(\geq\pi^0)\nu_{\tau})}{\Gamma(\tau^+\to\pi^+ K_S(\geq\pi^0)\nu_{\tau})+\Gamma(\tau^-\to\pi^- K_S(\geq\pi^0)\nu_{\tau})}$
was studied at $BABAR$ with the $\tau^+\tau^-$ data sample of $\int Ldt=476$~fb$^{-1}$.
The obtained result $A_{\rm CP}=(-0.36\pm 0.23\pm 0.11)\%$ is about $2.8$ standard deviations
from the SM expectation $A^{K^0}_{\rm CP}=(+0.36\pm 0.01)\%$.
At Belle CPV search was performed as a blinded analysis based on a $699~{\rm fb}^{-1}$ data sample.
Specially constructed asymmetry, which is a difference between the mean values of $\cos\beta\cos\psi$ for
$\tau^-$ and $\tau^+$ events, was measured in bins of $K^0_S\pi^-$ mass squared ($Q^2=M^2(K^0_S\pi)$):
\[ A_i^{CP}(Q^2_i) = \frac{\int\limits_{\Delta Q^2_i}\cos{\beta} \cos{\psi} \left( \frac{d\Gamma_{\tau^-}}{d\omega} - \frac{d\Gamma_{\tau^+}}{d\omega} \right) d\omega}{\frac{1}{2} \int\limits_{\Delta Q^2_i}\left(\frac{d\Gamma_{\tau^-}}{d\omega}+\frac{d\Gamma_{\tau^+}}{d\omega} \right)d\omega} \simeq \langle \cos\beta\cos\psi \rangle_{\tau^-} - \langle \cos\beta\cos\psi \rangle_{\tau^+}, \]
where $\beta$, $\theta$ and $\psi$ are the angles, evaluated from the measured parameters of the final hadrons,
$d\omega=dQ^2 d\!\cos\theta d\!\cos\beta$. In contrary to the decay-rate asymmetry the introduced
$A_i^{CP}(Q^2_i)$ is already sensitive to the CPV effects from the charged scalar boson exchange \cite{Kuhn:1996dv}.
No CP violation was observed and the upper limit on the CPV
parameter $\eta_S$ was extracted $|{\rm Im}(\eta_S)|<0.026$~at~90\% CL. Using this limit parameters
of the Multi-Higgs-Doublet models \cite{Grossman:1994jb,Choi:1994ch} can be constrained as
$|{\rm Im}(XZ^*)|<0.15~M_{H^\pm}^2/(1\,\mathrm{GeV}^{2}/c^4)$, where
$M_{H^\pm}$ is the mass of the lightest charged Higgs boson, the complex constants Z and X
describe the coupling of the Higgs boson to leptons and quarks respectively.
The other more complicated and most powerful method to extract
CPV parameter at $e^+ e^-$ $B$ factories is an unbinned maximum likelihood
fit of events in the full phase space. The main idea of this method is to consider
events where both taus decay to the particular final states.
One $\tau^{\mp}$ (signal side) decays to particular hadronic final state
(for example $\tau^\mp\to (K\pi)^\mp \nu$) and the other $\tau^{\pm}$
(tag side) decays to some well investigated mode with large branching fraction.
As a tag decay mode we can take $\tau^{\pm}\to\pi^{\pm}\pi^0\nu$
having the largest branching fraction, it also serves as spin analyser,
which allows one to be sensitive to the spin dependent part of the differential
decay width of signal decay using effect of spin-spin correlation of taus \cite{Tsai:1971vv}.
In the technique we analyze $e^+ e^-\to \tau^{\mp}\tau^{\pm}\to ((K\pi)^\mp\nu,~\pi^{\pm}\pi^0\nu)$
(or shortly $((K\pi)^\mp,~\pi^{\pm}\pi^0)$) events in the 12-dimensional
phase space. The probability density function is constructed from the total differential
cross section, which is a sum of spin independent term and spin-spin correlation term.
To write the total differential cross section we follow the approach developed in \cite{Fetscher:1990su,Tamai:2003he}.
The differential cross section of $e^+e^-\to\tau^+(\vec{\zeta}^{*+})\tau^-(\vec{\zeta}^{*-})$
reaction in the center-of-mass system (c.m.s.) is given by formula \cite{Tsai:1971vv}:
\[ \frac{d\sigma(\vec{\zeta}^{*-},\vec{\zeta}^{*+})}{d\Omega} =
\frac{\alpha^2}{64E^2_{\tau}}\beta_{\tau}(D_0+D_{ij}\zeta^{*-}_i \zeta^{*+}_j) \]
\[ D_0 = 1+\cos^2{\theta}+\frac{1}{\gamma^2_{\tau}}\sin^2{\theta} \]
\begin{equation}
D_{ij} = \left( \begin{array}{@{}c@{~~}c@{~~}c@{}}
(1+\frac{1}{\gamma_{\tau}^2})\sin^2{\theta} & 0 & \frac{1}{\gamma_{\tau}}\sin{2\theta} \\
0 & -\beta^2_{\tau}\sin^2{\theta} & 0 \\
\frac{1}{\gamma_{\tau}}\sin{2\theta} & 0 & 1+\cos^2{\theta}-\frac{1}{\gamma^2_{\tau}}\sin^2{\theta} \\
\end{array} \right)
\label{eqtotdif}
\end{equation}
where $\vec{\zeta}^{*\mp}$ is polarisation vector of $\tau^{\mp}$
in the $\tau^{\mp}$ rest frame (unitary vector along $\tau^{\mp}$ spin direction).
Asterisk marks parameters measured in the associated $\tau$ rest
frame. $\alpha$, $E_{\tau}$, $\gamma_{\tau}=E_{\tau}/M_{\tau}$,
$\beta_{\tau}=P_{\tau}/E_{\tau}$ and $\theta$ are fine structure constant, energy, Lorentz
factor, velocity of $\tau$ (in the units of $c$) and polar angle of
$\tau^-$ momentum direction, respectively.
Signal differential decay width is written in the form:
\[\frac{d\Gamma(\tau^\mp(\vec{\zeta}^*)\to (K\pi)^\mp\nu)}{dm^2_{K\pi}d\Omega^*_{K\pi}d\tilde{\Omega}_{\pi}}=\begin{array}{c} (A_0+{\eta_{CP}}A_1)+(\vec{B}_0+{\eta_{CP}}\vec{B}_1)\vec{\zeta}^* \\ (A_0+{\eta^*_{CP}}A_1)-(\vec{B}_0+{\eta^*_{CP}}\vec{B}_1)\vec{\zeta}^* \end{array},\]
where $\eta_{CP}$ is CPV sensitive parameter; $m_{K\pi}$, $\Omega^*_{K\pi}$ are $K\pi$
invariant mass and solid angle of the $K\pi$ system in the $\tau$ rest frame,
$\tilde{\Omega}_\pi$ is solid angle of the charged pion in the $K\pi$ rest frame.
$A_1$ and $A_2$ are form factors of the spin-independent part; $\vec{B}_0$ and $\vec{B}_1$ are
form factors of the spin-dependent part of the differential decay width.
The $\tau^{\pm}(\vec{\zeta'}^*)\to\rho^{\pm}(K^*)\nu(q^*)\to\pi^{\pm}(p^*_1)\pi^0(p^*_2)\nu(q^*)$
decay width reads (with unimportant for this analysis total normalization constant $\kappa_{\rho}$):
\begin{equation}
\frac{d\Gamma(\tau^{\pm}\to\pi^{\pm}\pi^0\nu)}{dm^2_{\pi\pi}d\Omega^*_{\rho}d\tilde{\Omega}_{\pi}}=A'\mp\vec{B}'\vec{\zeta'}^*,
\end{equation}
\[ Q^*=p^*_1-p^*_2,~K^*=p^*_1+p^*_2,~m^2_{\pi\pi}=K^{*2}, \]
\[ A'=\kappa_\rho W(m^2_{\pi\pi})(2(q,Q)Q^*_0-Q^2q^*_0),~\vec{B}'=\kappa_\rho W(m^2_{\pi\pi})(Q^2\vec{K}^*+2(q,Q)\vec{Q}^*),\]
\[ W(m^2_{\pi\pi})=|F_{\pi}(m^2_{\pi\pi})|^2\frac{p^*_{\rho}(m^2_{\pi\pi})\tilde{p}_{\pi}(m^2_{\pi\pi})}{M_{\tau}m_{\pi\pi}},~p^*_{\rho}=\frac{M_{\tau}}{2}\biggl(1-\frac{m^2_{\pi\pi}}{M^2_{\tau}}\biggr),\]
\begin{equation}
\tilde{p}_{\pi}=\frac{\sqrt{(m^2_{\pi\pi}-(m_{\pi}+m_{\pi^0})^2)(m^2_{\pi\pi}-(m_{\pi}-m_{\pi^0})^2)}}{2m_{\pi\pi}},
\end{equation}
where $p^*_{\rho}$, $\Omega^*_{\rho}$ are momentum and solid angle of $\rho$ meson in the $\tau$ rest frame,
$\tilde{p}_{\pi}$, $\tilde{\Omega}_{\pi}$ are momentum and solid angle of charged pion in the $\rho$ rest frame,
$F_{\pi}(m^2_{\pi\pi})$ is pion form factor \cite{Fujikawa:2008ma}.
As a result the total differential cross section for $((K\pi)^\mp,\rho^{\pm})$ events can be written as \cite{Tsai:1971vv}:
\begin{equation}
\frac{d\sigma((K\pi)^{\mp},\rho^{\pm})}{dm^2_{K\pi}d\Omega^*_{K\pi} d\tilde{\Omega}_\pi dm^2_{\pi\pi}d\Omega^*_\rho d\tilde{\Omega}_\pi d\Omega_\tau}=\frac{\alpha^2\beta_{\tau}}{64E^2_{\tau}}\biggl({\cal F}+\begin{array}{c} {\eta_{CP}} \\ {\eta^*_{CP}} \end{array}{\cal G}\biggr),
\end{equation}
\[ {\cal F}=D_0A_0A'-D_{ij}B_{0i}B'_j,~{\cal G}=D_0A_1A'-D_{ij}B_{1i}B'_j.\]
Experimentally we measure particle parameters in the c.m.s.,
hence visible differential cross section is given by:
\scriptsize
\begin{equation}
\frac{d\sigma((K\pi)^{\mp},\rho^{\pm})}{dp_{K\pi}d\Omega_{K\pi}dm^2_{K\pi}d\tilde{\Omega}_{\pi}dp_{\rho}d\Omega_{\rho}dm^2_{\pi\pi}d\tilde{\Omega}_{\pi}}=\sum\limits_{\Phi_1,\Phi_2}\frac{d\sigma((K\pi)^{\mp},\rho^{\pm})}{dm^2_{K\pi}d\Omega^*_{K\pi} d\tilde{\Omega}_\pi dm^2_{\pi\pi}d\Omega^*_\rho d\tilde{\Omega}_\pi d\Omega_\tau}\biggl|\frac{\partial (\Omega^*_{K\pi},\Omega^*_{\rho},\Omega_{\tau})}{\partial (p_{K\pi},\Omega_{K\pi},p_{\rho},\Omega_{\rho})}\biggr|,
\end{equation}
\normalsize
where the summation is done over the unknown $\tau$ direction, which
is determined with two-fold ambiguity by $\Phi_1$ and $\Phi_2$ angles.
They are calculated using parameters measured in the experiment \cite{Tamai:2003he}.
As a result $\eta_{\rm CP}$ can be extracted in the simultaneous unbinned maximum
likelihood fit of the $((K\pi)^-,\rho^+)$ and $((K\pi)^+,\rho^-)$ events in the
12-dimensional phase space. The advantage of this method is that we use the
whole information recorded in the experiment to extract CPV parameter. More
over this, with this method it is possible to study CPV in the spin-dependent
part of the decay width, i.e.~we can test NP models with the nontrivial $\tau$
spin-dependent effects. Similar method is used at $e^+ e^-$ $B$ factories to
measure Michel parameters in leptonic $\tau$ decays \cite{epifanov}.
|
\section{Introduction}\label{Sec:Introduction}
\IEEEPARstart{T}{i}me reversal is an adaptive transmission technique with applications in many areas, such as for instance wireless communication, imaging and sensing~\cite{Bogert-Timereversal-1957,Fink-Basic-1992,Nguyen-2005-Communlett-potential}. A typical time-reversal transmission operates as follows. First, channel sounding signals are emitted from multiple sources. Next, the channel sounding signals are received by a time reversal mirror (TRM) and time reversed. Finally, the time-reversed signals are re-transmitted by the TRM. If the channel from the TRM to the sources is reciprocal, the time-reversed signals retrace the incoming path and focus at the location of the initial source.
Recent research has shown that time-reversal can lead to electromagnetic far-field sub-wavelength focusing, which represents a novel technique to overcome the Rayleigh-diffraction limit~\cite{Fink-2007-Science-Farfield}. This unique mechanism has led to interesting applications in wireless sub-wavelength transmission and far-field super-resolution detection~\cite{Ge-Super-2009,Liao-subwavelength-2009}. So far, time-reversal sub-wavelength focusing has been achieved by transforming evanescent waves into propagating waves via appropriate manipulation of the angular spectrum of the involved electromagnetic waves. This has been done by placing randomly distributed metal scatterers in the near field of the sources~\cite{Fink-2007-Science-Farfield,Ge-Super-2009,Fusco-conjugation-2010}, or electromagnetically large multilayered dielectric structures in front of the TRM~\cite{Liao-subwavelength-2009}. However, the former approach is hardly reproducible~\cite{Ge-Super-2009} while the latter requires very bulky elements.
To overcome these issues, this paper propose a novel approach to achieve time-reversal sub-wavelength transmission. This approach is based on Radio Analog Signal Processing (R-ASP), which consist in processing signals in real-time using purely analog component\cite{Caloz-2013-Analog}. R-ASP technique offers promising benefits in terms of high speed, wide bandwidth, low power consumption, high reproducibility and low cost~\cite{Caloz-2013-Analog}, an represents thus a possible alternative to digital signal processing techniques for ultrafast radio systems. The proposed approach removes the complex and large electromagnetic structure of the aforementioned time-reversal techniques while offering superior performance. This is performed using compact and efficient analog pulse shaping circuits (PSCs)~\cite{Rulikowski-2008-Arbitrary,Khaleghi-2011-Timereversalantenna} to produce channel discrimination. This paper also provides an experimental validation of the technique.
\section{Time-Reversal Transmission with Analog Signal Processing Blocks in the Receivers}\label{Sec:Timereversal}
\subsection{Time-reversal Principle, Issue, and Proposed Solution}\label{SubSec:principle}
A wireless time-reversal system consists of a time-reversal transmitter and of a number of receivers, which may be placed at subwavelength distances from each other. The operation of the system includes two successive phases: a channel calibration phase and a data transmission phase. In the channel calibration phase, all the receivers sequentially send a channel sounding pulse to the transmitter. The transmitter receives all the sounding pulses, flips them in time (time reversal), and records them. These operations constitute the \emph{channel calibration} phase, which is so called because they allow the transmitter to acquire information on the distinct channels existing between itself and the different receivers. The \emph{data transmission} phase, typically using pulse position modulation~\cite{Ge-2011-subwavelegnth}, can then start. The transmitter radiates a wave constituted of the superposition of data signals modulated onto the different sounding signals destined to the corresponding receivers, and each receiver receives its intended signal while discarding the others thanks to the sounding codes.
If the distance between adjacent receivers is much less than the wavelength at the operation frequency, the system suffers from low spatial diversity, which leads to poor discrimination between the received signals. In order to solve this problem, we introduce here pulse shaping circuits (PSCs) with different impulse responses in each of the receivers, as shown in Fig.~\ref{Fig:fig1}. These PCSs increase the channel diversity for each receiver and hence enhance the throughput and reliability of the wireless system.
\begin{figure}
\begin{center}
\psfrag{T}[c][c][0.8]{Time-reversal}
\psfrag{R}[c][c][0.8]{transmitter}
\psfrag{M}[c][c][0.8]{TRM}
\psfrag{a}[c][c][0.8]{$h_{k}^\text{s}(t)$}
\psfrag{z}[c][c][0.8]{$d$}
\psfrag{A}[c][c][0.8]{Antenna-$1$}
\psfrag{B}[c][c][0.8]{Antenna-$2$}
\psfrag{D}[c][c][0.8]{Antenna-$k$}
\psfrag{E}[c][c][0.8]{Antenna-$K$-$1$}
\psfrag{F}[c][c][0.8]{Antenna-$K$}
\psfrag{p}[c][c][0.8]{PSC-$1$}
\psfrag{q}[c][c][0.8]{PSC-$2$}
\psfrag{r}[c][c][0.8]{PSC-$k$}
\psfrag{s}[c][c][0.8]{PSC-$K$-$1$}
\psfrag{t}[c][c][0.8]{PSC-$K$}
\psfrag{I}[c][c][0.8]{Receiver-$1$}
\psfrag{J}[c][c][0.8]{Receiver-$2$}
\psfrag{L}[c][c][0.8]{Receiver-$k$}
\psfrag{O}[c][c][0.8]{Receiver-$K$-$1$}
\psfrag{P}[c][c][0.8]{Receiver-$K$}
\psfrag{b}[c][c][0.7]{$h_1^\text{p}(t)$}
\psfrag{c}[c][c][0.7]{$h_{2}^\text{p}(t)$}
\psfrag{d}[c][c][0.7]{$h_{k}^\text{p}(t)$}
\psfrag{e}[c][c][0.7]{$h_{K-1}^\text{p}(t)$}
\psfrag{f}[c][c][0.7]{$h_{K}^\text{p}(t)$}
\includegraphics[width=1\columnwidth]{fig1.eps}
\caption{Wireless time-reversal system.}
\label{Fig:fig1}
\end{center}
\end{figure}
\subsection{Mathematical Description of the Proposed System}\label{SubSec:math}
\subsubsection{Channel Sounding Phase}\label{SubSec:sounding}
Receiver $k$ generates the channel sounding signal $s_k(t)$, passes it through its PSC, whose impulse response is $h_{k}^\text{p}(t)$ (`p' stands for `pulse'), and sends the resulting signal through its antenna. This radiated signal is
\begin{equation}
\label{eq:xk}
x_{k}(t)=s_k(t)\ast h_{k}^\text{p}(t),
\end{equation}
where `$\ast$' denotes the convolution product.
The time-reversal transmitter receives this signal, that has traveled across the space channel, with impulse response $h_{k}^\text{s}(t)$ (`s' stands for `space'), from receiver $k$ to itself. The received signal is then
\begin{equation}
\label{eq:yk}
y_{k}(t)=x_{k}(t)\ast h_{k}^\text{s}(t)=s(t)\ast h_{k}^\text{p}(t)\ast h_{k}^\text{s}(t),
\end{equation}
where \eqref{eq:xk} has been used in the second equality. The transmitter time-reverses the signal, which thus becomes
\begin{equation}
r_{k}(t)=y_{k}(-t)=s_{k}(-t)\ast h_{k}^\text {p}(-t)\ast h_{k}^\text {s}(-t),
\end{equation}
where \eqref{eq:yk} has been used in the second equality. This operation is sequentially repeated for all the receivers, i.e. for $k=1,2,\ldots,K$.
\subsubsection{Data Transmission Phase}\label{SubSec:data}
For each $k^\text{th}$ receiver, the transmitter pulse position modulates $r_k(t)$ by the corresponding data, which transforms $r_k(t)$ into
\begin{equation}
r_k^\text{m}(t)
=r_k(t-\tau_d)
=r_k(t)\ast\delta(t-\tau_d)
\end{equation}
(`m' stands for `modulation'), where $\tau_d$, $d=1,2,\ldots,D$ is the delay associated with the corresponding $D$-ary symbol of the data stream. The transmitter then sends the modulated signals through its antenna. All the signals, $k=1,2,\ldots,K$, can be transmitted \emph{simultaneously}, as they will be discriminated by the sounding codes, as shall be seen next.
Denoting the channel from the transmitter to the end of receiver $k$, $h_k(t)$, one has
\begin{equation}
h_k(t)=h_k^\text{s}(t)\ast h_k^\text{p}(t).
\end{equation}
The signal received by the $k^\text{th}$ receiver from the modulated signal $r_k^\text{m}(t)$, which is destined to it, is then
\begin{equation}
\label{eq:fkt}
\begin{split}
f_{k}(t)&= r_{k}^\text{m}(t)\ast h_{k}(t)\\
&=\{[s_k(-t)\ast h_{k}^\text {p}(-t)\ast h_{k}^\text{s}(-t)]\ast\delta(t-\tau_{kd})\}\\
&\qquad\qquad\ast [h_{k}^\text{s}(t)\ast h_{k}^\text{p}(t)],
\end{split}
\end{equation}
while the signal received by the $l^\text{th}$ ($l\neq k$) receiver from the modulated signal $r_k^\text{m}(t)$, which is not destined to it, is
\begin{equation}
\label{eq:fkl}
\begin{split}
f_{l}(t)&= r_{k}^\text{m}(t)\ast h_{l}(t)\\
&=\{[s_k(-t)\ast h_{k}^\text {p}(-t)\ast h_{k}^\text{s}(-t)]\ast\delta(t-\tau_{kd})\}\\
&\qquad\qquad\ast [h_{l}^\text{s}(t)\ast h_{l}^\text{p}(t)].
\end{split}
\end{equation}
Using the commutativity and associativity properties of the convolution product, expressions~\eqref{eq:fkt} and~\eqref{eq:fkl} may rewritten as
\begin{subequations}\label{eq:fklt_reform}
\begin{equation}
f_{k}(t)
=s_k(-t)\ast[h_k^\text{s}(-t)\ast h_{k}^\text{s}(t)\ast\delta(t-\tau_{kd})]\ast h_{k}^\text{p}(-t)\ast h_{k}^\text{p}(t),
\end{equation}
and
\begin{equation}
f_{l}(t)
=s_k(-t)\ast h_k^\text{p}(-t)\ast[h_{k}^\text{s}(-t)\ast h_{l}^\text{s}(t)\ast\delta(t-\tau_{kd})]\ast h_{l}^\text{p}(t).
\end{equation}
\end{subequations}
Since the the spatial channel responses $h_{k}^\text{s}(t)$ and $h_{l}^\text{s}(t)$ are highly correlated for a sub-wavelength array \cite{Ding-2013-subwavelegnth}, we can make the following approximation:
\begin{equation}
h_{k}^\text{s}(-t)\ast h_{k}^\text{s}(t)\ast\delta(t-\tau_{kd})\approx h_{k}^\text{s}(-t)\ast h_{l}^\text{s}(t)\ast\delta(t-\tau_{kd})=H(t).
\end{equation}
With this definition of $H(t)$, Eqs.~\eqref{eq:fklt_reform} can be re-written as
\begin{subequations}\label{eq:fkl_fin}
\begin{equation}\label{eq:fk_fin}
f_{k}(t)=z(t)\ast h_{k}^\text{p}(-t)\ast h_{k}^\text{p}(t),
\end{equation}
and
\begin{equation}\label{eq:fl_fin}
f_{l}(t)=z(t)\ast h_{k}^\text{p}(-t)\ast h_{l}^\text{p}(t),
\end{equation}
where
\begin{equation}
z(t)=s_k(-t)\ast H(t).
\end{equation}
\end{subequations}
Equations~\eqref{eq:fk_fin} and~\eqref{eq:fl_fin} reveal that the level of discrimination between $f_{k}(t)$ and $f_{l}(t)$ depends on the degree of dissimilarity between $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$. Maximizing discrimination implies minimizing the correlation between $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$.
Low correlated responses $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$ may be produced by digital signal processing (DSP) or analog signal processing (ASP) PSCs. However, DSP suffers from limited speed and high power consumption due to analog-digital/digital-analog converters. Therefore, we choose here R-ASP technology~\cite{Caloz-2013-Analog,Shulabh-Chipless-2011,Gupta-CSection-MT} to build the PSCs of the receivers.
\section{Channel Discrimination Enhancement \\ using Pulse Shaping}\label{Sec:Resolution}
Figure~\ref{Fig:fig2} shows the block diagram of the PSC of the $k^\text{th}$ receiver. The PSC consists of three blocks: 1)~an input 1:$M$ power splitter, which divides the input pulse into $M$ identical pulses; 2)~a delay block, consisting of $M$ passive non-dispersive delay lines which each delay the $i^\text{th}$ pulse by $\Delta\tau_{ki}$ ($i=1,2,\ldots,M$); 3)~an output $M$:1 power combiner, which adds up the delayed pulses into a single signal. Since the delay block is non-dispersive, the transfer function of the $k^\text{th}$ PSC is simply
\begin{equation}
\label{eq:Hk}
H_{k}^\text{p}(\omega)
=\sum_{i=1}^{M}a_{ki}e^{-j\omega\Delta\tau_{ki}},
\end{equation}
where $a_{ki}$ is the magnitude coefficient of the $i^\text{th}$ channel. Here, all the magnitude coefficients are assumed to be equal, i.e.
\begin{equation}\label{eq:aki}
a_{k1}=a_{k2}=\cdots=a_{ki}=\cdots=a_{kM-1}=a_{kM}=a.
\end{equation}
\begin{figure}
\begin{center}
\psfrag{a}[c][c][0.85]{Input}
\psfrag{b}[c][c][0.85]{~~~Output}
\psfrag{c}[c][c][0.85]{Power}
\psfrag{d}[c][c][0.85]{splitter}
\psfrag{e}[c][c][0.85]{Power}
\psfrag{f}[c][c][0.85]{combiner}
\psfrag{g}[c][c][0.85]{$a_{k1}, \Delta\tau_{k1}$}
\psfrag{h}[c][c][0.85]{$a_{k2}, \Delta\tau_{k2}$}
\psfrag{i}[c][c][0.85]{$a_{k3}, \Delta\tau_{ki}$}
\psfrag{j}[c][c][0.85]{$~a_{k(M-1)}, \Delta\tau_{k(M-1)}$}
\psfrag{k}[c][c][0.85]{$~a_{kM}, \Delta\tau_{kM}$}
\psfrag{M}[c][c][0.85]{Delay block}
\includegraphics[width=1\columnwidth]{fig2.eps}
\caption{$k^\text{th}$ pulse shaping circuit (PSC).}
\label{Fig:fig2}
\end{center}
\end{figure}
Since in addition the input pulse is equally split in $M$ pulses, the impulse response of the PSC is readily obtained as the inverse Fourier transform of~\eqref{eq:Hk} as
\begin{equation}\label{eq:hkpt}
h_{k}^\text{p}(t)=\sum_{i=1}^{M}a\delta(t-\Delta\tau_{ki}).
\end{equation}
In order to facilitate the forthcoming analysis, we arrange the delay lines such that
\begin{equation}\label{eq:tau}
\Delta\tau_{k1}\leq\Delta\tau_{k2}\leq\cdots\leq\Delta\tau_{ki}\leq\Delta\tau_{kM-1}\leq\Delta\tau_{kM}.
\end{equation}
As mentioned in Sec.~\ref{Sec:Timereversal}, the level of discrimination between $f_{k}(t)$ and $f_{l}(t)$ depends on the degree of dissimilarity between $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$. According to~\eqref{eq:fkl_fin}, $f_{k}(t)$ and $f_{l}(t)$ differ only by their respective terms $h_{k}^\text{p}(-t)\ast h_{k}^\text{p}(t)$ and $h_{k}^\text{p}(-t)\ast h_{l}^\text{p}(t)$, and these terms represent therefore the functions to be compared in order to find the conditions for maximal channel discrimination. These two terms are in fact the auto-correlation and cross-correlation functions of $h_{k}^\text{p}(t)$ and between $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$, respectively, which may be noted
\begin{subequations}\label{eq:corr}
\begin{equation}\label{eq:corrhkk}
c_{k,k}(t)=h_{k}^\text{p}(-t)\ast h_{k}^\text{p}(t)
\end{equation}
and
\begin{equation}\label{eq:corrhkl}
c_{k,l}(t)=h_{k}^\text{p}(-t)\ast h_{l}^\text{p}(t).
\end{equation}
\end{subequations}
Substituting~\eqref{eq:hkpt} into~\eqref{eq:corr} and using the even property of the Dirac delta function yields
\begin{subequations}\label{eq:exc}
\begin{equation}\label{eq:exckk}
\begin{split}
c_{k,k}(t)&=\sum_{i=1}^M\sum_{j=1}^{M}a^{2}\delta(t+\Delta\tau_{ki})\ast\delta(t-\Delta\tau_{kj})\\
&=\sum_{i=1}^M\sum_{j=1}^{M}a^{2}\delta(t-\Delta\tau_{kj}+\Delta\tau_{ki})\\
&=Ma^{2}\delta(t)+\sum_{i=1}^M\sum_{j=1,j\neq i}^{M}a^{2}\delta(t-\Delta\tau_{kj}+\Delta\tau_{ki})
\end{split}
\end{equation}
and
\begin{equation}\label{eq:exckl}
\begin{split}
c_{k,l}(t)&=\sum_{i=1}^M\sum_{j=1}^{M}a^{2}\delta(t+\Delta\tau_{ki})\ast\delta(t-\Delta\tau_{lj})\\
&=\sum_{i=1}^M\sum_{j=1}^{M}a^{2}\delta(t-\Delta\tau_{lj}+\Delta\tau_{ki})\\
&=\sum_{i=1}^{M}a^{2}\delta(t-\Delta\tau_{li}+\Delta\tau_{ki})\\
&\qquad+\sum_{i=1}^M\sum_{j=1,j\neq i}^{M}a^{2}\delta(t-\Delta\tau_{lj}+\Delta\tau_{ki}).
\end{split}
\end{equation}
\end{subequations}
By definition of the auto-correlation, $c_{k,k}(t)$ reaches its peak (maximum) at $t=0$.
Consider $c_{k,l}(t)$. If
\begin{equation}\label{eq:taun}
\Delta\tau_{l1}-\Delta\tau_{k1}=\cdots=\Delta\tau_{li}-\Delta\tau_{ki}=\cdots=\Delta\tau_{lM}-\Delta\tau_{kM}=C_{kl},
\end{equation}
where $C_{kl}$ is independent of $i,j$, $\forall k,l$, then
\begin{equation}\label{eq:cklnc}
c_{k,l}(t)=Ma^{2}\delta(t-C_{kl})+\sum_{i=1}^M\sum_{j=1,j\neq i}^{M}a^{2}\delta(t-\Delta\tau_{kj}+\Delta\tau_{ki}).
\end{equation}
In this case, comparison with~\eqref{eq:exckk} indicates that the peak of $c_{k,l}(t)$ will be as high as the peak of $c_{k,k}(t)$\footnote{Only the first term in~{eq:cklnc} and ~\eqref{eq:exckk} contribute to the peak of the function since in the second terms the Dirac delta function terms lead to maxima that are distributed in time.}, which means that $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$ are fully correlated.
If, in contrast,
\begin{equation}\label{eq:taun}
\Delta\tau_{l1}-\Delta\tau_{k1}\neq\cdots\neq\Delta\tau_{li}-\Delta\tau_{ki}\neq\cdots\neq\Delta\tau_{lM}-\Delta\tau_{kM},
\end{equation}
the maxima in the first term of~\eqref{eq:exckl} are completely spread out in time (all the terms contribute at different times), then this first term is much smaller than the first term in~\eqref{eq:exckk} and therefore $h_{k}^\text{p}(t)$ and $h_{l}^\text{p}(t)$ are minimally correlated.
Thus, maximal channel discrimination is achieved by designing the PSC such that condition~\eqref{eq:taun} is satisfied.
\section{System Benefits and Features}\label{sec:benefits}
The proposed PSC time reversal enhancement technique offers several compared to other techniques:
\begin{enumerate}
\item Compared with DSP techniques, its R-ASP nature provides larger bandwidth, faster response and lower cost.
\item In contrast to other time-reversal approaches, it does not require an electrically large time-reversal transmitter (also called TRM), since channel diversity is provided by pulse shaping instead of TRM aperture.
\item In contrast to other time-reversal approaches, it does not require complex electromagnetic structures (e.g. metallic ``brush-like'' structure in~\cite{Fink-2007-Science-Farfield} or multilayer radome~\cite{Liao-subwavelength-2009}), but simple and integrable components (the PSCs) which in addition provide large degrees of flexibilities to handle large numbers of channels.
\end{enumerate}
\section{Experiment Demonstration}\label{Sec:experiment}
Figure~\ref{fig:loop1} shows the schematic experimental setup, corresponding to~Fig.~\ref{Fig:fig1} for the case $K=2$. The experiments are carried out in an indoor environment. A modulated Gaussian pulse with $5$~GHz central frequency and $3$~GHz bandwidth is selected as the channel sounding pulse. The time-reversal transmitter uses an arbitrary waveform generator (Tektronix AWG7122B) to emulate the time-reversed sounding pulses and a $3.1-10.6$~GHz UWB antenna for transmitting them. The two receivers receive the signal using UWB antennas, shown in Fig.~\ref{fig:loop2} (similar to that of the transmitter), incorporate the PSCs shown in Fig.~\ref{fig:loop3} for pulse shaping, and are connected to a digital serial analyzer (Tektronix DSA72004B) for data analysis. The distance between the transmit and receive antennas is $10$~mm, which is far less than the wavelength of the operation: $f=6.5$ to $3.5$~GHz $\rightarrow$ $\lambda_0=46$ to $86$~mm $\rightarrow$ $d=\lambda_0/46$ to $\lambda_0/86$.
\begin{figure}
\begin{center}
\subfigure[]{\label{fig:loop1}
\psfrag{b}[l][c][0.75]{Time-reversal}
\psfrag{a}[l][c][0.75]{transmitter}
\psfrag{c}[c][c][0.75]{Antenna-$1$}
\psfrag{d}[c][c][0.75]{Antenna-$2$}
\psfrag{s}[l][c][0.75]{$d\leq \lambda_0$}
\psfrag{e}[l][c][0.75]{Receiving antennas}
\psfrag{i}[c][c][0.75]{PSC-$1$}
\psfrag{f}[c][c][0.75]{PSC-$2$}
\psfrag{g}[c][c][0.75]{Filter}
\psfrag{h}[c][c][0.75]{Filter}
\psfrag{N}[l][c][0.7]{LNA}
\psfrag{L}[l][c][0.7]{LNA}
\psfrag{D}[c][c][0.75]{DSA}
\psfrag{S}[c][c][0.75]{72004B}
\psfrag{A}[c][c][0.75]{AWG}
\psfrag{W}[c][c][0.75]{7122B}
\psfrag{C}[c][c][0.75]{Computer}
\psfrag{I}[c][c][0.75]{Indoor environment}
\psfrag{P}[c][c][0.75]{PA}
\includegraphics[width=1\columnwidth]{fig4a.eps}}
\subfigure[]{\label{fig:loop2}
\psfrag{a}[l][c][1]{\textcolor[rgb]{1.00,0.00,0.00}{$d\ll \lambda_0$}}
\psfrag{b}[l][c][1]{Antenna-$1$}
\psfrag{c}[l][c][1]{Antenna-$2$}
\includegraphics[width=0.7\columnwidth]{fig4b.eps}}
\subfigure[]{\label{fig:loop3}
\psfrag{a}[c][c][0.8]{Input block}
\psfrag{b}[c][c][0.8]{Output block}
\psfrag{c}[c][c][0.8]{Delay block}
\includegraphics[width=0.8\columnwidth]{fig3c.eps}}
\caption{Experiment. a) Schematic of the setup. b) Receiving antennas. c) One of the two pulse shaping circuits (PSCs).}\label{Fig:setup}
\end{center}
\end{figure}
The PSCs [Fig.~\ref{fig:loop3}] are built as follows. An input $1:4$ power splitter (DC to $8$~GHz operation frequency range) splits the input pulse into four identical pulses. The same device is used as the output $4$:$1$ power combiner to sum up the delayed pulses. The delay block consists of $4$ delay lines of different lengths selected so as to maximize channel discrimination according to~\eqref{eq:taun}. The corresponding parameters are listed in Tab.~\ref{Tab:Para}.
\begin{table}[t!]
\caption{Parameters of the PSCs}
\label{Tab:Para}
\centering
\begin{tabular}{c|c}
\hline
\bfseries PSC-$1$ & \bfseries PSC-$2$\\
\hline
Delay $\Delta \tau_{\text{1}i}$ & Delay $\Delta \tau_{\text{2}i}$ \\
\hline
0.9 ns & 0.3 ns\\
\hline
0.9 ns & 0.6 ns\\
\hline
0.9 ns & 0.9 ns\\
\hline
0.9 ns & 1.2 ns\\
\hline
\end{tabular}
\end{table}
The same equivalent channel sounding method, based on channel reciprocity, as in~\cite{Ge-2011-subwavelegnth,Ding-2013-subwavelegnth}, is used in the calibration phase, where the sounding signals are actually generated by the transmitter and measured by the receivers, for simplicity. The experimental procedure is as follows:
\begin{enumerate}
\item generate the channel sounding pulse, using the arbitrary wave generator, and transmit it with the time-reversal transmitter;
\item record the signals received by the receiving antennas using the digital serial analyzer, flip them and numerically modulate them using a computer;
\item transmit the modulated time-reversed signals from the time-reversal transmitter and record the signals received by both the target antenna and the non-target antenna.
\end{enumerate}
The experimental results are shown in Fig.~\ref{Fig:Case2}. Thanks to the PSCs, the waveform of the signal received by the target receiver is much higher and essentially identical to the transmitted one (not shown) while the signal received by the non-target receiver has a totally different waveform and is spread out in time with much lower temporal power density.
\begin{figure}
\begin{center}
\subfigure[]{\label{fig:a11}
\psfrag{A}[l][l][0.85]{Antenna-$1$}
\includegraphics[width=0.48\columnwidth]{fig5a.eps}}
\subfigure[]{\label{fig:a21}
\psfrag{A}[l][l][0.85]{Antenna-$2$}
\includegraphics[width=0.48\columnwidth]{fig5b.eps}}
\subfigure[]{\label{fig:a11}
\psfrag{A}[l][l][0.85]{Antenna-$1$}
\includegraphics[width=0.48\columnwidth]{fig6a.eps}}
\subfigure[]{\label{fig:a11}
\psfrag{A}[l][l][0.85]{Antenna-$2$}
\includegraphics[width=0.48\columnwidth]{fig6b.eps}}
\caption{Signals received by the antennas (experiment). a) and b) denote the case when antenna-$1$ is set to be the target antenna. c) and d) denote the case when antenna-$2$ is set to be the target} \label{Fig:Case2}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\subfigure[]{\label{fig:com1}
\psfrag{A}[l][l][0.85]{Antenna-$1$}
\includegraphics[width=0.48\columnwidth]{fig7a.eps}}
\subfigure[]{\label{fig:com2}
\psfrag{A}[l][l][0.85]{Antenna-$2$}
\includegraphics[width=0.48\columnwidth]{fig7b.eps}}
\caption{Same experiment as in Fig.~\ref{fig:a11} and ~\ref{fig:a21} but without pulse shaping circuits. a) Signal received by Antenna-1. b) Signal received by Antenna-2.} \label{Fig:compare}
\end{center}
\end{figure}
For comparison, the same experiment is carried out without PSCs. The results are shown in Fig.~\ref{Fig:compare}. In this case, as expected, the waveform of the signal received by the target receiver has lost its target features and is comparable to that received by the non-target receiver.
\section{Conclusions}\label{Sec:conclusion}
A novel approach to enhance time-reversal subwavelength transmission based on R-ASP PSCs is proposed, theoretically derived and experimentally validated. After a mathematical demonstration of the system, culminating with the determination of PSC conditions, experimental results are provided for validation, using two PSCs with different impulse responses to encode the channels between the transmitter and two receivers with sub-wavelength separation. The signals received by the antennas exhibit clear discrimination. Moreover, to the inherent analog nature of the PSCs, the time-reversal sub-wavelength transmission system is highly efficient and flexible. It may find applications in compact MIMO and higher resolution imaging systems.
\bibliographystyle{IEEEtran}
|
\section{A two-sided stochastic calculus}
\label{calculus}
We are neither competent nor it is our purpose to explain the stochastic calculus in this preliminary section. Several textbook and monographs are devoted to this topic, the author used a book of McKean \cite{McKean} and the relevant chapters in a book of Simon \cite{Simon}. We do not even aim to explain the two sided integral constructed by Pardoux and Protter \cite{Pardoux}, the reader should consult their article for details and proofs. We merely repeat what is relevant for our exposition and we gather several lemmas that we shall need for the proofs.
We consider a one dimensional Brownian motion $B_t,\, 0 \leq t \leq 1$ and the associated sigma algebra $\sigma(B_s,\,0 \leq s \leq t)$. For a continuous process $X_t$ adapted to the sigma algebra the forward \^{I}to integral of $X_t$ is defined as
$$
\int_0^s X_t {\rm d} B_t = \lim_{N \to \infty} \sum_{k=1}^{2^N} X_{\frac{k-1}{2^N}s}(B_{\frac{k}{2^N}s} - B_{\frac{k-1}{2^N}s}).
$$
It is an important part of the definition that the increment points to the future and hence $B_{\frac{k}{2^N}s} - B_{\frac{k-1}{2^N}s}$ and $X_{\frac{k-1}{2^N}s}$ are independent random variables. A consequence of this choice is that the integral, as a function of $s$, is a martingale and two basic formulas, cf. (\ref{itometry}),
$$
\mathbb{E}[\int_0^s X_t {\rm d} B_t] = 0, \quad \mathbb{E}[\left( \int_0^s X_t {\rm d} B_t \right)^2] = \int_0^s \mathbb{E}[X_t^2] {\rm d} t,
$$
hold true.
Backward \^{I}to integral is defined in an analogous manner. We consider a process $Y_t$ adapted to a sigma algebra $\sigma(B_s - B_1,\, t \leq s \leq 1)$ and we define the integral of $Y_t$ by
$$
\int_s^1 Y_t {\rm d} B_t = \lim_{N \to \infty} \sum_{k=1}^{2^N} Y_{s + (1-s) \frac{k}{2^N}} (B_{s + (1-s) \frac{k}{2^N}} - B_{s + (1-s) \frac{k-1}{2^N}}).
$$
Note that here the increments point to the future in order to ensure the independence with the integrand. The backward \^{I}to integral is a backward martingale as a function of $s$ and formulas corresponding to (\ref{itometry}) hold true,
$$
\mathbb{E}[\int_s^1 Y_t {\rm d} B_t] = 0, \quad \mathbb{E}[\left( \int_s^1 Y_t {\rm d} B_t \right)^2] = \int_s^1 \mathbb{E}[Y_t^2] {\rm d} t.
$$
We now consider particular processes $X_t,\,Y_t$ that arise as solutions of forward/backward stochastic differential equations,
\begin{align*}
X_t &= X_0 + \int_0^s b(X_t) {\rm d} t + \int_0^s \sigma(X_t) {\rm d} B_t ,\\
Y_t &= Y_0 + \int_s^1 c(Y_t) {\rm d} t + \int_s^1 \gamma(Y_t) {\rm d} B_t, \\
\end{align*}
for some continuous functions $b,\,c,\,\sigma,\,\gamma$. The integral in the first equation being the forward \^{I}to integral, while the integral in the second equation being the backward \^{I}to integral. Correspondingly the first equation has a unique solution $X_t$ that is a non-anticipatory martingale and the second equation has a unique solution $Y_t$ that is a backward martingale adapted to the associated sigma algebra.
We also use a differential form of these equations
$$
{\rm d} X_t = b(X_t) {\rm d} t + \sigma(X_t) {\rm d} B_t, \quad {\rm d} Y_t = -c(Y_t) {\rm d} t - \gamma(Y_t) {\rm d} B_t.
$$
Although the notation makes no distinction between the forward and the backward case one should keep in mind that these are distinct differentials.
A stochastic integral for joint functions of $X_t,\,Y_t$ was constructed in \cite{Pardoux}. Let $f(t,\,X_t,\,Y_t)$ be a continuous function of its arguments, then an integral
$$
\int_{s'}^s f(t,X_t,Y_t) {\rm d} B_t
$$
can be defined in such a way that if $f$ is independent of $Y_t$ (resp. $X_t$) then the integral coincides with the forward (resp. backward) \^{I}to integral. Furthermore the integral satisfies the following chain rule,
\begin{multline}
\label{chainrule}
f(s,\,X_s,\,Y_s) = f(s',\,X_{s'},\,Y_{s'}) + \int_{s'}^s \partial_t f(t,X_t,Y_t) {\rm d} t \\ + \int_{s'}^s \partial_{X} f(t,X_t,Y_t) {\rm d} X_t + \int_{s'}^s \partial_{Y} f(t,X_t,Y_t) {\rm d} Y_t \\ + \frac{1}{2} \int_s^{s'} \partial_{XX} f(t,X_t,Y_t) ({\rm d} X_t)^2 - \frac{1}{2} \int_s^{s'} \partial_{YY} f(t,X_t,Y_t) ({\rm d} Y_t)^2,
\end{multline}
where $(d X_t)^2 = \sigma^2(X_t) {\rm d} t$ is interpreted according to the \^{I}to rules $({\rm d} t)^2 = {\rm d} t {\rm d} B_t = 0$, $({\rm d} B_t)^2 = {\rm d} t$.
In the following paragraphs we employ the formula~(\ref{chainrule}) in a case with no second order derivatives to operator valued processes $X_t,\,Y_t$. Due to the simplicity of that case the operator valued extension is clear. The operator valued version is discussed in more details in \cite{Pardoux}.
To demonstrate the power of the calculus we show that equations (\ref{forward}, \ref{backward}) define the same propagator (the value of $\varepsilon$ is not important for the following considerations and we skip the index) and that this propagator satisfies the semigroup property
$$
U(s,0) = U(s,s') U(s',0).
$$
To this end we fix a time $t$ and let $X_s = U(s,0)$ be a solution of Eq.~(\ref{forward}) and $Y_s= \tilde{U}(t,s)$ be a solution of Eq.~(\ref{backward}). Then the above chain rule implies that for any $t \leq s \leq s'$ we have
$$
\tilde{U}(t,s) U(s,0) = \tilde{U}(t,s') U(s',0).
$$
By choosing $s = t$ and $s'=0$ we get the sought equivalence $\tilde{U}(t,0) = U(t,0)$. Upon erasing the tilde in the above equation we then establish the semigroup property.
In the following we will need two specific results concerning stochastic differential equations. The first is a particular version of the Duhamel formula, the second is a prior bound on stochastic integrals. We formulate the bound for the forward integral, the corresponding bound holds also for the backward integral.
\begin{lem}[Duhamel formula]
\label{duhamel}
The solution of the differential Eq.~(\ref{forward}) satisfies a relation
$$
U(s,\,s') = V(s,s') + \int_{s'}^s U(s,t) L_2(t) V(t,s') {\rm d} B_t,
$$
where $V(s,s')$ is the solution of a deterministic equation,
$$
{\rm d} V(s,s') = L_1(s) V(s,s') {\rm d} s, \quad V(s',s') = \mathbbm{1}.
$$
\end{lem}
{\bf Proof:} The proof is again an application of the chain rule (\ref{chainrule}). Pick $X_t = V(t,s'),\, Y_t = U(s,t)$ then for any $s \geq t \geq t' \geq s'$ the chain rule gives
$$
U(s,t) V(t,s') = U(s,t') V(t',s') - \int_{t'}^t U(s,x) L_2(x) V(x,s') {\rm d} B_x.
$$
The statement of the lemma then follows by choosing $t = s$ and $t' = s'$. \hfill $\square$
\begin{lem}[Prior estimates of stochastic integrals]
\label{bound}
Let $X_t \in \mathcal{H}$ be a non-anticipatory stochastic process, then the following estimates hold true:
\begin{enumerate}
\item[(a)]
$$
\mathbb{E}[ ||\int_0^1 X(s) {\rm d} B_s ||^{2n}] \leq (2n^2 - n)^n \mathbb{E}[\int_0^1 ||X_s||^{2n} {\rm d} s], \quad n \geq 1.
$$
\item[(b)]
\begin{equation}
\label{expbound}
Prob( || \int_0^1 X_s {\rm d} B_s ||^2 > \gamma) \leq e^{-\gamma \frac{1}{8 ||X||_\infty^2} + \frac{1}{4}},
\end{equation}
where $||X||_\infty := \sup_{0 \leq t \leq 1} ||X_t||_\infty$.
\end{enumerate}
\end{lem}
{\bf Proof:}
Denote $\Theta_t = \int_0^t X_s {\rm d} B_s$ and consider a real valued stochastic process $\zeta_t = (\Theta_t, \Theta_t)$. The stochastic differentiation of this process gives ${\rm d} \zeta_t = ((X_t, \Theta_t) + (\Theta_t, X_t)) {\rm d} B_t+ (X_t,X_t) {\rm d} t$, which is equivalent to an integral relation
\begin{equation}
\label{label}
\zeta_t - \int_0^t ||X_s||^2 {\rm d} s = \int_0^t ((X_s, \Theta_s) + (\Theta_s, X_s)) {\rm d} B_s.
\end{equation}
(a) We take the expectation of $d\zeta_t^n = n \zeta_t^{n-1} {\rm d} \zeta_t + 1/2n(n-1) \zeta_t^{n-2} {\rm d} \zeta_t {\rm d} \zeta_t$ to get an integral relation,
$$
\mathbb{E}[\zeta_t^n] = n \int_0^t \mathbb{E}[\zeta^{n-1}_s ||X_s||^2] {\rm d} s + \frac{n(n-1)}{2} \int_0^t \mathbb{E}[ \zeta_s^{n-2} ( (\Theta_s, \,X_s) + (X_s,\,\Theta_s))^2] {\rm d} s,
$$
between the moments. As a first observation note that all integrands are non-negative functions and hence $\mathbb{E}[\zeta_t^n]$ is a non-decreasing function of $t$. Now we employ the estimate $((X_s, \Theta_s) + (\Theta_s, X_s))^2\leq 4 ||X_s||^2 \zeta_s$ and the H\"{o}lder inequality to get
\begin{align*}
\mathbb{E}[\zeta_t^n] &\leq (n + 2n(n-1)) \int_0^t \mathbb{E}[ \zeta_s^{n-1} || X_s||^2] {\rm d} s \\
& \leq (2n^2 -n) \left( \int_0^t \mathbb{E}[\zeta_s^n] {\rm d} s \right)^\frac{n-1}{n} \left( \int_0^t ||X_s||^{2n} {\rm d} s \right)^{\frac{1}{n}} \\
&\leq (2n^2 -n) \left( \mathbb{E}[\zeta_t^n] {\rm d} s \right)^\frac{n-1}{n} \left( \int_0^t ||X_s||^{2n} {\rm d} s \right)^{\frac{1}{n}} .
\end{align*}
In the last inequality we also used $0 \leq t \leq 1$. Solving for the $n$-th moment establishes the first inequality of the lemma.
To prove (b) we will use a well-known prior estimate on stochastic integrals \cite[Chapter 2.3]{McKean}. Let $e_t, 0 \leq t \leq 1$ be a real non-anticipatory function and suppose that $\int_0^1 e_t^2 {\rm d} t < \infty$ then for any reals $\beta$ and $\alpha$ the following bound holds true,
\begin{equation}
\label{martin}
Prob\left[ \max_{0 \leq t \leq 1} ( \int_0^t e_s {\rm d} B_s - \frac{\alpha}{2} \int_0^t e_s^2 {\rm d} s ) > \beta \right] \leq e^{-\alpha \beta}.
\end{equation}
We note that the bound is an application of Doob's martingale inequality.
Applying the bound to Eq.(\ref{label}) then implies
$$
Prob\left[ \max_{0 \leq t \leq 1} \left( \zeta_t - \int_0^t ||X_s||^2 {\rm d} s - \frac{\alpha}{2} \int_0^t ((X_s, \Theta_s) + (\Theta_s, X_s))^2 {\rm d} s \right)> \beta \right] \leq e^{-\alpha \beta}.
$$
We claim that for $0 \leq t \leq 1$,
\begin{equation}
\label{stam}
\zeta_t - \int_0^t ||X_s||^2 {\rm d} s - \frac{\alpha}{2} \int_0^t ((X_s, \Theta_s) + (\Theta_s, X_s))^2 {\rm d} s \geq \zeta_t - ||X||_\infty^2 - 2 \alpha ||X||_\infty^2 \max_{0 \leq s \leq 1} \zeta_s.
\end{equation}
In particular whenever $\max RHS > \beta$ then also $\max LHS > \beta$ and the probability of an event $\max RHS > \beta$ is smaller then the probability of an event $\max LHS > \beta$. Combining this with the probability bound above we have
\begin{multline*}
Prob\left[ \max_{0 \leq t \leq 1} \zeta_t - ||X||_\infty^2 - 2 \alpha ||X_\infty||^2 \max_{0 \leq t \leq 1} \zeta_t> \beta \right] \\
= Prob \left[ \max_{0 \leq t \leq 1} \zeta_t > \frac{\beta + ||X||_\infty^2}{1 - 2 \alpha ||X||_\infty^2} \right] \leq e^{- \alpha \beta}.
\end{multline*}
Writing $\gamma = (\beta + ||X||_\infty^2)/(1 - 2 \alpha ||X||_\infty^2)$ and choosing the optimal $\alpha = 1/(4 ||X||_\infty^2)$ we get Bound~(\ref{expbound}).
It remains to prove Eq.~(\ref{stam}). The inequality follows from the inequality $ ||X_s||^2 \leq ||X||^2_\infty $ and the inequality
$$
((X_s, \Theta_s) + (\Theta_s, X_s))^2\leq 4 ||X||^2_\infty ||\Theta_s||^2 \leq 4 ||X||^2_\infty \max_{0 \leq s \leq 1} \zeta_s.
$$
Note that for $0 \leq t \leq 1$ an integral $\int_0^t$ of a positive constant can be bounded by that constant.
\hfill $\square$
An important consequence of the lemma is that
\begin{equation}
\label{Ocalculus}
\int_0^1 O_t(\varepsilon^n) {\rm d} B_t = O(\varepsilon^n),
\end{equation}
provided the moments of $||O_t(\varepsilon^n)||$ are uniformly bounded with respect to $t$.
Generalization of Lemma~\ref{bound}. is one of the main technical obstacles of a Banach space version of the theory. For finite dimensional spaces all norms are equivalent and the above bounds hold true up to a dimension dependent constant. On the other hand we do not know if such bounds are available in the infinite dimensional Banach spaces.
The adiabatic expansion, which is the main result of our paper, has a natural formulation in terms of the backward \^{I}to integral. On the other hand it is often easier -- not principally, just thanks to a larger degree of familiarity -- to perform calculations with a forward \^{I}to integral. Due to a special structure of integrals that appear in this work we can always convert a backward integral to a forward integral.
\begin{rem}
\label{conversion}
The backward stochastic integrals of the type $\int_0^s U(s,\,s') f(s') {\rm d} B_{s'}$, where $f$ is a deterministic function, can be converted into a forward integral thanks to the semigroup relation $U(s,\,s') =U(s,0) U(s',0)^{-1}$. The relation expresses the propagator in the future by a constant (with respect to the integration) times a propagator in the past. We still need to convert the backward to a forward integral.
To see in details how the conversion works we take a second look at the forward \^{I}to integral that we defined by
$$
I_- = \lim_{N \to \infty} \sum_{k=1}^{2^N} X_{\frac{k-1}{2^N}s}(B_{\frac{k}{2^N}s} - B_{\frac{k-1}{2^N}s}).
$$
Alternatively one can define\footnote{This is sometimes referred to as a backward integral, we do not use this name to avoid confusion.}
$$
I_+ = \lim_{N \to \infty} \sum_{k=1}^{2^N} X_{\frac{k}{2^N}s}(B_{\frac{k}{2^N}s} - B_{\frac{k-1}{2^N}s})
$$
and these two definitions are related by a quadratic variation of $X_t$,
$$
I_+ - I_- = \int_0^s {\rm d} X_t {\rm d} B_t = \int_0^s \sigma(X_t) {\rm d} t.
$$
For the integral under consideration this now implies --- we introduce back $\varepsilon$ as this will be useful at a later point in the article ---
\begin{multline}
\label{eq:6}
\int_0^s U_\varepsilon(s,\,s') f(s') {\rm d} B_{s'} - U_\varepsilon (s,0) \int_0^s U_\varepsilon(s',0)^{-1} f(s') {\rm d} B_{s'} \\ = - \frac{1}{\sqrt{\varepsilon}}U_\varepsilon (s,0) \int_0^s U_\varepsilon(s',0)^{-1} L_2(s') f(s') {\rm d} s',
\end{multline}
where we have used $\varepsilon {\rm d} U(s ,0)^{-1} = U(s ,0)^{-1} (-{\rm d} L(s) +L_2^2(s) {\rm d} s)$. The second line seems to diverge as $\varepsilon \to 0$, but in fact it is of order $1$, and provided all the inverse operators exist on the range of $L_2$ we have:
\begin{equation}
\label{eq:7}
\int_0^s U_\varepsilon(s,\,s') f(s') {\rm d} B_{s'} = U_\varepsilon (s,0) \int_0^s U_\varepsilon(s',0)^{-1} \tilde{f}(s') {\rm d} B_{s'} + O(\sqrt{\varepsilon}),
\end{equation}
with $\tilde{f}(s) = [1 + L_2(s) (L_1(s) - L_2^2(s))^{-1} L_2(s)] f(s)$. To see this we use
\begin{multline*}
\int_0^s U_\varepsilon(s',0)^{-1} L_2(s') f(s') {\rm d} s' = \varepsilon \int_0^s {\rm d} U_\varepsilon(s',0)^{-1} (-L_1(s') + L_2^2(s'))^{-1} L_2(s') f(s') \\
+ \sqrt{\varepsilon} \int_0^s U_\varepsilon(s',0)^{-1} L_2(s') (-L_1(s') + L_2^2(s'))^{-1} L_2(s') f(s') {\rm d} B_{s'}.
\end{multline*}
An integration by parts shows that the first line of the RHS of the formula is of order $\varepsilon$ and after plugging it into Eq.~(\ref{eq:6}) we obtain Eq.~(\ref{eq:7}).
\end{rem}
\section{Assumptions and basic results}
\label{basic}
We derive a solution of Eq.~(\ref{eq:1}) in the adiabatic limit $\varepsilon \to 0$ under three additional assumptions.
\begin{ass}
\begin{enumerate}
\item[(A)] For each $s$, $L_1(s) - 1/2 L_2^2(s)$ generates a contraction semigroup and $i L_2(s)$ is self-adjoint,
\item[(B)] $L_1(s), s \in (0,\,1)$ is a family of operators for which $0$ remains a uniformly isolated discrete eigenvalue,
\item[(C)] $\ker L_2(s) \supseteq \ker L_1(s),\,s\in (0,1) $.
\end{enumerate}
\end{ass}
Condition (A) is a sufficient and necessary condition of a stochastic version of Hille-Yosida theorem; It implies that $U_\varepsilon(s,\,s')$ is a contraction, i.e. $||U_\varepsilon(s,\,s')|| \leq1$. This prevents an exponential blow up of solutions, and it is a standard in the adiabatic theory \cite{Joye, AFGG}.
\begin{prop}[stochastic Hille-Yosida]
\label{shy}
Let $U(s,\,s')$ be a propagator associated to a stochastic differential equation
$$
{\rm d} U(s,s') = L_1 U(s,\,s') {\rm d} s + L_2 U(s,\,s') {\rm d} B_s, \quad U(s',s') = \mathbbm{1}.
$$
Then the following are equivalent
\begin{enumerate}
\item[(i)] $U(s,\,s')$ is a contraction, i.e. $||U(s,\,s')|| \leq 1$,
\item[(ii)] $L_1 - 1/2 L_2^2$ is a generator of a contraction semigroup and $i L_2$ is self-adjoint.
\end{enumerate}
Furthermore (ii) implies that $L_1$ is a generator of contractions.
\end{prop}
{\bf Proof:}
Without loss of generality we put $s' = 0$, and throughout the proof we denote $x(s) = U(s,\,0) x(0)$. The condition that $U(s,\,0)$ is a contraction is then equivalent to the statement that $||x(s)|| \leq ||x(0)||$ for all initial vectors $x(0)$.
(ii) $\implies$ (i): By \^{I}to rules we have
\begin{align*}
{\rm d} ||x(s)||^2 &= [ (L_1 x(s),\, x(s)) + (x(s),\, L_1 x(s)) + (L_2 x(s),\, L_2 x(s)) ] {\rm d} s \\
& \quad + [ (L_2 x(s), x(s)) + (x(s), L_2 x(s)) ] {\rm d} B_s \\
&= ((2 \mathrm{Re} L_1 - L_2^2)x(s), x(s)) {\rm d} s,
\end{align*}
where the last line is due to the assumption $L_2^* = - L_2$. Recall that $L_1 - 1/2 L_2^2$ is a generator of contraction on a Hilbert space if and only if it is dissipative, i.e. $\mathrm{Re}(L_1 - 1/2 L_2^2) \leq 0$. It then follows that ${\rm d} ||x(s)||^2 \leq 0$.
(i) $\implies$ (ii): We first prove that $L_2$ generates isometries, by proving that both $L_2$ and $-L_2$ generate contraction semigroups. Suppose to the contrary that there exists an interval $I = (I_-,\,I_+)$ such that for $\varphi \in I$ and some $x \in \mathcal{H}$ we have $||e^{\varphi L_2} x|| > ||x||$.
We consider the same decomposition of ${\rm d} L$ as in (i). We treat $L_1 - 1/2 L_2^2$ as a perturbation and express $U(s) \equiv U(s,0)$ be a Duhamel formula. Since the perturbation is deterministic this is the standard version of the formula,
$$
U(s) = e^{L_2 B_s} + \int_0^s e^{L_2 (B_s - B_{s'})} (L_1 - \frac{1}{2} L_2^2) U(s') {\rm d} s'.
$$
The event $E_{I,\,s} = \{ B_s \in I \quad \mbox{and} \quad I_- - 1 \leq B_t \leq I_+ + 1,\,0\leq t \leq s\}$ has a non-zero probability for any interval $I$ and any $s$. By choosing $s$ sufficiently small we can then achieve $||U(s) x|| > ||x||$, which is in contradiction with (i).
Since $L_2^* = - L_2$ we have
$$
{\rm d} ||x(s)||^2= \left[ (L_1x(s),x(s)) + (x(s),L_1x(s)) + (L_2x(s), L_2x(s))\right] {\rm d} s
$$
and since ${\rm d} ||x(0)|| \leq 0$ we conclude that $L_1 - 1/2 L^2_2$ is dissipative and hence generate a contraction semigroup.
The last claim of the proposition is not related to a classification of contraction semigroups. To prove it, observe that if $L_2$ is antiself-adjoint then $L_2^2$ is a generator of contractions. Hence $L_1 - 1/2 L_2^2 + 1/2 L^2_2$ is also a generator of contractions. Alternatively $L_1$ is a generator of the semigroup $\mathbb{E}[X_t]$.
\hfill $\square$
\begin{rem}
In a Banach space version of the proposition, the condition $i L_2$ is self-adjoint should be replaced by $L_2$ is a generator of isometries. The proof is technically more involved and requires a version of Trotter-Kato formula that does not seem to be available in the literature (\cite{Kurtz} assumes compact state space, while \cite{Gough2, Durr} assume the Hilbert space structure). In particular that if ${\rm d} L_j,\,j=1,2$ is the generator of a propagator $U_j(s,s')$, then the propagator $U(s,s')$ generated by a sum ${\rm d} L_1 + {\rm d} L_2$ can be expressed as
$$
U(s,s') = \lim_{N \to \infty} U_1(s,s_N) U_2(s,s_N) U_1(s_N, s_{N-1}) U_2(s_N, s_{N-1}) \dots U_1(s_1,s') U_2(s_1,s'),
$$
where $s \geq s_N \geq \cdots \geq s_1 \geq s'$ is any partition of the interval with a mesh going to $0$ as $N \to \infty$. This implies that if ${\rm d} L_1$ and ${\rm d} L_2$ generates contractions then so does ${\rm d} L_1 + {\rm d} L_2$.
\end{rem}
The gap condition, assumption (B), is also completely standard in the adiabatic theory. Since $L_1$ is a generator of contraction semigroup we have (see \cite{AFGG}) $\ker{L_1} \cap \mathop{\mathrm{ran}}{L_1} = 0$ and the gap condition implies
\begin{equation}
\label{decomposition}
\mathcal{B} = \ker L_1(s) \oplus \mathop{\mathrm{ran}} L_1(s).
\end{equation}
The rather restrictive condition (C) allows to define the slow manifold and we cannot imagine how it can be relaxed.
Before stating our results we shortly recall concepts from the adiabatic theory, see \cite{AFGG} or \cite{Te03} for a more thorough exposition. Let $P(s)$ be a $\mathcal{C}^1$ family of projections on $\mathcal{B}$ then the equation
$$
\pder{s} T(s,\,s') = [\dot P(s),\, P(s)] T(s,\,s'), \quad T(s',\,s') = \mathbbm{1}
$$
defines parallel transport on $\mathop{\mathrm{ran}} P(s)$. The name ``parallel transport" is justified by two crucial properties
\begin{enumerate}
\item[(i)] $T(s,\,s') P(s') = P(s) T(s,\,s')$,\\
\item[(ii)] A section $x(s) = T(s,\,s') x(s') \in \mathop{\mathrm{ran}} P(s)$ satisfies the equation
$$
P(s) \dot x(s) =0.
$$
\end{enumerate}
The parallel transport relevant to Eq.~(\ref{eq:1}) is given by the projection $P(s)$ on $\ker L_1(s) $ in the direction of $\mathop{\mathrm{ran}} L_1(s) $. This projection is well defined thanks to the decomposition Eq.~(\ref{decomposition}).
Henceforth $T(s,\,s')$ shall always refer to this particular projection, unless stated otherwise.
\begin{thm}
\label{thm:1}
Let $L_1(s),\,L_2(s)$ be $C^3$ families of operators satisfying assumptions (A)-(C). Then the differential equation $\varepsilon \,{\rm d} X(s) = {\rm d} L(s) X(s)$ admits solutions of the form
$$
X(s) = a_0(s) + \sqrt{\varepsilon} \int_0^s U_\varepsilon(s,\,s') L_2(s') b_1(s')\, {\rm d} W_{s'} + \varepsilon (a_1(s) + b_1(s)) + O(\varepsilon^{\frac{3}{2}}),
$$
where
\begin{align*}
a_0(s) &= T(s,\,0) a_0(0),\\
b_1(s) &= L_1(s)^{-1} \dot{a_0}(s),\\
a_1(s) &= \int_0^s T(s,\,s') P(s') \dot{b}_1(s')\, {\rm d} s',
\end{align*}
and the initial condition $a_0(0)$ belongs to $\ker L_1(0)$.
\end{thm}
We note that the integrand $U_\varepsilon(s,s')$ refers to the future and the integral is the backward \^{I}to integral.
The theorem is an immediate corollary of a more general Theorem~\ref{thwg} that describes the full expansion to all orders in $\varepsilon$. We feature it separately because we are not aware of any application of the expansion beyond the first order.
\section{Stochastic Schr\"{o}dinger equation}
\label{schrodinger}
The theorem may be applied to a driven stochastic Schr\"{o}dinger equation \cite[Chapter 5]{Holevo},
$$
\varepsilon {\rm d} \! \ket{\psi(s)} = -(i H(s) + \frac{1}{2} \Gamma(s)^2) \ket{\psi(s)} \, {\rm d} s - \sqrt{\varepsilon}i \Gamma(s) \ket{\psi(s)} {\rm d} B_s,
$$
where $\ket{\psi}$ is a vector in a Hilbert space and $H,\,\Gamma$ are self-adjoint operators. The equation generates unitary evolution and the average state $\bar \rho(s) = \mathbb{E}[\ket{\psi(s)}\!\bra{\psi(s)}]$ satisfies a Lindblad equation
\begin{equation}
\label{eq:lindblad}
\varepsilon \dot{\bar{\rho}}(s) = -i[H(s),\bar\rho(s)] + \Gamma(s) \bar\rho(s) \Gamma(s) - \frac{1}{2}(\Gamma^2(s) \bar\rho(s) + \bar\rho(s) \Gamma^2(s)).
\end{equation}
As in the deterministic case \cite[Section 3.1]{AFGG}, we need to subtract the dynamical phase before we can directly apply the adiabatic theorem. For an integrable function $E(s)$ and a square integrable function\footnote{The artificial square root in the definition of $\gamma$ was introduced in order to have the final results in the same form as in the Lindblad case.} $\sqrt{\gamma(s)}$ the transformation $ H(s) \to H(s) - E(s)$, $\Gamma(s) \to \Gamma(s) - \sqrt{\gamma(s)}$ transforms the solution of the stochastic Schr\"{o}dinger equation according to
$$
\ket{\psi(s)} \to e^{+i \frac{1}{\varepsilon} \int_0^s E(t) {\rm d} t + i \frac{1}{\sqrt{\varepsilon}} \int_0^s \sqrt{\gamma(t)} {\rm d} B_t} \ket{\psi(s)}.
$$
For simplicity we consider a $d$-dimensional Hilbert space and $H(s),\,\Gamma(s)$ with simple eigenvalues $E_0(s) = 0,\dots,E_{d-1}(s)$, $\sqrt{\gamma_0(s)} = 0,\,\dots,\,\sqrt{\gamma_{d-1}(s)}$ corresponding to a joint normalized eigenbasis $\ket{\psi_0(s)},\,\dots,\,\ket{\psi_{d-1}(s)}$. The eigenstate $\ket{\psi_k(s)}$ is determined only up to a phase and without loss of generality we assume that it is chosen in accordance with the parallel transport associated to the projection $\ket{\psi_k(s)}\! \bra{\psi_k(s)}$. Primarily, we shall study solutions
$\ket{\psi_\varepsilon(s)}$ of the stochastic Schr\"{o}dinger equation with an initial condition $\ket{\psi_\varepsilon(0)} = \ket{\psi_0(0)}$. Likewise we can study solutions with an initial condition $\ket{\psi_k(0)},\,k\in(1,\,d-1)$ after applying the above mentioned transformations.
Of particular interest is the tunneling out of the ground state defined as
\begin{align*}
T_\varepsilon(s) &= 1 - |\braket{\psi_0(s)}{\psi_\varepsilon(s)}|^2 \\
&= \sum_{k=1}^{d-1} |\braket{\psi_k(s)}{\psi_\varepsilon(s)}|^2.
\end{align*}
\begin{thm}
Let $H(s),\,\Gamma(s)$ be as above. Then the stochastic Schr\"{o}dinger equation admits a solution
$$
\ket{\psi_\varepsilon(s)} = \ket{\psi_0(s)} + \sqrt{\varepsilon} \sum_{k=1}^{d-1} \left(\int_0^s D_\varepsilon^{(k)}(s,\,s') t_k(s') {\rm d} B_{s'} \right) \ket{\psi_k(s)} + O(\varepsilon),
$$
where
$$
D_\varepsilon^{(k)}(s
,s')=e^{-i \frac{1}{\varepsilon} \int_{s'}^s E_k(t) {\rm d} t - i \frac{1}{\sqrt{\varepsilon}}\int_{s'}^s \sqrt{\gamma_k(t)} {\rm d} B_t} ,\quad t_k(s) = -i\sqrt{\gamma_k(s)}\frac{\braket{\psi_k(s)}{ \dot{\psi_0}(s)}}{-i E_k(s) - \frac{1}{2} \gamma_k(s)}.
$$
In particular for the tunneling we have $T_\varepsilon(s) = \varepsilon \sum_{k=1}^{d-1} T_k(s) + O(\varepsilon^\frac{3}{2})$,
$$
T_k(s) = \left| \int_0^s D_\varepsilon^{(k)}(s,s') t_k(s') {\rm d} B_{s'} \right|^2 .
$$
In the leading order, terms $T_k(s)$ are independent random variables, and each term has an exponential distribution with mean
\begin{equation}
\label{mean}
\mathbb{E}[ T_k(s)] = \int_0^s |t_k(s')|^2 {\rm d} s'.
\end{equation}
\end{thm}
{\bf Proof:}
Conditions (A)-(C) for $L_1(s) = -(i H(s) + \frac{1}{2} \Gamma(s)^2)$ and $L_2(s) = -i \Gamma(s)$ are clearly satisfied. $U_\varepsilon(s,s')$ is a unitary propagator. And the operator $L_1(s)$ has eigenvectors $\ket{\psi_k(s)}$ corresponding to simple discrete eigenvalues $-i E_k(s) - (1/2)\gamma_k(s)$.
In view of Theorem~\ref{thm:1} and the discussion above we then have in the leading order
\begin{equation}
\label{p2.1}
U_\varepsilon(s,\,s') \ket{\psi_k(s')} = D_\varepsilon^{(k)}(s,s') \ket{\psi_k(s)} + O(\sqrt\varepsilon).
\end{equation}
We proceed to the next order for the case with the initial condition $a_0(0) = \ket{\psi_0(s)}$. In order to do so we need to compute the coefficient $b_1(s)$. We express it in the joint eigenbasis of $H$ and $\Gamma$,
$$
b_1(s) = \sum_{k=1}^{d-1} \frac{\braket{\psi_k(s)}{\dot{\psi_0}(s)}}{-i E_k(s) - \frac{1}{2} \gamma_k(s)} \ket{\psi_k(s)}.
$$
It then follows from Theorem~\ref{thm:1} that
$$
\ket{\psi_\varepsilon(s)} = \ket{\psi_0(s)} + \sqrt{\varepsilon} \sum_{k=1}^{d-1} \left(\int_0^s U_\varepsilon(s,s') t_k(s') \ket{\psi_k(s')} {\rm d} B_{s'} \right) + O(\varepsilon),
$$
and by substituting from Eq.~(\ref{p2.1}) we obtain the first equation of the theorem. The expression for the tunneling is an immediate consequence. To compute the mean of the tunneling we use Formula~(\ref{itometry}).
It remains to show that transitions to different excited states are independent in the leading order and that the distribution of the tunneling is exponential. This will require some effort.
We recall that exponential probability distribution with mean $\mu$ has a probability density function $p(x) = \mu^{-1} e^{-\frac{x}{\mu}}$ and is uniquely characterized by its moments $\int p(x) x^n = n! \mu^n$. Our strategy is to compute the moments by establishing a recurrence relation between $\mathbb{E}[T_\varepsilon^n]$ and $\mathbb{E}[T_\varepsilon^{n-1}]$.
For convenience we first express the tunneling as a forward stochastic integral. Using the computation in Remark~\ref{conversion}, Eq.~(\ref{eq:7}), with $L_1 = ( -i E_k - 1/2 \gamma_k)$ and $L_2 = -i \sqrt{\gamma_k}$ we have
$$
\int_0^s D^{(k)}_\varepsilon(s,\,s') t_k(s') {\rm d} B_{s'} = D_\varepsilon^{(k)}(s,0) \int_0^s D_\varepsilon^{(k)}(0,\,s') r_k(s') {\rm d} B_{s'} + O(\sqrt{\varepsilon}),
$$
where $r_k(s) = -i\sqrt{\gamma_k(s)} \braket{\psi_k(s)}{ \dot{\psi_0}(s)} /(-i E_k(s) + \frac{1}{2} \gamma_k(s))$. We hence obtain a forward expression for the tunneling in the leading order,
$$
T_k(s) = \left| \int_0^s D_\varepsilon^{(k)}(0,s') r_k(s') {\rm d} B_{s'} \right|^2 .
$$
Note that $|t_k(s)|^2 = |r_k(s)|^2$, as it has to be for the mean to remain the same.
We start by considering a single transition $T_k(s)$. \^{I}to rules imply
$$
{\rm d} T_k(s) = \left(\int_0^s \bar D_\varepsilon^{(k)}(0,s') \bar r_k(s') {\rm d} B_{s'}\right) D_\varepsilon^{(k)}(0,s) r_k(s) {\rm d} B_{s} + c.c. + |r_k(s)|^2 {\rm d} s,
$$
and
$$
({\rm d} T_k(s))^2 = \left(\int_0^s \bar D_\varepsilon^{(k)}(0,s') \bar r_k(s') {\rm d} B_{s'}\right)^2 D_\varepsilon^{(k)}(0,s)^2 r^2_k(s) {\rm d} s
+c.c + 2 T_k(s) |r_k(s)|^2 {\rm d} s.
$$
Using integral version of ${\rm d} T^n = n T^{n-1} {\rm d} T + (1/2)n(n-1) T^{n-2} {\rm d} T {\rm d} T$ and taking the expectation value we have (use the first formula in Eq.~(\ref{itometry}))
\begin{multline*}
\mathbb{E}[T_k^n(s)] = n \int_0^s \mathbb{E}[T_k^{n-1}(s')] |r_k(s')|^2 {\rm d} s' + n (n-1) \int_0^s \mathbb{E}[T_k^{n-1}(s')] |r_k(s')|^2 {\rm d} s' \\
+ \frac{n(n-1)}{2} \int_0^s \mathbb{E}[T^{n-2}(s') \left(\int_0^{s'} \bar D_\varepsilon^{(k)}(0,s'') r_k(s'') {\rm d} B_{s''}\right)^2 D_\varepsilon^{(k)}(0,s')^2 r^2_k(s') {\rm d} s'] + c.c.
\end{multline*}
Integrating by parts with respect to the factor $e^{ -i \frac{2}{\varepsilon} \int_{0}^{s'} E_k(t) {\rm d} t} $ shows that the second line
is of order $\varepsilon^{1/2}$, whence
$$
\mathbb{E}[T^n_k(s)] = n^2 \int_0^s \mathbb{E}[T_k^{n-1}(s')] |r_k(s')|^2 {\rm d} s' + O(\varepsilon^{\frac{1}{2}}).
$$
Using this relation recursively we arrive at
\begin{align*}
\mathbb{E}[T^n_k(s)] &= (n!)^2 \int\limits_{0\leq s_1 \leq \dots \leq s_n \leq s} \prod_{i=1}^n |r_k(s_i)|^2 {\rm d} s_1 \dots {\rm d} s_n + O(\varepsilon^\frac{1}{2}) \\
&= n! \left(\int_0^s |r_k(s')|^2 {\rm d} s'\right)^n + O(\varepsilon^\frac{1}{2}),
\end{align*}
which is exactly the relation characterizing exponential distribution.
Now consider two terms $T_k(s),\,T_l(s)$ for $l \neq k$. By \^{I}to formula we have
\begin{align*}
\mathbb{E}[T_k(s) T_l(s)] &= \int_0^s \left(\mathbb{E}[{\rm d} T_k(s') T_l(s')] + \mathbb{E}[T_k(s') {\rm d} T_l(s')] + \mathbb{E}[{\rm d} T_k(s') {\rm d} T_l(s')] \right) \\
&=\int_0^s \left(|r_k(s')|^2\mathbb{E}[T_l(s')] + |r_l(s')|^2\mathbb{E}[T_k(s')] \right) {\rm d} s'+ O(\sqrt{\varepsilon}) \\
&= \mathbb{E}[T_k(s)] \mathbb{E}[T_l(s)] + O(\sqrt{\varepsilon}).
\end{align*}
That the last term on the RHS of the first line is of order $\varepsilon^{1/2}$ can be shown by integration by parts. Hence we showed that $T_k$ and $T_l$ are uncorrelated and we proceed to higher powers by induction. Suppose that $T_k^{n-1}$ and $T_l^m$ ($T_k^n$ and $T_l^{m-1}$) are uncorrelated to the leading order, then we have
\begin{align*}
\mathbb{E}[T_k^n T_l^m] &= \int \left( \mathbb{E}[{\rm d} (T_k^n) T_l^m] + \mathbb{E}[T_k^n {\rm d}(T_l^m)] + \mathbb{E}[{\rm d} (T_k^n) {\rm d} (T_l^m)] \right) \\
&= \int \left( n^2 |r_k|^2 \mathbb{E}[T_k^{n-1} T_l^m] + m^2 |r_l|^2 \mathbb{E}[T_k^{n} T_l^{m-1}] \right) + O(\sqrt{\varepsilon}) \\
&= \int \left({\rm d} (\mathbb{E}[T_k^n])\mathbb{E}[T_l^n] + \mathbb{E}[T_k^n] {\rm d} (\mathbb{E}[T_l^n]) \right) + O(\sqrt{\varepsilon} ) \\
&=\mathbb{E}[T_k^n]\mathbb{E}[T_l^n] + O(\sqrt{\varepsilon}).
\end{align*}
So to leading order $T_k$ and $T_l$ are independent, which finishes the proof.
\hfill $\square$
\begin{rem}
The main deficiency of the expansion in Theorem~\ref{thm:1} is that it involves the propagator itself, albeit in a higher order. It is straightforward, although cumbersome, to recursively eliminate the propagator.
We do not know of any more direct manner to derive higher order terms in the expansion.
\end{rem}
Formula (\ref{mean}) for the mean tunneling has been derived in \cite{AFGG} using the corresponding adiabatic Lindblad equation, Eq.~(\ref{eq:lindblad}), and subsequently used to study an optimal sweeping rate \cite{AFGG2} and Landau-Zener tunneling with dephasing \cite{AFGG3}. The mean tunneling is additive, which was interpreted as the tunneling in the dephasing case being local and unidirectional. The full statistics of the tunneling derived here, offers an unexpected twist. If the tunneling was additive it would have a Gaussian distribution, not an exponential one. It follows that only the mean tunneling is additive, while higher order cumulants exhibit non-local behavior typical for the Hamiltonian evolution.
\section{Full expansion and its proof}
\label{main}
Now we present the main theorem, that describes the expansion to all orders.
\begin{thm} \label{thwg}
Let $L_1(s), L_2(s)$ be $C^{N+2}$-families of operators satisfying Assumptions~(A)-(C). Then
\begin{enumerate}
\item The differential equation $\varepsilon\, {\rm d} X={\rm d} L(s)X$ admits solutions of the form
\begin{align} \label{solu}
X(s)=\sum_{n=0}^N\varepsilon^n\left( \varepsilon^{-1/2}\int_0^s U_\varepsilon(s,\,s') L_2(s') b_n(s')\, {\rm d} B_{s'} + a_n(s)+b_n(s)\right)+\varepsilon^{N}r_N(\varepsilon,s)
\end{align}
with
\begin{itemize}
\item $a_n(s) \in \ker L_1(s), \ b_n(s) \in \mathop{\mathrm{ran}} L_1(s)$.
\item initial data $x(0)$ is specified by arbitrary $a_n(0) \in \ker L_1(0)$; however, the $b_n(0)$ are determined
below by the $a_n(0)$ and together define the "slow manifold".
\end{itemize}
\item The coefficients are determined recursively through $(n=0, \dots, N)$
\begin{align}
b_0(s)&=0\,, \nonumber \\
a_n(s)&=T(s,0)a_n(0)+ \int_0^sT(s,s')\dot{P}(s')b_n(s')ds'\,, \label{an} \\
b_{n+1}(s)
&=L_1(s)^{-1}\left(\dot{P}(s)a_n(s)+P_{\perp}(s)\dot{b}_n(s)\right)\,. \label{bnpo}
\end{align}
\item The remainder is uniformly small in $\varepsilon$ and is of the form
$$
r_N(\varepsilon,s) = \sqrt{\varepsilon} \int_0^s r_N^{(2)}(\varepsilon, s') {\rm d} B_{s'} + \varepsilon r_N^{(1)}(\varepsilon,s) ,
$$
where $r_N^{(1)}(\varepsilon, s),\,r_N^{(2)}(\varepsilon, s)$ are uniformly bounded functions. In particular, $r_N(\varepsilon,s) = O(\sqrt{\varepsilon})$.
\end{enumerate}
\end{thm}
{\bf Proof:}
Since $L_1(s)$ is a generator of a contraction semigroup (see the last claim in Proposition~\ref{shy}) we can use the standard deterministic adiabatic theory for an equation $\varepsilon {\rm d} \tilde{X}(s) = L_1(s) \tilde{X}(s)$. Using the expansion in \cite[Theorem~6]{AFGG} the equation has a solution,
$$
\tilde{X}(s) = \sum_{n=0}^N\varepsilon^n\left( a_n(s)+b_n(s)\right)+\varepsilon^{N+1}r^{(1)}_N(\varepsilon,s),
$$
where $r^{(1)}_{N}(\varepsilon,s)$ is uniformly bounded.
By the Duhamel formula of Lemma~\ref{duhamel} we then have a solution of the stochastic equation,
\begin{multline*}
X(s)=\sum_{n=0}^N\varepsilon^n\left( \varepsilon^{-1/2}\int_0^s U_\varepsilon(s,\,s') L_2(s') b_n(s')\, {\rm d} B_{s'} + a_n(s)+b_n(s)\right) \\+\varepsilon^{N+1}r^{(1)}_N(\varepsilon,s) + \varepsilon^{N+\frac{1}{2}} \int_0^s U_\varepsilon(s,\,s') L_2(s') r^{(1)}_N(\varepsilon,s') {\rm d} B_{s'}.
\end{multline*}
This is exactly the expansion of the theorem with $r_N^{(2)}(\varepsilon, s') = U_\varepsilon(s,\,s') L_2(s') r^{(1)}_N(\varepsilon,s')$.
That $r_n^{(2)}(\varepsilon,s)$ is uniformly bounded (with probability $1$) follows from the assumption (A), which implies that $||U_\varepsilon(s,s')|| \leq 1$. That the error is of the order $O(\sqrt\varepsilon)$ follows from Lemma~\ref{bound}, or more precisely from a backward integration counterpart of the lemma. In fact, Lemma~\ref{bound}.(a) is sufficient for that conclusion, at the same time Lemma~\ref{bound}.(b) gives better error estimates.
\hfill $\square$
We conclude with several remarks regarding the generality of our exposition. Including several independent noises, i.e. $L_2 {\rm d} B \to \sum_k L_2^{(k)} {\rm d} B_k$ where $B_k$ are independent Brownian motions, is straightforward. In particular the tunneling Eq.~(3) turns into a sum over the noises, each giving an independent contribution to the tunneling. Boundedness of $L_1,\,L_2$ can surely be relaxed as well as the gap condition, Assumption~(B). We do not plan to elaborate on any of these generalizations. On the other hand it is important to allow generators $L_1(s),\,L_2(s)$ to depend on the Brownian motion, $B_t$, for $s \geq t \geq 0$. We hope to address this question in a further work.
\medskip\noindent
{\bf Acknowledgements.}
I thank Gian Michele Graf and Eddy Mayer-Wolf for fruitful discussions. A part of the work was done while I visited the Isaac Newton Institute in Cambridge, UK. Support by the Swiss National Science Foundation is acknowledged.
|
\section*{Supplementary material}
The quantum channel $\mathcal{E}$ introduced in the main text,
mapping the input state
$\rho$ into the output state $\rho'$, namely
$\rho'=\mathcal{E}(\rho)$, can be conveniently described
in the Fano representation (also known as the Bloch representation)
~\cite{fano,eberly,mahler,striniqpt,strinidiamondnorm}.
In the RWA regime, the quantum protocol described by $\mathcal{E}$,
transfers, up to a trivial unitary transformation,
the state $\rho$ from $\textsf{Q}_1$ to the cavity $\textsf{C}$,
and finally from $\textsf{C}$ to $\textsf{Q}_2$,
leaving $\textsf{Q}_1$ and $\textsf{C}$ in their ground state.
More precisely, if ${\bf r}^\prime=(x',y',z')$ are the Bloch ball coordinates
of the final state $\rho^\prime$ of $\textsf{Q}_2$ and ${\bf r}=(x,y,z)$ the
coordinates of the input state $\rho$ of $\textsf{Q}_1$, then
$x'=-x$, $y'=-y$, and $z'=z$. The state $\rho$ can therefore be
recovered from $\rho'$ after a rotation of angle $\pi$ about
the $z$-axis of its Bloch ball. Deviations from the ideal
quantum protocol appear when effects beyond the RWA cannot be neglected.
In the Fano form we write
$\rho=\frac{1}{2}\left(I^{(1)}+{\bf r}\cdot {\bf \sigma}^{(1)}\right)$ and
$\rho^\prime=\frac{1}{2}\left(I^{(2)}+{\bf r}^\prime\cdot {\bf \sigma}^{(2)}\right)$,
with $I^{(k)}$ identity operator for qubit $k$.
Due to the linearity of quantum mechanics the Bloch vectors
${\bf r}$ and ${\bf r'}$ are connected through an affine map $\mathcal{M}$
as follows:
\begin{equation}
\left[
\begin{array}{c}
{\bf r'}
\\
\hline
1
\end{array}
\right]
=
\mathcal{M}
\left[
\begin{array}{c}
{\bf r}
\\
\hline
1
\end{array}
\right]
=
\left[
\begin{array}{ccc}
{\bf M} & \Big \lvert & {\bf a} \\
\hline
{\bf 0}^T & \Big \lvert & 1
\end{array}
\right]
\left[
\begin{array}{c}
{\bf r}
\\
\hline
1
\end{array}
\right],
\label{eq:affine}
\end{equation}
where ${\bf M}$ is a $3\times 3$ real matrix,
${\bf r}$, ${\bf r'}$ and ${\bf a}$ real column vectors of
dimension $3$ and
${\bf 0}$ the null vector of the same dimension.
The Fano representation of quantum operations is physically transparent
since the Bloch vectors directly provide the expectation values
of polarization measurements.
While in general an affine map for a qubit depends
on twelve parameter~\cite{qcbook},
we found from the numerical simulation of the above described
quantum protocol the following structure
of ${\bf M}$ and ${\bf a}$:
\begin{equation}
{\bf M} = \left(
\begin{array}{ccc}
m_{xx} & m_{xy} & 0 \\
m_{yx} & m_{yy} & 0 \\
0 & 0 & m_{zz} \\
\end{array} \right), \qquad
{\bf a} = \left(
\begin{array}{c}
0 \\
0 \\
a_z \\
\end{array} \right).
\label{eq:Memoryless-Kraus-Operators}
\end{equation}
The dependence of the six non-zero parameters,
$m_{xx},m_{xy},m_{yx},m_{yy},m_{zz}$, and $a_z$,
is shown in Fig.~\ref{fig:blochg} as a function
of the parameter $g$. Note that in the RWA
limit ($g\ll 1$) we have $m_{xx}=m_{yy}=-1$,
$m_{zz}=1$, and $m_{xy}=m_{yx}=a_z=0$, as expected
for the ideal quantum state transfer protocol.
On the other hand, for $g\gtrsim 0.1$ significant
deviations from the ideal protocol are observed.
It is interesting to remark that the positions of peaks and valleys
matches those found in Fig.~2 of the main text for the coherent
information (note that here values of $g$ up to $g=1$ are considered).
Quantum channel $\mathcal{E}$ has an interesting
and non trivial
structure, since it is nonunital ($\mathcal{E}
(I)\ne I$ since $a_z\ne 0$) and matrix $M$ is not
symmetric ($m_{xy}\ne m_{yx}$).
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=16.0cm]{fig-supp.eps}
\end{center}
\caption{Non-zero parameters of the Fano representation of the
quantum channel $\mathcal{E}$, as a function of the
coupling strength $g$, for $T_1=T_2=\pi/2g$, $T_c=0$, and sudden switch
on/off of the couplings.}
\label{fig:blochg}
\end{figure*}
The geometrical meaning of the quantum channel $\mathcal{E}$ can
be understood from the decomposition of map
(\ref{eq:affine})-(\ref{eq:Memoryless-Kraus-Operators}) into
a sequence of elementary affine maps.
We first write $\mathcal{M}=\mathcal{M}_1\mathcal{M'}$,
where
\begin{equation}
\mathcal{M}_1=\left[
\begin{array}{cccc}
\cos\theta & 0 & 0 & 0 \\
0 & \cos\theta & 0 & 0 \\
0 & 0 & \cos^2\theta & \pm \sin^2\theta \\
0 & 0 & 0 & 1
\end{array}
\right],
\end{equation}
represents a displacements of the Bloch sphere along the
$z$-axis~\cite{striniepjd}.
Note that we have $\sin^2\theta=a_z$ or
$\sin^2\theta=-a_z$ depending on the sign of $a_z$.
In the first case the displacement of the center of the Bloch sphere is
along the positive direction of the $z$-axis and can be seen as representative
of zero temperature dissipation (amplidude damping
channel~\cite{qcbook}), in the latter case
the displacement is along the negative $z$-direction and can be seen as
thermal excitation.
The affine map $\mathcal{M}'$ represents a unital quantum channel. It reads as follows:
\begin{equation}
\mathcal{M'}=
\left[
\begin{array}{ccc}
{\bf M'} & \Big \lvert & {\bf 0} \\
\hline
{\bf 0}^T & \Big \lvert & 1
\end{array}
\right]
=
\left[
\begin{array}{cccc}
m_{xx}' & m_{xy}' & 0 & 0 \\
m_{yx}' & m_{yy}' & 0 & 0 \\
0 & 0 & m_{zz}'& 0 \\
0 & 0 & 0 & 1
\end{array}
\right],
\end{equation}
with $m_{ij}'=(\cos\theta) m_{ij}$ for $i,j=x,y$ and
$m_{zz}'=(\cos^2\theta) m_{zz}$.
Matrix ${\bf M'}$ can be written using the singular value
decomposition as
${\bf M'}={\bf O_1 D O_2^T}$,
with ${\bf O_1}$ and ${\bf O_2}$ rotation matrices and
${\bf D}$ diagonal scaling matrix.
Since ${\bf S}\equiv {\bf M' M'^T}=
{\bf O_1 D^2 O_1^T}$, the diagonal entries of
of ${\bf D}$ (known as the singular values) are the
square roots of the eigenvalues of the symmetric matrix ${\bf S}$.
We can therefore write the affine map
${\cal M'}$ as composition of three elementary maps,
${\cal M'}={\cal M}_2{\cal M}_3{\cal M}_4$,
with ${\cal M}_2$ and ${\cal M}_4$ rotations of the Bloch sphere
about the $z$ axis
and ${\cal M}_3$ deformation of the Bloch sphere into an ellipsoid
centred at the origin of the Bloch sphere and whose axes are directed
along $x$, $y$ and $z$. The lengths of the semi-axes of the ellipsoid
are the singular values of ${\bf M'}$.
To summarize, the overall quantum channel $\mathcal{E}$
is obtained by the composition of
a rotation of the Bloch sphere
(affine map $\mathcal{M}_4$), a deformation
of the Bloch sphere ($\mathcal{M}_3$), another rotation
($\mathcal{M}_2$), and a displacement of the Bloch sphere
($\mathcal{M}_1$). We need three real parameters to determine
$\mathcal{M}_3$, and one parameter for each of the other
transformations. Overall we have six real parameters, as also
clear from Eq.~(\ref{eq:Memoryless-Kraus-Operators}).
Finally, we point out that a set of Kraus operators for map
$\mathcal{E}$ can be easily obtained by composing the
Kraus representations for the elementary transformations
$\mathcal{M}_i$, ($i=1,...,4$), for which Kraus operators
are well known~\cite{qcbook}.
|
\section{Introduction}
Probing the quark substructure of scalar mesons is proven to be quite non-trivial (see \cite{pdg} - \cite{FJSS11}). The simple quark-antiquark model, which works well for pseudoscalars and vectors, does not explain the properties of the scalars below 1 GeV such as their light and inverted mass spectrum. The MIT bag model of Jaffe provides a theoretical foundation for understanding the properties of lowest-lying scalar mesons within a diquak-antidiquark picture. While scalars above 1 GeV are better treated within the quark model, they too show some signs of deviations from the quark-antiquark picture \cite{Mec}. A natural question then arises as to whether various underlying mixings among scalar mesons below and above 1 GeV are in any way responsible for these deviations. To answer this question it is necessary to investigate the global picture of scalars below and above 1 GeV and study the family relations among them. This objective was taken up in \cite{global} in which such family relations were studied in some detail within a generalized linear sigma model that includes two nonets of scalar mesons and two nonets of pseudoscalar mesons (a quark-antiquark nonet and a four-quark nonet). Prior works also include \cite{BFMNS01}, \cite{mixing}, \cite{thermo}, \cite{FJS08}, \cite{FJS05}-\cite{global},
\cite{FJSS11}.
The generalized linear sigma model was also applied to several semi-leptonic decays of $D_s^+(1968)$
measured by the CLEO collaboration \cite{Cleo}. These included $D_s^+(1968)\rightarrow f_0(980) e^+ v_e$, as well as $D_s^+(1968)\rightarrow \eta(\eta') e^+ v_e$. It was shown in \cite{FJSS11} that the model prediction for these semileptonic decays agrees well with the CLEO measurements. In the present work we study the predictions of the same model for $D_s^+(1968)\rightarrow K^0 e^+ v_e$, $D^+ \rightarrow \pi^0 e^+ v_e$,
$D^+ \rightarrow \eta (\eta') e^+ v_e$, $D^+ \rightarrow {\bar K}^0 e^+ v_e$, $D^0 \rightarrow \pi^- e^+ v_e$ and $D^0 \rightarrow K^- e^+ v_e$.
In section II we give a brief review of the hadronic ``weak currents" which are needed for the calculation. These are mathematically given by the so-called Noether currents of the sigma model Lagrangian being employed.
We work in the approximation where renormalization of these currents from the symmetry limit are neglected.
This means that there are no arbitrary parameters available to us.
In section III we give a detailed description of the calculation of the partial decay widths from the currents
discussed in section II. For this purpose we also use information on the scalar and pseudoscalar
meson masses and mixings obtained in \cite{global}. A short summary and discussion is given in section IV.
\section{The Hadronic Vector Currents}
The generalized linear sigma model of ref. \cite{global} describes the global picture of scalar and pseudoscalar mesons below and above 1 GeV in terms of two chiral nonets for scalars and two chiral nonets for pseudoscalars. These are a quark-antiquark nonet $M$ and a four-quark nonet $M'$ that in turn are expressed in terms of the corresponding scalar and pseudoscalar nonets
\begin{eqnarray}
M &=& S + i \phi, \nonumber \\
M' &=& S' + i \phi'.
\label{mmp}
\end{eqnarray}
The hadronic Noether vector currents, that are relevant to the semileptonic decays in the present work are (see Appendix A of \cite{su71}),
\begin{eqnarray}
V_{\mu a}^b&=&i\phi_a^c\stackrel{\leftrightarrow}{\partial_\mu}\phi_c^b +
i\tilde{S}_a^c\stackrel{\leftrightarrow}{\partial_\mu}\tilde{S}_c^b
+i(\alpha_a-\alpha_b)\partial_\mu\tilde{S}_a^b,
\nonumber \\
V_{\mu a}^{\prime b}&=&i\phi_a^{\prime c}\stackrel{\leftrightarrow}{\partial_\mu}\phi_c^{\prime b} +
i\tilde{S}_a^{\prime c}\stackrel{\leftrightarrow}{\partial_\mu}\tilde{S}_c^{\prime b}
+i(\beta_a-\beta_b)\partial_\mu\tilde{S}_a^{\prime b},
\label{su3currents}
\end{eqnarray}
wherein,
\begin{eqnarray}
S &=& \tilde{S} + \langle S \rangle,\quad\quad \langle S_a^b \rangle =\alpha_a \delta_a^b, \nonumber \\
S^\prime &=& \tilde{S}^\prime + \langle S^\prime \rangle,\quad\quad \langle {S^\prime}_a^b \rangle =\beta_a \delta_a^b,
\label{moresu3vevs}
\end{eqnarray}
where $\alpha$ and $\beta$ being the vacuum expectation values of fields $S$ and $S'$, respectively.
Therefore, total vector current is:
\begin{equation}
\left[ V_{\mu a}^b\right]^{\rm tot.}=V_{\mu a}^{b}+V^{\prime b}_{\mu a}.
\label{totalsu3currents}
\end{equation}
In this framework, the physical states become a linear combination (expressed by a rotation matrix) of the unprimed and primed fields.
The rotation matrices relevant to our study in this work are $R_\pi$, $R_K$ and $R_\eta$ defined as follows:
The
transformation between the physical $\pi^+$ and $\pi^{\prime +}$ fields and the
original
fields (say $\phi^+$ and $\phi'^+$) is \cite{FJS05}:
\begin{equation}
\left[
\begin{array}{c} \pi^+(137) \\
\pi'^+(1300)
\end{array}
\right]
=
R_\pi^{-1}
\left[
\begin{array}{c}
\phi_1^2 \\
{\phi'}_1^2
\end{array}
\right]
\label{R_pi}
\end{equation}
where $R_\pi$ is the pion rotation matrix.
Similarly, for the kaon system
\begin{equation}
\left[
\begin{array}{cc}
K^+(496)\\
{K'}^+(1460)
\end{array}
\right]
=
R_K^{-1}
\left[
\begin{array}{cc}
\phi_1^3\\
{\phi'}_1^3
\end{array}
\right]
\label{R_K}
\end{equation}
and for the eta system
\begin{equation}
\left[
\begin{array}{c}
\eta(547) \\
\eta(958) \\
\eta(1295) \\
\eta(1760)\\
\end{array}
\right]
=
R_0^{-1}
\left[
\begin{array}{cc}
\eta_a\\
\eta_b\\
\eta_c\\
\eta_d
\end{array}
\right],
\label{R_0}
\end{equation}
where
\begin{eqnarray}
\eta_a&=&\frac{\phi^1_1+\phi^2_2}{\sqrt{2}}
\hskip .4cm \rightarrow \hskip .4cm
n{\bar n},
\nonumber \\
\eta_b&=&\phi^3_3
\hskip 1.25cm \rightarrow \hskip .4cm
s{\bar s},
\nonumber \\
\eta_c&=& \frac{\phi'^1_1+\phi'^2_2}{\sqrt{2}}
\hskip .17cm \rightarrow \hskip .4cm
ns{\bar n}{\bar s},
\nonumber \\
\eta_d&=& \phi'^3_3
\hskip 1.15cm \rightarrow \hskip 0.4cm
nn{\bar n}{\bar n}.
\label{etafourbasis}
\end{eqnarray}
where on the right the schematic quark substructure is given (in which $n$ stands for non-strange up and down quarks). The rotation matrices have been determined in the global fit of ref. \cite{global} and will be used in the present work.
When extending this model to the case where a heavy flavor (such as the charm quark here) is added, the heavy spin zero mesons need to be considered as quark-antiquark states. This is based on the findings of ref. \cite{FJSS11} in which it is shown that the case of three flavors is special in the sense that
it is the only one in which a two quark-two antiquark field has the correct chiral
transformation property to mix (in the chiral limit) with $M$. Therefore, in the present generalized linear sigma model the kinetic term would then be written as:
\begin{equation}
{\cal L} = -\frac{1}{2} {\rm Tr}^4(\partial_\mu M \partial_\mu M^\dagger)
-\frac{1}{2} {\rm Tr}^3(\partial_\mu M^\prime \partial_\mu M^{\prime\dagger}),
\label{hybridlag}
\end{equation}
where the meaning of the superscript on the trace symbol is
that the first term should be summed over the heavy quark index as
well as the three light indices. This stands in contrast to the second term
which is just summed over the three light quark indices pertaining to the
two quark - two antiquark field $M^\prime$. Since the Noether
currents are sensitive only to these
kinetic terms in the model, the vector currents with
flavor indices 1 through 3 in this model are just the same as in
Eq.(\ref{totalsu3currents}) above. However if either or both flavor
indices take on the value 4 (referring to the heavy flavor) the current
will only have contributions from the field $M$, i.e.
\begin{equation}
\left[ V_{\mu 4}^a \right] ^{\rm tot.} = V_{\mu 4}^a=i\phi_4^c\stackrel{\leftrightarrow}{\partial_\mu}\phi_c^a +
i S_4^c\stackrel{\leftrightarrow}{\partial_\mu}S_c^a,
\label{heavycurrents}
\end{equation}
where the unspecified
indices can run from 1 to 4. The current is given in terms of the ``bare'' fields which are related to the
physical fields through the non-derivative terms (``potential'') terms of the effective Lagrangian. The connections between the ``bare'' and physical fields are given by the appropriate rotation matrices discussed above.
\section{Semileptonic $D$ decays}
\subsection{$D_s^+ \rightarrow K^0 e^+ \nu_e$}
The schematic diagram for this decay is shown in Fig. \ref{Ds_K0ev_FD}. The decay proceeds via the vector current described in Eq. (\ref{heavycurrents})
\begin{equation}
V_{\mu 4}^2=i\phi_4^3\stackrel{\leftrightarrow}{\partial_\mu}\phi_3^2 + \cdots =
i (R_K)_{11} D_s^+ \stackrel{\leftrightarrow}{\partial_\mu} {\bar K}^0 + \cdots
\end{equation}
where $R_K$ is the rotation matrix for the isodoublet pseudoscalars defined in (\ref{R_K}) and computed in \cite{global}. This rotation matrix projects the $d {\bar s}$ pair onto the wave function of the $K^0$. Using the model parameters found in \cite{global} and following the standard calculation of the decay width (Appendix A), the prediction of the model is given in Fig. \ref{Ds_K0ev_R} and compared with the experimental bounds for this decay width extracted from PDG \cite{pdg} (the horizontal lines). Clearly, the prediction of the model which is plotted versus $m[\pi(1300)]$ and for two values of $A_3/A_1$) falls within the experimental bounds. This provides further support for the effectiveness of the model and its predictions for the quark substructure of scalar and pseudoscalar mesons below and above 1 GeV \cite{global}.
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig1.eps}}
\caption[]{
Schematic diagram for semileptonic decay $D_s^+ \rightarrow K^0 e^+ \nu_e$. The rotation matrix $R_K$ is computed in ref. \cite{global} and projects the produced $d{\bar s}$ pair onto the wave function of $K^0$.
}
\label{Ds_K0ev_FD}
\end{figure}
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig2.eps}}
\caption{Prediction of the $MM'$ model for the partial decay width of $D_s^+ \rightarrow K^0 e^+ \nu_e$ versus $m[\pi(1300)]$ for values of $A_3/A_1$ equal to 20 (circles) and 30 (squares). The horizontal lines show the experimental range \cite{pdg}. }
\label{Ds_K0ev_R}
\end{figure}
\subsection{$D^+ \rightarrow \pi^0 e^+ \nu_e$, $D^+ \rightarrow \eta(\eta') e^+ \nu_e$ and $D^+ \rightarrow \eta(\eta') e^+ \nu_e$}
The schematic diagrams for these decays are shown in Fig. \ref{F_D+_FD} and proceed via production of a ${\bar d} d$ pair described by the vector current of Eq. (\ref{heavycurrents}). The ${\bar d} d$ pair then gets projected \cite{global} onto the wave function of the final meson by the appropriate rotation matrix $R_K$ (first decay), $R_0$ (second decay) and $R_K$ (last decay). The relevant vector current for the $\pi^0$ channel is
\begin{equation}
V_{\mu 4}^2=i\phi_4^2\stackrel{\leftrightarrow}{\partial_\mu}\phi_2^2 + \cdots =
-i {1\over\sqrt{2}} \left(R_\pi\right)_{11} D^+ \stackrel{\leftrightarrow}{\partial_\mu} \pi^0 + \cdots
\end{equation}
The extra $\sqrt{2}$ stems from the fact that
\begin{equation}
\pi^0 \propto \left(R_\pi^{-1}\right)_{11} {{u{\bar u} - d {\bar d}}\over \sqrt{2}}
\end{equation}
Therefore the projection of the $d {\bar d}$ onto the $\pi^0$ will include a division by $\sqrt{2}$. Similarly, the relevant vector current for the $\eta$ ($\eta'$) channel is
\begin{equation}
V_{\mu 4}^2 = i\phi_4^2\stackrel{\leftrightarrow}{\partial_\mu}\phi_2^2 + \cdots =
-i {1\over\sqrt{2}} \left(R_0\right)_{11(2)} D^+ \stackrel{\leftrightarrow}{\partial_\mu} \eta(\eta') + \cdots
\end{equation}
and for the ${\bar K}^0$ channel
\begin{equation}
V_{\mu 4}^3 = i\phi_4^2\stackrel{\leftrightarrow}{\partial_\mu}\phi_2^3 + \cdots =
i \left(R_K\right)_{11} D^+ \stackrel{\leftrightarrow}{\partial_\mu} K^0 + \cdots
\end{equation}
The prediction of the model for the decay width in the $\pi$ channel is given in Fig. \ref{D+_pi0ev_R} showing an order of magnitude agreement with experiment. The predictions for the decay width in the $\eta/\eta'$ channels are displayed in Fig. \ref{D+_etaev_R}. For the $\eta$-channel the agreement is at the level of order of magnitude, while for the $\eta'$ channel, for which only the upper experimental bound is given in PDG \cite{pdg}, the prediction is within the experimental range. Similarly, for the ${\bar K}^0$ channel (Fig. \ref{D+_K0ev_R}) the agreement with experiment is at the level of order of magnitude.
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig3a.eps}}
\hskip 1.5cm
{\includegraphics[width=7cm,clip=true]{fig3b.eps}}
\vskip1cm
{\includegraphics[width=7cm,clip=true]{fig3c.eps}}
\caption[]{
Schematic diagrams for semileptonic decays $D^+ \rightarrow \pi^0 e^+ \nu_e$ (top left); $D^+ \rightarrow \eta (\eta') e^+ \nu_e$ (top right); and $D^+ \rightarrow {\bar K}^0 e^+ \nu_e$ (bottom). The rotation matrices $R_0$, $R_\pi$ and $R_K$ are computed in ref. \cite{global} and project the produced $q{\bar q}$ pair onto the wave function of the final state meson.
}
\label{F_D+_FD}
\end{figure}
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig4.eps}}
\caption{Prediction of the $MM'$ model for the partial decay width of $D^+ \rightarrow \pi^0 e^+ \nu_e$ versus $m[\pi(1300)]$ for values of $A_3/A_1$ equal to 20 (circles) and 30 (squares). The horizontal lines show the experimental range \cite{pdg}.}
\label{D+_pi0ev_R}
\end{figure}
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig5a.eps}}
{\includegraphics[width=7cm,clip=true]{fig5b.eps}}
\caption{Prediction of the $MM'$ model for the partial decay width of $D^+ \rightarrow \eta e^+ \nu_e$ (left) and $D^+ \rightarrow \eta' e^+ \nu_e$ (right) versus $m[\pi(1300)]$ for values of $A_3/A_1$ equal to 20 (circles) and 30 (squares). The horizontal lines show the experimental range \cite{pdg} (for the $\eta'$ channel only upper experimental bound is known).}
\label{D+_etaev_R}
\end{figure}
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig6.eps}}
\caption{Prediction of the $MM'$ model for the partial decay width of $D^+ \rightarrow {\bar K}^0 e^+ \nu_e$ versus $m[\pi(1300)]$ for values of $A_3/A_1$ equal to 20 (circles) and 30 (squares). The horizontal lines show the experimental range \cite{pdg}.}
\label{D+_K0ev_R}
\end{figure}
\subsection{$D^0 \rightarrow \pi^- e^+ \nu_e$ and $D^0 \rightarrow K^- e^+ \nu_e$}
The schematic diagram for this decay is shown in Fig. \ref{D0_piKev_FD}. The decay proceeds via the vector current described in Eq. (\ref{heavycurrents}). For the $\pi^-$ channel,
\begin{equation}
V_{\mu 4}^2=i\phi_4^1\stackrel{\leftrightarrow}{\partial_\mu}\phi_1^2 + \cdots =
i (R_\pi)_{11} D^0 {\partial_\mu} \pi^+ + \cdots
\end{equation}
where $R_\pi$ is the rotation matrix for the isotriplet pseudoscalars computed in \cite{global} that projects the $d {\bar u}$ pair onto the wave function of the $\pi^-$. For the $K^-$ channel,
\begin{equation}
V_{\mu 4}^2=i\phi_4^1\stackrel{\leftrightarrow}{\partial_\mu}\phi_1^3 + \cdots =
i (R_K)_{11} D^0 {\partial_\mu} K^+ + \cdots
\end{equation}
where $R_K$ is the rotation matrix for the isodoublet pseudoscalars computed in \cite{global} that projects the $d {\bar u}$ pair onto the wave function of the $K^-$. The model predictions for both of these two channels are presented in Fig. \ref{D0_piKev_R}, and in both cases, there is an order of magnitude agreement.
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig7a.eps}}
\hskip 1.5cm
{\includegraphics[width=7cm,clip=true]{fig7b.eps}}
\caption[]{
Schematic diagrams for semileptonic decays $D^0 \rightarrow \pi^- e^+ \nu_e$ and $D^0 \rightarrow K^- e^+ \nu_e$. The rotation matrices $R_\pi$ and $R_K$ are computed in ref. \cite{global} and project the produced $d{\bar u}$ and $s{\bar u}$ pairs onto the wave functions of $\pi^-$ and $K^-$, respectively.
}
\label{D0_piKev_FD}
\end{figure}
\begin{figure}[htbp]
\centering
\rotatebox{0}
{\includegraphics[width=7cm,clip=true]{fig8a.eps}}
{\includegraphics[width=7cm,clip=true]{fig8b.eps}}
\caption
{Prediction of the $MM'$ model for the partial decay width of $D^0 \rightarrow \pi^- e^+ \nu_e$ (left) and $D^0 \rightarrow K^- e^+ \nu_e$ (right) versus $m[\pi(1300)]$ for values of $A_3/A_1$ equal to 20 (circles) and 30 (squares). The horizontal lines show the experimental range \cite{pdg}.}
\label{D0_piKev_R}
\end{figure}
\section{Summary and discussion}
The main objective of the present work was to further test the generalized linear sigma model of ref. \cite{global} that provides a global picture for the underlying two- and four-quark components of scalar and pseudoscalar mesons below and above 1 GeV. While a precision calculation of the partial decay widths for several semileptonic decays of $D$ mesons was not the primary objective, we saw that the global picture presented in \cite{global} predicts these decay widths in a reasonable agreement with the available experimental data. We interpret this qualitative agreement as further support for the global picture of scalars and pseudoscalars presented in \cite{global}. The calculations were done in the leading order of the model. We expect that additional corrections, such as, for example, the effect of higher order terms in the potential, or addition of scalar and pseudoscalar gluballs, will further refine these predictions. This framework can be applied to processes such as
$B_c^+ \rightarrow scalar + e^+ + \nu_e$ that might be useful for learning about mixing between
a $c\bar{c}$ scalar and the lighter three flavor scalars, or $B_s^0\rightarrow J/\psi f_0(980)$ measured by LHCb.
\section*{Acknowledgments} \vskip -.5cm
The work of
J.Schechter was supported in part by the U.
S. DOE under Contract no. DE-FG-02-85ER 40231. The work of R. J. was supported by a grant of the Ministry of National Education, CNCS-UEFISCDI, project number PN-II-ID-PCE-2012-4-0078.
|
\section{Introduction}
\label{sec_intro}
There has been a lot of interest recently in finitely presented algebras $A$ over a field $K$ defined by homogeneous semigroup relations, that is, semigroup algebras $K[S]$ where $S$ is a semigroup given by a finite presentation where all relations are length preserving. Numerous interesting semigroup algebras arise in this way, including: algebras yielding set-theoretic solutions to the Yang--Baxter equation and quadratic algebras of skew type (see for example \cite{Etingof1999, JespersBook2007, Cedo2010(3)} and \cite{Cedo2012(2), Gateva-Ivanova2003, Jespers2006}), algebras related to Young diagrams, representation theory and and algebraic combinatorics such as the Plactic and Chinese algebras (see \cite[Ch.~5]{lothaire_algebraic}, \cite{cedo_plactic, lascoux_plaxique} and \cite{Cassaigne2001, Jaszunska2011, Cedo2013}), and algebras defined by permutation relations (see \cite{Cedo2010, Cedo2010(2), Cedo2012}). In these examples, there are strong connections between the structure of the algebra $K[S]$ and that of the underlying semigroup $S$.
Further motivation for studying this class comes from other important semigroups in the literature that admit homogeneous presentations, such as the hypoplactic monoid \cite{Novelli2000}, shifted Plactic monoid \cite{Serrano2010}, monoids with the same multihomogeneous growth as the Plactic monoid \cite{duchamp_plactic}, trace monoids \cite{Diekert1997}, and the positive braid monoid \cite{Birman1998, Dehornoy1999}.
One of the fundamental first steps that is necessary when investigating a semigroup $S$ defined by homogeneous relations, and its associated semigroup algebra $K[S]$, is to find a good set of normal forms (canonical representatives over the generating set) for the elements of the monoid, and thus for elements of the algebra. (See the list of open problems in \cite[Section~3]{Cedo2010} for more on the importance of this problem in the context of semigroups defined by permutation relations.) Specifically we would like a set of normal forms that is a regular language, and we want to be able to compute effectively with these normal forms. Two situations where such a good set of normal forms does exist are for monoids that admit presentations by finite complete rewriting systems (see \cite{book_srs}), and for monoids and semigroups that are automatic (see \cite{Epstein1992, campbell_autsg}). Each of these properties also has implications for properties of the corresponding semigroup algebra. Indeed, if the semigroup admits a finite complete rewriting system, then the semigroup algebra admits a finite Gr\"{o}bner--Shirshov basis (see \cite{heyworth_rewriting} for an explanation of the connection between Gr\"obner--Shirshov bases and complete rewriting systems), while the automaticity of the semigroup implies that the algebra is an automaton algebra in the sense ofUfarnovskij; see \cite{Ufnarovskij} and \cite[Section~1]{Cedo2012(2)}.
Many of the examples of homogeneous semigroups mentioned above have been shown to admit presentations by finite complete rewriting systems, and have been shown to be biautomatic; see for example \cite{cgm_plactic, cgm_chinese, Cedo2012(2), guzelkarpuz_fcrs, kubat_grobner, Chen2008}.
It is natural to ask to what extent these results generalise to arbitrary homogeneous semigroups. One can ask:
Does every homogeneous semigroup admit a presentation by a finite complete rewriting system?
Is every such semigroup biautomatic? Within the class of homogeneous semigroups, what is the relationship between admitting a finite complete rewriting system and being biautomatic? (For general semigroups, these properties are independent; see \cite{otto_automonversus}). The aim of this paper is to make a comprehensive investigation of these questions. In fact, we shall consider two different strengths of automaticity, called automaticity and biautomaticity, and we shall also investigate the homotopical finiteness property of finite derivation type (FDT) in the sense of Squier \cite{Squier1994}, which is a key finiteness property that is satisfied by monoids that admit presentations by finite complete rewriting systems (full definitions of all of these concepts will be given in \fullref{Section}{sec_prelims}).
There are various degrees of homogeneity that one can impose on a semigroup presentation. We shall consider finite presentations $\pres{A}{\rel{R}}$ which are:
\begin{itemize}
\item homogeneous: relations are length preserving;
\item multihomogeneous: for each letter $a$ in the alphabet $A$, and for every relation $u=v$ from $\rel{R}$, the number of occurrences of the letter $a$ in $u$ equals the number of occurrences of the letter $a$ in $v$;
\item $n$-ary homogeneous: there is a fixed global constant $n$ such that for every relation $u=v$ in $\rel{R}$ we have $|u| = |v| = n$;
\item $n$-ary multihomogeneous: simultaneously $n$-ary homogeneous and multihomogeneous.
\end{itemize}
Of course, the most restricted class listed here is the class of $n$-ary multihomogeneous presentations.
For brevity, we introduce the following terminology for the four properties we are interested in: a monoid is
\begin{itemize}
\item \fcrs\ if it admits a presentation via a finite complete
rewriting system (with respect to some generating set);
\item \fdt\ if it has finite derivation type;
\item \biauto\ if it is biautomatic;
\item \auto\ if it is automatic.
\end{itemize}
We will also use the natural negated terms: non-\fcrs, non-\fdt, non-\biauto, and non-\auto.
\begin{table}[t]
\centering
\begin{tabular}{c@{}c@{}cc@{}c@{}cl} %
\toprule
\fcrs &$\implies$& \fdt & \biauto &$\implies$& \auto & Example\\
\midrule
Y && Y & Y && Y & Plactic monoid \\
Y && Y & N && Y & \fullref{Example}{eg:fcrsfdtnonbiautoauto}\\
Y && Y & N && N & \fullref{Example}{eg:fcrsfdtnonbiautononauto} \\
N && Y & Y && Y & \fullref{Example}{eg:nonfcrsfdtbiautoauto} \\
N && Y & N && Y & \fullref{Example}{eg:nonfcrsfdtnonbiautoauto} \\
N && Y & N && N & \fullref{Example}{eg:nonfcrsfdtnonbiautononauto} \\
N && N & Y && Y & \fullref{Example}{eg:nonfcrsnonfdtbiautoauto} \\
N && N & N && Y & \fullref{Example}{eg:nonfcrsnonfdtnonbiautoauto} \\
N && N & N && N & \fullref{Example}{eg:nonfcrsnonfdtnonbiautononauto} \\
\bottomrule
\end{tabular}
\caption{A catalogue of $n$-ary multihomogeneous monoids exhibiting all possible combinations of allowable properties.}
\label{tb:resume_properties}
\end{table}
\begin{figure}[t]
\centering
\begin{tikzpicture}[every node/.style={align=center}]
%
\path[rounded corners,fill=lightgray,opacity=.5] (2.75,5.9)--(-1.15,2)--(1,-0.15)--(4.9,3.75)--cycle;
\path[rounded corners,fill=lightgray,opacity=.5] (2.5,5.5)--(-1,2)--(0,1)--(3.5,4.5)--cycle;
\path[rounded corners,fill=lightgray,opacity=.5] (1.25,5.9)--(5.15,2)--(3,-0.15)--(-0.9,3.75)--cycle;
\path[rounded corners,fill=lightgray,opacity=.5] (1.5,5.5)--(5,2)--(4,1)--(0.5,4.5)--cycle;
%
\node[rotate=-45] at (2.85,4.85) {\fcrs};
\node[rotate=-45] at (3.68,4.68) {\fdt};
\node[rotate=45] at (1.15,4.85) {\biauto};
\node[rotate=45] at (0.28,4.68) {\auto};
%
\node[align=center,font=\small] (1) at (2,4) {Plactic\\[-1mm] monoid};
\node(2) at (1,3) {\fullref{Ex.}{eg:fcrsfdtnonbiautoauto}} edge (1);
\node(4) at (3,3) {\fullref{Ex.}{eg:nonfcrsfdtbiautoauto}} edge (1);
\node(3) at (0,2) {\fullref{Ex.}{eg:fcrsfdtnonbiautononauto}} edge (2);
\node(5) at (2,2) {\fullref{Ex.}{eg:nonfcrsfdtnonbiautoauto}} edge (2) edge (4);
\node(7) at (4,2) {\fullref{Ex.}{eg:nonfcrsnonfdtbiautoauto}} edge (4);
\node(6) at (1,1) {\fullref{Ex.}{eg:nonfcrsfdtnonbiautononauto}} edge (3) edge (5);
\node(8) at (3,1) {\fullref{Ex.}{eg:nonfcrsnonfdtnonbiautoauto}} edge (5) edge (7);
\node(9) at (2,0) {\fullref{Ex.}{eg:nonfcrsnonfdtnonbiautononauto}} edge (6) edge (8);
%
\end{tikzpicture}
\caption{The semilattice showing the relationship between examples.
By taking the free product of two examples, one obtains a new monoid
whose properties are given by taking the logical conjunction (that
is, the `and' operation) of corresponding properties of the original
example monoids. This corresponds to the meet operation in this semilattice.}
\label{fig:examplerelationships}
\end{figure}
We are interested in which combinations of these properties a homogeneous monoid can have. Since in general
\fcrs \ implies \fdt, and \biauto \ implies \auto, not all combinations will be possible. But, subject to these restrictions, our main results show that any possible combination is possible within the class of homogeneous monoids.
We then extend this result showing that for every possible combination of these properties, one can construct an $n$-ary multihomogeneous monoid with the given combination of properties (see \fullref{Theorem}{thm_main}). These results are summarised in \fullref{Table}{tb:resume_properties} and the relationship between the various examples is illustrated in \fullref{Figure}{tb:resume_properties}.
The paper is structured as follows: in \fullref{Section}{sec_prelims} we give the basic definitions and results needed in the paper.
In \fullref{Section}{sec:fundexs} we present the fundamental examples from which all of our other examples will be built, namely the Plactic monoid, and \fullref{Examples}{eg:fcrsfdtnonbiautoauto}, \ref{eg:fcrsfdtnonbiautononauto}, \ref{eg:nonfcrsfdtbiautoauto} and \ref{eg:nonfcrsnonfdtbiautoauto}.
\fullref{Section}{sec_freeproducts} contains general results about the behaviour of the properties under free products of monoids, which we then use to construct further examples (as illustrated in \fullref{Figure}{tb:resume_properties}).
In \fullref{Section}{sec_results} we prove some general results which are then applied to extend our result for homogeneous monoids to the more restricted class of $n$-ary multihomogeneous monoids.
\section{Preliminaries}
\label{sec_prelims}
The subsection below on derivation graphs, homotopy bases and finite derivation type is self-contained, but it can be complemented with \cite{OttoKobay_survey,Kobayashi5}. For
further information on automatic semigroups,
see~\cite{campbell_autsg}. We assume familiarity with basic notions of
automata and regular languages (see, for example,
\cite{hopcroft_automata}) and transducers and rational relations (see,
for example, \cite{berstel_transductions}).
For background on sting rewriting systems we refer the reader to \cite{baader_termrewriting,book_srs}.
\subsection{Words, rewriting systems, and presentations}
We denote the empty word (over any alphabet) by $\varepsilon$. For an
alphabet $A$, we denote by $A^*$ the set of all words over $A$. When
$A$ is a generating set for a monoid $M$, every element of $A^*$ can
be interpreted either as a word or as an element of $M$. For words
$u,v \in A^*$, we write $u=v$ to indicate that $u$ and $v$ are equal
as words and $u=_M v$ to denote that $u$ and $v$ represent the same
element of the monoid $M$. The length of $u \in A^*$ is denoted $|u|$,
and, for any $a \in A$, the number of symbols $a$ in $u$ is denoted
$|u|_a$. We denote by $u^\mathrm{rev}$ the reversal of a word $u$;
that is, if $u=a_1\cdots a_{n-1}a_n$ then $u^\mathrm{rev} =
a_na_{n-1}\cdots a_1$, with $a_i\in A$. If $\rel{R}$ is a relation on
$A^*$, then $\cgen{\rel{R}}$ denotes the congruence generated by
$\rel{R}$,
We use standard terminology and notation from the theory of string
rewriting systems; see \cite{book_srs} or \cite{baader_termrewriting}
for background reading.
A presentation is a pair $\pres{A}{\rel{R}}$ that defines [any monoid
isomorphic to] $A^*/\cgen{\rel{R}}$. The presentation
$\pres{A}{\rel{R}}$ is \defterm{homogeneous} (respectively,
\defterm{multihomogeneous}) if for every $(u,v) \in \rel{R}$, also
known as defining relations, and $a \in A$, we have $|u| = |v|$
(respectively, $|u|_a = |v|_a$). That is, in a homogeneous
presentation, defining relations preserve length; in a multihomogenous
presentation, defining relations preserve the numbers of each
symbol. A monoid is \defterm{homogeneous} (respectively,
\defterm{multihomogeneous}) if it admits a homogeneous
(respectively, multihomogeneous) presentation.
A \defterm{string rewriting system}, or simply a \defterm{rewriting
system}, is a pair $(A,\rel{R})$, where $A$ is a finite alphabet and
$\rel{R}$ is a set of pairs $(\ell,r)$, usually written $\ell
\rightarrow r$, known as \defterm{rewriting rules} or simply
\defterm{rules}, drawn from $A^* \times A^*$. The single reduction
relation $\rightarrow_{\rel{R}}$ is defined as follows: $u
\rightarrow_{\rel{R}} v$ (where $u,v \in A^*$) if there exists a
rewriting rule $(\ell,r) \in \rel{R}$ and words $x,y \in A^*$ such
that $u = x\ell y$ and $v = xry$. That is, $u \rightarrow_{\rel{R}} v$
if one can obtain $v$ from $u$ by substituting the word $r$ for a
subword $\ell$ of $u$, where $\ell \rightarrow r$ is a rewriting
rule. The reduction relation $\rightarrow^*_{\rel{R}}$ is the reflexive and
transitive closure of $\rightarrow_{\rel{R}}$. The subscript $\rel{R}$
is omitted when it is clear from context. The process of replacing a
subword $\ell$ by a word $r$, where $\ell \rightarrow r$ is a rule, is
called \defterm{reduction} by application of the rule $\ell \rightarrow
r$; the iteration of this process is also called reduction. A word $w
\in A^*$ is \defterm{reducible} if it contains a subword $\ell$ that
forms the left-hand side of a rewriting rule in $\rel{R}$; it is
otherwise called \defterm{irreducible}.
The rewriting system $(A,\rel{R})$ is \defterm{finite} if both $A$ and
$\rel{R}$ are finite. The rewriting system $(A,\rel{R})$ is
\defterm{noetherian} if there is no infinite sequence $u_1,u_2,\ldots$ of words from
$A^*$ such that $u_i \rightarrow u_{i+1}$ for all $i \in
\mathbb{N}$. That is, $(A,\rel{R})$ is noetherian if any process of
reduction must eventually terminate with an irreducible word. The
rewriting system $(A,\rel{R})$ is \defterm{confluent} if, for any
words $u, u',u'' \in A^*$ with $u \rightarrow^* u'$ and $u
\rightarrow^* u''$, there exists a word $v \in A^*$ such that $u'
\rightarrow^* v$ and $u'' \rightarrow^* v$.
It is well known that a noetherian system is confluent if and only if
all critical pairs resolve, where critical pairs are obtained by considering overlaps of left hand sides of the rewrite rules $\rel{R}$; see \cite{book_srs} for more details.
A rewriting system
that is both confluent and noetherian is \defterm{complete}.
If a monoid admits a presentation with respect to some generating set
that forms a finite complete rewriting system, the monoid is \fcrs.
The \defterm{Thue congruence} $\leftrightarrow^*_{\rel{R}}$ is the equivalence
relation generated by $\rightarrow_{\rel{R}}$. The elements of the monoid
presented by $\pres{A}{\rel{R}}$ are the $\leftrightarrow^*_{\rel{R}}$-equivalence
classes. The relations $\leftrightarrow^*_{\rel{R}}$ and $\cgen{\rel{R}}$ coincide.
Let $M$ be a homogeneous monoid. Let $\pres{A}{\rel{R}}$ be a
homogeneous presentation for $M$. Since none of the generators
represented by $A$ can be non-trivially decomposed, the alphabet $A$
represents a unique minimal generating set for $M$, and any generating
set must contain this minimal generating set. Any two words over $A$
representing the same element of $M$ must be of the same length. So
there is a well-defined function $\lambda : M \to \mathbb{N}^0$ where
$x\lambda$ is defined to be the length of any word over $A$
representing $x$. It is easy to see that $\lambda$ is a homomorphism.
\subsection{Derivation graphs, homotopy bases, and finite derivation type}
Given a monoid presentation $\pres{A}{\rel{R}}$ one builds a (combinatorial) $2$-complex $\mathcal{D}$, called the \emph{Squier complex}, whose $1$-skeleton has vertex set $A^*$ and edges corresponding to applications of relations from $\mathcal{R}$, and that has $2$-cells adjoined for each instance of ``non-overlapping'' applications of relations from $\mathcal{R}$ (see below for a formal definition of non-overlapping relations). There is a natural action of the free monoid $A^*$ on the Squier complex $\mathcal{D}$. A collection of closed paths in $\mathcal{D}$ is called a \emph{homotopy base} if the complex obtained by adjoining cells for each of these paths, and those that they generate under the action of the free monoid on the Squier complex, has trivial fundamental groups. A monoid defined by a presentation is said to have \emph{finite derivation type} (or \fdt\ for short) if the corresponding Squier complex admits a finite homotopy base. Squier \cite{Squier1} proved that the property \fdt\ is independent of the choice of finite presentation, so we may speak of \fdt\ monoids. The original motivation for studying this notion is Squier's result \cite{Squier1} which says that if a monoid admits a presentation by a finite complete rewriting system then the monoid must have finite derivation type. Further motivation for the study of these concepts comes from the fact that the fundamental groups of connected components of Squier complexes, which are called \emph{diagram groups}, have turned out to be a very interesting class of groups; see \cite{Guba}.
In more detail, with any monoid presentation $\mathcal{P} = \pres{A}{\rel{R}}$ we associate a
graph (in the sense of Serre \cite{SerreTrees}) as follows. The
\emph{derivation graph} of $\mathcal{P} $ is an infinite graph $\Gamma =
\Gamma(\mathcal{P}) = (V,E,\iota, \tau, ^{-1})$ with \emph{vertex set} $V =
A^*$, and \emph{edge set $E$} consisting of the collection of
$4$-tuples
\[
\{ (w_1, r, \epsilon, w_2): \ w_1, w_2 \in A^*, r \in R, \ \mbox{and} \ \epsilon \in \{ +1, -1 \} \}.
\]
The functions $\iota, \tau : E \rightarrow V$ associate with each edge
$\mathbb{E} = (w_1, r, \epsilon, w_2)$ (with $r=(r_{+1},r_{-1}) \in R$) its
initial and terminal vertices $\iota \mathbb{E} = w_1 r_{\epsilon} w_2$ and
$\tau \mathbb{E} = w_1 r_{- \epsilon} w_2$, respectively. The mapping $^{-1}
: E \rightarrow E$ associates with each edge $\mathbb{E} = (w_1, r,
\epsilon, w_2)$ an inverse edge $\mathbb{E}^{-1} = (w_1, r, -\epsilon,
w_2)$.
A path is a sequence of edges $\mathbb{P} = \mathbb{E}_1 \circ \mathbb{E}_2 \circ \ldots
\circ \mathbb{E}_n$ where $\tau \mathbb{E}_i = \iota \mathbb{E}_{i+1}$ for $i=1, \ldots,
{n-1}$. Here $\mathbb{P}$ is a path from $\iota \mathbb{E}_1$ to $\tau
\mathbb{E}_n$ and we extend the mappings $\iota$ and $\tau$ to paths
by defining $\iota \mathbb{P} = \iota \mathbb{E}_1$ and $\tau \mathbb{P} = \tau
\mathbb{E}_n$. The inverse of a path $\mathbb{P} = \mathbb{E}_1 \circ \mathbb{E}_2
\circ \ldots \circ \mathbb{E}_n$ is the path $\mathbb{P}^{-1} = \mathbb{E}_n^{-1}
\circ \mathbb{E}_{n-1}^{-1} \circ \ldots \circ \mathbb{E}_1^{-1}$, which is
a path from $\tau \mathbb{P}$ to $\iota \mathbb{P}$. A \emph{closed path} is
a path $\mathbb{P}$ satisfying $\iota \mathbb{P} = \tau \mathbb{P}$. For two paths
$\mathbb{P}$ and $\mathbb{Q}$ with $\tau \mathbb{P} = \iota \mathbb{Q}$ the composition
$\mathbb{P} \circ \mathbb{Q}$ is defined.
We denote the set of paths in $\Gamma$ by $P(\Gamma)$, where for each
vertex $w \in V$ we include a path $1_w$ with no edges, called the
\emph{empty path} at $w$.
The free monoid
$A^*$ acts on both sides of the set of edges $E$ of $\Gamma$ by
\[
x \cdot \mathbb{E} \cdot y = (x w_1, r, \epsilon, w_2 y)
\]
where $\mathbb{E} = (w_1, r, \epsilon, w_2)$ and $x, y \in A^*$. This
extends naturally to a two-sided action of $A^*$ on $P(\Gamma)$ where
for a path $\mathbb{P} = \mathbb{E}_1 \circ \mathbb{E}_2 \circ \ldots \circ \mathbb{E}_n$ we
define
\[
x \cdot \mathbb{P} \cdot y = (x \cdot \mathbb{E}_1 \cdot y) \circ (x \cdot \mathbb{E}_2 \cdot y)
\circ \ldots \circ (x \cdot \mathbb{E}_n \cdot y).
\]
If $\mathbb{P}$ and $\mathbb{Q}$ are paths such that $\iota \mathbb{P} = \iota \mathbb{Q}$ and
$\tau \mathbb{P} = \tau \mathbb{Q}$ then we say that $\mathbb{P}$ and $\mathbb{Q}$ are
\emph{parallel}, and write $\mathbb{P} \parallel \mathbb{Q}$. We use $\parallel\;
\subseteq P(\Gamma) \times P(\Gamma)$ to denote the set of all pairs
of parallel paths.
An equivalence relation $\sim$ on $P(\Gamma)$ is called a
\emph{homotopy relation} if it is contained in $\parallel$ and
satisfies the following four conditions.
\begin{figure}
\centering
\begin{tikzcd}
\null & \iota \mathbb{E}_1 \iota \mathbb{E}_2 \arrow[swap]{dl}{\mathbb{E}_1 \cdot \iota \mathbb{E}_2} \arrow{dr}{\iota \mathbb{E}_1 \cdot \mathbb{E}_2} & \\
\tau \mathbb{E}_1 \iota \mathbb{E}_2 \arrow[swap]{dr}{\tau \mathbb{E}_1 \cdot \mathbb{E}_2} & & \iota \mathbb{E}_1 \tau \mathbb{E}_2 \arrow{dl}{\mathbb{E}_1 \cdot \tau \mathbb{E}_2} \\
& \tau \mathbb{E}_1 \tau \mathbb{E}_2 &
\end{tikzcd}
\caption{Condition 1 of a homotopy relation.}
\label{fig_H1}
\end{figure}
\begin{enumerate}
\item If $\mathbb{E}_1$ and $\mathbb{E}_2$ are edges of $\Gamma$, then \[
(\mathbb{E}_1 \cdot \iota \mathbb{E}_2) \circ (\tau \mathbb{E}_1 \cdot \mathbb{E}_2) \sim
(\iota \mathbb{E}_1 \cdot \mathbb{E}_2) \circ (\mathbb{E}_1 \cdot \tau \mathbb{E}_2 ).
\]
\item For any $\mathbb{P}, \mathbb{Q} \in P(\Gamma)$ and $x,y \in A^*$
\[
\mathbb{P} \sim \mathbb{Q} \ \ \mbox{implies} \ \ x \cdot \mathbb{P} \cdot y \sim x \cdot \mathbb{Q} \cdot y.
\]
\item For any $\mathbb{P}, \mathbb{Q}, \mathbb{R}, \mathbb{S} \in P(\Gamma)$ with $\tau \mathbb{R} = \iota \mathbb{P} = \iota \mathbb{Q}$ and $\iota \mathbb{S} = \tau \mathbb{P} = \tau \mathbb{Q}$
\[
\mathbb{P} \sim \mathbb{Q} \ \ \mbox{implies} \ \ \mathbb{R} \circ \mathbb{P} \circ \mathbb{S} \sim \mathbb{R} \circ \mathbb{Q} \circ \mathbb{S}.
\]
\item If $\mathbb{P} \in P(\Gamma)$ then $\mathbb{P} \mathbb{P}^{-1} \sim 1_{\iota \mathbb{P}}$, where $1_{\iota \mathbb{P}}$ denotes the empty path at the vertex $\iota \mathbb{P}$.
\end{enumerate}
The idea behind condition~1 is the following. Suppose that a word
$w$ has two disjoint occurrences of rewriting rules in the sense that
$ w = \alpha r_\epsilon \beta \alpha' r_{\epsilon'}' \beta' $ where
$\alpha, \beta, \alpha', \beta' \in A^*$, $r, r' \in R$ and $\epsilon,
\epsilon' \in \{ -1, +1 \}$. Let $\mathbb{E}_1 = (\alpha, r, \epsilon,
\beta)$ and $\mathbb{E}_2 = (\alpha', r', \epsilon', \beta')$. Then the
paths
\[
\mathbb{P} = (\mathbb{E}_1 \cdot \iota \mathbb{E}_2)\circ (\tau \mathbb{E}_1 \cdot \mathbb{E}_2),
\quad
\mathbb{P}' = (\iota \mathbb{E}_1 \cdot \mathbb{E}_2)\circ (\mathbb{E}_1 \cdot \tau \mathbb{E}_2)
\]
give two different ways of rewriting the word $w = \alpha r_\epsilon
\beta \alpha' r_{\epsilon'}' \beta'$ to the word $w = \alpha
r_{-\epsilon} \beta \alpha' r_{-\epsilon'}' \beta'$, where in $\mathbb{P}$
we first apply the left-hand relation and then the right-hand, while
in $\mathbb{P}'$ the relations are applied in the opposite order; see
\fullref{Figure}{fig_H1}. We want to regard these two paths as being
essentially the same, and this is achieved by condition~1. This is
usually called \defterm{pull-down and push-up}.
For a subset $C$ of $\parallel$, the homotopy relation \emph{$\sim_C$
generated by $C$} is the smallest (with respect to inclusion)
homotopy relation containing $C$. The relation $\sim_0 =
\sim_\varnothing$ generated by the empty set $\varnothing$ is the
smallest homotopy relation. If $\sim_C$ coincides with $\parallel$,
then $C$ is called a \emph{homotopy base} for $\Gamma$. The
presentation $\pres{A}{\rel{R}}$ is said to have \emph{finite
derivation type} if the derivation graph $\Gamma$
of $\pres{A}{\rel{R}}$ admits a finite homotopy base. A finitely
presented monoid $S$ is said to have \emph{finite derivation type}, or
to be \fdt, if some (and hence any by \cite[Theorem~4.3]{Squier1})
finite presentation for $S$ has finite derivation type.
It is not difficult to see that a subset $C$ of $\parallel$ is a homotopy base of $\Gamma$ if and only if the set \[\{(\mathbb{P} \circ \mathbb{Q}^{-1}, 1_{\iota \mathbb{P}}): (\mathbb{P},\mathbb{Q}) \in C \}\] is a homotopy base for $\Gamma$. Thus we say that a set $D$ of circuits is a homotopy base if the corresponding set $\{ (\mathbb{P}, 1_{\iota \mathbb{P}}) : \mathbb{P} \in D \}$ is a homotopy base.
\subsection{Automaticity and biautomaticity}
\begin{definition}
Let $A$ be an alphabet and let $\$$ be a new symbol not in $A$. Define
the mapping $\delta_{\mathrm{R}} : A^* \times A^* \to ((A\cup\{\$\})\times (A\cup
\{\$\}))^*$ by
\begin{align*}
&(u_1\cdots u_m,v_1\cdots v_n) \mapsto\\
&\qquad\qquad\begin{cases}
(u_1,v_1)\cdots(u_m,v_n) & \text{if }m=n,\\
(u_1,v_1)\cdots(u_n,v_n)(u_{n+1},\$)\cdots(u_m,\$) & \text{if }m>n,\\
(u_1,v_1)\cdots(u_m,v_m)(\$,v_{m+1})\cdots(\$,v_n) & \text{if }m<n,
\end{cases}
\intertext{and the mapping $\delta_{\mathrm{L}} : A^* \times A^* \to ((A\cup\{\$\})\times (A\cup \{\$\}))^*$ by}
&(u_1\cdots u_m,v_1\cdots v_n) \mapsto \\
&\qquad\qquad\begin{cases}
(u_1,v_1)\cdots(u_m,v_n) & \text{if }m=n,\\
(u_1,\$)\cdots(u_{m-n},\$)(u_{m-n+1},v_1)\cdots(u_m,v_n) & \text{if }m>n,\\
(\$,v_1)\cdots(\$,v_{n-m})(u_1,v_{n-m+1})\cdots(u_m,v_n) & \text{if }m<n,
\end{cases}
\end{align*}
where $u_i,v_i \in A$.
\end{definition}
\begin{definition}
\label{def:autstruct}
Let $M$ be a monoid. Let $A$ be a finite alphabet representing a set
of generators for $M$ and let $L \subseteq A^*$ be a regular language such
that every element of $M$ has at least one representative in $L$. For
each $a \in A \cup \{\varepsilon\}$, define the relations
\begin{align*}
L_a &= \{(u,v): u,v \in L, {ua} =_M {v}\}\\
{}_aL &= \{(u,v) : u,v \in L, {au} =_M {v}\}.
\end{align*}
The pair $(A,L)$ is an \defterm{automatic structure} for $M$ if
$L_a\delta_{\mathrm{R}}$ is a regular language over $(A\cup\{\$\}) \times
(A\cup\{\$\})$ for all $a \in A \cup \{\varepsilon\}$. A monoid $M$ is
\defterm{automatic}, or \auto, if it admits an automatic structure
with respect to some generating set.
The pair $(A,L)$ is a \defterm{biautomatic structure} for $M$ if
$L_a\delta_{\mathrm{R}}$, ${}_aL\delta_{\mathrm{R}}$, $L_a\delta_{\mathrm{L}}$, and ${}_aL\delta_{\mathrm{L}}$ are regular
languages over $(A\cup\{\$\}) \times (A\cup\{\$\})$ for all $a \in A
\cup \{\varepsilon\}$. A monoid $M$ is \defterm{biautomatic}, or
\biauto, if it admits a biautomatic structure with respect to some
generating set. [Note that \biauto\ implies \auto.]
\end{definition}
Hoffmann \& Thomas have made a careful study of biautomaticity for
semigroups \cite{hoffmann_biautomatic}. They distinguish four notions
of biautomaticity for semigroups that require at least one of
$L_a\delta_{\mathrm{R}}$ and $L_a\delta_{\mathrm{L}}$ and at least one of ${}_aL\delta_{\mathrm{R}}$ and
${}_aL\delta_{\mathrm{L}}$ to be regular. These notions are all equivalent for
groups and more generally for cancellative semigroups
\cite[Theorem~1]{hoffmann_biautomatic} but distinct for semigroups
\cite[Remark~1 \& \S~4]{hoffmann_biautomatic}. \biauto\ clearly
implies all four Hoffmann--Thomas notions of biautomaticity. However,
we shall shortly prove that, within the class of homogeneous
monoids, any of the Hoffmann--Thomas notions of biautomaticity implies
\biauto\ (see \fullref{Proposition}{prop:htbiautoimpliesbiauto}).
In proving that $R\delta_{\mathrm{R}}$ or $R\delta_{\mathrm{L}}$ is regular, where $R$ is
a relation on $A^*$, a useful strategy is to prove that $R$ is a
rational relation (that is, a relation recognized by a finite
transducer \cite[Theorem~6.1]{berstel_transductions}) and then apply
the following result, which is a combination of
\cite[Corollary~2.5]{frougny_synchronized} and
\cite[Proposition~4]{hoffmann_biautomatic}:
\begin{proposition}
\label{prop:rationalbounded}
If $R \subseteq A^* \times A^*$ is rational relation and there is a
constant $k$ such that $\bigl||u|-|v|\bigr| \leq k$ for all $(u,v) \in
R$, then $R\delta_{\mathrm{R}}$ and $R\delta_{\mathrm{L}}$ are regular.
\end{proposition}
Now we shall prove some results on automaticity and biautomaticity for
the class of homogeneous monoids.
Unlike the situation for groups, both automaticity and biautomaticity
for monoids and semigroups are dependent on the choice of generating
set \cite[Example~4.5]{campbell_autsg}. However, for monoids,
biautomaticity and automaticity are independent of the choice of
\emph{semigroup} generating sets
\cite[Theorem~1.1]{duncan_change}. However, in the case of homogeneous
monoids, we do have independence of the choice of generating set:
\begin{proposition}
\label{prop:autochangegen}
Let $M$ be a homogeneous monoid that is \auto\ (respectively,
\biauto). Then for any finite alphabet $C$ representing a generating
set for $M$ there is a language $K$ over $C$ such that $(C,K)$ is an
automatic (respectively, biautomatic) structure for $M$.
\end{proposition}
\begin{proof}
We first consider the case for \auto.
Suppose $(B,L)$ is an automatic structure for $M$. Assume without loss that $(B,L)$ is an
automatic structure with uniqueness for $M$.
Notice that both the alphabet $B$ and the alphabet $C$ must contain a
subalphabet representing the unique minimal generating set of
$M$. Without loss of generality, assume that they both contain the
alphabet $A$ representing this minimal generating set. For each $b \in
B$, let $w_b \in A^*$ be such that $w_b =_M b$. Let $\rel{Q} \subseteq
B^* \times A^*$ be the relation
\[
\bigl\{(b_1b_2\cdots b_k,w_{b_1}w_{b_2}\cdots w_{b_k}) : k \in \mathbb{N}\cup \{0\}, b_i \in B\bigr\}.
\]
It is easy to see that $\rel{Q}$ is a rational relation. Let
\[
K = L \circ \rel{Q} = \bigl\{v \in A^* : (\exists u \in L)\bigl((u,v) \in \rel{Q}\bigr)\bigr\}.
\]
Let $a \in A \cup \{\varepsilon\}$. Then
\begin{align*}
(u,v) \in K_a &\iff u \in K \land v \in K \land ua =_{M} v\\
&\iff (\exists u',v' \in L)((u,u') \in \rel{Q} \land (v,v') \in \rel{Q} \land u'a =_{M} v')\\
&\iff (\exists u',v' \in L)((u,u') \in \rel{Q} \land (v,v') \in \rel{Q} \land (u',v') \in L_a)\\
&\iff (u,v) \in \rel{Q}^{-1} \circ L_a \circ \rel{Q}.
\end{align*}
Hence $K_a = \rel{Q}^{-1} \circ L_a \circ \rel{Q}$ is a rational
relation. Furthermore
\begin{align*}
(u,v) \in K_a &\implies ua =_M v \\
&\implies (ua)\lambda = v\lambda \\
&\implies u\lambda + a\lambda = v\lambda \\
&\implies |u| + 1 = |v|;
\end{align*}
thus $K_a\delta_{\mathrm{R}}$ is a regular language by
\fullref{Proposition}{prop:rationalbounded}.
For $c \in C - A$, let $u = u_1\cdots u_m \in A^*$ be such that $u =
c$; then $K_c\delta_{\mathrm{R}} = (K_{u_1}\circ K_{u_2}\circ\cdots \circ
K_{u_m})\delta_{\mathrm{R}}$ is regular by \cite[Proposition~2.3]{campbell_autsg}
and similarly $K_c\delta_{\mathrm{L}}$ is regular. Hence $(C,K)$ is an automatic
structure for $M$.
For \biauto, assume $(B,L)$ is a biautomatic structure for $M$ and
follow the above reasoning to show that each of the languages
$K_a\delta_{\mathrm{R}}$, ${}_aK\delta_{\mathrm{R}}$, $K_a\delta_{\mathrm{L}}$, and ${}_aK\delta_{\mathrm{L}}$ are regular.
\end{proof}
\begin{proposition}
\label{prop:htbiautoimpliesbiauto}
Let $M$ be a homogeneous monoid, let $B$ be a finite alphabet
representing a generating set for $M$, and let $L$ be a regular
language over $B$ such that for each $b \in B \cup \{\varepsilon\}$, at
least one of $L_b\delta_{\mathrm{R}}$ and $L_b\delta_{\mathrm{L}}$ and at least one of
${}_bL\delta_{\mathrm{R}}$ and ${}_bL\delta_{\mathrm{L}}$ is regular. Then $M$ is \biauto.
\end{proposition}
\begin{proof}
Suppose $L_b\delta_{\mathrm{R}}$ and ${}_bL\delta_{\mathrm{R}}$ are regular; the other cases are
similar. Then $M$ admits a right-biautomatic structure $(B,L)$. Thus
$L_b\delta_{\mathrm{R}}$ and ${}_bL\delta_{\mathrm{R}}$ are regular languages for all $b \in B
\cup \{\varepsilon\}$.
As in the proof of \fullref{Proposition}{prop:autochangegen}, the
alphabet $B$ must contain a subalphabet representing the unique
minimal generating set of $M$. Without loss of generality, assume that
$B$ contains the alphabet $A$ representing this minimal
generating set.
Construct the relation $\rel{Q} \subseteq B^* \times A^*$ as in the
proof of \fullref{Proposition}{prop:autochangegen}. Let $K = L \circ
\rel{Q}$.
Let $a \in A \subseteq B$. Then at least one of $L_a\delta_{\mathrm{R}}$ and
$L_a\delta_{\mathrm{L}}$ and at least one of ${}_aL\delta_{\mathrm{R}}$ and ${}_aL\delta_{\mathrm{L}}$ is
regular. In particular, $L_a$ and ${}_aL$ are rational relations. So
$K_a = \rel{Q}^{-1} \circ L_a \circ \rel{Q}$ and ${}_aK =
\rel{Q}^{-1}\circ {}_aL \circ \rel{Q}$ are rational relations. If
$(u,v)$ is in $K_a$ or ${}_aK$, then $|u| +1 = |v|$. Hence $K_a\delta_{\mathrm{R}}$,
$K_a\delta_{\mathrm{L}}$, ${}_aK\delta_{\mathrm{R}}$, and ${}_aK\delta_{\mathrm{L}}$ are all regular by
\fullref{Proposition}{prop:rationalbounded}. Since $a \in A \cup
\{\varepsilon\}$ was arbitrary, this proves that $(A,K)$ is a
biautomatic structure for $M$.
\end{proof}
Despite the positive results obtained so far, note that \auto\ does
not imply \biauto\ in the class of homogeneous monoids, as
we shall see below in
\fullref{Example}{eg:fcrsfdtnonbiautoauto}.
\section{Fundamental examples}
\label{sec:fundexs}
\subsection{An \fcrs, \fdt, non-\biauto, \auto\ homogeneous monoid}
In this subsection we present a homogeneous monoid that is \fcrs\ and thus
\fdt, is \auto, but is not \biauto. By considering the reversal
semigroup of this example we will get a homogeneous monoid that admits
a finite complete rewriting system but is not automatic.
\begin{example}
\label{eg:fcrsfdtnonbiautoauto}
Let $M$ be the monoid defined by the presentation $\pres{A}{\rel{R}}$,
where $A = \{a,b,c\}$ and $\rel{R}$ consists of the rewriting rules
\begin{align*}
cbab &\rightarrow cbcb, & cbbb &\rightarrow cbcb, & cbca &\rightarrow cacb,\\
cbaa &\rightarrow cbca, & cbba &\rightarrow cbca,\\
caab &\rightarrow cacb, & cabb &\rightarrow cacb,\\
caaa &\rightarrow caca, & caba &\rightarrow caca,
\end{align*}
\end{example}
\begin{proposition}
The monoid $M$ of \fullref{Example}{eg:fcrsfdtnonbiautoauto} is \fcrs.
\end{proposition}
\begin{proof}
The rewriting system $(A,\rel{R})$ is noetherian because every
rewriting rule either decreases the number of non-$c$ symbols or
decreases the number of symbols $b$ to the left of symbols $a$. To see
that it is confluent, notice that the only overlaps are those between
the left-hand side of $cbca \rightarrow cacb$ and the left-hand side of
a rule of the form $caxy \rightarrow cacy$, where $x,y \in
\{a,b\}$. However, we have
\[
cbcaxa \rightarrow \begin{cases} cacbxa \rightarrow cacbca \rightarrow cacacb \\
cbcaca \rightarrow cacbca \rightarrow cacacb
\end{cases}
\]
and
\[
cbcaxb \rightarrow \begin{cases} cacbxb \rightarrow cacbcb \\
cbcacb \rightarrow cacbcb.
\end{cases}
\]
Therefore $(A,\rel{R})$ is confluent.
\end{proof}
\begin{proposition}
The monoid $M$ of \fullref{Example}{eg:fcrsfdtnonbiautoauto} is \auto, but non-\biauto.
\end{proposition}
\begin{proof}
Let $L$ be the language of normal forms of $(A,\rel{R})$. Since
$(A,\rel{R})$ is finite, $L$ is regular. Let $u \in L$. Consider the
following cases separately:
\begin{enumerate}
\item $uc$ must also be in normal form, since no left-hand side of a
rewriting rule ends in $c$. Hence
\[
L_c = \{(u,uc) : u \in L\}.
\]
\item If $ub$ is not in normal form, then $ub$ must end with the
left-hand side of a rewriting rule. Hence $u = u'cxy$ for some $x,y
\in \{a,b\}$, then $ub = u'cxyb \rightarrow u'cxcb$. This word
$u'cxcb$ is in normal form since $u'cx$ (which is a prefix of $u$)
is in normal form and no rewriting rule has left-hand side $cxcb$
for any $x \in \{a,b\}$. Thus
\begin{align*}
L_b ={}& \{(u,ub) : ub \in L\} \\
\cup{}& \{(u'cxy,u'cxcb) : u'cxy \in L, x,y \in \{a,b\}\}.
\end{align*}
\item If $ua$ is not in normal form, then $ua$ must end with the
left-hand side of a rewriting rule and so either $u = u'cbc$ or $u =
u'cxy$ for some $x,y \in \{a,b\}$.
\begin{enumerate}
\item If $u = u''(cb)^\alpha c$ where $\alpha \geq 1$ is maximal. Then
$ua = u''(cb)^\alpha ca \rightarrow^* u''ca(cb)^\alpha$, which is in normal
form since $u''$ and $ca(ab)^\alpha$ are in normal form and the only
left-hand side of a rewriting rule of that ends in $ca$ is $cbca$,
and $\alpha$ is maximal.
\item If $u = u''cay$ then $ua = u''caya \rightarrow u''caca$ and this
word is in normal form since $u''ca$ is in normal form.
\item If $u = u''(cb)^\alpha y$ where $\alpha \geq 1$ is maximal then
$ua = u''(cb)^\alpha ya \rightarrow u''(cb)^\alpha ca \rightarrow^*
u''ca(cb)^\alpha$ and this word is in normal form since $u''ca$ is
in normal form and $\alpha$ is maximal.
\end{enumerate}
Therefore
\begin{align*}
L_a ={}& \{(u,ua) : ua \in L\} \\
\cup{}& \{(u''(cb)^\alpha c, u''ca(cb)^\alpha) : \alpha \in \mathbb{N}, u''(cb)^\alpha c \in L, u'' \notin A^*cb\} \\
\cup{}& \{(u''cay, u''caca) : y \in \{a,b\}, u''cay \in L\} \\
\cup{}& \{(u''(cb)^\alpha y, u''ca(cb)^\alpha) : \\
&\qquad\qquad y \in \{a,b\}, \alpha \in \mathbb{N}, u''(cb)^\alpha y \in L, u'' \notin A^*cb\}.
\end{align*}
\end{enumerate}
Note also that $L_\varepsilon = \{(u,u) : u \in L\}$. It is easy to see
that $L_x$ is a rational relation for any $x \in A \cup
\{\varepsilon\}$. If $(u,v)$ lies in one of these relations, then
$\bigl||u| - |v|\bigr| \leq 1$ and so $L_x\delta_{\mathrm{R}}$ is regular for all $x
\in A \cup \{\varepsilon\}$ by
\fullref{Proposition}{prop:rationalbounded}. Hence $M$ is \auto.
Suppose, with the aim of obtaining a contradiction, that $M$ is
\biauto. Then by \fullref{Proposition}{prop:autochangegen} it admits
a biautomatic structure $(A,L)$. In particular, $({}_cL \circ
{}_cL^{-1})\delta_{\mathrm{R}}$ is regular. Let $n$ be an even number exceeding the
number of states in an automaton recognizing $({}_cL \circ
{}_cL^{-1})\delta_{\mathrm{R}}$. Observe that
\[
{}_cL \circ {}_cL^{-1} = \{(u,v) \in L : cu =_M cv\}.
\]
Notice that $a^n b^{n+1}$ is not represented by any other word over $A$
and similarly $b^na^nb$ is not represented by any other word over
$A$. So $a^nb^{n+1},b^na^nb \in L$. Furthermore,
\begin{align*}
ca^nb^{n+1} &\rightarrow^* (ca)^{n/2}(cb)^{(n/2)+1} \\
\intertext{and}
cb^na^{n}b &\rightarrow^* (cb)^{n/2}(ca)^{n/2}cb \\
&\rightarrow^* (ca)^{n/2}(cb)^{(n/2)+1}
\end{align*}
and so $(a^nb^{n+1},b^na^nb) \in {}_cL \circ {}_cL^{-1}$. Since $n$
exceeds the number of states in an automaton recognizing $({}_cL \circ
{}_cL^{-1})\delta_{\mathrm{R}}$, we can pump within the first $n$ letters of
$(a^nb^{n+1},b^na^nb)\delta_{\mathrm{R}}$ to see that $(a^{n+ik}b^{n+1},b^{n+ik}a^nb)
\in {}_cL \circ {}_cL^{-1}$ for some $k \geq 1$ and for all $i \in \mathbb{N}$. But
\begin{align*}
ca^{n+2k}b^{n+1} &\rightarrow^* (ca)^{n/2+k}(cb)^{(n/2)+1} \\
\intertext{and}
cb^{n+2k}a^{n}b &\rightarrow^* (cb)^{n/2+k}(ca)^{n/2}cb \\
&\rightarrow^* (ca)^{n/2}(cb)^{(n/2)+k+1},
\end{align*}
which is a contradiction. So $M$ is not \biauto.
\end{proof}
\subsection{An \fcrs, \fdt, non-\biauto, non-\auto\ homogeneous monoid}
\begin{definition}
Let $S$ be a semigroup defined by a presentation
$\pres{A}{\rel{R}}$. Denote by $S^\mathrm{rev}$ the semigroup defined by the
presentation $\pres{A}{\rel{R}^\mathrm{rev}}$, where
$\rel{R}^\mathrm{rev}=\{(l^\mathrm{rev},r^\mathrm{rev}): (l,r)\in\rel{R}\}$, that is called
the \defterm{reversal semigroup} of $S$.
\end{definition}
\begin{example}
\label{eg:fcrsfdtnonbiautononauto}
Consider the monoid $M^\mathrm{rev}$ defined by the presentation
$\pres{A}{\rel{R}^\mathrm{rev}}$, where $\pres{A}{\rel{R}}$ is the
presentation defining \fullref{Example}{eg:fcrsfdtnonbiautoauto}.
\end{example}
It is clear that $M^\mathrm{rev}$ is also \fcrs\ and thus \fdt. Now, by
\cite[Lemma 3.4]{hoffmann_notions} we conclude that $M^{\mathrm{rev}}$ is
non-\auto\ and thus non-\biauto.
\subsection{A non-\fcrs, \fdt, \biauto, \auto\ homogeneous monoid}
The following homogeneous monoid was introduced by Katsura and
Kobayashi \cite[Example~3]{katsura_constructing}, who showed that it
is non-\fcrs, but is \fdt. We shall prove that it is \biauto\ and thus
\auto.
\begin{example}
\label{eg:nonfcrsfdtbiautoauto}
Let $A = \{a,b_i,c_i,d_i : i = 1,2,3\}$ and let $\rel{R}$ consist of
the rewriting rules
\begin{align}
b_ia &\rightarrow ab_i && i \in \{1,2,3\}, \label{eq:katsura:ba}\\
c_jb_j &\rightarrow c_1b_1 && j \in \{2,3\}, \label{eq:katsura:cb}\\
b_jd_j &\rightarrow b_1d_1 && j \in \{2,3\}. \label{eq:katsura:bd}
\end{align}
Let $M = \pres{A}{\rel{R}}$. Then $M$ is
\fdt\ \cite[\S~4]{katsura_constructing} but is
non-\fcrs\ \cite[Proposition~3]{katsura_constructing}.
\end{example}
\begin{proposition}
The monoid $M$ of \fullref{Example}{eg:nonfcrsfdtbiautoauto} is \biauto\ and thus \auto.
\end{proposition}
\begin{proof}
Let $\rel{S}$ consist of the following rewriting rules:
\begin{align}
b_ia &\rightarrow ab_i && i \in \{1,2,3\}, \label{eq:katsura2:ba} \\
c_ja^kb_j &\rightarrow c_1a^kb_1 && j \in \{2,3\}, k \in \mathbb{N} \cup \{0\}, \label{eq:katsura2:cab} \\
c_1a^kb_1a^\ell d_j &\rightarrow c_ja^{k+\ell}b_1d_1 && j \in \{2,3\}, k,\ell \in \mathbb{N} \cup \{0\}, \label{eq:katsura2:cabad} \\
b_jd_j &\rightarrow b_1d_1 && j \in \{2,3\}.\label{eq:katsura2:bd}
\end{align}
It is easy to see that every rule in $\rel{S}$ is a consequence of those in $\rel{R}$. In particular, using rules in $\rel{R}$, we have
\begin{multline*}
c_ja^kb_j \leftrightarrow c_ja^{k-1}b_ja \leftrightarrow \ldots \leftrightarrow c_jb_ja^k \\
\leftrightarrow c_1b_1a^k \leftrightarrow c_1ab_1a^{k-1} \leftrightarrow \ldots \leftrightarrow c_1a^kb_1;
\end{multline*}
and
\begin{multline*}
c_1a^kb_1a^\ell d_j \leftrightarrow \ldots \leftrightarrow c_1b_1a^{k+\ell}d_j \\
\leftrightarrow c_jb_ja^{k+\ell}d_j \leftrightarrow \ldots \leftrightarrow c_ja^{k+\ell}b_jd_j \leftrightarrow c_ja^{k+\ell}b_1d_1.
\end{multline*}
The rewriting system $(A,\rel{S})$ is noetherian since rules
\eqref{eq:katsura2:cab} decrease the number of symbols $b_j$ with $j
\in \{2,3\}$; rules \eqref{eq:katsura2:cabad} decrease the number of
symbols $d_j$ with $j \in \{2,3\}$ and do not increase the number of
symbols $b_j$ with $j \in \{2,3\}$; rules \eqref{eq:katsura2:bd}
decrease the numbers of symbols $b_j$ with $j \in \{2,3\}$; and rules
\eqref{eq:katsura2:ba} decrease the number of symbols $a$ to the right
of symbols $b_i$. Hence any rewriting using a rule in $\rel{S}$
decreases a word with respect to the right-to-left
length-plus-lexicographic order induced by any ordering of $A$ satisfying
$b_1 < b_j < a < d_1 < d_j$ (for $j \in \{2,3\}$).
To see that the $(A,\rel{S})$ is confluent, notice that there are three
possible overlaps of left-hand sides of rewriting rules: an overlap of
\eqref{eq:katsura2:ba} and \eqref{eq:katsura2:cab}, an overlap of
\eqref{eq:katsura2:bd} and \eqref{eq:katsura2:cab}, and an overlap of \eqref{eq:katsura2:ba} and \eqref{eq:katsura2:cabad}. However, critical
pairs resolve, since
\begin{align*}
c_ja^kb_ja &\rightarrow \begin{cases}
c_ja^{k+1}b_j \rightarrow c_1a^{k+1}b_1 \\
c_1a^kb_1a \rightarrow c_1a^{k+1}b_1
\end{cases}
\intertext{and}
c_ja^kb_jd_j &\rightarrow \begin{cases}
c_ja^kb_1d_1 \\
c_1a^kb_1d_j \rightarrow c_ja^kb_1d_1
\end{cases}
\intertext{and finally}
c_1a^kb_1a^\ell d_j &\rightarrow \begin{cases}
c_1a^{k+\ell}b_1d_1 \\
c_1a^{k+1}b_1a^{\ell-1} d_j \rightarrow c_1a^{k+\ell}b_1d_1.
\end{cases}
\end{align*}
So $(A,\rel{S})$ is confluent.
Let $L$ be the language of $\rel{S}$-irreducible words. That is,
\begin{align*}
L = A^* - A^*\Big(&\{b_1a,b_2a,b_3a\} \cup \{c_2,c_3\}a^*\{b_2,b_3\} \\
&\qquad\cup c_1a^*b_1a^*\{d_2,d_3\} \cup \{b_2d_2,b_3d_3\}\Big)A^*;
\end{align*}
thus $L$ is regular. If $w \in L$ and $x \in A$, then either $wx$ is
irreducible, or $wx$ can be rewritten to an irreducible word by a
single application of a rule from $\rel{S}$. Since a single rewriting
step can be carried out by a synchronous regular automaton, the
languages $L_x\delta_{\mathrm{R}}$ are regular. Similarly the languages ${}_xL\delta_{\mathrm{R}}$
are regular and so $(A,L)$ is a biautomatic structure for $M$.
\end{proof}
\subsection{A non-\fcrs, non-\fdt, \biauto, \auto\ homogeneous monoid}
In this section we will give an example of a homogeneous monoid that
is non-\fdt\ and thus non-\fcrs, but which is \biauto\ and thus \auto.
\begin{example}
\label{eg:nonfcrsnonfdtbiautoauto}
Let $A=\{a,b\}$ and let $\rel{R}$ be the rewriting system on $A\cup\{c\}$ consisting of the three rules:
\[
\begin{array}{lclcll}
K_a &:& ac & \rightarrow & ca & \\
K_b &:& bc & \rightarrow & cb & \\
C &:& cab& \rightarrow & cbb. &
\end{array}
\]
Let ${\mathcal{P}}$ be the presentation $\pres{A\cup\{c\}}{\rel{R}}$.
\end{example}
\begin{theorem}
\label{thm_exampleNotFDT}
Let $M$ be the monoid defined by the presentation $\mathcal{P}$ in \fullref{Example}{eg:nonfcrsnonfdtbiautoauto}.
The monoid $M$:
\begin{enumerate
\item has a set of normal forms $A^* \cup c^+b^*a^*$;
\item is \biauto\ and thus \auto;
\item is non-\fdt\ and thus non-\fcrs.
\end{enumerate}
\end{theorem}
Part~1 of \fullref{Theorem}{thm_exampleNotFDT} will follow from \fullref{Lemma}{lem_infCRS} below, and part~2 is proved in \fullref{Lemma}{lem_bio}. Then, the rest of the subsection will be devoted to proving that $M$ is non-\fdt, thus establishing part~3.
\begin{remark}
The methods we use here to prove that \fullref{Example}{eg:nonfcrsnonfdtbiautoauto} is not-\fdt \ are similar to those used in the proof of \cite[Theorem~1]{Gray2011}. In particular we will use the notion of critical peaks, and resolution of critical peaks, in our proof. We refer the reader to \cite[Section~2]{Gray2011} for the definitions of these concepts, and their connection with complete rewriting systems
and \fdt.
\end{remark}
Let us begin by fixing some of
the notation. We start by
adding to $\mathcal{P}$ infinitely many rules of the form
\[
\overline{C}_u \; : \; cuab \rightarrow cubb \; (u\in A^*)
\]
and denote by $\rel{R}'$ the set of all these rules. Notice first that
$\overline{C}_{\epsilon}$ is precisely the rule $C$ defined above and
that, for any word $u\in A^*$ the words $cuab$ and $cubb$ represent
the same element of the monoid $M$, since in the word $cuab$ we can
use relations of the form $K_x$ to pass the letter $c$ through the word $u$
from left to right,
then replace $cab$ by $cbb$ using the relation $C$, and finally move the $c$ back through $u$ again
from right to left
using
the relations $K_x$.
It follows that the presentations ${\mathcal{P}} = \pres{A\cup\{c\}}{\rel{R}}$ and
$\overline{{\mathcal{P}}}=\pres{A\cup\{c\}}{\rel{R} \cup \rel{R}'}$
are equivalent presentations, in the sense that two words $u, v \in (A\cup\{c\})^*$ are equivalent modulo the relations $\rel{R}$ if and only if they are equivalent modulo the relations $\rel{R} \cup \rel{R}'$. In particular, the monoid $M$ is also defined by the infinite presentation
\[
\overline{{\mathcal{P}}}=\pres{A\cup\{c\}}{\rel{R} \cup \rel{R}'}.
\]
\begin{lemma}
\label{lem_infCRS}
The infinite presentation $\overline{\mathcal{P}}$ is a complete presentation of $M$.
The set of irreducible words with respect to this complete rewriting system is
$A^* \cup c^+b^*a^*$.
\end{lemma}
\begin{proof}
The fact that $\overline{\mathcal{P}}$ is a presentation for $M$ follows from the comments made before the statement of the lemma. By considering the (left-to-right) length-plus-lexicographic ordering
on $\{a,b,c\}^*$ induced by $a>b>c$ one sees that the rewriting system
$\overline{\mathcal{P}}$ is noetherian. Also, the set of irreducible words under this
rewriting system is easily seen to be equal to the set $A^* \cup c^+b^*a^*$.
Finally, to prove that $\overline{\mathcal{P}}$ is confluent it suffices to consider all possible overlaps between left-hand sides of the rewriting rules $K_x$ ($x \in A$) and
$\overline{C}_u$ ($u \in A^*$), showing that all critical peaks arising from these overlaps resolve (see \cite[Section~2]{Gray2011}). There are three different ways in which these rewrite rules can overlap, giving rise to three types of critical peaks, all of which can be resolved; see \fullref{Figure}{Figure_resolutions}. This proves that $\overline{\mathcal{P}}$ is confluent and thus completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem_bio}
The monoid $M$ is \biauto.
\end{lemma}
\begin{proof}
Let $L = A^* \cup c^+b^*a^*$. We will prove that $(A\cup\{c\},L)$ is a
biautomatic structure for $M$. By the previous lemma, $L$ is a
language of unique representatives for $M$, so $L_\varepsilon$ is
regular. Next, notice that
\begin{align*}
L_a ={}& \{(u,ua) : u \in L\}\\
L_b ={}& \{(u,ub) : u \in A^*\} \cup \{(c^ib^ja^k,c^ib^{j+k+1}) : i \geq 1, j,k \geq 0\} \\
L_c ={}& \{(a^k,ca^k) : k \geq 0\} \cup \{(uba^k,cb^{|u|+1}a^k) : u \in A^*, k \geq 0\} \\
&\qquad\cup \{(c^ib^ja^k,c^{i+1}b^ja^k) : i \geq 1, j,k \geq 0\} \\
{}_aL ={}& \{(u,au) : u \in A^*\} \cup \{(c^ia^k,c^ia^{k+1}) : i\geq 1, k\geq 0\} \\
&\qquad\cup \{(c^ib^ja^k,c^ib^{j+1}a^k) : i,j \geq 1, k \geq 0\} \\
{}_bL ={}& \{(u,bu) : u \in A^*\} \cup \{(c^ib^ja^k,c^ib^{j+1}a^k) : i \geq 1, j,k \geq 0\} \\
{}_cL ={}& \{(a^k,ca^k) : k \geq 0\} \cup \{(uba^k,cb^{|u|+1}a^k) : u \in A^*, k \geq 0\} \\
&\qquad\cup \{(c^ib^ja^k,c^{i+1}b^ja^k) : i \geq 1, j,k \geq 0\}.
\end{align*}
All of these relations are rational, and so, since $M$ is homogeneous,
their images under $\delta_{\mathrm{R}}$ and $\delta_{\mathrm{L}}$ are regular by
\fullref{Proposition}{prop:rationalbounded}. So $(A,L)$ is a
biautomatic structure for $M$.
\end{proof}
Let $\Gamma$ denote the
derivation graph of $\mathcal{P}$, and $\overline{\Gamma}$ the derivation graph of $\overline{\mathcal{P}}$. Let $\Gamma_Z$ denote the connected components of $\Gamma$ with
vertex set the set of all words in $A\cup\{c\}$ with at least two occurrences of the letter $c$. Likewise let $\overline{\Gamma}_Z$ be the connected component of
$\overline{\Gamma}$ with the same vertex set as $\Gamma_Z$.
\begin{figure}[t]
\centering
\begin{tabular}{c@{\qquad}c}
\begin{tikzcd}[column sep=1.2em]
\null & xcuab \arrow[swap]{dl}{K_x \cdot uab} \arrow{dr}{x\cdot \overline{C}_{u}} & \\
cxuab \arrow[swap]{dr}{\overline{C}_{xu}} & \fdtpath{(\overline{CT1})} & xcubb \arrow{dl}{K_x\cdot ubb} \\
& cxubb &
\end{tikzcd}
&
\begin{tikzcd}[column sep=1.2em]
\null & cuabc \arrow[swap]{dl}{\overline{C}_{u} \cdot c} \arrow{dr}{cua\cdot K_b} & \\
cubbc \arrow[swap]{d}{cub\cdot K_b} & \fdtpath{(\overline{CT2})} & cuacb \arrow{d}{cu\cdot K_a\cdot b} \\
cubcb \arrow[swap]{dr}{cu\cdot K_b\cdot b } & & cucab \arrow{dl}{\ cu\cdot C} \\
& cucbb &
\end{tikzcd}
\end{tabular}
\begin{tikzcd}[column sep=1.2em]
\null & cuabvab \arrow[swap]{dl}{\overline{C}_{u}\cdot vab} \arrow{dr}{\overline{C}_{uabv}} & \\
cubbvab \arrow[swap]{dr}{\overline{C}_{ubbv}} & \fdtpath{(\overline{CT3})} & cuabvbb \arrow{dl}{\overline{C}_{u}\cdot vbb} \\
& cubbvbb &
\end{tikzcd}
\caption{Resolutions of the critical peaks in the derivation graph $\overline{\Gamma}$ of the presentation $\overline{\mathcal{P}}$. Here $x,y\in A$ and $u,v\in A^*$.}
\label{Figure_resolutions}
\end{figure}
Let us denote by $\overline{\cal C}$ the set of critical circuits of the form $\fdtpath{(\overline{CT1})}$ and $\fdtpath{(\overline{CT3})}$ displayed in \fullref{Figure}{Figure_resolutions} and denote by $\overline{\cal Z}$ the critical circuits of the form $\fdtpath{(\overline{CT2})}$. Observe that the critical circuits in $\overline{\cal Z}$ only appear in $\overline{\Gamma}_Z$ since the words labelling vertices in $\fdtpath{(\overline{CT2})}$ all contain two occurrences of the letter $c$.
The set of critical circuits $\overline{\cal C}\cup \overline{\cal Z}$ forms an infinite homotopy base for $\overline{\Gamma}$ (see \cite[Lemma~3]{Gray2011}).
We now want to use the infinite homotopy base $\overline{\cal C}\cup \overline{\cal Z}$ for $\overline{\Gamma}$ to obtain an infinite homotopy base for $\Gamma$. In order to do this we need to take the critical circuits $\fdtpath{(\overline{CT1})}$--$\fdtpath{(\overline{CT3})}$ and transform them into circuits in the derivation graph $\Gamma$ by replacing each occurrence of an edge $\overline{C}_u$ by a corresponding path in $\Gamma$.
As mentioned above when proving that $\mathcal{P}$ and $\overline{\mathcal{P}}$ are equivalent presentations,
the edges $\overline{C}_u$ can be realized in $\Gamma$ by paths $C_u$ which are defined inductively as follows: we first set ${C}_{\epsilon}$ to be the rule $C$, and then for $u=xu'$, with $x\in A$ and $u'\in A^*$ we set $C_u$ to be the path
\begin{align}
\label{eqn_C}
C_u: \quad
cxu'a b
\xrightarrow{K_x^{-1} \cdot u' a b}
xcu'a b
\xrightarrow{x \cdot {C}_{u'}}
xcu'b b
\xrightarrow{K_x \cdot u' b b}
cxu'b b.
\end{align}
So, $C_u$ is the path in $\Gamma$ from $cuab$ to $cubb$ given by commuting $c$ through $u$ using the relations $K_x$, applying the relation $C$ to transform $ucab$ into $ucbb$, and then commuting $c$ back through $u$ again using the relations $K_x$, ending at the vertex $cubb$.
Now let us define a mapping $\varphi$ from the set of paths $P(\overline{\Gamma })$ in $\overline{\Gamma}$ to the set of paths $P(\Gamma )$ in $\Gamma$.
Let $\varphi: P(\overline{\Gamma })\rightarrow P(\Gamma )$ be the map given by $(\alpha \cdot\overline{C}_u\cdot\beta) \varphi=\alpha \cdot C_u\cdot\beta$, for all $\alpha,\beta\in (A\cup\{c\})^*$ and $u\in A^*$, and defined to be the identity on every other edge of $\overline{\Gamma}$. Let ${\cal C} = (\overline{\cal C})\varphi$ and ${\cal Z}=(\overline{\cal Z})\varphi$.
Since $\overline{\cal C}\cup \overline{\cal Z}$ forms a homotopy base for $\overline{\Gamma}$ it follows that ${\cal C}\cup {\cal Z}$ is an infinite homotopy base for $\Gamma$ (see \cite[Lemma~4]{Gray2011}).
Observe that the infinite homotopy base ${\cal C}\cup {\cal Z}$ for $\Gamma$ is nothing more than the set of circuits in $\Gamma$ obtained by taking the set of circuits $\fdtpath{(\overline{CT1})}$--$\fdtpath{(\overline{CT3})}$ and replacing each occurrence of the edge $\overline{C}_u$ by the path $C_u$ defined in \eqref{eqn_C}. Let us denote this corresponding set of circuits ${\cal C}\cup {\cal Z}$ in $\Gamma$ by $\fdtpath{({CT1})}$--$\fdtpath{({CT3})}$.
The monoid $M$ is presented by the finite presentation $\mathcal{P}$ and associated to $\Gamma$, the derivation graph of $\mathcal{P}$, we have an infinite homotopy base ${\cal C}\cup{\cal Z}$. Notice that if $M$ was \fdt\ then we would have a finite subset ${\cal C}_0\cup{\cal Z}_0$ of ${\cal C}\cup{\cal Z}$ which would be a finite homotopy base for $\Gamma$. Our aim now is to show that this leads to a contradiction, and thus conclude that $M$ is not \fdt. In order to do this we shall now define a mapping from the set of paths $P(\Gamma)$ in $\Gamma$ into the integral monoid ring $\mathbb{Z}M$. Define $\Phi: P(\Gamma) \rightarrow {\mathbb{Z}M}$ to be the unique map which extends the mapping:
\begin{itemize}
\item $(\alpha \cdot K_a \cdot \beta)\Phi = \overline{\alpha}$, \quad $(\alpha \cdot K_b \cdot \beta)\Phi = \overline{\alpha}$, \quad and
\item $(\alpha \cdot C \cdot \beta)\Phi = \overline{\alpha }$,
\end{itemize}
where $\alpha,\beta \in (A\cup\{c\})^*$
and
$\overline{\alpha} \in M$ denotes the element of $M$ represented by the word $\alpha$,
to paths
in such a way that
\[
(\mathbb{P} \circ \mathbb{Q})\Phi = (\mathbb{P})\Phi + (\mathbb{Q})\Phi
\;\; \mbox{and} \;\;
(\mathbb{P}^{-1})\Phi = - (\mathbb{P})\Phi.
\]
The following basic properties of $\Phi$ are then easily verified for all paths $\mathbb{P}, \mathbb{Q} \in P(\Gamma)$ and words $\alpha,\beta \in (A\cup\{c\})^*$:
\begin{enumerate
\item $(\alpha \cdot \mathbb{P} \cdot \beta)\Phi = \overline{\alpha} \cdot(\mathbb{P})\Phi $
\item $(\mathbb{P} \circ \mathbb{P}^{-1})\Phi = 0$
\item $([\mathbb{P},\mathbb{Q}])\Phi = 0$ where
\[
[\mathbb{P},\mathbb{Q}] =
(\mathbb{P} \cdot \iota \mathbb{Q})
\circ
(\tau \mathbb{P} \cdot \mathbb{Q})
\circ
(\mathbb{P}^{-1} \cdot \tau \mathbb{Q})
\circ
(\iota \mathbb{P} \cdot \mathbb{Q}^{-1}).
\]
\item If $\mathbb{P} \sim_0 \mathbb{Q}$ then $(\mathbb{P})\Phi = (\mathbb{Q})\Phi$.
\end{enumerate}
Here, property~4 follows from properties~1, 2 and~3. Note that property~4 implies that $\Phi$ induces a well-defined map on the homotopy classes of paths of $\Gamma$.
In what follows we shall often omit bars from the top of words in the images under $\Phi$ and simply write words from $(A\cup\{c\})^*$ with the obvious intended meaning. Recall that $\overline{\cal C}$ is the set of critical circuits of the form $\fdtpath{(\overline{CT1})}$ and $\fdtpath{(\overline{CT3})}$, and that $\mathcal{C}=(\overline{\cal C})\varphi $ is the corresponding set of circuits in $\Gamma$.
Let $F=\{a,b\}^*$ denote the free monoid on the alphabet $\{a,b\}$.
\begin{lemma}
\label{lem_passtoZG}
If $M$ is \fdt\ then the submodule $\langle (\mathcal{C})\Phi \rangle_{\mathbb{Z}F}$, of the left $\mathbb{Z}F$-module $\mathbb{Z}F$,
generated by $(\mathcal{C})\Phi$
is a finitely generated left $\mathbb{Z}F$-module.
\end{lemma}
\begin{proof}
Assume that $M$ is \fdt\ and therefore $\mathcal{P}$ is \fdt. Since $\mathcal{C}\cup\mathcal{Z}$ is a homotopy base for its derivation graph $\Gamma$,
it follows that there are finite subsets $\mathcal{C}_0 \subseteq \mathcal{C}$ and $\mathcal{Z}_0 \subseteq \mathcal{Z}$ such that
$\mathcal{C}_0\cup \mathcal{Z}_0$ is a finite homotopy base for $\Gamma$. Let $\mathfrak{C} \in \mathcal{C}$ be arbitrary. We claim that $(\mathfrak{C})\Phi \in \langle (\mathcal{C}_0)\Phi \rangle_{\mathbb{Z}F}$. Once established, this will prove the lemma, since $(\mathcal{C}_0)\Phi$ is a finite subset of $\langle (\mathcal{C})\Phi \rangle_{\mathbb{Z}F}$.
By \cite[Lemma~2]{Gray2011}, since $\mathfrak{C}$ is a closed path in $\Gamma$ and
$\mathcal{C}_0\cup\mathcal{Z}_0$ is a homotopy base for $\Gamma$,
we can write
\begin{align}
\label{eqn_pathdecomp}
\mathfrak{C} \sim_0
\mathbb{P}_1^{-1} \circ (\alpha_1 \cdot \mathbb{Q}_1 \cdot \beta_1) \circ \mathbb{P}_1
\circ \cdots \circ
\mathbb{P}_n^{-1} \circ (\alpha_n \cdot \mathbb{Q}_n \cdot \beta_n) \circ \mathbb{P}_n,
\end{align}
where each $\mathbb{P}_i \in P(\Gamma)$, $\alpha_i, \beta_i \in (A\cup\{c\})^*$ and $\mathbb{Q}_i \in (\mathcal{C}_0\cup \mathcal{Z}_0)^{\pm 1}$.
Since the vertices of $\mathfrak{C}$ have exactly one $c$, and all the relations in the presentation $\mathcal{P}$ involve the letter $c$, it follows that $\alpha_i, \beta_i \in A^*$ and $\mathbb{Q}_i\in \mathcal{C}_0$, for all $i\in\{1,\ldots,n\}$. Applying $\Phi$ gives
$$ (\mathfrak{C})\Phi= \alpha_1(\mathbb{Q}_1)\Phi + \cdots +\alpha_n(\mathbb{Q}_n)\Phi\in \langle (\mathcal{C}_0)\Phi \rangle_{\mathbb{Z}F}$$ as claimed.
\end{proof}
To complete our proof, it remains to compute the subset $(\mathcal{C})\Phi$ of $\mathbb{Z}F$ and then prove that the submodule $\langle (\mathcal{C})\Phi \rangle_{\mathbb{Z}F}$ of $\mathbb{Z}F$ is not finitely generated as a left $\mathbb{Z}F$-module, where $F = \{a,b\}^*$.
Recall that $\overline{\cal C}$ is the set of critical circuits of the form $\fdtpath{(\overline{CT1})}$ and $\fdtpath{(\overline{CT3})}$, and that $\mathcal{C}=(\overline{\cal C})\varphi $.
So $\mathcal{C}$ is the set of closed paths $\fdtpath{(CT1)}$ and $\fdtpath{(CT3)}$ obtained by applying the mapping $\varphi$ to the closed paths $\fdtpath{(\overline{CT1})}$ and $\fdtpath{(\overline{CT3})}$, that is, obtained by taking each occurrence of $\overline{C}_u$ and replacing it by the path $C_u$. One can easily deduce from equation \eqref{eqn_C} that for any word $u \in A^*$ we have $C_{u}\Phi = u$. Using this fact, the result of computing $(\mathcal{C})\Phi$ is given in \fullref{Table}{table:circuitimages}.
\begin{table}[ht]
\centering
\begin{tabular}{c@{\hskip 1cm}l}
\toprule
Circuit type of $\mathbb{P}$ & \quad $(\mathbb{P})\Phi$ \\
\midrule
$\fdtpath{(CT1)}$ & $0_{\mathbb{Z}M}$ \\
$\fdtpath{(CT3)}$ & $u(ab-bb)v$ \\
\bottomrule
\end{tabular}
\caption{The images under $\Phi$ of the critical circuits from $\mathcal{C}$.}
\label{table:circuitimages}
\end{table}
\begin{lemma}
The submodule $\langle (\mathcal{C})\Phi \rangle_{\mathbb{Z}F}$ of $\mathbb{Z}F$, where
\[
(\mathcal{C})\Phi=\{ u(ab-bb)v \mid u,v\in A^*\}\cup\{0_{\mathbb{Z}M}\},
\]
is not finitely generated as a left $\mathbb{Z}F$-module, and therefore $M$ is not \fdt.
\end{lemma}
\begin{proof}
Suppose, with the aim of obtaining a contradiction, that $\langle (\mathcal{C})\Phi \rangle_{\mathbb{Z}F}$ is finitely generated as a left $\mathbb{Z}F$-module.
Then there exists a finite subset $X$ of $\langle (\mathcal{C})\Phi \rangle_{\mathbb{Z}F}$ such that $\langle X \rangle_{\mathbb{Z}F}\supseteq (\mathcal{C})\Phi $.
Let $n\in \mathbb{N}$ be the maximal length of a word $v$ where $u(ab-bb)v\in X$.
We shall show that $(ab-bb)a^{n+1}$ belongs to $(\mathcal{C})\Phi$ but not to $\langle X \rangle_{\mathbb{Z}F}$.
Suppose that in $\mathbb{Z}F$ we have \[\displaystyle (ab-bb)a^{n+1}=\sum_{i=1}^{k}\alpha_i(ab-bb)v_i,\] where $\alpha_i, v_i\in A^*$ and $|v_i|\leq n$, for $i=1,\ldots, k$.
Then for some $j\in\{1,\ldots, k\}$ we have either
$bba^{n+1} = \alpha_jabv_j$ or
$bba^{n+1} = \alpha_jbbv_j$ in the free monoid $F=\{a,b\}^*$, which
clearly contradicts the fact that $|v_i| \leq n$.
We conclude that $(\mathcal{C})\Phi$
is not finitely generated as a left $\mathbb{Z}F$-module, and it
then follows from \fullref{Lemma}{lem_passtoZG} that $M$ is not \fdt.
\end{proof}
\section{Free products of homogeneous monoids}
\label{sec_freeproducts}
In \fullref{Section}{sec:fundexs}, we gave four examples of
homogeneous monoids that possess certain combinations of the
properties \fcrs, \fdt, \biauto, and \auto. In this section, we use
free products to construct examples with the remaining possible
combinations of these properties. Note that if monoids $M_1$ and $M_2$
have homogeneous presentations $\pres{A_1}{\rel{R}_1}$ and
$\pres{A_2}{\rel{R}_2}$, then their free $M_1\ast M_2$ is defined by
the presentation $\pres{A_1 \cup A_2}{\rel{R}_1 \cup \rel{R}_2}$ and
is thus also homogeneous.
First we consider the interaction of the free product with
\biauto\ and \auto. It is known that the free product of two monoids
is \auto\ if and only if each of the monoids is \auto\ (\cite[Theorem
6.2]{campbell_autsg}, \cite[Theorem 1.2]{duncan_change}). This
result can be extended to biautomaticity. In this paper, we are
concerned only with homogeneous monoids, and the proofs are simpler in
this case:
\begin{proposition}
\label{prop:homogenousautofreeproduct}
Let $M_1$ and $M_2$ be homogeneous monoids. Then $M_1 \ast M_2$ is
\auto\ (respectively, \biauto) if and only if $M_1$ and $M_2$ are
\auto (respectively, \biauto).
\end{proposition}
\begin{proof}
Let $\pres{A_1}{\rel{R}_1}$ and $\pres{A_2}{\rel{R}_2}$ be homogeneous
presentations for $M_1$ and $M_2$, respectively.
Suppose $M_1 \ast M_2$ is \auto; the proof for \biauto\ is similar. By
\fullref{Proposition}{prop:autochangegen}, $M_1 \ast M_2$ admits an
automatic structure $(A_1 \cup A_2,L)$. Since $M_1$ is homogeneous,
there is no non-empty word over $A_1$ representing the identity of
$M_1$; hence every word in $(A_1\cup A_2)^*$ representing an element
of $M_2$ must lie in $A_2^*$. Thus $(A_2,L \cap A_2^*)$ is an
automatic structure for $M_2$; similarly, $(A_1,L\cap A_1^*)$ is an
automatic structure for $M_1$.
On the other hand, suppose $M_1$ and $M_2$ are \auto; again, the proof
for \biauto\ is similar. Then there are automatic structures
$(A_1,L_1)$ and $(A_2,L_2)$ for $M_1$ and $M_2$ respectively. Assume
furthermore that $(A_1,L_1)$ and $(A_2,L_2)$ are automatic structures
with uniqueness for $M_1$ and $M_2$. Let
\[
L = L_1\bigl((L_2 - \{\varepsilon\})(L_1 - \{\varepsilon\})\bigr)^*L_2.
\]
Following the reasoning in \cite[Proof of
Theorem~6.2]{campbell_autsg}, we see that $(A_1\cup A_2,L)$ is an
automatic structure for $M_1 \ast M_2$.
\end{proof}
Now we consider the interaction of free product with \fcrs\ within the
class of homogeneous monoids.
\begin{theorem}[{\cite[Theorem~D]{pride_rewriting}}]
\label{thm:fcrsfreefactorifnounits}
Let $M_1$ and $M_2$ be monoids. Suppose that $M_2$ has no non-trivial
left- or right-invertible elements. Then if $M_1 \ast M_2$ is \fcrs,
then $M_1$ is \fcrs.
\end{theorem}
\begin{corollary}
\label{corol:fcrsfreeprodhomog}
Let $M_1$ and $M_2$ be homogeneous monoids. Then $M_1 \ast M_2$ is \fcrs\ if and only if
$M_1$ and $M_2$ are \fcrs.
\end{corollary}
\begin{proof}
Suppose $M_1 \ast M_2$ is \fcrs. Since $M_1$ and $M_2$ are
homogeneous, neither contains any non-trivial left- or
right-invertible elements. So $M_1$ and $M_2$ are both \fcrs\ by
\fullref{Theorem}{thm:fcrsfreefactorifnounits}.
Now suppose that $M_1$ and $M_2$ are \fcrs; then they are presented by
finite complete rewriting systems $(A_1,\rel{R}_1)$ and
$(A_2,\rel{R}_2)$, respectively. Assume without loss of generality
that $A_1$ and $A_2$ are disjoint. Then $M_1 \ast M_2$ admits a
presentation via the rewriting system $(A_1 \cup A_2,\rel{R}_1 \cup
\rel{R}_2)$, which is easily seen to be complete.
\end{proof}
Finally, we recall the following result on the interaction of the free
product and \fdt:
\begin{theorem}[\cite{otto_modular,otto_modular_tr}]
\label{thm:fdtfreeprod}
Let $M_1$ and $M_2$ be monoids. Then $M_1 \ast M_2$ is \fdt\ if and
only if $M_1$ and $M_2$ is \fdt.
\end{theorem}
Now it is straightforward to use free products to construct the
remaining examples. In each example,
\fullref{Corollary}{corol:fcrsfreeprodhomog},
\fullref{Theorem}{thm:fdtfreeprod}, and
\fullref{Proposition}{prop:homogenousautofreeproduct} together show
that the free product has the desired properties.
\begin{example}
\label{eg:nonfcrsfdtnonbiautoauto}
Let $M_1$ be the \fcrs, \fdt, non-\biauto, \auto\ homogeneous monoid
from \fullref{Example}{eg:fcrsfdtnonbiautoauto}, and let $M_2$ be the
non-\fcrs, \fdt, \biauto, \auto\ homogeneous monoid from
\fullref{Example}{eg:nonfcrsfdtbiautoauto}. Then the homogeneous
monoid $M_1\ast M_2$ is non-\fcrs, \fdt, non-\biauto, and \auto.
\end{example}
\begin{example}
\label{eg:nonfcrsfdtnonbiautononauto}
Let $M_1$ be the \fcrs, \fdt, non-\biauto, non-\auto\ homogeneous monoid
from \fullref{Example}{eg:fcrsfdtnonbiautononauto}, and let $M_2$ be the
non-\fcrs, \fdt, \biauto, \auto\ homogeneous monoid from
\fullref{Example}{eg:nonfcrsfdtbiautoauto}. Then the homogeneous
monoid $M_1\ast M_2$ is non-\fcrs, \fdt, non-\biauto, and non-\auto.
\end{example}
\begin{example}
\label{eg:nonfcrsnonfdtnonbiautoauto}
Let $M_1$ be the \fcrs, \fdt, non-\biauto, \auto\ homogeneous monoid
from \fullref{Example}{eg:fcrsfdtnonbiautoauto}, and let $M_2$ be the
non-\fcrs, non-\fdt, \biauto, \auto\ homogeneous monoid from
\fullref{Example}{eg:nonfcrsnonfdtbiautoauto}. Then the homogeneous
monoid $M_1\ast M_2$ is non-\fcrs, non-\fdt, non-\biauto, and \auto.
\end{example}
\begin{example}
\label{eg:nonfcrsnonfdtnonbiautononauto}
Let $M_1$ be the \fcrs, \fdt, non-\biauto, non-\auto\ homogeneous monoid
from \fullref{Example}{eg:fcrsfdtnonbiautononauto}, and let $M_2$ be the
non-\fcrs, non-\fdt, \biauto, \auto\ homogeneous monoid from
\fullref{Example}{eg:nonfcrsnonfdtbiautoauto}. Then the homogeneous
monoid $M_1\ast M_2$ is non-\fcrs, non-\fdt, non-\biauto, and non-\auto.
\end{example}
\fullref{Figure}{fig:examplerelationships} shows the relationship
between the various examples. \fullref{Corollary}{corol:fcrsfreeprodhomog},
\fullref{Theorem}{thm:fdtfreeprod}, and
\fullref{Proposition}{prop:homogenousautofreeproduct} together show that by
taking a free product, one obtains a new monoid whose properties are
given by taking the logical conjunction (that is, the `and' operation)
of corresponding properties. Thus forming a free product corresponds to the
meet operation in the illustrated semilattice. Since the fundamental
examples given in \fullref{Section}{sec:fundexs} correspond to the
elements of this semilattice that have no non-trivial decomposition,
one sees that these examples are the smallest necessary set required
to obtain all examples of every possible combination of properties
using the free product.
\section{From homogeneous to $n$-ary multihomogeneous monoids}
\label{sec_results}
Thus far we have proved that for every possible combination of the
properties \fcrs, \fdt, \biauto, \auto, and their negations, there
exists a homogeneous monoid with exactly those properties. In this
section, we exhibit two constructions that allow us to turn a
homogeneous monoid with a particular combination of these properties
into an $n$-nary multihomogeneous with the same combination of
properties. The upshot of this is that we have the following result,
which follows by combining
\fullref{Theorem}{thm:from_homog_to_multihomog} and
\fullref{Corollary}{cor_multi}.
\begin{theorem}\label{thm_main}
For each possible combination of the properties \fcrs, \fdt, \biauto,
\auto, and their negations, there exists an $n$-ary multihomogeneous
monoid with exactly that combination of properties.
\end{theorem}
\subsection{Rees-ideal-com\-men\-su\-ra\-ble homogeneous mo\-noids}
We begin by introducing a new definition which is inspired by the
notions of abstractly com\-men\-su\-ra\-ble
groups \cite[\S\S~iv.27ff]{delaharpe_geometric} and Rees index for
semigroups. A subsemigroup $T$ of a given semigroup $S$ has finite
Rees index if $S\setminus T$ is finite. In that case the semigroup $S$
is said to be a small extension of $T$, and $T$ a large subsemigroup
of $S$. The main interest is that such semigroups share many important
properties (see \cite{ReesIndexSurvey} for a survey).
\begin{definition}
Two semigroups $S_1$ and $S_2$ are said to be \defterm{Rees-com\-men\-su\-ra\-ble} if there are finite Rees index subsemigroups $T_i\subseteq S_i$ (for $i=1,2$) that are isomorphic.
\end{definition}
Is is easy to verify that Rees-com\-men\-su\-ra\-bility is an equivalence relation on semigroups.
This notion can be naturally extended to ideals.
\begin{definition}
Two semigroups $S_1$ and $S_2$ are said to be \defterm{Rees-ideal-com\-men\-su\-ra\-ble} if there are finite Rees index ideals $U_i\subseteq S_i$ (for $i=1,2$) that are isomorphic.
\end{definition}
The idea behind these notions is that abstract
Rees-ideal-com\-men\-su\-ra\-ble semigroups share many important properties,
such as \fcrs, \fdt, \biauto, and \auto.
\begin{proposition}
\fcrs, \fdt, \biauto, and \auto\ are preserved under abstract Rees-ideal-com\-men\-su\-ra\-bility:
\end{proposition}
\begin{proof}
\begin{enumerate}
\item It is known that \fcrs\ is inherited by small extensions
\cite{Wang_small_extensions} and by large subsemigroups
\cite{Wong2011}. Although the result for small extensions was stated
in the context of monoids, it can be naturally extended to
semigroups. These two results imply that \fcrs\ is preserved under
abstract Rees(-ideal)-com\-men\-su\-ra\-bility.
\item It is known that \fdt\ is inherited by small extensions
\cite{Wang_small_extensions} of monoids, and by large semigroup
\emph{ideals} \cite{Malheiro_large}. We recall that the notion of
finite derivation type was first introduced for monoids, but it was
naturally extended to the semigroup case \cite{Malheiro_Rees}.
The result on small extensions can be easily adapted for the semigroup
case. Indeed, let $T$ be a semigroup and consider the monoid $T^1$
obtained from $T$ by adding an identity. If $T$ is \fdt\ the derivation
graph of $T^1$ can be obtained from the derivation graph of $T$ by
adding an extra connected component with a single vertex corresponding
to the empty word. Thus $T^1$ is \fdt. Now, if $S$ is a small
extension of the semigroup $T$, it turns out that the monoid $S^1$ is
a small extension of the monoid $T^1$. Therefore, by \cite[Theorem
2]{Wang_small_extensions} the monoid $S^1$ is \fdt. But $S$ is a large
ideal of $S^1$, and by \cite[Theorem 1]{Malheiro_large} we conclude
that $S$ is \fdt.
The two results on small extensions and large ideals of semigroups
show that \fdt\ is preserved under abstract
Rees-ideal-com\-men\-su\-ra\-bility.
\item By the analogue for \biauto\ of \cite[Theorem 1.1]{HTR02},
\biauto\ is inherited by small extensions and by large
subsemigroups, and therefore \biauto\ is preserved under
abstract Rees(-ideal)-com\-men\-su\-ra\-bility.
\item By \cite[Theorem 1.1]{HTR02}, \auto\ is inherited by small
extensions and by large subsemigroups, and therefore \auto\ is
preserved under abstract Rees(-ideal)-com\-men\-su\-ra\-bility.
\end{enumerate}
\end{proof}
[\fcrs, \biauto, and \auto\ are actually preserved under abstract
Rees-com\-men\-su\-ra\-bility, but for this paper, only the result for
abstract Rees-ideal-com\-men\-su\-ra\-bility is needed.]
The preceding result is important because, in the case of
(multi)homogeneous and $n$-ary (multi)homogeneous monoids, the
following result holds:
\begin{proposition}
Every finitely presented (multi)homogeneous monoid is
Rees-ideal-commen\-su\-ra\-ble to an $n$-ary (multi)homogeneous
monoid, where $n$ can be chosen arbitrarily as long as it is greater
than or equal to the length of the longest relation in $\rel{R}$.
\end{proposition}
\begin{proof}
Let $S=\pres{A}{\rel{R}}$ be a finite homogeneous presentation of a
monoid $M$. Let $n$ be greater than or equal to the maximum length of
a relation in $\rel{R}$. Let $I=\{[u]\in M : |u|\geq n\}$. Note that
$I$ is an ideal of finite Rees index in $M$.
Let $\rel{R}'=\{(u\ell v,urv): (\ell ,r)\in \rel{R}, u,v\in A^*, |u\ell v|=n\}$. Then $S'=\pres{A}{\rel{R}'}$ is an $n$-ary
(multi)homogeneous presentation.
Let $I'=\{[u]\in S': |u|\geq n\}$. The set $I'$ is a finite Rees index
ideal of $S'$.
Since $\rel{R}'$ is contained in the Thue congruence generated by
$\rel{R}$, we can define a map $\varphi: I' \rightarrow I$, with
$\left([u]_{\rel{R}'}\right)\varphi =[u]_{\rel{R}}$. This mapping is
clearly surjective, but also injective since, for any
$[u]_{\rel{R}'}$, $[v]_{\rel{R}'}\in I'$ such that
$\left([u]_{\rel{R}'}\right)\varphi
=\left([v]_{\rel{R}'}\right)\varphi$, we get
$[u]_{\rel{R}}=[v]_{\rel{R}}$, with $|u|=|v|\geq n$, and therefore
$[u]_{\rel{R}'}=[v]_{\rel{R}'}$. It is routine to check that $\varphi$
is a homomorphism. Thus $I$ and $I'$ are isomorphic finite Rees index
ideals of $M$ and $M'$. So $M$ and $M'$ are Rees-ideal-com\-men\-su\-ra\-ble.
\end{proof}
\begin{corollary}\label{cor_multi}
Let $P$ be a list of properties preserved under abstract
Rees-ideal-com\-men\-su\-ra\-bility. Then there exists a (multi)homogeneous
monoid satisfying every property in $P$ if and only if there exists
an $n$-ary (multi)ho\-mo\-ge\-neous monoid satisfying every property
in $P$.
\end{corollary}
Note that the list of properties $P$ can contain `negative' properties like `not finitely generated'. As an immediate consequence, we obtain the result we require:
\begin{theorem}
\label{thm:from_homog_to_n-ary_homog}
For each possible combination of the properties \fcrs, \fdt, \biauto,
\auto, and their negations, there exists an $n$-ary homogeneous monoid
with exactly that combination of properties.
\end{theorem}
\subsection{Embedding into a multihomogeneous monoid}
Let $\pres{A}{\rel{R}}$ be a presentation, where $A =
\{a_1,\ldots,a_n\}$. Define a homomorphism
\[
\phi : A^* \to \{x,y\}^*, \qquad a_i \mapsto x^{2}y^ixy^{n+1-i}.
\]
\begin{proposition}
If $\pres{A}{\rel{R}}$ is homogeneous, then $\pres{x,y}{\rel{R}\phi}$ is
multihomogeneous.
\end{proposition}
\begin{proof}
Suppose $\pres{A}{\rel{R}}$ is homogeneous. Let $(u,v) \in
\rel{R}$. Then $|u| = |v|$. Since $a_i\phi$ contains $3$ symbols $x$
and $n+1$ symbols $y$ for all $i \in \{1,\ldots,n\}$, it follows that
$|u\phi|_x = 3|u| = 3|v| = |v\phi|_x$ and $|u\phi|_y = (n+1)|u| =
(n+1)|v| = |v\phi|_y$. Hence $\pres{x,y}{\rel{R}\phi}$ is
multihomogeneous.
\end{proof}
The set $A\phi$ is a code in the alphabet $\{x,y\}^*$ in the sense that $A\phi$ is a set of free generators for $(A\phi)^*$.
Furthermore, a word $u\in \{x,y\}^*$ has a unique decomposition of the form $z_0 u_1 z_1 \cdots z_{n-1} u_n z_n$, where $u_i\in (A\phi)^*$ and any other factor of $u$ belonging to $(A\phi)^*$ is a factor of some $u_i$. Notice that $z_0$ and $z_n$ may be empty, and if $u_1,\ldots, u_n$ are non empty then also $z_1,\ldots, z_{n-1}$ are non empty.
\begin{proposition}
\label{prop:homogenousembedding}
The monoid $\pres{A}{\rel{R}}$ embeds into the monoid
$\pres{x,y}{\rel{R}\phi}$ via the map $u \mapsto u\phi$, and the words
over $\{x,y\}$ representing elements of (the image of) $\pres{A}{\rel{R}}$ are
precisely the words in $A^*\phi$.
\end{proposition}
\begin{proof}
Note first that $\phi$ is a well-defined homomorphism from
$\pres{A}{\rel{R}}$ to $\pres{x,y}{\rel{R}\phi}$ since words over $A$
related by the congruence $\cgen{\rel{R}}$ generated by $\rel{R}$ are
obviously mapped to words related by $\cgen{(\rel{R}\phi)}$.
Now let us prove that the map is injective. Let $w,w' \in A^*$ and
suppose that $(w\phi,w'\phi) \in (\rel{R}\phi)^\#$. Then there is a
sequence of elementary transitions
\[
w\phi = t_0 \leftrightarrow t_1 \leftrightarrow \ldots \leftrightarrow t_k = w'\phi,
\]
where $t_i \in \{x,y\}^*$ and $t_{i}$ is obtained from $t_{i-1}$ by
applying a relation in $\rel{R}\phi$ for all $i$. The aim is to prove by
induction on $i$ that $t_i = s_i\phi$ for some $s_i \in A^*$ and $s_i$
is obtained from $s_{i-1}$ by applying a relation in $\rel{R}$ for all
$i$.
This is trivially true for $i=0$. So suppose it is true for $i-1$; we
will prove it for $i$. By the induction assumption, $t_{i-1} =
s_{i-1}\phi$ for some $s_{i-1} \in A^*$. Now, either
\begin{itemize}
\item $t_{i-1} =
p(u\phi)q$ and $t_i = p(v\phi)q$ for some $(u,v)\in R$ and $p,q \in \{x,y\}^*$, or
\item $t_{i-1} =
p(v\phi)q$ and $t_i = p(u\phi)q$ for some $(u,v)\in R$ and $p,q \in \{x,y\}^*$.
\end{itemize}
The two cases are exacly parallel, so assume the first case
holds. Now, $t_{i-1} = s_{i-1}\phi$, where $s_{i-1} \in A^*$. So
$t_{i-1}$ decomposes as a concatenation of words in $A\phi$; that is,
words of the form $x^2y^jxy^{n+1-j}$ for various $j \in
\{1,\ldots,n\}$. Since subwords $x^2$ only occur at the start of such
words in $A\phi$, the $x^2$ at the start of $u\phi$ lies at the start
of a word from $A\phi$ in the decomposition of $t_{i-1}$. Since all
words in $A\phi$ have the same length, $u\phi$ must also finish at the
end of some word from $A\phi$ in the decomposition of $t_{i-1}$. Hence
$p$ and $q$ are (possibly empty) concatenations of words in $A\phi$,
and so $p = p'\phi$ and $q = q'\phi$ for some $p',q' \in A^*$. Thus
$t_i = p(v\phi)q = (p'vq')\phi= s_i\phi$, where $s_i = p'vq' \in
A^*$. Finally, note that $s_i$ is obtained from $s_{i-1}$ by applying
the relation $(u,v)$. This completes the induction step.
Finally, notice that this reasoning also shows that any word related
to $w\phi$ by $\cgen{(\rel{R}\phi)}$ must lie in $A^*\phi$. This
proves the last part of the proposition.
\end{proof}
The mapping $\phi$ can be naturally extended to the corresponding
presentations, and from those to the corresponding associated Squier
complexes. Also, a word $t$ can be obtained from a word $a$ by a
single application of a defining relation in $\rel{R}\phi$ if, and only if,
\begin{align*}
s &= z_0 u_1 z_1 \cdots z_{i-1} (p\phi)z_i\cdots z_{n-1} u_n z_n, \\
t &= z_0 u_1 z_1 \cdots z_{i-1} (q\phi)z_i\cdots z_{n-1} u_n z_n,
\end{align*}
where $q$ can be obtained from $p$ by a single application of a defining relation in $\rel{R}$. Notice that in each
connected component of the Squier complex
$\Gamma(\pres{x,y}{\rel{R}\phi})$, the words $z_0,\ldots,z_n$ as well
as $n$ are all fixed. Let us denote by
\begin{equation}
\label{edge_ripas}
\#u_1\#\cdots \#(\mathbb{E}\phi)_i\#\cdots \#u_n\#
\end{equation}
the edge corresponding to the application of this defining relation from $R\phi$ to obtain $t$ from $s$, where $\mathbb{E}$ is the edge corresponding to the application of the corresponding defining relation in $\rel{R}$ to obtain $q$ from $p$.
Therefore, if $\mathbb{P}$ is a nonempty path in the Squier complex $\Gamma(\pres{x,y}{\rel{R}\phi})$, by using pull-down and push-up, we have
\begin{equation}\label{path_ripas}
\mathbb{P} \sim_0 \mathbb{P}_1\circ \mathbb{P}_2 \circ \cdots \circ\mathbb{P}_n,
\end{equation}
where each $\mathbb{P}_i$ is a path of the form \eqref{edge_ripas}.
\begin{proposition}
The monoid $\pres{A}{\rel{R}}$ is \fdt\ if and only if the monoid
$\pres{x,y}{\rel{R}\phi}$ is \fdt.
\end{proposition}
\begin{proof}
For a set $C$ of closed paths in $\Gamma(\pres{A}{\rel{R}})$, let
$C\phi$ denote the set $\{\mathbb{P}\phi : \mathbb{P}\in C\}$. It is clear that,
given a closed path $\mathbb{P}$ in $\Gamma(\pres{A}{\rel{R}})$, we have
$\mathbb{P}\sim_C 1_{\iota\mathbb{P}}$ if, and only if, $\mathbb{P}\phi\sim_{C\phi}
1_{\iota(\mathbb{P}\phi)}$.
Let $\mathbb{P}$ be a non-empty closed path in
$\Gamma(\pres{x,y}{\rel{R}\phi})$. By \eqref{path_ripas} there are
closed paths $\mathbb{Q}_i$ in $\Gamma(\pres{A}{\rel{R}})$ such that
\[
\mathbb{P} \sim_0 \mathbb{P}_1\circ \mathbb{P}_2 \circ \cdots \circ\mathbb{P}_n,
\]
and each $\mathbb{P}_i$ has the form $\#u_1\#\cdots \#(\mathbb{Q}_i\phi)_i\#\cdots \#u_n\#$.
So if $C$ is a finite homotopy base for $\Gamma(\pres{A}{\rel{R}})$
then $\mathbb{P}_i\sim_{C\phi} 1_{\iota(\mathbb{P}\phi)}$, which in turn implies
that $\mathbb{P}\sim_{C\phi} 1_{\iota(\mathbb{P}\phi)}$. Consequently, $C\phi$ is
a finite homotopy base for $\Gamma(\pres{x,y}{\rel{R}\phi})$.
Conversely, let $D$ be a finite homotopy base (of closed paths) for
$\Gamma(\pres{x,y}{\rel{R}\phi})$. For each $\mathbb{Q}$ in $D$, arguing as
before, there are closed path $\mathbb{Q}_1,\ldots ,\mathbb{Q}_n$ in
$\Gamma(\pres{A}{\rel{R}})$, such that \[ \mathbb{Q} \sim_0 \mathbb{P}_1\circ
\mathbb{P}_2 \circ \cdots \circ\mathbb{P}_n,
\] and $\mathbb{P}_i=\#u_1\#\cdots \#(\mathbb{Q}_i\phi)_i\#\cdots \#u_n\#$. Denote by $C$ the finite set of all $\mathbb{Q}_i$'s, for all $\mathbb{Q}\in D$.
From \cite[Lemma~2]{Gray2011} we conclude that ${\sim_D}={\sim_{C\phi}}$.
Finally, from the first observation, we know that any closed path
$\mathbb{P}$ in $\Gamma(\pres{A}{\rel{R}})$ satisfies $\mathbb{P}\sim_C
1_{\iota\mathbb{P}}$, since $C\phi$ is a homotopy base of
$\Gamma(\pres{x,y}{\rel{R}\phi})$.
\end{proof}
\begin{proposition}
The monoid $\pres{A}{\rel{R}}$ is \auto\ (respectively,
\biauto) if and only if the monoid $\pres{x,y}{\rel{R}\phi}$ is
\auto (respectively, \biauto).
\end{proposition}
\begin{proof}
For convenience, let $M$ be the monoid $\pres{A}{\rel{R}}$ and $N$ be
the monoid $\pres{x,y}{\rel{R}\phi}$. We will prove the result for
\biauto; the result for \auto\ follows by considering
multiplication only on one side.
Suppose that $N$ is \biauto. By
\fullref{Proposition}{prop:autochangegen}, there is a biautomatic
structure $(\{x,y\},K)$ for $N$. By
\fullref{Proposition}{prop:homogenousembedding}, words over $\{x,y\}$
representing elements of the image under $\phi$ of $M$ are prescisely
those in $(A\phi)^*$. So $K \cap (A\phi)^*$ must map onto the image of
$M$.
Let $L = K\phi^{-1}$. The map $\phi$ and its converse $\phi^{-1}$ are
rational relations, so $L \subseteq A^*$ is regular. Since $K \cap
(A\phi)^*$ must map onto the image of $M$, the
language $L$ maps onto $M$. For any $a \in A \cup \{\varepsilon\}$,
\begin{align*}
(u,v) \in L_a &\iff u \in L \land v \in L \land ua =_M v \\
&\iff (\exists u',v' \in K)((u,u') \in \phi \land (v,v') \in \phi \land u'(a\phi) =_N v') \\
&\iff (\exists u',v' \in K)((u,u') \in \phi \land (v,v') \in \phi \land (u',v') \in K_{a\phi}) \\
&\iff (u,v) \in \phi \circ K_{a\phi} \circ \phi^{-1}.
\end{align*}
Since $K_{a\phi}$ is rational relation (since $(\{x,y\},K)$ is an
automatic structure for $N$), and $\phi$ is a rational relation, if
follows that $L_a$ is a rational relation. Since $M$ is homogeneous,
$(u,v) \in L_a \implies \bigl||u|-|v|\bigr| \leq 1$, and so $L_a\delta_{\mathrm{R}}$
and $L_a\delta_{\mathrm{L}}$ are regular by
\fullref{Proposition}{prop:rationalbounded}. Symmetrical reasoning
shows that ${}_aL\delta_{\mathrm{R}}$ and ${}_aL\delta_{\mathrm{L}}$ are regular. Hence $(A,L)$ is
a biautomatic structure for $M$.
Now suppose that $M$ is \biauto. By
\fullref{Proposition}{prop:autochangegen}, there is a biautomatic
structure $(A,L)$ for $M$. Let $J = \{x,y\}^* - L\phi$; so $J$
consists of all words over $\{x,y\}^*$ that represent elements of $N -
M\phi$. Notice in particular that $\varepsilon \notin J$. Let $K = K_1
\cup K_2 \cup K_3 \cup K_4$, where
\begin{align*}
K_1 &= \bigl(J(L\phi -\{\varepsilon\})\bigr)^* \\
K_2 &= \bigl(J(L\phi -\{\varepsilon\})\bigr)^*J \\
K_3 &= (L\phi - \{\varepsilon\})\bigl(J(L\phi -\{\varepsilon\})\bigr)^* \\
K_4 &= \bigl((L\phi -\{\varepsilon\})J\bigr)^*.
\end{align*}
So $K$ consists of alternating products of elements of $J$ and elements of $L\phi - \{\varepsilon\}$. Notice further that $K$ is regular. The aim is to prove that $(\{x,y\},K)$ is a biautomatic structure for $N$.
First, let $w \in \{x,y\}^*$ be a representative of some element of
$N$. Then $w$ is either the empty word, in which case it lies in $K_1$
and so in $L$, or $w$ uniquely factors in one of the four ways
\begin{multline*}
z_0u_1z_1\cdots z_{k-1}u_k,\quad z_0u_1z_1\cdots z_{k-1}u_kz_k,\\
u_1z_1\cdots z_{k-1}u_k,\quad u_1z_1\cdots z_{k-1}u_kz_k
\end{multline*}
where $z_i \in J$ and $u_i \in (A\phi)^* - \{\varepsilon\}$. For each $i$, there is a
word $u'_i \in A^*$ such that $u'_i\phi = u_i$. Since $(A,L)$ is a biautomatic structure for $M$, there is a word $u''_i \in L$ such that $u''_i =_M u'_i$. Let $v_i = u''_i\phi \in L\phi$ for each $i$. Then $v_i =_N u_i$ and so $w$ is equal to one of the four words
\begin{multline*}
z_0v_1z_1\cdots z_{k-1}v_k,\quad z_0v_1z_1\cdots z_{k-1}v_kz_k,\\
v_1z_1\cdots z_{k-1}v_k,\quad v_1z_1\cdots z_{k-1}v_kz_k,
\end{multline*}
which lie, respectively, in $K_1$, $K_2$, $K_3$, and $K_4$. So every element of $N$ has a representative in $K$.
The next step is to prove that $K_t\delta_{\mathrm{R}}, K_t\delta_{\mathrm{L}}, {}_tK\delta_{\mathrm{R}},
{}_tK\delta_{\mathrm{L}}$ are regular for all $t \in \{x,y,\varepsilon\}$. We will
show that $K_t\delta_{\mathrm{R}}$ and $K_t\delta_{\mathrm{L}}$ are regular; the other cases are
proved by symmetrical reasoning.
So let $w \in K$ and consider right-multiplying $w$ by
$t \in \{x,y,\varepsilon\}$. The word $w$ is either empty or it
uniquely factorizes in one of the forms
\begin{multline*}
z_0u_1z_1\cdots z_{k-1}u_k,\quad z_0u_1z_1\cdots z_{k-1}u_kz_k,\\
u_1z_1\cdots z_{k-1}u_k,\quad u_1z_1\cdots z_{k-1}u_kz_k
\end{multline*}
where $z_i \in J$ and $u_i \in L\phi - \{\varepsilon\}$. Consider the
corresponding factorizations of $wt$:
\begin{align*}
z_0u_1z_1\cdots z_{k-1}u_kt &= z_0u_1z_1\cdots z_{k-1}u_kz'_k, \qquad\text{where $z'_k = t \in J$;}\\
z_0u_1z_1\cdots z_{k-1}u_kz_kt &=
\begin{cases}
z_0u_1z_1\cdots z_{k-1}u_kz'_k, \\
\qquad \text{where $z'_k = z_kt$, if $z_kt \in J$;}\\
z_0u_1z_1\cdots z_{k-1}u'_k, \\
\qquad \text{where $(u_k,u'_k) \in L_{(z_kt)\phi^{-1}}\phi$, if $z_kt \in A\phi$} \\
z_0u_1z_1\cdots z_{k-1}u_kz'_ku'_{k+1}, \\
\qquad \text{where $(\varepsilon,u'_{k+1}) \in L_{vt\phi^{-1}}\phi$, if $z_k = z'_kv$,} \\
\qquad\qquad \text{where $z'_k \in J$ and $vt \in A\phi$}
\end{cases}\displaybreak[0]\\
u_1z_1\cdots z_{k-1}u_kt &= u_1z_1\cdots z_{k-1}u_kz'_k, \qquad\text{where $z'_k = t \in J$;}\\
u_1z_1\cdots z_{k-1}u_kz_kt &=
\begin{cases}
u_1z_1\cdots z_{k-1}u_kz'_k, \\
\qquad \text{where $z'_k = z_kt$, if $z_kt \in J$;}\\
u_1z_1\cdots z_{k-1}u'_k, \\
\qquad \text{where $(u_k,u'_k) \in L_{(z_kt)\phi^{-1}}\phi$, if $z_kt \in A\phi$} \\
u_1z_1\cdots z_{k-1}u_kz'_ku'_{k+1}, \\
\qquad \text{where $(\varepsilon,u'_{k+1}) \in L_{vt\phi^{-1}}\phi$, if $z_k = z'_kv$,} \\
\qquad\qquad \text{where $z'_k \in J$ and $vt \in A\phi$.}
\end{cases}
\end{align*}
Hence
\begin{align*}
K_t ={}& \bigl(J_\varepsilon(L\phi -\{\varepsilon\})_\varepsilon\bigr)^*(\varepsilon,t) \\
&\cup \bigl(J_\varepsilon(L\phi -\{\varepsilon\})_\varepsilon\bigr)^*J_t \\
&\cup \bigcup_{\substack{z \in \{x,y\}^*\\zt \in A\phi}}\bigl(J_\varepsilon(L\phi -\{\varepsilon\})_\varepsilon\bigr)^*J_\varepsilon (L\phi -\{\varepsilon\})_{(zt)\phi^{-1}}(z,\varepsilon)\\
&\cup \bigcup_{\substack{v \in \{x,y\}^*\\vt \in A\phi}}\bigl(J_\varepsilon(L\phi -\{\varepsilon\})_\varepsilon\bigr)^*J_v^{-1}\bigl(L_{(vt)\phi^{-1}}\phi \cap \{\varepsilon\} \times \{x,y\}^*\bigr)\displaybreak[0]\\
&\cup (L\phi - \{\varepsilon\})_\varepsilon\bigl(J_\varepsilon(L\phi -\{\varepsilon\})_\varepsilon\bigr)^*(\varepsilon,t) \\
&\cup \bigl((L\phi -\{\varepsilon\})J_\varepsilon\bigr)^*(L\phi -\{\varepsilon\})_\varepsilon J_t \\
&\cup \bigcup_{\substack{z \in \{x,y\}^*\\zt \in A\phi}}\bigl((L\phi -\{\varepsilon\})J_\varepsilon\bigr)^*(L\phi -\{\varepsilon\})_\varepsilon J_\varepsilon (L\phi -\{\varepsilon\})_{(zt)\phi^{-1}}(z,\varepsilon)\\
&\cup \bigcup_{\substack{v \in \{x,y\}^*\\vt \in A\phi}}\bigl((L\phi -\{\varepsilon\})J_\varepsilon\bigr)^*(L\phi -\{\varepsilon\})_\varepsilon J_v^{-1}\bigl(L_{(vt)\phi^{-1}}\phi \cap \{\varepsilon\} \times \{x,y\}^*\bigr).
\end{align*}
Noting that the various unions are all finite, this proves that $K_t$
is a rational relation. Since $N$ is homogeneous, $(u,v) \in
K_t \implies \bigl||u|-|v|\bigr| \leq 1$, and so $K_t\delta_{\mathrm{R}}$ and
$K_t\delta_{\mathrm{L}}$ are regular by \fullref{Proposition}{prop:rationalbounded}.
Hence $(\{x,y\},K)$ is a biautomatic structure for $N$.
\end{proof}
Recall that if a monoid is defined by a homogeneous presentation
$\pres{A}{\rel{R}}$, then the alphabet $A$ represents a unique minimal
generating set, and hence any generating set must contain it.
\begin{proposition}
If $\pres{B}{\rel{Q}}$ is another finite presentation for $\pres{A}{\rel{R}}$, with $A\subseteq B$, then
\[
\mathcal{P}=\pres{x,y,B}{\rel{Q},\ (a\phi,a)\ (\forall a\in A)}
\]
is a finite presentation defining $\pres{x,y}{\rel{R}\phi}$. Moreover,
the presentation $\pres{B}{\rel{Q}}$ is complete if and only if the
presentation $\cal P$ is complete. Hence $\pres{A}{\rel{R}}$ is
\fcrs\ if and only if $\pres{x,y}{\rel{R}\phi}$ is \fcrs.
\end{proposition}
\begin{proof}
For convenience, let $M$ be the monoid $\pres{A}{\rel{R}}$ and $N$ be
the monoid $\pres{x,y}{\rel{R}\phi}$.
Using Tietze transformations we obtain a new presentation for
$N$. Indeed, for each $a\in A$ we insert in the presentation
$\pres{x,y}{\rel{R}\phi}$ a generator $a$ and a relation $(a\phi,a)$,
thus obtaining a Tietze equivalent presentation
\[
\pres{x,y,A}{\rel{R}\phi,\ (a\phi,a)\ (\forall a\in A)}.
\]
Since $\phi$ is an homomorphism and again by Tietze transformations we get another presentation defining the same monoid:
\[
\pres{x,y,A}{\rel{R},\ (a\phi,a)\ (\forall a\in A)}.
\]
As it is possible to obtain from the presentation $\pres{A}{\rel{R}}$
the presentation $\pres{B}{\rel{Q}}$ using finitely many Tietze
transformations, we can obtain from the presentation
$\pres{x,y,A}{\rel{R},\ (a\phi,a)\ (\forall a\in A)}$ the presentation
$\cal P$ by using the same Tietze transformations.
Suppose that $\pres{B}{\rel{Q}}$ is also complete. Observe that
$\rel{Q}$ relates words from the alphabet $B$, and that a relation
from the set $\rel{E}=\{(a\phi,a):a\in A\}$ has left hand side in
$\{x,y\}^*$ and right hand side in $A$. Thus, if
$w\rightarrow_{\rel{Q}} w' \rightarrow_{\rel{E}} w''$, we can find
$\overline{w}$ such that $w\rightarrow_{\rel{E}} \overline{w}
\rightarrow_{\rel{Q}} w''$. Hence, $\rightarrow_{\rel{E}}$
quasi-commutes over $\rightarrow_{\rel{Q}}$, that is,
${\rightarrow_{\rel{Q}}} \circ {\rightarrow_{\rel{E}}} \subseteq
{\rightarrow_{\rel{E}}} \circ {\rightarrow^*_{\rel{Q}\cup\rel{E}}}$. By
\cite[Theorem 1]{BacDers86}, the rewriting system $(\{x,y\} \cup A,
\rel{Q}\cup\rel{E})$ is terminating if, and only if, both $\rel{Q}$
and $\rel{E}$ are terminating. By assumption $\rel{Q}$ is terminating
and it is easy to verify that $\rel{E}$ is also, by length-reduction.
Attending to the previous observation on the relations from $\rel{Q}$ and
$\rel{E}$, we can also deduce that whenever $w'\leftarrow^*_{\rel{Q}} w
\rightarrow^*_{\rel{E}} w''$, there exists $\overline{w}$ such that
$w'\rightarrow^*_{\rel{E}} \overline{w} \leftarrow^*_{\rel{Q}}
w''$. Therefore, the relations $\rightarrow^*_{\rel{E}}$ and
$\leftarrow^*_{\rel{Q}}$ commute, that is, ${\leftarrow^*_{\rel{Q}}}
\circ{\rightarrow^*_{\rel{E}}} \subseteq
{\rightarrow^*_{\rel{E}}}\circ{\leftarrow^*_{\rel{Q}}}$. By \cite[Lemma
2.7.10]{baader_termrewriting}, if $\rel{E}$ and
$\rel{Q}$ are confluent and $\rightarrow^*_{\rel{E}}$ and
$\leftarrow^*_{\rel{Q}}$ commute, then
$\rel{Q}\cup \rel{E}$ is also confluent. Since by assumption $\rel{Q}$
is confluent it remains to show that $\rel{E}$ is confluent.
Observe that, given left hand side of rules in $\rel{E}$, $a_i\phi$,
$a_j\phi$, they only overlap trivially, that is, when
$a_i\phi=a_j\phi$. Since $\phi$ is injective we get
$a_i=a_j$. Therefore, $\rel{E}$ has no critical pairs and thus is also
confluent.
Conversely, suppose that $\cal P$ is a complete presentation. It is
then clear that $\pres{B}{\rel{Q}}$ is terminating. Confluence also
holds from the fact that in $\cal P$ all critical pairs are resolved,
in particular, those arising from relations in $\rel{Q}$. Hence, each
resolution associated to relations from $\rel{Q}$, can only involve
relations from $\rel{Q}$, since left hand sides of rules in $\rel{E}$ belong to
$\{x,y\}^*$. Therefore, $\pres{B}{\rel{Q}}$ is complete.
\end{proof}
Combining the preceding three propositions gives the desired result:
\begin{theorem}
\label{thm:from_homog_to_multihomog}
For each possible combination of the properties \fcrs, \fdt, \biauto,
\auto, and their negations, there exists a multihomogeneous monoid
with exactly that combination of properties.
\end{theorem}
\bibliographystyle{alphaabbrv}
|
\section{Introduction}
\label{sec:introduction}
The Membrane Electrode Assembly (MEA) is the core of a Proton Exchange Membrane Fuel Cell (PEMFC). It consists of two symmetric catalyst layers (CL), placed at the anode and cathode sides and separated by a polymer electrolyte membrane (PEM), and of the gas diffusion layer~\cite{Eikerling2007, Vielstich2003, Weber2004}. Despite the tremendous progresses achieved in the past decades, the PEMFC is not yet largely commercialized. The most significant hurdles for large scale production include reduction of costs, improvement of power density and enhancement of durability~\cite{Borup2007, Peighambardoust2010}. It is currently consensual that further development of PEMFCs implies a directly understanding of the material properties at the molecular level, for each component of the MEA. In particular, regions of crucial importance are the catalyst layers, where different electrochemical reaction mechanisms take place~\cite{Litster2004,Mehta2003}. This includes two half-cell reaction mechanisms: {\em i)} the Hydrogen Oxidation Reaction (HOR), \ce{H2 -> 2H+ + 2e-} at the anode; and {\em ii)} the Oxygen Reduction Reaction (ORR), \ce{O2 + 4H+ + 4e- -> 2H2O} at the cathode~\cite{Markovic2002, Damjanovic1967, DeMorais2011}. The rates of those reaction mechanisms determine the efficiency of electrochemical conversion, which is directly related to the fuel cell performance~\cite{Rinaldo2010,Franco2008}. The most efficient choice of catalyst particles for enhancing reaction rates are Pt-based particles. The high cost associated to the amount of platinum required for the catalyst, particularly at the cathode, is one of the drawbacks of fuel cells~\cite{Stamenkovic2007,Gasteiger2005,Eikerling2007b, Eikerling2009}.
The CL performance also depends on the transport conditions for reactants and products moving from (to) other MEA components from (to) the catalyst surface inside the CL. A good cathode CL performance (similarly for the anode CL) may depend on: transport of protons from the membrane to the catalyst; electron conduction from the current collector to the catalyst; reactant gases from gas channels to the catalyst; and correct removal of water from the catalyst layer~\cite{Eikerling2007b}. In order to meet all requirements, a complex structure with interconnected pores for reactants diffusion, a phase for electron conduction and a path for proton transport must be considered in devising a CL~\cite{Malek2007, Malek2011a, More2006,Xie2010}.
The necessity of having a heterogeneous structure to satisfy all catalyst layer functionalities, implies the quest for new materials design to optimize the distribution of transport media, in order to reduce transport losses and produce the highest current density with a minimum amount of catalyst particles~\cite{Litster2004}. Effective properties mainly depend on the nature of the materials used and fabrication process applied. During the preparation of catalyst layer ink, Pt/C agglomerates, Nafion ionomer and solvent are mixed together. This process is highly empirical and uses poorly controlled processing methods, which are not based on any knowledge of physico-chemical processes at the molecular level~\cite{Wilson1992, Wilson1993, Wilson1995}.
Also, the CL is composed by materials characterized by very heterogeneous wetting properties, i.e., {\em hydrophilic} or {\em hydrophobic} character. The hydrophilicity of the CL plays an important role in fuel cell water management and it can be modified during the fabrication process~\cite{Li2009, Li2010}. Moreover, these wetting properties can be affected during fuel cell operation. The degradation mechanisms for these materials include ripening and compositional changes of catalyst due to corrosion, catalysts poisoning by adsorbed impurities, aging of the proton exchange electrolyte membrane, changes in the hydrophobic/hydrophilic properties of catalyst layer surfaces~\cite{Mashio2010, Borup2007,Chen2006, Wang2009}.
In Ref.~\cite{Borges2013} we introduced a mean-field-like model for the interaction of the hydrated Nafion ionomer with a substrate, characterized by a tunable degree of hydrophilicity. In particular, we focused on transport properties of water molecules in different regions of the film and demonstrated a high degree of heterogeneity.
We also gave a few hints about the dependence of some morphological features on the wetting properties of the substrate~\cite{Borges2013, Borges2013b}. Here, we consider a much more extended set of simulation data and a provide a complete picture of the produced ultra-thin films morphology. We performed a comprehensive Molecular Dynamics (MD) computer simulation investigation of the substrate effects on the ionomer ultra-thin film morphology at different hydration levels, considering as the control parameter the hidrophilicity degree of the substrate. We have analyzed quantitatively morphology and topology of the films, both at the interfaces with the solid support and air, and in the central layers far from the boundaries. We propose a general qualitative scenario for thin-films morphology in different hydration conditions and wetting nature of the support. We finally speculate about possible implications of our work on the optimization of the actual devices.
The paper is organized as follows. In Section~\ref{sec:catalyst layer} we provide an overview of experimental and computer simulation work relevant in the present context. In Sect.~\ref{sec:ionomer-model} we describe the atomistic model used for mimicking the hydrated ionomer and our effective model for the interaction of the ionomer with the substrate. We characterize the wetting properties of the support in terms of a contact angle. We finally give a few details on our computer simulations scheme. More technical details can be found in the Supplementary Information accompanying this paper. In Sect.~\ref{sec:morphology} we report our extended investigation of the morphology, while in Sect.\ref{sec:PEMFC-technology} we focus more in details on both the support/ionomer and ionomer/vacuum interfaces, discussing the implications of our findings on PEMFC technology. Finally, Sect.~\ref{sec:conclusions} contains our conclusions and possible perspectives on further work.
\section{The catalyst layer}
\label{sec:catalyst layer}
The CL structure is formed by platinum nanoparticles dispersed on a carbon matrix with impregnated Nafion ionomer~\cite{Malek2007, Malek2011a, More2006,Xie2010}. Nafion is a perfluorinated polymer which results from the copolymerisation of a tetrafluoroethylene backbone (Teflon) and perfluorovinyl ether groups, terminated by sulfonate group side-chains~\cite{Moore2004}. Nafion is characterized by a highly heterogeneous structure at the nanoscale, due to a spontaneous phase separation of the hydrophobic backbones and hydrophilic sulfonated side chains upon hydration~\cite{Gierke1981, Hsu1983,Yeager1981,Gebel1987, Gebel2000a,Young2002,Rubatat2002,Schmidt-Rohr2008,Elliott2011}. Nafion has been introduced as one of CL constituents for two reasons~\cite{Litster2004}: first, during the fabrication process it acts as a binder, playing an important role on the dispersion of Pt/C aggregates and, as a consequence, on the Pt utilization. Second, during fuel cell operation, it forms an extended proton-conductor network available for proton migration from (to) the membrane to (from) the catalyst sites. Nafion inside CL presents an inhomogeneous and non-continuous phase. It can be found as a well-dispersed ultra-thin film on the surface of carbon supports and Pt particles. Typically, this film is not uniformly distributed and has a thickness spanning the range $\sim 4$ to $20$ nm~\cite{More2006}.
The formation of Nafion ultra-thin films inside the catalyst layer has been analysed in numerous recent studies~\cite{Ma2007, Paul2011, Paul2011a, Paul2013, Wood2009, Dura2009, Masuda2009, Koestner2011, Eastman2012, Nagao2013, Kusoglu2014, Modestino2012, Modestino2013}. Structure and properties of these films significantly differ from those in the ionomer membrane (bulk). A detailed study based on variation of the ionomer film thickness and comparison with the membrane, has shown that some ionomer properties, {\em e. g.}, water uptake, swelling, water diffusion, respond differently to relative humidity. There is a critical thickness of around $60$~nm, where a transition from a bulk-like to confined ionomer is observed~\cite{Eastman2012}. Other experiments in thin-films adsorbed on \ce{Si2O}-terminated surfaces have underlined a proton conductivity which is lower than in the case of the bulk membrane~\cite{Paul2011,Paul2011a}. Also, Atomic Force Microscopy (AFM) experiments have shown that the ionomer orientation depends on the atomic arrangement of the substrate surface~\cite{Masuda2009,Koestner2011}. In the CL the Nafion ionomer is expected to self-organize in different forms, depending on the properties of the substrate. The impact of surface hydrophilicity on the ionomer properties have been recently subject of many studies, and there is experimental evidence that the change of wetting properties of the substrate is sufficient to affect Nafion film morphology~\cite{Modestino2012, Modestino2013, Bass2010, Bass2011}.
Modestino {\em et al.}~\cite{Modestino2012} have investigated the possibility to control structure and properties of Nafion thin films by modifying the wetting properties of the substrate. They prepared Nafion thin-films deposited on hydrophobic (OTS passivated Si) and hydrophilic (silicon) substrates, and investigated the impact of the internal morphology on water uptake. They found that thin films cast on hydrophobic substrates result in parallel orientation of ionomer channels, which retards the absorption of water from humidified environments. In contrast, films prepared on \ce{SiO2} result in isotropic orientation of these domains, thus favoring water adsorption and swelling of the polymer.
Wood {\em et al.}~\cite{Wood2009} observed multilayer structures of Nafion thin films in contact with smooth flat surfaces. These structures consist of separate hydrophobic and hydrophilic domains formed within the Nafion layer, when equilibrated with saturated \ce{D2O} vapor. Any strong interaction between a flat surface and Nafion is likely to lead to the polymer chains lying flat on that surface, which is completely different from any bulk Nafion morphologies proposed so far. When Nafion was in contact with a bare Pt surface, a hydrophobic Nafion region was found to form adjacent to a Pt film. In contrast, when a PtO monolayer was present, the hydrophobic backbone was pushed outward and the hydrophilic side chains were in contact with the PtO surface. These restructuring processes were reversible and strongly influenced by the polymer hydration. Dura {\em et al.}~\cite{Dura2009} performed Neutron Refractometry (NR) measurements in order to investigate the structure of Nafion in contact with \ce{SiO2}, \ce{Au} and \ce{Pt} surfaces. They showed that lamellar structures, composed of thin alternating water-rich and Nafion-rich layers, exist at the interface between \ce{SiO2} and the hydrated Nafion film. However, multilamellar structures do not exist at the Pt/Nafion or Au/Nafion (metallic) interfaces, where a single thin layer rich in water occurs. This difference indicates that Au and Pt surfaces have a lower affinity to the sulfonic acid/water phase than the more hydrophilic \ce{Si2O} surface. These structures where interpreted in terms of an interface-induced ordering of the ribbon-like aggregates or lamellae observed in Small-Angle X-Rays Scattering (SAXS) experiments of bulk Nafion. Therefore, the first Nafion-rich layer could be formed by closely packed ribbons or lamellae, oriented with their faces parallel to the substrate, and with successive layers of increasingly disordered character.
Molecular dynamics (MD) simulations can also provide insights in clarifying nanoscale structure and transport properties of Nafion at interfaces. Despite this evidence, only a few numerical studies have been dedicated to the above issues, partly due to the issue of convincingly parametrizing interaction force fields between Nafion and substrate materials. A few examples are reported in what follows.
Most part of computational work has focused on the behaviour of Nafion in the presence of carbon and platinum based materials~\cite{Balbuena2005,Lamas2006,Liu2008,Selvan2012,Selvan2008}. These simulations showed that Nafion strongly interacts with Pt nanoparticles, mainly through the hydrophilic sulfonic chains. Mashio {\em et al.}~\cite{Mashio2010} analysed the morphology of Nafion ionomer and water in contact with graphite surfaces. Because of the hydrophobic nature of the graphite sheet and ionomer backbones, Nafion ionomer was found to interact with the graphite sheet mainly via the backbones, whereas side chains were oriented away from the graphite sheet and water molecules were adsorbed at the sulfonic acid groups. The effect on structure and transport properties of the functionalization of graphitized carbon sheet with carboxyl ($COOH$) or carboxylate ($COO^-$) groups was also explored. The most significant effect on water and ionomer distributions was shown to come from the graphite sheet functionalized with ion groups. It was observed that the number of water molecules, hydronium ions and sulfonic acid groups in the vicinity of the graphite sheet increases with the application of the ionized functional groups. Overall, the structure and surface properties of carbon supports clearly affect the transport properties of proton and water.
\section{Modelling}
\label{sec:ionomer-model}
\subsection{The ionomer model}
\label{subsec:ionomer-model}
The Nafion polymer, is formed by a hydrophobic polytetrafluorothylene backbone ([\ce{-CF_2-CF_2}]) and intercalated perfluorinated side-chains, which are terminated by a strongly hydrophilic radical sulfonic acid group ($\ce{SO_3H}$). We consider a united-atom representation for $\ce{CF}$, $\ce{CF_2}$ and $\ce{CF_3}$ and a fully atomistic model for the \ce{SO3-} groups in the side-chain~\cite{Urata2005}. This mixed modelling scheme is commonly used to represent Nafion~\cite{Allahyarov2007,Allahyarov2009,Cui2007,Cui2008,Vishnyakov2000,Vishnyakov2001,Liu2010}.
The polymer backbone is formed by a linear chain of $160$ bonded monomers, which corresponds to a (completely extended) length of approximately $24$~nm. $10$ side-chains are uniformly distributed along the backbone. Each side-chain has $11$ atoms and a length of approximately $1$~nm. The spacing between adjacent side-chains has been chosen in order to match an equivalent weight $\sim 1143.05$~g/mol of $\ce{SO_3^-}$, a value typical for commercial Nafion 117.
The simulation starts from a configuration created by randomly placing 20 polymer chains, 200 hydronium ions and the number of water molecules set according to the desired water content $\lambda$. The system was equilibrated after a series of annealing and optimization runs. After the equilibration, trajectories of, at least, $5$~ns were generated for analyses. The total interaction energy of the system is the sum of non-bonded and intramolecular bonded terms. The force field parameters of our model are similar to the ones of the fully atomistic model of Venkatnathan~\cite{Venkatnathan2007} and adapted to the united-atom representation. The polymer backbone is charged neutral, while the sulfonic acid head groups are assumed to be fully ionized (\ce{SO3-} ). In order to preserve charge neutrality, flexible hydronium complexes (\ce{H3O+}) were added, with force field parameters and partial charges taken from~\cite{Kusaka1998}. Water molecules are described by the rigid extended Simple Point Charge (SPC/E) model~\cite{Berendsen1987}. A list of all parameters is given in Table~1 in the Supporting Information.
We tested the reliability of the ionomer model by performing various simulations of hydrated Nafion in the bulk and compared our results with those found in the literature. Our model is able to reproduce the general Nafion morphology and the correct dynamics of water and hydronium ions. For the reader interested, the main results of Nafion membrane are reported in Supporting Information.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{./figure1a.pdf}
\includegraphics[width=0.45\textwidth]{./figure1b.pdf}
\caption{
{\em Top:} 9-3 Lennard-Jones potential function for different values of $\epsilon_{w}$, with $\sigma_{w}=0.32$~nm. Most hydrophilic case on the bottom. {\em Bottom:} Simulated clusters formed by $3500$ water molecules in contact with the support characterized by increasing values of $\epsilon_{w}$. It is evident the increasingly hydrophilic nature of the interaction.
}
\label{fig1:93lj}
\end{figure}
\subsection{The interaction with the support}
\label{subsec:support-model}
The effect of confinement due to the presence of a solid phase characterized by given wetting properties is mimicked by the interaction potential of the ionomer with the support. The hydrophobic or hydrophilic character of a surface is related to nano-scale features, such as structure and polarity~\cite{Giovambattista2007,Castrillon2009,Nijmeijer1990}. Here we have considered a mean-field-like interaction ionomer/substrate, that allows us to precisely control the hydrophilic character of the substrate by using a unique tunable control parameter. This strategy has already been successfully applied in studies of molecular liquids at interfaces, like pure water in contact with perfectly smooth walls~\cite{Scheidler2002,Spohr1988}. All system units interact with an infinite smooth unstructured wall (the support), placed at $z=0$ and parallel to the $xy$-plane, {\em via} a $9\text{--}3$ Lennard Jones potential~\cite{Abraham1977}. This only depends on the distance, $z$, of the unit from the support:
\begin{equation}
V_{w}^\alpha(z)=\epsilon^\alpha_w\left[ \frac{2}{15}\left(\frac{\sigma_w^\alpha}{z}\right)^9-\left(\frac{\sigma_w^\alpha}{z}\right)^3\right]\theta(z_c-z),
\label{eq:wall}
\end{equation}
where $z_c=$~1.5 nm is a cut-off distance and $\theta$ is the Heaviside function.
\begin{table}[t]
\footnotesize
\centering
\begin{tabular}{l r r r r r r}
\hline\hline
$\epsilon_{w}$ (kcal/mole) & 0.125 & $^*$0.25 & 0.5 & $^*$1.0 & $^*$1.5 & $^*$2.0 \\
\textbf{$\theta$} (degrees) & $163.0$ & $151.3$ & $136.3$ & $100.9$ & $69.1$ & $29.7$ \\
\hline\hline
\end{tabular}
\caption{Values of the water droplet contact angles at the indicated values of $\epsilon_{w}$. We indicate with $^*$ the values of $\epsilon_{w}$ which we will consider in our analysis of the supported thin-films.}
\label{tb:contactangle}
\end{table}
The index $\alpha$ identifies complexes ($H_2O$, $H_30^+$, $SO_3^-$) with significant dipolar coupling to the (hydrophilic) support ($\alpha=\text{phyl}$), or units corresponding to the hydrophobic sections of the polymer ($\alpha=\text{phob}$) which, in contrast, interact very mildly. The energy well $\epsilon_w^\text{phob}=0.5$~kcal/mole is fixed and is the typical strength of the interaction of polymer units with a carbon sheet. This choice is justified by the observation that chemical and physical processes occurring at the surface, {\it e.g.} adsorption and chemical reactions in operating PEMFC, can affect surface polarity~\cite{Mashio2010, Giovambattista2007}. These polarity changes do not affect the interaction with the (apolar) backbone monomers in the same way they modify the interaction with water molecules. The impact of the polarity of the substrate is therefore expected to be more important on the wall/water than on wall/ionomer interactions. The hydrophilicity parameter $\epsilon_w^\text{phyl}=\epsilon_w$ is the control parameter, which was systematically varied in the range $0.125$ to $2.0$~kcal/mol. The typical interaction length scale $\sigma_w^\alpha=0.32$~nm in all cases. Examples of the potential of Eq.~(\ref{eq:wall}) at the indicated values of $\epsilon_w$ are shown in Fig.~\ref{fig1:93lj} (top).
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{./figure2a.pdf}
\includegraphics[width=0.45\textwidth]{./figure2b.pdf}
\caption{
{\em a)} Water droplet profiles at the indicated values of $\epsilon_{w}$. The solid lines are the results of the fitting procedure discussed in the text. {\em b)} Contact angles extracted from the droplet profiles. $\theta$ varies linearly in the investigated $\epsilon_{w}$ range.
}
\label{fig2:dropprofile}
\end{figure}
\begin{figure*}[]
\centering
\includegraphics[width=0.96\textwidth]{./figure3.pdf}
\caption{
Lateral views of typical snapshots of hydrated Nafion thin-films at $\lambda=22$, 11 and 6, formed in contact with supports at the indicated values of the contact angle. These range from strongly hydrophobic ($\theta=150^\circ$) to very hydrophilic nature ($\theta=30^\circ$). The typical films thickness is about 4.5~nm. Beads pertaining to backbones are shown in brown, those pertaining to side-chains are in yellow, \ce{SO_3} groups are in red, water molecules in blue and hydronium ions in white.
}
\label{fig3:side}
\end{figure*}
\subsection{Wetting properties of the support and water droplet contact angles}
\label{subsec:contact-angle}
In order to associate a physical meaning to the adopted choice for $\epsilon_{w}$ we have performed additional simulations of water droplets gently deposited on supports described by Eq.~(\ref{eq:wall}) and calculated the corresponding contact angles, $\theta$. By convention, a value of $\theta\le\pi/2$ corresponds to an hydrophilic support, while $\theta>\pi/2$ to an hydrophobic one.
Figure~\ref{fig1:93lj} (bottom) shows typical snapshots of the equilibrated water droplets at the values $\epsilon_{w}=$ 0.25, 1.0, 1.5 and 2.0~kcal/mol. Already from visual inspection, the increasing hydrophilic character of the support is evident. The contact angles can next be estimated by fitting the droplet profiles~\cite{Shi2009,Werder2003}. Droplets profiles for different values of $\epsilon_{w}$ are shown in Fig.~\ref{fig2:dropprofile}~(a). A circular best fit through these points is extrapolated to the wall surface and provides $\theta$. We compute $\theta$ for each value of $\epsilon_{w}$. In Fig.~\ref{fig2:dropprofile}~(b) we plot $\epsilon_w$-dependence of the contact angle, which is linear in the investigated range. The associated contact angles to the $\epsilon_{w}$ are displayed in Table~\ref{tb:contactangle}. We will often refer to these values in what follows.
Altogether, these data prove that our strategy is able to provide us with different scenarios for the wetting character of the substrate, ranging from strongly hydrophobic to very hydrophilic conditions. Note that these values are representative of specific materials studied in the past. For example, computer simulations of water droplets on a platinum surface shows a contact angle $\theta\simeq$~20-30$^\circ$ \cite{Shi2009}. In the case of carbon nanotubes, the contact angle varies in the range 100$^\circ$ to 106$^\circ$, while for graphite from 110$^\circ$ to 115$^\circ$~\cite{Werder2003, Werder2001}.
\section{Morphology of the hydrated ionomer thin-films}
\label{sec:morphology}
In Fig.~\ref{fig3:side} we show typical snapshots of the self-organized ionomer thin-films at the indicated values of hydration level and contact angles. Four hydrophilicity levels have been considered, encompassing very hydrophobic $(\theta \approx 150^\circ)$, intermediate $(\theta \approx 100^\circ)$, hydrophilic ($\theta \approx 70^\circ$) and strongly hydrophilic $(\theta \approx 30^\circ)$ supports. These contact angles correspond to interaction energies $\epsilon_{w}=$ 0.25, 1.0, 1.5 and 2.0~kcal/mol respectively, as detailed in Table~\ref{tb:contactangle}. The water contents considered are 6, 11 and 22. Those values are typical hydration level found in electrodes in fuel cell operation. Side-chains (yellow beads) terminated by the \ce{SO3-} groups (red beads), decorate the interface between the backbone (brown beads) and the hydrophilic domains formed by water molecules and hydronium ions (blue and white beads, respectively). This configuration is typical of the phase-separated structure present in the Nafion membrane (bulk). The films thickness is about 4.5~nm, for all cases. By visual inspection, it is clear that the hydrophilicity of the substrate indeed controls the global morphology of the film. Also, it is evident that morphology and connectivity of the hydrated domains within the film, changes significantly at different values of $\theta$ and $\lambda$. In what follows we report our analysis work and quantify these changes.
\subsection{Mass density distributions}
\label{subsec:mass density}
The structure of the ionomer film is first analysed in terms of the mass density profiles along the $z$-direction, perpendicular to the substrate. In Fig.~\ref{fig4:densityProfile} we show the polymer (left) and water (right) mass density distributions, $\rho_{p}(z)$ and $\rho_w(z)$ respectively, corresponding to snapshots of Fig.~\ref{fig3:side}. These curves clearly show important complementary changes on the distributions of water and polymer, following the value of $\theta$.
We first focus on films on top of strongly hydrophobic surfaces ($\theta=150^\circ$). In the highly hydrated film ($\lambda=22$), at short distances from the surface, {\em i.e.} $z<1$~nm, the presence of polymer is dominant, while $\rho_{w}(z)$ shows almost no presence of water molecules at distances $z<0.5$~nm (Fig.~\ref{fig4:densityProfile} (a) and (b)). In this region, $\rho_{p}(z)$ presents two well defined peaks. In the center of the film, {\em i.e.} at distances $1.0<z<3.0$~nm, $\rho_{w}(z)$ is at its maximum value, while $\rho_{p}(z)$ is at the minimum. This suggests the formation of water domains confined between polymer-rich layers localized at the bottom and on top of the film. When decreasing the degree of hydration ($\lambda=$~11 and 6) this layered structure is less evident and the distribution of the polymer is less localized. As indicated in Fig.~\ref{fig4:densityProfile}~(e) the polymer density profile only has a shallow minimum in the latter case.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{./figure4.pdf}
\caption{
Mass density profiles for polymer ($\rho_{p}(z)$) and water molecules ($\rho_{w}(z)$) in the considered thin-films at $\lambda = 22$, 11 and 6 at the indicated values of the contact angles $\theta$. $z$ is the distance from the support.
}
\label{fig4:densityProfile}
\end{figure}
In the case $\theta=100^\circ$, one starts to observe the presence of water molecules in direct contact with the substrate, as shown by the appearance of a peak in $\rho_{w}(z)$ at very short $z$. This suggests that a threshold exists at a value of the contact angle included in the range $100^\circ\div 150^\circ$, marking a transition from a completely hydrophobic to a mixed hydrophilic/hydrophobic character. In contrast, the polymer density profile shows the intensity of the first peaks are substantially decreased. Therefore, once water molecules start to adsorb at the support, the ionomer self-organizes by increasingly moving upward, and both species populate the substrate. With decreasing $\lambda$ this equilibrium is altered and the presence of polymer on the substrate is still dominant.
In the more hydrophilic cases ($\theta=70^\circ$ and $30^\circ{}$), the fraction of polymer in direct contact with the substrate is strongly reduced. At $\lambda=22$, the presence of ionomer is significant only for distances $z>2.5$~nm due to the presence of a large amount of water on the bottom which pushes the polymer upward, forming an ionomer layer on the upper part of the film. When $\lambda$ is lowered to 6, a significant fraction of the ionomer can be already found at a distance $z\simeq$~1~nm (Fig.~\ref{fig4:densityProfile}~(e)). In contrast, almost no water molecules are found in the middle of the film, in the range $1.0<z<2.5$~nm. This range encompasses the broad peak characterizing the polymer distribution and water molecules are concentrated in the region corresponding to a minimum of the polymer density profile.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{./figure5.pdf}
\caption{
Radial distribution functions calculated from water/water oxygen atoms ($g_{O_wO_w}(r)$) and water/hydronium oxygen atoms ($g_{O_wO_h}(r)$) at $\lambda=$~22, 11 and 6 and at the indicated values of $\theta$. Data for the membrane in the same hydration conditions are also shown, for comparison.
}
\label{fig5:gOw-film}
\end{figure}
For all cases the positions of the two peaks in the vicinity of the wall for both $\rho_p$ and $\rho_w$ (at 0.29 and 0.55~nm for water, and 0.33 and 0.76~nm for polymer, respectively) do not change neither with hydration nor with surface hydrophilicity. The positions of those peaks are directly controlled by the interaction of the chemical units with the wall and, more precisely, by the parameter $\sigma_{w}=$~0.32~nm in Eq.~(\ref{eq:wall}). The relative distances between the two peaks (0.26~nm and 0.46~nm) are comparable with the nearest-neighbours distances between water molecules and between polymer beads and other species, respectively. Also, the oscillations in density profile (layering) are a typical feature of liquids at the interface with smooth walls~\cite{Spohr1989, Lee1994}.
From the above analysis we can conclude that the modulation of the interaction with the support has indeed a strong impact on local density profiles and, as a consequence, on the morphology of the thin-films. Although it is not surprising that the support wetting behaviour grows due to an increasing hydrophilic character, the overall density profiles are complex and extremely variable. A deeper understanding of the morphological features of these thin-films implies a more detailed analysis, that we will discuss in what follows.
\subsection{Radial Distribution Functions}
\label{subsec:gdr}
In this Section, we explore in details the local structure of the thin-films in terms of 3-dimensional partial radial distribution functions, $g_{\alpha\beta}(r)$, between selected chemical species $\alpha$ and $\beta$, for all the investigated systems. The $g_{\alpha\beta}(r)$ are properly normalized to the entire film volume.
Fig.~\ref{fig5:gOw-film} shows the $g_{\alpha\beta}(r)$ for the oxygen atoms pertaining to water/water ($g_{O_wO_w}(r)$) and water/hydronium ($g_{O_wO_h}(r)$). We observe that the positions of the peaks are very similar to those for the membrane, while the intensity of the peaks, decreases when increasing the hydrophilicy of the substrate. The fist coordination number of water molecules around hydronium ions is reduced. For the case of $\lambda=22$, it decreases from 4.37 for $\theta=150^\circ$ and in the bulk, to 3.66 for $\theta=30^\circ$, indicating that a smaller number of water molecules is found in the vicinities of hydronium ions for the films formed on most hydrophilic supports.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{./figure6.pdf}
\caption{
Radial distribution functions for sulphur-sulphur ($g_{SS}(r)$), suphur/water ($g_{SO_w}(r)$), and sulphur/hydronium ($g_{SO_h}(r)$) for $\lambda=$~22, 11 and 6, at the indicated values of $\theta$. Data for the membrane are also shown, for comparison.}
\label{fig6:gSS-film}
\end{figure}
The local structure around the \ce{SO3-} groups is investigated considering the $g_{\alpha\beta}(r)$ of sulphur atoms with sulphur, $g_{SS}(r)$, and water, $g_{SO_w}(r)$, and hydronium, $g_{SO_h}(r)$, oxygen atoms. These data are shown in Fig.~\ref{fig6:gSS-film}. At variance with the cases of water and hydronium discussed above, the $g_{SS}(r)$ calculated for the different films are very different when compared to the bulk case. This effect is accentuated at $\lambda=22$ (Fig.~\ref{fig6:gSS-film}). For $\theta=30^\circ$, the first peak is located at $0.49$~nm and an additional peak exists at $\simeq 0.7$~nm. When the hydrophilicity degree decreases, for $\theta=100^\circ$ and $150^\circ$, the first peak is shifted to $0.58$~nm, while the second one transforms into a shoulder, approaching the structureless $g_{SS}(r)$ found in the membrane. This indicates that the ionomer formed on a hydrophilic support self-organizes in such a way to have the \ce{SO3-} groups at distances smaller than those found for more hydrophobic cases or in the membrane. Consequently the number of \ce{SO3-} ions lying together is larger in the case of $\theta=30^\circ$. A possible conclusion is that for highly hydrated films ($\lambda=22$) the interaction of the film with the substrate transforms a bulk-like local structure, where \ce{SO3-} groups are less constrained and more spaced, into a configuration where the \ce{SO3-} groups form compact ionic domains.
Both $g_{SO_w}(r)$ and $g_{SO_h}(r)$ exhibit strong correlations, similar to what is observed in the bulk (Figs.~\ref{fig6:gSS-film}). The first and second peaks are observed around $0.38$ and $0.60$~nm and these positions do not vary with the hydrophilicity of the support or with the hydration level of the film. Only the amplitude of those peaks show some changes with $\theta$ and $\lambda$. From the first shell coordination number of water molecules and hydronium ions around the sulphur atoms, we found that the number of water molecules surrounding the \ce{SO3-} decreases when the hydrophilicity of the substrate increases, while the opposite trend is observed for the hydronium. As it could be expected, these changes are more evident at $\lambda=22$, with water and hydronium coordination numbers varying respectively from $6.01$ and $1.45$ in the hydrophilic case, to $6.94$ and $0.9$ in the hydrophobic case. These findings are consistent with the picture based on the $g_{SS}(r)$ data. The number of water and hydronium molecules around the sulphur atoms is always correlated with the \ce{SO3-} agglomeration. Indeed, when the sulfonate ions are less agglomerated, they leave more space available for the water molecules to come closer to \ce{SO3-} groups. Consequently, the hydronium ions are increasingly solvated.
In summary, we have observed that for $\theta=30^\circ$ and $70^\circ$ sulphur atoms are found in compact agglomerates. As a consequence, around the \ce{SO3-} groups the number of water molecules decreases and the number of hydronium ions increases. This effect is more evident for the highly hydrated films ($\lambda=22$). We also conclude that the changes between the structure of the film and the membrane increases with the hydration level.
\begin{figure*}[t]
\centering
\includegraphics[width=0.6\textwidth]{./figure7.pdf}
\caption{
Probability distributions of $cos(\phi_{\ce{SO3-}})$, where $\phi_{\ce{SO3-}}$ is the angle formed by the \ce{SO3-} orientation vector $\hat{u}_{\ce{SO3-}}$ and the normal to the support, $\hat{z}$. The distributions are calculated in slabs of thickness $0.3$~nm parallel to the substrate and at the indicated distances from the support, $z$ (in nm). In the first slab, one can observe the inversion of the \ce{SO3-} orientation when decreasing $\theta$, as discussed in the text.
}
\label{fig7:so3orient}
\end{figure*}
\subsection{Molecular orientation profiles}
\label{subsec:orientation}
To further elucidate both global and local features of the deposited thin-films, orientational order of sulfonic acid groups in regions of the films at different distances from the support were extensively investigated. Similar information about the orientational order of water molecules has already been reported in Ref.~\cite{Borges2013}. There, we have shown that the orientation of water molecules is mainly driven by the interaction with the support, similar to the case of water molecules near Lennard-Jones smooth walls~\cite{Spohr1988,Glebov1997,Tatarkhanov2009}.
The orientation of the \ce{SO3-} groups at different distances from the support was quantified as follows: the films have been partitioned into partially overlapping slabs parallel to the support, with a thickness $\delta z=0.3$~nm. In each slab we have calculated the probability distributions $P(cos(\phi_{\ce{SO3-}}))$, with $cos(\phi_{\ce{SO3-}})=\hat{u}_\ce{SO3-}\cdot\hat{z}$. Here, $\hat{z}$ is the unit vector normal to the support and the unit vector $\hat{u}_{\ce{SO3-}}$ is oriented normal to the plane formed by the three oxygen atoms and points toward the sulphur atom. The \ce{SO3-} orientations at different distances from the support are crucial to elucidate the global ionomer orientation. As a reference, for $cos (\phi_{\ce{SO3-}}) = 1$, the three oxygen atoms face the support and lye in the $xy$-plane.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{./figure8.pdf}
\caption{
Average of $cos(\phi_{SO_3^-})$ as function of the distance from the surface in the films at (a) $\lambda=22$, (b) 11 and (c) 6.
}
\label{fig8:avcos-so3}
\end{figure}
In Fig.~\ref{fig7:so3orient} we show $P(cos(\phi_{\ce{SO3-}}))$ for the investigated films at the indicated values of $\theta$, $\lambda$ and distances from the support. Clearly, $P(cos(\phi_{\ce{SO3-}}))$ depends on the hydrophilic degree of the support. Focusing on the first layer, it is evident that in the most hydrophobic ($\theta=150^\circ$) and hydrophilic ($\theta=30^\circ$) cases, the \ce{SO3-} are oriented in opposite directions. In the first case, the side chains are oriented with the sulfonate groups pointing opposite to the substrate, while in the second case, they point toward the substrate. In the intermediate cases, ($\theta=70^\circ$ and $30^\circ$), the $P(cos(\phi_{\ce{SO3-}}))$ are peaked around $-0.5$. Therefore, the three oxygen atoms point in the direction of the ionomer, with the \ce{SO3-} vector forming an angle of about $60^\circ$ with the normal to the support. This orientation corresponds to side-chains aligned horizontally to the substrate. Side-chain orientational configurations parallel and orthogonal to the support are called "standing" and "lying", respectively, and have been also observed in previous simulations of the ionomer placed on top of platinum surfaces~\cite{Cheng2010,Selvan2008}.
When decreasing hydration, the degree of ionomer orientational order decreases. It is interesting to note that, in the case of $\theta=70^\circ$, the side-chains are first found in the lying position at $\lambda=22$ for gradually shifting to standing configurations, at $\lambda=6$. This indicates that water content also plays an important role in determining the side chains orientation. Indeed, in this particular low-$\lambda$ case, most part of water molecules are in contact with the substrate and, consequently, the ionomer self-organizes to maximize the fraction of \ce{SO3-} groups in direct contact with water. Details of the interface between water domains and side-chains will be further discussed below.
The data shown in Fig.~\ref{fig7:so3orient} also show that the \ce{SO3-} groups are characterized by different preferential orientations in different regions within the film. In order to be more specific on this point, the evolution of the average value $\langle cos(\phi_{\ce{SO3-}}) \rangle$ across the film is illustrated in Fig.~\ref{fig8:avcos-so3}. Interestingly, side-chains orientation inversions at particular distances are evident in some conditions. This inversion is particularly clear in the cases corresponding to $\lambda=22$ (Fig.~\ref{fig8:avcos-so3}~(a)) for $\theta=150^\circ$ and $100^\circ$. Here, $\langle cos(\phi_{\ce{SO3-}})\rangle$, which is negative in the regions close to the support, steadily increases across the central region of the film eventually assuming positive values in the regions furthest from the support. Also interesting are the cases of the films at $\lambda=6$ formed on very hydrophilic supports (Fig.~\ref{fig8:avcos-so3}~(c)). For $\theta=70^\circ$ and $30^\circ$, two inversions on the side-chain average orientation are observed. Strong correlations exist in this case with the water density profiles shown in Fig.~\ref{fig4:densityProfile}~(f). Indeed, we observe the minima of $\langle cos(\phi_{\ce{SO3-}})\rangle $ at $z\simeq$~2.25-2.75~nm, which have a significant overlap with the region where water pools have been observed ($z\simeq$~2.5-3.5~nm). This observation additionally supports the idea that side-chain orientation is mainly governed by the non-trivial distribution of water domains inside the film. An other observation originating from the data of Fig.~\ref{fig8:avcos-so3} is that at distances larger than $3$~nm, side-chain sulfonic acid groups always point toward the support, independently of the values of $\theta$ and $ \lambda$. This side-chain alignment on the top of the film is attributed in part to the ionomer/air interface. We will come back to this point in what follows.
In summary, our results demonstrate that the interaction of water molecules with the support determines the side-chains orientation. Indeed, the \ce{SO3-} groups must be embedded in water domains, to minimize the surface tension at the interface between the hydrophobic polymer backbone and water~\cite{Moore2004}. Therefore, although $\theta$ plays a mild role on the orientational properties of water molecules (as we demonstrated in ref.~\cite{Borges2013}), it has indeed a strong impact on side-chains orientation. This information is very important for the following, when we will propose a general qualitative picture for the morphology of supported Nafion thin-films. In the next Section we complete our investigation by characterizing the formation of ionic clusters across the film.
\begin{table}[b]
\centering
\begin{tabular}{ccccc}
\hline \hline
$\lambda\backslash\theta(^\circ)$ & $150$ & $100$ & $70$ & $30$\\
\hline \hline
22 & 1.71 & 1.87 & 3.02 & 2.96\\
11 & 3.38 & 3.10 & 3.65 & 3.33\\
6 & 5.85 & 5.64 & 4.96 & 4.99\\
\hline\hline
\end{tabular}
\caption{
Average \ce{SO3-} groups cluster sizes for the ionomer thin films at the indicated values of hydrophilicity degree $\theta$ and hydration level $\lambda$.}
\label{tb:clustersize}
\end{table}
\begin{figure}[b]
\centering
\includegraphics[width=0.45\textwidth]{./figure9.pdf}
\caption{
Average cluster size (left) and mass density distributions (right) for the \ce{SO3-} groups as function of distance from the support, $z$, at the indicated values of $\lambda$ and $\theta$.}
\label{fig9:ionic-clusterZ}
\end{figure}
\subsection{Formation of ionic clusters}
\label{subsec: ionic-clusters}
Above we have shown that films present different \ce{SO3-} packing features, {\em i.e.}, both coordination numbers and minimum distances between \ce{SO3-} groups (Fig.~\ref{fig6:gSS-film}) change for the different investigated cases. Here we conclude our analysis by focusing on the features of ionic clusters. This information is important for proposing a general picture for the morphology of the supported films in different hydration conditions and for different wetting nature of the substrate. We have identified the ionic clusters by identifying the \ce{SO3-} groups separated by a distance less than a cut-off $r_c=$~0.64~nm. The clustering analysis provide us with the probability distribution of the size of the clusters, i.e., the number of molecules pertaining to the same cluster. If a \ce{SO3-} group has no nearest neighbours within the cut-off distance, it is considered as an isolated cluster of size $1$.
In Table~\ref{tb:clustersize} we show the average cluster size for all the investigated films. At fixed $\theta$, the cluster size decreases when increasing water content, which is an expected effect due to film swelling: an increasing number of water molecules intercalates between adjacent side chains, therefore \ce{SO3-} groups form less compact agglomerates and isolated groups are found with a higher probability. The hydrophilicity degree also impacts the average cluster size in a non-trivial fashion, which possibly depends on the details of the morphology of the considered film. This result seems to be at odds with a visual inspection of the snapshots shown in Fig.~\ref{fig3:side}, where quite extended regions of condensation of \ce{SO3-} groups are evident in particular regions of the films. To better clarify this point, we computed the average clusters size in different regions of the film, as a function of the distance $z$ from the substrate. In Fig.~\ref{fig9:ionic-clusterZ} we plot the average cluster sizes $\langle S_{\ce{SO3-}}(z)\rangle$ (left), together with the sulfonic acid mass density distributions $\rho_{\ce{SO3-}}(z)$ (right). This helps us in underlining the regions where the presence of \ce{SO3-} groups is relevant. For all values of $\lambda$, at $\theta=30^\circ$ and $70^\circ$, the $\langle S_{\ce{SO3-}}(z)\rangle$ curves clearly indicate the formation of very extended clusters at distances larger than $2$~nm from the support, in the top part of the film, closer to the ionomer/air interface. This is consistent with the high \ce{SO3-} mass density in this region. However, we also note that, for the cases $\theta=150^\circ$ and $100^\circ$, the distribution of average cluster sizes does not show any pronounced peak, despite the presence of well defined maxima in the $\rho_{\ce{SO3-}}$ curves. In conclusion, the formation of \ce{SO3-} clusters seems not to be simply determined by the distribution of \ce{SO3-} but is apparently controlled by the details of the morphology of the film. Also, we emphasize that ionic clustering should play a crucial role on water dynamics. In general, \ce{SO3-} group cluster has a strong impact on hydrogen binding between side-chains, and determines both water binding and the different mechanisms of proton transport~\cite{Kreuer2000,Elliott2007}.
\subsection{Water clusters and connectivity}
\label{sec: water-domains}
We now focus on the topology of the domains formed by the water molecules, and investigate both shape and connectivity of the hydrated domains. We have characterized the water mass density distributions in planes parallel to the substrate, by partitioning the film in four slabs of thickness $1.2$~nm and computing the projected water density distributions on the $xy$-plane, averaged over the trajectory. Our data are plot in the form of color maps in Fig.~\ref{fig10:map-water}. Here a lighter color (yellow) identifies regions where water density is higher, while darker color characterizes regions where the presence of ionomer is significant.
\begin{figure}[]
\centering
\includegraphics[width=0.38\textwidth]{./figure10.pdf}
\caption{
Contour plots of water density for $\lambda=$ 22 (a) 11 (b) and 6 (c), calculated in slabs at the different indicated distances from the substrate.
}
\label{fig10:map-water}
\end{figure}
We first consider the maps in Fig.~\ref{fig10:map-water} for the most hydrophobic cases ($\theta=150^\circ$). Water is concentrated in the second and third slabs, and at $\lambda=22$, a quite homogeneous distribution suggests that water molecules form a unique layer parallel to the support and confined by two ionomer layers separated by a distance of about $\sim 2.4$~nm. The side-chains pertaining to the facing ionomer layers point toward the water layer, with Nafion chains adopting a "sandwich" morphology. In contrast, when decreasing water content, the water pool tends to be concentrated in the central region of the film, surrounded by the ionomer. This is particularly evident for $\lambda=6$, where water molecules form an elongated domain and seems to suggest an inverted micelles morphology, with ellipsoidal or cylindrical micelles shape oriented parallel to the substrate. In the intermediate case, $\theta=100^\circ$, although we do not observe any percolating water-rich region that could be considered as a continuous water layer, water can still form extended agglomerates in the three slabs closer to the wall. For $\lambda=6$, these water "pools" are well delimited and seem to be connected in adjacent slabs. We can also observe a few ionomer "barriers" (indicated by the darker color in the middle of the maps) connecting hydrophobic domains in adjacent slabs. At high hydration, $\lambda=22$, the formation of "pools" is less clear, water being quite homogeneously distributed in all regions, with the ionomer well hydrated everywhere.
In the most hydrophilic cases, $\theta=30^\circ$ and $70^\circ$, water distributions are similar, and the largest water domain forms in contact with the substrate, as expected. For $\lambda=22$, the amount of water is also significant in the second slab. This suggests that water forms a thick continuous layer between the substrate and the ionomer which accumulate on the top of the film, at the interface with air. As a result, these films adopt a completely phase-separated bi-layer configuration. When $\lambda$ decreases, water domains become less homogeneous already beyond the first considered layer, and the formation of disconnected pools in the middle of the film is observed. For $\lambda=6$, water is mostly concentrated in the first and third slab, suggesting a morphology with alternated water-poor and water-rich layers. Also, a single narrow water channel forms, directly connecting the two otherwise disconnected water domains. We finally observe that in all cases the fourth furthest slab is not populated by water molecules, consistent with a hydrophobic interface with air, mostly composed by the ionomer backbones with the side-chains pointing toward the substrate~\cite{Bass2010}.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{./figure11.pdf}
\caption{
Qualitative picture of film morphologies, at different values of $\theta$, ranging from highly hydrophobic (top) to very hydrophylic (bottom) and different hydration levels $\lambda$ (high and low hydration on left and right, respectively).
}
\label{fig11:scheme-film}
\end{figure}
\subsection{A qualitative picture for morphology}
\label{subsec:general-morphology}
Based on the analysis presented in the previous Sections we are now in the position to draw a general picture of the morphology of the supported hydrated Nafion thin films, at different hydration levels and for for varying wetting nature of the support. Despite the qualitative nature of our conclusions, this is the most important message of the present work. We schematically represent the expected morphology of the thin-films in the different conditions as cartoons in Fig.~\ref{fig11:scheme-film}. The \ce{SO3-} groups are represented by red beads, side chains by spring-like symbols and polymer backbones by solid black lines. Water pools are the blue domains. In summary, with reference to the wetting character of the support, we classify the typical morphologies in three classes:
{\bf 1.~Hydrophobic} The film at high hydration (left) shows a typical "sandwich" structure, constituted by a sequence of layers of different nature (Fig.~\ref{fig11:scheme-film}~(a)). This is in agreement with the experimental observations of Refs.~\cite{Dura2009,Wood2009}. Nafion backbones are therefore in direct contact with the substrate, with the sulfonic acid groups pointing upward, toward the water domain. On the top of the water pool, a reversed structure sulfonic groups/polymer backbone is observed, with a completely hydrophobic film/air interface. At low water content (right), the ionomer folds around the water domain, forming an inverted-micelle structure, which reminds the experimental observations of Refs.~\cite{Bass2010, Bass2011}. More precisely, in our simulations the ionomer folds into a inverted-micelle cylinders of diameter $\simeq 4$~nm and with the symmetry axis parallel to the support, as one can observe in the water maps in Figure~\ref{fig10:map-water}
{\bf 2.~Intermediate} In this case the ionomer film organizes into a configuration with interconnected water "pools" (Fig~\ref{fig11:scheme-film}~(b)). The film/substrate interface is characterized by both the presence of ionomer and water, while the film/air interface still has a hydrophobic character, with the side-chains of the ionomer pointing toward the substrate. Hydration level mostly impacts the size of water pools, which decreases by decreasing $\lambda$. In general, the local structure of the film in this case is very similar to the case of the membrane and no evident phase separation parallel to the support is present.
{\bf 3.~Hydrophilic} Thin films in contact with very hydrophilic substrates are organized in well-separated water and ionomer layers (Fig~\ref{fig11:scheme-film}~(c)). In high hydration conditions (left ) water floods the substrate and the ionomer accumulate at the top, with the hydrophobic polymer backbone in contact with air. For lower values of $\lambda$ (right), the ionomer approaches the support. This behavior is not driven by a direct interaction with the substrate, but rather indirectly due to the interaction of the side chains with the water layer in contact with the support. In this case the film can adopt a multilamellar configuration with multiple water layers parallel to the substrate and separated by ionomer domains. Adjacent water layers can be locally connected by water channels, which form dynamically but seem to be quite stable. This picture originating from our data is also consistent with the experimental observations of Refs.~\cite{Dura2009,Wood2009}, where the Authors discovered lamellar structures, formed close to hydrophilic substrates and composed of alternating water-rich and Nafion-rich thin layers.
We conclude this Section by observing that in this work we have considered very thin films of about $4.5$~nm and therefore showed that the wetting nature of the support strongly impacts morphology on length scales of the order of a few nanometers. However, we have also underlined that our qualitative picture seems to be in agreement with experimental observations on films of much larger thickness. We therefore conjecture that the structure of real films could be the results of a geometrical tiling, where the local building blocks are morphologies similar to the ones of Fig.~\ref{fig11:scheme-film}. How this tiling extends from the substrate to the ionomer/air interface in real systems is an open issue. In what follows we will discuss how the qualitative features summarized above can be relevant for PEMFC technology.
\section{Nafion thin-films morphology and PEMFC technology}
\label{sec:PEMFC-technology}
In this Section we discuss the relevance of our findings in the understanding of the catalyst layer features, a crucial issue in the PEMFC technology. From our analysis, the ionomer morphology is expected to impact the catalyst layer activity as follows. First, a strong effect can be envisaged on the transport features of water and hydronium complexes close to the catalyst and the catalyst/support interfaces. Indeed, we have shown in our previous publication~\cite{Borges2013} that complex morphology changes can result in a highly heterogeneous transport behaviour of water across the film. In particular, the extent of the heterogeneity seems to be directly controlled by the wetting character of the substrate and increases steadily by increasing the hydrophilicity character of the support~\cite{Borges2013}.
Second, our findings could also be relevant for a better understanding of the ionomer/catalyst interface. This is the region where the electrochemical reactions governing a PEMFC operation take place. In the actual device, two phenomena directly affect the reaction kinetics: adsorption of chemical species and the formation of the electrochemical double layer. Detailed descriptions of these mechanisms are not possible with our level of description, which cannot account for electrochemical activity. We can however speculate about the impact of the ionomer structural organization on these phenomena.
Third, analysis of the (top) film/air interface is relevant to understand its impact on the water and reactant gases transport inside catalyst layer pores (in the CL gas phase). The upper surface of the film plays an important role in the hydrophilicity of the catalyst layer pores, which in turn impacts water management during operation conditions. Moreover, the reactant gases in the gas phase ({\it e.g.} \ce{O2} and \ce{H2}) must cross the film in order to reach the catalyst surface where the reactions take place. Below we will describe the ionomer/air interface and its possible impact on the water and gases absorption and water management.
In what follows we explore these points in details, by characterizing the interfacial regions, {\it i.e.} immediately adjacent to the substrate and at the top of the film. We will first analyse ionomer adsorption and overall substrate coverage for different wetting nature of the support. Next, we will investigate the main features of the charge distribution close to the substrate. Finally, we will characterize the ionomer/air interface.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\textwidth,angle=-90,origin=c]{./figure12.pdf}
\caption{
Snapshots of the adsorption region, which extends to $z\simeq 0.56$~nm from the support. The backbone segments beads are plotted in brown, side-chains hydrophobic segments in yellow, the \ce{SO3-} groups in red, water molecules in blue and hydronium complexes in white.
}
\label{fig12:snapshot-adsorptionRegion}
\end{figure}
\subsection{Ionomer adsorption}
\label{subsec:film-wall}
In the CL, the catalyst (Pt and/or Pt-alloy) surfaces can react with water, hydronium ions or other chemical species~\cite{Subbaraman2010}. Although in this work electrochemical reactivity of the substrate is not accounted for, we are in the position to characterize the overall surface coverage. This should depend on the details of the ionomer distribution immediately adjacent to the substrate, which corresponds to the first peak in the mass density profiles of Fig.~\ref{fig4:densityProfile}. In Fig.~\ref{fig12:snapshot-adsorptionRegion} we show typical snapshots of the adsorption region, which extends to $z\simeq 0.56$~nm from the support. In the case of hydrophobic substrates, $\theta=150^\circ$, and at any degree of hydration, the ionomer is adsorbed via the backbone, as also observed in simulations of an ionomer adsorbed on graphitized carbon sheets~\cite{Mashio2010}. For the case of intermediate hydrophilicity, $\theta=100^\circ$, a more balanced presence of water, backbone segments and side-chains is observed. In the most hydrophilic cases, $\theta=70^\circ$ and $30^\circ$, limited adsorption of the ionomer is still observed, which takes place via the sulfonate groups (red beads in Fig.~\ref{fig12:snapshot-adsorptionRegion}).
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{./figure13.pdf}
\caption{
Ionomer backbone, water, side-chains, \ce{SO3-} and hydronium complexes coverage as a function of the substrate contact angle, at the indicated values of the hydration levels.
}
\label{fig13:coverage}
\end{figure}
The average substrate coverage for the ionomer backbones, \ce{H2O}, side-chains, \ce{H3O+} and \ce{SO3-} groups are shown in Fig.~\ref{fig13:coverage} for all thin-films investigated. The coverage is defined here as the number of molecules within the adsorption region per unit of area. The data in Fig.~\ref{fig13:coverage} clearly show an inversion of surface coverage following the hydrophilicity degree of the support. In contrast, water content does not seem to significantly modify ionomer backbone or water coverages. Indeed, by decreasing water content from $\lambda=22$ to $\lambda=6$, backbone coverage changes from $14.65$ to $15.72$~molecules/nm$^2$ for the most hydrophobic case, while water coverage reduces from $12.80$ to $11.56$~molecules/nm$^2$ for the most hydrophilic case. The reduction of water coverage is compensated by the increases of \ce{H3O+} and \ce{SO3-} coverages. \ce{SO3-} coverage increases from $0.003$ to $0.007$~molecules/nm$^2$ while the \ce{H3O+} coverage changes from $0.008$ to $0.015$~molecules/nm$^2$. Hence, the number of adsorbed \ce{SO3-} groups is higher for $\lambda=$~6 and 11, and they are well dispersed on the surface. In contrast, for $\lambda=22$, the \ce{SO3-} groups can be found in more agglomerated configurations. Overall, Figs.~\ref{fig12:snapshot-adsorptionRegion} and~\ref{fig13:coverage} further corroborate our previous observation of a transition from a predominant backbone coverage to predominant water coverage, when increasing the hydrophilic character of the substrate. However, even for most hydrophilic cases adsorption of the ionomer is still observed and occurs mainly via \ce{SO3-} groups. The adsorption of \ce{SO3-} is more evident when the hydration of the film is lower.
During PEMFC operation, oxidation and reduction reactions occurring on the top of catalyst surfaces strongly depend on surface coverage of reactants and spectator species~\cite{Franco2006,Malek2011a,DeMorais2011}. Our results shows that water molecules and hydronium ions can be found away from the catalyst surface, in the case where the wetting nature of the substrate is not favourable. The adsorption of the ionomer could block the adsorption of reactant species, reducing the area where the electrochemical reaction occurs. Note that this behaviour is usually overlooked when addressing the issue of increasing Pt utilization in PEMFCs.
Also important for PEMFC development is to clarify the impact of ionomer adsorption in ORR mechanism. It is well know that the kinetics of the ORR is sensitive to the nature of adsorption of spectator species~\cite{Markovic2002}. For example, specific adsorption of sulfonate anions has an important deactivation effect on the ORR. The extent of this feature correlates with the strength of the catalyst-sulfonate bond (the strenght of \ce{SO3-} adsorption)~\cite{Subbaraman2010, Subbaraman2010a}. Various factors can influence the chemical nature of \ce{SO3-} adsorption, including nature of the counter-cation, extent of \ce{SO3-} agglomeration within the ionomer, length and spacing between side chains adjacent along the backbone. Our results show that the \ce{SO3-} groups are adsorbed in different configurations, {\it e.g.} , both clustered and dispersed. This should affect the chemical nature of the \ce{SO3-} adsorption, and ultimately affect the electrochemical potential that drives the electrochemical reactions.
To conclude this Section, we observe that cell reactions are also governed by the structural properties of the Electrical Double Layer (EDL) formed close to the electrode surface~\cite{Quiroga2014}. Unfortunately, standard electrochemical theories normally used to describe the EDL, completely ignore the heterogeneous environment created by the adsorbed ionomer, which affects both charge and potential distributions~\cite{WangLRoudgarA2009, Krapf2006, Zhdanov2006561, Zhdanov2004, Zhdanov2008, Biesheuvel2009}. In contrast, our findings clearly show that the ionomer dictates the distribution of charges very close to the surface (as indicated by the ionic distributions shown in Fig.~\ref{fig9:ionic-clusterZ}) and, as a consequence, the over-potential at the reactant-electrode distance ($\sim$~0.2-0.5~nm) is also affected. Moreover, considering the different ionomer morphologies that may be found inside the CL, it is not much to say that the reaction rates are far from being uniformly distributed inside CL. Our results also strongly support the existence of a non-uniform spatial distribution of reaction rates, due to the complexity of the ionomer structure. An effective control of the ionomer morphology could therefore provide a valuable path for further development of PEMFC technology, for optimizing electrochemical interface and reducing ionomer inhibition.
\subsection{The ionomer/vacuum interface}
\label{subsec:film-air}
The morphology of the Nafion/vacuum interface has recently received special attention, also due to its importance in ionomer water uptake~\cite{Bass2011}. This interface includes the hydrophobic ionomer backbones which are exposed to the gas phase, and the underneath hydrophilic side-chains, pointing toward the water-rich domains. It is considered responsible for the so-called Schroeder's paradox, {\em i.e.,} a different Nafion water uptake from a liquid solvent or its vapour~\cite{Freger2009, Choi2003}.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\textwidth,angle=-90,origin=c]{./figure15.pdf}
\caption{
Colour maps of the wetting character of Nafion thin-film ionomer/vacuum interface. Hydrophobic and hydrophilic regions are in yellow and blue, respectively. The technique used for determining the maps is described in details in the text.
}
\label{fig:filmsurfacemap}
\end{figure}
\begin{figure*}[]
\centering
\includegraphics[width=0.9\textwidth]{./figure16.pdf}
\caption{
$xy$-contour maps of the $z$-position of atoms at the ionomer/vacuum interface for thin-films at $\theta=70^\circ$ and hydration levels $\lambda=$ 22 (a), 11 (b), and 6 (c).}
\label{fig:roughnessmap}
\end{figure*}
In order to explore the wetting nature of the ionomer/vacuum interface, we have determined spatial color maps of the local hydrophilic/hydrophobic character of the interface. In our calculations we have considered the atoms pertaining to polymer backbones and side-chains (different than sulfonate groups) as hydrophobic, while hydrophilic species included sulfonate groups, water molecules and hydronium ions. We have identified the ionomer/vacuum interface as the region with $3.0\le z\le 4.5$~nm. This region was partitioned in a regular grid, with cubic cells of volume $0.2\times0.2\times1.5$~nm$^3$, for all considered cases. We next attributed to each cell $i$ the difference in volume associated to hydrophobic and hydrophilic atoms in the cell, $\delta V^i = V^i_{phobic}-V^i_{philic}$. A negative value of $\delta V^i$ therefore corresponds to a mostly hydrophilic cell, a positive value to a hydrophobic one. The volume associated to each atom was computed by considering the value of the corresponding Lennard-Jones interaction parameter $\sigma$ as the effective diameter of the atom. We considered an average over an ensemble of $10^3$ configurations for each film.
In Fig.~\ref{fig:filmsurfacemap} we show the wetting maps for all films considered. The color range interpolates from strongly hydrophobic (yellow) to very hydrophilic (violet) regions. Thin films clearly present a compact and extended hydrophobic layer on the top in all cases, as already demonstrated above. However, violet regions are evident for $\theta=150^\circ$ and $100^\circ$ at high values of $\lambda$, which result from significant number of water molecules which accumulate immediately below the polymer backbone. In contrast, films with $\theta=70^\circ$ and $30^\circ$ present hydrated regions of very limited extent. These data suggest that the hydrophobicity of the ionomer/vacuum interface is particularly pronounced in the case of films formed on very hydrophilic substrates. At the lowest water contents, the films present similar surface hydrophobicity at all $\theta$ values.
Our results also suggest that tuning the film/substrate interaction can modify the Nafion ionomer/vacuum interface morphology. For instance, the substrate with $\theta=30^\circ$ determines an interface configuration where the entire water content is confined under the polymer, whereas the ionomer backbone forms a "crusty" hydrophobic layer. This "crust" should present high resistance to deformation, which could decrease water uptake and lead to transport losses during PEMFC operation. It could also prevent reactants (\ce{O2} and \ce{H2} coming from the CL pores) to cross the thin-film for reaching the catalyst sites. In contrast, the films formed on the substrate with $\theta=150^\circ$, are characterized by a configuration where a fraction of the ionomer backbone is in direct contact with the substrate. This reduces the concentration of polymer backbone at the interface with vacuum and, as a consequence, increases the presence of water. Clearly, this interface should be more favourable for water absorption, which is in contrast with the results of Ref.~\cite{Modestino2012} where, however, thin-films about $20$ times thicker than the ones considered here were investigated.
\begin{table}[b]
\centering
\begin{tabular}{cccc}
\hline \hline
$\theta(^\circ)\backslash\lambda$& $22$ & $11$ & $6$\\
\hline \hline
$150^\circ{}$ & 0.16 & 0.25 & 0.29 \\
$100^\circ{}$ & 0.21 & 0.56 & 0.29 \\
$70^\circ{}$ & 0.13 & 0.46 & 0.24 \\
$30^\circ{}$ & 0.25 & 0.44 & 0.30 \\
\hline\hline
\end{tabular}
\caption{
Roughness coefficient $R$~(nm) for the ionomer/vacuum interface, calculated as discussed in the text.
}
\label{tb:roughness}
\end{table}
The hydrophobic "crusty" ionomer/vacuum interface is characterized by a certain degree of roughness, which depends on the hydration conditions. Roughness can be quantified as the vertical deviation of the real surface compared to its ideal form, defined as the average vertical position of the interface. We can thus define a mean-squared roughness coefficient as $R^2=1/N \sum\limits_{i=1}^{N}(Z_i-\bar{Z})^2$, where $Z_i$ denotes the $z$-coordinate of the exposed atom $i$ at the interface, $\bar{Z}$ is the average $z$-position of the surface atoms, and $N$ is the number of the surface atoms~\cite{Huang2012}. Surface atoms are identified as those with no other atoms in a square prism of edge $0.1$~nm and height $5$~nm above them.
In Fig.~\ref{fig:roughnessmap} we show the $xy$-contour maps of the $z$-position of atoms at the ionomer/vacuum interface, for the case $\theta=70^\circ$, at the indicated values of $\lambda$. Table~\ref{tb:roughness} reports the values of $R$ for all films studied. The roughness of the films surface assumes values in the range $0.13\div 0.56$~nm, which can be compared to an experimental value of the roughness of Nafion films in contact with air of $0.35$~nm~\cite{Bass2010}. Interestingly, the roughness of the films at intermediate hydrophilicity, $\theta=100^\circ$, are slightly higher when compared to other films. This can be attributed to the disordered cluster configurations described above. According to Bass {\em et al.}~\cite{Bass2010}, the morphology of these interfaces is stable as long as the water vapour is not saturated. At that point, the hydrophobic layer should deform and the buried hydrophilic groups eventually migrate to the surface. However, when the surface is initially hydrophobic (especially at low water contents), the high energetic and kinetic barriers associated with the rearrangement of many chemical groups, may keep the ionomer kinetically trapped in this state for very long times~\cite{Bass2010}.
\section{Conclusions and Perspectives}
\label{sec:conclusions}
We have studied by Molecular Dynamics simulations the formation of Nafion ultra-thin films in contact with unstructured flat supports, characterized by their global wetting properties only. By tuning a single control parameter, $\epsilon_{w}$, we have been able to investigate in an unique framework an extended range of environments peculiar of the PEMFC catalyst layer, ranging from strongly hydrophobic (carbon-like) to very hydrophilic (platinum-like). The hydrophilicity degree of the substrate was estimated by computing the contact angle of a water droplet gently deposited on it. We considered four substrates, from strongly hydrophobic, through intermediate and mildly hydrophilic to very hydrophilic. Also, three hydration levels were considered, in order to investigate the role played by water content. Self-assembled instances of the thin-films corresponding to these very diverse conditions were analysed in details, in terms of their structural properties. Based on a very extended data sets, we have been able to propose a general picture for Nafion supported thin films morphology for variable wetting nature of the substrate and hydration conditions.
Our data show that variations in the hydrophilic character of the substrate have strong impact on film morphology. This ranges from a sandwich structure, where an extended water pool is sandwiched by ionomer sheets, to a bilayer configuration. In this case water floods the interface with the substrate and polymer mostly accumulate at the top, at the interface with air. By decreasing water content, films convert into inverted micelles and multilamellar structures, for hydrophobic and hydrophilic supports, respectively. We have also discovered that, in contrast to the sandwich structure, the bilayer structure shows large and compact \ce{SO3-} agglomerates, resulting in a poor hydration of \ce{H3O+} and \ce{SO3-} . Analysis of surface coverage showed a clear transition from predominant backbone coverage to predominant water coverage, when switching from hydrophobic to hydrophilic surfaces. Finally, we have shown that tuning the hydrophilicity of the substrate it is possible to modify the film/vapour interface.
The results presented in this work could be of interest for optimization of the catalyst layer performances and further development of PEMFC technology. We have shown that it is indeed possible to control the main morphological features of the films by tuning the wetting nature of the substrate. Therefore, the use of appropriate substrates could be highly attractive for controlling some aspects such as ionomer coverage, proton accessibility to the active surface, \ce{SO3-} adsorption, among others. This would optimize the electrode/electrolyte interface, in order to create electrochemical environment favourable to enhance cell reaction rates.
\bibliographystyle{achemso}
|
\section{Introduction}\label{sec:intro}
Ongoing data collection from the private sector along with federal, state, and local governments have produced massive quantities of data measured over geographic regions (areal data) and time. This unprecedented volume of spatio-temporal data contains a wide range of variables and has, thus, created unique challenges and opportunities for those practitioners seeking to capitalize on their full utility. For example, methodological challenges arise because these data exhibit variability over multiple processes (latent variables), geographies, and time-points. As such, the corresponding multivariate-spatio-temporal covariances may be quite complex, involving nonstationarity and interactions between different variables, regions, and times. To effectively model these complex dependencies we introduce the multivariate-spatio-temporal mixed effects model (MSTM).
Despite the wide availability of areal datasets exhibiting multivariate-spatio-temporal dependencies, the literature on modeling multivariate-spatio-temporal areal processes is relatively recent by comparison. For example, versions of a multivariate space-time conditional autoregressive (CAR) model have been proposed by \citet{carlinmst}, \citet{congdon}, \citet{pettitt}, \citet{zhuglm}, \citet{daniels}, and \citet{bestmst}, among others. However, these methodologies cannot accommodate data from multiple repeated surveys, and cannot efficiently model very-large-to-massive datasets. Additionally, these approaches impose separability assumptions (and in some cases independence), which are not appropriate for many settings, as these models fail to capture important interactions (and dependence) between different variables, regions, and times \citep{steinSep}.
In addressing the aforementioned issues, the proposed MSTM provides several methodological contributions, including a novel class of multivariate-spatio-temporal basis functions. Additionally, the MSTM framework introduces an extremely powerful dimension reduction approach for the multivariate-spatio-temporal setting. Specifically, we introduce a novel conditional (i.e., first-order) multivariate-spatio-temporal-dynamic-linear model specification, and propose an innovative parameter model to reduce a high-dimensional parameter space.
The foundation for these methodological advances consists of two general techniques for modeling dependencies. The first technique allows different random processes (latent variables) to share the same random effect. This idea has been effectively utilized in the spatial \citep{banerjee,johan,lindgren-2011,hughes,nychkaLK}, multivariate-spatial \citep{royle1999,finley,finley2}, and spatio-temporal \citep{stcar,wikle2001} settings. The second technique we consider partitions the joint likelihood into a product of more manageable conditional likelihoods; see, \citet{berlinerroyle} for the multivariate spatial setting, and \citet[][Chap. 7]{cressie-wikle-book} for the spatio-temporal setting.
To incorporate multivariate spatial dependencies in our MSTM framework we let the $L$ different variables under consideration share the same random effects at time $t$ (i.e., the first technique). Then, to build in temporal dependence we use a dynamic Bayesian hierarchical model (BHM) and specify conditional distributions for the temporal random effects at time $t$ given previous times (i.e., the second technique). Together, these two techniques are combined in a novel way to define the MSTM. In what follows, we demonstrate that our approach is extremely general and that many datasets can be effectively modeled using the MSTM framework.
Accounting for the possibility of ``big data'' is arguably one of the most important features to consider, since the size of modern spatio-temporal datasets creates unavoidable methodological difficulties. In fact, there is a large literature available on modeling large-to-massive spatial/spatio-temporal datasets and similar complications arise in the multivariate-spatio-temporal, though often exacerbated. The primary issue surrounding this literature is that there is a computational bottleneck that occurs when computing a {high-dimensional} Gaussian likelihood. {See \citet{cressie-wikle-book}, \citet{reviewmethods}, and the references therein for a comprehensive review of spatial and spatio-temporal models for ``big data.''} To illustrate the exceptional utility of our approach, we demonstrate that the MSTM can efficiently model massive data by considering a survey dataset consisting of 7,530,037 observations and 3,680 spatial fields, which we jointly model using the MSTM.
We are able to analyze this massive dataset using the MSTM framework, since the model can be specified to have a computationally advantageous reduced rank structure \citep[e.g., see][]{wikleHandbook}. This reduced rank structure is achieved, in part, by extending aspects of the model suggested by \citet{hughes} from the univariate spatial-only setting to the multivariate-spatio-temporal setting. Specifically, we extend the Moran's I (MI) basis functions to the multivariate-spatio-temporal setting \citep[see][for the spatial only case]{griffith2000, griffith2002, griffith2004, griffith2007,hughes,aaronp}. Further, we introduce a novel propagator matrix for a first-order vector autoregressive (VAR(1)) model, which we call the MI propagator matrix. Here, the propagator matrix of {the} VAR(1) is specified to have a desirable non-confounding property, which is similar to the specification of the MI basis functions in \citet{hughes}.
In addition to the MI basis functions and propagator matrix, we also propose an extension of the spatial random effects covariance parameter model used in \citet{hughes} and \citet{aaronpBayes}, which we call the MI prior. Here, we interpret the MI prior as a rescaling of the covariance matrix that is specified to be close (in Frobenius norm) to a ``target precision'' matrix. This parameterization significantly reduces the dimension of the parameter space, thereby reducing the computation needed for fully Bayesian inference. Furthermore, this target precision matrix can be sensibly chosen based on knowledge of the underlying spatial process.
Given these methodological advances, the MSTM can be used to effectively address numerous statistical modeling and analysis problems in the context of multivariate-spatio-temporal areal data. In particular, we use the MSTM to model nonseparable and nonstationary covariances, to combine data from multiple repeated surveys, and to analyze a massive multivariate-spatio-temporal areal dataset. Although, in this article, we mainly focus on these particular applications, the MSTM is tremendously flexible and can be readily adapted to other settings.
The remainder of this article is organized as follows. In Section~\ref{sec:MSTM}, we introduce and define the MSTM, which includes our proposed MI propagator matrix. Then, in Section~\ref{sec:parmod} we introduce the parameter model for the covariance matrix of the random effects term and show explicitly how one can incorporate knowledge of the spatial process into this parameter model. Next, in Section~\ref{sec:app} we demonstrate the use of the MSTM through three distinct modeling applications, namely: analyzing cancer mortality rates using data obtained from the National Cancer Institute (NCI); combining unemployment rates from the American Community Survey (ACS, US Census Bureau) and the Local Area Unemployment Survey (LAUS, Bureau of Labor Statistics (BLS)), and modeling multivariate-spatio-temporal data from the Longitudinal Employer-Household Dynamics (LEHD) program (US Census Bureau). These examples are used to demonstrate that we can model nonseparable and nonstationary multivariate-spatio-temporal covariances, combine data from multiple surveys, and process massive data using the MSTM. We end with a discussion in Section~\ref{sec:disc}. For convenience of exposition, details surrounding some of the technical results and the MCMC algorithm are left to the Appendix.
\section{The Multivariate-Spatio-Temporal Mixed Effects Model}\label{sec:MSTM}
The MSTM is defined hierarchically using the familiar data model, process model, and parameter model organization \citep{berlinhier,cressie-wikle-book}. In this section, we provide the specific details.
\subsection{The Data Model} The data model for the MSTM is defined as,
\begin{equation}
\label{data:model}
Z_{t}^{(\ell)}(A) = Y_{t}^{(\ell)}(A) + \epsilon_{t}^{(\ell)}(A);\hspace{5pt} \ell = 1,\ldots,L, \hspace{5pt}t = T_{L}^{(\ell)},\ldots,T_{U}^{(\ell)}, \hspace{5pt}A \in D_{\mathrm{P},t}^{(\ell)},
\end{equation}
\noindent
where $\{Z_{t}^{(\ell)}(\cdot)\}$ represents multivariate-spatio-temporal data. The components of (\ref{data:model}) are defined and elaborated as follows:
\begin{enumerate}
\item The subscript ``$t$'' denotes discrete time, and the superscript ``$\ell$'' indexes different variables of interest (e.g., unemployment rates). There are a total of $L$ variables of interest (i.e., $\ell = 1,\ldots,L$) and we allow for a different number of observed time-points for each of the $L$ variables of interest (i.e., for variable $\ell$, $t = T_{L}^{(\ell)},\ldots,T_{U}^{(\ell)}$).
\item We require $T_{L}^{(\ell)},\ldots,T_{U}^{(\ell)}$ to be on the same temporal scale (e.g., monthly, quarterly, etc.) for each $\ell$, $T_{L}^{(\ell)} \le T_{U}^{(\ell)}$, $\mathrm{min}\left(T_{L}^{(\ell)}\right) = 1$, and $\mathrm{max}\left(T_{U}^{(\ell)}\right) = T \ge 1$.
\item The set $A$ represents a generic areal unit. For example, a given set $A$ might represent a state, county, or a Census tract. Denote the collection of all $n_{t}^{(\ell)}$ observed areal units with the set $D_{\mathrm{O},t}^{(\ell)}\equiv \{A_{t,i}^{(\ell)}: i = 1,\ldots,n_{t}^{(\ell)}\}$; $\ell = 1,\ldots,L$. The observed data locations are different from the prediction locations $D_{\mathrm{P},t}^{(\ell)}\equiv \{A_{t,j}^{(\ell)}: j = 1,\ldots,N_{t}^{(\ell)}\}$; that is, we consider predicting on a spatial support that may be different from $\{D_{\mathrm{O},t}^{(\ell)}\}$. Additionally, denote the number of prediction locations at time $t$ as $N_{t} =\sum_{\ell = 1}^{L}N_{t}^{(\ell)}$, and the total number of prediction locations as ($n \equiv \sum_{t = 1}^{T}n_{t}$) $N \equiv \sum_{t = 1}^{T}N_{t}$. In a similar manner, the number of observed locations at time $t$ and total number of observations are given by $n_{t} =\sum_{\ell = 1}^{L}n_{t}^{(\ell)}$ and $n \equiv \sum_{t = 1}^{T}n_{t}$, respectively.
\item The random process $Y_{t}^{(\ell)}(\cdot)$ represents the $\ell$-th variable of interest at time $t$. For example, $Y_{t}^{(\ell)}(\cdot)$ might represent the cancer mortality rate for females at time $t$. The stochastic properties of $\{Y_{t}^{(\ell)}(\cdot)\}$ are defined in Section~\ref{sec:process}. Latent processes like $\{Y_{t}^{(\ell)}(\cdot)\}$ have been used to incorporate spatio-temporal dependencies \citep[e.g., see][]{cressie-wikle-book}, which we modify to the multivariate-spatio-temporal setting.
\item It is assumed that $\epsilon_{t}^{(\ell)}(\cdot)$ is a white-noise Gaussian process with mean zero and known variance var(\(\epsilon_{t}^{(\ell)}(\cdot)\))= $v_{t}^{(\ell)}(\cdot)$ for $\ell = 1,\ldots,L,$ and $t = T_{L}^{(\ell)},\ldots,T_{U}^{(\ell)}$. The presence of $\{\epsilon_{t}^{(\ell)}(\cdot)\}$ in (\ref{data:model}) allows us to take into account that we do not perfectly observe $\{Y_{t}^{(\ell)}(\cdot)\}$, and instead observe a {noisy} version $\{Z_{t}^{(\ell)}(\cdot)\}$. In many settings, there is information that we can use to define $\{\epsilon_{t}^{(\ell)}(\cdot)\}$ (e.g., information provided by a statistical agency). In particular, {variances} are provided by the statistical agency {(e.g., NCI, US Census Bureau, and BLS provide survey variance estimates)}. If one does not account for this extra source of variability
then the total variability of the process $\{Y_{t}^{(\ell)}(\cdot)\}$ may be underestimated. For example, \citet{finley} show that if one ignores white-noise error in a Gaussian linear model then one underestimates the total variability of the latent process of interest.
\end{enumerate}
\subsection{The Process Model}\label{sec:process}
The process model for MSTM is defined as,
\begin{align}
\label{process:model}
Y_{t}^{(\ell)}(A) = \mu_{t}^{(\ell)}(A) + \textbf{S}_{t}^{(\ell)}(A)^{\prime}\bm{\eta}_{t} + \xi_{t}^{(\ell)}(A);&\hspace{5pt} \ell = 1,\ldots,L, t = T_{L}^{(\ell)},\ldots,T_{U}^{(\ell)},A \in D_{\mathrm{P},t}^{(\ell)}.
\end{align}
\noindent
In (\ref{process:model}), $Y_{t}^{(\ell)}(\cdot)$ represents the $\ell$-th spatial random process of interest at time $t$, which is modeled by three terms on the right-hand side of (\ref{process:model}). The first term (i.e., $\{\mu_{t}^{(\ell)}(\cdot)\}$) is a fixed effect, which is unknown, and requires estimation. We set $\mu_{t}^{(\ell)}(\cdot)\equiv \textbf{x}_{t}^{(\ell)}(\cdot)^{\prime}\bm{\beta}_{t}$, where $\textbf{x}_{t}^{(\ell)}$ is a known $p$-dimensional vector of covariates and $\bm{\beta}_{t}\in \mathbb{R}^{p}$ is a fixed unknown parameter vector; $\ell = 1,\dots,L$ and $t = 1,\ldots,T$. In general, we allow both $\textbf{x}_{t}^{(\ell)}$ and $\bm{\beta}_{t}$ to change over time; however, in practice one must assess whether or not this is appropriate for their application.
The second term on the right-hand side of (\ref{process:model}) (i.e., $\{\textbf{S}_{t}^{(\ell)}(\cdot)^{\prime}\bm{\eta}_{t}\}$) represents multivariate-spatio-temporal dependencies; in Section~\ref{sec:dynamics}, we provide the stochastic properties of $\{\bm{\eta}_{t}\}$. Here, the $r$-dimensional vectors of multivariate-spatio-temporal basis functions $\textbf{S}_{t}^{(\ell)}(\cdot)\equiv (S_{t,1}^{(\ell)}(\cdot),\ldots,S_{t,r}^{(\ell)}(\cdot))^{\prime}$ are pre-specified for each $t = 1,\ldots,T$ and $\ell = 1,\ldots,L$. {In principle,} the $r$-dimensional vector $\textbf{S}_{t}^{(\ell)}(\cdot)$ can belong to any class of spatial basis functions; however, we shall use the Moran's I (MI) basis functions \citep{griffith2000, griffith2002, griffith2004, griffith2007,hughes,aaronp}. The MI basis functions are a class of functions used to model areal data in a reduced dimensional space (i.e., $r \ll n$). This feature allows for fast computation of the distribution of $\{\bm{\eta}_{t}\}$, which can become computationally expensive for large $r$ \citep{hughes, aaronp}. This will be particularly useful for the datasets in Section~\ref{sec:app}, which {for one of our examples has 7,530,037 observations}. Additionally, the MI basis functions allow for nonstationarity in space and, for areal data, this is a desirable property (e.g., see \citet{banerjee-etal-2004} and the references therein).
Now, the $r$-dimensional vectors of MI basis functions $\{\textbf{S}_{t}^{(\ell)}(\cdot)\}$ are equivalent to the first $r$ eigenvectors of the MI operator \citep[see,][]{hughes}. That is, the MI operator at time $t$ is defined as
\begin{equation}
\textbf{G}(\textbf{X}_{t},\textbf{A}_{t}) \equiv\left(\textbf{I}_{N_{t}} - \textbf{X}_{t}\left(\textbf{X}_{t}^{\prime}\textbf{X}_{t}\right)^{-1}\textbf{X}_{t}^{\prime}\right)\textbf{A}_{t}\left(\textbf{I}_{N_{t}} - \textbf{X}_{t}\left(\textbf{X}_{t}^{\prime}\textbf{X}_{t}\right)^{-1}\textbf{X}_{t}^{\prime}\right);\hspace{5pt}t = 1,\ldots,T,
\end{equation}
where the $N_{t}\times p$ matrix $\textbf{X}_{t} \equiv \left(\textbf{x}_{t}^{(\ell)}(A): \ell = 1,\ldots,L, A \in D_{\mathrm{P},t}^{(\ell)}\right)^{\prime}$, $\textbf{I}_{N_{t}}$ is an $N_{t}\times N_{t}$ identity matrix, and $\textbf{A}_{t}$ is the $N_{t}\times N_{t}$ adjacency matrix corresponding to the edges formed by $\{D_{\mathrm{P},t}^{(\ell)}:\ell = 1,\ldots,L\}$. From the spectral representation, $\textbf{G}(\textbf{X}_{t},\textbf{A}_{t}) = \bm{\Phi}_{X,G,t}\bm{\Lambda}_{X,G,t}\bm{\Phi}_{G,t}^{\prime}$, we denote the $N_{t}\times r$ real matrix formed from the first $r$ columns of $\bm{\Phi}_{X,G,t}$ as $\textbf{S}_{X,t}$. Additionally, set the row of $\textbf{S}_{X,t}$ that corresponds to variable $\ell$ and areal unit $A$ equal to $\textbf{S}_{t}^{(\ell)}(A)$.
The third term on the right-hand side of (\ref{process:model}) (i.e., $\{\xi_{t}^{(\ell)}(\cdot)\}$) represents fine-scale variability and is assumed to be Gaussian white-noise with mean-zero and unknown variance $\{\sigma_{\xi,t}^2\}$. In general, $\{\xi_{t}^{(\ell)}(\cdot)\}$ represents the left-over variability not accounted for by $\{\textbf{S}_{t}^{(\ell)}(\cdot)^{\prime}\bm{\eta}_{t}\}$. In the setting where the variance $\{v_{t}^{(\ell)}(\cdot)\}$ is unknown, one should model the sum $\{\epsilon_{t}^{(\ell)}(\cdot) + \xi_{t}^{(\ell)}(\cdot)\}$ due to identifiability issues between $\{\epsilon_{t}^{(\ell)}(\cdot)\}$ and $\{\xi_{t}^{(\ell)}(\cdot)\}$ \citep[e.g., see][]{banerjee, finley,finley2}. Minor adjustments can be made to our methodology to allow for this. One might also consider modeling spatial covariances in $\{\xi_{t}^{(\ell)}(\cdot)\}$. Again, minor adjustments to our methodology could be used to incorporate, for example, a CAR model \citep[][Chap. 3]{banerjee-etal-2004}, tapered covariances \citep[][pg. 108]{cressie} or block diagonal covariances \citep{steinr} in $\{\xi_{t}^{(\ell)}(\cdot)\}$.
\subsection{Temporal Dynamics for the Latent Process}\label{sec:dynamics} We assume $\bm{\eta}_{t}$ is generated using a VAR(1) {model} \citep[][Chap. 7]{cressie-wikle-book}:
\begin{equation}\label{autoregressive}
\bm{\eta}_{t} = \textbf{M}_{t}\bm{\eta}_{t-1} + \textbf{u}_{t};\hspace{5pt}t = 2,3,\ldots,T
\end{equation}
\noindent
where for $t = 1,2,\ldots,T$ the $r$-dimensional random vector $\bm{\eta}_{t}$ is Gaussian with mean-zero and has an unknown $r\times r$ covariance matrix $\textbf{K}_{t}$; $\textbf{M}_{t}$ is a $r \times r$ known propagator matrix (see discussion below); and $\textbf{u}_{t}$ is an $r$-dimensional Gaussian random vector with mean-zero and unknown $r\times r$ covariance matrix $\textbf{W}_{t}$, and is independent of $\bm{\eta}_{t-1}$.
First order vector autoregressive models may offer more realistic structure with regards to interactions across space and time. This is a feature that cannot be included in the alternative modeling approaches discussed in Section~\ref{sec:intro}. Additionally, the VAR(1) model has been shown to perform well (empirically) in terms of both estimation and prediction for surveys repeated over time \citep{Jones,Bell,Feder}.
The $r$-dimensional random vectors $\{\bm{\eta}_{t}\}$ are not only used to model temporal dependencies in $\{Y_{t}^{(\ell)}(\cdot)\}$, but are also used to model multivariate dependencies. Notice that the random effect term $\bm{\eta}_{t}$ is common across all $L$ processes. Allowing for a common random effect term between different processes is a simple way to induce dependence \citep[][Chap. 7.4]{cressie-wikle-book}. This strategy has been used by \citet{royle1999}, \citet{finley}, and \citet{finley2} in the multivariate spatial setting { and has been extended here.}
We are now left to specify the $r\times r$ real matrices in the set $\{\textbf{M}_{t}\}$. The problem of confounding provides motivation for the definition of the MI basis functions $\{\textbf{S}_{X,t}^{(\ell)}(\cdot)\}$ \citep{griffith2000, griffith2002, griffith2004, griffith2007,Reich,hughes}. In a similar manner, the problem of confounding manifests in a spatio-temporal VAR(1) model and can be addressed through careful specification of $\{\textbf{M}_{t}\}$. To see this, substitute (\ref{autoregressive}) into (\ref{process:model}) to obtain,
\begin{equation}\label{matrix:process}
\textbf{y}_{t} = \textbf{X}_{t}\bm{\beta}_{t} + \textbf{S}_{X,t}\textbf{M}_{t}\bm{\eta}_{t-1} + \textbf{S}_{X,t}\textbf{u}_{t} + \bm{\xi}_{t};\hspace{5pt} t = 2,\ldots, T
\end{equation}
where the $N_{t}$-dimensional latent random vectors $\textbf{y}_{t}\equiv (Y_{t}^{(\ell)}(A): \ell = 1,\ldots,L, A \in D_{\mathrm{P},t}^{(\ell)})^{\prime}$ and $\bm{\xi}_{t}\equiv (\xi_{t}^{(\ell)}(A): \ell = 1,\ldots,L, A \in D_{\mathrm{P},t}^{(\ell)})^{\prime}$. The specification of $\{\textbf{S}_{X,t}\}$ using MI basis functions implies that there are no issues with confounding between $\{\bm{\beta}_{t}\}$ and $\{\textbf{u}_{t}\}$; however, depending on our choice for $\{\textbf{M}_{t}\}$ there might be issues with confounding between $\bm{\eta}_{t-1}$ and the $(p+r)$-dimensional random vector $\bm{\zeta}_{t} \equiv (\bm{\beta}_{t}^{\prime}, \textbf{u}_{t}^{\prime})^{\prime}$; $t = 2,\ldots,T$. Then, rewriting (\ref{matrix:process}), we get
\begin{equation}\label{matrix:process2}
\textbf{S}_{X,t}^{\prime}(\textbf{y}_{t}-\bm{\xi}_{t}) = \textbf{B}_{t}\bm{\zeta}_{t} + \textbf{M}_{t}\bm{\eta}_{t-1};\hspace{5pt} t = 2,\ldots, T,
\end{equation}
where the $r\times(p+r)$ matrix $\textbf{B}_{t}\equiv (\textbf{S}_{X,t}^{\prime}\textbf{X}_{t}, \textbf{I})$. The representation in (\ref{matrix:process2}) gives rise to what we call the MI propagator matrix, which is defined in an analogous manner to the MI basis functions. Using the spectral representation of $\textbf{G}(\textbf{B}_{t},\textbf{I}_{r}) = \bm{\Phi}_{G,B,t}\bm{\Lambda}_{G,B,t}\bm{\Phi}_{G,B,t}^{\prime}$ we set the $r\times r$ real matrix $\textbf{M}_{t}$ equal to the first $r$ columns of $\bm{\Phi}_{G,B,t}$ for each $t$, which is denoted with $\textbf{M}_{B,t}$.
Notice that there are no restrictions on $\{\textbf{M}_{B,t}\}$ to {mathematically guarantee} that $\textbf{M}_{B,t}$ does not become ``explosive'' as $t$ increases. Thus, one should investigate whether or not this is the case when using this model for forecasting. One should also be aware that the covariates $\{\textbf{X}_{t}\}$ inform the MI propagator matrices. Importantly, $\{\textbf{M}_{B,t}\}$'s (functional) dependence with $\{\textbf{X}_{t}\}$ and $\{\textbf{S}_{X,t}\}$ implies nonstationarity in time.
Also, notice that we do not treat $\textbf{M}_{t}$ as an unknown parameter matrix to be estimated. Instead, we chose a specific form for $\{\textbf{M}_{t}\}$, namely $\{\textbf{M}_{B,t}\}$, that avoids confounding between $\{\bm{\eta}_{t}\}$ and $\{\bm{\zeta}_{t}\}$. As a result, the final form of $\{\textbf{M}_{B,t}\}$ might not be interpretable from a spatial point of view. We address this in Section~\ref{sec:parmod}, where constraints are added to the parameter model so that $\mathrm{cov}(\bm{\eta}_{t}) = \textbf{M}_{B,t}\textbf{K}_{t-1}\textbf{M}_{B,t}^{\prime} + \textbf{W}_{t}$ is { spatially interpretable.}
\section{The Parameter Model}\label{sec:parmod}
At this stage, one could specify any desired prior for the $r\times r$ covariance matrices $\textbf{K}_{t}$ and $\textbf{W}_{t}$. We propose a novel specification for $\textbf{K}_{t}$ and $\textbf{W}_{t}$ that provides an extension of the MI prior used by \citet{griffith2000,griffith2002,griffith2004}, \citet{griffith2007}, \citet{hughes}, and \citet{aaronp}. The MI prior for $\textbf{K}_{1}$ is given by $\textbf{K}_{1} = \sigma_{K}^{2} \textbf{S}_{X,1}^{\prime}\textbf{Q}_{1}\textbf{S}_{X,1}$, where $\sigma_{K}^{2}>0$ is unknown, $\textbf{Q}_{1} = \bm{1}_{N_{1}} - \textbf{A}_{1}$, and $\bm{1}_{N_{1}}$ is a $N_{1}$-dimensional vector of 1s. Notice that the MI prior is defined in the spatial only setting where $t = 1 = T$. Hence, we extend this prior to the multivariate-spatio-temporal setting. Moreover, we show that this extension allows one to incorporate knowledge of the spatial process.
Now, a reasonable criticism of the MI basis function is that we may be ignoring important sources of variability by restricting $\{\textbf{S}_{t}\}$ to a column space that is linearly independent of $\{\textbf{X}_{t}\}$; that is, for each $t$ requiring $\textbf{S}_{t}\in \mathcal{C}(\textbf{P}_{X,t})$, where $\textbf{P}_{X,t}\equiv \textbf{X}_{t}(\textbf{X}_{t}^{\prime}\textbf{X}_{t})^{-1}\textbf{X}_{t}$ and the column space of $\textbf{P}_{X,t}$ is denoted as $\mathcal{C}(\textbf{P}_{X,t})$. To see this, rewrite (\ref{process:model}) and let $\textbf{S}_{t} = [\textbf{H}_{X,t}, \textbf{L}_{X,t}]$ and $\bm{\eta}_{t}\equiv (\bm{\kappa}_{X,t}^{\prime},\bm{\delta}_{X,t})^{\prime}$ so that
\begin{align}\label{decomp2}
\textbf{y}_{t} &= \textbf{X}_{t}\bm{\beta}_{t} + \textbf{H}_{X,t}\bm{\kappa}_{X,t} + \textbf{L}_{X,t}\bm{\delta}_{X,t}+ \bm{\xi}_{t}; \hspace{5pt}t = 2,\ldots,T.
\end{align}
\noindent
Here, the $N_{t}\times h$ matrix $\textbf{H}_{X,t}\in \mathcal{C}(\textbf{P}_{X,t})$, the $N_{t}\times l$ matrix $\textbf{L}_{X,t}\in \mathcal{C}(\textbf{P}_{X,t})^{\perp}$, $h$ and $l$ are non-negative integers, $\bm{\kappa}_{X,t}$ is a $h$-dimensional Gaussian random vector, and $\bm{\delta}_{X,t}$ is a $l$-dimensional Gaussian random vector; $t = 2,\ldots,T$. The decomposition in (\ref{decomp2}) is the space-time analogue of the decomposition used for discussion in \citet{Reich} and \citet{hughes}. The use of MI basis functions is equivalent to setting $h$ equal to $r$, $\textbf{H}_{X,t} = \textbf{S}_{X,t}$, and $\textbf{L}_{X,t}$ equal to a $n_{t}\times l$ matrix of zeros for each $t$. As a result, the model based on MI basis functions ignores the variability due to $\{\bm{\delta}_{X,t}\}$. In a similar manner, one can argue that the MI propagator matrix may also ignore other sources of variability.
To address this concern we consider specifying $\{\textbf{K}_{t}\}$ as positive semi-definite matrices that are ``close'' to target precision matrices (denoted with $\textbf{P}_{t}$ for $t = 1,\ldots,T$) that do not ignore these sources of variability; critically, the use of a target precision matrix allows us to reduce the parameter space. Then, let $\textbf{K}_{t} = \sigma_{K}^{2}\textbf{K}_{t}^{*}(\textbf{P}_{t})$, where $\sigma_{K}^{2}>0$ is unknown and
\begin{equation}\label{Kstar}
\textbf{K}_{t}^{*}(\textbf{P}_{t})=\underset{\textbf{C}}{\mathrm{arg\hspace{5pt}min}}\left\lbrace ||\textbf{P}_{t} - \textbf{S}_{X,t}\textbf{C}^{-1}\textbf{S}_{X,t}^{\prime}||_{\mathrm{F}}^{2}\right\rbrace;\hspace{5pt} t = 1,\ldots,T.
\end{equation}
\noindent
Here, $||\cdot||_{\mathrm{F}}$ denotes the Frobenius norm. In (\ref{Kstar}), we minimize the Frobenius norm across the space of positive semi-definite matrices. In a similar manner, if $\textbf{P}_{t}\equiv \textbf{P}$, $\textbf{X}_{t}\equiv \textbf{X}$, $\textbf{S}_{X,t}\equiv \textbf{S}_{X}$, $\textbf{M}_{B,t}\equiv \textbf{M}_{B}$, $\textbf{K}_{t}\equiv \textbf{K}$, and $\textbf{W}_{t}\equiv \textbf{W}$ then we let $\textbf{K} = \sigma_{K}^{2}\textbf{K}^{*}(\textbf{P})$, where $\sigma_{K}^{2}>0$ is unknown and
\begin{equation}\label{Kstarnot}
\textbf{K}^{*}(\textbf{P})=\underset{\textbf{C}}{\mathrm{arg\hspace{5pt}min}}\left\lbrace ||\textbf{P} - \textbf{S}_{X}\textbf{C}^{-1}\textbf{S}_{X}^{\prime}||_{\mathrm{F}}^{2}\right\rbrace.
\end{equation}
\noindent
In the following proposition we show how to compute $\textbf{K}_{t}^{*}$ in (\ref{Kstar}) for $t = 1,\ldots,T$ and $\textbf{K}^{*}$ in (\ref{Kstarnot}).\\
\noindent
\textit{Proposition 1:} Let $\bm{\Phi}_{k}$ be a generic $n \times r$ real matrix such that $\bm{\Phi}_{k}^{\prime}\bm{\Phi}_{k}=\textbf{I}_{r}$, $\textbf{C}$ be a generic $r \times r$ positive definite matrix, $\textbf{P}_{k}$ be a generic $n\times n$ positive definite matrix, and let $k = 1,\ldots,K$. Then, the value of $\textbf{C}$ that minimizes $\sum_{k=1}^{K}||\textbf{P}_{k} - \bm{\Phi}_{k}\textbf{C}^{-1}\bm{\Phi}_{k}^{\prime}||_{F}^{2}$ within the space of positive semi-definite covariances is given by,
\begin{equation}\label{minfrobs}
\textbf{C}^{*} = \left\lbrace\mathcal{A}^{+}\left(\frac{1}{K}\sum_{k = 1}^{K} \bm{\Phi}_{k}^{\prime}\textbf{P}_{k}\bm{\Phi}_{k}\right)\right\rbrace^{-1},
\end{equation}
\noindent
where $\mathcal{A}^{+}$($\textbf{R}$) is the best positive approximate \citep{Higham} of a real square matrix $\textbf{R}$. Similarly, the value of $\textbf{C}$ that minimizes $\sum_{k=1}^{K}||\textbf{P}_{k} - \bm{\Phi}_{k}\textbf{C}\bm{\Phi}_{k}^{\prime}||_{F}^{2}$ within the space of positive semi-definite covariances is given by,
\begin{equation}\label{minfrobs2}
\mathcal{A}^{+}\left(\frac{1}{K}\sum_{k = 1}^{K} \bm{\Phi}_{k}^{\prime}\textbf{P}_{k}\bm{\Phi}_{k}\right).
\end{equation}
\noindent
\textit{Proof:} See {Appendix A}. \\
\noindent
If we let $K=T$, $\bm{\Phi}_{k}=\textbf{S}_{X,k}$ for each $k$, then the corresponding expression of (\ref{minfrobs}) yields $\textbf{K}_{t}^{*}$ in (\ref{Kstar}). Likewise, if we let $K=1$, $\bm{\Phi}_{1}=\textbf{S}_{X}$, then the corresponding expression of (\ref{minfrobs}) yields $\textbf{K}^{*}$ in (\ref{Kstarnot}).
With both $\{\textbf{K}_{t}\}$ and $\{\textbf{M}_{t}\}$ specified we can solve for $\{\textbf{W}_{t}\}$. That is, using the VAR(1) model
\begin{equation}\label{Wstar}
\textbf{W}_{t} = \textbf{K}_{t} - \textbf{M}_{B,t}\textbf{K}_{t-1}\textbf{M}_{B,t}^{\prime}\equiv \sigma_{K}^{2}\textbf{W}_{t}^{*}; \hspace{5pt} t = 2,\ldots,T,
\end{equation}
\noindent
or
\begin{equation}\label{Wstarnot}
\textbf{W} = \textbf{K} - \textbf{M}_{B}\textbf{K}\textbf{M}_{B}^{\prime}\equiv \sigma_{K}^{2}\textbf{W}^{*},
\end{equation}
\noindent
in the case where $\textbf{X}_{t}\equiv \textbf{X}$, $\textbf{S}_{X,t}\equiv \textbf{S}_{X}$, $\textbf{M}_{B,t}\equiv \textbf{M}_{B}$, $\textbf{K}_{t}\equiv \textbf{K}$, and $\textbf{W}_{t}\equiv \textbf{W}$. In (\ref{Wstar}) and (\ref{Wstarnot}), the $r\times r$ matrices $\textbf{W}_{t}^{*} = \textbf{K}_{t}^{*} - \textbf{M}_{B,t}\textbf{K}_{t-1}^{*}\textbf{M}_{B,t}^{\prime}$ and $\textbf{W}^{*} = \textbf{K}^{*} - \textbf{M}_{B}\textbf{K}^{*}\textbf{M}_{B}^{\prime}$; $t = 2,\ldots,T$. It is important to note that the $r\times r$ matrices in the set $\{\textbf{W}_{t}^{*}\}$ (or the $r\times r$ matrix $\textbf{W}^{*}$) may not necessarily positive semi-definite. If $\textbf{W}_{t}^{*}$ is not positive semi-definite for some $t$ then we suggest using the best positive approximate. This is similar to ``lifting'' adjustments suggested by \citet{kang-cressie-shi-2010} in the spatio-temporal setting.
There are many choices for the ``target precision'' matrices $\{\textbf{P}_{t}\}$ in (\ref{Kstar}) and (\ref{Wstar}). For example, one might use the CAR model and let $\textbf{P}_{t} = \textbf{Q}_{t}$, where $\textbf{Q}_{t} = \bm{1}_{N_{t}} - \textbf{A}_{t}$ and $\bm{1}_{N_{t}}$ is a $N_{t}$-dimensional vector of 1s; $t = 1,\ldots,T$. This allows one to incorporate neighborhood information into the priors for $\textbf{K}$ and $\textbf{W}$. In the case where the areal units are small and regularly spaced, one might consider the many spatio-temporal covariance functions that are available (e.g., see \citet{Gneitingcorr}, \citet{HuangCressie2}, and \citet{steinSep}). An empirical Bayesian approach might be considered and an estimated precision (or covariance) matrix might be used (e.g., see \citet{guttorpandpeterson}).
An additional motivation for the restrictions in (\ref{Kstar}) and (\ref{Kstarnot}) is that the MI prior can be interpreted as a special case. This is formally stated in Corollary 1 below.\\
\noindent
\textit{Corollary 1:} Let $\textbf{S}_{X,1}$ be the MI propagator matrix, and $\textbf{C}$ be a generic $r \times r$ positive definite matrix. Then, the value of $\textbf{C}$ that minimizes $||\textbf{Q}_{1} - \textbf{S}_{X,1}\textbf{C}\textbf{S}_{X,1}^{\prime}||_{F}^{2}$ within the space of positive semi-definite covariances is given by,
\begin{equation}\label{minfrobscor}
\textbf{C}^{*} = \mathcal{A}^{+}\left(\textbf{S}_{X,1}^{\prime}\textbf{Q}_{1}\textbf{S}_{X,1}\right).
\end{equation}
\noindent
\textit{Proof:} Let $K = 1$, $\bm{\Phi}_{1} = \textbf{S}_{X,1}$, and $\textbf{P}_{1} = \textbf{Q}_{1}$. Then, apply Proposition 1.\\
\noindent
If $\textbf{S}_{X,1}^{\prime}\textbf{Q}_{1}\textbf{S}_{X,1}$ is positive definite, then (\ref{minfrobscor}) is equal to the MI prior. \citet{aaronpBayes} show that $\textbf{S}_{X,1}^{\prime}\textbf{Q}_{1}\textbf{S}_{X,1}$ is positive definite as long as an intercept is included in the definition of $\textbf{X}_{1}$.
The prior distributions for these parameters are specified to be Gaussian (for $\{\bm{\beta}_{t}\}$) and inverse gamma (IG) priors (for $\sigma_{K}^{2}$, $\sigma_{W}^{2}$, and $\{\sigma_{\xi,t}^{2}\}$). This will allow us to use conjugacy to obtain exact expressions for the full-conditionals within a Gibbs sampler. See {Appendix B} for the details regarding the MCMC algorithm.
\section{Applications}\label{sec:app} To illustrate the variety of random processes that can be modeled using the MSTM, we consider three important problems. In doing so, we demonstrate that the MSTM provides an extremely viable solution to many important problems found in different corners of the spatial statistics literature.
The first problem is modeling temporal nonstationarity. This is a long-standing problem within the time-series literature \citep[e.g., see][]{Dahlhaus}, and we extend it to the multivariate-spatio-temporal setting. This is demonstrated in Section~\ref{sec:cancer}, where we consider US cancer mortality data.
The second problem is multivariate-spatio-temporal prediction (by prediction we mean estimating latent random processes) using data from multiple surveys. The MSTM is flexible enough to solve this problem; here, we only need to set $\{v_{t}^{(\ell)}(\cdot)\}$ (i.e., the variance of $\{\epsilon_{t}^{(\ell)}\}$) equal to the {survey variance provided by each of the available statistical agencies}. In general, there are many models used to combine surveys in the time-series setting \citep[e.g., see][]{Jones,Bell,Feder}, but to our knowledge nothing has been proposed that would be suitable for the multivariate-spatio-temporal data setting. In fact, much of the literature involves simplifications that are not appropriate for our setting; for example, \citet{Merkouris1} and \citet{dorfman} assume marginal independence between surveys and \citet{Elliott}, \citet{Keller}, and \citet{lohr} consider a dual frame design in the non-spatio-temporal setting. \citet{wang} and \citet{Giorgi} consider a fully Bayesian approach to combine surveys in the time series and spatial settings, respectively. For other examples of Bayesian approaches to combining multiple surveys see \citet{Raghunathan}, \citet{Bryant}, and the references therein. Although there has been considerable methodological development for combining data from multiple surveys, none of the previously mentioned approaches incorporate multivariate spatio-temporal dependencies. As such, in Section~\ref{sec:comb}, we use the MSTM in a fully Bayesian framework to combine unemployment rates from ACS and LAUS.
In practice, multivariate-spatio-temporal datasets can be extremely large. Hence, the third problem we consider is spatial prediction using massive datasets. Spatial datasets are becoming increasingly larger and, as a result, new methodologies {(that are not directly applicable in the multivariate-spatio-temporal setting)} are being proposed to address the computational bottleneck involved with spatial prediction using large spatial datasets. Specifically, the Gaussian likelihood involves the computation of an inverse and a determinant of a $n\times n$ covariance matrix; a task that is on the order of $n^{3}$ computations. The reduced rank structure that we impose in (\ref{process:model}) allows us to avoid computing the likelihood of a large dimensional Gaussian random vector if $r \ll n$. In Section~\ref{sec:massive}, we analyze a massive multivariate-spatio-temporal dataset consisting of average monthly income estimates obtained from the US Census Bureau's {LEHD} survey.
For all three of the following examples, the Gibbs sampler, provided in Appendix~B, was run for 10,000 iterations with a burn-in of 1,000 iterations. Convergence of the {Markov chain Monte Carlo algorithm} was assessed visually using trace plots of the sample chains, with no lack of convergence detected.
\subsection{An Analysis of U.S. Cancer Mortality Rates}\label{sec:cancer} The NCI is a division of the National Institutes of Health (NIH). They generate a {wide range} of cancer statistics recorded over the US, which we use to provide an analysis of mortality rates due to cancer (per 100$\--$thousand) by gender (see, http://statecancerprofiles.cancer.gov/). These mortality rates are recorded over each state from 1975 to 2010; that is, let $\ell = 1$ indicate females, $\ell = 2$ indicate males, $D_{\mathrm{O},t}^{(\ell)}$ consists of each state in the US, $T_{L}^{(\ell)} = 1$, and $T_{U}^{(\ell)} = 35$ (for notational convenience we shift the time period from 1975$\--$2010 to 1$\--$35). In Figures 1(a) and 1(b), we provide selected maps of the mortality rates as estimated by NCI. Here, we see that in general NCI estimates higher cancer mortality rates in the east coast than in the west coast for both women and men.
The primary goals of our analysis are: to specify the MSTM in a way that allows for the complex dependencies found in US cancer mortality rates; to estimate cancer mortality rates; and to interpret estimates from the MSTM. Recent studies have compared cancer mortality forecasts based on methods with both stationary
and nonstationary models in time \citep{cancertime,cancertime2}. These studies suggest that the assumption of nonstationarity in time is reasonable and that one would also expect interactions between space and time. Consequently, we specify the MSTM in (\ref{data:model}) and (\ref{process:model}) to allow for these properties.
To obtain nonstationarity in time we require the MI propagator matrix to differ at different time points. Upon inspection of the definitions of $\{\textbf{S}_{X,t}\}$ and $\{\textbf{M}_{B,t}\}$ we see that this can be achieved by letting $\textbf{x}_{t}^{(\ell)}$ change over time. Hence, for illustrative purposes we make the following specifications. Let $\textbf{x}_{t}^{(\ell)}(A)\equiv (1,I(\ell = 2),c_{1}, c_{2}, t, I(\ell = 2)\times c_{1}, I(\ell = 2)\times c_{2})^{\prime}$ be the 7$\--$dimensional vector of known covariates, where $\textbf{c} \equiv (c_{1}, c_{2})^{\prime}$ are the x$\--$y coordinates of the centroid of the generic areal unit $A$. Notice that we allow for two-way interactions between gender so that the changes in $\mu_{t}^{(\ell)}(\cdot)$ per unit increase in time and $\textbf{c}$ are different for each gender. Additionally, we let $r=12$, which is roughly 10$\%$ of the available MI basis functions at each time point $t$. Let $\{\textbf{Q}_{t}\}$, defined below (\ref{Wstar}), be the target precision matrix.
Preliminary analyses ({QQ plots} and histograms) indicate that the assumption of normality appears reasonable. Hence, we apply the MSTM to the NCI cancer mortality rates and in Figures 1(c)$\--$1(f), we provide selected maps of the predicted cancer mortality rates and their corresponding posterior mean squared prediction error (MSPE) for women and men, respectively. Here, we see that in general there are higher cancer mortality rates in the east coast than in the west coast. For males, this pattern is more apparent.
Next, in Figure~2 we plot the posterior mean of the regression parameter corresponding to time (i.e., $\beta_{3,t}$, where $\bm{\beta}_{t}=(\beta_{1,t},...,\beta_{7,t})^{\prime}$). Recall that the presence of time-varying covariates implies nonstationarity in time for the MSTM. Hence, if we observe a pattern that suggests that $\beta_{3,t}\equiv 0$ then the assumption of nonstationarity in time may not be reasonable. We see in Figure~2 that the mean cancer mortality rate decreases exponentially over time. Additionally, the 95$\%$ {pointwise credible intervals} do not contain zero at any time point, which suggests that nonstationarity in time is a reasonable assumption.
\subsection{Combining Missouri Unemployment Rates From ACS and LAUS}\label{sec:comb} {The Bureau of Labor Statistics administers the LAUS program}, which provides unemployment statistics recorded over the US. LAUS defines unemployment as all out-of-work individuals who are available to have a job and have sought work in the past four weeks of taking the survey. In this section, we provide an analysis of these unemployment rates (in percent) recorded over each county in the state of Missouri from 1990 to 2012; that is, $D_{\mathrm{O},t}^{(1)}$ consists of Missouri counties, $T_{L}^{(1)} = 1$, and $T_{U}^{(1)} = 23$ (for notational purposes we shift the time period from 1990$\--$2012 to 1$\--$23).
However, LAUS is not the only data source available for unemployment rates over the US. In particular, ACS also provides 1$\--$year period estimates of the unemployment rates from 2005 to 2012 over counties in Missouri (see http://factfinder2.census.gov/). ACS shares the same definition of unemployment as LAUS. In addition to unemployment rates, we also consider a demographic variable that is related to unemployment rates; specifically, we analyze 1$\--$year ACS period estimates of median household income ($\ell=2$) from 2005 to 2012. In Figures 3(a)$\--$3(c), we provide selected maps of the LAUS unemployment rate, and the unemployment rate and median income as estimated by ACS. The ACS estimates have considerably less spatial coverage than the LAUS estimates, since ACS does not provide 1$\--$year period estimates for every county of Missouri. Additionally, ACS estimates are available across multiple variables, which is not the case for LAUS.
The primary goals of our analysis are to estimate and interpret the unemployment rates and determine whether or not combining the estimates from LAUS and ACS leads to a substantial improvement in the estimates. One difficulty with interpreting ACS estimates is that they often have large margins of error \citep{speilman}. The fact that ACS estimates can have large sampling variability makes it an interesting example in that the incorporation of LAUS may lead to more precise estimates of Missouri unemployment rates.
Preliminary analyses using {QQ plots} and histograms indicate that the logit (log) of the unemployment rates (median income) is roughly Gaussian. Since we assume that the underlying data is Gaussian we treat the logit (log) of the unemployment rates (median income) as $\{Z_{t}^{(\ell)}(\cdot)\}$ in (\ref{data:model}). That is, let $R_{t}^{(\ell)}(\cdot)$ represent the estimates from ACS and LAUS; $\ell = 1,2$ and $t = 1,...,23$. The logit transformation is given by $\mathrm{logit}(w) \equiv \mathrm{log}\{w/(1-w)\}$ for a generic real number $w\in [0,1]$. Then, we define $Z_{t}^{(1)}(\cdot) \equiv \mathrm{logit}(R_{t}^{(1)}(\cdot))$ and $Z_{t}^{(2)}(\cdot) \equiv \mathrm{log}(R_{t}^{(1)}(\cdot))$ for $t = 1,...,23$. The survey variance estimates are approximated on the transformed scale using the Delta-method (e.g., see \citet{delta}).
For illustrative purposes we make the following specifications. Set the target precision matrices equal to $\{\textbf{Q}_{t}\}$ as previously described below (\ref{Wstar}). Let $\textbf{x}_{t}^{(1)}(A)\equiv (1,c_{1}, c_{2})^{\prime}$, where we recall $(c_{1}, c_{2})^{\prime}$ are the x$\--$y coordinates of the centroid of the generic areal unit $A$. Also, let $r=10$, which is roughly 10$\%$ of the available MI basis functions at each time point $t$. In Figures 3(d) and 3(e) we present a selected map of the predicted unemployment rate and the associated estimates of MSPE of the unemployment rate (on the original scale). The values of MSPE at each location is small (on the order of $10^{-7}$); thus, we appear to be obtaining precise estimates of the hidden process. In Figure 4, we plot $(1/|D_{\mathrm{P},t}^{(1)}|)\times \sum_{A}\widehat{Y}_{t}^{(1)}(A)$ versus year $t$, where $\widehat{Y}_{t}^{(\ell)}(\cdot)$ denotes the predictions based on the MSTM. This plot shows a decrease in the unemployment rate until the early 2000s followed by an increase. Then, in 2007 there was a sharp increase in the unemployment rate until 2010. This conforms to intuition since 2007 marks the start of the US housing crisis \citep[e.g., see][]{holt}.
{Now, denote the multivariate spatial predictors based solely on ACS (LAUS) estimates with $\widehat{Y}_{t}^{(\ell,1)}$ ($\widehat{Y}_{t}^{(\ell,2)}$); $t = 1,...,23$ and $\ell = 1,...,T$. Consider the relative leave-one survey-out (RLS) criterion
\begin{equation*}
\mathrm{RLS}(m)\equiv \frac{\sum_{j = 1}^{J}\sum_{t = 1}^{T}\underset{A \in D_{\mathrm{O},t}^{(\ell)}}{\sum}(Y_{t}^{(1)}(A; \textbf{z})^{[j]} - \widehat{Y}_{t}^{(1,m)}(A))^{2}}{\sum_{j = 1}^{J}\sum_{t = 1}^{T}\underset{A \in D_{\mathrm{O},t}^{(1)}}{\sum}(Y_{t}^{(1)}(A; \textbf{z})^{[j]} - \widehat{Y}_{t}^{(1)}(A))^{2}};\hspace{5pt} m = 1,2,
\end{equation*}
where $Y_{t}^{(\ell)}(A; \textbf{z})^{[j]}$ represents the $j$$\--$th MCMC replicate of $Y_{t}^{(\ell)}(A)$ computed using the entire $n$$-$dimensional data-vector $\textbf{z} \equiv (Z_{t}^{(\ell)}(A): \ell = 1,...,L, t = T_{L}^{(\ell)},...,T_{U}^{(\ell)}, A \in D_{\mathrm{P},t}^{(\ell)})^{\prime}$ and $j = 1,...,J$. That is, $\mathrm{RLS}(m)$ is the MSPE of $\widehat{Y}_{t}^{(1)}$ relative to the MSPE of $\widehat{Y}_{t}^{(1,m)}$ for survey $m$. Values of RLS that are smaller than 1 indicate that combining surveys does not lead to an improvement in MSPE, while values larger than 1 indicate an improvement in MSPE. Also, if $1<\mathrm{RLS}(2)< \mathrm{RLS}(1)$ then this indicates that Survey 1 benefits more from combining surveys (in terms of reducing MSPE) than Survey 2. For this example, $\mathrm{RLS}(2) = 3.03\times 10^5$ and $\mathrm{RLS}(1) = 3.01\times 10^7$, which is considerably larger than 1. Hence, we see a dramatic improvement in the MSPE when using both surveys as opposed to using a single survey. Additionally, we see that ACS benefits more from combining surveys than LAUS, since $1<\mathrm{RLS}(2)< \mathrm{RLS}(1)$.}
\subsection{Predicting Average Monthly Income Using a Massive Dataset}\label{sec:massive} We demonstrate the use of MSTM using a massive multivariate-spatio-temporal dataset
made up of data obtained from the LEHD program, which is administered by the US Census Bureau. The LEHD program provides public-use survey data by combining Census Bureau survey data on employers and employees. Public access data on several earnings and other economic variables are available quarterly on various geographies of the US (see, http://www.census.gov/ces/dataproducts/).
In this section, we consider the average monthly income by individuals with steady jobs for each quarter from 1990 to 2013 ($T = 92$), and by US counties ($\{D_{\mathrm{P},t}^{(\ell)}\}$ and $\{D_{\mathrm{O},t}^{(\ell)}\}$ both consist of US counties). These income estimates are available by industry (see Appendix~C for a list) and gender. Each industry/gender combination identifies a unique multivariate-spatio-temporal field; hence, $\ell = 1,...,L = 40$. In total, there are 7,530,037 observations over the entire US in this dataset, which we jointly analyze using the MSTM. We present a subset of this dataset in Figures 5(a) and 5(b). We see that the average monthly income is relatively constant across each county of the state of Missouri, and that men tend to have higher average monthly income than women. This pattern is consistent across the different spatial locations, industries, and time-points.
The primary goals of our analysis are to estimate the average monthly income, and determine whether or not it is computationally feasible to use the MSTM for a dataset of this size. Preliminary analyses indicate that the log average monthly income is roughly Gaussian. Since we assume that the underlying data is Gaussian we treat the log of the average income as $\{Z_{t}^{(\ell)}(\cdot)\}$ in (\ref{data:model}). The survey variance estimates are approximated on the transformed scale using the Delta-method (e.g., see \citet{delta}).
For illustration, we make the following specifications. Set the target precision matrix equal to $\{\textbf{Q}_{t}\}$ as previously described below (\ref{Wstar}). Let $\textbf{x}_{t}^{(\ell)}(A)\equiv (1,I(\ell = 1),...,I(\ell = 39), I(g = 1)\times I(\ell = 1),...,I(g = 1)\times I(\ell = 39))^{\prime}$, where $g = 1,2$ indexes males and females, respectively, and recall $I(\cdot)$ is the indicator function. Also, let $r=20$, which is roughly 30$\%$ of the available MI basis functions at each time point $t$. Using the MSTM with these specifications we predict $L\times T = 40\times 92 = 3,680$ different spatial fields. The CPU time required to compute these predictions is approximately 2.3 days, and all of our computations were performed on a dual 10 core 2.8 GHz Intel Xeon E5-2680 v2 processor, with 256 GB of RAM. Of course, additional efforts in efficient programming may result in faster computing; however, these results indicate that it is practical to use the MSTM to analyze massive data from a computational point of view.
Although we modeled the entire US simultaneously, for illustration, we present maps of predicted monthly income for the state of Missouri, for each gender, for the education industry, and for the 92-nd quarter (Figures~5(c) and 5(d)). The prediction maps are essentially constant over the state of Missouri, where women tend to have a predicted monthly income of slightly less than 1,200 dollars and men consistently have a predicted monthly income of about 1,800 dollars. As observed in Figure~5(a) and (b), there is a clear pattern where men have higher predicted monthly income than women. These predictions appear reasonable since the maps of the posterior square root MSPE, in Figures~5(e) and 5(f), indicate we are obtaining precise predictions. Additionally, upon comparison of Figures~5(a) and 5(b) to Figures 5(c) and 5(d), we see that the predictions reflect the same general pattern in the data. These results are similar across the different states, industries, and time-points.
To further corroborate the patterns in the MSTM predictions we fit the univariate spatial model from \citet{hughes}, which is currently the alternative model for spatial prediction of large areal datasets. We fit the univariate spatial model from \citet{hughes} to the data in Figures 5(a) and 5(b) with $r=62$ basis functions ($100\%$ of the available basis functions) and obtain the prediction maps (not shown). Notably, the predictions are also fairly constant around 1,200 and 1,800 dollars. Moreover, the MSPE of the \citet{hughes} predictions (summed over all US counties) is 3.23 times larger than the MSPE of the predictions from the MSTM summed over all US counties. This may be due, in part, to the fact that the model in \citet{hughes} does not incorporate dependencies that arise from different variables and time-points.
It should be mentioned that, despite the inherent computational issues, having an abundance of data has advantages. For example, notice in Figure~5(b) that LEHD does not release data at two counties of Missouri for men in the education industry during quarter 92. Although these values are missing for this variable and time-point, LEHD releases data at these two counties (for men in the education industry) for 43 different quarters. Hence, with the observed values from 43 different spatial fields, we reduce the variability of predictions at the two missing counties during the 92-nd quarter (compare Figure~5(b) to 5(f)).
\section{Discussion}\label{sec:disc} We have introduced fully Bayesian methodology to analyze areal datasets with multivariate-spatio-temporal dependencies. In particular, we introduce the multivariate-spatio-temporal mixed effects model (MSTM). {To date, little} has been proposed to model areal data that exhibit multivariate-spatio-temporal dependencies; furthermore, the available alternatives (see, \citet{carlinmst} and \citet{daniels}) do not allow for certain complexities in cross-covariances. Hence, the MSTM provides an important addition to the multivariate-spatio-temporal literature.
Our modeling decisions are made in an effort to allow the MSTM to be applied to a wide array of datasets. For example, we use a reduced rank approach to allow for massive multivariate-spatio-temporal datasets. Additionally, we allow for nonstationary and nonseparable multivariate-spatio-temporal dependencies, which is appropriate for many settings. This is achieved, in part, through a novel propagator matrix for a first-order vector autoregressive (VAR(1)) model, which we call the MI propagator matrix. This propagator matrix is an extension of the MI basis function \citep{griffith2000, griffith2002, griffith2004, griffith2007,hughes,aaronp} from the spatial only setting to the multivariate-spatio-temporal setting. We motivate both the MI basis function and the MI propagator matrix as an approximation to a target precision matrix, that allows for (1) computationally efficient statistical inference and (2) identifiability of regression parameters.
We also make an effort to allow practitioners to incorporate knowledge of the spatial process into the MSTM. Specifically, we propose an extension of the MI prior to the spatio-temporal case. This extension shows that the covariance matrix of the random effect is close (in Frobenius norm) to a ``target precision'' matrix, which is chosen based on knowledge of the underlying spatial process. In general, this contribution has implications for defining informative parameter models for high-dimensional spatio-temporal processes.
To demonstrate the effectiveness and broad applicability of our approach, we consider three motivating examples. In the first application, we analyze US cancer mortality rates using the MSTM. Here, nonstationary and nonseparablity are realistic assumptions \citep{cancertime,cancertime2}, which can easily be incorporated into the MSTM. Estimates of the MSPE indicate that the proposed BHM leads to precise predictions and estimates. Furthermore, the results of this study suggest that the assumption of nonstationarity in time is reasonable.
In the second example, we consider combining {data from multiple repeated surveys}, which is a topic of general interest. To demonstrate this, we consider unemployment rates from both ACS and LAUS. Here, the MSTM is used to combine ACS and LAUS estimated unemployment rates in Missouri. Estimates of the MSPE indicate that combining these surveys using the proposed BHM leads to a more precise estimate of unemployment rate than using each survey individually.
In the third example we consider a massive dataset of monthly income. The dataset consists of 7,530,037 observations, which is used to predict 3,680 different spatial fields consisting of all the counties in the US. The recorded CPU time for this example was 2.3 days, which indicates that it is reasonable to use the MSTM for massive data.
There are many opportunities for future research. In particular, the parameter model introduced in Section~\ref{sec:parmod} is of independent interest. In our applications, we let $\{\textbf{Q}_{t}\}$ give the target precision. However, one could conceive of many different ``target precisions'' built from deterministic models, for say, atmospheric variables. Another avenue {for} future research is to develop the MI propagator matrix, which was only provided for VAR(1). One could easily use this strategy for other time series models.
\section*{Acknowledgments} This research was partially supported by the U.S. National Science Foundation (NSF) and the U.S. Census Bureau under NSF grant SES$\--$1132031, funded through the NSF-Census Research Network (NCRN) program.
\section*{Appendix A: The Proof of Proposition 1}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}
By definition of the Frobenius norm
\begin{align}\label{frobnormproof}
\nonumber
&\sum_{k = 1}^{K}||\textbf{P}_{k} - \bm{\Phi}_{k}\textbf{C}^{-1}\bm{\Phi}_{k}^{\prime}||_{F}^{2} = \sum_{k = 1}^{K}\mathrm{trace}\left\lbrace\left(\textbf{P}_{k} - \bm{\Phi}_{k}\textbf{C}^{-1}\bm{\Phi}_{k}^{\prime}\right)^{\prime}\left(\textbf{P}_{k} - \bm{\Phi}_{k}\textbf{C}^{-1}\bm{\Phi}_{k}^{\prime}\right)\right\rbrace \\
\nonumber
&= \sum_{k = 1}^{K}\left\lbrace\mathrm{trace}\left(\textbf{P}_{k}^{\prime}\textbf{P}_{k}\right) - 2\times \mathrm{trace}\left( \bm{\Phi}_{k}^{\prime}\textbf{P}_{k}^{\prime}\bm{\Phi}_{k}\textbf{C}^{-1}\right) + \mathrm{trace}\left(\textbf{C}^{-2}\right)\right\rbrace\\
\nonumber
& = \sum_{k = 1}^{K}\mathrm{trace}\left(\textbf{P}_{k}^{\prime}\textbf{P}_{k}\right)- K\times \mathrm{trace}\left\lbrace \left(\frac{1}{K}\sum_{k = 1}^{K}\bm{\Phi}_{k}^{\prime}\textbf{P}_{k}\bm{\Phi}_{k}\right)^{2}\right\rbrace\\
&+K\times||\textbf{C}^{-1} - \frac{1}{K}\sum_{k = 1}^{K}\bm{\Phi}_{k}^{\prime}\textbf{P}_{k}\bm{\Phi}_{k}||_{F}^{2}.
\end{align}
\noindent
It follows from Theorem 2.1 of \citet{Higham} that the minimum of (\ref{frobnormproof}) is given by Equation (10) in the main document. In a similar manner, if one substitutes $\textbf{C}$ for $\textbf{C}^{-1}$ in (\ref{frobnormproof}) then we obtain the result in Equation (11) in the main document.
\section*{Appendix B: Full-Conditionals for the Gibbs Sampler}
\renewcommand{\theequation}{B.\arabic{equation}}
\setcounter{equation}{0}
The model that we use for multivariate-spatio-temporal data is given by:
\begin{align}\label{summary}
\nonumber
&\mathrm{Data\hspace{5pt}Model:}\hspace{5pt}Z_{t}^{(\ell)}(A)\vert \bm{\beta}_{t}, \bm{\eta}_{t}, \xi_{t}^{(\ell)}(\cdot)\stackrel{\mathrm{ind}}{\sim} \mathrm{Normal}\left(\textbf{x}_{t}^{(\ell)}(A)^{\prime}\bm{\beta}_{t} + \textbf{S}_{X,t}^{(\ell)}(A)^{\prime}\bm{\eta}_{t} + \xi_{t}^{(\ell)}(A), v_{t}^{(\ell, m)}(A)\right);\\
\nonumber
&\mathrm{Process\hspace{5pt}Model\hspace{5pt}1:}\hspace{5pt} \bm{\eta}_{t}\vert \bm{\eta}_{t-1},\textbf{M}_{B,t},\textbf{W}_{t}\sim \mathrm{Gaussian}\left(\textbf{M}_{B,t}\bm{\eta}_{t-1}, \textbf{W}_{t}\right);\\
\nonumber
&\mathrm{Process\hspace{5pt}Model\hspace{5pt}2:}\hspace{5pt} \bm{\eta}_{1}\vert \textbf{K}_{1} \sim \mathrm{Gaussian}\left(\bm{0}, \textbf{K}_{1}\right);\\
\nonumber
&\mathrm{Process\hspace{5pt}Model\hspace{5pt}3:}\hspace{5pt} \xi_{t}^{(\ell)}(\cdot)\vert \sigma_{\xi,t}^{2} \stackrel{\mathrm{ind}}{\sim} \mathrm{independent\hspace{4pt}Normal}\left(0, \sigma_{\xi,t}^{2}\right);\\
\nonumber
&\mathrm{Parameter\hspace{5pt}Model\hspace{5pt}1:}\hspace{5pt} \bm{\beta}_{t} \sim \mathrm{Gaussian}\left(\bm{\mu}_{\beta}, \sigma_{\beta}^{2}\textbf{I}_{p}\right);\\
\nonumber
&\mathrm{Parameter\hspace{5pt}Model\hspace{5pt}2:}\hspace{5pt} \sigma_{\xi,t}^{2} \sim \mathrm{IG}\left(\alpha_{\xi}, \beta_{\xi}\right);\\
\nonumber
&\mathrm{Parameter\hspace{5pt}Model\hspace{5pt}3:}\hspace{5pt} \sigma_{K,t}^{2} \sim \mathrm{IG}\left(\alpha_{K}, \beta_{K}\right);\hspace{5pt} \ell = 1,...,L, t = T_{L}^{(\ell)},...,T_{U}^{(\ell)},A \in D_{\mathrm{P},t}^{(\ell)},
\end{align}
\noindent
where $\sigma_{\beta}^{2}>0$, $\alpha_{\xi}>0$, $\alpha_{K}>0$, $\beta_{\xi}>0$, and $\beta_{K}>0$. In Section 4 the prior mean of $\bm{\mu}_{\beta}$ is set equal to a $p$$\--$dimensional zero vector, and the corresponding variance $\sigma_{\beta}^{2}$ is set equal to $10^{15}$ so that the prior on $\{\bm{\beta}_{t}\}$ is vague. In Section 4, we also specify $\alpha_{\xi}$, $\alpha_{K}$, $\beta_{\xi}$, and $\beta_{K}$ so that the prior distributions of $\sigma_{\xi,t}^{2}$ and $\sigma_{K,t}^{2}$ are vague. Specifically, we let $\alpha_{\xi}=\alpha_{K}=2$, and $\beta_{\xi}=\beta_{K}=1$; here, the IG(2,1) prior is interpreted as vague since it has infinite variance.
We now specify the full-conditionals for the process variables (i.e., $\{\bm{\eta}_{t}\}$ and $\{\xi_{t}^{(\ell)}(\cdot)\}$) and the parameters (i.e., $\{\bm{\beta}_{t}\}$, $\{\sigma_{\xi,t}^{2}\}$, and $\sigma_{K}^{2}$).\\
\noindent
\textit{Full-Conditionals for Process Variables:} Let the $n_{t}$$\--$dimensional random vectors $ \bm{\textbf{z}}_{t} \equiv \left(Z_{t}^{(\ell)}(A): \ell\right.$ $\left. = 1,...,L, A \in D_{\mathrm{O},t}^{(\ell)}\right)^{\prime}$, $ \bm{\xi}_{t} \equiv \left(\xi_{t}^{(\ell)}(A): \ell = 1,...,L, A \in D_{\mathrm{O},t}^{(\ell)}\right)^{\prime}$, and the $n_{t}\times p$ matrix $\textbf{X}_{t} \equiv \left(\textbf{x}_{t}^{(\ell)}(A):\right.$ $\left.\ell = 1,...,L, m = 1,...,M^{(\ell)}, A \in D_{\mathrm{O},t}^{(\ell)}\right)^{\prime}$; $t = 1,...,T$. Then, we update the full-conditional for $\bm{\eta}_{1:T} \equiv \left(\bm{\eta}_{t}^{\prime}: t = 1,...,T\right)^{\prime}$ at each iteration of the Gibbs sampler using the Kalman smoother. We accomplish this by performing the following steps:
\begin{enumerate}
\item Find the Kalman filter using the shifted measurements $\{\widetilde{\textbf{z}}_{t}: \widetilde{\textbf{z}}_{t} = \textbf{z}_{t}-\textbf{X}_{t}\bm{\beta}_{t}-\bm{\xi}_{t}\}$\citep{shumway,cart1994,schnatter94,cressie-wikle-book}. That is, for $t = 1,...,T$ compute
\begin{enumerate}
\item $\bm{\eta}_{t\vert t}^{[j]} \equiv E\left(\bm{\eta}_{t}\vert \widetilde{\textbf{z}}_{1:t}, \bm{\theta}_{t}^{[j]}\right)$
\item $\bm{\eta}_{t\vert (t-1)}^{[j]} \equiv E\left(\bm{\eta}_{t}\vert \widetilde{\textbf{z}}_{1:(t-1)}, \bm{\theta}_{t}^{[j]}\right)$
\item $\textbf{P}_{t\vert t}^{[j]} \equiv \mathrm{cov}\left(\bm{\eta}_{t}\vert \widetilde{\textbf{z}}_{1:t}, \bm{\theta}_{t}^{[j]}\right)$
\item $\textbf{P}_{t\vert (t-1)}^{[j]} \equiv \mathrm{cov}\left(\bm{\eta}_{t}\vert \widetilde{\textbf{z}}_{1:(t-1)}, \bm{\theta}_{t}^{[j]}\right)$,
\end{enumerate}
\noindent
where $\textbf{P}_{1\vert 1}^{[j]} = (\sigma_{K}^{[j]})^{2}\textbf{K}^{*}$ and $\bm{\theta}_{t}^{[j]}$ represents the $j$$\--$th MCMC draw of $\bm{\theta}_{t}$ and $\sigma_{K}^{2}$, respectively.
\item Sample $\bm{\eta}_{T}^{[j+1]} \sim \mathrm{Gaussian}\left(\bm{\eta}_{T\vert T}^{[j]}, \textbf{P}_{T\vert T}^{[j]}\right)$.
\item For $t = T-1,T-2,...,1$ sample\\$\bm{\eta}_{t}^{[j+1]} \sim \mathrm{Gaussian}\left(\bm{\eta}_{t\vert t}^{[j]} + \textbf{J}_{t}^{[j]}(\bm{\eta}_{t+1}^{[j]} - \bm{\eta}_{t+1\vert t}^{[j]}), \textbf{P}_{t\vert t}^{[j]} - \textbf{J}_{t}^{[j]}\textbf{P}_{t+1\vert t}^{[j]}(\textbf{J}_{t}^{[j]})^{\prime} \right)$,\\where $\textbf{J}_{t}^{[j]} \equiv \textbf{P}_{t\vert t}^{[j]} \textbf{M}_{t}^{\prime}(\textbf{P}_{t+1\vert t}^{[j]})^{-1}$.
\end{enumerate}
Notice that within each MCMC iteraction we need to compute Kalman filter and Kalman smoothing equations. This adds more motivation for reduced rank modeling; that is, if $r$ is large (i.e., if $r$ is close in value to $n$) this step is not computationally feasible.
The remaining process variable $\{\xi_{t}^{(\ell)}(\cdot)\}$ can also be computed efficiently \citep{ravishank}. The full conditional for $\{\xi_{t}^{(\ell)}(\cdot)\}$ is given by: $\bm{\xi}_{t} \sim \mathrm{Gaussian}\left(\bm{\mu}_{\xi,t}^{*},\bm{\Sigma}_{\xi.t}^{*}\right),$ where $\bm{\Sigma}_{\xi,t}^{*} \equiv \left(\textbf{V}_{t} + \sigma_{\xi}^{2}\textbf{I}_{N_{t}}\right)^{-1}$, $\bm{\mu}_{\xi,t}^{*} \equiv \bm{\Sigma}_{\xi}^{*}\times \textbf{V}_{t}^{-1}\times(\textbf{z}_{t}-\textbf{X}_{t}\bm{\beta}_{t}-\textbf{S}_{t}\bm{\eta}_{t})$, $\textbf{V}_{t} \equiv \mathrm{diag}\left(v_{t}^{(\ell)}(A): \ell = 1,...,L, A \in D_{\mathrm{O},t}^{(\ell)}\right)$, and $\textbf{S}_{t} \equiv
\left(\textbf{S}_{t}^{(\ell)}(A): \ell = 1,...,\right.$ $\left. L,A \in D_{\mathrm{O},t}^{(\ell)}\right)^{\prime}$; $t = 1,...,T$.\\
\noindent
\textit{Full-Conditionals for the Parameters:} Similar to the full-conditional for $\{\xi_{t}^{(\ell)}(\cdot)\}$ \citep{ravishank} we also have the following full-conditional for ${\bm{\beta}_{t}}$:
$\bm{\beta}_{t} \sim \mathrm{Gaussian}\left(\bm{\mu}_{\beta,t}^{*},\bm{\Sigma}_{\beta,t}^{*}\right)$, where $\bm{\Sigma}_{\beta,t}^{*} \equiv \left( \textbf{X}_{t}^{\prime}\textbf{V}_{t}^{-1}\textbf{X}_{t}+ \sigma_{\beta}^{-2}\textbf{I}_{p} \right)^{-1}$, and
$\bm{\mu}_{\beta,t}^{*} \equiv \bm{\Sigma}_{\beta}^{*}\times \textbf{X}_{t}^{\prime}\textbf{V}_{t}^{-1}(\textbf{z}_{t}-\bm{\xi}_{t}-\textbf{S}_{t}\bm{\eta}_{t})$; $t = 1,...,T$.
Finally, the exact form of the full-conditionals for $\sigma_{K}^{2}$, $\sigma_{W}^{2}$, and $\{\sigma_{\xi,t}^{2}\}$ can also be found in a straightforward manner. It follows that the full conditionals for $\sigma_{K}^{2}$, $\sigma_{W}^{2}$, and $\sigma_{\xi,t}^{2}$ are IG($Tr/2 + 2$, $1 + \bm{\eta}_{1}^{\prime}\textbf{K}_{1}^{*-1}\bm{\eta}_{1}/2 + \sum_{t=2}^{T}(\bm{\eta}_{t}-\textbf{M}_{t}\bm{\eta}_{t-1})^{\prime}\textbf{W}_{t}^{*-1}(\bm{\eta}_{t}-\textbf{M}_{t}\bm{\eta}_{t-1})/2$), and IG($n/2 + 2$, $1 + \bm{\xi}_{t}^{\prime}\bm{\xi}_{t}/2$) (for $t = 1,...,T$), respectively.
\section*{Appendix C: List of Industries used in Section 4.3} We list the different industries that were jointly analyzed in Section 4.3 below.
\singlespacing
\begin{enumerate}
\item Agriculture, Forestry, and Fishing and Hunting
\item Mining, Quarrying, and Oil and Gas Extraction
\item Utilities
\item Construction
\item Manufacturing
\item Wholesale Trade
\item Retail Trade
\item Transportation and Warehousing
\item Information
\item Finance and Insurance
\item Real Estate, and Rental and Leasing
\item Professional, Scientific, and Technical Services
\item Management of Companies and Enterprises
\item Administrative, and Support, Waste Management, and Remediation Services
\item Educational Services
\item Health Care and Social Assistance
\item Arts, Entertainment, and Recreation
\item Accommodation and Food Services
\item Public Administration
\item Other Services
\end{enumerate}
\singlespacing
\bibliographystyle{jasa}
|
\section{INTRODUCTION}
\label{sec:intro}
The Planet Formation Imager (PFI) is a future world facility that will image the process of planetary formation, especially the formation of giant planets. It will complement the Atacama Large Millimetre Array (ALMA) which will image the influence of planetary formation on disks, instruments like SPHERE and GPI, which will image mature planetary systems, and long-term radial velocity monitoring programs, which are now beginning to collect statistics on Jupiter analogs\cite{Wittenmyer14}. PFI will have an angular resolution and sensitivity sufficient to resolve structures smaller than the sphere of gravitational influence of a newly formed giant planet - its {\em Hill sphere} defined by:
\begin{equation}
r_H = a \sqrt[3]{\frac{m_p}{M_s}},
\end{equation}
where $a$ is the semi-major axis of the planetary orbit, $m_p$ is the planet mass and $M_s$ is the star mass. This resolution requirement for a Jupiter-mass planet at a 1\,AU separation is approximately 0.05 AU. In order to truly image the range of separations most likely to form planets, a 20 AU field of view is required\cite{Kraus12}, and the instrument would have to be sensitive enough to carry out a comprehensive survey of solar-type stars in nearby star forming regions at ~140\,pc (e.g. Upper Scorpius and Taurus). For the purposes of this paper, we will assume that these science requirements apply as a minimum to the astronomical N-band (10--13\,$\mu$m) as justified below, and are taking the following as the instrumental requirements:
\begin{itemize}
\item Fringe tracking for mid-infrared coherencing as faint as H=9.0, N=7.5 (a young solar-type star, e.g. LkCa~15, in Taurus or Upper Scorpius).
\item 1-$\sigma$ sensitivities of N=15 in 10 hours of integration, including any ``speckle-noise'' or cross-talk, at a working angle of 7 milli-arcsec or further from an N=7.5 central star.
\item An angular resolution of 0.35\,milli-arcsec or better (0.05\,AU at 140\,pc).
\item An imaging field of view of at least 140\,milli-arcsec radius.
\end{itemize}
The details of these requirements will be debated over the coming months and years by the science working group of the Planet Formation Imager collaboration, but these are a at least one set of requirement for a fully-functional instrument. We note that the adaptive optics imaging work of e.g. Follette et al\cite{Follette13} for the SR~21 transitional disk system shows that, at least at moderately low spatial resolution, scattering by small grains can give a relatively large signal from complex structures not necessarily associated with a disk or planets. Thermal radiation from a newly-formed planet and its environment should overwhelm scattered light from low-mass optically-thin dust structures, and it is the thermal radiation that we will focus on in this paper.
\begin{figure}
\includegraphics[width=\textwidth]{schematic3.jpg}
\caption{A system level diagram of the heterodyne PFI concept. For each of N telescopes, the wavelengths are split in 3: 10--13\,$\mu$m radiation for heterodyne interferometry,
1.1--1.8\,$\mu$m radiation for fringe tracking, and 0.6--1.0\,$\mu$m radiation for adaptive optics wavefront sensing. The 1.1--1.8\,$\mu$m radiation is transported by fiber to long
vacuum delay lines and a conventional fringe tracker. The 10--13\,$\mu$m radiation is mixed with a laser frequency comb, split in polarisation and sent through a high-resolution
spectrograph. The 4\,THz output bandwidth of each pair of telescopes, for each polarisation is correlated in real time, using the 1.1--1.8\,$\mu$m fringe tracker for post-processing
coherencing.}
\label{figSchematic}
\end{figure}
The luminosity output by a forming proto-Jupiter is thought to be of order $2 \times 10^{-4}$\,L$_{\odot}$, over a timescale of a few $\times 10^5$ years\cite{Lissauer09}. The planetary temperature is of order 1500\,K with a background temperature in the disk of order 150\,K, a stellar temperature of order 4000\,K and a stellar luminosity of order 1\,$L_\odot$. Naeively, this would correspond to a required contrast of $4 \times 10^{-4}$ at the blackbody peak at $\sim$2\,$\mu$m (8.4 magnitudes) and $3 \times 10^{-3}$ at 11.5\,$\mu$m (6.4 magnitudes), however this does not take into account re-processing of radiation by the circumplanetary disk and environment, which makes the longer wavelengths more suitable. If the complete $\sim2 \times 10^{-4}$\,L$_{\odot}$ is re-processed by a 500\,K disk, then the disk/star ratio at 11.5\,$\mu$m could be as much as 0.05 (3 magnitudes). Conversely, this reduction in effective temperature of the planet to 500\,K by circumplanetary material (still well within the Hill sphere) would be enough to render 2\,$\mu$m useless, although we concede that alternative designs focusing on the 3--4\,$\mu$m region may also be suitable, and would be less influenced by the background radiation from the protoplanetary disk.
The resolution requirement of one Hill sphere radius is not just to enable the separation of signals from neighbouring planets and to maximise the ability to detect bright point-source like emission from the planet. It is also essential to separate the planetary emission from background disk thermal radiation. If, for example, a planet is forming in a disk gap that is only as wide as its Hill sphere, then the planetary signal would rapidly become dwarfed by the emission from the disk at a few Hill radii.
In this paper, we argue that a heterodyne design is competitive to direct detection designs for detecting this thermal radiation and should be considered as one of the options for PFI. This design has a low-order adaptive optics system creating a collimated beam which is mixed with a laser frequency comb. This mixed signal is then fed into a spectrograph disperses the mixed signal onto an linear array of individual detectors, ideally one for each polarisation. We begin by describing the overall interferometer architecture and sensitivity, then move into plausible designs for the individual components. A schematic system diagram is given in Figure~\ref{figSchematic}.
\section{ARRAY FORMAT}
At a central N-band wavelength of 11.5\,$\mu$m, the angular resolution requirement (taken to be $\lambda/B$) corresponds to maximum baseline of 7\,km. Allowing for some super-resolution might enable baselines to be reduced to 3\,km, but not without a reduction in sensitivity due to possible contamination by the disk emission as discussed above. The imaging requirement for PFI is essential because complex structures in the disk could mask planetary signals if ambiguous modelling of interferometric data are needed. The minimum number of baselines is roughly the number of resolution elements across the final image - 400 in our case. This drives PFI to consider a large number of telescopes - of order 30 or more, noting that the number of baselines $N_B = N_T (N_T-1)/2$ with $N_T$ the number of telescopes. This number of telescopes is inadequate for snapshot imaging, and requires earth rotation synthesis to get the full $\sim$400 x 400 resolution element image. Snapshot imaging fidelity is arguably not required, as the smallest timescale required to be resolved is $\sim$ half the rotation period of the proto-Jupiter, or a few hours. Given the need for ground-based observations sensitive to atmospheric conditions to avoid high air-masses, it is also important that earth rotation synthesis can be achieved in only a few hours of observing and not a full night. Taken together, these requirements mean that a 3-arm spiral (e.g. Figure~\ref{figFormat}) or a Y-shaped array with a format similar to the VLA is close to optimal.
\begin{figure}
\includegraphics[width=0.5\textwidth]{format.jpg}
\caption{An example 30-telescope array configuration, with maximum baseline 7\,km, minimum baseline 82\,m and excellent imaging and baseline-bootstrapping properties.}
\label{figFormat}
\end{figure}
The aperture size of the individual telescope elements also can not be so large that the $\sim$20\,AU radius field of view is outside the primary beam. For a 11.5$\mu$m wavelength, a requirement of 20\,AU/140\,pc $< \lambda/D$ means that the telescope diameter of individual array elements should not be larger than 8\,m.
\section{SENSITIVITY}
Perhaps the most important requirement for sensitivity is the telescope diameter. For an array limited by the number of telescope elements (e.g. if the beam-combiners and delay-line architecture are most important in the total cost), the appropriate question is : `` what is the minimum telescope size required to achieve the sensitivity goals?". For an array where the cost is driven by the telescope cost, we note that more small telescopes are cheaper\cite{vanBelle04}, so the appropriate question is: ``what is the number of telescopes of fixed size required to achieve the sensivity goals?". We will consider both these questions - the minimum telescope diameter if there are 30 telescopes in the array, and the number of telescopes in the array if they are all 2\,m telescopes.
The sensitivity of Heterodyne detection is given by\cite{Hale97}:
\begin{equation}
\left( \frac{S}{N} \right )_h = \frac{P_\nu}{h\nu} \sqrt{2 \Delta \nu t},
\end{equation}
where $P_\nu$ is the received power per polarisation per unit bandwidth (Hz). Note that the thermal background does not appear in this equation, because the quantum noise for heterodyne detection is equivalent to a thermal background of $T=h\nu/k_B$, or 1250\,K at 11.5\,$\mu$m.
A simple thought experiment where a large aperture is split into many small ones, and the voltages added prior to forming an autocorrelation, makes it easy to see that this formula applies to an interferometer or to single-dish heterodyne detection, so long as al the amplitude, phase and auto-correlation information is available. In the case of unavailable (or useless for imaging) auto-correlations and antenna phases, this signal-to-noise is reduced by a factor of $(N_T-1)/N_T$. Taking the 11.5\,$\mu$m flux of Vega as 17\,Jy per polarisation, we then arrive at the overall signal-to-noise per polarisation at 11.5\,$\mu$m:
\begin{equation}
\left( \frac{S}{N} \right )_h = 9.8 \times 10^{-6} {\rm m}^{-2} A_T \eta (N_T-1) \sqrt{2 \Delta \nu t}~10^{-0.4 m_N},
\end{equation}
where the total collecting area per telescope is $A_T$, the target Vega magnitude is $m_N$ and the system efficiency is $\eta$. For a system bandwidth $\Delta \nu$=4\,THz and $\eta=0.35$ with a dual-polarisation design (or $\eta=0.5$ with a single-polarisation design), the requirement for (S/N)=1 for $m_N = 15$ in 10 hours gives a minimum telescope diameter of 4\,m for $N_T$=30. With the same bandwidth and system efficiency, but 2\,m telescopes, the minimum number of telescopes is 136. If the efficiency $\eta$ is halved, then the required number of telescopes doubles. This clearly shows that the heterodyne efficiency must be maximised.
\section{SPECTROGRAPH DESIGN}
Starlight from a low-order adaptive optics system is first mixed with the Laser Frequency Comb (LFC) with 3\,GHz line spacing via a beamsplitter of low (e.g. 5\%) reflectivity, that reflects the 5\% of wasted starlight toward a cold stop. Given that gratings are only efficient in one polarisation, the most efficient time to split the polarisations is immediately after mixing with the LFC. We note that if e.g. the LFC is circularly polarised, it will effectively mix with both outputs of a linear polarising beamsplitter.
This mixed radiation is then dispersed using a spectrograph that has a grating used at 40 degree angle of incidence with a 110mm pupil, with each comb line landing on a separate single cooled HgCdTe detector with a 2 GHz bandwidth. This is roughly the maximum achievable bandwidth for a HgCdTe photodiode\cite{Rogalski05}. Although not simple mechanically, a large number of these photodiodes (2000 in our case) could be placed next to each other, injected by individual micro lenses, and a separate 2\,GHz output extracted from each photodiode. An example configuration is shown in Figure~\ref{figSpectrograph}.
\begin{figure}
\includegraphics[width=0.7\textwidth]{spectrograph.jpg}
\caption{An example spectrograph configuration that would give $\sim$0.4mm physical separation between laser lines and would be able to have the appropriate resolving power of R$\sim$17000 in order to separate the
comb lines of 3\,GHz spacing. The input pupil in this case is 110\,mm, the focal-length of the spherical mirror 1\,m and the grating Littrow angle 40 degrees. Pupils as large as 200\,mm could be accommodated in this arrangement with smaller grating angles, but would require more space.}
\label{figSpectrograph}
\end{figure}
Unfortunately, the starlight landing on each individual detector element is dispersed, so that only the wavelengths that exactly match the comb lines will have a perfect spatial overlap. The example spectrograph here would have R$\sim$17000, which is enough to have the comb-lines separated by $2\lambda/D_{\rm pupil}$, with the starlight 1\,GHz from each comb line having a reduced overlap integral by a factor of 0.65. There are also small effects of cross-talk (see Figure~\ref{figOverlap}).
\begin{figure}
\includegraphics[width=0.6\textwidth]{overlap.jpg}
\caption{The dispersed image-plane, showing the comb lines (solid lines), starlight at the +/- 1\,GHz bandpass edges (dashed lines) and the edges of one resolution element (dotted lines).}
\label{figOverlap}
\end{figure}
\section{LASER FREQUENCY COMB}
Laser frequency combs, in development for more then 15 years\cite{Holzwarth00} are now available commercially in the visible and near-infrared wavelength ranges. In order to produce a mid-infrared laser frequency comb, one solution is to mix a 1.55\,$\mu$m now commercially available comb (e.g. the FC1500-250-WG from Menlo Systems) filtered by a Fabry-Perot\cite{Steinmetz09}, with a stabilised tuneable ~1400nm diode laser. Difference frequency generation is achievable using non-linear materials such as GaSe, producing comb lines with a target 3 GHz separation throughout the 10-13\,$\mu$m bandpass. We estimate that a conversion efficiency of $\sim$0.1\% is achievable, which could produce power after of order 1\,$\mu$W per comb line after mixing. Over a 2\,GHz bandwidth, the laser power is not a significant source of noise as long as many more than 2$\times 10^9$ photons/s are detected by each HgCdTe photodiode - equivalent to a laser power of $>$30\,pW, a requirement which is relatively easy to meet. The more relevant requirement for the LFC and detector system is that the Noise Equivalent Power (NEP) due the quantum noise from the laser fluctuations in each comb mode far exceeds the NEP of the detector itself. This is certainly achieved by 77\,K photodiodes of $<$0.1\,mm size and NEP values less than $10^{-13}$\,W\,Hz$^{1/2}$ and comb powers of order 1\,$\mu$W per comb line, but it could prove to be a difficult requirement to meet with a full system model and high-bandwidth photodiodes.
\section{CORRELATOR REQUIREMENTS}
A key difference between this correlator and radio correlators is that there may be no need for many channels within each 2\,GHz bandpass, as the $\sim$30\,km\,s$^{-1}$ resolving power of each 2\,GHz channel is at least enough to separate strong lines form the continuum, and to measure spectral features due to dust. With a 30 telescope design, the 4\,THz full bandwidth of a heterodyne PFI requires $\sim$350 times more processing power than ALMA, i.e. correlating few-bit samples at 2.8\,PHz. Assuming that a PHz for few-bit samples equivalent to a Pflop, this processing requirement is already less than the 10 most powerful supercomputers today. The power requirement (approaching the dominant cost for super-computers) is approximately 2\,MW, which is smaller than typical operating costs of a typical single large telescope today. For this reason, especially once typical computer speed and power requirement scaling laws are taken into account, we do not expect the correlator to be a significant cost component to a heterodyne PFI for a construction date $\sim$10 or more years in the future.
\section{FRINGE TRACKER}
Even in the optimistic case with telescopes of 8\,m diameter and $\eta = 0.35$ for a dual polarisation design, signal-to-noise at N=7.5 star on a single baseline is only $\sim$0.2 in a 100\,ms coherence time. This means that out-of-band fringe-tracking is needed, with the fringe phase corrected in post-processing, at least over a timescale of $\sim$10 seconds. This means that a companion direct-detection beam combiner is required. Once the signal-to-noise at 11.5\,$\mu$m significantly exceeds 1 (as expected in 10\,s), phase referencing at 11.5$\mu$m can be used, with closure-phase techniques. Note that water vapour seeing is not expected to be significant: a 10\,s timescale corresponds to only approximately 0.15 radians RMS phase error due to water vapour seeing at 11.5\,$\mu$m with fringe tracking at 1.6\,$\mu$m\cite{Colavita04}.
Our conceptual design for this companion instrument is a H-band (1.65\,$\mu$m) system with length-matched and controlled single-mode fibers for beam transport and a delay line system incorporating both a fixed and variable delay line components, and with Gaussian beam waists occurring part-way through each delay system. With a 6\,km Fresnel length for a 1.65\,$\mu$m wavelength and a $\sim$0.1\,m diameter beam, relatively small vacuum pipes ($\sim$0.2\,m diameter) can be used for the $\sim$30 fixed delay components.
Based on simple SNR considerations and anticipated full system H-band efficiencies exceeding 1\% (including effects of fiber coupling), fringe tracking on a H=7.5 star should in principle be relatively straightforward for telescope diameters greater than 1\,m and modern, low noise detectors\cite{Finger10}. However, with the last generation of detectors, active fringe tracking has proved moderately difficult, with magnitude limits of H$\sim$7.5 on the challenging end for the VLTI with the 1.8\,m Auxiliary telescopes\cite{Sahlmann10}, but certainly anticipated as routine with adaptive optics and new detectors. For this reason, a minimum telescope diameter for 2\,m is needed for PFI.
The fringe tracker itself should almost certainly neither be a pairwise nor an all-in-one combiner. Baselines longer than $\sim$2\,km over-resolve solar-type young stars in nearby star forming regions ($\sim$0.15\,milli-arcsec diameters), so the longest baselines would not be useful for fringe tracking. Given the likelihood of at least one telescope failure, a pairwise combiner is also likely not optimal, with an appropriate compromise being one where each telescope is combined with $\sim$4 other telescopes (generally nearest neighbours). This would still enable data from the fringe tracker to be used for short-wavelength imaging at a spatial resolution exceeding the long wavelength light.
\section{COST CONSIDERATIONS}
The largest clear identifiable cost for the array is the telescopes themselves. 30 telescopes of 4\,m diameter is a total size comparable to the Giant Magellan Telescope (GMT) in terms of collecting area. Given the scaling laws of van Belle et al.\cite{vanBelle04}, one might expect the cost of the telescope component for many 4\,m telsscopes to be smaller by a factor of $\sim$2 than GMT, especially given that the telescopes only have to be designed once and constructed many times. In any case, as long as the ratio between instrumentation and telescope costs are not so different for PFI as for extremely large telescopes, a total construction cost in the range 0.5 to 1\$B USD appears plausible. The lower limit on the cost envelope for PFI is the moving mass (structural steel), glass (area) and building (high end commercial space) costs, which is of order 0.1\$B USD - this is of course unrealistically small.
\section{DIRECT-DETECTION COMPARISON}
For the same bandwidth and an assumed efficiency $\eta=0.35$, there is a factor of $\sim$10 improvement in signal-to-noise for direct detection interferometry over heterodyne interferometry for a single telescope\cite{Hale97}. This efficiency may be unrealistic for direct detection in particular, given typical overall efficiencies of interferometers significantly lower than this. A direct detection interferometer could be highly optimised in principle, with bare gold gratings giving 99\% reflectivity at mid-infrared wavelengths and losses due to surface roughness being negligible.
Where an interferometer needs to image a complex scene, direct detection suffers from an additional loss where each beam has to be split many ways in pairwise combination, or where an all-in-one combiner, one has to measure signal from each pair of telescopes in the presence of photon and background noise from all other telescopes. We will assume that an all-in-one combiner is used: this additional noise factor is roughly
$\sqrt{N_T}$, meaning that with 30 telescopes, direct detection still retains a factor of $\sim$2 advantage in principle over heterodyne interferometry.
However, the beam-combiner could not have a very simple design, as the field of view for PFI requires a spectral resolution of at least R$\sim$400. One example design would be an integral field unit with 100 spectral resolution elements behind a 30-aperture non-redundant aperture mask. The complexity of such a system is roughly equivalent to the high resolution spectrograph needed for heterodyne interferometry. Hence, a first-order assumption is that with the exception of the additional $\sim$10--15 mirror beam train, the direct-detection and heterodyne techniques have the same efficiency. Overall, the direct detection scheme might then (optimistically) have $\sim$80\% of the dispersed heterodyne technique's efficiency.
A cost unique to the direct detection scheme is the vacuum pipes for beam transport. The maximum pipe length of 4\,km, has a corresponding Fresnel length of 0.21m at 11.5$\mu$m. Assuming no optics along the pipe, this would require a minimum pipe inner diameter of $\sim$0.4\,m. In the beam-combining laboratory, dual off-axis paraboloid mirrors specific to each telescope would need to re-image the pupil and send the beams into the vacuum delay lines.
The delay line beam size is also required to be scaled up with cross sectional area proportional to wavelength, with a beam diameter of at least $\sim$0.3\,m as well, perhaps requiring a 0.7\,m diameter vacuum pipe for the delay line. Assuming a cost scaling proportional to cross-sectional area, delay lines appropriate to 11.5\,$\mu$m would have $\sim$7 times the cost of delay line appropriate to H-band (1.65\,$\mu$m).
\section{CONCLUSIONS}
We have demonstrated that heterodyne interferometry is a plausible competitor to direct detection interferometry for the Planet Formation Imager concept. Critical to the design is mixing starlight with a frequency comb laser in each telescope, dispersing the light in a high-resolution spectrograph and detecting the light mixed with each line of the frequency comb individually. Fringe tracking using direct detection interferometry is still required, but this is significantly simplified by transporting the beams from the telescopes to the delay lines by fiber. The overall cost of the project is likely comparable to the low-end of ELT budgets.
\acknowledgments
M. Ireland would like to acknowledge support from the Australian Research Council under the Future Fellowships scheme (FT130100235).
|
\section{Introduction}
The Wiener Sausage problem \cite{KolmogoroffLeontowitsch:1933}
is concerned with determining the volume traced out (the sausage) by a $d$-dimensional
sphere attached to a Brownian particle in $d$ dimensions. The problem is
illustrated in \Fref{2d_example} in dimension $d=2$. It has been studied
extensively in the literature \cite{Montroll:1964,Spitzer:1964,KacLuttinger:1974,DonskerVaradhan:1975,LeGall:1986,BerezhkovskiiMakhnovskiiSuris:1989,vanWijlandCaserHilhorst:1997,Sznitman:1998,vandenBergBolthausendenHollander:2001,Spitzer:2001} from a probabilistic point of view and has
a very wide range of applications, such as
medical physics \cite[for example]{DagdugBerezhkovskiiWeiss:2002},
chemical engineering \cite[for example]{EggersdorferPratsinis:2014}
or
ecology \cite[for example]{Visser:2007}.
On the lattice, the volume of the Sausage translates to the number of
\emph{distinct} sites visited \cite{Torney:1986}. In
this work, we present an alternative, field-theoretic approach which is particularly
flexible with respect to boundary conditions and observables
with the aim to characterise and resolve the technical challenges in such an undertaking, not with the aim to improve upon the existing theory of the Wiener Sausage.
The approach has the additional appeal that, somewhat similar to percolation
\cite{StaufferAharony:1994} where all non-trivial features are due to the imposed definition
of clusters as being composed of occupied sites connected via open bonds
between nearest neighbours, the ``interaction'' in the present case is
one imposed
in retrospect. After all,
the Brownian particle studied is free and not affected by any form of
interaction. Yet, the observable requires us to discount returns, \latin{i.e.}\@\xspace
loops in the trajectory of the particle, thereby inducing an interaction
between the particle's past and present.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\linewidth]{rw_wiener_cropped_0_241.png}
\end{center}
\caption{\flabel{2d_example}Example of the Wiener Sausage problem in two
dimensions. The blue area has been traced out by the Brownian particle
attached to a disc
shown in red.}
\end{figure}
Before describing the process to be analysed in further detail, we want
to point out that some of the questions pursued in the following are
common to the field-theoretic re-formulation of stochastic processes
\cite{Doi:1976,Peliti:1985,Cardy:1999,Cardy:2008,TaeuberHowardVollmayr-Lee:2005,Taeuber:2014}. Against the
background of a field theory of the Manna Model
\cite{Manna:1991a,Dhar:1999c} one of us recently
developed, the features we wanted to understand were: 1) ``Fermionic'',
``excluded volume'' or ``hard-core interaction''
processes \cite[for example]{Hinrichsen:2000a}, \latin{i.e.}\@\xspace processes where lattice sites have a certain carrying
capacity (unity in the present case) that cannot be exceeded.
2) Systems with boundaries, \latin{i.e.}\@\xspace lack
of momentum conservation in the vertices. 2') Related to that, how different
modes couple in finite, but translationally invariant systems (periodic boundary conditions). 3) The
special characteristics of the propagator of the immobile species.
4) Observables that are spatial or spatio-temporal integrals of
densities.
The Wiener Sausage incorporates all of the above and
because it is exactly solvable or has been characterised by very
different means
\cite{KolmogoroffLeontowitsch:1933,DonskerVaradhan:1975,BerezhkovskiiMakhnovskiiSuris:1989,vandenBergBolthausendenHollander:2004}, it also gives access to a better understanding of the
renormalisation process itself.
In the following section we will describe the process we are
investigating and contrast it with the original Wiener Sausage. In
\Sref{FieldTheory} we will introduce the field-theoretic description up
to tree level, which is complemented by \Sref{beyond_tree_level}, where
we perform a one-loop renormalisation procedure. It will turn out that
there are no further corrections beyond one loop and our perturbative
results may thus be regarded as exhaustive. \Sref{semi_infty_strip} and
\Sref{infinite_cylinder} are dedicated to calculations in finite systems. \Sref{discussion} contains a discussion of the results mostly from a field-theoretic point of view, with \Sref{summary_results} however focusing on a summary of this work with regard to the original Wiener Sausage problem.
\section{Model}\slabel{model}
Originally, the Wiener Sausage is concerned with the moments or
generally statistical properties as a function of time of the
volume traced out by a sphere of fixed given (say, unit) radius, which is attached
to a Brownian particle. This volume is thus the set of point within a
certain distance to the particle's trajectory. Our field-theoretic
approach will not recover that process, but one that can reasonably be
assumed to reside in the same universality class. One may take the view
that the field-theoretic description is merely a different view on the
same phenomenon, namely the Wiener Sausage.
To motivate the field theory and link it to the original problem, we will
distinguish three different models:
(i) The original Wiener Sausage in
terms of a sphere dragged by a Brownian particle
\cite{KolmogoroffLeontowitsch:1933}, (ii) a discrete time random walker on a lattice,
where the Sausage becomes the set of distinct sites visited
\cite{Torney:1986}, (iii) a Brownian particle
in the continuum that spawns immobile offspring with a finite rate and
subject to a finite carrying capacity.
In the following, we will first describe how the phenomenon
on the lattice, (ii), relates to the original Wiener Sausage, (i), and then
how the field theory, (iii), relates to the lattice model, (ii).
The asymptote in long times $t$ of the number of distinct sites visited by
a discrete time random walker on a lattice ((ii) above) is expected to converge to
that of the volume $V(t)$ over the volume $V_0$ of the sphere in the original
process ((i) above), provided the walker returns repeatedly, so that the shape of
the sphere and the structure of the lattice respectively do not enter
into the shape and size of the volume visited.
Frequent returns are realised in the limit of long
times and below $d=2$ dimensions.
In that case, the walker
on the lattice becomes a discretised version of the original Wiener Sausage,
as the particle drags a sphere that is small compared to the volume
traced out.
Indeed, in one dimension, $d=1$, the expected
volume of the Wiener Sausage in units of the volume of the sphere is
dominated by $V(t)/V_0 \sim \sqrt{4 tD/(\pi b^2)}$, where $t$ is the time, $D$ is the
diffusion constant and $b$ the radius of the sphere, whereas the
expected number of distinct sites visited by a random walker
after $n$ steps is dominated $\sqrt{8n/\pi}$ \cite{Torney:1986}. The two
expressions are identical for $t=n$ and $D=2b^2$, the effective
diffusion constant of a random walker taking one step of distance $2b$
in one of the $2$ directions in each time step. Above $d=2$ dimensions, the walker is
free, \latin{i.e.}\@\xspace self-intersection of the trace becomes irrelevant on larger
time scales. The number of distinct sites visited and the Wiener Sausage
volume therefore both scale linearly in $t$ and $n$ respectively.
However, the (non-universal) proportionality factor, \latin{e.g.}\@\xspace
$\lim_{t\to\infty} V(t)/(V_0 t)$ for the original Wiener Sausage,
is affected by the microscopic details such as the self-intersection of
the sphere or the lattice structure of the random walker.
We proceed to relate the process on the lattice (ii) to a
Brownian particle spawning immobile offspring (iii).
To this end, we first describe (ii) in the language of reaction and
diffusion. In (ii), an ``active'' particle (species ``A'', the active
species) performs a random walk on a lattice. Simultaneously, the
particle spawns immobile offspring particles (species ``B'', the blue
ink traces of A shown in \Fref{2d_example}, below sometimes
referred to as a ``substrate particle'') at every site visited, provided
that the site is not already occupied by an immobile B particle. In
other words, A spawns exactly one B at every newly visited site, so that
the number of B particles deposited becomes a proxy for the number of
distinct sites visited. In
dimensions less than $2$ the A particle will return to every site visited
arbitrarily often in the limit of long times.
A finite spawning
probability
will therefore change the number of B particles deposited only at the
fringes of the set of sites visited, without, however, changing the
asymptotics of the number of B particles in the system as a function of
time. If $\overline{n}_B(\gpvec{x},t)$ is the number of B particles at position $\gpvec{x}$
on the lattice and time $t$, the probability with which B particles are spawned by an A
particle may be written as $\overline{\gamma}(1-\overline{n}_B(\gpvec{x},t))$, so that deposition
occurs with probability $\overline{\gamma}$ if no B particle is present and not at all
otherwise.
At this stage, we may introduce a \emph{carrying capacity} $\overline{n}_0$,
which determines the maximum number of B particles deposited on any
site, by making the spawning probability drop from $\overline{\gamma}$
to $0$ linearly in the particle number $\overline{n}_B(\gpvec{x},t)$, \latin{i.e.}\@\xspace like
\begin{equation}\elabel{def_carrying_capacity_lattice}
\overline{\gamma}\frac{\overline{n}_0-\overline{n}_B(\gpvec{x},t)}{\overline{n}_0} \ .
\end{equation}
In the process (ii) discussed so far, $\overline{n}_0$
is unity, but from what has been discussed above, $\overline{n}_0>1$ will result in
each (frequently) revisited site carrying $\overline{n}_0$ immobile B particles.
The meaning of the carrying capacity in relation to the field theory is
further discussed in \Sref{carrying_capcity}.
To see the relation between the third process, (iii) and the discrete
time, discrete space process (ii), we first introduce continuous time in
the latter. Random hopping which used to occur once in every time step
now becomes a Poisson process with a certain rate, say $H$, as does the
spawning of immobile offspring, with now takes place with rate
$\gamma(\overline{n}_0-\overline{n}_B(\gpvec{x},t))/\overline{n}_0$. In the limit of $\gamma\gg H$ all
distinct sites visited will carry $\overline{n}_0$ immobile offspring. However, in
dimensions $d<2$ sites are visited repeatedly, so that even a finite
deposition (attempt) rate $\gamma$ yields the same asymptotic
occupation. In dimensions $d>2$ the number of B particles deposited
will, on the other hand, be proportional to the rate $\gamma$.
The expression $\gamma(\overline{n}_0-\overline{n}_B(\gpvec{x},t))/\overline{n}_0$ may be written
as $\gamma - \overline{\kappa} \overline{n}_B(\gpvec{x},t)$ where $\overline{\kappa}=\gamma/\overline{n}_0$ is a
discount rate. In this interpretation (the view adopted in the
perturbation theory below), deposition takes place unhindered
with rate $\gamma$ while \emph{unlimited} (and thus supposedly
suppressed) deposition is discounted by $\overline{\kappa} \overline{n}_B(\gpvec{x},t)$, \latin{i.e.}\@\xspace
with a rate proportional to the occupation and inversely proportional to
the carrying capacity.
It remains to take the continuum (space) limit to arrive at process
(iii) to be written as a field theory, where occupation numbers
$\overline{n}_B(\gpvec{x},t)$ and $\overline{n}_A(\gpvec{x},t)$ for B and A particles
respectively turn into
occupation densities, \latin{i.e.}\@\xspace fields, namely
$n_B(\gpvec{x},t)$ and $n_A(\gpvec{x},t)$.
Moreover, the carrying capacity $\overline{n}_0$ turns into
a carrying density capacity, $n_0$, so that the discount mentioned
above is now
parameterised by $\kappa=\gamma/n_0$ (a rate per density). The
deposition thus occurs with rate
\begin{equation}
\gamma\left(1-\frac{n_B(\gpvec{x},t)}{n_0}\right) = \gamma - \kappa n_B(\gpvec{x},t) \ .
\elabel{def_carrying_capacity}
\end{equation}
The random movement of the A particle is now parameterised by the
diffusion constant $D$, which may be obtained as the hopping rate $H$ over
the squared lattice spacing in the limit of the latter going to $0$.
\subsection{Intermediate summary}
The long-winded discussion above serves as a justification as to why we expect the
field theory of (iii) to produce a phenomenon in the same universality class as
the original Wiener Sausage. Starting from the original Wiener Sausage
(i), we have motivated why the process on the lattice, (ii), can be
regarded as a discretised version of (i) and introduced (iii) as its
continuum approximation.
In the course of the justification, we
made use of some of the details of the processes involved, such as
repeated returns to sites in process (ii). The field theoretic
description of the universality class of the Wiener Sausage, however,
may be derived without
recourse to these details, simply by observing that the volume traced
out by the sausage is proportional to the length of the trajectory with
multiple visits discounted, corresponding to the number of immobile
B particles deposited by a Brownian particle (of species A), if its spawning rate is
moderated down in the presence of B particles.
To summarise, process (iii), to be cast in a Liouvillian and thus a
field theory below,
is defined as follows: The Brownian particle A
freely diffuses with diffusion constant $D$ and possibly subject to
boundary conditions. While diffusing, the particle
can spawn offspring with Poissonian rate $\gamma$ which, however, belong
to an immobile second species B. If $n_B(\gpvec{x},t)$ is the density of
these particles, the deposition is linearly regulated down in their
presence, according to $\gamma-\kappa n_B(\gpvec{x},t)$ with
$\kappa=\gamma/n_0$. Here, $\gamma$ is the deposition rate and $n_0$ is the
carrying (density) capacity.
It is convenient in the field theory to allow for spontaneous extinction
with rate (or ``mass'') $r$.
Ignoring boundary conditions, the propagator of the Brownian particle
(species A, the ``activity'') takes the familiar form $1/(-\imath \omega + D \gpvec{k}^2
+ r)$ where $\omega$ and $\gpvec{k}$ parameterise frequency and momentum
(wave number) coordinates, respectively. The propagator of the immobile species takes the
form $1/(-\imath \omega + \epsilon')$, where $\epsilon'$ is the rate of
spontaneous extinction of $B$ therticles and the limit $\epsilon'\to0^+$
is implied to establish causality, as often done in field theories.
The key observable, corresponding to the volume of the sausage,
is the total number of immobile particles in the
system after a given time $t$, \latin{i.e.}\@\xspace the spatial integral over
$n_B(\gpvec{x},t)$.
The engineering dimension of $\kappa$ is a rate per density, which in
comparison to the engineering dimension of the diffusion constant, an
area per time, reveals the upper critical dimension of $2$.
Alternatively, this can be seen from the density of unhindered
deposition as a function of time, $(\gamma t)/(Dt)^{d/2}$, \latin{i.e.}\@\xspace in the
absence of discounts, $\kappa=0$ or $n_0\to\infty$.
In what follows, we will characterise process (iii) field theoretically.
The Liouvillian of the process is split into a linear part,
\Eref{harmonic_part}, discussed in \Sref{harmonic_part}, and
a non-linear part \Eref{nonlinear_part}, discussed in
\Sref{non-linearity}. The linear part of the Liouvillian can be
constructed from the propagators mentioned above and vice versa.
The Liouvillian will enter into a path-integral,
which can be used to generate all correlation and vertex functions. The
path integral itself is to be evaluated perturbatively in the
non-linearity, which, for example, instantly indicates that the
non-linearity has no bearing on the propagator of particle A.
Above two dimensions, the non-linearity causes an ultraviolet divergence, \Eref{one_loop},
which has its origin
in the increasingly sharp divergence in $t$ of the density of a random
walker at the origin.\footnote{The integral \eref{one_loop} is in fact
the time-integrated density of a random walker at the origin, subject to
extinction $r+\epsilon'$, namely $\int_0^{\infty}\dint{t}
\exp{-(r+\epsilon')t}(4\pi Dt)^{-d/2}=\Gamma(1-d/2) (4\pi
D)^{-d/2}(r+\epsilon')^{-(1-d/2)}$.} However, in dimensions above $2$
the non-linearity is infrared irrelevant and so all long-time,
long-range observables are covered by the tree-level. In dimension below
$2$ no ultraviolet divergence occurs and the infrared can be regularised
using finite masses $r$ and $\epsilon'$. We will therefore work in
dimensions $2-\epsilon$, with $\epsilon>0$, known as dimensional
regularisation (of the ultraviolet).
Initially, the density fields will be studied on an infinite domain
without boundaries. However, in \Sref{semi_infty_strip} we will also
consider an infinite slab and in \Sref{infinite_cylinder} an infinite
cylinder. We will use Fourier transforms to write the fields in the
infinite domain and suitably chosen Fourier series
in the presence of open (Dirichlet) or cylindrical boundaries. These
transforms and series are discussed in
\Sref{Fourier} and used later in the respective sections.
We will demonstrate in the following that the field theory recovers
exact results of the original Wiener Sausage as far as universal
exponents are concerned, but also with respect to some amplitudes
(namely the leading order term of the volume of the sausage in one
dimension as a function of time and the leading order of the volume as a
function of the system size of the infinite slab). Firstly, the present results
confirm that the logistic term \Eref{def_carrying_capacity} is capable
of capturing the constraints due to the carrying capacity. At a more
technical level, the
calculations for (partially) finite systems (infinite slab and cylinder)
involve different propagators, which under renormalisation can lead to new
non-linearities. Similar to the classic case discussed in
\cite{LubenskyRubin:1975} this problem, however, will be avoided. The
results for these more complicated boundary conditions show very
interesting crossover behaviour. Finally, from a physical point of
view, it is particularly interesting that the infrared regularisation of
the immobile species, $\epsilon'$, a neccessary ingredient as to preserve causality in
the absence of diffusion, can in principle be used to regularise the
theory as a whole, \latin{i.e.}\@\xspace without the need of a particle mass $r$.
Before introducing the field theory of the present model in
\Sref{FieldTheory}, we discuss in the following
briefly the intricacies of the fermionic nature of the B particles.
\subsection{Finite carrying capacity}\slabel{carrying_capcity}
To fully understand the effect and consequences of the carrying
capacity, it is best to reconsider the process on the lattice. A
carrying capacity of $\overline{n}_0=1$
in \Eref{def_carrying_capacity_lattice}
switches off the deposition of B particles
in their presence in a rather dramatic fashion, implementing a constraint that is
normally referred to as \emph{fermionic}, because there is never more
than one B particle deposited on a site. Raising $\overline{n}_0$ allows the spawning
rate to drop linearly in the occupation in an otherwise \emph{bosonic} setup.
While this may raise suspicion and invite the criticism of a fudge,
as
demonstrated below, such a bosonic regularisation may be interpreted as
the fermionic case on a lattice with a particular connectivity, \latin{i.e.}\@\xspace the
attempted regularisation is the original, fermionic case in disguise,
suggesting that no such regularisation is needed.
Some authors \cite[and references therein]{vanWijland:2001} avoid terms like
\Eref{def_carrying_capacity_lattice} or \Eref{def_carrying_capacity} by
expanding a suitable expression for $\delta_{1,\overline{n}_B(\gpvec{x},t)}$, a
Kronecker $\delta$-function. \Eref{def_carrying_capacity_lattice} and
\Eref{def_carrying_capacity} are \emph{not} leading order
terms in an expansion. For $\overline{n}_0=1$ and before taking any other approximation (\latin{e.g.}\@\xspace
continuous space and density or removing irrelevant terms in the field theory) a logistic term
like \eref{def_carrying_capacity_lattice} is a representation of the original process as exact as
one involving the Kronecker $\delta$-function. For $\overline{n}_0>1$ a logistic term gives rise
to a model that may be strictly different compared to one with a sharp carrying capacity
implemented by, say, a Heaviside step-function, $\theta(\overline{n}_0-\overline{n}_B(\gpvec{x},t))$, but nonetheless
one that may be of equal interest.
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.40]
\foreach \r in {1,3,5,7}
{
\node[circle,fill=white,minimum size=10pt] (0-\r) at (0,\r) {};
\node[circle,fill=white,minimum size=10pt] (21-\r) at (21,\r) {};
\foreach \c in {3,6,9,12,15,18}
{
\node[draw,circle,minimum size=10pt] (\c-\r) at (\c,\r) {};
}
}
\foreach \t in {1,3,5,7}{
\foreach \r in {1,3,5,7}{
\foreach \c in {0,3,6,9,12,15,18}{
\myp=\c
\advance\myp by 3
\path (\c-\r) edge (\the\myp-\t);
}
}
}
\pgfdeclarefading{faderight}
{\tikz\fill[left color=transparent!100,right
color=transparent!0] (0,0) rectangle (40bp,40bp); }
\pgfdeclarefading{fadeleft}
{\tikz\fill[left color=transparent!0,right
color=transparent!100] (0,0) rectangle (40bp,40bp); }
\fill[path fading=faderight,fill=white] (19.5,0) rectangle (21,8);
\fill[path fading=fadeleft,fill=white] (1.5,0) rectangle (0,8);
\draw[<->,thick](0,0) -- (21,0);
\node (L) at (10.5,-0.6) {\large $L$};
\draw[<->,thick](21,0.8) -- (21,7.3);
\node (n) at (22.5,4) {\large $\overline{n}_0=4$};
\end{tikzpicture}
\end{center}
\caption{\flabel{funny_lattice}A one dimensional lattice of size $L$ and carrying capacity
$\overline{n}_0=4$ corresponds to the lattice shown above, where the carrying capacity
of the former is implemented by expanding each site into a column of
$\overline{n}_0$
sites. The Brownian particle can jump from every site to all sites in the
neighbouring columns. In the new lattice, the carrying capacity \emph{per
site}
is unity, the carrying capacity \emph{per column} is $\overline{n}_0$.}
\end{figure}
Large $\overline{n}_0$ on the other hand, softens the cutoff, because
spawning does not drop from suddenly from $\overline{\gamma}$ to $0$ but is
more and more suppressed. One might
therefore be inclined to study the problem in the limit of large $\overline{n}_0$.
At closer inspection, however, it turns out that such increased $\overline{n}_0$
does not present a qualitative change of the problem: Having $\overline{n}_0>1$
is \emph{as
if} each site was divided into $\overline{n}_0$ spaces. When the Brownian particle
jumps from site to site it arrives in one of those $\overline{n}_0$ spaces, only
$\overline{n}_0-\overline{n}_B(\gpvec{n},t)$ of which are empty, so that an offspring can be left behind.
The process with carrying capacity $\overline{n}_0>1$ therefore corresponds to the
process with a carrying capacity of unity per space on a lattice where $\overline{n}_B(\gpvec{n},t)$ describes the number
of immobile offspring in each ``nest'' or column of such spaces, as illustrated in
\Fref{funny_lattice}. In effect, the carrying capacity $\overline{n}_0>1$ is
implemented \emph{per column}, leaving the original fermionic constraint
of at most one offspring \emph{per space (or site)} in place.
In other words, even when a carrying
capacity $\overline{n}_0\gg1$ is introduced to smoothen the fermionic constraint, it is still
nothing else but the original constraint $\overline{n}_0=1$ on a different lattice.
This led us to believe that there is no qualitative difference in
$\overline{n}_0=1$ or any other finite value of $\overline{n}_0$. In the following,
the field theory will retain the carrying capacity
$n_0$ because it is an interesting parameter ($n_0\to\infty$ switches
the interaction off) and a ``marker'' of the
interaction. It may be set to any positive value.
\section{Field theory}
\slabel{FieldTheory}
In order to cast the model introduced above in a field-theoretic
language, we take the Doi-Peliti \cite{Doi:1976,Peliti:1985} approach without going
through too many technical details. There are a number
of reviews and extremely useful tutorials available
\cite{Cardy:1999,Cardy:2008}.
In the following the mobile particle is of species ``A'', performing
Brownian motion with (nearest neighbour) hopping rate $H$, which translates to diffusion constant
$D=H/(2d)$ on a $d$-dimensional hypercubic lattice. We expect universal scaling in the large time and space limit.
To regularise the infrared,
we also introduce an extinction rate $r$. A's creation operator is
$\creat{a}{\gpvec{x}}$, its annihilation operator is $\annih{a}{\gpvec{x}}$.
The immobile species is ``B'', spawned with rate $\gamma$ by species A.
Its creation operator is
$\creat{b}{\gpvec{x}}$, its annihilation operator is $\annih{b}{\gpvec{x}}$,
both commuting with the creation and annihilation operators of species A.
The immobile species goes extinct with rate $\epsilon'$, which allows us
to have a Fourier transform and to restore causality (possible annihilation, \latin{i.e.}\@\xspace existence, only after
creation) even without spontaneous extinction, once we take the limit $\epsilon'\to0$.
\subsection{Fourier Transform}\slabel{Fourier}
After replacing the operators by real fields, the Gaussian (harmonic)
part of the resulting path integral can be performed, once the fields
have been Fourier transformed. We will use the sign and notational
convention of
\begin{multline}\elabel{fourier_transform}
\phi(\gpvec{x},t) \\
=
(2\pi)^{-1} \int_{-\infty}^{\infty} \dint{\omega}
(2\pi)^{-1} \int_{-\infty}^{\infty} \dint{k_1}\ldots
(2\pi)^{-1} \int_{-\infty}^{\infty} \dint{k_d}
\ \phi(\gpvec{k},\omega) \exp{-\imath\omega t + \imath\gpvec{k}\gpvec{x}}\\
=
\int \dintbar{\omega}\ddintbar{k}
\ \phi(\gpvec{k},\omega) \exp{-\imath\omega t + \imath\gpvec{k}\gpvec{x}} \ .
\end{multline}
The field $\phi(\gpvec{k},\omega)$ corresponds to the annihilator
$\annih{a}{\gpvec{x}}$ of the active particles, the field $\tilde{\phi}(\gpvec{k},\omega)$ to the
Doi-shifted creator $\creatDoi{a}{\gpvec{x}}=\creat{a}{\gpvec{x}}-1$.
Correspondingly, $\psi(\gpvec{k},\omega)$ and $\tilde{\psi}(\gpvec{k},\omega)$
replace $\annih{b}{\gpvec{x}}$ and $\creatDoi{b}{\gpvec{x}}=\creat{b}{\gpvec{x}}-1$,
respectively.
It is instructive to consider a second set of orthogonal
functions at this stage.
Placing the process in a space that has a \emph{finite extension} along
one axis means that
boundary conditions have to be met, which is more conveniently done in
one eigensystem rather than another. Below, we will consider
an infinite slab with finite thickness $L$, \latin{i.e.}\@\xspace
$d$-dimensional spaces which are infinitely
extended (using the orthogonal functions and transforms introduced
above) in $\tilde{d}=d-1$ dimensions, while along one axis, the boundaries are open, \latin{i.e.}\@\xspace the particle
density of species A vanishes at the (two parallel, $\tilde{d}$-dimensional) boundaries and outside. This Dirichlet boundary condition
is best met using eigenfunctions $\sqrt{2/L} \sin(q_n z)$ with $q_n=\pi
n/L$
and $n=1,2,\ldots$, making it complete and orthonormal because
\begin{equation}\elabel{efuncs_sin}
\frac{2}{L} \int_0^L\dint{z}\, \sin(q_n z) \sin(q_m z) = \delta_{n,m} \ .
\end{equation}
In passing, we have introduced
the finite linear length of the space, $L$.
Purely for ease of notation and in order to keep expressions in finite
systems dimensionally as similar as possible to those in infinite ones,
\Eref{fourier_transform}, we will transform as follows:
\begin{subeqnarray}{\elabel{convenient_choices}}
\phi(z)&=&\frac{2}{L}\sum_{n=1}^\infty \sin(q_n z) \phi_n\\
\phi_n&=&\int_0^L \dint{z} \sin(q_n z) \phi(z)
\end{subeqnarray}
using
\begin{subeqnarray}{}
\frac{2}{L}\sum_{n=1}^\infty \sin(q_n y) \sin(q_n z) & = & \delta(z-y)\\
\int_0^L \dint{z} \sin(q_n z) \sin(q_m z) & = & \frac{L}{2} \delta_{m,n} \ ,
\elabel{two_terms_Dirichlet}
\end{subeqnarray}
where $\delta(z-y)$ is the usual Dirac $\delta$ function for
$z-y\in(0,L)$ but to be replaced by the periodic Dirac comb
$\sum_{m=-\infty}^\infty \delta(z-y+mL)$ for arbitrary $z-y$. For ease
of notation, we have omitted the time dependence of $\phi(\gpvec{x},t)$ as
well as $\tilde{d}$ components other than $z$. The other fields, $\tilde{\phi}$,
as well as $\psi$ and $\tilde{\psi}$ transform correspondingly. The
spatial transform of the latter is subject to some convenient choice,
because the immobile species is not constrained by a boundary condition.
It will turn out that, as
expected in finite size scaling, the lowest mode $q_1=\pi/L$ plays the
r{\^o}le of a temperature like variable, controlling the distance to the
critical point.
We will also briefly study systems which are infinitely extended in
$\tilde{d}$ dimensions and periodically closed in one.
In the periodic dimension, the spectrum of conveniently chosen
eigenfunctions $\sqrt{1/L} \exp{\imath k_n y}$ is discrete with
$k_n=2\pi n/L$ and $n\in\gpset{Z}$,
\begin{equation}\elabel{efuncs_exp}
\frac{1}{L} \int_0^L\dint{y}\, \exp{\imath k_n y} \exp{\imath k_m y} = \delta_{n+m,0} \ .
\end{equation}
Again, we transform slightly asymmetrically (in $L$),
\begin{subeqnarray}{\elabel{convenient_choices_periodic}}
\phi(z)&=&\frac{1}{L}\sum_{n=-\infty}^\infty \exp{\imath k_n z} \phi_n\\
\phi_n&=&\int_0^L \dint{z} \exp{-\imath k_n z} \phi(z)
\end{subeqnarray}
with
\begin{subeqnarray}{}
\frac{1}{L}\sum_{n=-\infty}^\infty \exp{\imath k_n y} \exp{-\imath k_n z} & = & \delta(z-y)
\elabel{exp_sum}\\
\int_0^L \dint{z} \exp{\imath k_n z} \exp{-\imath k_m z} & = & L \delta_{m,n} \ ,
\end{subeqnarray}
where again $\delta(z-y)$ is to be replaced by a Dirac comb if considered for $z-y\notin
(0,L)$. Again, time and $\tilde{d}$ spatial coordinates were omitted.
Similar transforms apply to the other fields.
There is a crucial difference between eigenfunctions $\exp{\imath k_ny}$ and
$\sin(q_nz)$, as the former conserves momenta in vertices, whereas the
latter does not:
\begin{subequations}
\elabel{delta_versus_epsilon_nml_all}
\begin{equation}
\int_0^L \dint{y}
\exp{\imath k_n y} \exp{\imath k_m y} \exp{\imath k_{\ell} y} =
L \delta_{n+m+\ell,0} \elabel{three_terms_periodic}
\end{equation}
while
\begin{equation}
\int_0^L \dint{y}
\sin(q_n y) \sin(q_m y) \sin(q_{\ell} y) = L \epsilon_{nm\ell} \ ,
\elabel{three_terms_Dirichlet}
\end{equation}
where
\begin{equation}
\epsilon_{nm\ell} \elabel{def_epsilon_nml}
=
\left\{
\begin{array}{ll}
\frac{1}{2\pi} \left(
\frac{1}{n+m-\ell}
+\frac{1}{n-m+\ell}
+\frac{1}{-n+m+\ell}
-\frac{1}{n+m+\ell}
\right)
& \text{for $\ell+m+n$ odd}\\
\quad & \quad \\
0 & \text{otherwise}
\end{array}
\right.
\end{equation}
\end{subequations}
with $q_n=\pi n/L>0$, $n\in\gpset{N}^+$ and $k_n=2\pi n/L$, $n\in\gpset{Z}$ (sign unconstrained) as
introduced above.
Having made convenient choices such as \Eref{convenient_choices}, we
will carry on using the Fourier transforms of the bulk
\Eref{fourier_transform}, which is
easily re-written for Dirichlet boundary conditions using
\Eref{convenient_choices},
simply by replacing each integral over $\dintbar{k}$ by $(2/L)\sum_n$
and similar for periodic boundary conditions,
\Eref{convenient_choices_periodic}. Only the non-linearity,
\Sref{non-linearity}, is expected to require further careful analysis as
$\epsilon_{nm\ell}$ of \Eref{three_terms_Dirichlet} is structurally far more demanding than
$\delta_{n+m+\ell,0}$ of \Eref{three_terms_periodic}.
\subsection{Harmonic Part}\slabel{harmonic_part}
Following the normal procedure \cite[for example]{Cardy:1998}, the harmonic
part $\mathcal{L}_0$ of the Liouvillian $\mathcal{L}=\mathcal{L}_0+\mathcal{L}_1$ reads
\begin{equation}\elabel{harmonic_part}
\mathcal{L}_0 =
\tilde{\phi} \partial_t \phi
+ D \nabla \tilde{\phi} \nabla \phi
+ r \tilde{\phi} \phi
+ \tilde{\psi} \partial_t \psi
+ \epsilon' \tilde{\psi} \psi \ .
\end{equation}
The non-linear part $\mathcal{L}_1$, \Eref{nonlinear_part}, is
discussed in \Sref{non-linearity}. The harmonic part, $\mathcal{L}_0$,
describes the diffusive evolution of the density field of A particles,
represented by $\phi$ and $\tilde{\phi}$,
which diffuse with diffusion constant $D$ and get
spontaneously extinct with rate $r$, as well as the evolution of immobile
particles B, represented by densities $\psi$ and $\tilde{\psi}$, which do
not diffuse but get extinct with rate $\epsilon'$.
After Fourier transforming and without further ado the harmonic part of the path integral
\[
\int
\mathcal{D}\phi
\mathcal{D}\tilde{\phi}
\mathcal{D}\psi
\mathcal{D}\tilde{\psi}
\exp{-\int \ddintbar{\!k}\,\dintbar{\omega}\ \mathcal{L}}
\]
can be performed, producing the two bare propagators
\begin{subequations}\elabel{bare_propagator}
\begin{align}\elabel{bare_activity_propagator}
\ave{\phi(\gpvec{k},\omega)\tilde{\phi}(\gpvec{k}',\omega')}_0
&=
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')\delta\mkern-6mu\mathchar'26(\omega+\omega')}{-\imath\omega+D\gpvec{k}^2+r}
\ \hat{=}\ \tikz[baseline=-2.5pt]{\draw[activity] (0.6,0) -- (-0.6,0);}\\
\ave{\psi(\gpvec{k},\omega)\tilde{\psi}(\gpvec{k}',\omega')}_0
&=
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')\delta\mkern-6mu\mathchar'26(\omega+\omega')}{-\imath\omega+\epsilon'}
\ \hat{=}\ \tikz[baseline=-2.5pt]{\draw[substrate] (0.6,0) -- (-0.6,0);}
\elabel{substrate_propagator}
\end{align}
\end{subequations}
where $\delta\mkern-6mu\mathchar'26(\omega+\omega')=\delta(\omega+\omega')/(2\pi)$ and
$\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')=\delta(\gpvec{k}+\gpvec{k}')/(2\pi)^d$.
Below, we will refer to the propagator
of the diffusive particles as the ``activity propagator'' and to the one
for the immobile species as the ``substrate propagator'' (or
``activity'' and ``substrate legs'', respectively).
As the propagation of the active particles is unaffected by the
deposition of immobile particles, the activity propagator does not
renormalise
$\ave{\phi\tilde{\phi}}=\ave{\phi\tilde{\phi}}_0$. The same is true for the
immobile species, which might be spawned by active particles, however,
once deposited remains inert, $\ave{\psi\tilde{\psi}}=\ave{\psi\tilde{\psi}}_0$.
The Fourier
transform \Eref{fourier_transform} of the latter produces
$\delta(\gpvec{x}-\gpvec{x}')\theta(t-t')$ in the limit $\epsilon'\to0$, with $\theta(x)$ denoting the
Heaviside $\theta$-function as one would expect (with $\gpvec{x},t$
being the position and time of ``probing'' and $\gpvec{x}',t'$ position and time of
creation).
At this stage, there is no interaction and no transmutation,
$\ave{\tilde{\psi}(\gpvec{k},\omega)\phi(\gpvec{k}',\omega')}=0$. Diffusing
particles A happily co-exist with immobile ones.
\subsection{Non-Linearity}\slabel{non-linearity}
The harmonic part of the Liouvillian, $\mathcal{L}_0$, discussed in the preceding section
covers the diffusive motion and spontaneous extinction of A particles (fields $\phi$ and
$\tilde{\phi}$) and the spontaneous extinction of the resting B particles
(fields $\psi$ and
$\tilde{\psi}$). In the following, we will discuss the non-linear
(interacting) part of the Liouvillian, $\mathcal{L}_1$, which
introduces the spawning of B particles by the A particle, subject to the
constraint of the finite carrying capacity, which establishes an effective
interaction between
previously deposited
particles and any new particle to be deposited.
As discussed in \Sref{carrying_capcity}, spawning is moderated down in the
presence of B particles to $\gamma(1-n_B(\gpvec{x},t)/n_0)$. At the level of
a master equation, this conditional deposition gives a non-linear contribution of
\begin{multline}\elabel{master_eqn}
\partial_t \mathcal{P}(\ldots,n_A,n_B,\ldots) = \text{harmonic terms} +\\
\sum_{\gpvec{n}} n_A \gamma\left(1-\frac{n_B-1}{n_0}\right)
\mathcal{P}(\ldots,n_A,n_B-1,\ldots)\\
- n_A \gamma\left(1-\frac{n_B}{n_0}\right) \mathcal{P}(\ldots,n_A,n_B,\ldots) \ ,
\end{multline}
where, for convenience, the problem is considered for individual lattice
sites $\gpvec{n}$ which contain $n_A=n_A(\gpvec{n})$ particles of species A and
$n_B$ particles of species B. The contributions by harmonic terms,
namely diffusion of A particles and spontaneous extinction of both, as
discussed in the previous section, have been omitted. The first term in the sum describes the
creation of a B particle in the presence of $n_B-1$ of those to make up $n_B$ in total, the second term makes the
B particle number exceed $n_B$, $n_B\to n_B+1$. If
\begin{equation}
\ket{\Psi}(t)=\sum_{\{n_A,n_B\}}
\mathcal{P}(\ldots,n_A,n_B,\ldots)
\prod_{\gpvec{n}} \creat[n_A]{a}{\gpvec{n}}
\prod_{\gpvec{n}} \creat[n_B]{b}{\gpvec{n}}
\ket{0} \ ,
\end{equation}
where the sum runs over all states of the entire lattice, then the
conditional deposition produces the contribution
\begin{multline}\elabel{non-linearities}
\partial_t \ket{\Psi}(t)=\text{bilinear terms} + \\
\sum_{\gpvec{n}}
\gamma\ \creatDoi{b}{\gpvec{n}}\,\creat{a}{\gpvec{n}}\annih{a}{\gpvec{n}}
-\frac{\gamma}{n_0}\ \creatDoi{b}{\gpvec{n}}\creat{b}{\gpvec{n}}\annih{b}{\gpvec{n}}\,\creat{a}{\gpvec{n}}\annih{a}{\gpvec{n}}
\ ,
\end{multline}
where we have used the commutator,
$(\creat{b}{}\annih{b}{}-1)\creat{b}{}=\creat[2]{b}{}\annih{b}{}$ and
the Doi-shifted creation operator, $\creat{b}{}=\creatDoi{b}{}+1$, as
well as the particle number operator $\creat{b}{}\annih{b}{}$.
Although using Doi-shifted operators throughout gives rise to a rather
confusing six non-linear
vertices, the resulting field theory does not turn out as messy as one
may expect. However, we need to allow for different renormalisation,
therefore introducing six different couplings below.
Replacing $\creat{a}{}$ by $1+\creatDoi{a}{}$ in
the first term of the sum generates the bilinearty
$\annih{a}{}\creatDoi{b}{}$, which we will parameterise in the following
by $\tau$, corresponding to a transmutation of
an active particle to an immobile one. Transmutation is obviously spurious; it does not actually take place but will allow us in the Doi-shifted setup (and thus with the corresponding left vacuum \cite{Cardy:1999,Cardy:2008}) to probe for substrate particles (using $\annih{b}{}$) after creating an active one (using $\creat{a}{}$) without having to probe for the latter (using $\annih{a}{}$). There is no advantage in moving
that to the bilinear part $\mathcal{L}_0$, because the determinant of
the bilinear matrix $M$ in
\begin{equation}
\mathcal{L}'_0=
\left(
\begin{array}{l}
\tilde{\phi}\\
\tilde{\psi}
\end{array}
\right)^T
\underbrace{
\left(
\begin{array}{cc}
-\imath \omega + D \gpvec{k}^2 + r & 0 \\
-\tau & -\imath \omega + \epsilon'
\end{array}
\right)}_{M}
\left(
\begin{array}{l}
\phi\\
\psi
\end{array}
\right)
\end{equation}
is unaffected by $\tau\ne0$ and therefore none of the propagators
mentioned above change. One may therefore treat \emph{all} terms
(including the bilinear transmutation) resulting
from the interaction
perturbatively, with transmutation
\begin{multline}\elabel{deri_transmutation}
\ave{\psi(\gpvec{k},\omega)\tilde{\phi}(\gpvec{k}',\omega')}_0\\
=\int \ddintbar{\!k''}\, \dintbar{\omega''}
\ \ave{\psi(\gpvec{k},\omega)\tilde{\psi}(\gpvec{k}'',\omega'')}
\tau
\ave{\phi(\gpvec{k}'',\omega'')\tilde{\phi}(\gpvec{k}',\omega')}\\
=\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')\delta\mkern-6mu\mathchar'26(\omega+\omega')}{-\imath\omega+\epsilon'}
\ \tau \ \frac{1}{-\imath\omega+D\gpvec{k}^2+r}
\ \hat{=}\ \tikz[baseline=-2.5pt]{
\draw[activity] (0.6,0) -- (0,0) node[at end,above] {$\tau$};
\draw[substrate] (0,0) -- (-0.6,0);
}
\end{multline}
that is present regardless of the carrying capacity $n_0$. At this stage
it is worth noting the sign of $\tau$ (and $\sigma$ below) as positive,
\latin{i.e.}\@\xspace the perturbative expansion will generate terms with pre-factors
$\tau$ (and $\sigma$ below).
The only
other non-linearity independent from the carrying capacity $n_0$
is the vertex
$\creatDoi{b}{}\creatDoi{a}{}\annih{a}{}$ (or $\tilde{\psi}\tilde{\phi}\phi$) in the following parameterised
by the coupling constant $\sigma$. Diagrammatically, it may be written
as the (amputated vertex)
\begin{equation}\elabel{sigma_diagram}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0,0) -- (-0.4,0) node [at start,above] {$\sigma$};
\draw[Aactivity] (0.4,0) -- (0,0);
\draw[substrate] (-130:0.4) -- (0,0);
} \ ,
\end{equation}
and can be thought of as spawning, rather than transmutation parameterised
by $\tau$.
According to \Eref{non-linearities}, there are
four non-linearities with bare-level couplings of $\gamma/n_0$,
generated by replacing the regular creation operators by their
Doi-shifted counterparts,
$\creat{a}{\gpvec{n}}=1+\creatDoi{a}{\gpvec{n}}$ and
$\creat{b}{\gpvec{n}}=1+\creatDoi{b}{\gpvec{n}}$,
in
$\frac{\gamma}{n_0}\ \creatDoi{b}{\gpvec{n}}\creat{b}{\gpvec{n}}\annih{b}{\gpvec{n}}\,\creat{a}{\gpvec{n}}\annih{a}{\gpvec{n}}$.
Each spawns at least one substrate particle, but more importantly, it
also annihilates at least one substrate particle as it ``probes for'' its
presence. The two simplest and most important (amputated) vertices are the ones introduced above with a
``wriggly tail added'',
\begin{equation}\elabel{mu_kappa}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0.5,0) -- (0,0) node[at end,above] {$\lambda$};
\draw[substrate] (0,0) -- (-0.5,0);
\draw[substrate] (-50:0.5) -- (0,0);
}
\qquad
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0.5,0) -- (0,0) node[at end,above] {$\kappa$};
\draw[Aactivity] (0,0) -- (-0.5,0) node[at end,above] {};
\draw[substrate] (-130:0.5) -- (0,0);
\draw[substrate] (-50:0.5) -- (0,0);
}
\end{equation}
where we have also indicated their coupling. By mere inspection, it is
clear that those two vertices can be strung together, renormalising the
left one. In fact, $\kappa$ is the one and only coupling that
renormalises all non-linearities ($\sigma$,$\lambda$,$\kappa$,$\chi$ and
$\xi$), including itself.
Two more vertices are generated,
\begin{equation}\elabel{rho_nu}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0.5,0) -- (0,0) node[at end,above] {$\ \ \chi$};
\draw[substrate] (0,0) -- (-0.5,0);
\draw[substrate] (-50:0.5) -- (0,0);
\draw[substrate] (130:0.5) -- (0,0);
}
\qquad
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0.5,0) -- (0,0) node[at end,above] {$\ \ \xi$};
\draw[Aactivity] (0,0) -- (-0.5,0) node[at end,above] {};
\draw[substrate] (-130:0.5) -- (0,0);
\draw[substrate] (-50:0.5) -- (0,0);
\draw[substrate] (130:0.5) -- (0,0);
} \ ,
\end{equation}
which become important only for higher order correlation functions of
the substrate particles, because there is no vertex annihilating more
than one of them --- correlations between substrate particles are present but not relevant
for the dynamics. Notably, there is no vertex that has more incoming
than outgoing
substrate legs. Finally, we note that the sign with which $\lambda$,
$\kappa$, $\chi$ and $\xi$ are generated in the perturbative expansion
is negative.
For completeness, we state the interaction part of the Liouvillian (see
\Eref{harmonic_part})
\begin{equation}\elabel{nonlinear_part}
\mathcal{L}_1 =
- \tau \tilde{\psi} \phi
- \sigma \tilde{\psi} \tilde{\phi} \phi
+ \lambda \tilde{\psi} \psi \phi
+ \kappa \tilde{\phi} \tilde{\psi} \psi \phi
+ \chi \tilde{\psi}^2 \psi \phi
+ \xi \tilde{\phi} \tilde{\psi}^2 \psi \phi \ ,
\end{equation}
with
\begin{equation}
\tau=\sigma=\gamma
\qquad \text{and} \qquad
\lambda=\kappa=\chi=\xi=\gamma/n_0
\elabel{bare_level_couplings}
\end{equation}
at bare level.
\newcommand{\dimensionof}[1]{\left[#1\right]}
\subsection{Dimensional analysis}
Determining the engineering dimensions of the coupling introduced above
is part of the ``usual drill'' and will allow us to determine the upper
critical dimension and to remove irrelevant couplings. Without dwelling
on details, analysis of the harmonic part, \Eref{harmonic_part},
reveals that $\dimensionof{D}=\mathtt{L}^2/\mathtt{T}$ (as expected for a diffusion
constant) and
$\dimensionof{r}=\dimensionof{\epsilon'}=1/\mathtt{T}$ (as expected for all extinction rates), with
$\dimensionof{x}=\mathtt{L}$, a length,
and $\dimensionof{t}=\mathtt{T}$, a time. In real time and real space,
$\dimensionof{\tilde{\phi}\phi}=\dimensionof{\tilde{\psi}\psi}=\mathtt{L}^{-d}$.
Performing the Doi-shift in \Eref{non-linearities} first and introducing
couplings for the non-linearities as outlined above allows for two
further independent dimensions, say spawning
$\dimensionof{\sigma}=\mathtt{A}$ and transmutation $\dimensionof{\tau}=\mathtt{B}$ (both originally
equal to the rate $\gamma$),
which implies
$\dimensionof{\lambda}=\mathtt{A}^{-1}\mathtt{B} \mathtt{L}^d\mathtt{T}^{-1}$,
$\dimensionof{\kappa}=\mathtt{L}^d\mathtt{T}^{-1}$,
$\dimensionof{\chi}=\mathtt{L}^d\mathtt{B}$,
$\dimensionof{\xi}=\mathtt{L}^d\mathtt{A}$,
as well as
$\dimensionof{\psi}=\mathtt{T}\mathtt{A}\mathtt{L}^{-d}$, $\dimensionof{\tilde{\psi}}=\mathtt{A}^{-1}\mathtt{T}^{-1}$,
$\dimensionof{\phi}=\mathtt{A} \mathtt{B}^{-1}\mathtt{L}^{-d}$, $\dimensionof{\tilde{\phi}}=\mathtt{A}^{-1}\mathtt{B}$ in
real space and time.
As far as the field theory is concerned, the only constraint is to
retain the diffusion constant on large scales, which implies $\mathtt{T}=\mathtt{L}^2$.
As a result, the non-linear coupling $\kappa$ (originally
$\gamma/n_0$) becomes irrelevant in dimensions $d>d_c$, as
expected with upper critical dimension $d_c=2$.
The two independent engineering dimensions $\mathtt{A}$ and $\mathtt{B}$ will be used in
the analysis below in order to maintain the existence of the associated
processes of transmutation and spawning, which are expected to govern
the tree level. If we were to argue that they become irrelevant above
a certain upper critical dimension, the density of offspring and its
correlations would
necessarily vanish everywhere.\footnote{Strictly, as we will demonstrate
below, $n$-point correlation
functions can be constructed with $\tau$ only, say
$\tikz[baseline=-2.5pt]{
\draw[tSactivity] (0.6,0) -- (0,0) node[at end,above] {$\tau$};
\draw[tsubstrate] (0,0) -- (-0.6,0);
\draw[tsubstrate] (-0.8,0) -- (-1.4,0);
\crossblob{-0.7,0};
}$
in
\Eref{second_order_diagrams}. However, it is clear that the density of
the active walker and its immobile offspring will remain correlated,
which is mediated by $\sigma$, \Eref{sigma_diagram}.}
Even though we may want to exploit the ambiguity in the engineering dimensions
\cite{LeeCardy:1995,TaeuberHowardVollmayr-Lee:2005} in the scaling
analysis (however, consistent with the results above),
in the following section we will make explicit use of the
Doi-shift when deriving observables, which means that both $\tilde{\phi}$ and $\tilde{\psi}$ are
dimensionless (in real space and time),
$\dimensionof{\tilde{\phi}}=\dimensionof{\tilde{\psi}}=1$, which implies
$\mathtt{A}=\mathtt{T}^{-1}$ and $\mathtt{A}=\mathtt{B}$.
As expected, $\tau$ is then a rate (namely the
transmutation rate) and so is $\sigma$,
$\dimensionof{\tau}=\dimensionof{\sigma}=1/\mathtt{T}$. Also not unexpectedly, the
remaining four couplings all end up having the same engineering
dimension,
$\dimensionof{\lambda}=\dimensionof{\kappa}=\dimensionof{\chi}=\dimensionof{\xi}=\mathtt{L}^d\mathtt{T}^{-1}$,
as suggested by $\gamma/n_0$, which is a rate
per density ($\gamma$ being the spawning rate and $n_0$ turning into a
carrying capacity density as we take the continuum limit).
\subsection{Observables at tree level: Bulk}\slabel{tree_level}
The aim of the present work is to characterise the volume of the Wiener
Sausage field-theoretically.
As discussed in \Sref{model}, this is done not in terms of an actual spatial
volume, but rather in terms of the \emph{number} of spawned immobile
offspring.
In this section, we define the relevant observables in terms
of the fields introduced above. This is best done at tree level,
presented in the following, before
considering loops and the subsequent renormalisation. While the
tree level is the theory valid above the upper critical dimension, it is
equivalently the theory valid in the absence of any physical
interaction, \latin{i.e.}\@\xspace the theory of $n_0\to\infty$. We introduce the
observables first in the presence of a mass $r$, which amounts to
removing the particle after a time of $1/r$ on average.
If $v^{(1)}(\gpvec{x};\gpvec{x}^*)$ is the density of substrate particles at $\gpvec{x}$ in a
particular realisation of the process at the end of
the life time of the diffusive particle which started at $\gpvec{x}^*$, the volume of the Sausage is
$V=\int\ddint{x} v^{(1)}(\gpvec{x};\gpvec{x}^*)$. The ensemble average is then just
$\ave{V}=\int\ddint{x} \ave{v^{(1)}(\gpvec{x};\gpvec{x}^*)}$, where
$\ave{\bullet}$
denotes the ensemble average of $\bullet$ and the dependence on
$\gpvec{x}^*$ drops out in the bulk. Alternatively (as done below), one may
consider a distribution\footnote{This is a distribution of
experiments with \emph{one} initial particle each, rather than a
``smeared out'' initial particle, whose ``constituents'' would be able
to ``see'' each other's traces.} $d(\gpvec{x}^*)$ of initial starting points
$\gpvec{x}^*$, over which
an additional expectation, denoted by an overline, $\spave{\bullet}$, has to be taken.
Higher moments require higher order correlation functions
\begin{equation}
\ave{V^n} = \int\ddint{x_1}\ldots\ddint{x_n}
\ave{\spave{v}^{(n)}(\gpvec{x}_1,\ldots,\gpvec{x}_n)} \ ,
\end{equation}
where
\begin{equation}
\ave{\spave{v}^{(n)}(\gpvec{x}_1,\ldots,\gpvec{x}_n)}
= \int \dint{\gpvec{x}^*} d(\gpvec{x}^*)
\ave{v^{(n)}(\gpvec{x}_1,\ldots,\gpvec{x}_n;\gpvec{x}^*)}
\end{equation}
and $\ave{v^{(n)}(\gpvec{x}_1,\ldots,\gpvec{x}_n;\gpvec{x}^*)}$
denotes the $n$-point correlation function of
the substrate particle density generated by a single diffusive particle
started at $\gpvec{x}^*$. Equivalently in $\gpvec{k}$-space
\begin{equation*}
\ave{\spave{v}^{(n)}(\ldots)}
\!=\!
\int \ddint{x^*} d(\gpvec{x}^*)
\ave{v^{(n)}(\ldots;\gpvec{x}^*)}
\!=\!
\int \ddintbar{k^*} d(-\gpvec{k}^*) \ave{v^{(n)}(\ldots;\gpvec{k}^*)}.
\end{equation*}
Given that $\creat{b}{\gpvec{x}}\annih{b}{\gpvec{x}}$ is
the particle density operator, that correlation function is the
expectation
\begin{multline}
\ave{v^{(n)}(\gpvec{x}_1,\ldots,\gpvec{x}_n;\gpvec{x}^*)}
=
\lim_{t_1,t_2,\ldots,t_n\to\infty}
\left\langle
\psi^{\dagger}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)
\right.
\\
\left.
\times\psi^{\dagger}(\gpvec{x}_2,t_2)\psi(\gpvec{x}_2,t_2)
\times \ldots \times
\psi^{\dagger}(\gpvec{x}_n,t_n)\psi(\gpvec{x}_n,t_n)
\times \phi^{\dagger}(\gpvec{x}^*,t_0)
\right\rangle
\end{multline}
with only a single,\footnote{Obviously, an entirely different set of
diagrams will be generated by having more than one initial particle,
clearly so in the presence of a finite carrying capacity, where the
trail of one suppresses the trail of the other, $\tikz[baseline=-3pt]{
\draw[tsubstrate](-0.5,0) -- (0,0);
\draw[tactivity](0.7,0) -- (0,0);
\draw[tactivity](0.7,-0.2) -- (0.3,-0.2);
\draw[tsubstrate](0.3,-0.2) -- (0,0);
} \propto \sigma\tau$. One can see that this diagram
contains an infrared divergence for $d\ge4$
\cite{vandenBergBolthausendenHollander:2004}.}
initial, diffusive particle started at $\gpvec{x}^*,t_0$. The multiple
limits on the right are needed so we measure deposition due to the
active particle left \emph{after} its lifetime.
As the present
phenomenon is time-homogeneous, $t_0$ will not feature explicitly, but
rather enter in the difference $t_i-t_0$, each of which diverges as the
limits are taken.
In principle, only a single
limit is needed, $t=t_1=t_2=\ldots=t_n\to\infty$, but as discussed
below, equal times leave some ambiguity that can be avoided.
For $n=1$, the relevant correlation function is
$\ave{
\psi^{\dagger}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)
\phi^{\dagger}(\gpvec{x}^*,0)}$, which leaves us with four terms after
replacing by Doi-shifted creation operators,
\begin{multline}\elabel{v_1_term_by_term}
\ave{
\psi^{\dagger}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)
\phi^{\dagger}(\gpvec{x}^*,0)}
\\
=
\ave{\psi(\gpvec{x}_1,t_1)}
+\ave{\tilde{\psi}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)}\\
+\ave{\psi(\gpvec{x}_1,t_1)\tilde{\phi}(\gpvec{x}^*,0)}
+\ave{\tilde{\psi}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)\tilde{\phi}(\gpvec{x}^*,0)} \ .
\end{multline}
Pure annihilation, $\ave{\psi}$, vanishes --- it is
the expected density of substrate particles in the vacuum, as no active
particle has been created first. The
expectation
$\ave{\tilde{\psi}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)}\propto\theta(t_1-t_1)$ vanishes as
well, for $\theta(0)=0$ (effectively the It{\=o} interpretation
of the time derivatives, \cite{Taeuber:2014}) is needed in order to make the
Doi-Pelitti approach meaningful. The field $\tilde{\psi}(\gpvec{x}_1,t_1)$
in the density $\tilde{\psi}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)$ is meant to \emph{re}-create the particle annihilated by the operator
corresponding to $\psi(\gpvec{x}_1,t_1)$. For the same reason,
$\ave{\tilde{\psi}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)\tilde{\phi}(\gpvec{x}^*,0)}$
vanishes, even when a vertex,
\[
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0.7,0) -- (0,0) node[at start, right] {$0$};
\draw[substrate] (0,0) -- (-0.7,0) node[at end, left] {$t_1$};
\draw[substrate] (-50:0.7) -- (0,0) node[at start,right] {$t_1$};
}
\]
is available. In fact, to contribute,
any occurrence of
$\tilde{\psi}(\gpvec{x}_1,t_1)$ requires an occurrence of $\psi(\gpvec{x}_2,t_2)$
with $t_2>t_1$. What remains of \Eref{v_1_term_by_term} is therefore
only
\begin{equation}\elabel{tree-level_V1}
\ave{\psi(\gpvec{x}_1,t_1)\tilde{\phi}(\gpvec{x}^*,0)}=\tikz[baseline=-2.5pt]{
\draw[activity] (0.6,0) -- (0,0);
\draw[substrate] (0,0) -- (-0.6,0);
}\ .
\end{equation}
Taking the Fourier transform
of \Eref{deri_transmutation},
\begin{multline}\elabel{FT_deri_transmutation}
\int\dintbar{\omega_0}\exp{-\imath\omega_0 t_0}
\int\dintbar{\omega_1}\exp{-\imath\omega_1 t_1}
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}_0+\gpvec{k}_1)\delta\mkern-6mu\mathchar'26(\omega_0+\omega_1)}{-\imath\omega_1+\epsilon'}
\ \tau \ \frac{1}{-\imath\omega_1+D\gpvec{k}_1^2+r}
\\
=
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}_0+\gpvec{k}_1) \theta(t_1-t_0)
\tau}{D\gpvec{k}_1^2+r-\epsilon'} \left(
\exp{-\epsilon'(t_1-t_0)} - \exp{-(D\gpvec{k}_1^2+r)(t_1-t_0)}
\right)
\end{multline}
reveals the general mechanism of
\begin{equation}\elabel{general_limt_mechanism}
\lim_{t\to\infty}\lim_{\epsilon'\to0}
\int\dintbar{\omega}
\frac{g(\omega)\exp{-\imath\omega t}}{-\imath\omega+\epsilon'}
=g(0) \ ,
\end{equation}
provided $g(\omega)$ itself has no pole at the origin, as otherwise
additional residues that survive the limit $t\to\infty$ would
have to be considered.
In \Eref{FT_deri_transmutation}
the starting point of the walker still enters via $\gpvec{k}_0$. If that ``driving''
is done with a distribution of initial starting points $d(\gpvec{k}_0)$, the
resulting deposition is given by
\begin{equation}\elabel{tau_integral}
\ave{\spave{v}^{(1)}(\gpvec{k})}
= \int \ddintbar{k_0} \ave{v^{(1)}(\gpvec{k};\gpvec{k}_0)}\ d(-\gpvec{k}_0)
= \frac{\tau d(\gpvec{k})}{D\gpvec{k}^2+r}
=
\left.
\tikz[baseline=-2.5pt]{
\draw[Sactivity] (0.6,0) -- (0,0) node[at end,above] {$\tau$};
\draw[substrate] (0,0) -- (-0.6,0);
\draw[very thick,color=red,fill=white] (0.6,0) circle(0.1cm);
}\right|_{\omega=0}
\end{equation}
where the little circle on the right indicates the ``driving'' which ``supplies'' a
certain momentum distribution.
More specifically, an initial
distribution of $d(\gpvec{x})=\delta(\gpvec{x}-\gpvec{x}^*)$ has Fourier components
$$\int\ddint{x} d(\gpvec{x}) \exp{-\imath \gpvec{k}_0 \gpvec{x}}
=\exp{-\imath \gpvec{k}_0 \gpvec{x}^*}
=d(\gpvec{k}_0)$$ and the
resulting deposition is distributed according to
$$\ave{\spave{v}^{(1)}(\gpvec{k};\gpvec{x}^*)}= \tau \exp{-\imath \gpvec{k}
\gpvec{x}^*}/(D\gpvec{k}^2+r)\ .$$
In an infinite system, the position of the initial driving should not
and will not enter --- to calculate the volume of the Sausage, we will
evaluate at $\gpvec{k}=0$. The same applies for the time of when the initial
distribution of particles is made.
In principle it would give rise to an
additional factor of $\exp{-\imath \omega t^*}$, but we will evaluate at
$\omega=0$.
Evaluating at $\gpvec{k}=0$ in the bulk produces the volume integral over the offspring
distribution, \latin{i.e.}\@\xspace the expected volume $V$ of the Sausage, in the absence of a
limiting carrying capacity,
\begin{equation}\elabel{V1_bulk}
\ave{V}=\ave{\spave{v}^{(1)}(\gpvec{k}=0;\gpvec{x}^*)}=\frac{\tau}{r}\ ,
\end{equation}
which corresponds to the
na{\"i}ve expectation of the (number) deposition rate $\tau$ multiplied by
the survival time of the random walker $1/r$. From this expression it is
also clear that the ``volume'' calculated here is, as expected, dimensionless.
Following similar arguments for $n=2$, the relevant diagrams are
\begin{multline}\elabel{second_order_diagrams}
\ave{
\psi^{\dagger}(\gpvec{x}_2,t_2)\psi(\gpvec{x}_2,t_2)
\psi^{\dagger}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)
\phi^{\dagger}(\gpvec{x}_0,t_0)
} \\
=
\ave{
\psi(\gpvec{x}_2,t_2)
\tilde{\psi}(\gpvec{x}_1,t_1)\psi(\gpvec{x}_1,t_1)
\tilde{\phi}(\gpvec{x}_0,t_0)
}
+
\ave{
\psi(\gpvec{x}_2,t_2)
\psi(\gpvec{x}_1,t_1)
\tilde{\phi}(\gpvec{x}_0,t_0)
} \\
\hat{=}\ \tikz[baseline=-2.5pt]{
\draw[Sactivity] (0.6,0) -- (0,0) node[at end,above] {$\tau$};
\draw[substrate] (0,0) -- (-0.6,0);
\draw[substrate] (-0.8,0) -- (-1.4,0);
\crossblob{-0.7,0};
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (-1.0,0) -- (-0.5,0) node[at end, above] {$\tau$};
\draw[Sactivity] (0,0) -- (-0.5,0);
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\sigma$};
\draw[substrate] (-130:0.7) -- (0,0);
} \ ,
\end{multline}
where the symbol $\tikz[baseline=-2.5pt]{\crossblob{0,0}}$ represents
$\tilde{\psi}(\gpvec{x},t)\psi(\gpvec{x},t)$, which is a convolution in Fourier
space,
\begin{equation}
\tikz[baseline=-2.5pt]{\crossblob{0,0}}
\ \hat{=}
\int\ddint{x}\dint{t} \tilde{\psi}(\gpvec{x},t)\psi(\gpvec{x},t)
\exp{\imath\omega t-\imath\gpvec{k} \gpvec{x}}
=
\int\ddintbar{k'}\dintbar{\omega'}
\tilde{\psi}(\gpvec{k}',\omega')\psi(\gpvec{k}-\gpvec{k}',\omega-\omega')
\end{equation}
so that
\begin{equation}\elabel{twiddle_blob_twiddle}
\tikz[baseline=-2.5pt]{
\draw[substrate] (0,0) -- (-1,0) node[at start,below] {$\gpvec{k}_0,\omega_0$};
\draw[substrate] (-1.2,0) -- (-2.2,0) node [at end, below] {$\gpvec{k}_2,\omega_2$};
\crossblob[$\gpvec{k}_1,\omega_1$]{-1.1,0};
} =
\frac{
\delta\mkern-6mu\mathchar'26(\omega_0+\omega_1+\omega_2)
\delta\mkern-6mu\mathchar'26(\gpvec{k}_0+\gpvec{k}_1+\gpvec{k}_2)
}{
(-\imath\omega_2 + \epsilon')
( \imath\omega_0 + \epsilon')
} \ ,
\end{equation}
which in real space and time gives a
$\delta(\gpvec{x}_2-\gpvec{x}_1)\delta(\gpvec{x}_1-\gpvec{x}_0)\theta(t_2-t_1)\theta(t_1-t_0)$, corresponding to an
immobile particle deposited at $t_0$ and $\gpvec{x}_0$, found later at time $t_1>t_0$ and $\gpvec{x}_1=\gpvec{x}_0$
and left there to be found again at time $t_2>t_1$ and $\gpvec{x}_2=\gpvec{x}_1=\gpvec{x}_0$.
The effect of taking the limits $t_i\to\infty$ is the same as for the first
moment, namely it results in $\omega_i=0$. The same holds here, except that in
diagrams containing the convolution, the result depends on the order in which
the limits are taken. This can be seen in the factor
$\theta(t_2-t_1)\theta(t_1-t_0)$, as one naturally expects from this diagram:
The first probing must occur after creation and the second one after the
first. A diagram like the second in \Eref{second_order_diagrams} does not carry a constraint like that.
Each of the diagrams on the right hand side of
\Eref{second_order_diagrams} appears twice, as the external fields can be
attached in two different ways. When evaluating at $\gpvec{k}_1=\gpvec{k}_2=0$
this would lead to the same (effective) combinatorial factor of $2$ for
both diagrams.
However, taking
the time limits in a particular order
means that one labelling of the first diagram results in a vanishing
contribution. The resulting combinatorial factors are
therefore $1$ for
$\tikz[baseline=-2.5pt,scale=0.7]{
\draw[tSactivity] (0.6,0) -- (0,0);
\draw[tsubstrate] (0,0) -- (-0.6,0);
\draw[tsubstrate] (-0.8,0) -- (-1.4,0);
\crossblob{-0.7,0};
}$
and $2$ for
$\tikz[baseline=-2.5pt,scale=0.7]{
\draw[tsubstrate] (-1.0,0) -- (-0.5,0);
\draw[tSactivity] (0,0) -- (-0.5,0);
\draw[tSactivity] (0.5,0) -- (0,0);
\draw[tsubstrate] (-160:0.7) -- (0,0);
}$, \latin{i.e.}\@\xspace
\begin{equation}\elabel{V2_bulk}
\ave{V^2}=\frac{\tau}{r}\left(1+2\frac{\sigma}{r}\right) \ ,
\end{equation}
again dimensionless. Given that $\tau=\sigma=\gamma$ initially,
\Eref{non-linearities}, the above may be written
$\gamma/r+2\gamma^2/r^2$. Unsurprisingly, the moments correspond to
those expected for a Poisson process with rate $\gamma$ taking place
during the exponentially distributed lifetime of the particle, subject
to a Poisson process with rate $r$. The resulting moment generating
function is simply
\begin{equation}\elabel{MGF_tree_level}
\mathcal{M}(x)=
\frac{r/\gamma}{r/\gamma+1-\exp{x}}
\end{equation}
with $\ave{V^n}=\left.\frac{\mathrm{d}^n}{\mathrm{d} x^n}\right|_{x=0} \mathcal{M}(x)$
reproducing all moments once $\tau=\sigma=\gamma$.
Carrying on with the diagrammatic expansion,
higher order moments can be constructed correspondingly. At tree level (or
$n_0\to\infty$ equivalently), there are no further vertices contributing.
Determining $\ave{v^{(n)}(\gpvec{k}_1,\ldots,\gpvec{k}_n;\gpvec{k}_0)}$ is therefore merely a
matter of adding substrate legs, $\tikz[baseline=-2.5pt]{\draw[tsubstrate]
(0,0) -- (-0.5,0);}$, either by adding a convolution,
$\tikz[baseline=-2.5pt]{\crossblob{0,0}}$, or by
branching with coupling $\sigma$. For example,
\begin{equation}\elabel{tree_v3}
\ave{v^{(3)}}
\hat{=}
\tikz[baseline=-2.5pt]{
\draw[Sactivity] (0.6,0) -- (0,0) node[at end,above] {$\tau$};
\draw[substrate] (0,0) -- (-0.6,0);
\draw[substrate] (-0.8,0) -- (-1.4,0);
\draw[substrate] (-1.6,0) -- (-2.2,0);
\crossblob{-0.7,0};
\crossblob{-1.5,0};
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (-1.1,0) -- (-0.5,0) node[at end, above] {$\tau$};
\draw[Sactivity] (0,0) -- (-0.6,0);
\draw[substrate] (-1.9,0) -- (-1.3,0);
\crossblob{-1.2,0};
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\sigma$};
\draw[substrate] (-130:0.7) -- (0,0);
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (-1.0,0) -- (-0.5,0) node[at end, above] {$\tau$};
\draw[Sactivity] (0,0) -- (-0.6,0);
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\sigma$};
\draw[substrate] (-130:0.6) -- (0,0);
\draw[substrate] (-130:1.4) -- (-130:0.8);
\crossblob{-130:0.7};
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (-1.0,0) -- (-0.5,0) node[at end, above] {$\tau$};
\draw[Sactivity] (0,0) -- (-0.6,0);
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\sigma$};
\draw[substrate] (-130:0.6) -- (0,0);
\begin{scope}[xshift=0.5cm]
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\sigma$};
\draw[substrate] (-130:0.6) -- (0,0);
\end{scope}
}
.
\end{equation}
Upon taking the limits, effective combinatorial factors become $1$, $3$,
$3$ and $6$ respectively, so that
\begin{equation}\elabel{V3_bulk}
\ave{V^3}=
\frac{\tau}{r} \left( 1
+ 6 \frac{\sigma}{r}
+ 6 \left(\frac{\sigma}{r}\right)^2
\right) \ ,
\end{equation}
and similarly
\begin{subeqnarray}{\elabel{higher_bulk_moments}}
\ave{V^4}&=&\frac{\tau}{r} \left(
1
+ 14 \frac{\sigma}{r}
+ 36 \left(\frac{\sigma}{r}\right)^2
+ 24 \left(\frac{\sigma}{r}\right)^3
\right)\\
\ave{V^5}&=&\frac{\tau}{r} \left(
1
+ 30 \frac{\sigma}{r}
+ 150 \left(\frac{\sigma}{r}\right)^2
+ 240 \left(\frac{\sigma}{r}\right)^3
+ 120 \left(\frac{\sigma}{r}\right)^4
\right)\\
\ave{V^6}&=&\frac{\tau}{r} \left(
1
+ 62 \frac{\sigma}{r}
+ 540 \left(\frac{\sigma}{r}\right)^2
+ 1560 \left(\frac{\sigma}{r}\right)^3
+ 1800 \left(\frac{\sigma}{r}\right)^4
\right.\nonumber \\
&& \left. + 720 \left(\frac{\sigma}{r}\right)^5 \right) \ .
\end{subeqnarray}
In general, the leading order behaviour in small $r$ at tree level in the bulk is dominated
by diagrams with the largest number of branches, \latin{i.e.}\@\xspace the largest power of $\sigma$, like the
right-most term in \Eref{tree_v3}, so that
\begin{equation}
\ave{V^m} \propto m! \tau\sigma^{m-1} r^{-m} \ ,
\elabel{general_scaling_tree}
\end{equation}
which is essentially determined by the time the active particle survives.
\subsection{Observables at tree level: open boundary conditions}\slabel{OBC}
Nothing changes diagrammatically when considering the observables
introduced above in systems with open boundary conditions along one axis. As
$n_0\to\infty$ does not pose a constraint, it makes no difference
whether the system is periodically closed (in $d=2$ a finite cylinder) or
infinitely extended (infinite slab) along the
other
axes --- these directions simply do not matter for the observables
studied, except when the diffusion constant enters. What makes the difference to the considerations in the bulk,
\Sref{tree_level}, are \emph{open} dimensions, in the following fixed to
\emph{one}, so that the number of infinite (or, at this stage
equivalently, periodically closed) directions is $\tilde{d}=d-1$; in the following $\gpvec{k},\gpvec{k}'\in\gpset{R}^{\tilde{d}}$.
While the diagrams obviously remain unchanged, their interpretation
changes because of the orthogonality relations as stated in
\Eref{two_terms_Dirichlet} and
\Eref{three_terms_Dirichlet} or, equivalently, the lack of momentum
conservation due to the absence of translational invariance.
Replacing the propagators by
\begin{subequations}\elabel{bare_propagators_open}
\begin{align}\elabel{bare_activity_propagator_open}
\ave{\phi_n(\gpvec{k},\omega)\tilde{\phi}_m(\gpvec{k}',\omega')}_0
&=
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')\delta\mkern-6mu\mathchar'26(\omega+\omega')\frac{L}{2}\delta_{n,m}}{-\imath\omega+D\gpvec{k}^2+Dq_n^2+r}
\ \hat{=}\ \tikz[baseline=-2.5pt]{\draw[activity] (0.6,0) -- (-0.6,0);}\\
\ave{\psi_n(\gpvec{k},\omega)\tilde{\psi}_m(\gpvec{k}',\omega')}_0
&=
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')\delta\mkern-6mu\mathchar'26(\omega+\omega')\frac{L}{2}\delta_{n,m}}{-\imath\omega+\epsilon'}
\ \hat{=}\ \tikz[baseline=-2.5pt]{\draw[substrate] (0.6,0) --
(-0.6,0);} \ ,
\end{align}
\end{subequations}
where a \emph{single} open dimension causes the appearance of the
indices $n$ and $m$,
results in the one point function
\[
\ave{\spave{v}^{(1)}_n(\gpvec{k})}
= \frac{\tau d_n(\gpvec{k})}{D\gpvec{k}^2+Dq_n^2+r}
=
\left.
\tikz[baseline=-2.5pt]{
\draw[Sactivity] (0.6,0) -- (0,0) node[at end,above] {$\tau$};
\draw[substrate] (0,0) -- (-0.6,0);
\draw[very thick,color=red,fill=white] (0.6,0) circle(0.1cm);
}\right|_{\omega=0} \ ,
\]
where the index $n$ refers to the Fourier-$\sin$ component as discussed
in \Sref{Fourier}.
If driving (\latin{i.e.}\@\xspace initial deposition) is uniform (homogeneous) along the open, finite
axis, its Fourier transform is
$d_n(\gpvec{k})=\delta\mkern-6mu\mathchar'26(\gpvec{k})\int_0^L\dint{z}\sin(q_nz)/L=2\delta\mkern-6mu\mathchar'26(\gpvec{k})/(q_nL)$ for odd $n$ and vanishes
otherwise. As for the periodic or infinite dimensions, the distribution
of the driving does not enter into $\ave{V^n}$, as momentum conservation
implies that the only amplitudes of the driving that matter are that of the
$\gpvec{k}=0$ or $k_0=0$ modes, \Eref{fourier_transform} and \Eref{efuncs_exp}.
Integrating $(2/L)\sum_n \ave{\spave{v}^{(1)}_n(\gpvec{k})} \sin(q_nz)$ over
the interval $[0,L]$ produces
\cite{Mathematica:8.0.1.0}
\begin{multline}\elabel{V_open_details}
\ave{V}=\frac{2}{L}\sum_{n\,\text{odd}} \frac{2}{q_n}
\frac{\tau}{Dq_n^2+r} \frac{2}{Lq_n} \\
=
\frac{8\tau}{\pi^4 D}L^2
\sum_{n\,\text{odd}}
\frac{1}{n^2}\frac{1}{n^2+\frac{rL^2}{D\pi^2}}
=\frac{\tau}{r}\left(1-\sqrt{\frac{4D}{rL^2}}
\tanh\left( \sqrt{\frac{rL^2}{4D}}\right) \right) \ .
\end{multline}
In the limit of large $L$ this result recovers \Eref{V1_bulk}, which
would be less surprising if $L\to\infty$ would simply restore the bulk,
which is, however, not the case, because as the driving is uniform, some of it
always takes place ``close to'' the open boundaries. However, open
boundaries matter only up to a distance of $\sqrt{D/r}$ from the
boundaries, \latin{i.e.}\@\xspace the fraction of walkers affected by the open boundaries
is of the order $\sqrt{D/r}/L$.
The limit
$r\to0$ gives $\ave{V}=\tau L^2/(12 D)$, matching results for the
average residence time of a random walker on a finite lattice with
cylindrical boundary conditions using $D=1/(2d)$
\cite{Pruessner_aves:2013}. Sticking with $r\to0$,
calculating higher order moments for uniform driving is straight-forward, although
somewhat tedious. For example, the two diagrams contributing to
$\ave{v^{(2)}}$ are
\begin{multline}\elabel{convolution_tree}
\tikz[baseline=-2.5pt]{
\draw[activity] (1,0) -- (0,0) node[at end, above] {$\tau$} node[at start,below] {$\gpvec{k}_0,\omega_0,n$};
\draw[substrate] (0,0) -- (-1,0);
\draw[substrate] (-1.2,0) -- (-2.2,0) node [at end, below] {$\gpvec{k}_2,\omega_2,l$};
\crossblob[$\gpvec{k}_1,\omega_1,m$]{-1.1,0};
} \\
=
\frac{
\tau \delta\mkern-6mu\mathchar'26(\omega_0+\omega_1+\omega_2)
\delta\mkern-6mu\mathchar'26(\gpvec{k}_0+\gpvec{k}_1+\gpvec{k}_2)
L \epsilon_{nm\ell}
}{
(-\imath\omega_2 + \epsilon')
( \imath\omega_0 + D\gpvec{k}_0^2 + Dq_n^2 +r)
( \imath\omega_0 + \epsilon')
}
\end{multline}
and
\begin{multline}\elabel{branch_tree}
\tikz[baseline=-2.5pt]{
\draw[substrate] (-1.8,0) -- (-0.9,0) node[at end, above] {$\tau$} node [at start, below] {$\gpvec{k}_2,\omega_2,l$};
\draw[Sactivity] (0,0) -- (-0.9,0);
\draw[Sactivity] (0.9,0) -- (0,0) node[at end, above] {$\sigma$} node[at start,below] {$\gpvec{k}_0,\omega_0,n$};
\draw[substrate] (-130:0.7) -- (0,0) node [at start, below] {$\gpvec{k}_1,\omega_1,m$};
} \\
=
\frac{
\tau \sigma \delta\mkern-6mu\mathchar'26(\omega_0+\omega_1+\omega_2)
\delta\mkern-6mu\mathchar'26(\gpvec{k}_0+\gpvec{k}_1+\gpvec{k}_2)
L \epsilon_{nm\ell}
}{
(-\imath\omega_2 + D\gpvec{k}_2^2 + Dq_{\ell}^2 +r)
(-\imath\omega_2 + \epsilon')
( \imath\omega_0 + D\gpvec{k}_0^2 + Dq_n^2 +r)
( \imath\omega_0 + \epsilon')
} \ .
\end{multline}
Using
\begin{subeqnarray}{\elabel{messy_sums}}
2\pi \sum_{\substack{nm\ell\\\text{odd}}}
\frac{1}{n^3}
\frac{1}{m}
\frac{1}{\ell}
\epsilon_{nm\ell} &=& \frac{1}{6} \left(\frac{\pi}{2}\right)^6 \\
2\pi \sum_{\substack{nm\ell\\\text{odd}}}
\frac{1}{n^3}
\frac{1}{m}
\frac{1}{l^3}
\epsilon_{nm\ell} &=& \frac{1}{15} \left(\frac{\pi}{2}\right)^8 \ ,
\end{subeqnarray}
where $n,m,l\in\{1,3,5,\ldots\}$ (as driving is uniform and the sausage volume is an integral over the entire system),
then produces
\begin{equation}\elabel{V2_tree_open}
\ave{V^2}=\frac{\tau L^2}{12D} \left(1+\frac{\sigma L^2}{5D}\right) \ .
\end{equation}
This may be compared to the known expressions for the moments of the
number of distinct sites visited by a random walker within $n$ \emph{moves}
\cite[in particular Eq.~(A.14)]{Torney:1986}, which contains logarithms even in three dimensions, where the present tree level results are valid. This is, apparently,
caused by constraining the length of the Sausage by limiting the number
of moves, rather than a Poissonian death rate.
Performing the summations \Eref{messy_sums} is straight-forward, but
messy and tedious.\footnote{GP would like to thank Aman Pujara and
Songhongyang Yuan for their help.} The relevant sums converge rather quickly,
for the third moment
producing (by summing numerically over $200$ terms for each index), for example
\begin{multline}\elabel{V3_tree_open}
\ave{V^3}
=
\frac{\tau L^2}{12D}
\left(
1.00002196165\ldots+
0.60000307652\ldots \frac{\sigma L^2}{D}\right.\\
\left. + 0.060714286977\ldots \frac{\sigma^2 L^4}{D^2}
\right).
\end{multline}
Just like in the bulk for small $r$, \Eref{higher_bulk_moments},
the diagrams dominating large $L$ are the tree-branch-like diagrams such
as \Eref{branch_tree}, with highest power of $\sigma$, rather than those involving convolutions,
\Eref{convolution_tree}. Each new branch produces a factor $L^2$, so in
general
\begin{equation}\elabel{general_scaling_open_tree}
\ave{V^m}\propto \tau \sigma^{m-1} L^{2m} D^{-m}\ ,
\end{equation}
as in \Eref{general_scaling_tree} essentially determined by the time the
particle stays on the lattice.
Similar to the bulk, the lack of interaction allows the volume moments
of the Sausage to be determined on the basis of the underlying Poisson
process. In the case of homogeneous drive, the $m$th moment of the residence
time $t_r$ of a Brownian particle diffusing on an open interval of length $L$ is
\begin{equation}\elabel{exact_residence_moments}
\ave{t_r^m}=
\frac{8 m!}{\pi^{2(m+1)} D^m} L^{2m}\sum_{n\,\text{odd}} n^{-2(m+1)}
\end{equation}
and the moment generating function of the Poissonian deposition with
rate $\gamma$ is just $\mathcal{M}(z)=\exp{-\gamma t_r (1-\exp{z})}$, so that
$\ave{V^m}=\ave{\mathrm{d}^m \mathcal{M}(z)/\mathrm{d} z^m|{z=0}}$, reproducing the
results above such as
\begin{equation}
\ave{V^3}=\frac{\gamma L^2}{12D}
\left(
1+
\frac{3 \gamma L^2}{5D}+
\frac{17 \gamma^2 L^4}{280 D^2}\ ,
\right)
\end{equation}
confirming, in particular, the high accuracy of the leading order term
in $L$, as $17/280=0.06071428571428571428\ldots$.
\section{Beyond tree level}\slabel{beyond_tree_level}
Below $d_c=2$ the additional vertices parameterised by $\lambda$, $\kappa$,
$\chi$ and $\xi$, \Eref{mu_kappa} and \Eref{rho_nu} respectively, have to
be taken into account. Because $\kappa$ is the only vertex that has the
same number of incoming and outgoing legs, it is immediately clear that
its presence can, and, in fact, will contribute to the renormalisation
of all other vertices, say
\begin{equation}\elabel{kappa_renormalises_sigma}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0,0) -- (-0.4,0) node [at start,above] {$\sigma$};
\draw[Aactivity] (0.4,0) -- (0,0);
\draw[substrate] (-130:0.4) -- (0,0);
}
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\sigma$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\sigma$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\sigma$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\end{scope}
\end{scope}
\begin{scope}[xshift=-1.6cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+ \ldots \ ,
\end{equation}
but in particular itself:
\begin{equation}\elabel{kappa_renormalises_itself}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (0,0) -- (-0.4,0) node [at start,above] {$\kappa$};
\draw[Aactivity] (0.4,0) -- (0,0);
\draw[substrate] (-130:0.4) -- (0,0);
\draw[substrate] (-50:0.4) -- (0,0);
}
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\kappa$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\kappa$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\kappa$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\end{scope}
\end{scope}
\begin{scope}[xshift=-1.6cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+ \ldots \ .
\end{equation}
Among the vertices
introduced in \Sref{non-linearity}, namely
$\tau$
$\tikz[baseline=-2.5pt]{\draw[tsubstrate](0,0) -- (-0.4,0);\draw[tAactivity] (0,0) -- (0.4,0);}$,
$\sigma$
$\tikz[baseline=-2.5pt]{\draw[tAactivity] (-0.4,0) -- (0.4,0);\draw[tsubstrate](-150:0.4) -- (0,0);}$,
$\lambda$
$\tikz[baseline=-2.5pt]{\draw[tsubstrate](0,0) -- (-0.4,0);\draw[tAactivity] (0,0) -- (0.4,0);\draw[tsubstrate](-30:0.4) -- (0,0);}$,
$\kappa$
$\tikz[baseline=-2.5pt]{\draw[tAactivity] (-0.4,0) -- (0.4,0);\draw[tsubstrate](-30:0.4) -- (0,0);\draw[tsubstrate](-150:0.4) -- (0,0);}$,
$\chi$
$\tikz[baseline=-2.5pt]{\draw[tsubstrate](0,0) -- (-0.4,0);\draw[tAactivity] (0,0) -- (0.4,0);\draw[tsubstrate](160:0.4) -- (0,0);\draw[tsubstrate](-30:0.4) -- (0,0);}$
and
$\xi$
$\tikz[baseline=-2.5pt]{\draw[tAactivity] (-0.4,0) -- (0.4,0);\draw[tsubstrate](-30:0.4) -- (0,0);\draw[tsubstrate](-150:0.4) -- (0,0);\draw[tsubstrate](160:0.4) -- (0,0);}$,
none has an outgoing activity leg if it
does not have an incoming activity leg, and all have at least as many
outgoing substrate legs as they have incoming substrate legs. Apart from
$\kappa$, each vertex has either more outgoing substrate legs than
incoming ones or fewer outgoing activity legs than incoming ones.
Combining them in any form will thus never result in a diagram
contributing to the renormalisation of $\kappa$, which has one leg of each kind.
Combinations of other vertices gives rise to ``cross-production'', say
$\chi$,
$\tikz[baseline=-2.5pt]{
\draw[tAactivity] (0.5,0) -- (0,0);
\draw[tsubstrate] (-0.5,0) -- (0,0);
\draw[tsubstrate] (-20:0.5) -- (0,0);
\draw[tsubstrate] (160:0.5) -- (0,0);
}$, by $\lambda\xi$,
$\tikz[baseline=-2.5pt]{
\draw[tactivity] (1,0) -- (0,0);
\draw[tsubstrate] (0,0) -- (-0.5,0);
\draw[tsubstrate] (0,0) to[bend right,in=-140,out=-40] (1.,0);
\begin{scope}[xshift=1.cm]
\draw[tAactivity] (0.5,0) -- (0,0);
\draw[tAactivity] (0,0) -- (-0.5,0);
\draw[tsubstrate] (-20:0.5) -- (0,0);
\draw[tsubstrate] (160:0.5) -- (0,0);
\end{scope}
}$, but none of these terms contains more than one loop without the
involvement of $\kappa$. As for the generation of higher order vertices,
it is clear that the number of outgoing substrate-legs (on the left)
can never be decreased by combining vertices, because within every
vertex the number of outgoing substrate legs is at least that of
incoming substrate legs. In particular $\tikz[baseline=-2.5pt]{
\draw[tAactivity](0.4,0) -- (-0.4,0);
\draw[tsubstrate] (-20:0.4) -- (0,0);
}$
does not exist. A vertex like that, combined, say, with $\sigma$ to
form the bubble
$\tikz[baseline=-2.5pt]{
\draw[tactivity] (0.8,0) -- (0.4,0);
\draw[tactivity] (-0.4,0) -- (-0.8,0);
\draw[tactivity] (0.4,0) to[bend left,in=-140,out=-40] (-0.4,0);
\draw[tsubstrate] (-0.4,0) to[bend right,in=-140,out=-40] (0.4,0);
}$, which renormalises the propagator, suggests the
diffusive movement of active particles is affected by the presence of
substrate particles. This is, by definition of the original problem, not
the case.
Because no active particles are generated solely by a combination of
substrate particles, none of the vertices has more outgoing then
incoming activity legs. Denoting the tree level coupling of the proper vertex (with amputated legs)
\begin{equation}
\Gammaai{m}{n}{a}{b} =
\tikz[baseline=-2.5pt]{
\begin{scope}[rotate=0]
\draw [decorate,decoration={brace,amplitude=5pt}] (-128:1.4cm) -- (-172:1.4) node[pos=0.5,left,xshift=-0.07cm,yshift=-0.14cm] {$a$};
\draw[substrate] (-130:0.3) -- (-130:1.3);
\path [postaction={decorate,decoration={raise=0ex,text along path,
text align={center},
text={|\large|....}}}]
(-155:1.2cm) arc (-155:-130:1.2cm);
\draw[substrate] (-155:0.3) -- (-155:1.3);
\draw[substrate] (-170:0.3) -- (-170:1.3);
\end{scope}
\begin{scope}[rotate=-60]
\draw [decorate,decoration={brace,amplitude=5pt}] (-128:1.4cm) -- (-172:1.4) node[pos=0.5,left,xshift=-0.07cm,yshift=0.14cm] {$m$};
\draw[Aactivity] (-130:0.3) -- (-130:1.3);
\path [postaction={decorate,decoration={raise=0ex,text along path,
text align={center},
text={|\large|....}}}]
(-155:1.2cm) arc (-155:-130:1.2cm);
\draw[Aactivity] (-155:0.3) -- (-155:1.3);
\draw[Aactivity] (-170:0.3) -- (-170:1.3);
\end{scope}
\begin{scope}[rotate=120]
\draw [decorate,decoration={brace,amplitude=5pt}] (-128:1.4cm) -- (-172:1.4) node[pos=0.5,right,xshift=0.07cm,yshift=-0.14cm] {$b$};
\draw[substrate] (-130:0.3) -- (-130:1.3);
\path [postaction={decorate,decoration={raise=0ex,text along path,
text align={center},
text={|\large|....}}}]
(-155:1.2cm) arc (-155:-130:1.2cm);
\draw[substrate] (-155:0.3) -- (-155:1.3);
\draw[substrate] (-170:0.3) -- (-170:1.3);
\end{scope}
\begin{scope}[rotate=180]
\draw [decorate,decoration={brace,amplitude=5pt}] (-128:1.4cm) -- (-172:1.4) node[pos=0.5,right,xshift=0.07cm,yshift=0.14cm] {$n$};
\draw[Aactivity] (-130:0.3) -- (-130:1.3);
\path [postaction={decorate,decoration={raise=0ex,text along path,
text align={center},
text={|\large|....}}}]
(-155:1.2cm) arc (-155:-130:1.2cm);
\draw[Aactivity] (-155:0.3) -- (-155:1.3);
\draw[Aactivity] (-170:0.3) -- (-170:1.3);
\end{scope}
\draw[thick,fill=white] (0,0) circle (0.3cm)
node {};
}
\end{equation}
of the correlation function
\begin{multline}
\Gai{m}{n}{a}{b}
\left(
\{\gpvec{k}_1,\ldots,\gpvec{k}_{m+n+a+b};
\omega_1,\ldots,\omega_{m+n+a+b}\};
D,r,\tau,\sigma,\lambda,\kappa,\chi,\xi
\right)\\
=
\left\langle
\underbrace{\phi(\gpvec{k}_1,\omega_1)\ldots\phi(\gpvec{k}_m,\omega_m)}_{\textcolor{red}{m} \text{ terms}}
\underbrace{\psi\ldots\psi}_{\textcolor{olive}{a} \text{ terms}}
\underbrace{\tilde{\phi}\ldots\tilde{\phi}}_{\textcolor{red}{n} \text{ terms}}
\underbrace{\tilde{\psi}\ldots\tilde{\psi}}_{\textcolor{olive}{b} \text{ terms}}
\right\rangle
\elabel{general_G}
\end{multline}
by $\gammaai{m}{n}{a}{b}$, the topological conditions on the vertices
can be summarised as $a\ge b$, $b\le1$, $n=1$, $m\le1$, $m+a\ge 1$,
which means that there are in fact only four different types of
vertices, namely $n=1$, $b=0,1$ and $m=0,1$, whereas $a$ is hitherto
undetermined. For future reference, we note
\begin{subequations}
\elabel{tau_etc_as_gamma}
\begin{align}
\tau & = \gammaai{0}{1}{1}{0} & \sigma & = \gammaai{1}{1}{1}{0}\\
\lambda & = \gammaai{0}{1}{1}{1} & \kappa & = \gammaai{1}{1}{1}{1}\\
\chi & = \gammaai{0}{1}{2}{1} & \xi & = \gammaai{1}{1}{2}{1}
\end{align}
\end{subequations}
Dimensional analysis gives
\[
\dimensionof{\gammaai{m}{n}{a}{b}}=\dimensionof{\Gammaai{m}{n}{a}{b}}=\mathtt{L}^{d(n+b-1)}
\mathtt{T}^{a-b-1} \mathtt{A}^{m-n+a-b} \mathtt{B}^{n-m}\ .
\]
Because diffusion is to be maintained, it follows that $\mathtt{T}=\mathtt{L}^2$, yet, as indicated
above, the dimensions of $\mathtt{A}$ and $\mathtt{B}$ are to some extent a matter of
choice. Leaving them undetermined results in $d(n+b-1)+2(a-b)\le2$ for
$\Gammaai{m}{n}{a}{b}$ to be relevant in $d$ dimensions. Setting, on the
other hand, $\mathtt{A}=\mathtt{B}=\mathtt{T}^{-1}$ (see above) results in $d(n+b-1)\le2$. As $n=1$, this
implies $(d-2)b+2a\le2$ and $db\le2$, respectively. In both cases, the
upper critical dimension for a vertex with $b\ge1$ and thus $a\ge1$ to
be relevant is $d_c=2$. On the other hand, no loop can be formed if
$b=0$, so above $d=2$ (where $b=1$ is irrelevant) there are no one-particle irreducibles
contributing to any of the $\Gammaai{m}{n}{a}{b}$ and so the set of
couplings introduced above, $\tau$, $\sigma$, $\lambda$, $\kappa$, $\chi$
and $\xi$ remains unchanged. As far as Sausage moments are concerned,
$\lambda$, $\kappa$, $\chi$ and $\xi$ do not enter, as there is no vertex
available to pair up the incoming substrate leg on the right. The
tree level results discussed in \Sref{tree_level} therefore are the
complete theory in $d>d_c=2$.
Below $d_c=2$, the dimensional analysis depends on the choice one makes
for $\mathtt{A}$ and $\mathtt{B}$. If they remain independent, then the only relevant
vertices that are topologically possible are those with $a\le1$,
removing $\chi$ and $\xi$ from the problem. However, it is entirely
consistent (and one may argue, even necessary) to assume $\mathtt{A}=\mathtt{B}=\mathtt{T}^{-1}$,
resulting in no constraint on $a$ at all. Not only are therefore vertices
for all $a$ relevant, what is worse, they are all generated as
one-particle irreducibles. For example, the reducible
diagram
$\tikz[baseline=-2.5pt]{
\draw[tsubstrate] (-1.0,0) -- (-0.5,0);
\draw[tSactivity] (0,0) -- (-0.5,0);
\draw[tSactivity] (0.5,0) -- (0,0);
\draw[tsubstrate] (-160:0.7) -- (0,0);
}$
contributing to $\ave{v^{(2)}}$ at tree level, \Sref{tree_level},
possesses, even
at one loop, two one-particle irreducible
counterparts in $d<2$,
\[
\tikz[baseline=-2.5pt]{
\draw[Sactivity] (0,0) -- (-1,0) node[at end, above] {$\ \ \chi$};
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\sigma$};
\draw[Bsubstrate] (-1,0) to[bend right,in=-90,out=-90] (0,0);
\draw[substrate] (-1,0)+(140:0.5) -- +(0,0);
\draw[substrate] (-1,0)+(220:0.5) -- +(0,0);
}
\qquad
\tikz[baseline=-2.5pt]{
\draw[Sactivity] (0,0) -- (-1,0) node[at end, above] {$\lambda$};
\draw[Sactivity] (0.5,0) -- (0,0) node[at end, above] {$\ \ \xi$};
\draw[Bsubstrate] (-1,0) to[bend right,in=-90,out=-90] (0,0);
\draw[substrate] (-1,0)+(180:0.5) -- +(0,0);
\draw[substrate] (-0,0)+(140:0.5) -- +(0,0);
}
\]
contributing to the corresponding proper vertex. Such diagrams exist for
all $a$, so, in principle, all these couplings have to be allowed for in
the Liouvillian and all have to be renormalised in their own right. The
good news is, however,
that the $Z$-factor of $\kappa$ (see below) contains all infinities of all couplings exactly once, \latin{i.e.}\@\xspace the renormalisation of all couplings can be related to that of $\kappa$ by a
diagrammatic vertex
identity, see \Sref{WardTakahashi}.
\subsection{Renormalisation}\slabel{renorm}
Without further ado, we will therefore carry on with renormalising $\kappa$
only. As suggested in \Eref{kappa_renormalises_itself}, this can be done
to all orders, in a geometric sum. The one and only
relevant integral is\footnote{We have written explicitly $\kappa$ vertices,
including the amputated legs. At this stage it is unimportant which
coupling forms the loop, but this will change when
we study infinite slabs in \Sref{semi_infty_strip}.}
\begin{multline}\elabel{one_loop}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (3,0) -- (2.4,0);
\draw[Aactivity] (0,0) -- (-0.6,0);
\draw[substrate] (0,0) -- (-120:0.7);
\draw[substrate] (2.4,0)+(0,0) -- +(-60:0.7);
\draw[activity] (2.4,0) -- (0,0) node[above,pos=0.5]
{$\gpvec{k}-\gpvec{k}',\omega-\omega'$} node[below,pos=0.5,yshift=-0.1cm] {$\gpvec{k}',\omega'$};
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (2.4,0);
} \\
= \kappa^2 \int \ddintbar{k'}\dintbar{\omega'}
\frac{1}{-\imath\omega'+\epsilon'}
\frac{1}{-\imath(\omega-\omega')+D(\gpvec{k}-\gpvec{k}')^2+r}\\
=
\frac{\kappa^2}{(4\pi)^{d/2}D}\left(\frac{r+\epsilon'-\imath\omega}{D}\right)^{-\epsilon/2}
\Gamma\left(\frac{\epsilon}{2}\right) \ ,
\end{multline}
where $\epsilon=2-d$ and we have indicated the \emph{total} momentum
$\gpvec{k}$ (\latin{i.e.}\@\xspace the sum of the momenta delivered by the two incoming legs) and the \emph{total} frequency $\omega$ going through
it.\footnote{Here and in the following we obviously choose
$\{\gpvec{k}_i,\omega_i\}=\{0,0\}$ in the renormalisation condition, \latin{i.e.}\@\xspace
$\gammaai{m}{n}{a}{b}_{\mathscr{R}} = \Gammaai{m}{n}{a}{b} (\{0,0\})$.}
This integral has the remarkable property that it is independent of $\gpvec{k}$,
because of the $\gpvec{k}$-independence of the substrate propagator. While
the latter conserves momentum in the bulk by virtue of
$\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')$ in \Eref{substrate_propagator}, its amplitude
does not depend on $\gpvec{k}$. Even if there were renormalisation of the
activity propagator it would therefore not affect its
$\gpvec{k}$-dependence, \latin{i.e.}\@\xspace $\eta=0$, whereas its $\omega$ dependence may be
affected, \latin{i.e.}\@\xspace $z\ne2$ would be possible.
The expression $((r+\epsilon'-\imath\omega)/D)^{1/2}$ can be identified
as an inverse length; it is the infrared regularisation (or more precisely the normalisation point, $R=1$, \Eref{def_R}) that can, in the
present case, be implemented either by considering finite time
($\omega\ne0$), spontaneous extinction of activity ($r>0$) or, notably,
spontaneous extinction (evaporation) of substrate particles
($\epsilon'>0$). In order to extract exponents, it is replaced by
the arbitrary inverse length scale $\mu$. We will return to the case
$\mu=\sqrt{-\imath\omega/D}$ in \Sref{scaling}, \latin{e.g.}\@\xspace \Eref{V_t_exact}. For the time being,
the normalisation point is
\begin{equation}\elabel{def_mu}
\mu^2=\frac{r}{D}
\end{equation}
with $\epsilon'\to0$, $\omega\to0$.
The renormalisation conditions are then (see \Eref{tau_etc_as_gamma})
\begin{subequations}
\elabel{tau_R_etc}
\begin{align}
\tau_{\mathscr{R}} & = \Gammaai{0}{1}{1}{0}\left(\{0,0\}\right) & \sigma_{\mathscr{R}} & = \Gammaai{1}{1}{1}{0}\left(\{0,0\}\right)\\
\lambda_{\mathscr{R}} & = \Gammaai{0}{1}{1}{1}\left(\{0,0\}\right) & \kappa_{\mathscr{R}} & = \Gammaai{1}{1}{1}{1}\left(\{0,0\}\right)\\
\chi_{\mathscr{R}} & = \Gammaai{0}{1}{2}{1}\left(\{0,0\}\right) & \xi_{\mathscr{R}} & = \Gammaai{1}{1}{2}{1}\left(\{0,0\}\right)
\end{align}
\end{subequations}
where $\{0,0\}$ indicates that the vertex is evaluated at vanishing momenta and frequencies. Defining $Z=\kappa_{\mathscr{R}}/\kappa$ allows all renormalisation to be expressed in terms of $Z$, as detailed in \Sref{WardTakahashi}.
Starting with only one loop,
the renormalisation of $\kappa$,
\Eref{kappa_renormalises_itself}, is therefore
$\kappa_{\mathscr{R}} = \kappa - \kappa^2 W$ with
\begin{equation}
W = \frac{\Gamma\left(\frac{\epsilon}{2}\right)}{(4\pi)^{d/2}D}
\mu^{-\epsilon}
\elabel{def_W}
\end{equation}
or $\kappa_{\mathscr{R}}=\kappa Z$ with $Z=1-\kappa W$.
Introducing the dimensionless
coupling $g=\kappa W/\Gamma(\epsilon/2)$ with $g_{\mathscr{R}}=gZ$ gives
$Z=1-g\Gamma(\epsilon/2)$, which may be approximated to one loop by
$Z=1-g_{\mathscr{R}}\Gamma(\epsilon/2)$.
Keeping, however, \emph{all} loops in \Eref{kappa_renormalises_itself},
this last expression is no longer an approximation:
if all terms in
\Eref{kappa_renormalises_itself} are retained, $Z$ becomes a
geometric sum in $g$,
\begin{equation}\elabel{Z_exact}
Z=1-\kappa W+(\kappa W)^2-\ldots=\frac{1}{1+\kappa W} =
\frac{1}{1+g\Gamma(\epsilon/2)} = 1-g_{\mathscr{R}}\Gamma(\epsilon/2) \ ,
\end{equation}
incorporating all parquet diagrams \cite{Herbut:2007}.
The resulting
$\beta$-function is
$\beta_g(g) = \mathrm{d} g_{\mathscr{R}}/\mathrm{d} \ln \mu|_g g_{\mathscr{R}} = -\epsilon g_{\mathscr{R}} - \kappa W \beta_g$
and therefore
\begin{equation}
\beta_g(g)
= \frac{-\epsilon g_{\mathscr{R}}}{1+\kappa W}
= - \epsilon g_{\mathscr{R}} Z
= -\epsilon g_{\mathscr{R}} \left(1-g_{\mathscr{R}} \Gamma(\epsilon/2)\right) \ .
\end{equation}
The last statement is exact to all orders; the non-trivial fixed point
in $\epsilon>0$ is exactly $g_{\mathscr{R}}^*=1/\Gamma(\epsilon/2)\approx\epsilon/2$, which is when the
$Z$-factor vanishes (as $g$ diverges in small $\mu$).
\subsubsection{Ward-Takahashi and vertex identities}\slabel{WardTakahashi}
Different vertices and therefore the renormalisation of different
couplings can be related to each other by Ward-Takahashi identities.
They are usually
constructed by considering global symmetries \cite{Zinn-Justin:1997}, such as
the invariance of the Liouvillian under \cite{BenitezWschebor:2013}
\begin{align}
\phi &\to \phi (1+\delta) & \tilde{\phi} &\to \tilde{\phi} (1+\delta)^{-1} \\
\psi &\to \psi (1+\delta) & \tilde{\psi} &\to \tilde{\psi} (1+\delta)^{-1} \ ,
\end{align}
to be considered for small $\delta$,
which produces an identity on couplings involving an odd number of
fields,
\begin{equation}
(n-m+b-a) \Gammaai{m}{n}{a}{b}=
\int \ddint{x} \dint{t} \left(
-\sigma \frac{\partial}{\partial\sigma}
+\lambda \frac{\partial}{\partial\lambda}
-\xi \frac{\partial}{\partial\xi}
\right) \Gammaai{m}{n}{a}{b} \ .
\elabel{WardTakahashi}
\end{equation}
The identities derived in the following are certainly consistent with
\Eref{WardTakahashi}, but derived at diagrammatic level. To start with,
we reiterate that \Eref{kappa_renormalises_itself} contains \emph{all} contributions (and to all
orders) to
$\Gammaai{1}{1}{1}{1}$, the renormalised vertex $\kappa$.
Repeating
for $\sigma$,
$\tikz[baseline=-2.5pt]{\draw[tAactivity] (-0.4,0) -- (0.4,0);\draw[tsubstrate](-150:0.4) -- (0,0);}$,
and $\lambda$,
$\tikz[baseline=-2.5pt]{\draw[tsubstrate](0,0) -- (-0.4,0);\draw[tAactivity] (0,0) -- (0.4,0);\draw[tsubstrate](-30:0.4) -- (0,0);}$,
the diagrammatic, topological argument presented for $\kappa$ after
\Eref{kappa_renormalises_itself},
it turns out that diagrams contributing to their
renormalisation are essentially identical to those contributing to
$\kappa$, as shown in \Eref{kappa_renormalises_sigma}.
Using the same notation as in
\Eref{tau_R_etc}, we note that $\kappa_{\mathscr{R}}=\kappa Z$ implies
$\sigma_{\mathscr{R}}=\sigma Z$ and $\lambda_{\mathscr{R}}=\lambda Z$, \latin{i.e.}\@\xspace
\begin{equation}
\lambda_{\mathscr{R}}=\frac{\lambda}{\kappa}\kappa_{\mathscr{R}} \qquad
\sigma_{\mathscr{R}}=\frac{\sigma}{\kappa}\kappa_{\mathscr{R}} \ .
\end{equation}
The renormalisation of the coupling $\tau$ breaks with that pattern as
\begin{equation}\elabel{tau_R}
\tau_{\mathscr{R}} = \tau \left( 1 + \frac{\sigma\lambda}{\kappa \tau} (Z-1)\right) \ ,
\end{equation}
because the tree level contribution $\tau$, \Eref{deri_transmutation}, has higher
order
corrections such as
$\tikz[baseline=-2.5pt]{
\draw[tsubstrate] (-0.3,0) -- (0,0);
\draw[tAactivity] (1,0) -- (1.3,0);
\draw[tactivity] (1,0) -- (0,0);
\draw[tsubstrate] (0,0) to[bend right,in=-140,out=-40] (1.,0);
}$, which do not contain $\tau$ itself, but rather the combination
$\lambda\sigma$. However, at bare level, $\sigma=\tau$ and $\lambda=\kappa$, so that in the present case
\begin{equation}\elabel{tau_Ward_id}
\tau_{\mathscr{R}} = \frac{\tau}{\kappa} \kappa_{\mathscr{R}} \ .
\end{equation}
A different issue affects the renormalisation of $\chi$ and $\xi$. For
example, the
latter acquires contributions from any of the diagrams shown in
\Eref{kappa_renormalises_itself}
by ``growing an outgoing substrate leg'', $\tikz[baseline=-5.5pt]{
\draw[tsubstrate] (160:0.4) -- (0,0);}$, on any of the $\kappa$
vertices,
\begin{multline}\elabel{diagrams_xi}
\tikz[baseline=-2.5pt]{
\draw[substrate] (130:0.4) -- (0,0);
\draw[Aactivity] (0,0) -- (-0.4,0) node [at start,above] {$\ \xi$};
\draw[Aactivity] (0.4,0) -- (0,0);
\draw[substrate] (-130:0.4) -- (0,0);
\draw[substrate] (-50:0.4) -- (0,0);
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (-0.4,0)+(130:0.4) -- +(0,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\kappa$} node [at end,above] {$\ \xi$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (0.4,0)+(130:0.4) -- +(0,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\ \xi$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
}
\\
+
\tikz[baseline=-2.5pt]{
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\kappa$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate] (-0.4,0)+(130:0.4) -- +(0,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\ \xi$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (-0.4,0)+(130:0.4) -- +(0,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\kappa$} node [at end,above] {$\ \xi$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+
\tikz[baseline=-2.5pt]{
\draw[substrate] (0.4,0)+(130:0.4) -- +(0,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at start,above] {$\ \xi$} node [at end,above] {$\kappa$};
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[substrate,xshift=0.4cm] (-50:0.4) -- (0,0);
\draw[Aactivity] (0.8,0) -- (0.4,0);
\begin{scope}[xshift=-0.8cm]
\draw[Bsubstrate] (-0.4,0) to[bend right,in=-90,out=-90] (0.4,0);
\draw[activity] (0.4,0) -- (-0.4,0) node [at end, above] {$\kappa$};
\begin{scope}[xshift=-0.4cm]
\draw[Aactivity] (0,0) -- (-0.4,0);
\draw[substrate] (-130:0.4) -- (0,0);
\end{scope}
\end{scope}
}
+ \ldots
\end{multline}
whereas contributions from
$\tikz[baseline=-2.5pt]{
\draw[tactivity] (1,0) -- (0.5,0);
\draw[tactivity] (0.5,0) -- (0,0);
\draw[tsubstrate] (0,0) to[bend right,in=-140,out=-40] (1.,0);
\draw[tsubstrate] (0.5,0)+(160:0.5) -- +(0,0);
}$, generated by $\sigma \mathrm{d}/\mathrm{d} r$ are UV finite and therefore
dropped. Given that \Eref{diagrams_xi} are the only contributions to the
renormalisation of $\xi$, it reads
\begin{equation}
\xi_{\mathscr{R}}=2 \xi \frac{\mathrm{d} \kappa_{\mathscr{R}}}{\mathrm{d} \kappa} - \xi \frac{\kappa_{\mathscr{R}}}{\kappa}
\end{equation}
and correspondingly for the one-particle irreducible contributions to $\chi_{\mathscr{R}}$
\begin{equation}
\chi_{\mathscr{R}}=
2 \chi \frac{\mathrm{d} \kappa_{\mathscr{R}}}{\mathrm{d} \kappa} - \chi \frac{\kappa_{\mathscr{R}}}{\kappa}
\ ,
\end{equation}
where we have used $\chi-\xi\lambda/\kappa=0$.
From \Sref{renorm}, it is straight forward to show that
\begin{equation}
\frac{\mathrm{d} \kappa_{\mathscr{R}}}{\mathrm{d} \kappa} = Z^2
\end{equation}
and we can therefore summarise
\begin{subequations}
\elabel{renorm_summary}
\begin{align}
\tau_{\mathscr{R}} & = \tau Z & \sigma_{\mathscr{R}} & = \sigma Z \\
\lambda_{\mathscr{R}} & = \lambda Z & \kappa_{\mathscr{R}} & = \kappa Z \\
\chi_{\mathscr{R}} & = \chi (2Z^2-Z) & \xi_{\mathscr{R}} & = \xi (2Z^2-Z)
\end{align}
\end{subequations}
In $d<2$, the only proper vertices $\Gammaai{n}{m}{a}{b}$ to consider are
those with $n=1$, $b\le1$, $m\le1$ and arbitrary $a$. The
renormalisation for all of them can be
traced back to that of $\Gammaai{1}{1}{1}{1}$.
It is a matter of straight-forward algebra to demonstrate this explicitly. As these couplings play no further r{\^o}le for the observables analysed henceforth, we
spare the reader a detailed account.
\subsection{Scaling}\slabel{scaling}
We are now in the position to determine the scaling of \emph{all}
couplings. For the time being, we will focus solely, however, on calculating the first moment of the Sausage volume.
We have noted earlier (\Sref{beyond_tree_level}), that the governing
non-linearity is $\kappa$ and have already introduced the corresponding
dimensionless, renormalised coupling $g_{\mathscr{R}}$ and found its fixed point
value. Following the standard procedure \cite{Taeuber:2014}, we define the \emph{finite}, dimensionless,
renormalised vertex functions
\begin{multline}
\Gammaai{m}{n}{a}{b}\left(
\{\gpvec{k}, \omega\};
D, r, \tau, \sigma, \lambda, \kappa, \chi, \xi
\right)\\
=
\mu^{2-d(n+b-1)} D
\GammaaiR{m}{n}{a}{b}\left(
\{\gpvec{k}, \omega\};
R, T_{\mathscr{R}}, s_{\mathscr{R}}, \ell_{\mathscr{R}}, g_{\mathscr{R}}, c_{\mathscr{R}}, x_{\mathscr{R}}; \mu
\right) \ ,
\elabel{def_GammaaiR}
\end{multline}
where $\{\gpvec{k}, \omega\}$ denotes the entire set of momenta and frequency arguments and $\mu$ is an arbitrary inverse scale.
In principle, there could be more bare couplings and there are certainly more generated, at least in principle, see \Sref{WardTakahashi}.
The vertex functions can
immediately be related to their arguments via \Eref{tau_R_etc} and \Eref{tau_etc_as_gamma}:
\begin{subequations}
\elabel{T_R_etc}
\begin{align}
R & = r D^{-1} \mu^{-2} & \elabel{def_R}\\
T_{\mathscr{R}} & = \tau Z D^{-1} \mu^{-2} & S_{\mathscr{R}} & = \sigma Z D^{-1} \mu^{-2}\\
\ell_{\mathscr{R}} & = \lambda Z D^{-1} \mu^{-\epsilon} (4\pi)^{d/2} & g_{\mathscr{R}} & = \kappa Z D^{-1} \mu^{-\epsilon} (4\pi)^{d/2}\\
c_{\mathscr{R}} & = \chi (2Z^2-Z) D^{-1} \mu^{-\epsilon} & x_{\mathscr{R}} & = \xi (2Z^2-Z) D^{-1} \mu^{-\epsilon} \ ,
\end{align}
\end{subequations}
where the normalisation point is $R=1$. Because
\[
\lim_{g_{\mathscr{R}}\to g_{\mathscr{R}}^*}
\frac{\mathrm{d}}{\mathrm{d} \ln \mu} \ln Z = \epsilon ,
\]
$Z$ scales in $\mu$ like $Z\propto\mu^{\epsilon}$.
The asymptotic solution (of the
Callan-Symanzik equation)
\begin{multline}
z^{2-d(n+b-1)} D\\
\times \GammaaiR{m}{n}{a}{b}\left(
\{\gpvec{k}, \omega\};
R z^{-2},
T_{\mathscr{R}} z^{-2+\epsilon},
s_{\mathscr{R}} z^{-2+\epsilon},
\ell_{\mathscr{R}}^*,
g_{\mathscr{R}}^*,
c_{\mathscr{R}}^*,
x_{\mathscr{R}}^*;
\mu z
\right) \\
= \text{asymptotically constant in small $z$}
\end{multline}
can be combined with
the dimensional analysis of the renormalised vertex function, which gives
\begin{multline}
\GammaaiR{m}{n}{a}{b}\left(
\{\gpvec{k}, \omega\};
R, T_{\mathscr{R}}, s_{\mathscr{R}}, \ell_{\mathscr{R}}, g_{\mathscr{R}}, c_{\mathscr{R}}, x_{\mathscr{R}}; \mu z
\right)\\
=
\GammaaiR{m}{n}{a}{b}\left(
\left\{\frac{\gpvec{k}}{z}, \frac{\omega}{z^2}\right\};
R, T_{\mathscr{R}}, s_{\mathscr{R}}, \ell_{\mathscr{R}}, g_{\mathscr{R}}, c_{\mathscr{R}}, x_{\mathscr{R}}; \mu
\right) \ ,
\end{multline}
to give, using $z^2=r$ and \Eref{def_GammaaiR},
\begin{multline}
\Gammaai{m}{n}{a}{b}\left(
\{\gpvec{k}, \omega\};
D, r, \tau, \sigma, \lambda, \kappa, \chi, \xi
\right)\\
=
r^{1-\frac{d}{2}(n+b-1)}
\Gammaai{m}{n}{a}{b}\left(
\{\frac{\gpvec{k}}{\sqrt{r}}, \frac{\omega}{r}\};
D, 1,
\tau r^{-\frac{2-\epsilon}{2}},
\sigma r^{-\frac{2-\epsilon}{2}},
\lambda, \kappa,
\chi,
\xi
\right)
\end{multline}
As far as scaling (but not amplitudes) is concerned,
the tree level results apply to the right hand side as its mass $r$ is finite, \latin{i.e.}\@\xspace
\begin{equation}
\Gammaai{1}{1}{0}{0}\left(\{0, 0\}; D, r, \tau, \sigma, \lambda, \kappa, \chi, \xi \right)
= r^1 \times 1
\end{equation}
and
\begin{equation}
\Gammaai{0}{1}{1}{0}\left(\{0, 0\}; D, r, \tau, \sigma, \lambda, \kappa, \chi, \xi \right)
\propto r^1 \times \tau r^{-\frac{2-\epsilon}{2}}
\end{equation}
so that following \Eref{V1_bulk}
\begin{equation}\elabel{V_bulk_r}
\ave{V}
=
\frac{\Gammaai{0}{1}{1}{0}}{\Gammaai{1}{1}{0}{0}}
\propto
\tau r^{-d/2} \ .
\end{equation}
If $r^{-1}$ is interpreted as the observation time $t$, the result
\begin{equation}\elabel{V_bulk_t}
\ave{V} \propto t^{d/2}
\end{equation}
in $d<2$ (and $\ave{V} \propto t$ in $d>2$,
\Eref{V1_bulk}) recovers the earlier result in \cite{BerezhkovskiiMakhnovskiiSuris:1989},
including the logarithmic corrections expected at the upper critical
dimension. Eqs.~\eref{V_bulk_r} and \eref{V_bulk_t} are the first two key results for the field theory of the Wiener Sausage reported in the present work. We will now further explore the results and their implications.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\draw[->] (-5,0) -- (5,0) node[at end, below] {$x$};
\draw[line width=0.3cm] (-3,0.3cm) -- (3,0.3cm);
\draw[color=white,ultra thick] (-3,0.3cm) -- (3,0.3cm);
\draw (-3.8,0.3) -- (3.8,0.3);
\draw [|<->|] (3,0.7) -- (3.8,0.7) node[pos=0.5,above] {$\frac{V_0}{2}$};
\draw (3.8,0.4)+(-0.1,0) arc (90:-90:0.1);
\draw [|<->|] (-3,0.7) -- (-3.8,0.7) node[pos=0.5,above] {$\frac{V_0}{2}$};
\draw (-3.8,0.2)+(0.1,0) arc (90:-90:-0.1);
\draw [|<->|] (-3,0.7) -- (3,0.7) node[pos=0.5, above] {points actually visited};
\end{tikzpicture}
\end{center}
\caption{\flabel{1D_footprint}The volume of the Wiener Sausage in one dimension is the length
covered by the Brownian particle (the set of all points actually
visited) plus the volume $V_0$ of the sphere the Brownian
particle is dragging (indicated by the two rounded bumpers).}
\end{figure}
In $d=1$, it is an exercise in complex analysis (albeit lengthy) to determine the amplitude of the first moment. To make contact with established results in the literature, we study the sausage in one dimension after finite time $t$. Following the tree level results Eqs.~\eref{tree-level_V1}, \eref{tau_integral} and \eref{V1_bulk} we now have
\begin{equation}\elabel{V1_integral}
\ave{V}(t) = \int \dint{x_1}
\tikz[baseline=-2.5pt]{
\draw[activity] (0.8,0) -- (0.2,0);
\draw[substrate] (-0.2,0) -- (-0.8,0);
\draw[thick,fill=white] (0,0) circle (0.2cm);
}
= \int \dintbar{\omega} \frac{1}{-\imath \omega + \epsilon'} \tau Z \frac{1}{-\imath \omega + r} \ ,
\end{equation}
where the space integral is taken by setting $\gpvec{k}=0$ and the driving has been evaluated to $d(0)=1$, see \Eref{tau_integral}.
The $Z$-factor is given by \Eref{Z_exact}, but $\mu$ should be replaced
by $\sqrt{-\imath\omega/D}$, as we will consider the double limit
$r,\epsilon'\to0$, but at finite $\omega$, which is the \emph{total}
frequency flowing through the diagram, \Eref{one_loop},
so for $d=1=\epsilon$
\begin{equation}\elabel{Z_exact_d_1}
Z=
\frac{1}{1+\kappa \sqrt{\imath / (4 D \omega)}}
\end{equation}
which for small $\omega$ and therefore large $t$ (which we are interested in) is dominated by $2 \sqrt{-\imath D \omega)}/\kappa$.
Keeping only that term, the integral in \Eref{V1_integral} can be performed and gives
\begin{equation}\elabel{V_t_exact}
\ave{V}(t) = \frac{\tau}{\kappa} 4 \sqrt{\frac{tD}{\pi}} \ .
\end{equation}
On the lattice, \latin{i.e.}\@\xspace before taking the continuum limit, sites have no volume and
the ratio $\tau/\kappa$ is just the carrying capacity $n_0$. Setting that to
unity one recovers, up to the additive volume mentioned above, see \Fref{1D_footprint},
the result in the continuum by
Berezhkovskii, Makhnovskii and Suris
\cite[Eq.~(10)]{BerezhkovskiiMakhnovskiiSuris:1989} which coincides with the
asymptote on the lattice \cite{Montroll:1964,Torney:1986}.
Given the difference in the process and the course a field-theoretic
treatment taken, in particular the continuum limit, one might argue that this is a mere
coincidence. In fact, attempting a similar calculation for the amplitude of the second
moment does not suggest that it can be recovered.
As for higher moments of the volume, in addition to the two diagrams mentioned in \Eref{second_order_diagrams}, there is now also
\begin{equation}
\tikz[baseline=-2.5pt]{
\draw[substrate] (-130:1.3) -- (-130:0.6);
\draw[substrate] ( 130:1.3) -- ( 130:0.6);
\draw[Sactivity] (1.3,0) -- (0.6,0);
\draw[thick,fill=white] (0,0) circle (0.6cm) node {$\Gammaai{0}{1}{2}{0}$};
}
\end{equation}
and $\Gammaai{0}{1}{2}{0}= \chi\sigma(3Z-2Z^2-1)/\kappa$. However, as
above, the second moment is dominated in small $r$ by the second, tree-like term in
\Eref{second_order_diagrams}, which gives to leading order
\begin{equation}\elabel{V2_bulk_r}
\ave{V^2} \propto 2 \frac{\tau Z \sigma Z }{r^2} \propto 2 \tau \sigma r^{-d} \ ,
\end{equation}
as $Z\propto r^{\epsilon/2}$. Higher order moments follow that pattern
$\ave{V^m}\propto Z^m$,
and as
dimensional consistency is maintained by the dimensionless product
$r D^{d/\epsilon} \kappa^{-2/\epsilon}$ entering $Z$, Eqs.~\eref{def_mu}, \eref{def_W} and \eref{Z_exact},
\[
Z=\frac{1}{1+\Gamma\left(\frac{\epsilon}{2}\right) (4\pi)^{-d/2} (r D^{d/\epsilon} \kappa^{-2/\epsilon})^{-\epsilon/2}}
\]
the general result is
\begin{equation}\elabel{general_scaling}
\ave{V^m} \propto
m! \tau \sigma^{m-1} r^{-m} \left( \frac{r D^{d/\epsilon}}{\kappa^{2/\epsilon}} \right)^{\epsilon m/2}
=
m! \tau \sigma^{m-1} r^{-md/2} \left( \frac{D^{d/2}}{\kappa} \right)^{m} \ ,
\end{equation}
for $d<2$ with $r\hat{=} 1/t$. Compared to
\Eref{general_scaling_tree} the diffusion constant is present again, as
the coverage depends not only on the survival time (determined by $r$),
but also on the area explored during that time.
\subsection{Infinite slab}\slabel{semi_infty_strip}
In the following, we study the renormalisation of the present field
theory on an infinite slab, \latin{i.e.}\@\xspace a lattice that is finite and open (Dirichlet
boundary conditions) along \emph{one} axis and infinite in $\tilde{d}=d-1$ orthogonal
dimensions. The same setup was considered at tree level in \Sref{OBC}.
Again, there
are no
diagrammatic changes, yet the renormalisation procedure itself requires
closer attention.
Before carrying out the integration of the relevant loop,
\Eref{one_loop}, we make a mild adjustment with respect to the set of
orthogonal functions that we use for the substrate and the activity.
While the latter is subject to Dirichlet boundary conditions in the
present case, naturally leading to the set of $\sin(q_n z)$
eigenfunctions introduced above, the former is not afflicted with such a
constraint, \latin{i.e.}\@\xspace in principle one may choose whatever set is most
convenient\footnote{The existence of a $0$-mode as the space integral is
one feature to consider.} and suitable. As general as that statement is, there are,
however, some subtle implications; to start with, whatever
representation is used in the harmonic part of the Hamiltonian must
result in the integrand factorising, so that the path integral over the
Gaussian can be performed. In the presence of transmutation, that
couples the choice of the set for one species to that for the other.
With a suitable choice, all propagators fulfil orthogonality relations
and therefore conserve momentum, \latin{i.e.}\@\xspace they are proportional to
$\delta_{n,m}$ (in case of the basis $\sin(q_nz)$),
$\delta_{n,-m}$ (basis $\exp{\imath k_nz}$)
and/or
$\delta(\gpvec{k}+\gpvec{k}')$ (basis $\exp{\imath \gpvec{k} z}$), which is obviously a welcome
simplification of the diagrams and their corresponding integrals and
sums.
This constraint
can be relaxed by considering transmutation
only perturbatively, \latin{i.e.}\@\xspace removing it from the harmonic part. However, if
different eigenfunctions are chosen for different species, transmutation
vertices are no longer momentum conserving; if we choose, as we will
below, $\sin(q_n z)$ for the basis of the activity and $\exp{\imath k_m z}$,
then the proper vertex of $\tau$ comes with
\begin{equation}\elabel{def_Delta}
\int_0^L \dint{z} \exp{\imath k_n z} \sin(q_m z) =
L \Delta_{n,m}
\end{equation}
and a summation of the $n$ and $m$, connecting from the sides,
\Eref{deri_transmutation}, \latin{i.e.}\@\xspace
\begin{multline}\elabel{full_tau_open}
\tikz[baseline=-2.5pt]{
\draw[activity] (1,0) -- (0,0) node[at end,above] {$\tau$}
node[at start, below] {$\ell$} node [at end, below, xshift=0.2cm] {$m$};
\draw[substrate] (0,0) -- (-1,0) node [at start, below, xshift=-0.2cm] {$n$}
node[at end, below] {$p$};
}
\hat{=}\ \frac{2}{L}\sum_{n=1}^\infty \frac{1}{L} \sum_{m=-\infty}^\infty
\int \ddMOneintbar{\!k}\, \dintbar{\omega''}\\
\times \ \ave{\psi_p(\gpvec{k},\omega)\tilde{\psi}_n(\gpvec{k}'',\omega'')}
\ \tau L \Delta_{n,m}
\ave{\phi_m(\gpvec{k}'',\omega'')\tilde{\phi}_{\ell}(\gpvec{k}',\omega')}\\
=
\frac{\delta\mkern-6mu\mathchar'26(\gpvec{k}+\gpvec{k}')\delta\mkern-6mu\mathchar'26(\omega+\omega')}{-\imath\omega+\epsilon'}
\ \tau L \Delta_{p,\ell} \ \frac{1}{-\imath\omega+D\gpvec{k}^2+D q_{\ell}^2+r}
\end{multline}
where the $m\in\gpset{Z}$ refers to the index of the eigenfunction used for
the activity and $n\in\gpset{N}^+$ to the eigenfunction of the substrate field. The fact that $\Delta_{p,l}$ has
off-diagonal elements indicates that momentum-conservation is broken.
Obviously, in the presence of boundaries, translational invariance is
always broken, but that does not necessarily result in a lack of
momentum conservation in bare propagators, as it does here. However, it
always results in a lack of momentum conservation in vertices with more
than two legs, as only exponential eigenfunctions have the property that their products
are eigenfunctions as well. If propagators renormalise through these
vertices, they will eventually inherit the non-conservation, \latin{i.e.}\@\xspace
allowing them to have off-diagonal elements from the start will become a necessity in the process of renormalisation.
While the transmutation vertex introduced above may appear unnecessarily messy, it does not renormalise and does not require much further attention.
Rewriting the four-point vertex $\kappa$ in terms of the two different
sets of eigenfunctions, however, proves beneficial. Introducing
\begin{equation}\elabel{def_U}
\int_0^L \dint{z}
\sin(q_n z) \exp{\imath k_m z} \exp{\imath k_k z} \sin(q_{\ell} z)
=
L U_{n,m+k,\ell}
\end{equation}
means that the relevant loop is
\begin{multline}\elabel{one_loop_open}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (3,0) -- (2.4,0) node[at start, above,xshift=0.1cm] {$\ell$};
\draw[Aactivity] (0,0) -- (-0.6,0) node[at end, above,xshift=-0.1cm] {$n$};
\draw[substrate] (0,0) -- (-120:0.7) node[at end,
above,xshift=-0.3cm,yshift=-0.2cm] {$m_2$};
\draw[substrate] (2.4,0)+(0,0) -- +(-60:0.7) node[at end,
above,xshift=0.3cm,yshift=-0.2cm] {$m_1$};
\draw[activity] (2.4,0) -- (0,0)
node[above,pos=0.5] {$\gpvec{k}\!-\!\gpvec{k}',\omega\!-\!\omega',n'$}
node[below,pos=0.5,yshift=-0.07cm] {$\,\gpvec{k}'\!,\omega'\!,m'$};
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (2.4,0);
} \\
=
\kappa^2 \frac{2}{L^2}\sum_{m'=-\infty}^{\infty} \sum_{n'=1}^{\infty}
\int \ddMOneintbar{k'}\dintbar{\omega'}
L^2
U_{n,m_2-m',n'}
U_{n',m'+m_1,\ell}\\
\times
\frac{1}{-\imath\omega'+\epsilon'}
\frac{1}{-\imath(\omega-\omega')+D(\gpvec{k}-\gpvec{k}')^2+D q_{n'}^2+r} \ .
\end{multline}
Contrary to \Eref{one_loop}, it is now of great importance to know with
which couplings (here two $\kappa$ couplings)
this loop was formed, because different couplings require different
``tensors'', like $U_{n,m+k,\ell}$ in the present case. For example, the coupling $\sigma$ comes
with
$\int_0^L \dint{z} \sin(q_n z) \exp{\imath k_m z} \sin(q_{\ell} z)$. The
actual technical difficulty to overcome, however, is the possible
renormalisation of $U_{n,m,\ell}$ itself, as there is no guarantee that the
right hand side of \Eref{one_loop_open} is proportional to $U_{n,m,\ell}$.
In other words, the sum \Eref{kappa_renormalises_itself} may be of the
form $\kappa (L U_{n,m+k,\ell} + \kappa W L U'_{n,m+k,\ell} + \kappa^2 W^2
L U''_{n,m+k,\ell}+\ldots)$, with $U'_{n,m+k,\ell}\ne U'_{n,m+k,\ell}$ \latin{etc.}\@\xspace, rather than $L U_{n,m+k,\ell} \kappa (1+\kappa W +
\kappa^2 W^2 + \ldots)$, which would spoil the renormalisation process.
Carrying on with that in mind, the integrals over $\omega'$ and $\gpvec{k}'$ are identical to the ones
carried out in \Eref{one_loop} and therefore straight-forward. The
summation over $m'$ is equally simple, because that index features only
in $U_{n,m,\ell}$ and \Eref{exp_sum} implies
\begin{multline}\elabel{UU_integral}
\frac{1}{L}
\sum_{m'}
L^2
U_{n,m_2-m',n'}
U_{n',m'+m_1,\ell}
\\
=
\int_0^L \dint{z}
\sin(q_n z)
\exp{\imath k_{m_2} z}
\exp{\imath k_{m_1} z}
\sin(q_{\ell} z)
\sin^2(q_{n'} z) \ .
\end{multline}
Using that identity in \Eref{one_loop_open} allows us to write
\begin{multline}\elabel{one_loop_open_next_step}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
} =
\frac{\kappa^2}{(4\pi D)^{\frac{d-1}{2}}}
\Gamma\left(\frac{3-d}{2}\right)\\
\times \int_0^L \dint{z}
\sin(q_n z)
\exp{\imath k_{m_2} z}
\exp{\imath k_{m_1} z}
\sin(q_{\ell} z)\\
\times
\frac{2}{L}\sum_{n'=1}^{\infty}
\left(D q_{n'}^2 + r + \epsilon' -\imath \omega\right)^{\frac{d-3}{2}}
\sin^2(q_{n'} z) \ .
\end{multline}
It is only that last sum that requires further investigation. In
particular, if we were able to demonstrate that it is essentially
independent of $z$, then the preceding integral becomes
$L U_{n,m_1+m_2,\ell}$ and this contribution to the renormalisation of
$\kappa U_{n,m_1+m_2,\ell}$ is proportional to $U_{n,m_1+m_2,\ell}$.
\newcommand{\operatorname{Li}}{\operatorname{Li}}
The remaining summation in \Eref{one_loop_open_next_step}
can be performed \cite{Mathematica:8.0.1.0} to leading order in the small\footnote{\fnlabel{open_recovers_bulk}For $\rho$
large, $\sum_{n'=1}^{\infty} (n'^2 + \rho)^{\frac{d-3}{2}} \approx
\rho^{\frac{d-2}{2}} \sqrt{\pi}\Gamma(\epsilon/2)/\Gamma((3-d)/2)$, the
open system recovers the results in the bulk, \Eref{one_loop}.}
dimensionless quantity
$\rho=L^2 \left(r + \epsilon' -\imath \omega\right)/(\pi^2 D)$,
\begin{multline}\elabel{open_sum}
\sum_{n'=1}^{\infty} (n'^2 + \rho)^{\frac{d-3}{2}} \sin^2(q_{n'} z) \\
=
\mathchoice{\frac{1}{2}}{(1/2)}{\frac{1}{2}}{(1/2)} \zeta(3-d) - \fourth
\left(
\operatorname{Li}_{3-d}\left(\exp{\frac{2\pi \imath z}{L}}\right)
+
\operatorname{Li}_{3-d}\left(\exp{-\frac{2\pi \imath z}{L}}\right)
\right) + \mathcal{O}(\rho)
\end{multline}
with $\zeta(3-d)=\zeta(1+\epsilon)=(1/\epsilon) + \mathcal{O}(1)$, the Riemann
$\zeta$-function, and $\operatorname{Li}_s(z)$ the polylogarithm with \cite{Mathematica:8.0.1.0}
\begin{equation}\elabel{polylogarithm}
\operatorname{Li}_{1+\epsilon}\left(\exp{\frac{2\pi \imath z}{L}}\right)
+
\operatorname{Li}_{1+\epsilon}\left(\exp{-\frac{2\pi \imath z}{L}}\right)
=-\ln(4 \sin^2(z\pi/L))+\mathcal{O}(\epsilon) \ ,
\end{equation}
so that the leading order behaviour in $\epsilon$ of
\Eref{one_loop_open_next_step} is in fact
\begin{equation}\elabel{one_loop_open_final_step}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
} =
\frac{2\kappa^2 L U_{n,m_1+m_2,\ell}}{(4\pi)^{d/2}D}
\left(\frac{L}{\pi}\right)^{\epsilon}
\zeta(3-d) + \mathcal{O}(\epsilon^0,\rho)
\end{equation}
to leading order in $\epsilon$, where we have used $\Gamma((3-d)/2)=\sqrt{\pi}+\mathcal{O}(\epsilon)$,
anticipating no singularities around $d=3$.
Approximating $2\zeta(3-d)\approx\Gamma(\epsilon/2)$
the $Z$-factor for the renormalisation of $\kappa$ in a system with open
boundaries in one dimension is therefore unchanged, \latin{cf.}\@\xspace Eqs.~\eref{one_loop} and \eref{one_loop_open_final_step}, provided
$\mu=\pi/L$. Of course, that result holds only as long as
$\rho\ll 1$ is small enough, in particular $r\ll D/L^2$, \latin{i.e.}\@\xspace sudden
death by extinction is rare compared to death by reaching the boundary.
In the case of more frequent deaths by extinction, or, equivalently,
taking the thermodynamic limit in the finite, open dimension, extinction is
expected to take over eventually and the bulk results above apply,
\Sref{scaling}. Although there is an effective change of mechanism (bulk
extinction versus reaching the edge), there is no dimensional crossover.
The renormalisation of $\tau$ involves the $\kappa$-loops characterised
above, as well as $\sigma$ and $\lambda$, which, in principle, have to
be considered separately; after all, the loop they form has a structure,
$\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[substrate] (0,0) -- (-0.25,0);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
}$,
that deviates from the structure studied above,
$\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
}$,
\Eref{one_loop_open_final_step}.
In principle, there is (again) no guarantee that the diagrams
contributing to the renormalisation of $\tau$ all have the same dependence
on the external indices, \latin{i.e.}\@\xspace whether they are all proportional to
$\Delta_{n,m}$, \Eref{def_Delta}. By definition, however, \Eref{def_U}
\begin{equation}\elabel{U2Delta}
\frac{2}{L}\sum_{n=1,3,\ldots}^{\infty} L U_{n,m,\ell}\frac{2}{q_n} =
L \Delta_{m,l} \ ,
\end{equation}
\latin{i.e.}\@\xspace one leg is removed by evaluating at $m_1=0$ (see the diagram in \Eref{one_loop_open}) and one by performing the
summation. Applying this operation to all diagrams appearing in
\Eref{kappa_renormalises_itself} produces all diagrams renormalising
$\tau$. Provided that $\sigma=\tau$ and $\lambda=\kappa$, the renormalisation of $\tau$ is therefore linear in that of $\kappa$ and \Eref{tau_Ward_id} remains valid, \latin{i.e.}\@\xspace
the renormalisation procedure outlined above for $\tau$ and
$\kappa$ remains intact.
In principle, further attention is required for
the renormalisation of higher order vertices, but as long as only (external)
substrate legs are attached,
$\tikz[baseline=-2.5pt]{\draw[tsubstrate] (0,0) -- (-0.5,0);}$,
their index $m_n$ can be absorbed into the sum of the indices of the
substrate legs present: Just like any external leg can take up momentum or
frequency, such new legs shift the index used in the internal summation
such as the one in \Eref{one_loop_open}, but that does not affect the
renormalisation provided that it is done at vanishing external momenta, so that the external momenta do not move the poles of the propagators involved.
We conclude that all
diagrammatic vertex
identities of \Sref{WardTakahashi} remain unchanged. As for
the scaling of the Sausage volume, comparing
\Eref{one_loop_open_final_step} to \Eref{one_loop} and identifying
$\mu=\pi/L$ or $r=\pi^2 D/L^2$ means that now
\begin{equation}\elabel{general_scaling_open}
\ave{V^m} \propto
m! \tau \sigma^{m-1}
\left(\frac{L}{\pi}\right)^{md}
\kappa^{-m}
\end{equation}
for $d<2$,
compared to \Eref{general_scaling}. Noticeably, compared to the tree
level \Eref{general_scaling_open_tree}, the diffusion constant is absent ---
in dimensions $d<2$ each point is visited infinitely often, regardless
of the diffusion constant. Even though the deposition in the present
setup is Poissonian, what determines the volume of the sausage is not the
time it takes the active
particles to drop off the lattice, $\propto L^2/D$, but the competition between
deposition parameterised by $\tau$ and $\sigma$ and its inhibition by $\kappa$.
The scaling $\ave{V^m} \propto L^{md}$ for $d<2$ suggests that the Wiener Sausage
is a ``compact'' $d$ dimensional object in dimensions $d<2$, whereas
$\ave{V^m} \propto L^{2m}$ at tree level, $d>2$, \Sref{OBC}. The Wiener
Sausage may therefore be seen as a two-dimensional object projected into
a $d$-dimensional space.
The obvious interpretation of
$r=\pi^2 D/L^2$ in \Eref{general_scaling_open} is that of $\pi/L$ being the lowest mode in the denominator of
the propagator \Eref{bare_activity_propagator_open} in the presence of open boundaries compared to
(effectively) $\sqrt{r/D}$ at $\gpvec{k}=0$ in
\Eref{bare_activity_propagator}.
It is interesting to determine the amplitude of the scaling in $L$ with
one open boundary, not least in order to determine whether the finding
of \Eref{V_t_exact} being identical to the result known in the literature is a mere
coincidence. Technically, the route to take differs from
\Eref{V_open_details}, because in \Sref{OBC} both substrate as well as
activity were represented in the $\sin$ eigensystem. However,
integrating over $L$ (for uniform driving and in order to determine the volume)
amounts to evaluating the matrix $\Delta_{p,\ell}$ in
\Eref{full_tau_open} at $p=0$ and in that case $L \Delta_{p,\ell}=2/q_{\ell}$
for $\ell$ odd and $0$ otherwise, which reproduces \Eref{V_open_details} at
$r=0$, with $\tau$ replaced by $\tau_{\mathscr{R}}$:
\begin{equation}
\ave{V}=\frac{2}{L}\sum_{n\,\text{odd}} \frac{2}{q_n}
\frac{\tau_{\mathscr{R}}}{Dq_n^2} \frac{2}{Lq_n}
=
\frac{8\tau_{\mathscr{R}}}{\pi^4 D}L^2
\sum_{n\,\text{odd}}
\frac{1}{n^4}
=
\frac{\tau_{\mathscr{R}}}{12 D}L^2 \ .
\end{equation}
To determine $\tau_{\mathscr{R}}=\tau Z$ we replace $W$ in Eqs.~\eref{one_loop}, \eref{def_W} and \eref{Z_exact} by
$2 (L/\pi)^{\epsilon} \zeta(3-d)\Gamma((3-d)/2)/(\sqrt{\pi}D(4\pi)^{d/2})$, according to
\Eref{one_loop_open_final_step}, so that asymptotically in large $L$
\begin{equation}\elabel{V_open_L}
\ave{V}=\frac{\pi^{(5-d)/2}2^d \tau}{24 \zeta(3-d)
\Gamma\left(\frac{3-d}{2}\right)
\kappa}L^d
\end{equation}
which for $d=1$ reproduces the exact result (for uniform driving)
\begin{equation}\elabel{V_open_exact}
\ave{V}=\frac{ \tau}{2 \kappa}L \ ,
\end{equation}
which is easily confirmed from first principles.
However, repeating the calculation for driving at the centre, $x^*=L/2$,
gives $d_n=(-1)^{(n-1)/2}$ for $n$ odd and $0$ otherwise, so that in
$d=1$ after some algebra
\begin{equation}\elabel{V_open_centre}
\ave{V}(x^*=L/2)=\frac{3 \tau}{4 \kappa}L \ ,
\end{equation}
which is somewhat off the exact amplitude of $\ln(2)=0.69314718\ldots$
compared to $3/4$. This is apparently due to the renormalisation of $U_{n,m,\ell}$ in \Eref{one_loop_open_final_step} being correct only up to $\mathcal{O}(\epsilon^0)$, but that problem may require further investigation.
\subsection{Infinite cylinder: crossover}\slabel{infinite_cylinder}
At tree level, \Sref{OBC}, it makes no technical difference to study the Sausage on
a finite cylinder or an infinite slab, because the relevant
observables require integration in space which amounts to evaluating at
$k_n=0$ or $\gpvec{k}=0$ resulting in the same expression, \latin{e.g.}\@\xspace
\Eref{V1_bulk} in both cases.
When including interaction, however, it does matter whether the lattice
studied is infinite in $d-1$ dimensions or periodically closed.
Clearly a periodically closed axis has a $0$-mode and does therefore not impose an effective cutoff in $\gpvec{k}$.
In that respect, periodic closure is identical to infinite extent, while physically it is not (just like at tree level).
One may therefore wonder how periodic closure differs from infinite extent mathematically: How does a finite cylinder differ from an infinite strip?
As a
first step to assess the effect, we \emph{replace} the open dimension (axis) by
a periodically closed one. One may regard this as an unfortunate kludge
--- after all, what we are really interested in is a system that is
finite in two dimensions, namely open in one and periodically closed in the
other. However, if the aim is to study finite size scaling in
$2-\epsilon$ dimensions, then two finite dimensions are already
$\epsilon$ too many.
However, the setup of an infinitely long (in $d-1$ dimensions)
periodically closed tube with circumference $L$ does address the problem in question, namely the difference of $\gpvec{k}=0$ in an infinitely extended axis versus $k_n=0$ in a finite but periodic closed dimension. In addition, an infinite cylinder compared to an infinite strip has
translational invariance restored in the periodic dimension, and therefore
the vertices even for a finite system dramatically simplified.
The physics of a $d$-dimensional system with one
axis periodically closed is quite clear: At early times, or,
equivalently, large extinction rates $r\gg D/L^2$, the periodic closure is
invisible and so the scaling is that of a $d$-dimensional (infinite)
bulk system as described in \Sref{scaling}, $\ave{V^m}\propto
r^{-md/2}$. But when the walker starts to re-encounter, due to the
periodic closure, sites visited earlier, this ``dimension will
saturate'' and so for very small $r$, it will display the scaling of
an infinite $d-1$-dimensional lattice.
Just like for the setup in \Sref{tree_level}, it is most convenient to
study the system for small but finite extinction rate $r$. The integrals
to be performed are identical to \Eref{one_loop_open}, but both sums
have a pre-factor of $1/L$, \Eref{convenient_choices_periodic}, (rather than one having $1/L$ and
the other $2/L$, \Eref{convenient_choices}) and $L U_{n,m,l}$ has the much simpler
Kronecker form
\begin{equation}\elabel{def_Utilde}
\int_0^L \dint{z}
\exp{\imath k_n z} \exp{\imath k_m z}
\exp{\imath k_k z} \exp{\imath k_{\ell} z}
=
L \tilde{U}_{n,m+k,\ell}
= L \delta_{n+m+k+\ell,0} \ .
\end{equation}
Most importantly the expression corresponding to \Eref{UU_integral}
sees
$\sin^2(q_{n'}z)$ replaced by unity, because the bare propagator
corresponding to \Eref{bare_activity_propagator_open} carries a factor
$L\delta_{n+m,0}$, \Eref{efuncs_exp}, rather than
$L\delta_{n,m}/2$, \Eref{efuncs_sin}, which
results in $n'$ of $\tilde{U}_{n,m_2-m',n'}$ to pair up with $-n'$ in
$\tilde{U}_{-n',m'+m_1,\ell}$.
For easier comparison, we will keep $L\tilde{U}_{n,m+k,\ell}$ in the following.
We thus have
(see \Eref{one_loop_open_next_step})
\begin{multline}\elabel{one_loop_cyl_next_step}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
}\\ =
\frac{\kappa^2 L \tilde{U}_{n,m+k,\ell}}{(4\pi D)^{\frac{d-1}{2}}}
\Gamma\left(\frac{3-d}{2}\right)
\frac{1}{L}\sum_{n'=-\infty}^{\infty}
\left(D k_{n'}^2 + r + \epsilon' -\imath \omega\right)^{\frac{d-3}{2}} \ .
\end{multline}
Comparing \Eref{one_loop_cyl_next_step} to \Eref{one_loop_open_next_step}, \Eref{open_sum} and
\Eref{one_loop_open_final_step} and re-arranging terms gives for small
$\tilde{\rho}=L^2 \left(r + \epsilon' -\imath \omega\right)/(4 \pi^2 D)$
\begin{multline}\elabel{one_loop_cyl_final_step}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
} =
\frac{2\kappa^2 L \tilde{U}_{n,m_1+m_2,\ell}}{(4\pi)^{d/2} D}
L^{\epsilon}
\Gamma\left(\frac{3-d}{2}\right)\\
\times
\left\{
(2\pi)^{-\epsilon}
\frac{\zeta(3-d)}{\sqrt{\pi}}
+
\sqrt{\pi}
\left(
\frac{r+\epsilon'-\imath \omega}{D}L^2
\right)^{\frac{d-3}{2}}
\right\}
+ \mathcal{O}(\epsilon^0,\tilde{\rho})\\
\end{multline}
and for large $\tilde{\rho}$
\begin{multline}\elabel{one_loop_cyl_final_step_large_circ}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
} =
\frac{\kappa^2 L \tilde{U}_{n,m_1+m_2,\ell}}{(4\pi)^{d/2} D}
\left(\frac{r+\epsilon'-\imath\omega}{D}\right)^{-\frac{\epsilon}{2}}
\Gamma\left(\frac{\epsilon}{2}\right)
+ \mathcal{O}(\epsilon^0,\tilde{\rho}^{\frac{d-3}{2}})
\end{multline}
using
\begin{subnumcases}{\elabel{cyl_sum}
\hspace*{-1cm}
\sum_{n'=-\infty}^{\infty} (n'^2 + \tilde{\rho})^{\frac{d-3}{2}} =}
\tilde{\rho}^{\frac{d-3}{2}} + 2 \zeta(3-d) + \mathcal{O}(\tilde{\rho}) &
for $\tilde{\rho} \ll 1$ \elabel{cyl_sum_small}\\
\tilde{\rho}^{\frac{d-2}{2}}
\frac{\sqrt{\pi} \Gamma\left(\frac{2-d}{2}\right)}{\Gamma\left(\frac{3-d}{2}\right)}
+ \mathcal{O}\left(\tilde{\rho}^{\frac{d-3}{2}}\right)&
for $\tilde{\rho} \gg 1$ \elabel{cyl_sum_large} \ .
\end{subnumcases}
The asymptotics above are responsible for all the interesting features
to be discussed in the following. Firstly, intuition seems to play
tricks: One may think that for small $\tilde{\rho}$ in the sum on the left of \Eref{cyl_sum}, it will always be
large compared to $n'=0$ and always be small compared to $n'\to\infty$.
In fact, one might think there is no difference at all between large or
small $\tilde{\rho}$ and be tempted to approximate the sum immediately by an
integral, $\sum (n'^2 + \tilde{\rho})^{(d-3)/2}\approx
\int_{-\infty}^{\infty} \dint{n'} (n'^2 + \tilde{\rho})^{(d-3)/2} =
\tilde{\rho}^{\frac{d-2}{2}} \sqrt{\pi}
\Gamma\left(\frac{2-d}{2}\right)/\Gamma\left(\frac{3-d}{2}\right)$.
That, however, produces only the second line, \Eref{cyl_sum_large}.
The crucial difference is that in a sum each summand actually
contributes, whereas in an integral the integrand is weighted by the
integration mesh. So, the summand $(n'^2 + \tilde{\rho})^{(d-3)/2}$ has to
be evaluated for $n'=0$, producing $\tilde{\rho}^{\frac{d-3}{2}}$ in
\Eref{cyl_sum_small}, which dominates the sum for $d<2$ (even $d<3$, but
the series does not converge for $2<d$, and, in fact, is not needed as
no IR divergences appear in
$d>2$) and
$\tilde{\rho}\to0$. The remaining terms can actually be evaluated for
$\tilde{\rho}=0$, producing $2\zeta(3-d)$. The integral, which the (Riemann) sum converges to for large $\tilde{\rho}$, on the other hand,
is strictly proportional to $\tilde{\rho}^{\frac{d-2}{2}}$ and therefore
much less divergent than then sum for small $\tilde{\rho}\to0$ and $d<2$.
Of the two regimes $\tilde{\rho}\gg1$ and $\tilde{\rho}\ll1$ the former is more
easily analysed. Setting $\epsilon'-\imath\omega=0$ for the time being,
we notice that $\tilde{\rho}\propto L^2 r$ suggests, somewhat
counter-intuitively, that large $r$, which shortens the lifetime of the
walker, has the same effect as large $L$, which prolongs the time it
takes the walker to explore the system. Both effects are, however, of
the same nature: They prevent the walker from ``feeling'' the
periodicity of the system. In that case, the walker displays bulk
behaviour and in fact, \Eref{one_loop_cyl_final_step_large_circ} is the same as
\Eref{one_loop}.
The other regime, $\tilde{\rho}\ll1$ is richer. At $d<2$ and fixed $L$,
\Eref{one_loop_cyl_final_step} displays a
crossover between the two additive terms on the right hand
side. Stretching the expansion \eref{cyl_sum_small} beyond its stated limits, for
intermediate values of $r$ or $L$, $\tilde{\rho}\approx 1$, the first
term on the right hand side of \Eref{one_loop_cyl_final_step} dominates
and the scaling
behaviour is that of an \emph{open} infinite slab of linear extent
$L$, \Eref{one_loop_open_final_step}. This is because at moderately
large $r$ (or, equally, short times $t$), the walker is not able to
fully explore the infinitely extended directions. But rather than ``falling off'' as
in the system with open boundaries, it starts crossing its
own path due to the periodic boundary conditions, at which point the
scaling like a $d$-dimensional bulk lattice ($\tilde{\rho}\gg1$) ceases and turns into that of a $d$-dimensional open one ($\tilde{\rho}\approx1$). The crossover can
also be seen in \Eref{cyl_sum_small}, which for $d<2$ is dominated by $2\zeta(3-d)$ for
large $\tilde{\rho}$ and by $\rho^{(d-3)/2}$ for small
$\tilde{\rho}$.
As $r$ gets even
smaller (or $t$ increases), $\tilde{\rho}\to0$, the scaling is dominated by the infinite
dimensions, of which there are $\tilde{d}=d-1$, \latin{i.e.}\@\xspace the scaling is that of
a bulk system with $\tilde{d}$ dimensions as discussed in \Sref{renorm}, in
particular \Eref{one_loop}. In this setting, the walker explores an
infinitely long,
thin cylinder, which has effectively degenerated into an infinitely long
line. While the (comparatively) small circumference of the cylinder
remains accessible this is fully explored very quickly compared to the
progress in the infinite directions.
To emphasise the scaling of the last two regimes,
one can
re-write
\Eref{one_loop_cyl_final_step} as
\begin{multline}\elabel{one_loop_cyl_final_step_simplified}
\tikz[baseline=-2.5pt]{
\draw[Aactivity] (1.25,0) -- (1.,0);
\draw[Aactivity] (0,0) -- (-0.25,0);
\draw[substrate] (0,0) -- (-120:0.3);
\draw[substrate] (1,0)+(0,0) -- +(-60:0.3);
\draw[Sactivity] (1,0) -- (0,0);
\draw[substrate] (0,0) to[bend right,in=-120,out=-60] (1.,0);
} =
\frac{2 \kappa^2 L \tilde{U}_{n,m_1+m_2,\ell}}{(4\pi)^{d/2}D}
\frac{\Gamma\left(\frac{1+\epsilon}{2}\right)}{\sqrt{\pi}}
\left(\frac{L}{2\pi}\right)^{\epsilon}
\zeta(1+\epsilon) \\
+
\frac{\kappa^2 L \tilde{U}_{n,m_1+m_2,\ell}}{(4\pi)^{\tilde{d}/2}LD}
\left(
\frac{r+\epsilon'-\imath \omega}{D}
\right)^{-\tilde{\epsilon}/2}
\Gamma\left(\frac{\tilde{\epsilon}}{2}\right)
+
\mathcal{O}(\epsilon^0,\tilde{\rho}) \ ,
\end{multline}
with
$\tilde{\epsilon}=1+\epsilon=3-d$,
$\tilde{d}=d-1$.
Here, the first term displays the behaviour of the infinite slab
discussed above (\Sref{semi_infty_strip},
\Eref{one_loop_open_final_step}, $\zeta(3-d)\propto1/\epsilon$, but $L/\pi$ there and $L/(2\pi)$ here) and the second term that of a
bulk-system with $\tilde{d}$ dimensions, \Eref{one_loop}; the infrared singularity
$(r+\epsilon'-\imath \omega)^{-\tilde{\epsilon}/2}$ is in fact accompanied
by the corresponding ultraviolet singularity $\Gamma(\tilde{\epsilon}/2)$,
exactly as if the space dimension was reduced from $d$ to $\tilde{d}=d-1$.
The second term also
reveals an additional factor $1/L$
compared to \eref{one_loop}.\footnote{The factor $L$ in front of
$\tilde{U}_{n,m_1+m_2,\ell}$ should be regarded as part of the latter, as
the tree level obviously comes with the same pre-factor, see \Eref{def_Utilde}.
Dimensional consistency is maintained by $((r+\epsilon'-\imath\omega)/D)^{-\tilde{\epsilon}/2}$, which has dimension $\mathtt{L}^{1+\epsilon}$.} This expression
determines the factor $W$, which enters the $Z$-factor inversely,
$Z\propto L r^{\tilde{\epsilon}/2}$,
\Eref{Z_exact}, \latin{i.e.}\@\xspace in the present setting, the Sausage volume scales
like $(\tau/r) L r^{\tilde{\epsilon}/2} = \tau L r^{-\tilde{d}/2}$. The
scaling in $t$ is found by replacing $r$ by
$1/t$, or more precisely by $\omega$
and Fourier transforming according to \Eref{V1_integral}, which results
in the scaling
$\ave{V}\propto L t^{1-\tilde{\epsilon}/2}=Lt^{\tilde{d}/2}$.
\section{Summary and discussion}\slabel{discussion}
Because the basic process analysed above is very well understood and has
a long-standing history \cite{KolmogoroffLeontowitsch:1933,Spitzer:1964,KacLuttinger:1974,DonskerVaradhan:1975,LeGall:1986,BerezhkovskiiMakhnovskiiSuris:1989,vanWijlandCaserHilhorst:1997,Sznitman:1998,vandenBergBolthausendenHollander:2001,Spitzer:2001}, this work may add not so much to the
understanding of the process itself, was it not for a field-theoretic
re-formulation, which is particularly flexible and elegant. The price is
a process that ultimately differs from the original model. In hindsight,
the agreement of the original Wiener Sausage problem with the process
used here to formulate the problem field-theoretically deserves further scrutiny.
In the following, we first summarise our findings above with respect to the original Wiener Sausage problem, before discussing in further detail the field-theoretic insights.
\subsection{Summary of results in relation to the original Wiener
Sausage}\slabel{summary_results}
The original Wiener Sausage problem is concerned with the volume traced out by
a finite sphere attached to a Brownian particle. In the present analysis, this
has been replaced by a Brownian particle attempting to spawn immobile offspring
at Poissonian rate $\sigma$. The attempt fails if such immobile particles are
present already.
On the lattice, this process amounts to a variant
of the number of distinct sites visited \cite{Torney:1986}.
Above, the field-theoretic treatment has been carried out perturbatively to one
loop for dimensions $d<2=d_c$, but it turns out that there are no higher order
loops to be considered. In any dimension, by construction and as a matter of universality, the large time and
space asymptotes of the original Wiener Sausage,
the process on the lattice and the field theory are expected to
coincide at least as far as exponents
are concerned.
The tree level of the field theory describes the phenomenon without
interaction, \latin{i.e.}\@\xspace ignoring returns. The resulting observables are the asymptotes
of the Wiener Sausage volume in dimensions above $d=2$. The moments found in
the bulk, Eqs.~\eref{V1_bulk}, \eref{V2_bulk}, \eref{V3_bulk},
\eref{higher_bulk_moments} and generally \eref{general_scaling_tree}, $\ave{V^m} \propto m! \tau\sigma^{m-1} r^{-m}$,
coincide with those from the exact moment
generating function \Eref{MGF_tree_level} of the process ignoring return,
obtained by probabilistic considerations.
In the infinite slab, the field
theory still produces exact results (of the process ignoring return), such as
Eqs.~\eref{V_open_details} and \eref{V2_tree_open}, although higher moments are
tedious to calculate in closed form, \Eref{V3_tree_open}. Again, they are easily
verified using generating functions, such as \Eref{exact_residence_moments}, which
also confirms the general form \Eref{general_scaling_open_tree},
$\ave{V^m}\propto \tau \sigma^{m-1} L^{2m} D^{-m}$, determined
field-theoretically.
Below two dimensions, infrared divergences occur in the perturbation theory,
which need to be controlled by a finite extinction rate $r$ (or $\epsilon'$).
It turns out that all orders can be dealt with at once, because ``parquet diagrams''
\cite{Herbut:2007} can be summed over in a geometric (Dyson) sum, such as
\Eref{kappa_renormalises_itself}. We can therefore expect exact universal exponents
of asymptotes, whereas amplitudes are generally non-universal and can be
affected by field-theoretically irrelevant terms. In the bulk, the
asymptotes Eqs.~\eref{V_bulk_r}, \eref{V_bulk_t},
\eref{V2_bulk_r} and generally \eref{general_scaling},
$\ave{V^m} \propto m! \tau \sigma^{m-1} r^{-md/2} (D^{d/2}/\kappa)^{m}$,
reproduce the (leading order) exponents as known in the literature
\cite{BerezhkovskiiMakhnovskiiSuris:1989}. In one dimension, the first moment
of the volume, \Eref{V_t_exact}, reproduces the asymptote (in large $t$) in the
continuum \cite{BerezhkovskiiMakhnovskiiSuris:1989} and on the lattice
\cite{Montroll:1964,Torney:1986}. Even the amplitude is reproduced correctly.
The bulk calculations can be modified to apply to the infinite slab, producing
\Eref{general_scaling_open}, $\ave{V^m} \propto m! \tau \sigma^{m-1}
(L/\pi)^{md} \kappa^{-m}$. However, the renormalisation in
this case is correct only to leading order in $\epsilon$, as terms of order $\epsilon^0$, such as
\Eref{polylogarithm}, were omitted (whereas in the bulk, the
$Z$-factor was exact, \Eref{Z_exact_d_1} or \Eref{Z_exact}). In one dimension, \latin{i.e.}\@\xspace
when the walker can explore only a finite interval, the amplitude of the first
moment for uniformly distributed initial starting points, \Eref{V_open_L} at $d=1$, coincides with the exact result,
\Eref{V_open_exact}. However, placing the particle
initially at the centre results in an amplitude, \eref{V_open_centre}, that
differs from the exact result.
Unless one is prepared to allow for a space-dependent $\kappa$ (whose
space dependence is in fact irrelevant in the field-theoretic sense) as suggested in \Eref{one_loop_open_next_step} for the infinite slab,
one cannot expect the resulting amplitudes to recover the exact
results. That \Eref{V_open_exact} does so nevertheless, may be explained by
the ``averaging effect'' of the uniform driving, given that
\[
\int_0^L \dint{z}
\left(
\operatorname{Li}_{1+\epsilon}\left(\exp{\frac{2\pi \imath z}{L}}\right)
+
\operatorname{Li}_{1+\epsilon}\left(\exp{-\frac{2\pi \imath z}{L}}\right)
\right)
=
0 \ ,
\]
see \eref{open_sum}.
As alluded to above, the field-theoretic description of the Wiener
Sausage is very elegant, not least because the diagrams have an
immediate interpretation. For example,
$\tikz[baseline=-2.5pt]{
\draw[tactivity](0.4,0) -- (0,0);
\draw[tactivity](0,0) -- (-0.4,0);
\draw[tsubstrate] (-140:0.4) -- (0,0);
}$
corresponds to a substrate particle deposited while the active particle
is propagating. Correspondingly,
$\tikz[baseline=-2.5pt]{
\draw[tactivity] (0,0) -- (-0.3,0);
\draw[tsubstrate] (-140:0.4) -- (0,0);
\draw[tactivity] (1.3,0) -- (1,0);
\draw[tactivity] (1,0) -- (0,0);
\draw[tsubstrate] (0,0) to[bend right,in=-140,out=-40] (1.,0);
}$ is the suppression of a deposition as the active particle encounters
an earlier deposition --- the active particle
\emph{returns} to a place it has been before. All loops can therefore be
contracted along the wavy line,
$\tikz[baseline=-2.5pt]{
\draw[tsubstrate] (-0.4,0) -- (0,0);}$,
to produce a \emph{trajectory}, say
$\tikz[baseline=-2.5pt]{
\draw[tAactivity] (0,0) -- (-0.3,0);
\draw[tAactivity,rounded corners] (0,0) to[bend right,in=-140,out=-40] (0,0.5) to[bend
right,in=-140,out=-40] (0,0);
\draw[tAactivity] (0.3,0) -- (0,0);
\draw[tsubstrate] (-140:0.4) -- (0,0);
}$ or more strikingly just
$\tikz[baseline=-2.5pt]{
\draw[tAactivity] (0,0) -- (-0.3,0);
\draw[tAactivity,rounded corners] (0,0) to[bend right,in=-140,out=-40] (0,0.5) to[bend
right,in=-140,out=-40] (0,0);
\draw[tAactivity] (0.3,0) -- (0,0);
}$, illustrating that the loop integrals calculated above, in fact capture
the probability of a walker to return: $W\propto\omega^{-\epsilon/2}$,
\Eref{def_W}, which in the time domain gives
$t^{-d/2}$.
\subsection{Original motivation}
The present study was motivated by a number of ``technicalities'' which
were encountered by one of us during the study of a more complicated
field theory. The \textbf{first issue}, as mentioned in the introduction, was the
``fermionic'' or excluded-volume interaction. In a first step, that was
generalised to an arbitrary carrying capacity $n_0$, whereby the deposition
rate of immobile offspring varies smoothly in the occupation number
until the carrying capacity is reached.
It was argued above,
\Fref{funny_lattice}, that the constraint to a finite but large carrying
capacity $n_0$, which may be conceived as less brutal than setting
$n_0=1$, can be understood as precisely the latter constraint, but on a
more complicated lattice.
Even though the field theory was constructed
in a straight-forward fashion, the perturbative implementation of the constraint,
namely by effectively discounting depositions that should not have
happened in the first place, make it look like a minor miracle that it
produces the correct scaling (and even the correct amplitudes in some
cases). We conclude that the present
approach is perfectly suitable to implement excluded volume constraints.
It is interesting to vary $n_0$ in the expressions obtained for the
volume moments. At first it may not be obvious that, for example, the
first volume moments in one dimension, \Eref{V_t_exact} and \Eref{V_open_exact}, are linear in
$n_0$, because $\kappa=\tau/n_0$, \Eref{bare_level_couplings}. Given
that $\kappa$ enters the $m$th moment $\ave{V^m}$ as $\kappa^{-m}$,
\Eref{general_scaling} and \Eref{general_scaling_open}, the carrying capacity
therefore enters through $\kappa=\gamma/n_0$ as $n_0^{m}$. Even though
the carrying capacity enters smoothly into the deposition rate (or,
equivalently, the suppression of the deposition), in dimensions $d<2$ each site is visited
infinitely often and is therefore ``filled up to the maximum'' with
offspring particles, as if the carrying capacity was a hard cutoff (\latin{i.e.}\@\xspace as if the deposition rate were constant until the occupation reaches the
carrying capacity). The volume of each sausage therefore increases by a
factor $n_0$ in dimensions $d<2$ and is independent of it (as $\kappa$
does not enter) in $d>2$.
The \textbf{second issue} to be investigated was the presence of open boundaries.
This is, obviously, not a new problem as far as field theory is
concerned in general, but in the present case being able to change
boundary conditions exploits the flexibility of the field-theoretical
re-formulation of the Wiener Sausage and allows us to probe results in
a very instructive way.
It is often said that translational invariance corresponds to momentum
conservation in $\gpvec{k}$-space, but the present study highlights some subtleties.
As far as bare propagators are concerned, open,
periodic, or, in fact, reflecting boundary conditions all allow it to be
written with a Kronecker-$\delta$ function. In that sense, \emph{bare} propagators
do not lose momentum. Momentum, however, is generally not conserved in
vertices, \latin{i.e.}\@\xspace vertices with more than two legs do not come with a simple
$\delta_{n+m+\ell,0}$, but rather in a form such as \Eref{def_epsilon_nml} or \Eref{def_U}.
These more complicated expressions are present even at tree level,
\Eref{V2_tree_open}. This touches on an interesting feature,
namely that non-linearities are present even in
dimensions above the upper critical dimension --- they have to, as
otherwise the tree level lacks a mechanism by which immobile offspring
are deposited.
Below the upper critical dimension, the lack of momentum conservation
has three major consequences: Firstly, each vertex comes with a summation
and so a loop formed of two vertices, \Eref{one_loop_open}, requires not only one
summation ``around the loop'' but a second one accounting for another
index, which is no longer fixed by momentum conservation. This is a
technicality, but one that requires more and potentially serious computation. Secondly,
and more seriously, the very structure of the vertex might change. For
example, at bare level $\kappa$ comes with a factor $LU_{n,m+k,\ell}$,
but that $U_{n,m+k,\ell}$ might change under renormalisation.
Finally, the third and probably most challenging consequence is the loss
of momentum conservation in the propagator. While a lack of
translational invariance may not be a problem at bare level, the
presence of non-momentum conserving vertices can render the propagators
themselves non-momentum conserving --- provided the propagators
renormalise at all (see the discussion after \Eref{full_tau_open}), which they do not in the present case, as far as the
two shown in \Eref{bare_activity_propagator} are concerned. However,
$\tikz[baseline=-2.5pt]{
\draw[tactivity] (0.4,0) -- (0,0);
\draw[tsubstrate] (0,0) -- (-0.4,0);
}$
parameterised by $\tau$
has every right to be called a propagator and it does renormalise.
Luckily, however, it never features within loops, so the complications
arising from its new structure can be handled within observables and does
not spoil the renormalisation process itself.
A consequence of the Dirichlet boundary conditions is the existence of a
lowest, non-vanishing mode, $q_1=\pi/L$, \Eref{general_scaling_open}, which, in fact, turns out to
play the r{\^o}le of the effective mass --- just like the minimum of the
inverse propagator, $(-\imath\omega+D\gpvec{k}^2+r)$, the ``gap'', is $r$ in
the bulk, it is $Dq_1^2+r$ in the presence of Dirichlet boundary
conditions, and thus does not vanish even when $r=0$. This is a nice
narrative, which is challenged, however, when periodic boundary
conditions are applied. At tree
level, when the interaction is switched off, periodic boundaries cannot
be distinguished from an infinite system, and so we would evaluate
at tree level an infinite and a periodic system both at $\gpvec{k}=0$ and
$k_n=0$ respectively, producing exactly the same expectation (for exactly the right reason).
The situation is different beyond tree level. Periodic or open, the
system is finite. However,
periodic boundaries do not drain active particles, so the lowest wave
number vanishes, $k_n=0$. To control the infrared (in the infinite directions), a finite extinction rate $r$
is necessary, which effectively competes with the system size $L$ via
$\tilde{\rho}\propto L^2 r/D$, \Eref{one_loop_cyl_final_step} and
\Eref{one_loop_cyl_final_step_large_circ}. If $\tilde{\rho}$ is large, bulk
behaviour $\propto \tilde{\rho}^{-\epsilon/2}$ is recovered,
\Eref{one_loop_cyl_final_step_large_circ}, as is the case in the open system (see footnote \fnref{open_recovers_bulk} before \Eref{open_sum}). For
moderately small values, $\zeta(3-d)\propto 1/\epsilon$ dominates,
\Eref{cyl_sum_small}, a signature of a $d$-dimensional system with \emph{open}
boundaries, \Eref{one_loop_open_final_step}. In that case, scaling amplitudes are in fact $\propto
L^{\epsilon}$, \Eref{one_loop_cyl_final_step_simplified}. However, the
presence of the $0$-mode allows for a different asymptote as $\tilde{\rho}$
is lowered further, the bulk-like term governing the $d-1=\tilde{d}$
infinite dimensions takes over, $\propto L^{-1}((r+\epsilon'-\imath\omega)/D)^{-\tilde{\epsilon}/2}$.
It is the appearance of that term and \emph{only} that term
which distinguishes periodic from open boundary conditions.
So, the narrative of ``lowest wave number corresponds to mass'' is
essentially correct. In open systems, it dominates for all small masses.
In periodic systems, the scaling of the lowest non-zero mode competes with that of a $d-1$-dimensional bulk system due to the presence of a $0$-mode in the periodic dimension, which asymptotically drops out.
The \textbf{third point} that was to be addressed in the present work
were the special properties of a propagator of an immobile species. The
fact that the propagator is, apart from $\delta(\gpvec{k}+\gpvec{k}')$,
\Eref{substrate_propagator},
independent of the momentum is physically relevant as the particles
deposited stay where they have been deposited and so the walker has to truly
return to a previous spot in order to interact. Also, deposited particles are not
themselves subject to any boundary conditions --- this is the reason for
the ambiguity of the eigenfunctions that can be used for the fields of
the substrate particles. If deposited particles were to ``fall off'' the lattice,
the volume of the sausage on a finite lattice cannot be determined by
taking the $\omega\to0$ limit.
It is interesting to see what happens to the crucial integral
\Eref{one_loop}
when the immobile propagator is changed to $(-\imath\omega+\nu\gpvec{k}^2+\epsilon')^{-1}$:
\begin{multline}\elabel{one_loop_with_extras}
\kappa^2 \int \ddintbar{k'}\dintbar{\omega'}
\frac{1}{-\imath\omega'+\nu\gpvec{k}^2+\epsilon'}
\frac{1}{-\imath(\omega-\omega')+D(\gpvec{k}-\gpvec{k}')^2+r}\\
=
\frac{\kappa^2}{(4\pi)^{d/2}(D+\nu)}\left(\frac{r+\epsilon'-\imath\omega+\frac{D\nu}{D+\nu}\gpvec{k}^2}{D+\nu}\right)^{-\epsilon/2}
\Gamma\left(\frac{\epsilon}{2}\right) \ ,
\end{multline}
which at external momentum $\gpvec{k}=0$ is \Eref{one_loop} with $D$
replaced by $D+\nu$.
The integral thus remains essentially unchanged, just that the effective diffusion
constant is adjusted by $D\to D+\nu$.
A slightly bigger surprise is the fact that $\epsilon'$, the IR
regulator of the substrate propagator, is just as good an IR regulator
as $r$, the IR regulator of the activity propagator. The entire field
theory and thus all the physics discussed above, does not change when
the ``evaporation of walkers'' is replaced by ``evaporation of substrate
particles''. The stationarity of both in infinite systems is obviously
due to two completely different processes, which, however, have the same
effect on the moments of the Sausage Volume: If $r$ is finite, then
a walker eventually disappears, living behind the trace of substrate
particles, which stay indefinitely. If $\epsilon'$ is finite, then
stationarity is maintained as substrate particles disappear while new
ones are produced by an ever wandering walker.
Finally, the \textbf{fourth issue} to be highlighted in the present work
was that of observables which are spatial integrals of densities. These
observables have a number of interesting features. As far as space is
concerned, eigenfunctions with a $0$-mode immediately give access to
integrals over all space. However, open boundaries force us to perform a
summation (and an awkward looking one, too, say \Eref{V_open_details}).
\subsection{Future work}
Two interesting extensions of the present work deserve brief
highlighting. Firstly, the Wiener Sausage may be studied on networks:
Given a network or an ensemble thereof, how many distinct sites are
visited as a function of time. The key ingredient in the analysis is the
lattice Laplacian, which provides a mathematical tool to describe the
diffusive motion of the walker. The contributions $\gpvec{k}^2$ and $q_n^2$
in the denominator of the propagator, \Eref{bare_activity_propagator}
and \Eref{bare_activity_propagator_open},
are the squared eigenvalues of the Laplacian operator in the continuum
and, in fact, of the lattice Laplacian, for, say, a square
lattice. The integrals in $\gpvec{k}$-space and, equivalently, sums like
\Eref{convenient_choices} and \Eref{V_open_details} should be seen as
integrating over all eigenvalues $\gpvec{k}^2$, whose density in $d$
dimensions is proportional
to $|\gpvec{k}|^{d-1}$. It is \emph{that} $d$ which determines the scaling
in, say, $\ave{V} \propto t^{d/2}$ for $d<2$. In other words, if
$|\gpvec{k}|^{d_s-1}$
is the density of eigenvalues (the density of states) of the lattice Laplacian, then the Wiener
Sausage volume scales like $t^{d_s/2}$ (and the probability of return
like $t^{-d_s/2}$).
Provided the propagator does not acquire an anomalous dimension, which could depend on $d_s$ in
a complicated way, the difference between a field theory on a regular
lattice with dimension $d$ and one on a complicated graph with spectral dimension
$d_s$ is captured by replacing $d$ by $d_s$ \cite[p.~23]{Dorogovtsev:2010}.
We confirmed this finite size scaling of the Wiener Sausage on four different fractal lattices.
The second interesting extension is the addition of processes, such as
branching of the walkers itself. In that case they not only interact
with their past trace, but also with the trace of ancestors and successors.
This field theory is primarily dominated by the branching ratio,
say $s$,
and
$\lambda$,
whereas $\kappa$, $\chi$ and $\xi$ are irrelevant.
Preliminary results suggest that $d_c=4$ \cite[see
also]{vandenBergBolthausendenHollander:2004} in this case and again
$\ave{V}\propto L^{2-\epsilon}$, this time, however, with
$\epsilon=4-d$. Higher moments seem to follow
$\ave{V^m}\propto L^{(m-1)d+2-\epsilon}=L^{md-2}$. The latter result
suggests that the dimension of the cluster formed of sites visited is
that of the underlying lattice.
\begin{acknowledgement}
The authors would like to thank Andy Thomas and Niall Adams for computing support,
Hoai Nguyen Huynh for the numerical results on the Wiener Sausage on fractals and
Aman Pujara and
Songhongyang Yuan for computational support.
SN would like to thank Imperial College and in particular GP for their hospitality and the German National Academic Foundation for its generous support of SN's stay at Imperial College.
This work has been supported by the EPSRC Mathematics Platform grant EP/I019111/1.
\end{acknowledgement}
|
\section{Introduction}
With the rocketing development of the Internet, we are confronted with the problem of information overload \cite{1,2}. In order to break through this dilemma, various recommender algorithms \cite{3,4,5,6,7,8,9}, which attempts to predict users' interests by analyzing their historical activities, have been proposed. So far, the collaborative filtering (CF) algorithm \cite{10,11} has been one of the successful recommendation algorithms, which is designed based on the assumption that users with similar preferences will rate similar objects. When predicting the potential interests of a given user, the CF algorithm firstly identifies the neighborhood of each user by calculating similarities between all pairs of users, and then makes recommendations based on the neighbors' selections. It is well known that the most important ingredient in determining the performance of the CF algorithm is how to precisely define the similarities between each pair of users \cite{12,13}. Based on the user-object bipartite network, the cosine similarity \cite{14} is the most widely used index to quantify the proximity of users' tastes. In addition, Sarwal {\it et al.} \cite{15} proposed the item-based CF algorithm by comparing different items. Deshpande and Karypis \cite{16} proposed the item-based top-$N$ CF algorithm, in which items were ranked according to the frequency of appearing in the set of similar items and the top-$N$ ranked items were returned. Luo {\it et al.} \cite{17} introduced the concepts of local and global user similarity based on surprisal-based vector similarity
and the concepts of maximum distance in graph theory.
Recently, some physical dynamics, such as random walks \cite{6,7} and heat conduction \cite{18}, have found their applications in user or item similarity measurement to generate recommendation algorithms. Liu {\it et al.} \cite{7} embedded the random-walks process into the CF algorithm to calculate the user similarity and found that the random-walk-based CF algorithm had remarkable accuracy. By taking into account the second-order correlation of the objects and users, Zhou {\it et al} \cite{19} proposed improved CF algorithms by depressing the influence of mainstream preferences. The simulation results show that both accuracy and diversity of the improved CF algorithms could be enhanced greatly. By tuning the similarity from neighbors to the target user, Liu {\it et al.} \cite{20} analyzed the relationship between the similarity direction and the performance of the CF algorithm and found that emphasizing the small-degree users' recommendation powers could not only accurately identify users' interests, but also increase the algorithmic capability of finding diverse objects. Inspired by the idea that both second-order user correlations and the similarity direction affect the accuracy and diversity of the CF algorithms, we investigate the effect of the similarity direction on the second-order CF algorithm. Based on the statistical properties of the user correlation network and the similarity direction effect, we present a modified CF algorithm, namely the directed second-order CF (HDCF) algorithm. The experimental results on the data sets Movielens and Netflix show that only by changing the direction of the similarity obtained by the second-order user correlation, the accuracy greatly outperforms state-of-the-art CF methods, which suggests that the similarity direction is definitely a significant factor for second-order information filtering.
\begin{center}
\begin{figure}[t]
\center\scalebox{0.75}[0.75]{\includegraphics{figure2.eps}}
\caption{(Color Online) The diversity $S$ and the average object
degrees $\langle k\rangle$ vs. $\lambda$ of the HDCF, HCF and MHCF
algorithms for Movielens and Netflix data sets when recommendation
list equals to $L=10$. (a)-(b) exhibit the diversity $S$ vs.
$\lambda$, and (c)-(d) show the average object degrees $\langle
k\rangle$ vs. $\lambda$. At the optimal cases, both diversity $S$
and popularity $\langle k\rangle$ of HDCF are much better than the
ones of MHCF algorithms. All the data sets points are averaged over
ten independent runs with different data set divisions.}
\label{Fig2}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[h]
\center\scalebox{0.75}[0.75]{\includegraphics{figure3.eps}}
\caption{(Color Online) Precision $P$ and Recall $R$ of the HDCF,
HCF and MHCF algorithms for Movielens and Netflix data sets when the
recommendation list length $L=10$. (a)-(b) exhibit the precision $P$
vs. $\lambda$, and (c)-(d) report recall $R$ vs. $\lambda$. At the
optimal cases, both $P$ and $R$ of HDCF for Movelens and Netflix are
larger than the ones of MHCF algorithm. All the data points are
averaged over ten independent runs with different data set
divisions.} \label{Fig3}
\end{figure}
\end{center}
\section{Related works}
\subsection{Bipartite network and the standard CF algorithm}
An information filtering system could be characterized by a user-object bipartite network which consists of a set of user nodes denoted as $U=\{u_{1},u_{2},\ldots,u_{n}\}$, the object nodes as $O=\{o_{1},o_{2},\ldots,o_{m}\}$ and the links between these two sets, which are indicated by $E=\{e_{1},e_{2},\ldots,e_{p}\}$. The bipartite network containing $n$ users and $m$ objects can be represented by an adjacent matrix $A=\{a_{ij}\}\in R^{m,n}$, where $a_{ij}=1$ if user $u_{j}$ collects object $o_{i}$ and $a_{ij}=0$ otherwise.
Inspired by the random-walks process presented by Zhou {\it et al.} \cite{6}, Liu {\it et al.} \cite{7} proposed a CF algorithm based on the random-walks (denoted by CF). It is assumed that a certain amount of resource is associated with each user, and the weight $s_{ij}$
represents the proportion of the resource user $u_j$ would like to distribute to user $u_i$. The weight $s_{ij}$ representing the fraction of initial resource user $u_j$ ultimately gives to user $u_i$ can be defined as
\begin{equation}\label{equation1}
s_{ij}=\frac{1}{k_{u_j}}\sum_{l=1}^m\frac{a_{li}a_{lj}}{k_{o_l}},
\end{equation}
where $k_{u_j}=\sum\nolimits_{\alpha=1}^{m}a_{\alpha j}$ and $k_{o_l}=\sum\nolimits _{i=1}^{n}a_{li}$ denote the degrees of user $u_j$ and object $o_l$. For the user-object pair $(i,\alpha)$, if user $u_i$ has not collected object $\alpha$ yet (i.e., $a_{\alpha i}=0$), the predicted score $v_{i \alpha}$, is given as
\begin{equation}\label{equation2}
v_{i\alpha}=\frac{\sum_{j=1}^n s_{ji}a_{\alpha
j}}{\sum_{j=1}^ns_{ji}},
\end{equation}
where $s_{ji}$ indicates the similarity from the target user $u_i$ to its neighbor $u_j$. To the target user $u_i$, when the predicted scores $v_{i\alpha}$ among all the objects he/her has not collected are calculated, all these scores will be sorted in descending order, and finally those on the top will be recommended.
\subsection{Second-order user correlation effect analysis}
\begin{figure*}
\begin{equation}
\begin{array}{ccl}
h_{ij} & = & s_{ij} + \lambda \sum_u{s_{iu}s_{uj}},\\[5pt]
& = & \frac{1}{k_{j}}\sum_{o=1}^m\frac{a_{o i}a_{o j}}{k_{o}}
+ \lambda
\sum_{u=1}^{n}\Big(\frac{1}{k_{u}}\sum_{o=1}^m\frac{a_{oi}a_{ou}}{k_{o}}\Big)\Big(\frac{1}{k_{j}}\sum_{o=1}^m\frac{a_{ou}a_{oj}}{k_{o}}\Big)\\
& = & \frac{1}{k_{j}}\Big\{\sum_{o=1}^m\frac{a_{o i}a_{o j}}{k_{o}}
+ \lambda
\sum_{u=1}^{n}\Big(\frac{1}{k_{u}}\sum_{o_1=1}^m\sum_{o_2=1}^m\frac{a_{o_1i}a_{o_1u}}{k_{o_1}}\frac{a_{o_2u}a_{o_2j}}{k_{o_2}}\Big)\Big\}.\\
\end{array}
\end{equation}
\end{figure*}
The correlation between two users is a reflection of their similar tastes or preferences, therefore, for two arbitrary users, the specific interests should contribute more to the similarity measurement than the mainstream preferences \cite{add1}. In addition, two users sharing many mainstream preferences have high second-order similarity. Liu {\it et al.} \cite{8} proposed an effective method to depress the influence of mainstream preferences by considering the second-order similarity, where the similarity matrix is given by
\begin{equation}\label{equation8}
{\bf H}={\bf S}+\lambda {\bf S}^{2},
\end{equation}
and ${\bf H}=\{h_{ij}\}_{n,n}$ is the user similarity matrix obtained by taking into account the second-order user correlation. Here the value range of the parameter $\lambda$ is $(-1,0)$ in order to improve the algorithmic accuracy.
\subsection{The directed second-order CF algorithm}
\revision{To users $u_i$ and $u_j$, the similarity from user $u_j$ to user $u_i$, $h_{ij}$ representing the amount of initial resource $u_j$ evenly transferred to $u_i$ could be written as Eq.3. It is unlikely these quantities are exactly the same for each pair of users, therefore, $h_{ij}\neq h_{ji}$ in most cases. In addition, one has the following relationship
\begin{equation}
\frac{h_{ij}}{h_{ji}} = \frac{k_i}{k_j}.
\end{equation}
If $k_i>k_j$, then $h_{ij}>h_{ji}$ and vice versa. In other words, the similarities from small-degree users to large-degree users would be larger than the ones from the opposite direction. Since that the degree of most users in the real world is very small, which means that the large-degree users would frequently be identified as small-degree users' friends, the second-order CF algorithm would emphasize the large-degree users' recommendation powers, leading to the high similarity of most users' recommendation lists.
Furthermore, the second-order CF algorithm using the similarity from the target user to his neighbors, does not match up to the core idea of the standard CF algorithm. Therefore we could enhance the small-degree users' effects by reversing the user similarity direction from neighbors to the target user} and present the directed second-order CF (denoted by HDCF) algorithm, which could be described as follows
\begin{description}
\item[(i)] To user $i$, calculating the user similarity from all other users $\{h_{ij}\} (j=1,2,\cdots,n)$ in terms of the Eq.4;
\item[(ii)]
For each user $i$, calculating the predicted scores for the
uncollected objects,
\begin{equation}\label{equation9}
v_{i\alpha}=\frac{\sum_{j=1}^n h_{ij}a_{\alpha
j}}{\sum_{j=1}^n h_{ij}},
\end{equation}
\item[(iii)] Sorting the uncollected objects in the descending order of the predicted scores, and those objects on the top will be recommended.
\end{description}
\subsection{The maximum-similarity-based CF algorithm}
Actually the properties of the real datasets could also affect the algorithmic performance. In other words, although the performance of the HDCF algorithm outperforms the state-of-the-art CF algorithms, it may only happen in some certain datasets whose second-order similarities from neighbors to the target user are more effective than the ones in the reverse direction. Thus, we present a maximum-similarity-based CF (denoted by MHCF) algorithm to investigate the effect of second-order similarity magnitude, and the predicted score $v_{i\alpha}$ is given by
\begin{equation}\label{equation10}
v_{i\alpha}^{m}=\frac{\sum_{j=1}^n h^{\rm max}_{ij}a_{\alpha
j}}{\sum_{j=1}^nh^{\rm max}_{ij}},
\end{equation}
where $h^{\rm max}_{ij}$ is defined as the larger second-order similarity between users $u_i$ and $u_j$
\begin{equation}\label{equation11}
h^{\rm max}_{ij}=\max\{h_{ij}, h_{ji}\}.
\end{equation}
\section{Experimental results}
\subsection{Data description}
In this paper, we implement our experiments on two benchmark data sets, {\it Movielens} \footnote{http://www.Movielens.com} and {\it Netflix} \footnote{http://http.netflix.com}. Movielens consists of 6040 users and 3952 movies (objects). Netflix is a random sample of the original Netflix data set, which consists of 6000 movies and 10000 users, and 824802 ratings. The users of the data sets vote on movies with discrete ratings from 1 (i.e., worst) to 5 (i.e., best). Here we apply a coarse graining method \cite{6}: A movie is considered to be collected by a user only if the given rating is three or more. In this way, the Movielens data have 836478 edges, and the Netflix data have 701947 edges. Table 1 gives the basic statistical properties of the data sets. In order to evaluate the algorithmic performance, we randomly divide data set $E$ into two parts: $E=E^{T} \cup E^{P}$, where $E^T$ is the training set treated as known information, and the other one $E^P$ is the probe set, whose information is not allowed to be used for prediction. We treat 90\% percent of the ratings as the training set, and the remaining 10\% part compose the probe set. Then five different metrics are employed to test the algorithm performance, including average ranking score, popularity, diversity, precision and recall.
\begin{table}
\footnotesize
\caption{Basic statistical properties of the tested data sets.}
\begin{center}
\begin{tabular} {ccccc}
\hline \hline
Data Sets & Users & Objects & Links & Sparsity \\ \hline
MovieLens & 6040 & 3952 & 836,478 & $3.50\times 10^{-2}$\\
Netflix & 10,000 & 6,000 & 701,947 & $1.17\times 10^{-2}$ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\subsection{Performance metrics}
1) {\bf Average ranking score} \cite{22}
Since the average ranking score doesn't rely on the length of recommendation list, we use it to measure the ability of the algorithm to produce a good uncollected object ranking list that matches the target user's preference. For an arbitrary user $u_i$, if in the training set the object $o_{\alpha}$ is not collected by user $u_i$, while the entry $(i,\alpha)$ is in the probe set, we use the rank of the object $o_\alpha$ in the recommendation list to evaluate accuracy. Therefore, the mean value of the positions, called {\it average ranking score} $\langle r \rangle$, averaged over all the entries in the probe set, can be used to evaluate the algorithmic accuracy
\begin{equation}\label{equation12}
\langle r\rangle=\frac{1}{n}\sum_{i=1}^n\Big(\frac{\sum_{(u_i,
o_\alpha)\in E^p}r_{i\alpha}}{q-k_{u_i}}\Big),
\end{equation}
where $E^{p}$ is the edge set existing in the probe set and $q$ is the number of objects in the probe set. The smaller the average ranking score, the higher the algorithmic accuracy, and vice versa.
2) {\bf Diversity} \cite{23}
Personalized recommendation algorithms should not only present accurate prediction, but also generate different recommendations to different users according to their specific tastes or habits. Therefore, besides accuracy, the {\it diversity} measured by the mean value of {\it Hamming distance} $S$, is taken into account to evaluate the strength of personalization. If the overlapped number of objects in $u_i$ and $u_j$ recommendation list $L$ is $Q_{ij}$, their Hamming
distance could be quantified as:
\begin{equation}\label{equation13}
S= 1-\langle Q_{ij}(L)\rangle/L,
\end{equation}
Generally speaking, a more personalized recommendation list should have larger Hamming distance to other lists. Accordingly, we adopt the mean value of Hamming distance $S=\langle H_{ij} \rangle$, averaged over all the user pairs, to measure the strength of the algorithmic diversity. The largest $S=1$ indicates recommendation to all users are completely different. While the smallest $S=0$ means all of the recommendations are exactly the same.
3) {\bf Popularity} \cite{22}
An accurate and diverse recommender system is expected to help users find the niche or unpopular objects that are hard for them to identify. Since there are countless channels which advertise the popular movies, such as the Internet, TV, newspaper, etc., uncovering the very specific preference, corresponding to unpopular objects, is much more significant than simply picking out what a users likes from the top of the list. The metric {\it popularity} is introduced
to quantify the ability of an algorithm to generate the unexpected recommendation lists, which is defined as the average collected times over all recommended objects:
\begin{equation}\label{equation14}
\langle k\rangle =
\frac{1}{n}\sum_i\Big(\frac{1}{L}\sum_{o_\alpha\in
O^L_i}k_{o_\alpha}\Big),
\end{equation}
where $o_{i}^{L}$ is user $i$'s recommendation list with length $L$. A smaller average degree $\langle k\rangle$, corresponding to less popular objects, is preferred since those small-degree objects are hard to be found by users themselves.
\begin{table*}
\scriptsize
\caption{Algorithmic performances for Movielens and Netflix
data sets when $p=0.9$, including the average ranking score
$\langle r\rangle$, and diversity $S$, popularity $\langle k\rangle$, precision $P$ and
recall $R$
corresponding to the length of recommendation list $L=10$.
CF is the collaborative filtering algorithm based on
random-walks proposed in the Ref.7; DCF is the directed random-walks-based
CF algorithm whose similarity measurement is from neighbors
to the target user ($\beta_{\rm opt}=3.2$ for Movielens and
$\beta_{\rm opt}=2.0$ for Netlix) \cite{20}; HCF is an improved
CF algorithm, in which the user similarity is based
on the random-walks, and the second-order correlation is
involved ($\lambda_{\rm opt}=-0.85$ for Movielens and
$\lambda_{\rm opt}=-0.8$ for Netflix) \cite{25}; MHCF is the high-order
CF algorithm whose similarity is defined as the larger one between
two users ($\lambda_{\rm opt}=-0.95$ for Movielens and
$\lambda_{\rm opt}=-0.9$ for Netflix); HDCF is
the presented new algorithm in this paper. Each number is
obtained by a averaging over ten runs of independently
random division of training set and probe set.}
\begin{center}
\begin{tabular}{ccc|cccccccccccc}
\hline\hline
& & Algorithms && $\langle r\rangle$ & & $\langle k\rangle$ & & $S$ & & $P$ & & $R$ &\\ \hline
& & CF & & 0.1055 & & 1818 & & 0.7068 && 0.1025 && 0.1031&\\
& & DCF & & 0.0853 & & 1259 & & 0.9191 && 0.1053 && 0.1165&\\
&{\it Movielens} & HCF & & 0.0828 & & 1397 & & 0.9112 && 0.1436 && 0.1607&\\
& & MHCF & & 0.0791 & & 1623 & & 0.8424 && 0.1388 && 0.1456&\\
& & HDCF & &{\bf 0.0767}& & 1545 & & 0.8707 && 0.1422 & & 0.1602&\\
\hline
& & CF & & 0.050 & & 2813 & & 0.7001 && 0.0917 && 0.1365&\\
& & DCF & & 0.045 & & 2506 & & 0.8236 && 0.0967 && 0.1640&\\
&{\it Netflix} & HCF & & 0.0434 & & 2531 & & 0.8535 && 0.1269 && 0.2083&\\
& & MHCF & &{\bf 0.0402}&& 2814 && 0.7737 && 0.1210 && 0.1915&\\
& & HDCF & & {\bf 0.0402}&& 2731 && 0.7997 && 0.1211 && 0.1998&
\\ \hline\hline
\end{tabular}
\end{center}
\end{table*}
4) {\bf Precision and Recall} \cite{ad23,24}
Since real users are usually cared only about the top part of the recommendation list, a more practical method is to consider the number of a user's deleted links contained in the top-$L$ positions. Based on our concerns, we may take into account either how many of these top $L$ places are possessed by deleted links, or how many of user's deleted links have been recovered in this way. Thus, precision and recall can satisfy our requests. For an arbitrary user $u_i$, the precision and recall of the recommendation, $P_i(L)$ and $R_i(L)$, are defined as
\begin{equation}\label{equation15}
P_i(L)=\frac{d_i(L)}{L},\ \ R_i(L)=\frac{d_i(L)}{D_i}
\end{equation}
where $d_i(L)$ indicates the relevant objects in the top-$L$ positions of the recommendation list that collected by user $u_i$ in the probe set, and $D_i$ is the total number of user $u_i$'s relevant objects. Averaging the individual precision and recall over all users, we obtain the mean value {\it Precision} $P(L)$ and {\it Recall} $R(L)$ of the algorithm on one data set
\begin{equation}\label{equation16}
P(L)= \frac{1}{n}\sum_i\frac{d_i(L)}{L},\ \ R(L)=
\frac{1}{n}\sum_i\frac{d_i(L)}{D_i}.
\end{equation}
A larger precision corresponds to a better performance, and the larger recall corresponds to the better performance.
\subsection{Performance}
Table 2 shows the comparisons of five metrics among different CF algorithms at cutoff 10. The CF algorithm based on random-walks proposed by Liu {\it et al.} \cite{7} is denoted by CF, DCF represents the directed random-walks-based CF algorithm \cite{20}, HCF is the second-order-based CF algorithm \cite{25}, MHCF stands for the high-order-based CF algorithm whose similarity is defined as the larger one and the presented new algorithm in this paper considering the similarity direction and the second-order user correlation is called HDCF. Comparing with the HCF algorithm, the average ranking score $\langle r\rangle$ of the HDCF algorithm is reduced from 0.0828 to 0.0767 for Movielens and from 0.0434 to 0.0402 for Netflix, at the optimal values of $\lambda$. Apparently, by only taking into account the similarity direction of the second-order correlation without any other information, the accuracy of our algorithm outperforms the standard second-order CF algorithm, and even better than state-of-the-art CF algorithm. Clearly, enhancing the small-degree users' recommendation powers by changing the similarity direction could provide fairly accurate recommendation results.
Fig.1 represents the diversity $S$ and the popularity $\langle k\rangle$ as a function of $\lambda$ when the recommendation list $L=10$ for Movielens and Netflix, respectively. From Fig.1(a)-(b), one could find that, for different data sets, the diversity $S$ is negatively correlated with $\lambda$, which indicates that involving the directed second-order similarity makes the recommendation lists more diverse. Compared with the results of the standard CF algorithm, when recommendation list length $L=10$, for the Movielens data set, the diversity $S=0.8707$ at the optimal $\lambda_{\rm opt}=-0.8$, and for the Netflix, the diversity $S=0.7997$ at the optimal $\lambda_{\rm opt}=-0.7$, which is improved by 23.2\% and 14.2\% respectively. Fig.1(c)-(d) exhibit a positive relation between popularity $\langle k\rangle$ and $\lambda$, thus to depress the influence of mainstream preferences, we should give more opportunity to the less popular objects. Comparing with the results of standard CF algorithm corresponding, when $L=10$, the popularity $\langle k\rangle=1545$ for Movielens and $\langle k\rangle=2731$ for Netflix at their optimal values, which is reduced by 15\%, and by 2.9\% respectively. So the presented algorithm with negative $\lambda$ has the capability to provide more diverse recommendation lists and excavate unpopular objects.
Since users always pay more attention to the top of the recommendation lists, from Fig.2 we could see that the precision $P$ and recall $R$ are also very good. Comparing with the standard CF algorithm, when $L=10$ with the optimal parameter corresponding to the lowest ranking score ,the $P$ is roughly improved by 38.7\% and 32\%, and $R$ is approximately enhanced by 55.4\% and 46\% for Movielens and Netflix respectively.
In general, the specific common interests contribute more to the similarity measurement between two users than their mainstream preferences, and the directed similarity calculated by the random-walks process is reverse to the initial users' degrees. Thus, the standard CF algorithm may count repeatedly the attributes
of the popular objects which meet the tastes of most users and would assign more power for the larger-degree users to the small-degree users, which decreases the accuracy, increases the average object degrees and reduces the diversity. Our presented algorithm with negative $\lambda$ parameter and new user similarity direction could depress the influence of mainstream preferences and enhance the small-degree users' recommendation powers, which increases the
prediction accuracy greatly, as well as gives high chances to less popular objects and help users find diverse objects, leading to better algorithm performance.
\section{Conclusions and discussions}
The CF algorithms are one of the most successful information filtering algorithms and have been extensively implemented to many online applications. In this paper, by considering the second-order user correlations and the similarity direction, we present a directed second-order CF algorithm by tuning the user similarity from neighbors to the target user to emphasize the recommendation powers of small-degree users. The experimental results for MovieLens and Netflix data sets show that the new algorithm could indeed generate a more favorable recommendation performance. Compared with the CF algorithm based on random-walks, the average ranking score could be improved to 0.0767 and 0.0402, which is enhanced by 27.3\% and 19.1\% for MovieLens and Netflix respectively. Additionally, the diversity, precision and recall are also enhanced greatly. The possible reasons for the pretty good performance of HDCF algorithm lie in the fact that second-order similarity could depress the influence of mainstream preferences on the target user, and tuning the similarity direction of the CF algorithm from the neighbors to the target user could enhance small-degree users' recommendation power to improve the accuracy and help users uncover less popular objects.
Since we only need to change the similarity direction of the tradition second-order CF algorithm without any other information, it is very easy for this algorithm to be used in real information filtering systems. How to better provide personalized recommendations by taking into account the mixing pattern \cite{26} of the network for diverse users is a long-standing challenge in modern information science. Any method to this issue may intensively change our society, economic and life in the near future. We believe our current work will enlighten readers in this interesting and promising direction.
\section{Acknowledgments}
We acknowledge {\it GroupLens} Research Group for providing us {\it MovieLens} data and the Netflix Inc. for {\it Netflix} data. This work is partially supported by NSFC (61374177, 71371125 and 71171136), MOE Project of Humanities and Social Science (13YJA630023) and Shanghai First-class Academic Discipline Project (S1201YLXK).
|
\section{Introduction}\label{sec_intro}
Circumstellar discs are ubiquitous in the star formation process. The disc
fraction around young stars approaches 100\% in star-forming regions with
ages less than 1 Myr, and then steadily declines to 10\% or less for regions
10 Myr or older \citep[e.g.,][]{wyatt2008:disks}. Disc formation is a
natural consequence of the conservation of angular momentum in a collapsing
system and has long been expected to form early in the protostellar stage.
We note that, in this paper, the protostellar stage refers to the
evolutionary stage when protostars (and their discs if they have already
formed) are still embedded in and accreting from their parent cores
\citep[in other words, the Class 0 and I evolutionary stages; see, e.g., ][for a recent review on evolutionary stages]{dunham2014:ppvi}. Such
early disc formation has been
confirmed by \citet{terebey1984:model}, who modified the
\citet{shu1977:sis} solution for the inside-out collapse of a singular
isothermal sphere to include rotation. The formation of relatively large,
massive, Keplerian discs very early in the protostellar stage is also found by
non-magnetic simulations of collapsing cores
\citep[e.g.,][]{vorobyov2011:disks}.
However, simulations of magnetized collapsing cores have found that efficient
magnetic braking removes angular momentum from the system and may completely
suppress disc formation in the protostellar stage
\citep[e.g.,][]{allen2003:magneticbraking,hennebelle2008:magneticbraking,mellon2008:magneticbraking,seifried2011:magneticbraking}. Many potential solutions
have been proposed to reduce the effects of magnetic braking and allow discs
to form, including magnetic flux loss through various mechanisms,
non-ideal MHD effects, outflow-induced envelope clearing, turbulence, and
misalignment between the magnetic field and rotation axis
\citep[e.g.,][]{mellon2008:magneticbraking,machida2011:magneticbraking,seifried2012:magneticbraking,li2013:misalignment,joos2012:misalignment,krumholz2013:misalignment,machida2014:disks}.
The latter has recently been tested by two observational studies which reached
conflicting results on the alignment (or lack thereof) between magnetic
fields and rotation axes \citep{hull2013:misalignment,chapman2013:misalignment}.
While discs must form at some point, how early they form and how quickly they
grow in mass and size remain key open questions. Observations of discs
surrounding embedded protostars are difficult because the discs are deeply
embedded in their parent cores, and are especially challenging in the Class
0 stage when most of the system mass is still in the core. One of the earliest
studies of a Class 0 protostellar disc was presented by
\citet{harvey2003:b335}, who modeled interferometer millimeter continuum
observations of B335 and identified the presence of a compact component
with a mass of 0.004 \msun, which they inferred to be a disc.
Similar inferences of Class 0 discs through analysis and modeling of compact
components in interferometer (sub)millimeter continuum observations have been
presented by other authors
\citep[e.g.,][]{looney2003:disks,chiang2012:l1157,zapata2013:iras16293},
although the results are model-dependent and the samples are too small to
draw any statistical conclusions.
Two recent surveys have used larger samples (10 -- 20 protostars) to argue
that discs form very early in the Class 0 stage
\citep{jorgensen2009:prosac,enoch2011:disks}, although both studies reach such
sample sizes by using very
simple methods for separating disc and core emission. On the other hand,
\citet{maury2010:pdbi} found no evidence for large discs with sizes greater
than $\sim$100 AU in an interferometer millimeter continuum survey of five
Class 0 protostars and argued that their data were inconsistent with
simulations that neglected magnetic braking. However, they did not directly
address the nature of their continuum detections, in particular whether they
arise from discs or inner, dense cores. Only three Class 0
systems have confirmed detections of discs through direct detection of
Keplerian velocity profiles in millimeter spectral line observations: L1527
\citep{tobin2012:l1527,tobin2013:l1527}, VLA 1623A
\citep{murillo2013:vla1623,murillo2013:vla1623a},
and R CrA IRS7B \citep{lindberg2014:rcra}.
Taken together, these studies emphasize that the properties and even the
existence of embedded protostellar discs remains quite uncertain, especially
at the youngest (Class 0) stages of protostellar evolution.
The formation time and early evolution of circumstellar discs has several
consequences for both star and planet formation. If discs do indeed form early
in the embedded stage, mass from the infalling core
can pile up in them until they become
gravitationally unstable and fragment. These fragments can be driven onto
the star through torques associated with spiral arms in the discs, causing
short accretion bursts and a general cycle of episodic protostellar mass
accretion
\citep[e.g.,][]{vorobyov2005:bursts,vorobyov2006:bursts,vorobyov2010:bursts}.
Such bursts can have significant effects on the chemistry of the surrounding
core and planet-forming disc
\citep[][]{lee2007:chemistry,kim2011:cb130,kim2012:co2,visser2012:chemistry,vorobyov2013:chemistry}.
Additionally, the large changes in accretion rates lead to large changes in
accretion luminosity, and the episodic mass accretion process driven by disc
fragmentation predicted by \citet{vorobyov2010:bursts} provides a viable
solution to the protostellar luminosity problem
\citep{dunham2010:evolmodels,dunham2012:evolmodels}, whereby
protostars are underluminous compared to simple theoretical expectations
\citep{kenyon1990:luminosities,kenyon1994:luminosities,kenyon1995:luminosities}.
Moreover, fragments can be ejected from discs into the intracluster
medium, producing a population of proto-brown dwarfs and very-low-mass protostars \citep{BV2012},
or settle onto stable orbits, giving rise to wide-separation giant planet and brown dwarf
companions \citep{VB2010gp,Vorob2013}.
Finally, solid cores forming in the fragment interiors may provide an alternative scenario
for the icy and terrestrial planet formation \citep{Nayakshin2010}.
Clearly, the significance of all of these effects due to disc fragmentation
will depend on when exactly, if at all,
discs with sufficient mass to fragment actually
form. While observational constraints are currently quite limited, as
discussed above, the largest surveys to date of protostellar discs
\citep{jorgensen2009:prosac,enoch2011:disks} depend
on very simple assumptions to separate disc and core emission and derive
disc masses. If these methods successfully recover disc masses they offer a
promising method of assessing the formation and early evolution of protostellar
discs for statistically significant samples.
In this paper we use synthetic observations of simulated
protostellar core+disc systems to evaluate whether or not these assumptions
successfully recover disc masses, and, based on those results, assess
whether protostellar discs are sufficiently massive to fragment given the
current observational constraints. The organization of this paper is as
follows: We describe the coupling between the hydrodynamic simulations and
radiative transfer models used to generate the synthetic observations in
\S \ref{sec_model}, we evaluate the extent to which simple assumptions can
be adopted to recover intrinsic disc masses from our synthetic
interferometric observations in \S \ref{sec_results}, we discuss the
implications and limitations of our results in \S \ref{sec_discussion} and
\S \ref{sec_limit}, respectively, and we summarize our results in
\S \ref{sec_summary}.
\section{Description of the Model}\label{sec_model}
To assess whether or not recent interferometer surveys of protostars are
capable of successfully recovering the intrinsic masses of protostellar discs,
we use models that are based on a coupling of the two-dimensional,
numerical hydrodynamical simulations of collapsing cores presented by
\citet{vorobyov2010:bursts} with the two-dimensional, evolutionary radiative
transfer models of collapsing cores presented by
\citet{dunham2010:evolmodels}. To briefly summarize, the
\citet{vorobyov2010:bursts} hydrodynamical simulations follow for several Myr
the collapse of a non-magnetized
cloud core with a fixed initial mass and initial angular
momentum from the starless phase, through the embedded phase, and into the T
Tauri phase, with a self-consistent calculation of the instantaneous accretion
rates onto the disc and protostar and thus the masses of the core, disc, and
protostar as a function of time. The \citet{dunham2010:evolmodels}
evolutionary models use the two-dimensional, axisymmetric, Monte Carlo
dust radiative transfer code RADMC
\citep{dullemond2000:radmc,dullemond2004:radmc}
to calculate the dust temperature profiles, including both internal heating
from the protostar and external heating from the interstellar radiation field.
Accompanying routines are then used to generate spectral energy distributions
(SEDs) and images through ray-tracing and to calculate the complex
interferometer visibilities. We adopt the dust opacities of
\citet{ossenkopf1994:oh5}
appropriate for thin ice mantles after $10^5$ yr of coagulation at a gas
density of $10^6$ cm$^{-3}$
(OH5 dust; see \S \ref{sec_results_opacity} for a discussion
of the effects of grain growth in discs on the opacity law and resulting
disc mass calculations), and include
isotropic scattering off dust grains. SEDs and complex interferometer
visibilities at each timestep are calculated for nine different inclinations
ranging from $i=5-85$\degree\ in steps of 10\degree,
where $i=0$\degree\ corresponds to a pole-on (face-on) system and
$i=90$\degree\ corresponds to an edge-on system.
\input{tab1.tex}
\subsection{Hydrodynamical Simulations}\label{sec_simulations}
Our numerical hydrodynamics model for the formation and evolution
of a young stellar object is described in detail
in \citet{vorobyov2010:bursts} and \citet{vorobyov2013:chemistry}.
Here, we briefly review the main concepts. The model
includes a protostar, described by a stellar evolution
code, and a protostellar disc plus infalling envelope, both
described by a numerical hydrodynamics code. Both codes are coupled in real time,
but due to computational constraints the stellar evolution code is invoked only
every 25~yr to update the properties of the forming star.
We use the thin-disc approximation complemented by a calculation
of the vertical scale height in both the disc and envelope
determined in each computational cell using an assumption
of local hydrostatic equilibrium. The resulting model has a flared
structure with the vertical scale height increasing with radial distance.
Both the disc and envelope receive a fraction of the irradiation energy from
the central protostar.
The main physical processes
taken into account when computing the evolution of the
disc and envelope include viscous and shock heating, irradiation
by the forming star, background irradiation, radiative cooling
from the disc surface, and self-gravity. The corresponding
equations of mass, momentum, and energy transport are
\begin{equation}
\label{cont}
\frac{{\partial \Sigma }}{{\partial t}} = - \nabla_p \cdot
\left( \Sigma \bl{v}_p \right),
\end{equation}
\begin{eqnarray}
\label{mom}
\frac{\partial}{\partial t} \left( \Sigma \bl{v}_p \right) &+& \left[ \nabla \cdot \left( \Sigma \bl{v_p}
\otimes \bl{v}_p \right) \right]_p = - \nabla_p {\cal P} + \Sigma \, \bl{g}_p + \\ \nonumber
& + & (\nabla \cdot \mathbf{\Pi})_p,
\label{energ}
\end{eqnarray}
\begin{equation}
\frac{\partial e}{\partial t} +\nabla_p \cdot \left( e \bl{v}_p \right) = -{\cal P}
(\nabla_p \cdot \bl{v}_{p}) -\Lambda +\Gamma +
\left(\nabla \bl{v}\right)_{pp^\prime}:\Pi_{pp^\prime},
\end{equation}
where subscripts $p$ and $p^\prime$ refers to the planar components $(r,\phi)$
in polar coordinates, $\Sigma$ is the mass surface density, $e$ is the internal energy per
surface area,
${\cal P}$ is the vertically integrated gas pressure calculated via the ideal equation of state
as ${\cal P}=(\gamma-1) e$ with $\gamma=7/5$,
$\bl{v}_{p}=v_r \hat{\bl r}+ v_\phi \hat{\bl \phi}$ is the velocity in the
disc plane, and $\nabla_p=\hat{\bl r} \partial / \partial r + \hat{\bl \phi} r^{-1}
\partial / \partial \phi $ is the gradient along the planar coordinates of the disc.
The gravitational acceleration in the disc plane, $\bl{g}_{p}=g_r \hat{\bl r} +g_\phi \hat{\bl \phi}$, takes into account self-gravity of the disc, found by solving for the Poisson integral,
and the gravity of the central protostar.
Turbulent viscosity is taken into account via the viscous stress tensor
$\mathbf{\Pi}$. We parameterize the magnitude of kinematic viscosity $\nu$ using
the $\alpha$-prescription with a spatially and temporally uniform $\alpha=5\times 10^{-3}$.
The radiative cooling $\Lambda$ in equation~(\ref{energ}) is determined using the diffusion
approximation of the vertical radiation transport in a one-zone model of the vertical disc
structure \citep{Johnson2003}, while the radiative heating $\Gamma$
is calculated using the irradiation temperature at the disc surface determined
by the stellar and background black-body irradiation. For more details the reader
is referred to \citet{vorobyov2010:bursts} and \citet{vorobyov2013:chemistry}.
For the initial gas surface density $\Sigma$ and angular velocity $\Omega$ distributions
we take those typical of pre-stellar cores formed as a result of the slow expulsion
of magnetic field due to ambipolar diffusion \citep{Basu1997}
\begin{equation}
\Sigma={r_0 \Sigma_0 \over \sqrt{r^2+r_0^2}}\:,
\label{dens}
\end{equation}
\begin{equation}
\Omega=2\Omega_0 \left( {r_0\over r}\right)^2 \left[\sqrt{1+\left({r\over r_0}\right)^2
} -1\right].
\label{omega}
\end{equation}
Here, $\Omega_0$ and $\Sigma_0$ are the angular velocity and gas surface
density at the center of the core and $r_0 =\sqrt{A} c_{\rm s}^2/\pi G \Sigma_0 $
is the radius of the central plateau, where $c_{\rm s}$ is the initial sound speed in the core.
The gas surface density distribution described by equation~(\ref{dens}) can
be obtained (to within a factor of unity) by integrating the
three-dimensional gas density distribution characteristic of
Bonnor-Ebert spheres with a positive density-perturbation amplitude A \citep{Dapp2009}.
The value of $A$ is set to 1.2 and the initial gas temperature is set to 10~K.
\begin{figure*}
\centering
\includegraphics[width=16cm]{f1.eps}
\caption{{\it Left Panels:} Time evolution of the core, star, and disc masses
for models 1 (top) and 2 (bottom). {\it Middle Panels:} Time evolution of the
protostellar mass accretion rates for models 1 (top) and 2 (bottom).
{\it Right Panels:} Time evolution of the disc radius for models 1 (top) and
2 (bottom). In each
panel the thick gray lines mark, from left to right, the Class 0, Class 0/I,
and Class I timesteps considered in this paper (see \S \ref{sec_results} for
details).}
\label{fig_models}
\end{figure*}
In this paper two different models with different initial conditions
are coupled with the radiative transfer models, as described below in
\S \ref{sec_radtrans}. Table \ref{tab_models} lists the initial conditions
for each model, including the initial central angular velocity ($\Omega_0$),
the initial outer radius of the core ($r_0$), the initial central surface
density ($\Sigma_0$), the radius of the central region with an initially flat
surface density profile ($r_0$), the initial core
mass, and the initial ratio of rotational to gravitational energy ($\beta$).
Figure \ref{fig_models} plots the time evolution of the core, star, and
disc masses, protostellar mass accretion rates, and disc radii for each model.
These two models are chosen to represent opposite extremes: model 1 has a
near-solar mass core that forms a sub-solar mass star, whereas model 2
initially has a much lower core mass, giving birth to a very-low mass star.
As a consequence, model 1 forms a much larger, more massive disc than model 2
\citep[see Figure \ref{fig_models} and][for details]{vorobyov2010:bursts}.
Nevertheless, both discs are gravitationally unstable and prone to
fragmentation, owing to rather high initial values of $\beta$, and also reveal
highly variable accretion rates caused by disc gravitational instability.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f2.eps}}
\caption{Gas surface density images (in~log~g~cm$^{-2}$)
in model~1 (left column) and model~2 (right column) in the
Class 0, Class 0/I, and Class I timesteps considered in this paper (see
\S~\ref{sec_results} for details). The time in parentheses is counted from the
formation of the first hydrostatic core. The yellow lines outline the discs
according to the adopted critical surface density of
$\Sigma_{\rm cr}=0.5$~g~cm$^{-2}$. The red line in the lower left panel
outlines the disc assuming a lower $\Sigma_{\rm cr}$ of 0.1~g~cm$^{-2}$ (see
text for details). Fragments forming in the disc in the Class 0/I and I
stages are clearly visible.}
\label{fig_image}
\end{figure}
Figure~\ref{fig_image} shows the gas surface density distribution
in model~1 (left column) and model~2 (right column) in the
Class 0, Class 0/I, and Class I timesteps considered in this paper (see
\S \ref{sec_results} for details). For both models, only the inner several
hundred AU are shown; the full computational area is approximately 10 times
larger. The yellow lines outline the disc in each stage as identified using
the algorithm described in \citet{dunham2012:evolmodels}, which is based on a
critical surface density for the disc-to-envelope transition of
$\Sigma_{\rm cr}=0.5$~g~cm$^{-2}$. We found that this algorithm works well in
the Class 0/I stage, but may somewhat overestimate the disc mass in the Class
0 stage (when the inner part of the infalling envelope may have densities
exceeding $\Sigma_{\rm cr}$) and somewhat underestimate the disc mass in the
Class I stage (when the disc spreads out and its density drops below
$\Sigma_{\rm cr}$ in the outer parts). For instance, a lower value of
$\Sigma_{\rm cr}=0.1$~g~cm$^{-2}$ would result in larger disk sizes, as
illustrated by the red curve in the bottom-left panel of
Figure~\ref{fig_image}. However, the net increase in the disc mass is only
$4\%$, owing to a rather low density at the disc periphery.
\subsection{Radiative Transfer Calculations}\label{sec_radtrans}
We couple the simulations described above with radiative transfer models
to calculate the time evolution of the SEDs of cores collapsing according
to the predictions of the simulations, as well as images and interferometer
visibilities at all wavelengths.
The coupling between the simulations and radiative transfer models was first
described by \citet{dunham2012:evolmodels}. We give a brief overview here but
refer to \citet{dunham2012:evolmodels} for more
details. Since the simulations do not provide the full volume density
structure, in order to set up the physical structure
of the radiative transfer models we take analytic profiles for the core and
disc density structure and re-scale them at each timestep according to known
parameters from the simulations (see \S \ref{sec_limit_structure} for a
discussion of the limitations of this approach and prospects for future
improvements). These parameters include the initial core
mass (\mcore) and radius (\renv),
along with the time evolution of the core mass, disc mass (\mdisc),
protostellar mass (\mstar), disc outer radius (\rdisc), accretion rate onto
the protostar (\mdotstar), and accretion rate onto the disc (\mdotdisc).
For the core, we adopt the density profile given by the
\citet{terebey1984:model} solution for the collapse of a slowly rotating core.
This solution gives a core that is initially a spherically symmetric, singular
isothermal sphere with a density distribution $n \propto r_{core}^{-2}$,
identical to the classic \citet{shu1977:sis} solution. As collapse proceeds,
the solution takes on two forms: an outer solution that is similar to the
non-rotating, spherically symmetric solution and an inner solution that
exhibits flattening of the density profile. This model is parameterized by the
initial angular velocity of the core and the time since the formation of the
protostar. At each timestep we truncate the solution at the given initial
\renv\ for each model and
then renormalize the density profile so that the core mass matches that given
by the simulations. Once the material at the outer core boundary beings
collapsing, which occurs once the infall
radius\footnote{The infall radius is the radius within which the core is
collapsing. It starts at the center and moves outward at the sound speed.}
exceeds the initial outer radius, we use the velocity profiles given by the
\citet{terebey1984:model} solution to allow the core radius to decrease.
The disc structure follows a power law in the
radial coordinate and a Gaussian in the vertical coordinate, with the density
profile given by:
\begin{equation}\label{eq_disc_density_profile}
\rho_{disc}(s,z) = \rho_0 \left(\frac{s}{s_o} \right)^{-n}
exp\left[ -\frac{1}{2} \left(\frac{z}{H_s}\right)^2 \right] \qquad ,
\end{equation}
where $z$ is the distance above the midplane ($z=rcos\theta$, with $r$ and
$\theta$ the usual radial and zenith angle spherical coordinates), $s$ is the
distance in the midplane from the origin ($s= \sqrt{r^2 - z^2}$), and
$\rho_0$ is the density at the reference midplane distance $s_0$.
The quantity $H_s$ is the disc scale height and is given by
$H_s = H_o \left(\frac{s}{s_o}\right)^\gamma$, where $H_o$ is the scale height
at $s_o$. The parameter $\gamma$ describes how the scale height changes with
$s$ and sets the flaring of the disc. We set $H_o = 10$ AU at $s_o = 100$ AU
and $\gamma = 1.25$ (see \citet{dunham2010:evolmodels} and references therein
for a more detailed discussion of these quantities), giving flared discs very
similar to those found in the hydrodynamical simulations (see Figure 11 of
\citet{vorobyov2010:bursts}). The disc surface density profile, calculated by
integrating Equation \ref{eq_disc_density_profile} over the vertical
coordinate $z$, has a radial power-law index of $\Sigma(s) \propto s^{-p}$,
where $p = n - \gamma$. Unlike our previous work
\citep{dunham2010:evolmodels} where we set $p$ to a fixed value,
here we determine $p$ by fitting a power-law to
the azimuthally-averaged surface density profile given by the
simulations, and use this to set $n$
($p=-1.68, -1.44$ for the Class 0 and I evolutionary stages of model 1,
respectively, and $p=-1.77, -1.36$ for the Class 0 and I evolutionary stages
of model 2, respectively). Finally, the disc inner
radius is set equal to the dust destruction radius, calculated
(assuming spherical, blackbody dust grains) as
\begin{equation}
R_{disc}^{in} = \sqrt{\frac{L_*}{16 \pi \sigma T_{dust}^4}} \qquad ,
\end{equation}
where $L_*$ is the protostellar luminosity (see below) and $T_{dust}$ is the
dust destruction temperature
\citep[assumed to be 1500 K; e.g.,][]{cieza2005:tdust}.
The total internal luminosity of the protostar and
disc at each point in the collapse from core to star contains six components:
(1) luminosity arising from accretion from the core directly onto the
protostar, (2) luminosity arising from accretion from the core onto the disc,
(3) luminosity arising from accretion from the disc onto the protostar, (4)
disc ``mixing luminosity'' arising from luminosity released when newly
accreted material mixes with existing disc material, (5) luminosity arising
from the release of energy stored in differential rotation of the protostar,
and (6) photosphere luminosity arising from gravitational contraction and
deuterium burning. The first five components are calculated following
\citet{adams1986:protostars}, using direct inputs from the simulations for
the accretion rates, masses, and sizes of the protostar, disc, and core
(further details can be found in \citet{young05:evolmodels},
\citet{dunham2010:evolmodels}, and \citet{dunham2012:evolmodels}).
The photospheric luminosity and protostellar radius are
calculated from the stellar evolution code
of \citet{BC2010} coupled to the main numerical hydrodynamic code
(see \citet{vorobyov2013:chemistry}
for details).
While the total amount of accretion luminosity is correct,
we do not include viscous disc heating since the radiative transfer code can
only incorporate heating due to one central, internal source of photons and
an external radiation field. Viscous heating is only significant within the
inner $20-30$ AU of protostellar discs, and even in those locations it only
increases the disc temperatures by much less than a factor of two (see, e.g.,
Figure 4 of \citet{vorobyov2010:disks}). Since the discs we consider are 100
AU or larger, even at the earliest times (see right panels of Figure
\ref{fig_models}), small deviations to warmer temperatures in the inner discs
than calculated from our radiative transfer models will have a negligible
impact on our results.
Finally, we also include external luminosity arising from heating of the core
by the Interstellar Radiation Field (ISRF). As in our previous papers, we
adopt the \citet{black1994:isrf} ISRF, modified in the ultraviolet to
reproduce the \citet{draine1978:isrf} ISRF, and then extincted by $A_V = 0.5$
of dust with properties given by \citet{draine1984:dust} to simulate extinction
by the surrounding lower density environment.
\section{Results}\label{sec_results}
The two largest surveys of protostellar discs to date, and the only two surveys
to include Class 0 protostars, are those presented by
\citet{jorgensen2009:prosac} and \citet{enoch2011:disks}.
Both studies were based on
measuring the total 1.1 or 1.3 mm interferometric amplitude at baselines of
50 k$\lambda$ for protostellar sources at distances ranging from 125 to
415 pc. At these distances, emission on baselines of 50 \kl\
corresonds to spatial scales of approximately $600-2000$ AU, and the basic
assumption made is that this emission has completely resolved out
the surrounding core but not yet begun to resolve out the disc. Under this
assumption, and further assuming that the observed emission is from
optically thin, isothermal dust at a temperature of 30 K, disc masses can then
be calculated assuming a standard gas-to-dust ratio of 100. A number of open
questions surrounding these assumptions must be answered before any general
conclusions can be drawn about the formation and early evolution of
protostellar discs, including:
\begin{enumerate}
\item Is it true that the total amplitude measured at baselines of 50
k$\lambda$ is free of contamination from the surrounding core and recovers
the total emission from the disc, and how do these answers depend on the
distance to the source?
\item Is it valid to assume that the (sub)millimeter emission is optically
thin, or is mass missed due to optically thick emission?
\item Is 30 K the best temperature to assume?
\item How do the answers to all of these questions depend on the source
inclination and the wavelength of the observations?
\end{enumerate}
While \citet{jorgensen2009:prosac} partially addressed the
first question by attempting to estimate and remove contaminating emission from
the surrounding core, it is unclear how accurate their corrections are since
they used simple, one-dimensional models to derive them.
To answer these questions, we take each of the models listed in Table
\ref{tab_models} and, for source distances ranging from 100 pc to 500 pc in
steps of 50 pc, calculate the complex interferometer visibilities at four
wavelengths: 3.4 mm, 1.3 mm, 0.7 mm (700 \um), and 0.35 mm (350 \um)
(approximately 90, 230, 430, and 850 GHz, respectively). The visibilities
are calculated using an accompanying routine in the RADMC package and are
generated assuming the Atacama Large Millimeter Array (ALMA)
primary beam at each of the wavelengths listed
above, for nine inclinations ranging from 5\degree\ to 85\degree\ in steps of
5\degree. The contributions to the visibility amplitudes from the
core and disc are calculated separately by coupling the hydrodynamic
simulations with the radiative transfer models a second time with the disc
mass set to zero,
generating a set of core-only visibilities, and then subtracting their
amplitudes from the combined core+disc amplitudes to generate a set of
disc-only amplitudes. We
consider three timesteps for each model: one early in the Class 0 stage when
only 25\% of the initial core mass has accreted onto
the protostar and/or disc, one at the Class 0/I boundary when 50\% of the
initial core mass has accreted onto the protostar and/or disc, and one late
in the Class I stage when 75\% of the initial core mass has accreted onto
the protostar and/or disc. While we caution that there is not always a
one-to-one correspondence between observational Class and physical Stage
due to the effects of geometry and episodic accretion
\citep[e.g.,][]{dunham2014:ppvi}, for simplicity here we use the term Class
to refer to both. In the following subsections we use these
synthetic observations to answer the questions listed above.
\subsection{Contamination From the Surrounding Core}\label{sec_results_core}
To evaluate whether or not the total amplitude measured at a baseline of 50
\kl\ is free of contamination from the surrounding core and truly
represents emission from only the disc, we define the quantity $f_D$ as
the ratio of the amplitude at 50 \kl\ in the disc-only visibilities to the
amplitude at 50 \kl\ in the combined (core+disc) visibilities. Figure
\ref{fig_fd} plots $f_D$ versus distance to the source for each timestep of
each model for each of the four wavelengths considered in this paper,
at an inclination angle of 45\degree. For
model 1, which has a relatively massive disc, there is almost no contamination
from the core. Even during the Class 0 stage when the core still contains
the majority of the system mass, at the largest distances where the core is
the least well-resolved, at most only 20\% of the amplitude measured at
50 \kl\ arises from the core.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f3.eps}}
\caption{$f_D$, the ratio of the amplitude at 50 \kl\ in the disc-only model
to the amplitude at 50 \kl\ in the disc+core model, plotted as a function
of distance to the source for distances ranging from 100 to 500 pc. The
amplitudes are calculated for a source inclination angle of 45\degree.}
\label{fig_fd}
\end{figure}
As seen in Figure \ref{fig_fd},
the contamination level is somewhat higher for model 2, which is not
surprising given that this model forms a less massive disc. During the Class 0
stage, between 10\% -- 60\% of the observed amplitude at 50 \kl\ arises
from the surrounding core, with the highest contamination levels occuring
for submillimeter observations of sources at the largest distances. At
the Class 0/I boundary the contamination levels are lower, in the range of
0\% -- 35\%, and they are even lower in the Class I stage, ranging between
0\% and 20\%. These results for both models do not show any significant
dependence on inclination angle.
Overall, we find that the surrounding core does not significantly contaminate
the visibility amplitudes at 50 \kl. For cores that form relatively massive
discs the contamination is almost completely negligible for all but the
earliest times and largest source distances, where it can reach up to 20\%.
For cores that form less massive discs, up to $\sim$60\% of the amplitude
may actually arise from the core rather than the disc for Class 0 sources
at distances of 400 -- 500 pc observed at submillimeter wavelengths. For
any combination of closer distances, more evolved sources, and longer
wavelength observations the contamination is less.
These results are in general agreement
with those of \citet{jorgensen2009:prosac}, who used simple, one-dimensional
models to estimate that cores contribute about 10\% -- 30\% of the observed
interferometer amplitudes at 50 \kl.
\subsection{Resolving Out the Disc}\label{sec_results_resolve}
Depending on both the size of the disc and the distance to the source, the
interferometer amplitudes at 50 \kl\ may not trace the full mass of the disc
if the disc is partially resolved. Indeed, \citet{enoch2011:disks} noted
that, due to the disc being partially resolved out at 50 \kl, the inferred
disc mass for one source in their sample
was 30\% lower than the mass obtained from detailed radiative
transfer modeling performed in an earlier study \citep{enoch2009:disks}.
Since the \citet{vorobyov2010:bursts}
simulations form relatively large, massive discs, we evaluate here whether
disc masses calculated from the amplitues at 50 \kl\ are underestimated due
to the discs being partially resolved out. For each timestep of each
model we define the quantity $f_{TD}$ as the ratio of amplitudes in the
disc-only model at 50 \kl\ to that at 0 \kl. In this sense $f_{TD}$ measures
the fraction of total disc flux recovered at 50 \kl.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f4.eps}}
\caption{$f_{TD}$, the ratio of the amplitude at 50 \kl\ in the disc-only
model to the total amplitude of the disc-only model (the amplitude at 0 \kl),
plotted as a function of distance to the source for distances ranging from 100
to 500 pc. The amplitudes are calculated for a source inclination angle of
45\degree.}
\label{fig_ftd}
\end{figure}
Figure \ref{fig_ftd} plots $f_{TD}$ versus distance to the source for each
timestep of each model for each of the four wavelengths considered in this
paper, at an inclination angle of 45\degree\ (the choice of inclination
angle is not found to have any significant effect on the results). In the
Class 0 stage the amplitudes at 50 \kl\ do recover essentially all of the
disc emission except for sources closer than 200 pc, where up to 20\% of the
emission is resolved out in models with larger and more massive discs.
However, as the disc grows larger at later evolutionary stages, more of
the disc emission is resolved out. For model 2, the disc remains small
enough such that essentially no emission is resolved out for source distances
greater than $\sim$ 200 pc; for closer sources, up to 20\% is resolved out
at the Class 0/I boundary, and up to 30\% late in the Class I stage.
For model 1, on the other hand, the disc becomes partially resolved out
for all but the largest source distances. Up to 50\% of the total disc
emission is resolved out at the Class 0/I boundary for observations at
3.4 and 1.3 mm (and even more at submillimeter wavelengths), and up to 60\%
late in the Class I stage (again, even more at submillimeter wavelengths).
Combining the results of the last two sections, interferometer amplitudes at
50 \kl\ will overestimate the emission from a disc by up to about a factor of
two in the worst cases due to contaminating emission from the surrounding
core. On the other hand, they will also underestimate the emission from a disc
by a similar factor due to partially resolving out the disc, as long as the
observations are obtained at millimeter wavelengths. For large
enough samples of objects over different ranges of source distances and
intrinsic disc properties, these effects will at least partially cancel each
other out. For observations obtained at submillimeter wavelengths,
however, up to 80\% of the disc emission can be resolved out, and as this
effect is larger than the contamination from the surrounding core,
submillimeter observations used to measure disc masses in this manner will
systematically underestimate the true masses.
\subsection{Optical Depth, Temperature, and Inclination}\label{sec_results_mass}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f5.eps}}
\caption{The vertically integrated optical depth through the disc for each
timestep of each model, plotted as a function of the midplane radial distance.
The optical depth is calculated and displayed for each of the four wavelengths
considered in this paper. The thick gray line in each panel marks the limit
between optically thick ($\tau > 1$) and optically thin ($\tau < 1$). The
optical depths are calculated using OH5 opacities (see text for details).}
\label{fig_tau}
\end{figure}
In this section we evaluate the effects of optical depth, temperature, and
inclination on our ability to recover the actual disc masses from our
synthetic observations. The observed disc emission will only trace the
total mass of the disc if the emission remains optically thin throughout the
full vertical and radial extent of the disc. This condition is known to be
false in older Class II discs where the inner regions are often optically
thick \citep[e.g.,][]{andrews2005:disks}. To evaluate whether the protostellar
discs predicted by the \citet{vorobyov2010:bursts} simulations are also
partially optically thick,
Figure \ref{fig_tau} plots, for each wavelength considered in this
paper, the vertically integrated optical depth through the disc for each
timestep of each model as a function of the midplane radial distance.
For self-consistency the optical depths are calculated using OH5 dust
opacities, since these opacities were used in the radiative transfer models to
generate the synthetic observations.
The vertically integrated optical depth at a fixed midplane radial distance
increases with decreasing wavelength, as expected given that the opacities
also increase with decreasing wavelength. In all cases, the discs are not
optically thin all the way down to the midplane for all radii. For model 1,
the disc emission becomes optically thin beyond a few AU for the millimeter
wavelengths but
only beyond ten to a few tens of AU for the submillimeter wavelengths.
For model 2, which forms a smaller, less massive disc that is
consequently less optically thick, the disc is generally optically thin
beyond a few AU except for the shortest wavelengths and latest evolutionary
stages, where it is only optically thin beyond about 10 AU.
These results demonstrate that the discs formed by the
\citet{vorobyov2010:bursts} simulations are not fully optically thin,
in agreement with other numerical studies (see e.g. \citet{Rice2010}).
However, they are not optically thick enough to cause large mass
underestimates. For a disc with a surface density profile
$\Sigma(s) \propto s^{-p}$, the mass contained within the midplane
radius $s$ is $M(<s) \propto s^{2-p}$. At each timestep the disc size (Figure
\ref{fig_models}), the power-law index (see \S \ref{sec_radtrans}), and the
midplane radius where the emission transitions to optically thin (Figure
\ref{fig_tau}) can be combined to calculate the fraction of total mass
contained within the optically thick region of the disk. For the shortest,
most optically thick wavelength considered here (0.35 mm), we calculate this
fraction to be 41\%, 23\%, and 20\% for the
Class 0, 0/I, and I timesteps of model 1, respectively, and 45\%, 14\%, and
14\% for model 2. Thus the total disc mass underestimates due to optically
thick emission are less than a factor of two even at the shortest submillimeter
wavelengths (except for edge-on lines-of-sight; see below).
With these results in mind,
we calculate, at each wavelength, the total disc mass inferred
from the synthetic interferometer amplitude at 50 \kl\ for each timestep of
each model. The mass is given by
\begin{equation}\label{eq_dustmass}
M = 100 \frac{d^2 S_{\nu}(50~\kl)}{B_{\nu}(T_D) \kappa_{\nu}} \quad ,
\end{equation}
where $S_{\nu}(50~\kl)$ is the interferometer amplitude at 50 \kl,
$B_{\nu}(T_D)$ is the Planck
function at the assumed isothermal dust temperature $T_D = 30$ K,
$\kappa_{\nu}$ is the dust opacity, $d$ is the distance to the source, and
the factor of 100 is the assumed gas-to-dust ratio. We again adopt the OH5
dust opacities for self-consistency with the radiative transfer models
(see \S \ref{sec_results_opacity} for discussion of the opacities), and use
the synthetic interferometer visibilities calculated for a source distance of
250 pc.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f6.eps}}
\caption{Disc mass calculated from the interferometer visibilities at 50
\kl\ at each wavelength, for each timestep of each model, calculated according
to Equation \ref{eq_dustmass} (see text for details) and plotted versus the
inclination assumed to generate the synthetic visibilities. In each panel the
thick gray line shows the actual disc mass.}
\label{fig_mass}
\end{figure}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f7.eps}}
\caption{Ratio of actual to calculated disc mass at each wavelength,
for each timestep of each model, plotted versus the inclination assumed to
generate the synthetic interferometer visibilities.}
\label{fig_massratio}
\end{figure}
Figure \ref{fig_mass} plots the resulting disc masses calculated at each
wavelength as a function of inclination, and Figure \ref{fig_massratio}
plots the ratio of actual to calculated disc masses, again as a function of
inclination.
In both Figures, the calculated disc masses are seen to converge
toward a single value for all inclinations less than $\sim$60\degree\
as the wavelength increases. The increasing underestimates in disc mass at
progressively shorter wavelengths is due to a combination of two factors:
(1) the emission is progressively more optically thick, and (2) the
emission is progressively farther from the Rayleigh-Jeans limit, thus
temperature overestimates cause progressively larger disc mass underestimates.
The importance of the latter factor is emphasized since we have already shown
that the underestimates due to optically thick emission are less than a factor
of two even at the shortest wavelengths.
In four of the six panels the masses calculated from the
longest, most optically thin wavelengths converge to lower masses than the
actual disc masses. These underestimates grow more significant with later
evolutionary stages, reaching up to factors of two late in the Class I stage,
and are due to the fact that the assumed isothermal temperature of 30 K is
often too high. As the discs are not truly isothermal,
the correct temperature to assume in Equation \ref{eq_dustmass} is the
mass-weighted average temperature of the optically thin portions of the disc.
Table \ref{tab_temperatures} lists these temperatures for each timestep of
each model. The temperatures decrease as the evolutionary stages increase,
since as the disc evolves it grows in mass and size and is optically thick
to larger radii (see Figure \ref{fig_tau}), leading to more shielding of the
optically thin regions. Consequently, masses calculated
assuming $T_D = 30$ K lead to progressively larger underestimates as
evolutionary stage increases.
A similar trend of decreasing disc temperatures leading
to increasing mass underestimates with evolutionary stage was also noted by
\citet{jorgensen2009:prosac}. The small overestimate (less than 20\%) in the
disc mass for the Class 0 stage of model 2 is due to the temperature being
slightly above the assumed 30 K (35 K, see Table \ref{tab_temperatures}).
\input{tab2.tex}
Assuming a random distribution of source orientations on the sky, 50\% of all
sources will be viewed at inclinations of 60\degree\ or greater.
For such sources, the lines-of-sight will pass partially or fully through the
disc midplane where the densities (and thus optical depths) are the largest.
As illustrated by Figures \ref{fig_mass} and \ref{fig_massratio}, the disc
mass underestimates due to the combination of optically thick emission
and incorrect isothermal dust temperatures for sources
with inclinations greater than $\sim$60\degree\ are generally less than 50\%
for observations at millimeter wavelengths, but can reach up to factors of
5--10 or larger for submillimeter wavelengths.
\subsection{Dust Opacity Law}\label{sec_results_opacity}
As noted above, the radiative transfer models used to generate the
synthetic interferometer visibilities assumed the OH5 dust opacity law,
so to be self-consistent we also adopted the OH5 dust opacities when
calculating disc masses according to Equation \ref{eq_dustmass}.
Defining $\beta$ as the power-law index of the dust opacity law
($\kappa \propto \nu^{\beta}$, where $\kappa$ is the dust opacity at
frequency $\nu$), optically thin emission will scale as
$I \propto \nu^{\alpha}$, where $\alpha$ is the spectral index and is given
by $\alpha = 2 + \beta$. For OH5 dust, $\beta = 1.75$ over all
submillimeter and millimeter wavelengths, implying that the emission should
have a spectral index of 3.75. However, numerous recent observational
studies of protostars have found $\alpha < 3.75$ over a variety of spatial
scales \citep[e.g.,][]{kwon2009:beta,melis2011:l1527,shirley2011:mustang,scaife2012:disks,chiang2012:l1157,tobin2013:l1527}.
In an extreme example, \citet{tobin2013:l1527} measured $\alpha = 2$ between
3.4 and 0.87 mm over baselines ranging from 10 -- 500 \kl, implying a flat
dust opacity law ($\beta = 0$).
There are two ways to lower the spectral index: partially optically thick
emission and grain growth.
It is well-established that dust grains can quickly grow to
millimeter sizes or larger in Class II discs
(older discs than those considered in this study),
and that such growth can change the opacity law by decreasing $\beta$
\citep[see][for a recent review]{testi2014:ppvi}. Whether or not
grains have grown sufficiently large to change the opacity law in protostellar
discs remains an open question.
Since we have already shown that the discs
in these models are partially optically thick, especially at shorter
wavelengths, we first evaluate whether optical depth alone can explain these
recent results of $\alpha < 3.75$ in observations of protostars.
Figure \ref{fig_spectralindex} plots the spectral index, $\alpha$, calculated
from each wavelength pair, for each timestep of each model, as a function of
baseline distance. Since grain growth is not an option in models with a fixed
dust opacity law, and $\beta = 1.75$ over all submillimeter and millimeter
wavelengths for OH5 dust opacities,
any deviations from $\alpha = 3.75$ are due to
optical depth effects. We find that $\alpha$ is lower when calculated from
shorter wavelengths that are more optically thick, and is less than 3.75
even when calculated only from millimeter wavelengths. Similar results are
found when three or four wavelengths are used to calculate $\alpha$, as long
as the first and last wavelengths remain the same. While these results do
suggest that optical depth lowers the observed $\alpha$, the
\citet{tobin2013:l1527} results are most comparable to our results for $\alpha$
calculated from 3.4 and 0.7 mm, and since we find $\alpha \sim 3-3.5$ at these
wavelengths, optical depth effects alone are unable to explain their results.
Furthermore, while the source studied by
\citet{tobin2013:l1527} is at a nearly edge-on
inclination and our results are presented for an inclination of 45\degree,
we find qualitatively similar results even for an inclination as high as
85\degree\ in that the observed $\alpha$ is generally larger than two.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{f8.eps}}
\caption{Spectral index, $\alpha$, calculated from each wavelength pair for
each timestep of each model, as a function of baseline distance. Calculations
over different wavelength ranges are shown, with the legend in each panel
showing the two wavelengths used to calculate $\alpha$ for each line. A
source inclination of 45\degree\ was assumed for these calculations.}
\label{fig_spectralindex}
\end{figure}
If grains have grown to larger sizes in discs, as suggested by combining our
results with the observational studies cited above
\citep[see also the discussion in][]{tobin2013:l1527}, the opacity law will
flatten. This will generally lead to smaller dust opacities at submillimeter
wavelengths and larger dust opacities at millimeter wavelengths. To give a
specific example, the dust opacity law of \citet{ricci2010:disks},
which includes grains that have grown up to 1 mm,
predicts $\beta=0.25$ and $\kappa = 2.5$ cm$^2$ g$^{-1}$ at 850 \um. This
opacity law is within a factor of 1.3 of that predicted by the OH5 dust
opacity law at 850 \um\ (1.89 cm$^2$ g$^{-1}$), but leads to opacities that
are a factor of 3 smaller at 350 \um\ and factors of 2.5 and 10.5 larger at
1.3 and 3.4 mm, respectively. If the \citet{ricci2010:disks} opacity law
represented the true opacity law in the disc, but the OH5 dust opacities were
erroneously adopted to calculate disc masses,
the resulting masses would underestimate the true disc masses at the
shortest wavelengths but overestimate them at the longest wavelengths.
Future work must concentrate on determining
whether the \citet{tobin2013:l1527} result of a very low spectral index for
L1527 is an anomaly or the norm for protostars, and use these results to
set detailed constraints on the opacity law for dust in protostellar discs.
\section{Discussion}\label{sec_discussion}
\subsection{Reliability of Protostellar Disc Mass Measurements}\label{sec_discussion_reliability}
\begin{figure*}
\centering
\includegraphics[width=18cm]{f9.eps}
\vspace{-1.42cm}
\caption{
Protostellar disc mass distributions for each timestep of each model. In each
panel the solid histograms show the intrinsic distributions taken directly
from the simulation data, thus the solid histograms are the same in each panel
for a given model. The dashed histograms show the synthetic disc mass
distributions over the full evolution of each model calculated from the
synthetic interferometer amplitude at 50 \kl\ (see text for details).
In the top two rows the assumed isothermal dust temperature is 30 K for
all timesteps, whereas in the bottom two rows it is 30 K for timesteps in the
Class 0 evolutionary stage and 15 K for timesteps in the Class I evolutionary
stage. Each panel is labeled with the model, wavelength, and isothermal dust
temperatures adopted to calculate disc masses in the Class 0 and I evolutionary
stages.
}
\label{fig_mdisc_mean}
\end{figure*}
\input{tab3.tex}
To summarize the above sections, we find that,
on average, the competing effects of contamination from the surrounding core
and partially resolving the disc will tend to cancel each other
at millimeter wavelengths but lead to underestimates at submillimeter
wavelengths, by up to factors of 4 in the worst cases. Optically thick
emission, colder dust temperatures than the assumed $T_D = 30$ K, and the
effects of inclination can lead to underestimates by up to factors of $2-3$ at
millimeter wavelengths and up to an order of magnitude or larger at
submillimeter wavelengths, especially 350 \um.
Finally, flatter dust opacity laws than the assumed
OH5 opacity law, as implied by recent observations, will cause further
disc mass underestimates at the shortest submillimeter wavelengths and
overestimates at the longest millimeter wavelengths, with the exact magnitude
dependent on the details of the intrinsic dust opacity law.
In order to understand the net sum of all of these effects
(except for grain growth, which we are unable to evaluate with our fixed
dust opacity law), Figure \ref{fig_mdisc_mean} plots various intrinsic and
synthetic disc mass distributions for each model. The intrinsic distributions
are taken directly from the simulation data and represent the true disc mass
distribution over each model run; thus they are the same
in each panel for a given model. The synthetic distributions are assembled by
calculating, for each wavelength and each inclination,
the disc mass throughout the full evolution of the model (in 2000
yr timesteps) from the synthetic interferometer amplitude at 50 \kl,
assuming optically thin, isothermal emission according to Equation
\ref{eq_dustmass}. The contribution of each
inclination to the final distribution for each wavelength of each model is
weighted by solid angle (see \citet{dunham2010:evolmodels} for details).
In the top two rows the assumed isothermal dust temperature is 30 K for
all timesteps, whereas in the bottom two rows it is 30 K for timesteps in the
Class 0 evolutionary stage and 15 K for timesteps in the Class I evolutionary
stage. The mean of each distribution is listed in Table \ref{tab_masses}.
When the isothermal $T_D$ is held fixed at 30 K independent of evolutionary
stage, the calculated disc mass distributions are clearly lower limits to
the intrinsic distributions, with mean values lower by factors of
$1.5-6$ depending on wavelength and model. However, when two values of
$T_D$ are adopted, 30 K for Class 0 sources and 15 K for Class I sources,
the calculated distributions are much more representative of the intrinsic
distributions (mean values within 50\%).
\subsection{Existence of Fragmenting Protostellar Discs}\label{sec_discussion_fragmenting}
\begin{figure*}
\centering
\includegraphics[width=16cm]{f10.eps}
\caption{Distribution of protostellar disc masses. In each panel, the shaded
histogram shows the combined distribution of observed disc masses
from \citet{jorgensen2009:prosac}, \citet{enoch2011:disks}, and
\citet{eisner2012:disks}, with the error bars showing the statistical
($\sqrt{N}$) uncertainties, and the dashed histogram shows the distribution of
protostellar disc masses predicted by the \citet{vorobyov2010:bursts}
simulations. The IMF used to calculate the distributions from the simulations
are those of \citet[][left panels]{kroupa2002:imf} and
\citet[][right panels]{chabrier2005:imf} (see text for details).
The top two panels show the observed disc mass distribution calculated assuming
a fixed isothermal dust temperature of $T_D = 30$ K independent of evolutionary
stage. The bottom two panels show the observed disc mass distribution
calculated assuming $T_D = 30$ K for Class 0 sources and $T_D = 15$ K for
Class I sources.
}
\label{fig_mdisc_histogram}
\end{figure*}
Given the results presented in this paper, what constraints on protostellar
disc masses can be set by existing surveys? Figure \ref{fig_mdisc_histogram}
plots the combined distribution of protostellar disc masses from
\citet{jorgensen2009:prosac}, \citet{enoch2011:disks}, and
\citet{eisner2012:disks}, all of which are based on observations at
1.1 or 1.3 mm.
In the top two panels, disc masses were taken directly from the first two
surveys, who used the method evaluated here with a fixed $T_D$ of 30 K, and
were calculated from the \citet{eisner2012:disks} survey based on their
reported visibility amplitudes for each source, again assuming a fixed
$T_D$ of 30 K. We note that these calculated disc masses generally agree
to within a factor of two with masses derived by \citet{eisner2012:disks}
based on radiative transfer modeling of each source, which we take to be
excellent agreement considering they fit to a relatively coarsely sampled
model grid. In the bottom two panels, the disc masses from all three
surveys are recalculated assuming $T_D = 30$ K for Class 0 sources and
$T_D = 15$ K for Class I sources. Altogether, 43 disc mass
measurements are plotted; an additional four non-detections in the combined
sample are not shown in Figure \ref{fig_mdisc_histogram}.
The observed distribution of protostellar disc masses spans more than two
orders of magnitude, from less than 10$^{-2}$ \msun\ to greater than 1 \msun,
with a peak slightly above 10$^{-2}$ \msun\ (although, as demonstrated
by the error bars in Figure \ref{fig_mdisc_histogram}, the shape of this
distribution is still quite uncertain due to small-number statistics).
Figure \ref{fig_mdisc_histogram} also plots the distribution of protostellar
disc masses predicted by the \citet{vorobyov2010:bursts} simulations, which
often form discs that are sufficiently massive to fragment and drive
short-lived accretion bursts from the disc onto the protostar.
This distribution is calculated as the fraction of total time spent at each
disc mass for an ensemble of 23 simulation runs spanning a large range of
initial core masses and angular momenta (see Table 1 of
\citet{dunham2012:evolmodels} for a full list). The contribution of each
simulation to the total distribution is weighted by either the
\citet[][left]{kroupa2002:imf} or \citet[][right]{chabrier2005:imf} stellar
initial mass function (IMF) based on the final mass of the star formed in each
simulation (see \citet{dunham2012:evolmodels} for details of this weighting).
Comparing the two distributions in the top panels reveals that the relatively
massive discs predicted by the \citet{vorobyov2010:bursts} non-magnetic
hydrodynamical simulations are observed in approximately 50\% of all
protostars. However, since the observed disc mass distribution in these
panels was assembled a fixed $T_D = 30$ K, our results from \S
\ref{sec_discussion_reliability} imply that it is a lower limit to the
intrinsic protostellar disc mass distribution. Comparing the two
distributions in the bottom panels, where $T_D = 15$ K for Class I sources
and thus the observed distribution should closely match the intrinsic
distribution, the discs predicted by the \citet{vorobyov2010:bursts}
simulations are observed in approximately 70\% of all protostars. In both the
top and bottom panels somewhat better agreement is found when
the simulations are weighted by the \citet{kroupa2002:imf} IMF, which gives
increased weight to simulations that form lower-mass stars.
Based on these comparisons, and given that the \citet{vorobyov2010:bursts}
simulations are prone to form fragmenting discs, particularly when the disk
mass exceeds $\sim$0.07 \msun\ \citep{Vorob2013}, we conclude
that at least some protostellar discs are likely sufficiently massive to
fragment. The fact that the observed and simulation disc mass distributions
do not show perfect overlap (in particular, the observed distribution extends
to lower masses) may be explained by magnetic braking acting to
remove angular momentum and suppress disc formation in some cases.
Additionally, we caution that the surveys published to date
are generally biased toward massive and luminous protostars that are thus
bright at (sub)millimeter wavelengths, and may not be representative of all
embedded protostars. Larger, more representative surveys are needed to test
these findings. We are currently working on assembling such samples, on
including the effects of magnetic braking into the numerical models, and on
searching for other observational signatures of fragmenting discs beyond
their total masses, such as those predicted by \citet{vorobyov2013:fragments},
including direct detection of spiral structure and fragments and characteristic
SED features.
\subsection{Recommendations for Future Studies}\label{sec_discussion_recommendations}
Given the results presented in this paper, we make the following two
recommendations for future studies that seek to use interferometer continuum
observations of protostars to measure the protostellar disc mass distribution
for larger samples than those considered to date:
\begin{enumerate}
\item Observe at a wavelength of approximately 1 mm. Shorter wavelengths will
be more optically thick and more prone to errors due to incorrect temperature
assumptions (since they are farther outside of the Rayleigh-Jeans limit), and
longer wavelengths will be more prone to disc mass overestimates due to
larger opacities caused by grain growth.
\item Adopt two different isothermal dust temperatures depending on the
evolutionary stage of each protostar: 30 K for Class 0 protostars and 15 K
for Class I protostars.
\end{enumerate}
\section{Limitations and Caveats}\label{sec_limit}
\subsection{Physical Structure}\label{sec_limit_structure}
While both the hydrodynamical simulations and radiative transfer calculations
are performed using two-dimensional codes, they use different coordinate
systems. In the hydrodynamical simulations, the thin-disc
approximation is used to calculate the evolution of the gas surface density and
velocity profiles as a function of cylindrical radius, $s$, and azimuthal
angle, $\phi$,
whereas the radiative transfer calculations assume axisymmetry to calculate the
dust temperature distribution as a function of radius, $r$, and zenith angle,
$\theta$, in
spherical coordinates. As a result, the radiative transfer calculations
are able to match the exact mass, radius, and azimuthally-averaged surface
density profile of the disc but do not recover its axial structure,
in particular spiral density waves and fragments.
As described in \S \ref{sec_radtrans}, we match the disc surface density
profiles
in the simulations and radiative transfer calculations by fitting a power-law
to the azimuthally-averaged surface density profile predicted by the
simulations. However, inspection of the simulations reveal that the profiles
are generally better described by a broken power-law, with a much steeper
index at large radii, than by a single power-law over the full extent of the
disc. We included
such a broken power-law in the radiative transfer calculations to test its
effects but found that it had no significant effects on our results, since
it changes slightly the distribution of mass within the disc but not the total
mass recovered by the simulated observations.
However, we do acknowledge that there are limitations related to our inability
to include the full, non-axisymmetric structure of the disc in our radiative
transfer calculations. One such limitation is the effect of disc fragments.
In the most extreme cases, the discs in the \citet{vorobyov2010:bursts}
simulations form fragments that contain up to 50\% of the total mass of the
disc. Such fragments will be extremely optically thick and thus essentially
invisible to (sub)millimeter interferometric observations.
Thus, in addition to the effects discussed in \S \ref{sec_results},
hidden fragments will cause disc masses calculated from interferometer
visibility amplitudes measured at 50 \kl\ to be further underestimated.
The full magnitude of these underestimates can only be evaluated once the
simulations are coupled to three-dimensional radiative transfer calculations
that capture the full disc structure; we are currently working on such models
and will present the results in a forthcoming paper.
Another limitation is the effect that grain growth will have on the assumption
that the gas and dust are well-coupled, which we adopt when using the gas
surface density from the simulations to set the dust surface density profile
in the radiative transfer calculations (as described in \S
\ref{sec_radtrans}). If dust grains in protostellar discs have already grown
to millimeter sizes, the dust surface density profile can deviate significantly
from the gas due to preferential concentration of these large grains within
density enhancements \citep[such as fragments, clumps, and spiral arms;][]{rice2004:graingrowth,boley2010:graingrowth,testi2014:ppvi}.
Thus we again emphasize the critical need for future work devoted to
understanding the dust properties in protostellar discs to go along with recent
advances in our understanding of dust properties in older, more evolved
(i.e., Class II) discs \cite[see][for a recent review]{testi2014:ppvi}. Such
work requires high-angular resolution, high-sensitivity, multi-wavelength
observations of embedded protostellar discs and should thus be viable in the
coming years with ALMA and the upgraded Very Large Array (VLA).
\subsection{Confirming Keplerian Rotation}\label{sec_limit_keplerian}
Emission on baselines of 50 \kl\ corresponds to spatial scales of approximately
$500-2500$ AU for sources at distances ranging from $100-500$ pc, thus we
must emphasize that such observations do not spatially resolve the discs.
As discussed in \S \ref{sec_results_resolve}, this is actually a strength as
it allows this method to probe the total emission from the disc, but at the
cost of assuming that the unresolved component does in fact arise from a
rotationally supported, Keplerian disc. To test whether emission on 50 \kl\
baselines truly does imply the presence of a disc-like structure, we generated
synthetic interferometric observations of an alternative version of model 1
with the disc removed and the core inner radius set to 1 AU rather than the
outer radius of the disc, which is generally 100 AU or larger (see
Figure \ref{fig_models}). The resulting disc masses calculated from the
amplitudes at 50 \kl\ for this model ranged from 5 $\times 10^{-3}$ \msun\
early in the Class 0 stage to less than $10^{-3}$ \msun\ late in the Class I
stage. As most observations performed to date suggest discs with masses
greater than $10^{-2}$ \msun\ (see Figure \ref{fig_mdisc_histogram}),
even cores with such extreme inner radii
have a negligible effect on the observed disc mass distribution derived
assuming all emission on 50 \kl\ baselines arises from discs. These
results are in general agreement with those of \citet{jorgensen2009:prosac},
who also found that cores do not contribute significantly to the total
emission at 50 \kl, even in cases where the inner radii are as small as 25 AU.
They are also in agreement with earlier work that arrived at
similar conclusions using much simpler models
\citep[e.g.,][]{terebey1993:disks}. Non-magnetized cores simply do not
contain enough mass on the small spatial scales probed by 50 \kl\ baselines.
While these results demonstrate that the structures probed by interferometer
observations at baselines of 50 \kl\ can not be explained as simply the inner,
dense parts of cores, they still do not prove that the structures are due to
rotationally supported, Keplerian discs. Magnetically supported
``pseudodiscs'' form on size scales of a few hundred to a few thousand AU
in magnetized cores \citep[e.g.,][]{galli1993:bfields,chiang2008:pseudodisks},
and unresolved
continuum observations are unable to distinguish between the two types of
objects.
We emphasize that the observed disc mass distribution
assembled using the methods investigated in this study must be confirmed with
detections of Keplerian rotation signatures in molecular line observations.
While such observations are now possible
\citep[e.g.,][]{jorgensen2009:prosac,tobin2012:l1527,murillo2013:vla1623a,yen2013:discs,harsono2014:discs}, they push existing facilities such as the
Submillimeter Array (SMA), Combined Array for Research in Millimeter-wave
Astronomy (CARMA), and Plateau de Bure Interferometer (PdBI)
to their limits and are only possible for the brightest and most massive
sources, especially in the Class 0 stage. However, they should become routine
in the near future with the technical capabilities of ALMA and the Northern
Extended Millimeter Array (NOEMA). Confirming Keplerian rotation and
measuring the radii out to which such motions are detected is a high priority
for future observations.
\section{Summary}\label{sec_summary}
In this paper we have coupled hydrodynamical simulations of the collapse of
dense, rotating cores into protostars with radiative transfer
evolutionary models to
generate synthetic observations as the simulations evolve. We have used
this coupling to investigate the extent to which a simple method for
measuring protostellar disc masses, namely the assumption that the
(sub)millimeter interferometer amplitude measured at a baseline of 50 \kl\
arises from optically thin, 30 K thermal dust emission in the disc with
no contamination from the surrounding core, recovers the intrinsic masses of
the discs formed in the simulations. We summarize our main results as follows:
\begin{itemize}
\item Overestimates due to contamination from the surrounding core and
underestimates due to partially resolving out the disc tend to cancel each
other at millimeter wavelengths but can lead to net underestimates by up to
factors of 4 at submillimeter wavelengths.
\item Optically thick emission, colder dust temperatures than the assumed
$T_D = 30$ K, and the effects of inclination can lead to underestimates by
up to factors of 2--3 at millimeter wavelengths and up to an order of magnitude
or larger at submillimeter wavelengths, especially 350 \um.
\item Flatter dust opacity laws due to grain growth will lead to further
disc mass underestimates at the shortest submillimeter wavelengths and
overestimates by up to factors of 10 or larger at the longest millimeter
wavelengths (3 mm or longer), with the exact magnitude dependent on the
details of the true opacity law in the disc.
\item The optically thin portions of protostellar discs are generally cooler
in the Class I stage than the Class 0 stage since Class I discs are typically
larger and more optically thick, and thus more shielded. Protostellar disc mass
distributions assembled from observations where the dust temperature is fixed
at 30 K independent of evolutionary stage will be lower limits to the
intrinsic distributions, with mean values lower by factors of $1.5-6$
depending on wavelength and intrinsic disc properties.
If temperatures of 30 K for Class 0 sources and 15 K for Class I sources are
adopted, the observed distributions will be more representative of the
intrinsic distributions, with mean values within 50\%.
\item Comparing the observed disc mass distribution from three different
surveys at 1 mm to that assembled from
the \citet{vorobyov2010:bursts} hydrodynamical simulations suggests that
the disc masses predicted by the simulations are observed in
approximatley 50\% -- 70\% of all protostars.
These results suggest that at least some protostellar discs are likely
sufficiently massive to fragment, although larger samples are needed to
further quantify this result.
\item We make the following two recommendations for future studies that seek
to use interferometer continuum observations of protostars to measure the
protostellar disc mass distribution:
\begin{enumerate}
\item Observe at a wavelength of approximately 1 mm.
\item Adopt a variable dust temperature of 30 K for Class 0 sources and 15 K
for Class I sources.
\end{enumerate}
\end{itemize}
These results must be confirmed with larger, less biased observational
samples. Additionally, this method is not capable of distinguishing between
rotataionlly supported, Keplerian discs and magnetically supported
``pseudodiscs'', thus future molecular line observations are critically needed
to confirm Keplerian rotation in protostars found to have discs with the
method investigated here. However, the current evidence does support the
hypothesis that disc fragmentation may play a significant role in the early
evolution of protostellar systems.
We thank the anonymous referee for comments and suggestions
that have improved the quality of this work.
This research has made use of NASA's Astrophysics Data System (ADS)
Abstract Service and the IDL Astronomy Library hosted by the NASA Goddard Space
Flight Center. MMD acknowledges support from the Submillimeter Array through
an SMA postdoctoral fellow. The simulations were performed
on the Shared Hierarchical Academic Research Computing Network
(SHARCNET), on the Atlantic Computational Excellence Network (ACEnet),
and on the Vienna Scientific Cluster (VSC-2).
This project was partly supported by the Russian Ministry of Education and
Science Grant (state assignment) 3.961.2014/K.
HGA acknowledges support from the NSF through grant AST-0845619.
\bibliographystyle{mn2e.bst}
|
\section{Introduction}
In this work, we are interested in solving numerically the Cahn-Hilliard-Navier-Stokes (CHNS) phase field model that describes the interface dynamics of a binary incompressible and macroscopically immiscible Newtonian fluids with matched density and viscosity in a bounded domain $\Omega \subseteq \mathbb{R}^d, d=2, 3$. The non-dimensional system takes the explicit form as, cf. \cite{KKL2004}
\begin{align}
&\phi_t+ \nabla \cdot (\phi \mathbf{u})=\nabla \cdot(M(\phi) \nabla \mu), \quad \text{ in } \Omega_T \label{CH} \\
&\mu=f_0^\prime(\phi)-\epsilon^2 \Delta \phi, \quad \text{ in } \Omega_T \label{CP} \\
&\mathbf{u}_t-\frac{1}{Re} \Delta \mathbf{u}+\mathbf{u}\cdot \nabla \mathbf{u}+\nabla p=-\frac{\epsilon^{-1}}{{We}^\ast} \phi \nabla \mu, \quad \text{in } \Omega_T \label{NS}\\
&\nabla \cdot \mathbf{u}=0, \quad \text{in } \Omega_T \label{div}
\end{align}
where $\mathbf{u}$ is the velocity field, $p$ is a modified pressure, $\phi $ is the phase field variable (order parameter), $\mu$ the chemical potential, $f_0(\phi)$ is the quartic homogeneous free energy density function
$
f_0(\phi)=\frac{1}{4}(1-\phi^2)^2,
$
and $\Omega_T:=\Omega \times (0,T)$ with $T>0$ a fixed constant.
$Re$ is the Reynolds number; ${We}^\ast$ is the modified Weber number that measures the relative strengths of the kinetic and surface energies \cite{KKL2004}; $\epsilon$ is a dimensionless parameter that measures capillary width of the diffuse interface; $M(\phi)$ is the mobility function that incorporates the diffusional Peclet number $Pe$. We refer to \cite{KAD2007, LoTr1998} for the detailed non-dimensionalization of the CHNS system.
We close the system with the following initial and boundary conditions
\begin{align}\label{BCs}
\mathbf{u}&=0, \quad \text{on } \partial \Omega \times (0,T) \\
\nabla \phi \cdot \mathbf{n}=\nabla \mu \cdot \mathbf{n} &=0, \quad \text{on } \partial \Omega \times (0,T) \\
(\mathbf{u}, \phi)|_{t=0}&=(\mathbf{u}_0, \phi_0), \quad \text{in } \Omega.
\end{align}
Here $\mathbf{n}$ denotes the unit outer normal vector of the boundary $\partial \Omega$.
It is clear that the CHNS system \eqref{CH}-\eqref{div} under the above boundary conditions is mass-conservative,
\begin{align}\label{Mass-c}
\frac{d}{dt}\int_\Omega \phi\, dx=0,
\end{align}
and energy-dissipative
\begin{align}\label{ConEnL}
\frac{d}{dt}E_{tot}(\mathbf{u}, \phi)=-\frac{1}{Re}\int_\Omega |\nabla \mathbf{u}|^2\,dx -\frac{\epsilon^{-1}}{{We}^\ast} \int_\Omega M(\phi)|\nabla \mu|^2\, dx,
\end{align}
where the total energy $E_{tot}$ is defined as
\begin{align}\label{Etot}
E_{tot}(\mathbf{u}, \phi)=\int_{\Omega}\frac{1}{2}|\mathbf{u}|^2\, dx+ \frac{1}{{We}^\ast}\int_{\Omega} \big( \frac{1}{\epsilon} f_0(\phi)+\frac{\epsilon}{2}|\nabla \phi|^2\big)\, dx.
\end{align}
The first term on the right hand side of equation \eqref{Etot} is the total kinetic energy, and the term, denoted by $E_{f}$ throughout, is a measure of the surface energy of the fluid system.
The CHNS phase field model \eqref{CH}-\eqref{div} is proposed
as an alternative of sharp interface model to describe the dynamics of two phase, incompressible, and macroscopically immiscible Newtonian fluids with matched density, cf. \cite{HoHa1977, GPV1996, AMW1998, LoTr1998, LiSh2003}.
In contrast to the sharp interface model, the diffuse interface model recognizes the micro-scale mixing and hence treats the interface of two fluids as a transition layer with small but non-zero width $\epsilon$. Although the region is thin, it may play an important role during topological transition like interface pinchoff or reconnection \cite{LoTr1998}. One then introduces an order parameter $\phi$, for instance the concentration difference, which takes the value $1$ in the bulk of one fluid and $-1$ in regions filled by the other fluid and varies continuously between $1$ and $-1$ over the interfacial region. One can view the zero level set of the order parameter as the averaged interface of the mixture. Thus, the dynamics of the interface can be simulated on a fixed grid without explicit interface tracking, which renders the diffuse interface method an attractive numerical approach for deforming interface problems. The CHNS diffuse interface model has been successfully employed for the simulations of two-phase flow in various contexts. We refer the readers to \cite{AMW1998, KKL2004} and references therein for its diverse applications.
In this work, we assume that $m_1 \leq M(\phi) \leq m_2$ for constants $0<m_1 \leq m_2$. We point out that the degenerate mobility function may be more physically relevant, as it guarantees the order parameter stays within the physical bound $\phi \in [-1, 1]$ \cite{Boyer1999}, though uniqueness of weak solutions is still open even for the Cahn-Hilliard equation. Recent numerical experiments \cite{ZhWa2010} also indicate that the Cahn-Hilliard equation with degenerate mobility may be more accurate for immiscible binary fluids. Numerical resolution of the degenerate case is a subtle matter and beyond the scope of our current work (cf. \cite{BBG1999, CeGa2013} for the case of Cahn-Hilliard equation).
There are several challenges in solving the system \eqref{CH}-\eqref{div} numerically. First of all, the small interfacial width $\epsilon$ introduces tremendous amount of stiffness into the system (large spatial derivative within the interfacial region). It demands the numerical scheme to be unconditionally stable so that the stiffness can be handled
with ease. The resulting numerical scheme tends to be nonlinear and therefore poses challenge in proving unconditionally unique solvability. A popular strategy in discretizing the Cahn-Hilliard equation (Eqs \eqref{CH}--\eqref{CP}) in time is based on the convex-splitting of the free energy functional $E_f$, i.e., treating the convex part of the functional implicitly and concave part explicitly, an idea dates back to Eyre \cite{Eyre1998}. The design of convex-splitting scheme yields not only unconditional stability, but also unconditionally unique solvability for systems with symmetric structures\cite{Wise2010, CSW2013}. However, the variational approach for proving unique solvability (see the references above) is not applicable to the CHNS system since the advection term in Navier-Stokes equation (Eq.\eqref{NS}) breaks the symmetry. In addition, the stiffness issue naturally requires adaptive mesh refinement in order to reduce the computational cost. Secondly, when it comes to solving the Navier-Stokes equation, one always faces the difficulty of the coupling between velocity and pressure. The common practice is to use the well-known Chorin-Temam type pressure projection scheme, see \cite{GMS2006} for a general review. Lastly, higher order scheme is always preferable from the accuracy point-of-view. Yet, it is a challenge to design higher order scheme for a nonlinear system while maintaining the unconditional stability.
There have been many works on the numerical resolution of the CHNS system, see a comprehensive summary by Shen \cite{Shen2012}. Here we survey several papers that are especially relevant to ours. In \cite{KKL2004}, Kim, Kang and Lowengrub proposed a conservative, second-order accurate fully implicit discretization of the CHNS system. The update of the pressure in the Navier-Stokes equation is based on an approximate pressure projection method. To ensure the unconditional stability, they introduce a non-linear stabilization term to the Navier-Stokes solver. The scheme is strongly coupled and highly nonlinear, for which they design a multigrid iterative solver. The authors point out (without proof) that a restriction on the time-step size may be needed for the unique solvability of the scheme. In \cite{Feng2006}, Feng analyses a first-order in time, fully discrete finite element approximation of the CHNS system. He shows that his scheme is unconditionally energy-stable and convergent, but gives no analysis on unique solvability.
Kay, Styles and Welford \cite{KSW2008} also studied a first-order in time, finite element approximation of CHNS system. In contrast to Feng's scheme, the velocity in the Cahn-Hilliard equation \eqref{CH} is discretized explicitly at the discrete time level. Thus the computation of the Cahn-Hilliard equation is fully decoupled from that of Navier-Stokes equation. Moreover, the unique solvability of the overall scheme can be established easily by exploring the gradient flow structure of the Cahn-Hilliard equation. However, a CFL condition has to be imposed for the scheme to be stable. See \cite{Minjeaud2013} for an operator-splitting strategy in decoupling the computation of Cahn-Hilliard equation and Navier-Stokes equation which still preserves the unconditional stability (without decoupling the pressure and velocity). Dong and Shen \cite{DoSh2012} recently derived a fully decoupled linear time stepping scheme for the CHNS system with variable density, which involves only constant matrices for all flow variables.
However, there is no stability analysis on their numerical scheme.
In this paper, we propose a novel second order in time numerical scheme for Cahn-Hilliard-Navier-Stokes phase field model with matched density. The scheme is based on second order convex-splitting for the Cahn-Hilliard equation and pressure-projection for the Navier-Stokes equation. This scheme satisfies a modified energy law which mimics the continuous version of the energy law \eqref{ConEnL}, and is therefore unconditionally stable. Moreover, we prove that the scheme is unconditionally uniquely solvable at each time step by exploring the monotonicity associated with the scheme. Thanks to the weak coupling of the scheme, we design an efficient Picard iteration procedure to further decouple the computation of Cahn-Hilliard equation and Navier-Stokes equation. We implement the scheme by the mixed finite element method. Ample numerical experiments are performed to validate the accuracy and efficiency of the numerical scheme.
The possibility of such a scheme is alluded in Remark 5.5 \cite{Shen2012}. A similar scheme without pressure-correction for Cahn-Hilliard-Brinkman equation is proposed in the concluding remarks of \cite{CSW2013}.
The rest of the paper is organized as follows. In section 2, we give the discrete time, continuous space scheme. We prove the mass-conservation, unconditional stability and unconditionally unique solvability in section 3. In section 4, the scheme is further discretized in space by mixed finite element approximation. An efficient Picard iteration procedure is proposed to solve the fully discrete equations. Finally, We provide some numerical experiments in section 5 to validate our numerical scheme.
\section{A Discrete Time, Continuous Space Scheme}
Let $\delta t >0$ be a time step size and set $t^k=k\delta t$ for $0\leq k \leq K=[T/\delta t]$. Without ambiguity, we denote by $(f,g)$ the $L^2$ inner product between functions $f$ and $g$. Also for convenience, the following notations will be used throughout this paper
\begin{subequations}\label{notas}
\begin{align}
&\phi^{k+\frac{1}{2}}=\frac{1}{2}(\phi^{k+1}+\phi^k), \quad \widetilde{\phi}^{k+\frac{1}{2}}=\frac{3\phi^k-\phi^{k-1}}{2}, \\
&\overline{\mathbf{u}}^{k+\frac{1}{2}}=\frac{\overline{\mathbf{u}}^{k+1}+\mathbf{u}^k}{2}, \quad \widetilde{\mathbf{u}}^{k+\frac{1}{2}}=\frac{3\mathbf{u}^k-\mathbf{u}^{k-1}}{2}.
\end{align}
\end{subequations}
We propose the semi-implicit, semi-discrete scheme in strong form as follows:
\begin{align}
&\frac{\phi^{k+1}-\phi^k}{\delta t}=\nabla \cdot \big(M(\widetilde{\phi}^{k+\frac{1}{2}})\nabla \mu^{k+\frac{1}{2}}-\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}^{k+\frac{1}{2}}\big),\label{2ndCHNSa}\\
&\mu^{k+\frac{1}{2}}=\frac{1}{2}\big((\phi^{k+1})^2+(\phi^k)^2\big)\phi^{k+\frac{1}{2}}-\widetilde{\phi}^{k+\frac{1}{2}}-\epsilon^2 \Delta \phi^{k+\frac{1}{2}}, \label{2ndCHNSb}\\
&\frac{\overline{\mathbf{u}}^{k+1}-\mathbf{u}^k}{\delta t}-\frac{1}{Re}\Delta \overline{\mathbf{u}}^{k+\frac{1}{2}}+B(\widetilde{\mathbf{u}}^{k+\frac{1}{2}}, \overline{\mathbf{u}}^{k+\frac{1}{2}}) =-\nabla p^k-\frac{\epsilon^{-1}}{{We}^\ast}\widetilde{\phi}^{k+\frac{1}{2}} \nabla \mu^{k+\frac{1}{2}},\label{2ndCHNSc}\\
&\left\{
\begin{aligned}
&\frac{\mathbf{u}^{k+1}-\overline{\mathbf{u}}^{k+1}}{\delta t} + \frac{1}{2}\nabla(p^{k+1}-p^k)=0, \\
&\nabla \cdot \mathbf{u}^{k+1}=0,
\end{aligned} \label{2ndCHNSd}
\right.
\end{align}
with boundary conditions
\begin{align}\label{2ndCHNSe}
&\nabla \phi^{k+1} \cdot \mathbf{n}|_{\partial \Omega}=0, \quad \nabla \mu^{k+\frac{1}{2}}\cdot \mathbf{n}|_{\partial \Omega}=0,\quad \overline{\mathbf{u}}^{k+\frac{1}{2}}|_{\partial \Omega}=0, \quad \mathbf{u}^{k+1} \cdot \mathbf{n} |_{\partial \Omega}=0.
\end{align}
Here $B(\mathbf{u}, \mathbf{v}):=(\mathbf{u}\cdot \nabla)\mathbf{v}+\frac{1}{2}(\nabla \cdot \mathbf{u})\mathbf{v}$ is the skew-symmetric form of the nonlinear advection term in the Navier-Stokes equation \eqref{2ndCHNSc}, which is first introduced by Temam \cite{Temam1969}. In the space continuous level, $\nabla \cdot \widetilde{\mathbf{u}}^{k+\frac{1}{2}}=0$, thus $B(\widetilde{\mathbf{u}}^{k+\frac{1}{2}}, \overline{\mathbf{u}}^{k+\frac{1}{2}})=\widetilde{\mathbf{u}}^{k+\frac{1}{2}} \cdot \nabla \overline{\mathbf{u}}^{k+\frac{1}{2}} $, which amounts to a second order semi-implicit discretization of the advection term. The skew symmetric form $B(\mathbf{u}, \mathbf{v})$ induces a trilinear form $b$ defined as, $\forall \mathbf{u}, \mathbf{v}, \mathbf{w} \in \mathbf{H}^1_0(\Omega)$
\begin{align}\label{trilinear}
b(\mathbf{u},\mathbf{v},\mathbf{w})=(B(\mathbf{u}, \mathbf{v}), \mathbf{w})=\frac{1}{2}\{(\mathbf{u}\cdot \nabla \mathbf{v}, \mathbf{w})-(\mathbf{u}\cdot \nabla \mathbf{w}, \mathbf{v})\}.
\end{align}
It follows immediately that $b(\mathbf{u},\mathbf{v},\mathbf{v})=0$ for any $\mathbf{u}, \mathbf{v} \in \mathbf{H}^1_0 (\Omega)$. This skew symmetry holds regardless of whether $\mathbf{u}, \mathbf{v}$ are divergence-free or not, which would help to preserve the stability when the scheme is further discretized in space.
The overall scheme \eqref{2ndCHNSa}--\eqref{2ndCHNSd} is based on the Crank-Nicolson time discretization and the second order Adams-Bashforth extrapolation. We note that the term $\frac{1}{2}\big((\phi^{k+1})^2+(\phi^k)^2\big)\phi^{k+\frac{1}{2}}-\widetilde{\phi}^{k+\frac{1}{2}}$ from the chemical potential equation \eqref{2ndCHNSb} is a second order approximation of the nonlinear term $f_0^\prime(\phi)$ (Eq.\eqref{CP}), which is derived according to a convex-splitting of the free energy density function $f_0(\phi)$. To see this, we rewrite $f_0(\phi)$ as the sum of a convex function and a concave function
$$
f_0(\phi)=f_v(\phi)+f_c(\phi):= \frac{1}{4}\phi^4+\big(-\frac{1}{2}\phi^2+\frac{1}{4}\big),
$$
and accordingly $f_0^\prime(\phi)=f_v^\prime(\phi)+f_c^\prime(\phi)$. The idea of convex-splitting is to use explicit discretization for the concave part (i.e. $f_c^\prime(\widetilde{\phi}^{k+\frac{1}{2}})$) and implicit discretization for the convex part. Thus we approximate $f_v^\prime(\phi^{k+\frac{1}{2}})$ by the Crank-Nicolson scheme
$$
f_v^\prime(\phi^{k+\frac{1}{2}})\approx \frac{f_v(\phi^{k+1})-f_v(\phi^k)}{\phi^{k+1}-\phi^k}=\frac{1}{2}[(\phi^{k+1})^2+(\phi^k)^2]\phi^{k+\frac{1}{2}}.
$$
Such a second order convex-splitting scheme is originally proposed and analysed in \cite{HWWL2009, BLWW2013} in the context of phase field crystal equation, see also \cite{SWWW2012} for applications in thin film epitaxy. We point out one can also approximate $f_0^\prime (\phi^{k+\frac{1}{2}})$ directly by Crank-Nicolson scheme \cite{KKL2004, GoHu2011} which would yield unconditional stability. The design of convex-splitting scheme enables us to prove not only unconditional stability but also unconditionally unique solvability of the overall scheme.
Eqs. \eqref{2ndCHNSc} and \eqref{2ndCHNSd} comprise the second order incremental pressure projection method
of Van Kan type \cite{vanKan1986} with linear extrapolation for the nonlinear advection term. The viscous step (Eq. \eqref{2ndCHNSc}) solves for an intermediate velocity $\overline{\mathbf{u}}^{k+1}$ (or, equivalently $\overline{\mathbf{u}}^{k+\frac{1}{2}}$) which is not divergence-free. The projection step (Eq. \eqref{2ndCHNSd}) is amount to \linebreak $\mathbf{u}^{k+1}=P_{\mathbf{H}} \overline{\mathbf{u}}^{k+1}$, where $P_{\mathbf{H}}$ is the Leray projection operator into $\mathbf{H}$:
\begin{align}
\mathbf{H}:=\{\mathbf{v} \in \mathbf{L}^2(\Omega); \nabla \cdot \mathbf{v}=0; \mathbf{v} \cdot \mathbf{n}|_{\partial \Omega}=0\}. \nonumber
\end{align}
The projection equation \eqref{2ndCHNSd} can also be solved in two sub-steps: first through a Pressure Poisson equation for the pressure increment
\begin{equation} \label{PrePoi}
\left \{
\begin{aligned}
&\Delta (p^{k+1}-p^k)=\frac{2}{\delta t}\nabla \cdot \widetilde{\mathbf{u}}^{k+1}, \\
&\nabla (p^{k+1}-p^k) \cdot \mathbf{n}|_{\partial \Omega}=0.
\end{aligned}
\right.
\end{equation}
and then by an algebraic update for velocity
\begin{align} \label{AupVel}
&\mathbf{u}^{k+1}=\widetilde{\mathbf{u}}^{k+1}-\frac{\delta t}{2} \nabla (p^{k+1}-p^k).
\end{align}
Variants of such a splitting method are analyzed in \cite{Shen1996} where it is shown (discrete time, continuous space) that the schemes are second order accurate for velocity in $l^2(0, T; L^2(\Omega))$ but only first order accurate for pressure in $l^\infty(0,T; L^2(\Omega))$. The loss of accuracy for pressure is due to the artificial boundary condition (cf. Eq. \eqref{PrePoi}) imposed on pressure \cite{ELi1995}. We also remark that the Crank-Nicolson scheme with linear extrapolation is a popular time discretization for the Navier-Stokes equation. We refer to \cite{Ingram2013} and references therein for analysis on this type of discretization.
Note that the projection step (Eq. \eqref{2ndCHNSd}) is decoupled from the rest of the equations. Moreover, the coupling between Eqs. \eqref{2ndCHNSa}-\eqref{2ndCHNSc} is fairly weak, thanks to the semi-implicit discretization. We see that the Cahn-Hilliard equation \eqref{2ndCHNSa} and \eqref{2ndCHNSb} is coupled with the Navier-Stokes equation \eqref{2ndCHNSa} only through the velocity $\overline{\mathbf{u}}^{k+\frac{1}{2}}$ in the advection term of Eq. \eqref{2ndCHNSa} and the chemical potential $\mu^{k+\frac{1}{2}}$ in the elastic forcing term of Eq. \eqref{2ndCHNSc}. On the one hand, this allows us to use a Picard iteration procedure on velocity to further decouple the computation of the nonlinear Cahn-Hilliard equation from the linear Navier-Stokes equation, see Section 4 for details. On the other hand, owing to the special design, we are able to show the unconditionally unique solvability of the system \eqref{2ndCHNSa}-\eqref{2ndCHNSc} by a monotonicity argument (cf. Section 3). In fact, one can define a solution operator $\phi^{k+1}(\mu^{k+\frac{1}{2}}): \mu^{k+\frac{1}{2}} \rightarrow \phi^{k+1}$ from equation \eqref{2ndCHNSb}. Likewise, equation \eqref{2ndCHNSc} gives rise to a solution operator $\overline{\mathbf{u}}^{k+\frac{1}{2}}(\mu^{k+\frac{1}{2}}): \mu^{k+\frac{1}{2}} \rightarrow \overline{\mathbf{u}}^{k+\frac{1}{2}}$. As a result, the system \eqref{2ndCHNSa}-\eqref{2ndCHNSc} reduces to a scalar equation in terms of the unknown $\mu^{k+\frac{1}{2}}$
\begin{align*}
\phi^{k+1}(\mu^{k+\frac{1}{2}})-\phi^k + \delta t\nabla \cdot \big(\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}^{k+\frac{1}{2}}(\mu^{k+\frac{1}{2}})\big) -\delta t\nabla \cdot \big(M(\widetilde{\phi}^{k+\frac{1}{2}})\nabla \mu^{k+\frac{1}{2}} \big)=0.
\end{align*}
The key here is to recognize that the left-hand side of the above equation defines a strictly monotone operator $T(\mu)$, in the sense that
\begin{align*}
\left\langle T(\mu)-T(\nu), \mu-\nu \right\rangle \geq 0,
\end{align*}
with equal sign if and only if $\mu=\nu$. Thus one can invoke the Browder-Minty Lemma \ref{BrowderMinty} (see Section 3) to prove the unique existence of such a solution $\mu^{k+\frac{1}{2}}$. We remark that the variational approach \cite{Wise2010, CSW2013} is not directly applicable for the unique solvability of the Cahn-Hilliard-Navier-Stokes system \eqref{2ndCHNSa}-\eqref{2ndCHNSd}. In both cases (Cahn-Hilliard-Hele-Shaw, Cahn-Hilliard-Brinkman), the approach relies on the symmetry of the underlying systems which breaks down in the Navier-Stokes equation due to the nonlinear advection.
\section{Properties of the scheme}
In this section, we summarize the properties of the discrete time, continuous space scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSe}, namely mass-conservation, unconditional stability and unconditionally unique solvability. It will be clear from the proof, that these properties will be preserved when the scheme is combined with any consistent Galerkin type spatial discretization schemes.
First of all, one can readily obtain that the scheme is mass-conservative.
\begin{proposition}
The scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSd} equipped with the boundary condition \eqref{2ndCHNSe} satisfies the mass-conservation, i.e.,
\begin{align*}
\int_{\Omega}\phi^{k+1}dx =\int_{\Omega}\phi^k dx, \quad k=0, 1, \cdots K-1.
\end{align*}
\end{proposition}
Next, we show that our numerical scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSe} is unconditionally stable, thus allowing for large time stepping. Recall the definition of the total energy functional $E_{tot}(\mathbf{u}, \phi)$ in Eq. \eqref{Etot}.
\begin{proposition}
The scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSd} with the boundary condition \eqref{2ndCHNSe} satisfies the modified energy law
\begin{align}\label{ModEnergyLaw}
&\Big\{E_{tot}(\mathbf{u}^{k+1}, \phi^{k+1})+\frac{\epsilon^{-1}}{4{We}^\ast}||\phi^{k+1}-\phi^k||_{L^2}^2+\frac{\delta t^2}{8}||\nabla p^{k+1}||_{L^2}^2\Big\} \nonumber\\
& -\Big\{E_{tot}(\mathbf{u}^{k}, \phi^{k})+\frac{\epsilon^{-1}}{4{We}^\ast}||\phi^{k}-\phi^{k-1}||_{L^2}^2+\frac{\delta t^2}{8}||\nabla p^{k}||_{L^2}^2\Big\}\nonumber \\
&=-\delta t\frac{\epsilon^{-1}}{{We}^\ast}||\sqrt{M}\nabla \mu^{k+\frac{1}{2}}||_{L^2}^2-\delta t \frac{1}{Re}||\nabla \overline{\mathbf{u}}^{k+\frac{1}{2}}||_{L^2}^2-\frac{\epsilon^{-1}}{4{We}^\ast}||\phi^{k+1}-2\phi^k+\phi^{k-1}||_{L^2}^2.
\end{align}
Thus it is \textbf{unconditionally stable}.
\end{proposition}
\begin{proof}
One first takes the $L^2$ inner product of Eq. \eqref{2ndCHNSa} with $\delta t \mu^{k+\frac{1}{2}}$ to obtain
\begin{align}\label{2ndCH1}
&\big(\phi^{k+1}-\phi^k, \mu^{k+\frac{1}{2}} \big)=-\delta t||\sqrt{M}\nabla \mu^{k+\frac{1}{2}}||_{L^2}^2+\delta t\big(\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}^{k+\frac{1}{2}},\nabla \mu^{k+\frac{1}{2}}\big).
\end{align}
Next, multiplying Eq. \eqref{2ndCHNSb} by $(\phi^{k+1}-\phi^k)$, performing integration by parts and using the the following identity
\begin{eqnarray}
& &\big(\widetilde{\phi}^{k+\frac{1}{2}},\phi^{k+1}-\phi^k\big)\nonumber \\
&=& \frac{1}{2}\big(3\phi^k-\phi^{k-1},\phi^{k+1}-\phi^k\big) \nonumber \\
&=&\frac{1}{2}\big(\phi^{k+1}+\phi^k,\phi^{k+1}-\phi^k\big) -\frac{1}{2}\big(\phi^{k+1}-2\phi^k+\phi^{k-1},\phi^{k+1}-\phi^k\big)\nonumber \\
&=&\frac{1}{2}\big(||\phi^{k+1}||_{L^2}^2-||\phi^k||_{L^2}^2\big)-\frac{1}{4}\{||\phi^{k+1}-\phi^k||_{L^2}^2-||\phi^k-\phi^{k-1}||_{L^2}^2\nonumber \\
& & +||\phi^{k+1}-2\phi^k+\phi^{k-1}||_{L^2}^2\}, \nonumber
\end{eqnarray}
one deduces
\begin{align}\label{2ndCH2}
&-\big(\phi^{k+1}-\phi^k, \mu^{k+\frac{1}{2}}\big)+\big(f_0(\phi^{k+1})-f_0(\phi^k),1\big)+\frac{\epsilon^2}{2}(||\nabla \phi^{k+1}||_{L^2}^2-||\nabla \phi^k||_{L^2}^2)\nonumber \\
&+\frac{1}{4}\{||\phi^{k+1}-\phi^k||_{L^2}^2-||\phi^{k}-\phi^{k-1}||_{L^2}^2+||\phi^{k+1}-2\phi^k+\phi^{k-1}||_{L^2}^2\}=0,
\end{align}
where one has utilized the definition of $f_0(\phi)=\frac{1}{4}(\phi^2-1)^2$.
Summing up Eqs. \eqref{2ndCH1} and \eqref{2ndCH2} gives
\begin{align}\label{2ndCHf}
&\big(f_0(\phi^{k+1})-f_0(\phi^k),1\big)+\frac{\epsilon^2}{2}(||\nabla \phi^{k+1}||_{L^2}^2-||\nabla \phi^k||_{L^2}^2)+\frac{1}{4}(||\phi^{k+1}-\phi^k||_{L^2}^2-||\phi^{k}-\phi^{k-1}||_{L^2}^2)\nonumber \\
&=-\frac{1}{4}||\phi^{k+1}-2\phi^k+\phi^{k-1}||_{L^2}^2-\delta t||\sqrt{M}\nabla \mu^{k+\frac{1}{2}}||_{L^2}^2+\delta t\big(\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}^{k+\frac{1}{2}},\nabla \mu^{k+\frac{1}{2}}\big),
\end{align}
Now we turn to the Navier-Stokes part. Taking the $L^2$ inner product of Eq. \eqref{2ndCHNSc} with $\overline{\mathbf{u}}^{k+\frac{1}{2}}\delta t$ and using the skew-symmetry of the trilinear form $b$ in \eqref{trilinear}, one gets
\begin{align}\label{2ndNS1}
&\frac{1}{2}(||\overline{\mathbf{u}}^{k+1}||_{L^2}^2-||\mathbf{u}^k||_{L^2}^2)+\delta t \frac{1}{Re}||\nabla \overline{\mathbf{u}}^{k+\frac{1}{2}}||_{L^2}^2
=-\delta t \big(\nabla p^k, \overline{\mathbf{u}}^{k+\frac{1}{2}}\big)-\delta t \frac{\epsilon^{-1}}{{We}^\ast}\big(\widetilde{\phi}^{k+\frac{1}{2}} \nabla \mu^{k+\frac{1}{2}}, \overline{\mathbf{u}}^{k+\frac{1}{2}}\big).
\end{align}
Testing the first equation in \eqref{2ndCHNSd} by $\mathbf{u}^{k+1} \delta t$ and performing integration by parts yield
\begin{align}\label{2ndNS2}
&\frac{1}{2}(||\mathbf{u}^{k+1}||_{L^2}^2-||\overline{\mathbf{u}}^{k+1}||_{L^2}^2+||\mathbf{u}^{k+1}-\overline{\mathbf{u}}^{k+1}||_{L^2}^2)=0,
\end{align}
where one has utilized explicitly the divergence-free condition $\nabla \cdot \mathbf{u}^{k+1}=0$.
Next, we rewrite the projection step Eq. \eqref{2ndCHNSd} as
\begin{align*}
\frac{\mathbf{u}^{k+1}+\mathbf{u}^k-2\overline{\mathbf{u}}^{k+\frac{1}{2}}}{\delta t} + \frac{1}{2}\nabla(p^{k+1}-p^k)=0.
\end{align*}
Testing the above equation with $\frac{\delta t^2}{2}\nabla p^k$, one arrives at
\begin{align}\label{2ndNS3}
\frac{\delta t^2}{8}\big\{||\nabla p^{k+1}||_{L^2}^2-||\nabla p^{k}||_{L^2}^2-||\nabla(p^{k+1}-p^k)||_{L^2}^2\big\}=\delta t\big(\nabla p^k, \overline{\mathbf{u}}^{k+\frac{1}{2}}\big).
\end{align}
On the other hand, it follows directly from Eq. \eqref{2ndCHNSd} that
\begin{align}\label{2ndNS4}
\frac{{\delta t}^2}{8}||\nabla(p^{k+1}-p^k)||_{L^2}^2=\frac{1}{2}||\mathbf{u}^{k+1}-\overline{\mathbf{u}}^{k+1}||_{L^2}^2.
\end{align}
Now summing up Eqs. \eqref{2ndNS1}-\eqref{2ndNS3} and in view of Eq. \eqref{2ndNS4}, one obtains
\begin{align}\label{2ndNSf}
&\frac{1}{2}(||\mathbf{u}^{k+1}||_{L^2}^2-||\mathbf{u}^k||_{L^2}^2)+\frac{\delta t^2}{8}\big\{||\nabla p^{k+1}||_{L^2}^2-||\nabla p^{k}||_{L^2}^2\big\}
\nonumber \\
&=-\delta t \frac{1}{Re}||\nabla \overline{\mathbf{u}}^{k+\frac{1}{2}}||_{L^2}^2-\delta t \frac{\epsilon^{-1}}{{We}^\ast}\big(\widetilde{\phi}^{k+\frac{1}{2}} \nabla \mu^{k+\frac{1}{2}}, \overline{\mathbf{u}}^{k+\frac{1}{2}}\big).
\end{align}
The energy law \eqref{ModEnergyLaw} then follows from summing up the multiple of Eq. \eqref{2ndCHf} by $\frac{\epsilon^{-1}}{{We}^\ast}$ and Eq. \eqref{2ndNSf}.
\end{proof}
\begin{remark}
Heuristically, $E_{tot}(\mathbf{u}^{k+1}, \phi^{k+1})+\frac{\widetilde{W}e^{-1}}{4\epsilon}||\phi^{k+1}-\phi^k||_{L^2}^2+\frac{\delta t^2}{8}||\nabla p^{k+1}||_{L^2}^2$ is a second order approximation of $E_{tot}(\mathbf{u}^{k+1}, \phi^{k+1})$, as one can write
\begin{align*}
||\phi^{k+1}-\phi^k||_{L^2}^2=\delta t^2||(\phi^{k+1}-\phi^k)/\delta t||_{L^2}^2,
\end{align*}
and $(\phi^{k+1}-\phi^k)/\delta t$ is an approximation of $\phi_t$ at $t^{k+1}$.
\end{remark}
To prove the unconditionally unique solvability of Eqs. \eqref{2ndCHNSa}-\eqref{2ndCHNSe}, we write them in a weak form. Note that the pressure equation \eqref{2ndCHNSd} is completely decoupled from the rest of the equations. Thus one only needs to establish the unique solvability of Eqs. \eqref{2ndCHNSa}-\eqref{2ndCHNSc}. Once $\overline{\mathbf{u}}^{k+1}$ or equivalently $\overline{\mathbf{u}}^{k+\frac{1}{2}}$ is known, one can find $\mathbf{u}^{k+1}$ and $p^{k+1}$ by either solving a Darcy problem as Eq. \eqref{2ndCHNSd} or solving a pressure Poisson equation and an update of the velocity as described in Eqs. \eqref{PrePoi}-\eqref{AupVel}. Hereafter, we denote by $L^2_0(\Omega)$ an $L^2$ subspace with mean zero, i.e., $L^2_0(\Omega):= \{f\in L^2(\Omega); \int_\Omega f dx=0\}$.
\begin{definition}\label{def1}
Given that $\phi^k, \phi^{k-1} \in H^1(\Omega)$, $\mathbf{u}^k, \mathbf{u}^{k-1} \in \mathbf{H^1}(\Omega)$, and $p^k\in H^1(\Omega)\cap L_0^2(\Omega)$ for $k=1,2, \cdots K=[T/\delta t]$, the triple $\{\phi^{k+1}, \mu^{k+\frac{1}{2}}, \overline{\mathbf{u}}^{k+\frac{1}{2}}\}$ is said to be a weak solution of Eqs. \eqref{2ndCHNSa}-\eqref{2ndCHNSc} if they satisfy
\begin{align*}
&\phi^{k+1} \in H^1(\Omega), \quad \mu^{k+\frac{1}{2}} \in H^1(\Omega),
\quad \overline{\mathbf{u}}^{k+\frac{1}{2}} \in \mathbf{H^1_0}(\Omega),
\end{align*}
and there hold, $\forall v\in H^1(\Omega), \varphi \in H^1(\Omega), \mathbf{v} \in \mathbf{H^1_0}(\Omega)$,
\begin{align}
&\big(\phi^{k+1}-\phi^k,v\big)+\delta t \big(M(\widetilde{\phi}^{k+\frac{1}{2}})\nabla \mu^{k+\frac{1}{2}}, \nabla v\big)-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}{\overline{\mathbf{u}}}^{k+\frac{1}{2}}, \nabla v\big)=0, \label{2ndCHNSwa}\\
&\big(\mu^{k+\frac{1}{2}}, \varphi \big)=\frac{1}{4}\bigg([(\phi^{k+1})^2+(\phi^k)^2](\phi^{k+1}+\phi^k), \varphi\bigg)-\big(\widetilde{\phi}^{k+\frac{1}{2}}, \varphi \big) \nonumber \\
&+\frac{\epsilon^2}{2} \big(\nabla (\phi^{k+1}+\phi^k), \nabla \varphi \big), \label{2ndCHNSwb}\\
&2\big({\overline{\mathbf{u}}}^{k+\frac{1}{2}}-\mathbf{u}^k, \mathbf{v} \big)+\delta t \frac{1}{Re}\big( \nabla{\overline{\mathbf{u}}}^{k+\frac{1}{2}}, \nabla\mathbf{v}\big)+\delta t b(\widetilde{\mathbf{u}}^{k+\frac{1}{2}}, {\overline{\mathbf{u}}}^{k+\frac{1}{2}}, \mathbf{v}) \nonumber \\
& =-\delta t\Big(\nabla p^k, \mathbf{v}\Big)-\delta t \frac{\epsilon^{-1}}{{We}^\ast} \big(\widetilde{\phi}^{k+\frac{1}{2}} \nabla \mu^{k+\frac{1}{2}}, \mathbf{v}\big), \label{2ndCHNSwc}
\end{align}
\end{definition}
where the trilinear form $b$ is defined in \eqref{trilinear}.
We will mainly use the well-known Browder-Minty lemma in establishing the unconditionally unique solvability of Eqs. \eqref{2ndCHNSwa}-\eqref{2ndCHNSwc}, cf. \cite{ReRo2004}, p.364, Theorem 10.49.
\begin{lemma}[Browder-Minty]\label{BrowderMinty}
Let $X$ be a real, reflexive Banach
space and let $T : X \rightarrow X^\prime$ (the dual space of $X$) be bounded, continuous, coercive and monotone.
Then for any $g \in X^\prime$ there exists a solution $u \in X$ of the equation
\begin{align*}
T(u)=g.
\end{align*}
If further, the operator $T$ is strictly monotone, then the solution $u$ is unique.
\end{lemma}
We observe that Eqs. \eqref{2ndCHNSwa}-\eqref{2ndCHNSwc} are coupled together through $\mu^{k+\frac{1}{2}}$. It is possible to rewrite the system equivalently as a scalar equation in terms of unknown $\mu^{k+\frac{1}{2}}$. To do so, we introduce two solution operators $\phi^{k+1}(\mu^{k+\frac{1}{2}}): \mu^{k+\frac{1}{2}} \rightarrow \phi^{k+1}$ and $\overline{\mathbf{u}}^{k+\frac{1}{2}}(\mu^{k+\frac{1}{2}}): \mu^{k+\frac{1}{2}} \rightarrow \overline{\mathbf{u}}^{k+\frac{1}{2}}$ by solving equations \eqref{2ndCHNSwb} and \eqref{2ndCHNSwc}, respectively, for a given source function $\mu^{k+\frac{1}{2}} \in H^1(\Omega)$. Specifically, one can establish the following lemmas.
\begin{lemma}[solvability of Eq. \eqref{2ndCHNSwb}]\label{wellposed2nd}
Given a source function $\mu^{k+\frac{1}{2}} \in H^1(\Omega)$ and known functions $\phi^k, \phi^{k-1} \in H^1(\Omega)$, there exists a unique solution $\phi^{k+1} \in H^1(\Omega)$ to Eq. \eqref{2ndCHNSwb}. Moreover, the solution is bounded and depends continuously on $\mu^{k+\frac{1}{2}}$ in the weak topology.
\end{lemma}
\begin{lemma}[solvability of Eq. \eqref{2ndCHNSwc}]\label{wellposed3rd}
Given a source function $\mu^{k+\frac{1}{2}} \in H^1(\Omega)$, known functions $\phi^k, \phi^{k-1} \in H^1(\Omega)$ and $\mathbf{u}^k, \mathbf{u}^{k-1} \in \mathbf{H}^1(\Omega)$, there exists a unique solution $\overline{\mathbf{u}}^{k+\frac{1}{2}} \in \mathbf{H}^1_0(\Omega)$ to Eq. \eqref{2ndCHNSwc}. In addition, the solution is bounded and depends continuously on $\mu^{k+\frac{1}{2}}$ in the strong topology.
\end{lemma}
It will be clear from the proof of Proposition \ref{Solva} below that the unique solvability of Eq. \eqref{2ndCHNSwb} can be proved by using Browder-Minty Lemma \ref{BrowderMinty} as well. The boundedness and continuity of the solution readily follow from the fact that Eq. \eqref{2ndCHNSwb} is a semilinear elliptic equation for $\phi^{k+1}$ with cubic nonlinearity. Lemma \ref{wellposed3rd} can be proved by invoking the Lax-Milgram Theorem. We omit the details here for conciseness.
With the help of Lemma \ref{BrowderMinty}, Lemma \ref{wellposed2nd} and Lemma \ref{wellposed3rd}, one can prove the unique existence of a weak solution in the sense of Definition \ref{def1}.
\begin{proposition}\label{Solva}
Assume that $\phi^k, \phi^{k-1} \in H^1(\Omega)$, $\mathbf{u}^k, \mathbf{u}^{k-1} \in \mathbf{H}^1(\Omega)$, and $p^k \in H^1(\Omega)$ are known functions for $k=1,2, \cdots, K-1$. Then there exists a unique weak solution to Eqs. \eqref{2ndCHNSa}-\eqref{2ndCHNSc} in the sense of Def. \ref{def1}
\end{proposition}
\begin{proof}
Here for notational simplicity, we will temporarily omit the the superscripts on $\phi^{k+1}, \mu^{k+\frac{1}{2}}, \overline{\mathbf{u}}^{k+\frac{1}{2}}$.
For any $\mu \in H^1(\Omega)$, one defines an operator $T: H^1(\Omega) \rightarrow (H^1(\Omega))^\prime$ such that
\begin{align}\label{Operator}
\langle T(\mu), v \rangle := \big(\phi-\phi^k,v\big)+\delta t \big(M\nabla \mu, \nabla v\big)-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}, \nabla v\big), \quad \forall v \in H^1(\Omega),
\end{align}
where $\langle, \rangle$ is the duality pairing between $(H^1(\Omega))^\prime$ and $H^1(\Omega)$, $\phi$ and $\overline{\mathbf{u}}$ are the unique solutions to Eqs. \eqref{2ndCHNSwb} and \eqref{2ndCHNSwc} that are defined in Lemma \ref{wellposed2nd} and Lemma \ref{wellposed3rd}, respectively.
It readily follows that
\begin{align*}
|\langle T(\mu), v \rangle| \leq C(\delta t)\big( ||\phi||_{L^2} +||\phi^k||_{L^2} +||\nabla \mu||_{L^2} +||\widetilde{\phi}^{k+\frac{1}{2}}||_{H^1}||\overline{\mathbf{u}}||_{H^1}\big)||\mathbf{v}||_{H^1},
\end{align*}
where we have used the boundedness of the mobility function $m_1 \leq M \leq m_2$ for constants $0<m_1 \leq m_2$.
Thus the boundedness of the operator $T$ follows from the boundedness of $\phi$ and $\overline{\mathbf{u}}$ as functions of $\mu$ in Lemmas \ref{wellposed2nd} and \ref{wellposed3rd}. Similarly, one can verify that the operator $T: H^1(\Omega)\rightarrow (H^1(\Omega))^\prime$ is continuous as a consequence of the continuity of $\phi$ and $\overline{\mathbf{u}}$ on $\mu$.
For the monotonicity, one obtains from the definition of $T$ in \eqref{Operator}
\begin{align}\label{Mono}
&\langle T(\mu)-T(\nu), \mu-\nu \rangle = \big(\phi(\mu)-\phi(\nu),\mu-\nu\big)+\delta t ||\sqrt{M}\nabla(\mu-\nu)||_{L^2}^2 \nonumber\\
&-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}[\overline{\mathbf{u}}(\mu)-\overline{\mathbf{u}}(\nu)], \nabla (\mu-\nu)\big), \quad \forall \mu, \nu \in H^1(\Omega),
\end{align}
where $\phi(\nu)$ and $\overline{\mathbf{u}}(\nu)$ are solutions to Eqs. \eqref{2ndCHNSwb} and \eqref{2ndCHNSwc}, respectively, with a given source function $\nu$.
For the first term on the right hand side of \eqref{Mono}, one subtracts Eq. \eqref{2ndCHNSwb} with source functions $\mu$ and $\nu$ respectively to get
\begin{align*}
\big(\mu-\nu, \varphi\big)&=\frac{1}{4}\int_\Omega(\phi(\mu)-\phi(\nu))[(\phi(\mu)+\phi(\nu))^2+(\phi(\mu)+\phi^k)^2+(\phi^k+\phi(\nu))^2]\varphi\, dx \\
&+\frac{\epsilon^2}{2}\big(\nabla(\phi(\mu)-\phi(\nu)), \nabla \varphi\big), \quad \forall \varphi \in H^1(\Omega).
\end{align*}
By taking $\varphi=\phi(\mu)-\phi(\nu)$ in the above equation, one concludes that
\begin{align}\label{Mono1st}
\big(\mu-\nu, \phi(\mu)-\phi(\nu)\big) \geq 0,
\end{align}
and that the equality holds if only if $\mu=\nu$ thanks to the uniqueness of solutions to Eq. \eqref{2ndCHNSwb} in Lemma \ref{wellposed2nd}.
By the linearity of Eq. \eqref{2ndCHNSwc}, the third term on the right hand side of Eq. \eqref{Mono} can be written as
\begin{align}\label{Mono2nd}
-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}[\overline{\mathbf{u}}(\mu)-\overline{\mathbf{u}}(\nu)], \nabla (\mu-\nu)\big)=\epsilon{We}^\ast\{2||\overline{\mathbf{u}}(\mu)-\overline{\mathbf{u}}(\nu)||_{L^2}^2+\frac{\delta t}{Re}||\nabla(\overline{\mathbf{u}}(\mu)-\overline{\mathbf{u}}(\nu))||_{L^2}^2\},
\end{align}
where the convective term vanishes thanks to the skew-symmetry of the form $b$. In view of \eqref{Mono1st} and \eqref{Mono2nd}, one sees
\begin{align}
\langle T(\mu)-T(\nu), \mu-\nu \rangle \geq 0,
\end{align}
with equality if only if $\mu=\nu$. This establishes the strict monotonicity of the operator $T$.
We next turn to the coercivity of the operator $T$. One has
\begin{align}\label{Coer}
\langle T(\mu), \mu \rangle = \big(\phi-\phi^k,\mu\big)+\delta t \big(M\nabla \mu, \nabla \mu\big)-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}, \nabla \mu\big), \quad \forall \mu \in H^1(\Omega).
\end{align}
Taking the test function $\varphi=\phi-\phi^k$ in Eq. \eqref{2ndCHNSwb}, one obtains
\begin{align}\label{Coer1st}
&\big(\phi-\phi^k, \mu\big) \nonumber \\
&=\frac{1}{4} \int_\Omega \phi^4-(\phi^k)^4\, dx -\int_\Omega \widetilde{\phi}^{k+\frac{1}{2}}(\phi-\phi^k)\, dx+\frac{\epsilon^2}{2}\int_\Omega |\nabla \phi|^2-|\nabla \phi^k|^2\, dx. \nonumber \\
&\geq \frac{1}{8} \int_\Omega \phi^4\, dx +\frac{\epsilon^2}{2}\int_\Omega |\nabla \phi|^2\,dx-C(\epsilon, \Omega)\big(||\widetilde{\phi}^{k+\frac{1}{2}}||_{L^2}^2+||\phi^k||_{H^1}+1\big)
\end{align}
Similarly, one can take the test function $\mathbf{v}=\overline{\mathbf{u}}$ in Eq. \eqref{2ndCHNSwc} to get
\begin{align}\label{Coer2nd}
-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}\overline{\mathbf{u}}, \nabla \mu\big)&=\epsilon {We}^\ast \{2||\overline{\mathbf{u}}||_{L^2}^2+\delta t ||\nabla \overline{\mathbf{u}}||_{L^2}^2- \delta t \big(2\mathbf{u}^k-\nabla p^k, \overline{\mathbf{u}}\big)\} \nonumber \\
&\geq C(\epsilon, {We}^\ast, \delta t)\{||\overline{\mathbf{u}}||_{L^2}^2+||\nabla \overline{\mathbf{u}}||_{L^2}^2-(||\mathbf{u}^k||_{L^2}^2+||\nabla p^k||_{L^2}^2)\}.
\end{align}
Collecting inequalities \eqref{Coer1st} and \eqref{Coer2nd}, one finds that Eq. \eqref{Coer} becomes
\begin{align}\label{Coer2}
\langle T(\mu), \mu \rangle \geq C||\nabla \mu||_{L^2}^2+\frac{1}{8} ||\phi||_{L^4}^4 +\frac{\epsilon^2}{2} ||\nabla \phi||^2_{L^2}+C(||\overline{\mathbf{u}}||_{L^2}^2+||\nabla \overline{\mathbf{u}}||_{L^2}^2)- C,
\end{align}
where again the boundedness of the mobility function has been invoked.
To have coercivity in $H^1(\Omega)$, one needs to bound the average $m(\mu):=\frac{1}{|\Omega|}\int_\Omega \mu\, dx$ appropriately. For this, one takes the test function $\varphi=1$ in Eq. \eqref{2ndCHNSwb}.
\begin{align*}
\Big|\int_\Omega \mu\, dx\Big| &\leq \frac{1}{4}\int_\Omega |\phi|^3+|\phi|| \phi^k|^2+|\phi^k \phi^2|\, dx+ C(||\phi^k||_{L^3}^3+||\widetilde{\phi}^{k+\frac{1}{2}}||_{H^1}) \\
& \leq C(||\phi||_{L^3}^3+||\phi||_{L^2}^2+||\phi^k||_{L^2}||\phi||_{L^4}^2)+C \nonumber \\
& \leq C(||\phi||_{L^3}^3+||\phi||_{L^2}^2+||\phi||_{L^4}^3)+C \nonumber \\
&\leq C(||\phi||_{L^4}^3+||\phi||_{L^4}^2)+ C,
\end{align*}
where one has applied Young's inequality.
It readily follows that
\begin{align}\label{Coer21}
|m(\mu)|^\frac{4}{3} \leq C||\phi||_{L^4}^4+ C
\end{align}
Thus by using Poincar\'{e} inequality, one gets from \eqref{Coer2} and \eqref{Coer21} that
\begin{align}
\langle T(\mu), \mu \rangle \geq C||\mu||_{H^1}^{\frac{4}{3}}-C,
\end{align}
which implies the coercivity of $T$.
Now Browder-Minty Lemma \ref{BrowderMinty} yields that there exists a unique solution $\mu^{\star} \in H^1(\Omega)$ such that $\langle T(\mu^{\star}), v \rangle=0, \forall v \in H^1(\Omega)$. In view of the definition of $T$ in \eqref{Operator}, one sees $\mu^\star \in H^1(\Omega)$ and the corresponding $\phi^{\star} \in H^1(\Omega), \overline{\mathbf{u}}^{\star} \in \mathbf{H}_0^1$ uniquely solve the system \eqref{2ndCHNSwa}-\eqref{2ndCHNSwc}.
\end{proof}
\section{Mixed Finite Element Formulation}
We now discretize the time-discrete scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSe} in space by finite element method.
Let $\mathcal{T}_h$ be a quasi-uniform triangulation of the domain $\Omega$ of mesh size $h$. We introduce $\mathbf{X}_h$ and $Y_h$ the finite element approximations of $\mathbf{H}_0^1(\Omega)$ and $H^1(\Omega)$ respectively based on the triangulation $\mathcal{T}_h$. In addition, we define $M_h=Y_h \cap L^2_0(\Omega):= \{q_h \in Y_h; \int_{\Omega}q_h dx=0\}$. We assume that $Y_h \times Y_h$ is a stable pair for the biharmonic operator in the sense that there holds the inf-sup condition
\begin{align*}
\sup_{\phi_h \in Y_h} \frac{(\nabla \phi_h, \nabla \varphi_h)}{||\phi_h||_{H^1}} \geq c||\varphi_h||_{H^1}, \quad \forall \varphi_h \in Y_h.
\end{align*}
We also assume that $\mathbf{X}_h$ and $Y_h$ are stable approximation spaces for velocity and pressure in the sense of
\begin{align*}
\sup_{\mathbf{v}_h \in \mathbf{X}_h} \frac{(\nabla \cdot \mathbf{v}_h, q_h)}{||\mathbf{v}_h||_{H^1}} \geq c||q_h||_{L^2}, \quad \forall q_h \in Y_h.
\end{align*}
It is pointed out \cite{GuQu1998b} that the inf-sup condition is necessary for the stability of pressure even though one may solve the projection step as a pressure Poisson equation.
Then the fully discrete finite element formulation for scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSe} reads: find \linebreak $(\phi_h^{k+1}, \mu^{k+\frac{1}{2}}_h, \overline{\mathbf{u}}_h^{k+\frac{1}{2}}, p_h^{k+1}, \mathbf{u}_h^{k+1}) \in Y_h \times Y_h \times \mathbf{X}_h \times M_h \times \mathbf{X}_h$ such that for all $(v_h, \varphi_h, \mathbf{v}_h, q_h) \in Y_h \times Y_h \times \mathbf{X}_h \times Y_h$ there hold
\begin{align}
&\big(\phi^{k+1}_h-\phi^k_h,v_h\big)+\delta t \big(M\nabla \mu^{k+\frac{1}{2}}_h, \nabla v_h\big)-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}_h\overline{\mathbf{u}}^{k+\frac{1}{2}}_h, \nabla v_h\big)=0, \label{d2ndCHNSwa}\\
&\big(\mu^{k+\frac{1}{2}}_h, \varphi_h \big)=\frac{1}{4}\bigg([(\phi^{k+1}_h)^2+(\phi^k_h)^2](\phi^{k+1}_h+\phi^k_h), \varphi_h\bigg)-\big(\widetilde{\phi}^{k+\frac{1}{2}}_h, \varphi_h \big) \nonumber\\
&+\frac{\epsilon^2}{2} \big(\nabla (\phi^{k+1}_h+\phi^k_h), \nabla \varphi_h \big), \label{d2ndCHNSwb}\\
&\big(2\overline{\mathbf{u}}^{k+\frac{1}{2}}_h, \mathbf{v}_h \big)+\delta t \frac{1}{Re}\big( \nabla\overline{\mathbf{u}}^{k+\frac{1}{2}}_h, \nabla \mathbf{v}_h\big)+\delta t b\big(\widetilde{\mathbf{u}}^{k+\frac{1}{2}}_h, \overline{\mathbf{u}}^{k+\frac{1}{2}}_h, \mathbf{v}_h\big) =-\delta t\big(\nabla p^k_h, \mathbf{v}_h\big) \nonumber \\
&+\big(2\mathbf{u}^k_h, \mathbf{v}_h \big)-\delta t \frac{\epsilon^{-1}}{{We}^\ast} \big(\widetilde{\phi}^{k+\frac{1}{2}}_h \nabla \mu^{k+\frac{1}{2}}_h, \mathbf{v}_h\big), \label{d2ndCHNSwc} \\
&\big(\mathbf{u}^{k+1}_h-\widetilde{\mathbf{u}}^{k+1}_h, \mathbf{v}_h\big)+ \frac{\delta t}{2}\big(\nabla (p_h^{k+1}-p_h^k), \mathbf{v}_h\big)=0, \label{d2ndCHNSwd} \\
&\big(\nabla \cdot \mathbf{u}^{k+1}_h, q_h\big)=0. \label{d2ndCHNSwe}
\end{align}
The notations used here are defined in \eqref{notas} and Eq. \eqref{trilinear}.
The properties of the time-discrete scheme \eqref{2ndCHNSa}-\eqref{2ndCHNSe} (i.e., mass-conservation, unconditional stability and unconditionally unique solvability) are preserved by the fully discrete formulation \eqref{d2ndCHNSwa}-\eqref{d2ndCHNSwe}. Note that Eqs. \eqref{d2ndCHNSwd}-\eqref{d2ndCHNSwe} amount to solving the projection step \eqref{2ndCHNSd}-\eqref{2ndCHNSe} as a Darcy problem. This formulation is shown \cite{GuQu1998a} to yield an optimal condition number for the pressure operator associated with finite element spatial discretizations. An alternative way of solving Eqs. \eqref{d2ndCHNSwd}-\eqref{d2ndCHNSwe} is the so-called "approximate projection" (cf. \cite{Codina2001} and references therein)
\begin{align*}
&\big(\nabla(p^{k+1}_h-p^k_h), \nabla q_h\big)=\frac{2}{\delta t} \big(\widetilde{\mathbf{u}}^{k+1}_h, \nabla q_h \big), \forall q_h \in Y_h\\
&\big(\mathbf{u}_h^{k+1}, \mathbf{v}_h\big)=\big(\widetilde{\mathbf{u}}_h^{k+1}-\frac{\delta t}{2}\nabla(p_h^{k+1}-p_h^k), \mathbf{v}_h\big), \forall \mathbf{v}_h \in \mathbf{X}_h.
\end{align*}
One can still prove the unconditional stability of the scheme with the approximate projection, see the reference above. In our numerical experiment, we observe the $L^2$ error of the pressure is indeed smaller in the former case, though at the expense of more memory consumed due to the coupling between the velocity and pressure.
Note that the only nonlinear term appears in the chemical potential equation \eqref{d2ndCHNSwb}. We thus adopt a Picard iteration procedure on velocity to decouple the computation of the nonlinear Cahn-Hilliard equation \eqref{d2ndCHNSwa} and \eqref{d2ndCHNSwb} from that of the linear Navier-Stokes equation \eqref{d2ndCHNSwc}. Denote by $i$ the Picard iteration index. Specifically, given the velocity $\overline{\mathbf{u}}^{k+\frac{1}{2}, i}$, we solve for $\phi^{k+1, i+1}, \mu^{k+\frac{1}{2}, i+1}$ from the Cahn-Hilliard equation \eqref{d2ndCHNSwa} -- \eqref{d2ndCHNSwb} by Newton's method. As $\mu^{k+\frac{1}{2}, i+1}$ is available, we can then proceed to solve for $\overline{\mathbf{u}}^{k+\frac{1}{2}, i+1}$ from the linear equations \eqref{d2ndCHNSwc}. We repeat this procedure until the relative difference between two iterations within a fixed tolerance. We summarize this procedure in four steps as follows:
\noindent \textbf{Step $1$}: given $\overline{\mathbf{u}}^{k+\frac{1}{2}, i}$, find $(\phi^{k+1, i+1}_h, \mu^{k+\frac{1}{2}, i+1}_h) \in Y_h \times Y_h$ such that $\forall (v_h, \varphi_h) \in Y_h \times Y_h$
\begin{align*}
& \big(\phi^{k+1, i+1}_h-\phi^k_h,v_h\big)+\delta t \big(M\nabla \mu^{k+\frac{1}{2}, i+1}_h, \nabla v_h\big)-\delta t \big(\widetilde{\phi}^{k+\frac{1}{2}}_h\overline{\mathbf{u}}^{k+\frac{1}{2}, i}_h, \nabla v_h\big)=0, \\
& \big(\mu^{k+\frac{1}{2}, i+1}_h, \varphi_h \big)=\frac{1}{4}\bigg([(\phi^{k+1, i+1}_h)^2+(\phi^k_h)^2](\phi^{k+1, i+1}_h+\phi^k_h), \varphi_h\bigg)-\big(\widetilde{\phi}^{k+\frac{1}{2}}_h, \varphi_h \big) \\
&+\frac{\epsilon^2}{2} \big(\nabla (\phi^{k+1, i+1}_h+\phi^k_h), \nabla \varphi_h \big),
\end{align*}
\noindent \textbf{Step $2$}: find $\overline{\mathbf{u}}^{k+\frac{1}{2}, i+1} \in \mathbf{X}_h$ such that $\forall \mathbf{v}_h \in \mathbf{X}_h$
\begin{align*}
&\big(2\overline{\mathbf{u}}^{k+\frac{1}{2}, i+1}_h, \mathbf{v}_h \big)+\delta t \frac{1}{Re}\big( \nabla\overline{\mathbf{u}}^{k+\frac{1}{2}, i+1}_h, \nabla \mathbf{v}_h\big)+\delta t b\big(\widetilde{\mathbf{u}}^{k+\frac{1}{2}}_h, \overline{\mathbf{u}}^{k+\frac{1}{2}, i+1}_h, \mathbf{v}_h\big) =-\delta t\big(\nabla p^k_h, \mathbf{v}_h\big) \\
&+\big(2\mathbf{u}^k_h, \mathbf{v}_h \big)-\delta t \frac{\epsilon^{-1}}{{We}^\ast}\big(\widetilde{\phi}^{k+\frac{1}{2}}_h \nabla \mu^{k+\frac{1}{2}, i+1}_h, \mathbf{v}_h\big),
\end{align*}
\noindent \textbf{Step $3$}: find $\overline{\mathbf{u}}^{k+\frac{1}{2}}_h \in \mathbf{X}_h$ by repeating Step 1 and Step 2 until the relative error in $L^2$ of $(\overline{\mathbf{u}}^{k+\frac{1}{2}, i+1}_h-\overline{\mathbf{u}}^{k+\frac{1}{2}, i}_h)$ is within a fixed tolerance.
\noindent \textbf{Step $4$}: find $\mathbf{u}_h^{k+1} \in \mathbf{X}_h, p_h^{k+1} \in Y_h$ (equivalently, $p_h^{k+1}-p_h^{k}$) such that $\forall \mathbf{v}_h \in \mathbf{X}_h, q_h \in Y_h$
\begin{align*}
&\big(\mathbf{u}^{k+1}_h-\widetilde{\mathbf{u}}^{k+1}_h, \mathbf{v}_h\big)+ \frac{\delta t}{2}\big(\nabla (p_h^{k+1}-p_h^k), \mathbf{v}_h\big)+\big(\nabla \cdot \mathbf{u}^{k+1}_h, q_h\big)=0.
\end{align*}
We remark that our scheme is a two step method.
One can solve for $\phi^1_h, \mu^1_h, \mathbf{u}^1_h, p^1_h$ through a coupled first order scheme (see, for example, \cite{Feng2006, KSW2008}) to initialize the second order scheme. Numerical simulations in \cite{KKL2004} suggest that at least 4 grid elements across the interfacial region of thickness $\sqrt{2}\epsilon$ are needed for accuracy. To improve the efficiency of the algorithm, we explore the capability of adaptive mesh refinement of FreeFem++ (cf. \cite{Hecht2012}) in which a variable metric/Delaunay automatic meshing algorithm is implemented.
\section{Numerical Experiments}
In this section, we perform some standard tests to gauge our numerical algorithm. For simplicity, we will use P1--P1 function spaces for $Y_h \times Y_h$ , and P1b--P1 mixed finite element spaces for $\mathbf{X}_h \times Y_h$ . It is well-known \cite{Ciarlet2002, ABF1984} that these approximation spaces satisfy the inf-sup conditions for the biharmonic operator and Stokes operator, respectively. In principle, any inf-sup compatible approximation spaces for biharmonic operator and Stokes operator can be used, for example, P2--P2 for $Y_h \times Y_h$, and Taylor-Hood P2--P1 for $\mathbf{X}_h \times Y_h$.
\subsection{Convergence, energy dissipation, mass conservation}
Here we provide some numerical evidence to show that our scheme is second order accurate, energy-dissipative and mass-conservative.
As the Cahn-Hilliard equation does not have a natural forcing term, we verify the second order convergence of the scheme by a Cauchy convergence test. We consider the problem in a unit square domain $\Omega=[0,1]\times[0,1]$. The initial conditions are taken to be
\begin{align*}
& \phi_0=0.24\cos(2\pi x)\cos(2\pi y)+0.4\cos(\pi x)\cos(3\pi y), \\
& \mathbf{u}_0=(-\sin(\pi x)^2\sin(2\pi y), \sin(\pi y)^2\sin(2\pi x)).
\end{align*}
We impose no-slip no penetration boundary conditions for velocity, and homogeneous Neumann boundary condition for $\phi$ and $\mu$ .
The final time is $T=0.1$, the grid in space is uniform $h=\frac{\sqrt{2}}{2^n}$ ($2^n$ grid points in each direction), for $n$ from $5$ to $9$, and the refinement path is taken to be $\delta t= \frac{0.2}{\sqrt{2}}h$. The other parameters are $\epsilon=0.04$, $M=0.1$, ${We}^\ast=25$ , $Re=100$. We calculate the the rate at witch the Cauchy difference converges to zero in the $L^2$ norm. The errors and convergence rates are given in Table \ref{CauchyP1} . The results show that the scheme is of second order accuracy for $\phi$ and $\mathbf{u}$ in $L^2$ norm, and the rate of convergence for pressure $p$ appear to be only first order.
\begin{table}[h!]
\begin{center}
\begin{tabular}{c c c c c c c c}
\hline
& $32-64$ & rate & $64-128$ & rate • &$128-256$ & rate & $256-512$ \\
\hline
$\phi$• & $4.14e-3$ & $1.90$ & $1.11-3$ & $1.97$ & $2.83e-4$ & $1.99$& $7.12e-5$ \\
$u$ & $7.21e-4$• &$2.08$ &$1.70e-4$ &$ 2.04$ & $4.16e-5$ &$ 2.02$ & $1.03e-5$ \\
$v$ & $6.99e-4$ & $ 2.11 $ & $1.62e-4$ & $ 2.05$ &$3.93e-5$ & $2.02$ & $9.71e-6$ \\
$p$ & $2.05e-3$ & $1.75$ & $6.10e-4$& $ 1.62$ & $1.98e-4$ & $ 1.44$ &$7.27e-5$ \\
\hline
\end{tabular}
\caption{Cauchy convergence test; errors are measured in $L^2$ norm; $2^n$ grid points in each direction for $n$ from $5$ to $9$, $\delta t= \frac{0.2}{2}h$, $\epsilon=0.04$, $M=0.1$, ${We}^\ast=25$ , $Re=100$.}
\label{CauchyP1}
\end{center}
\end{table}
Next, we verify numerically that the total energy of the system is non-increasing at each time step. We define two discrete energy functional at discrete time $t=k \delta t$ according to Proposition \ref{ModEnergyLaw}
\begin{align*}
&E^{h,t}=\int_{\Omega}\frac{1}{2}|\mathbf{u}_h^k|^2\, dx+ \frac{1}{{We}^\ast}\int_{\Omega} \big( \frac{1}{\epsilon} f_0(\phi_h^k)+\frac{\epsilon}{2}|\nabla \phi_h^k|^2\big)\, dx, \\
&E^{h,t}_{app}=E^{h,t}+\frac{\epsilon^{-1}}{4{We}^\ast}\int_{\Omega}|\phi^{k}_h-\phi^{k-1}_h|^2 dx+\frac{\delta t^2}{8}\int_{\Omega}|\nabla p^{k}_h|^2 dx.
\end{align*}
In the calculation, we take $\delta t= 0.005$, $h=\frac{\sqrt{2}}{128}$ and a constant mobility $M=1.0$. The other parameters are the same as ones in the Cauchy convergence test. Fig. \ref{DEn} shows that both of the discrete energy functional $E^{h,t}$ and $E^{h,t}_{app}$ are indeed non-increasing in time. Moreover, since $E^{h,t}_{app}$ is a second order approximation of $E^{h,t}$ in terms of $\delta t$, the qualitative evolution behaviour of $E^{h,t}$ and $E^{h,t}_{app}$ is virtually the same.
\begin{figure}
\centering
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=\textwidth]{TrueEn.pdf}
\caption{Evolution of $E^{h,t}$}
\label{DEnTrue}
\end{subfigure}%
~
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=\textwidth]{NumEn.pdf}
\caption{Evolution of $E^{h,t}_{app}$}
\label{DEnNum}
\end{subfigure}
\caption{Time evolution of the discrete energy; $\delta t= 0.005$, $h=\frac{\sqrt{2}}{128}$, $M=1.0$, $\epsilon=0.04$, ${We}^\ast=25$ , $Re=100$.}
\label{DEn}
\end{figure}
In Fig. \ref{Mass}, we show the time evolution of the discrete mass $\int_{\Omega}\phi_h^k dx$ associated with the energy dissipation test (Fig. \ref{DEn}). Note that $\int_{\Omega}\phi_0 dx=0$. After projection into the finite element space P1, we have $\int_{\Omega} \phi_h^0dx =8.14e-06$. Fig. \ref{Mass} shows that the exact value is preserved during the evolution, which verifies that our scheme is conservative.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\linewidth]{Mass.pdf}
\caption{Time evolution of the discrete mass $\int_\Omega \phi_h^k dx$; the parameters are given in Fig. \ref{DEn}.}
\label{Mass}
\end{figure}
\subsection{Shape relaxation}
Here we use the CHNS system \eqref{CH}-\eqref{div} to simulate the relaxation of an isolated shape in a two-phase flow system. The initial shape is a small square located in the middle of the domain (cf. Fig. \ref{IniSh}). For velocity, we set both the initial condition and boundary condition to be zero. We impose homogeneous Neumann boundary conditions fro $\phi$ and $\mu$. The parameters are $\epsilon=0.005$, ${We}^\ast=200$, $M(\phi)=0.1\sqrt{(1-\phi^2)^2+\epsilon^2}$, $Re=10$, $\delta t=0.005$. In space, we explore the adaptive mesh refinement of FreeFem++ (cf. \cite{Hecht2012}) which uses a variable metric/Delaunay automatic meshing algorithm. Specifically, we adapt the mesh according to the Hessian of the order parameter such that at least four grid cells are located across the diffuse interface.
\begin{figure}[h!]
\centering
\includegraphics[width=0.3\linewidth]{ShaRIni.png}
\caption{The initial shape of the order parameter for simulations of shape relaxation.}
\label{IniSh}
\end{figure}
Since the initial velocity is zero, the initial total energy of the system is the surface energy. Due to the effect of surface tension and the isotropy of the mobility, isolated irregular shape will relax to a circular shape. This relaxation is observed in Fig. \ref{ShaRE}. We also show the effectiveness of the adaptive mesh refinement at $t=0.02$ and $t=0.4$ in Fig. \ref{AdpM}.
\begin{figure}[h!]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.3\linewidth]{SurD0d02.pdf}& \includegraphics[width=0.3\linewidth]{SurD0d1.pdf}\\
$t=0.02$ & $t=0.1$ \\
\includegraphics[width=0.3\linewidth]{SurD004.pdf}& \includegraphics[width=0.3\linewidth]{SurD010.pdf}\\
$t=0.4$ & $t=1$
\end{tabular}
\caption{Shape relaxation of surface tension driven flow; $\epsilon=0.005$, ${We}^\ast=200$, $M(\phi)=0.1\sqrt{(1-\phi^2)^2+\epsilon^2}$, $Re=10$, $\delta t=0.005$; Adaptive mesh refinement is explored for spatial discretization.}
\label{ShaRE}
\end{figure}
\begin{figure}[h!]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.3\textwidth]{Mesh0d02.pdf}& \includegraphics[width=0.3\textwidth]{Mesh004.pdf}\\
mesh at $t=0.02$ & mesh at $t=0.4$
\end{tabular}
\caption{Adaptive mesh refinement associated with shape relaxation in Fig. \ref{ShaRE} at $t=0.02, 0.4$; $\epsilon=0.005$, $4$ grid elements are placed across the interfacial area.}
\label{AdpM}
\end{figure}
Next, we demonstrate the effect of imposed shear on shape relaxation. The initial configuration of order parameter is given in Figure \ref{IniSh}. For velocity, we take the initial data to be the Stokes solution to the lid driven cavity problem and for boundary data we take $\mathbf{u}|_{y=1}=\big(x(1-x), 0\big)$ and zero otherwise. We set $Re=100$ and the rest of the parameters are the same as in the case of surface tension driven flow (Fig. \ref{ShaRE}). The relaxation of the shape under shear driven flow and the associated flow field are reported in Fig. \ref{ShearD}. As the flow goes clockwise, the shape travels slightly to the left. Meanwhile, the shape elongates to an ellipse with the major axis along north-west direction.
\begin{figure}[h!]
\centering
\begin{tabular}{ccc}
$t=0.02$ & \includegraphics[width=0.26\linewidth]{Phi0d02.pdf}& \includegraphics[width=0.29\linewidth]{V0d02.png}\\
$t=0.1$& \includegraphics[width=0.26\linewidth]{Phi0d2.pdf} &\includegraphics[width=0.29\linewidth]{V0d2.png} \\
$t=0.4$ & \includegraphics[width=0.26\linewidth]{Phi0d4.pdf}&\includegraphics[width=0.29\linewidth]{V0d4.png} \\
$t=1$ &\includegraphics[width=0.26\linewidth]{Phi1.pdf} &\includegraphics[width=0.29\linewidth]{V1.png}
\end{tabular}
\caption{Shape relaxation under shear driven flow and the flow field; The applied shear is on the upper boundary with a shear rate of $x(1-x)$; $Re=100$, $\epsilon=0.005$, ${We}^\ast=200$, $M(\phi)=0.1\sqrt{(1-\phi^2)^2+\epsilon^2}$.}
\label{ShearD}
\end{figure}
\subsection{Spinodal decomposition}
The CHNS system \eqref{CH}-\eqref{div} can be used as a model for spinodal decomposition of a binary fluid, cf. \cite{KKL2004}. Here we examine the effect of the excess surface tension, defined as $\gamma=\frac{1}{{We}^\ast}$, on coarsening during spinodal decomposition. The initial velocity is zero $\mathbf{u}_0=0$. For the initial condition of the phase field variable, we take a random field of values $\phi_0=\bar{\phi}+r(x,y)$ with an average composition $\bar{\phi}=-0.05$ and random $r \in [-0.05, 0.05]$. We take no-slip no penetration boundary condition for velocity and homogeneous Neumann boundary condition for $\phi$ and $\mu$. The parameters are $\epsilon=0.005$, $M(\phi)=0.1\sqrt{(1-\phi^2)^2+\epsilon^2}$, $Re=10$, $\delta t=0.005$, $h=\frac{\sqrt{2}}{256}$. In Fig. \ref{Spi}, we show some snapshots of the filled contour of $\phi$ in gray scale (white color $\phi\approx 1.0$, black color $\phi \approx -1.0$) at different times with $\gamma=0, 0.1\epsilon, 1.0 \epsilon$, respectively. The case of $\gamma=0$, corresponds to purely Cahn-Hilliard equation with no fluid motion, is included for comparison purpose.
\begin{figure}[h!]
\centering
\begin{tabular}{cccc}
$t=4$ & \includegraphics[scale=0.168]{0t=4.png}& \includegraphics[scale=0.18]{0D1t=4.png} &\includegraphics[scale=0.18]{1t=4.png} \\
$t=9$ & \includegraphics[scale=0.168]{0t=9.png}& \includegraphics[scale=0.18]{0D1t=9.png} &\includegraphics[scale=0.18]{1t=9.png}\\
$t=20$ & \includegraphics[scale=0.168]{0t=20.png}& \includegraphics[scale=0.19]{0D1t=20.png} &\includegraphics[scale=0.19]{1t=20.png}\\
$t=40$ & \includegraphics[scale=0.168]{0t=40.png}& \includegraphics[scale=0.19]{0D1t=40.png} &\includegraphics[scale=0.19]{1t=40.png}\\
& $\gamma=0$& $\gamma=0.1 \epsilon$ &$\gamma=1.0 \epsilon$
\end{tabular}
\caption{Snapshots of coarsening of a binary fluid during spinodal decomposition with $\gamma = 0.1\epsilon$ (second column), $1.0 \epsilon$ (third column), respectively; The case of $\gamma=0$ (first column) is included for comparison purpose; The rest of the parameters are $\epsilon=0.005$, $M(\phi)=0.1\sqrt{(1-\phi^2)^2+\epsilon^2}$, $Re=10$, $\delta t=0.005$, $h=\frac{\sqrt{2}}{256}$.}
\label{Spi}
\end{figure}
After a rapid initial phase separation (not shown in Fig. \ref{Spi}), the dynamics of the CHNS system are dominated by the slow process of coarsening. There are several physical mechanisms in the CHNS system that contribute to the coarsening process: bulk diffusion, surface diffusion, and hydrodynamic convection. Note that our chosen regularized degenerate mobility $M(\phi)$ limits the bulk diffusion (of order $\epsilon$) at the late stage of the coarsening process. In comparison to the coarsening process governed by the Cahn-Hilliard equation with no fluid motion (the first column of Fig. \ref{Spi}), we find that the hydrodynamic effect speeds up the coarsening process by promoting the droplets coalescence, the larger $\gamma$, the more dramatic the coalescence effect. The effect is less discernible in the case of $\gamma=0.1 \epsilon$ (${We}^\ast=2000$). Indeed, the morphology for $\gamma=0$ (${We}^\ast=\infty$) and $\gamma=0.1\epsilon$ are nearly identical over the evolution. One can even observe the evaporation-condensation effect (Ostwald ripening) for the scattered isolated drops at $t=4, 9$. In contrast, for $\gamma=\epsilon$ (${We}^\ast=200$) at the same time, the morphology is less deformed and exhibits rich connection (fewer isolated drops). Moreover, as time evolves, the scattered islands quickly merge together.
Coarsening rate can be tied with surface energy decay rate.
The domain size of one phase $L$ (physical length scale) can be defined as a suitable negative norm of the order parameter \cite{KoOt2002}. Recall that the surface energy $E_f$ is defined as
$$
E_f=\frac{1}{{We}^\ast}\int_{\Omega} \big( \frac{1}{\epsilon} f_0(\phi)+\frac{\epsilon}{2}|\nabla \phi|^2\big)\, dx.
$$
Thus the surface energy $E_f$ is proportional to the average interfacial circumference in 2D (interfacial area in 3D), at least near equilibrium where the order parameter roughly has a profile of hyperbolic tangent function \cite{Shen2012}. It follows from the conservation of volume that the spatially averaged surface energy should scale like the inverse of the domain diameter $L$. This heuristic argument suggests that the decay rate of the surface energy can be used as a proxy of the phase coarsening rate. One can also motivate this argument from the standpoint of sharp interface limit. The exact relation between $L$ and $E_f$ is an inequality established rigorously in \cite{KoOt2002}. Fig. \ref{CoarsenRate} shows the correlation between the surface energy decay rate $E_f$ (thus coarsening rate) and the excess surface tension parameter $\gamma$. It is observed that larger surface tension $\gamma$ (smaller ${We}^\ast$) yields faster coarsening rate, which agrees with the energy law \eqref{ModEnergyLaw}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\linewidth]{rate3.pdf}
\caption{Loglog plot of the surface energy $E_f$ as a function of time (solid lines) for simulations in Fig. \ref{Spi}. Here we include the case $\gamma=0.01\epsilon$ for comparison purpose. The dash lines are fitted functions $c_1 t^{-0.216}$ ($\gamma=0.01\epsilon$), $c_2 t^{-0.239}$ ($\gamma=0.1\epsilon$) and $c_3t^{-0.304}$ ($\gamma=1.0\epsilon$), respectively.}
\label{CoarsenRate}
\end{figure}
For a large system of a binary fluid at late stage of spinodal decomposition, it is expected \cite{Siggia1979, MGG1985} that the coarsening rate would obey a dynamical scaling law: $L(t) \propto t ^\alpha$, where $L(t)$ is the average domain size of one phase. Nevertheless, in 2D such a scaling law is open to debate (see the recent work \cite{OSS2013} and references therein). Here we run our scheme on a domain $\Omega=[0,200] \times [0,200]$ for a final time up to $10^4$. The parameters are: $\epsilon=1.0$, $M=1.0$, ${We}^\ast=1.0$, $Re=1.0$. The initial and boundary conditions for $\phi$ and $\mathbf{u}$ are set similarly as above. We take $\delta t =0.5$ and $h=\sqrt{2}$ for $t \leq 1000$, and $\delta =1.0$ and adaptive mesh refinement for $t \in [10^3, 10^4]$. We plot the decay of the surface energy $E_f$ in log-log scale in Fig. \ref{Coarsening}, which reveals roughly a decay rate of $\frac{1}{2}$ at the late stage of coarsening. This result corroborates the $t^{\frac{1}{2}}$ growth law for the average domain size proposed in \cite{MGG1985}. Note that the wall effect becomes influential when $t$ approaches $10^4$ at which large islands occupy the boundary of the domain.
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\linewidth]{coarsening.pdf}
\caption{Loglog plot of the surface energy $E_f$ as a function of time (solid line) for the CHNS system; $\Omega=[0, 200] \times [0, 200]$, $\epsilon=1.0$, $M=1.0$, ${We}^\ast=1.0$, $Re=1.0$; The red dash line has a slope of $-\frac{1}{2}$. }
\label{Coarsening}
\end{figure}
\section{Conclusions}
In this paper, we have presented a novel second order in time numerical method for the Cahn-Hilliard-Navier-Stokes system that models two-phase flow with matched density.
The method is efficient since we decoupled the pressure from the velocity and phase field, and the coupling between the velocity field and the phase field is weak.
We have shown in a rigorous fashion that the scheme is unconditionally stable and uniquely solvable. Fully discrete numerical methods effected with finite-element method are also presented and analyzed with similar conclusions. To the best of our knowledge, this is the first second-order scheme that decouples the pressure and the velocity and phase field variables while maintaining unconditional stability and unique solvability.
Several numerical experiments are performed to test the accuracy of the scheme. We verify numerically that our scheme is conservative, energy- dissipative, and is of second order accuracy in $L^2$ norm. We demonstrate the effectiveness of our scheme incorporated with adaptive mesh refinement by simulating the shape relaxation with and without applied shear. Finally, we also investigates the effect of surface tension on the coarsening rate of spinodal decomposition of a binary fluid. In particular, our long time numerical simulation suggests a growth rate of $t^{\frac{1}{2}}$ for a large system at late stage, which agrees with \cite{MGG1985}.
There are numerous potential extensions of the current work. The design of second-order in time scheme that decouples the pressure, velocity and phase field completely, and is unconditionally stable and uniquely solvable is very desirable. The extension of the current scheme to the case of unmatched density, or to the case of coupled Cahn-Hilliard-Stokes-Darcy system that models two-phase flow in karstic geometry would also be interesting \cite{HSW13}.
From the theoretical side, the rigorous error analysis of the scheme, especially with adaptive mesh, is a very attractive but challenging topic.
\section*{Acknowledgments}
This work was completed while Han was supported as a Research Assistant on an NSF grant (DMS1312701). The authors also acknowledge the support of NSF DMS1008852, a planning grant and a multidisciplinary support grant from the Florida State University.
\bibliographystyle{plain}
|
\section{Fast Sketches for Matroid Rank Functions}
If $v$ is a matroid rank function, we obtain a $\Omega(\sqrt n)$-sketch of it with only $\tilde O(n)$ value queries.
\begin{proposition}
There exists an algorithm that produces an $\tilde O(\sqrt n)$ sketch of a matroid rank function $v$ using only $O\left(n\cdot \log^2\left(n\right)\right)$ value queries.
\end{proposition}
\begin{proof} (sketch) First, notice that every matroid rank function is well bounded so it is enough to consider this case only. We use the same implementation of the $1$-XOS oracle as in Corollary \ref{cor-submodular} and provide a fast implementation of $CARD(\cdot)$ with an approximation ratio of $1$. Given a set $T$, our implementation finds whether there exists an item $j$ that can be added to $T$ with a marginal value $v(\{j\}+T)-v(T)=1$. Starting from the empty set, $CARD(\cdot)$ will keep adding items to $T$ until there is no item with a marginal of $1$, or until $|T|=k$. Finding an item with a marginal $1$ can be done with $O(\log n )$ queries: consider the marginal value of $N-T$: if it is $0$, by monotonicity there is no item that can be added to $T$ with a marginal of $1$. Else, partition the set of items $N-T$ into two equal sized sets, find the one with the positive marginal value, and recurse until a single item $j$ with a positive marginal value is found. The number of queries we made to find $j$ is $O(\log n)$. The total number of queries that $CARD(\cdot)$ makes is $O(k\cdot \log n)$. Correctness of the algorithm follows from the augmentation property of the matroid: as long as there is an independent set $I$ with $|I|>|T|$ by the augmentation property of matroids there must be an item $j$ such that $T\cup \{j\}$ is an independent set, and thus the marginal value of $j$ given $T$ is $1$.
Observe that Step \ref{iterations} of the sketching algorithm will only perform $O\left(k\right)=O\left(\log\left(n\right)\right)$ iterations, since $r\in\{0,1\}$. In addition, we can remove stage \ref{remove_light} which is now redundant, as the current implementation of $CARD(\cdot)$ never adds items with marginal value of 0.
The overall running time of the loop in \ref{loop} is $O\left(n\cdot \log\left(n\right)\right)$: the running time of each iteration of $CARD(\cdot)$ is $O\left(d_{i}\cdot \log\left(n\right)\right)$, where $d_{i}$ is the rank of the remaining items in $N'$ on the $i^{th}$ iteration of \ref{loop}. Note that $\underset{i} \sum{d_{i}}\leq n$ since the sets are disjoint and clearly $\underset{T_{i}\in \mathcal T_{k}} \sum{|T_{i}|}\leq n$. We conclude that the running time of Step \ref{loop} is indeed $O\left(n\cdot \log\left(n\right)\right)$, and from Lemma \ref{queries_number} we know that the additional number of value queries used by the algorithm is $O\left(n\right)$, thus the algorithm uses a total of $O\left(n\cdot \log^2\left(n\right)\right)$ value queries and achieves an approximation ratio of $O(\sqrt{n}\cdot \log\left(n\right))$.
\end{proof}
\section{The Construction}
We now describe an algorithm that sketches a well-bounded valuation $v$. Using this algorithm we can obtain sketches for any valuation function by applying Lemma \ref{lemma-well-bounded-reduction}. Two oracles are used in this algorithm, and we assume that they are given to us as black boxes. The first oracle $CARD(N',k)$ gets as input a subset of the items $N'\subseteq N$ and an integer $k$ and returns a subset $T\subseteq N'$ such that $v(T)\geq \frac {\max_{S\subseteq N',\,|S|\leq k}v\left(S\right)} \alpha$ (i.e., maximization subject to a cardinality constraint).
The second oracle $\beta-XOS(S)$ gets a bundle $S$ and returns a $\beta$-approximate XOS clause of $S$ with respect to the valuation $v$. We will use known implementations of these oracles to obtain our efficient sketches for subadditive and submodular valuations. In the appendix we also show how to obtain sketches for matroid rank function with only $\tilde O(n)$ value queries. For that we provide a new fast and simple implementation of $CARD(N',k)$ for this class. Since $v$ is well bounded we may assume that for all $j$, $1\leq v(\{j\})\leq n^2$.
\subsubsection*{The Sketch Construction Algorithm}
\begin{enumerate}
\item \label{iterations} For every $k=\sqrt n, 2\cdot \sqrt n, 4\cdot \sqrt n, \ldots, n$ and $r=1,2,4,\ldots, n^2$:
\begin{enumerate}
\item Let $\mathcal T_{k,r}=\emptyset$. Let the set of ``heavy'' items be $H_{k,r}=\left\{ i\in N\,|\; v\left(\left\{ i\right\} \right)\geq\frac{k\cdot r}{\sqrt{n}}\right\} $ and the set of ``light'' valued items be $L_r=N-H_{k,r}$. Let $N'=L_r$.
\item \label{loop} While $T=CARD(N',k)$ is such that $v(T)\geq \frac {k\cdot r} {2\alpha}$:
\begin{enumerate}
\item \label{remove_light} Let $a$ be the additive valuation that $\beta-XOS(T)$ returns. Let $T'\subseteq T$ be the set of items $j\in T$ with $a(\{j\})\geq \frac {r} {4\alpha\cdot \beta}$.
\item Add $T'$ to $\mathcal T_{k,r}$. Let $N'=N'-T'$.
\end{enumerate}
\end{enumerate}
\item For every item $j$, return $v(\{j\})$. For every $k,r$ iterated above return $\mathcal T_{k,r}$.
\end{enumerate}
The above algorithm constructs a sketch for a valuation $v$. For a given bundle $S$, we will use the algorithm's output to return its approximate value:
\begin{enumerate}
\item Let $max-singleton= \max_{j\in S}v\left(\{j\}\right)$,
$max-intersection= \max_{k,r,T\in \mathcal T_{k,r}}|T\cap S|\cdot\frac{r}{4\alpha\cdot\beta}$.
\item Return $\max\{max-singleton,\;max-intersection\}$.
\end{enumerate}
\subsubsection*{Analysis}
We now analyze the performance of the algorithm. We start with an helpful claim:
\begin{claim} \label{subsets}
Fix a well bounded subadditive valuation $v$, and consider some bundle $S$, and let $S^{r'}$ be its $r$-projection, where $|S^{r'}|=|S^{r'}\cap L_r|=k'$. Then the algorithm will keep adding subsets to $\mathcal T_{k',r'}$ as long as $|S^{r'}\cap\left(\underset{i}{\bigcup}T_{i}\right)|<\frac{k'}{2}$. Furthermore, at the end of the algorithm we have that $|\mathcal T_{k',r'}|\leq \frac{4\alpha\cdot \beta\cdot \sqrt{n}}{2\beta-1}$.
\end{claim}
\begin{proof}
Consider the disjoint subsets $T_{1},\ldots,T_{l}\in \mathcal T_{k',r'}$. Since we assume $|S^{r'}|=|S^{r'}\cap L_r|=k'$, as long as we did not use at least half of the items in $S^{r'}$ to construct the sets in $\mathcal T_{k',r'}$, then there is some subset $T\subseteq N'$ with $v\left(T\right)\geq\frac{k'\cdot r'}{2}$ and thus the oracle $CARD(\cdot)$ must return some set with value at least $\frac{k'\cdot r'}{2\alpha}$. We conclude that the loop in Step \ref{loop} will not stop before we get that $|S^{r'}\cap\left(\underset{i}{\bigcup}T_{i}\right)|\geq \frac{k'}{2}$.
For the second part of the claim, note that every subset $T$ returned by $CARD(\cdot)$ is valued at least $\frac{k\cdot r}{2\alpha}$. In Step \ref{remove_light} we remove from $T$ the items with value less than $\frac{r}{4\alpha\cdot \beta}$, thus the maximal loss of value in this step will be $\frac{k\cdot r}{4\alpha\cdot \beta}$ and we get that $v\left(T'\right)\geq v\left(T\right)- \frac{k\cdot r}{4\alpha\cdot \beta}= \frac{\left(2\beta-1\right)\cdot k\cdot r}{4\alpha\cdot \beta}$. Observe that the sets consist of only ``light'' items, and since we only deal with subadditive valuations $v\left(\underset{j\in T'} \bigcup{j}\right)\leq \underset{j\in T'} \sum{v\left(\{j\}\right)}$ and we conclude that $|T'|\geq \frac{\frac{\left(2\beta-1\right)\cdot k\cdot r}{4\alpha\cdot \beta}}{\frac{k\cdot r}{\sqrt{n}}}=\frac{\left(2\beta-1\right)\cdot \sqrt{n}}{4\alpha\cdot \beta}$. Since $|N|=n$ we can upper bound the number of the disjoint subsets with $\frac{4\alpha\cdot \beta\cdot \sqrt{n}}{2\beta-1}$.
\end{proof}
The next two lemmas analyze the running time and the approximation ratio of the sketch as a function of the properties of the oracles. We will later consider specific implementations of these oracles and will provide explicit bounds on the performance of the algorithm for each class.
\begin{lemma} \label{queries_number}
For every well bounded subadditive valuation $v$ there is an implementation of the algorithm in which the number of times that the algorithm calls $CARD(\cdot)$ and $\beta-XOS(\cdot)$ is $O(\alpha \cdot \sqrt{n}\cdot \log^2 n)$. In addition, the algorithm makes $O(n)$ value queries.
\end{lemma}
\begin{proof}
Stage \ref{iterations} of the algorithm performs $k\cdot r$ iterations, where $k=O\left(\log\left(n\right)\right)$ and $r=O\left(\log\left(n\right)\right)$. We claim that the loop in \ref{loop} iterates $\frac{4\alpha\cdot \beta\cdot \sqrt{n}}{2\beta-1}$ times: we know from Claim \ref{subsets} that for every $k',r'$, $|\mathcal T_{k',r'}|\leq \frac{4\alpha\cdot \beta\cdot \sqrt{n}}{2\beta-1}$ and this is also the maximal number of iterations of the loop. Overall the number of calls to the oracles $CARD(\cdot)$ and $\beta-XOS(\cdot)$ is indeed $O(\alpha \cdot \sqrt{n}\cdot \log^2 n)$. The additional value queries the algorithm uses are to determine the "heavy" items set and to return the value of every singleton, which overall requires $n$ singleton value queries.
\end{proof}
\begin{lemma}
For every well bounded subadditive valuation $v$, the algorithm returns an $O(\alpha\cdot \beta\cdot \sqrt{n} \cdot \log n)$-sketch of $v$.
\end{lemma}
\begin{proof}
Fix a bundle $S$. By Lemma \ref{core} the core of $S$ gives us a $\log\left(n\right)$-approximation to $v(S)$, and we claim that the algorithm constructs sketches of each $r$-projection of $S$, and of $C\left(S\right)$ in particular. Let $|C\left(S\right)|=k'$ and $r'$ be the projection index of the core set, then $k'\cdot r' \leq a\left(C\left(S\right)\right)\leq 2\cdot k'\cdot r'$, where $a$ is the additive valuation which is an approximate XOS clause of $S$. Note that if $H_{k',r'}\cap S\neq \emptyset$ then there exists a singleton that gives $O\left(\sqrt{n}\right)$-approximation for $a\left(C\left(S\right)\right)$, and its value will be returned in stage 2 of the algorithm (observe that if $k'<\sqrt{n}$ then by the pigeonhole principle and subadditivity such ``heavy'' item must exist).
If $H_{k',r'}\cap S= \emptyset$, Claim \ref{subsets} gives us that $|C\left(S\right)\cap\left(\underset{i}{\bigcup}T_{i}\right)|\geq \frac{k'}{2}$ and also that $|\mathcal T_{k',r'}|\leq\frac{4\alpha\cdot \beta\cdot \sqrt{n}}{2\beta-1}$. Thus, by the pigeonhole principle, $\exists T_{i}\in\mathcal T_{k',r'}$ such that $|T_{i}\cap C\left(S\right)|\geq \frac{\frac{k'}{2}}{\frac{4\alpha\cdot \beta\cdot \sqrt{n}}{2\beta-1}}=\frac{\left(2\beta-1\right)\cdot k'}{8\alpha\cdot \beta\cdot \sqrt{n}}$. Observe that the value that the algorithm outputs is at least $|T_{i}\cap C\left(S\right)|\cdot\frac{r'}{4\alpha\cdot\beta} \geq \frac{\left(2\beta-1\right)\cdot k'}{8\alpha\cdot \beta\cdot \sqrt{n}}\cdot\frac{r'}{4\alpha\cdot\beta}\geq\frac{\left(2\beta-1\right)\cdot \frac{a\left(C\left(S\right)\right)}{2}}{32\alpha^{2}\cdot \beta^{2}\cdot \sqrt{n}}\geq \frac{ v\left(S\right)}{\frac{128\alpha^{2}\cdot\beta^{3}}{2\beta-1}\cdot \sqrt{n}\cdot \log\left(n\right)}$.
\end{proof}
\subsubsection*{Applications}
By providing specific efficient implementations for the oracles used in the algorithm, we get our main results:
\begin{corollary}\label{cor-submodular}
There exists an algorithm that produces an $\tilde O(\sqrt n)$ sketch of every monotone submodular valuation $v$ using only $O\left(n^\frac{3}{2}\cdot \log^{3}\left(n\right)\right)$ value queries.
\end{corollary}
\begin{proof}
We first prove for a well-bounded valuation $v$. We use a recent result \cite{BV14} that presents an implementation for the oracle $CARD(\cdot)$ for monotone submodular functions with $\alpha=\frac{1}{1-\frac{1}{e}-\epsilon}$ while using $O\left(\frac{n}{e}\cdot \log\left(\frac{n}{e}\right)\right)$ value queries\footnote{Another possibility is using the classic greedy algorithm \cite{NWF78} which gives $\left(1-\frac{1}{e}\right)$-approximation, but its running time is $O\left(n^2\right)$ which results in a slower overall running time.}.
In addition, the implementation of the $\beta-XOS(\cdot)$ oracle for monotone submodular functions is straightforward - given a bundle $S$, we return an additive function $a$ that assigns every item in $S$ with its marginal value, that is: $v\left(i_{1}\right)=v\left(S\right)-v\left(S\backslash \{i_{1}\}\right),\ldots ,v\left(i_{j}\right)=v\left(S\backslash \underset{k<j} \bigcup{i_{k}}\right)-v\left(S\backslash \left(\underset{k<j} \bigcup{i_{k}}\cup\{i_{j}\}\right)\right)$, and get an XOS clause with $\beta=1$ \cite{DNS05}. That requires additional $O\left( n\right)$ value queries. Combining with claim 3.1, we conclude that our algorithm makes $O\left(n^\frac{3}{2}\cdot \log^{3}\left(n\right)\right)$ value queries, and achieves an approximation ratio of $O(\sqrt n \cdot \log n)$. The result for any valuation follows by applying Lemma \ref{lemma-well-bounded-reduction}.
\end{proof}
We do not know whether we can construct sketches for general submodular functions using only $\tilde O(n)$ value queries. However, if $v$ is further known to be a matroid rank function we are able to sketch it using only $\tilde O(n)$ value queries. We postpone this result to the appendix.
\begin{corollary}
There exists an algorithm that produces an $\tilde O(\sqrt n)$ sketch of every subadditive valuation $v$ using $O(n)$ demand and value queries.
\end{corollary}
\begin{proof}
We first prove the result for a well-bounded valuation $v$. For the implementation of the oracle $CARD(\cdot)$ we use a former result \cite{DPS11} that gives an approximation of $\alpha =2$ while using $O\left(\log\left(n\right)\right)$ demand queries. The oracle $\beta-XOS(\cdot)$ can be implemented using a construction that returns an approximate XOS clause with $\beta =O(\log\left(n\right))$ while using $O\left(\log\left(n\right)\right)$ demand queries \cite{D07}. We therefore achieve an approximation ratio of $ O(\sqrt{n}\cdot \log^3 n)$ using $O\left(\sqrt{n}\cdot \log^3\left(n\right)\right)$ demand queries and $O(n)$ value queries. The result for a general valuation $v$ follows by applying Lemma \ref{lemma-well-bounded-reduction}.
\end{proof}
\section{Preliminaries}
In this paper we consider valuation functions $v:\,2^{N}\rightarrow\mathbb{R}_{+}$ where $N=\left\{ {1},\ldots,{n}\right\} $. We assume that $v$ is normalized ($v(\emptyset)=0$) and monotone (for every $S\subseteq T\subseteq N$ it holds that $v\left(S\right)\leq v\left(T\right)$). In addition, we consider the following restrictions on the valuation functions:
\begin{itemize}
\item \textbf{Subadditivity:} for every $S,T\subseteq N, v\left(S\right)+v\left(T\right)\geq v\left(S\cup T\right)$.
\item \textbf{$\beta$-XOS:} There exists a set of additive functions $a_{1},\ldots,a_{t}$ such that for every bundle $S$ the following two conditions hold:
\begin{itemize}
\item
Let $a_{l}\left(S\right)=\max_{r}a_{r}(S)$, then $a_{l}\left(S\right)\leq v\left(S\right)\leq \beta\cdot a_{l}\left(S\right)$.
\item
For every $T\subseteq S$, $v\left(T\right)\geq \underset{j\in T} \sum{a_l\left(j\right)}$.
\end{itemize}
Given a valuation function which is $\beta-XOS$ and a bundle $S$, an approximate XOS clause of $S$ will be an additive function for which both the conditions above hold. A function $v$ which is $\beta-XOS$ with $\beta=1$ is called simply an XOS function.
\item \textbf{Submodularity:} for every $S,T\subseteq N, v\left(S\right)+v\left(T\right)\geq v\left(S\cup T\right)+v\left(S\cap T\right)$. An equivalent and more intuitive definition of submodular functions
is the property of decreasing marginal values: for every $X\subseteq Y\subseteq N$,
$z\in N\backslash Y$ it holds that $v\left(Y\cup\left\{ z\right\} \right)-v\left(Y\right)\leq v\left(X\cup\left\{ z\right\} \right)-v\left(X\right)$.
\item \textbf{Matroid rank function:} Let $M=\left(\mathcal N, \mathcal I\right)$ be a matroid. The rank function of the matroid, $r:\,2^{\mathcal N}\rightarrow\mathbb{N}$, is defined for every $S\subseteq \mathcal N$ as $r\left(S\right)=\max_{T\subseteq S}\{|T|,\,T\in\mathcal I\}$.
\end{itemize}
Every XOS function is also a subadditive function, and every submodular function is also an XOS function \cite{LLN01}. Furthermore, every subadditive function is $O(\log n)$-XOS \cite{D07}. It is also well known that every matroid rank function is submodular.
We say that a function $\tilde v:2^N\rightarrow \mathbb R_+$ is an $\alpha$-\emph{sketch} of $v$ if $\tilde v$ can be represented in $poly(n)$ space and for every bundle $S$ we have that $\frac {v(S)} \alpha \leq \tilde v(S) \leq v(S)$. We assume that $v$ is represented by a black box that can answer a specific type of queries. One type is \emph{value query}: given $S$, return $v(S)$. Another type is \emph{demand query}: given prices per item $p_1,\ldots, p_j$, return a most profitable bundle $S\in \arg\max_T v(T)-\Sigma_{j\in T}p_j$.
Recall that a sketch $\tilde v$ has to be represented in $poly(n)$ space. However, if the ratio $\frac {\max_j{v(\{j\})}} {\min_j{v(\{j\})}}$ is large (e.g., super exponential in $n$), even writing down approximate values for $\max_j{v(\{j\})}$ and $\min_j{v(\{j\})}$ requires too many bits. To this end, we say that a valuation $v$ is \emph{well bounded} if $\frac {\max_j{v(\{j\})}} {\min_j{v(\{j\})}}\leq n^2$. We first show that we can focus on sketching well-bounded valuations (a non-algorithmic version of this lemma appeared in \cite{BDFK+12}):
\begin{lemma}\label{lemma-well-bounded-reduction}
Let $A$ be an algorithm that produces an $\alpha$-sketch of well bounded subadditive valuations. Then, there is an algorithm $A'$ that produces an $O(\alpha)$-sketch for any subadditive valuation. Moreover, $A'$ runs $A$ on $n$ sets of items $T_1,\ldots, T_n$, where each item $j$ appears in $O(1)$ sets, in addition to $n$ value queries.
\end{lemma}
Denote by $q$ the number of queries that $A$ makes. Notice that as long as $q\geq n$ queries, the number of queries that $A'$ makes is $O(q)$.
\begin{proof}(proof of Lemma \ref{lemma-well-bounded-reduction})
We construct the $T_i$'s as follows. Without loss of generality, assume that $v(\{1\})\geq v(\{2\}) \geq \ldots \geq v(\{n\})$. For each item $j$, let $N_j=\{j'|\frac {v(\{j\})} {v(\{j'\}}\leq n^2,j\leq j' \}$. Let $T_1=N_1$. Let $j_2$ be the item with smallest index such that $\frac {v(\{1\})} {v(\{j_2\})}\geq \frac n 2$. Let $T_2=N_{j_2}$. Similarly, Let $j_3$ be the item with the smallest index such that $\frac {v(\{j_2\})} {v(\{j_3\})}\geq \frac n 2$, let $T_3=N_{j_3}$ and so on.
Notice that the conditions in the statement of the lemma hold for $T_1,\ldots, T_n$. The sketch that $A'$ produces is obtained by running $A$ on each set $T_i$ separately. All that is left to show is that we can compute any set $S$ to within a factor of $O(\alpha)$. Towards this end, consider some set $S$ and let $j$ be the item with the smallest index in $S$. Let $i$ be the largest index such that $j\in N_i$. Let $S'=S-N_{i}$. Since by assumption we have that we get an $\alpha$-sketch for $S\cap N_{i}$, it suffices to show is that $v(S')\leq \frac {v(S)} {2}$ and this concludes the proof since by subadditivity we then have that $v(S\cap N_{i})\geq \frac {v(S)} {2}$.
Observe that by construction for each $j'\in S'$ we have that $\frac {v(\{j\})} {v(\{j'\})}\geq \frac {n^2 } {\frac n 2}=2n$. Thus, by subadditivity it holds that $v(S')\leq \Sigma_{j'\in S'}v(\{j'\})\leq n\cdot \frac {v(\{j\})} {2n} \leq \frac {v(S)} {2}$, as needed.
\end{proof}
\begin{definition}
Fix a $\beta$-XOS function $v$. Let $S\subseteq N$ be a bundle, and fix some approximate clause $a$ of $S$. For $r\in \mathbb{N}$, the \emph{$r$-projection} of $S$, denoted by $S^{r}$, is defined as
$S^{r}=\{i\in S\,|\; r=2^{l}\leq a\left(i\right)< 2^{l+1}=2\cdot r\}$.
We also define the \emph{core} of $S$, denote $C\left(S\right)$, as the highest valued $r$-projection of $S$, that is: $C\left(S\right)=S^{r'}$, where $a\left(S^{r'}\right)=\max_{r}a\left(S^{r}\right)$.
\end{definition}
\begin{claim} \label{core}
For every bundle $S$ and a well bounded valuation $v$, $v\left(C\left(S\right)\right)\geq a\left(C\left(S\right)\right)\geq \frac{v\left(S\right)}{\beta\cdot 2\log\left(n\right)}$, where $a$ is the additive valuation which is an approximate XOS clause of $S$.
\end{claim}
\begin{proof}
Let $y$ be the number of non-empty projections of $a$.
Recall that $a\left(S\right)\leq \sum_{r}a\left(S^{r}\right)\leq \beta \cdot a\left(S\right)$,
and thus by the pigeonhole principle
we conclude that $\exists r\; s.t.\; a\left(S^{r}\right)\geq\frac{a\left(S\right)}{y}$
and that same $r$ satisfies $v\left(S^{r}\right)\geq a\left(S^{r}\right)\geq\frac{a\left(S\right)}{y}\geq \frac{v\left(S\right)}{\beta \cdot y}$. Since $v$ is well bounded and since the projections were defined on the ranges between powers of $2$, there are at most $2\log\left(n\right)$ non-empty projections. The claim follows by the discussion above.
\end{proof}
\section{Introduction}
In this paper we tackle the problem of sketching valuation functions. Consider a function $v:2^N\rightarrow \mathbb R_+$ that is defined over a ground set of $N$, $|N|=n$, elements. Representing $v$ requires specifying $2^n$ numbers, which is too large for most problems. For example, if the function is submodular then we usually require that optimization algorithms will run in time $poly(n)$.
One way to overcome the representation obstacle is to assume that $v$ is given as a black box and develop algorithms that query $v$ no more than $poly(n)$ times. Goemans et al \cite{GHIM09} suggest the following alternative approach: instead of $v$ we can use $\tilde v$ -- a function in which for all $S$, $\tilde v(S)$ is close to $v(S)$ and $\tilde v$ can be succinctly described. More formally, $\tilde v:2^N\rightarrow \mathbb R_+$ is an $\alpha$-\emph{sketch}\footnote{The paper \cite{GHIM09} says that in this case $\tilde v$ \emph{approximates $v$ everywhere}, but in this paper we use the term \emph{sketch} that was suggested later in \cite{BDFK+12}.} of $v$ if $\tilde v$ can be represented in polynomial space and for every bundle $S$ we have that $\frac {v(S)} \alpha \leq \tilde v(S) \leq v(S)$.
Goemans et al \cite{GHIM09} show that not only that every submodular function $v$ admits an $\tilde O(\sqrt n)$-sketch, but also that if $v$ is given to us as a black box we can find this sketch by making only $poly(n)$ value queries (given $S$, what is $v(S)$?). Their construction is technical and involved. Very roughly speaking, existence of a good sketch is proved by considering a polymatroid associated with $v$, and finding an ellipsoid $E$ that approximates well the polymatroid using John's theorem. Algorithmically finding an ellipsoid that approximates $E$ is done by repeatedly solving several instances of a certain convex optimization problem, to get closer and closer to $E$. We refer the reader to \cite{GHIM09} for full details on their approach.
Balcan and Harvey \cite{BH11} prove that the construction above is essentially the best possible, in the sense that there are submodular functions (which are even rank functions of a certain family of matroids) for which no $n^{\frac 1 3-\epsilon}$-sketch exists. Thus, the focus of this research line has switched to finding good sketches for larger classes of valuations\footnote{A very related question is learning valuation functions -- see, e.g., \cite{BFIW11,BH11,FV13, FK14}.}. Balcan et al. \cite{BFIW11} and Badanidiyuru et al \cite{BDFK+12} show that $\tilde O(\sqrt n)$-sketches exist for the larger class of subadditive valuations. Furthermore, these sketches can be found using a polynomial number of demand queries (the power of value queries is weak in this setting -- it is known that with polynomially many value queries only trivial sketches of subadditive functions can be obtained \cite{BDFK+12}). These papers follow the basic ellipsoidal-based approach of \cite{GHIM09}, developing it even further to support subadditive functions.
Our contribution is two-fold. First, we present a radically different proof for all the above positive results. Our proof is much simpler, uses first principles only, and in particular avoids complicated geometric arguments. Moreover, although all of the above algorithmic sketches run formally in polynomial time, the hidden constants are very large. A concrete benefit of our proof is that it enables us to develop much faster sketching algorithms, hence joining the recent effort of obtaining faster versions of fundamental algorithms in submodular optimization (e.g, \cite{BV14, BBCL14}).
\begin{theorem} The following two statements hold:
\begin{itemize}
\item There exists an algorithm that finds $\tilde O(\sqrt n)$-sketch for any submodular function using $O\left(n^\frac{3}{2}\cdot \log^{3}\left(n\right)\right)$ value queries.
\item There exists an algorithm that finds $\tilde O(\sqrt n)$-sketch for any subadditive function using $O(n)$ value and demand queries.
\end{itemize}
\end{theorem}
In fact, we only provide and analyze a single algorithm that calls in a black box manner to subroutines such as maximizing a function subject to a cardinality constraint. We obtain both parts of the theorem by using known implementations of these black boxes. In addition, we use the same algorithm to obtain $\tilde O(\sqrt n)$-sketch for matroid rank functions using only $\tilde O(n)$ value queries.
\subsection*{An Overview of our Construction}
We now give a brief overview of our construction. Recall that a valuation $a$ is \emph{additive} if for every set $S$ we have that $a(S)=\Sigma_{j\in S}a(\{j\})$. Also, recall that a valuation function $v$ is XOS if there exist additive valuations $a_1,\ldots, a_t$ such that $v(S)=\max_ra_r(S)$. (It is known that every XOS function is also a subadditive function, and every submodular function is also an XOS function.) For simplicity of presentation, in the next few paragraphs we focus on the case where for each item $j$ and valuation $a_r$ we have that\footnote{This is less restrictive than it sounds -- not only that this class contains all matroid rank functions, but moreover one can apply standard arguments to show that the existence of an $\alpha$-sketch for this case implies an $\tilde O(\alpha)$-sketch for the general XOS valuations. However, these arguments do not yield an efficient algorithm.} $a_r(\{j\})\in \{0,1\}$.
Let us now discuss a procedure that given a value $k$ correctly sketches all sets $S$ such that $v(S)=k$. If $k\leq \sqrt n$, to approximate the value of a given bundle $S$ it is enough to decide whether there exists a single item $j\in S$ such that $v(\{j\})=1$ or whether for all $j\in S$ we have that $v(\{j\})=0$. In the former case, our sketch can return that the value of $S$ is $1$, and we maintain a $\sqrt n$-approximation. In the latter, we know that $v(S)$ is $0$. Notice that the information needed to decide between the cases can be written down using only $n$ bits, where the $j$'th bit specifies whether $v(\{j\})=1$ or not.
The heart of the proof is the trickier case where $k>\sqrt n$. In this case our sketch will include a maximal number of disjoint bundles $T_1,\ldots, T_l\subseteq N$ where for each bundle $T_i$ we have that $v(T_i)=\frac k 2$ and $|T_i|=\frac k 2$. Observe that since the $T_i$'s are disjoint and since $\frac k 2\geq \frac {\sqrt n} 2$, we have that $l\leq 2\sqrt n$. Consider some bundle $S$ where $v(S)=k$ and let $R\subseteq S$, $|R|=k$, be such that $v(R)=k$. Notice that such $R$ exists since the value of every item in any of the additive valuations is in $\{0,1\}$. Furthermore, this also implies that for each $R'\subseteq R$ we have that $v(R')=|R'|$.
Let $R'=R\cap (\cup_iT_i)$. First, observe that $|R'|\geq k/2$. If this were not the case, we have that $|R-R'|\geq k/2$. This is a contradiction to our assumption that $T_1,\ldots, T_l$ is a maximal disjoint set of bundles such that $v(T_i)=\frac k 2$, since we could have added any subset of size $\frac k 2$ from $R-R'$ to the $T_i$'s. Since there are $l\leq 2\sqrt n$ disjoint sets, there must be one set $T_i$ such that $|R'\cap T_i| \geq \frac {\frac k 2} {2\sqrt n}=\frac {k} {4\sqrt n}$. Thus, for every given $S$ our sketch will return the size of largest intersection between $S$ and the $T_i$'s, and this size is guaranteed to be within a factor of $O(\sqrt n)$ of $v(S)$.
Finally, we run the procedure above for all possible values of $1\leq k\leq n$. Given a bundle $S$, we return the maximum value of $S$ that was computed by any of the sketches constructed by the procedure. By the discussion above, this value is within a factor of $O(\sqrt n)$ of $v(S)$.
\input{prelim}
\input{construction}
\bibliographystyle{plain}
|
\section{Introduction}
Language is the most characteristic trait of human communication
but takes on many heterogeneous forms. Dialects, in particular, are linguistic varieties which differ
phonologically, gramatically or lexically in geographically separated regions \cite{cha98}.
However, despite its fundamental importance and many recent developments, the way language varies spatially is still poorly understood.
Traditional methodological approaches in the study of regional dialects are based on interviews and questionnaires administered by a researcher to a small number (typically, a few hundred) of selected speakers known as informants \cite{lab05}. Based on the answers provided, linguistic atlases are generated that are naturally limited in scope and subject to the particular choice of locations and informants and perhaps not completely free of unwanted influences from the dialectologist.
Another approach is the use of mass media corpora which provide a wealth of information on language usage but suffer from the tendency of media and newspapers to use standard norms (the "BBC English" for example) \cite{bau04} that limits their usefulness for the study of informal local variations.
On the other hand, the recent rise of online social tools has resulted in an unprecedented avalanche of content that is naturally and organically generated by millions or tens of millions of geographically distributed individuals that are likely to speak in vernacular and do not feel constrained to use standard linguistic norms. This, combined with the widespread usage of GPS enabled smartphones to access social media tools provides a unique opportunity to observe how languages are used in everyday life and across vast regions of space.
In this work, we use a large dataset of geolocated Tweets to study local language variations across the world.
Similar datasets have recently been used to map public opinion and social behavior \cite{borge2011structural,tumasjan2010predicting, culotta2010towards,salathe11-1,salathe12-2,kulshrestha2012geographic,mislove2011understanding,hong_2011}
and to analyze planetary language diversity~\cite{moc13}
Preliminary results demonstrating the feasibility of this approach have thus far been limited to considering only few words or just a few geographical areas \cite{eis10,bam12}.
Here, we move beyond the mere proof of concept and provide a detailed global picture of spatial variants for a specific language. For definiteness, we choose Spanish as it is not only one of the most spoken in the world but it has the added advantage of being spatially distributed across several continents\cite{ethnologue,pen00}.
Several other languages such as Mandarin or English have more native speakers or higher supra-regional status but their use is hindered by the limited local availability of Twitter (Mandarin) or a high abundance of homographs that percludes a detailed lexicographic analysis (English).
\section{Methods}
We used the Twitter gardenhose to gather an unbiased sample of all tweets written in Spanish that contained GPS information over the course of over two years. Language detection was performed using the state of the art Chromium Compact Language Detector\cite{chromium} software library.
The resulting dataset contained over $5\times 10^{7}$ geolocated tweets written in Spanish distributed across the world (see Fig. ~\ref{fig_map}). As expected, most tweets are localized in Spain, Spanish America and extensive areas of the United States.
These results are consistent with recent sociolinguistic data~\cite{stewart,moreno}, providing an initial level of validation to our approach. Interestingly, we also find significant contributions from major non-Spanish-speaking cities in Latin America and Western Europe,
likely due to considerable population of temporary settlers and tourists. See Ref. \cite{moc13} for further details and results on this dataset.
\begin{figure*}[!ht]
\begin{centering}
\includegraphics[width=1.0\textwidth]{figs/spanish_crop.pdf}
\caption{\label{fig_map} {\bf Spanish tweet locations.} The overwhelming majority of Spanish tweets are located in Spain and Spanish America but significant contributions arise in certain US states and major Western European and Brazilian cities.}
\end{centering}
\end{figure*}
Traditional approaches in dialectology have preferred rural, male informants while modern analyses include interactions with urban speakers regardless of age and gender. On average, Twitter users are young, urban\cite{smi10} and more likely to be technologically savvy thus providing more modern perspective on the use of language.
To be able to determine exactly what the major local varieties of Spanish are, we use a list of concepts and utterances selected from an exhaustive study of lexical variants in major Spanish-speaking cities. Reference \cite{varilex} provides a comprehensive list of possible words representing several concepts, such as "popcorn", "car", "bus", etc. We selected a subset of concepts that minimized possible semantic ambiguities by ensuring that they contained no common words\footnote{The complete list of words for each concept studied can be accessed at~\url{http://www.bgoncalves.com/languages/spanish.html}}.
In our initial set of Tweets we observed $7.5\times 10^{5}$ geolocated instances where words from our catalogue were used. Individual instances were then agregated geographically into cells of $0.25^{\verb+"+}\times 0.25^{\verb+"+}$, which corresponds to an approximate area of $25\times 25$~km$^2$ in the equator.
\begin{figure}[!h]
\begin{centering}
\includegraphics[width=0.45\columnwidth]{figs/computador.png}\includegraphics[width=0.45\columnwidth]{figs/auto.png}
\caption{\label{fig_popcorn_car} {\bf Geographical distribution of the dominant word for the concepts Computer (left) and Car (right).} Map locations are colored
according to the most common expression found in the corresponding cell. The area of the circle is proportional to the number of tweets}
\end{centering}
\end{figure}
Finally, we define the dominant word for each concept in each geographical cell by a simple majority rule and generate a $M=N_{cells}\times N_{words}$ matrix where element $M_{ij}$ is $1$ when word $j$ is the dominant for a given concept in cell $i$ and $0$ otherwise. The resulting matrix has $N_{cells}=1135$ rows and $N_{words}=131$ columns and constitutes the dataset used for the analysis presented in the remainder of this paper.
\section{Results and discussion}
Figure~\ref{fig_popcorn_car} illustrates two illustrative concepts ('computer' and 'car') that are both associated to multiple utterances. Each utterance is represented with a different color. We draw a circle centered on each cell with an area proportional to the number of tweets that use the corresponding expression\footnote{The corresponding maps for the other concepts in our catalog can be seen at: \url{http://www.bgoncalves.com/languages/spanish.html}}
It is clear from the map that some expressions (\textit{computadora}, \textit{ordernador}, \textit{computador}) are strongly clustered in space, allowing us to easily define regional dialects characterized by the set of dominant words used to express the concepts in our list. Due to the unique resolution of our data we could limit the isoglosses (boundaries) of the regions corresponding to each concept-word with a high degree of precision. However, the isoglosses corresponding to different concepts can overlap and bundle rendering any simple arrangement of dialect areas almost impossible.
The natural way to overcome this difficulty and characterize the various regional dialects present in modern day Spanish is to apply machine learning (ML) approaches to automatically cluster the $M$ matrix and identify which cells are closely related to one another. We start by applying Principal Component Analysis to reduce the dimensionality of the matrix $M$. PCA determines the linear combinations of the columns (features in ML literature) of the matrix that explain most of the variance observed in the rows (observations). We find that by projecting the data onto the $40$ principal components (see Fig.~\ref{PCA}) we are able to maintain over $94\%$ of the variance in the data while reducing by $2/3$ the dimension of the matrix with clear numerical advantages.
\begin{figure*}[!h]
\begin{centering}
\includegraphics[width=0.45\columnwidth]{figs/variance-PCA.pdf}
\caption{\label{PCA} {\bf Cumulative variance explained as a function of the number of components} With $40$ components (vertical blue line) we are able to maintain over $94\%$ of the variance present in the data while significantly reducing the matrix size.}
\end{centering}
\end{figure*}
\subsection{Superdialects}
\begin{figure*}[!h]
\begin{centering}
\includegraphics[width=0.95\columnwidth]{figs/population.pdf}
\caption{\label{population} {\bf Characterization of the two superdialects} A) $f\left(K\right)$ and silhouette statistics as a function of $K$. B) Geographical representation of the two clusters, $\alpha$ (red) and $\beta$ (blue). For visualization purposes we increased the size of each cell. The name of main cities corresponding to superdialect $\alpha$ are shown for clarity. C) Population distribution of the cells corresponding to each cluster.}
\end{centering}
\end{figure*}
The task of identifying meaningful clusters in this matrix is now simplified. We proceed by applying the well known $K$-means\cite{kmeans} algorithm that iteratively refines the position of the centers of $K$ clusters until it finds a stable set of locations. The main dificulty of utilizing this algorithm lies in identifying the correct number $K$ of clusters to utilize. Here, we apply the $f\left(k\right)$ metric introduced by Pham \emph{et al.} to establish the best value for $K$. We run $K$-means with values of $K$ up to $20$ using $100$ different random initializations and depict the results in Fig.~\ref{population} A). For verification purposes, we also plot the value of the Silhouette~\cite{silhouette} of the clusters found with each value of $K$. Both metrics agree that $2$ is the correct number of clusters (both curves show a extremum at that point), leading to two clusters of size $179$ (cluster $\alpha$) and $956$ (cluster $\beta$), respectively.
A geographic plot of the location of the cells belonging to each clusters ($\alpha$ and $\beta$) provides a fundamental clue to their meaning (see Fig.~\ref{population} B)).
Strikingly, we find a profound correlation between location of cells belonging to cluster $\alpha$ (red dots) and areas of high population density. We validate this idea using estimates of the population living within each cell provided by the LandScan dataset. Hence, we plot the population distribution boxplot for each cluster in Fig.~\ref{population} C). The results clearly confirm our intuition. Cluster $\alpha$ corresponds to cells with a typical population that is significantly larger than cluster $\beta$. This suggests a natural lexical bipartition of Spanish into two superdialects. Superdialect $\alpha$ is utilized by speakers in main American and Spanish cities and corresponds to an international variety with a strongly urban component while superdialect $\beta$ is comprised mostly of rural areas and small towns.
Our result provides some evidence that the increasing globalization of major languages leads to an homogenization that is especially apparent for the active lexicon \cite{lopez}. Cities (our superdialect $\alpha$)
naturally exert an intrinsic linguistic centripetal force that favors dialect unification, smoothing possible lexical differences. This leveling process present in all countries (thereby its international denomination) is reinforced by the rapid increase of worldwide social ties and the powerful influence of mass media precisely located in important metropolitan areas (Madrid, Mexico City, Miami)~\cite{tru}.Several other sociolinguistic aspects (prestige, higher educational status) also have a role that is more visible in urban environments.
In contrast, rural areas (superdialect $\beta$) are generally more conservative and keep a larger number of characteristic lexical items and native words. As a result, the dialectal area corresponding to superdialect $\beta$ is much more geographically diverse and can be further split, as discussed below.
\begin{figure}[!h]
\begin{centering}
\includegraphics[width=0.85\columnwidth]{figs/clusters_final.pdf}
\caption{\label{cluster} {\bf Characterization of major cluster $\beta$} Geographical representation of regional dialects. For visualization purposes we increased the size of each cell. Three well separated regions are indicated with dashed lines.}
\end{centering}
\end{figure}
\subsection{Regional dialects}
The size imbalance between the two clusters when combined with our intuition suggest that we can also employ the statistical procedure discussed above
to further divide the largest cluster ($\beta$). We apply $K$-means recursively until the remaining cluster has a similar size to the previous ones. In the end, we obtain six well defined clusters that we display in Fig.~\ref{cluster}.
Clearly, three regions can be distinguished. Yellow dots span a wide area covering Mexico, Central America, the Caribbean
and north-western areas of South America. Green dots correspond to the Southern Cone while blue dots are almost exclusively accumulated
within Spain. The first region is quite diverse. In fact, smaller cells can be aggregated into two additional clusters
(depicted with magenta and orange dots in Fig.~\ref{cluster}).
Interestingly, the magenta and orange dots seem to be localized in the Mexican plateau,
the interior of Central America and Andean Colombia, in contrast with the speech of Venezuela,
the Antilles and coastal areas represented with yellow dots. This division
between highland and lowland varieties agrees with classifications
discussed previously in the linguistics literature \cite{cotton}.
The two regions marked in Fig.~\ref{cluster} partly reflect the settlement patterns and the formal colonial Spanish
administration within the Empire. Conquerors and settlers occupied first the territories of Mexico, Peru and the Caribbean,
and only much later colonists established permanent residence in the Southern Cone, which stayed away
from prestigious linguistic norms. This strong cultural heritage that can still be observed, centuries later,
in our datasets deserves to be further analysed in future works.
\section{Conclusions}
Using a large dataset of user generated content in vernacular Spanish, we analyse the diatopic structure of modern day Spanish language at the lexical level. By applying standard machine learning techniques, we find, for the first time, two large Spanish varieties which are related to, respectively, international and local speeches. We can also identify regional dialects and their approximate isoglosses.
Our results are relevant to empirically understand how languages are used in real life across vastly different geographical regions.
We believe that our work has considerable latitude for further applications in the computational study of linguistics, a field full of rewarding opportunities.
One can envisage much deeper analyses pointing the way towards new developments in sociolinguistic studies
(bilingualism, creole varieties). Our work is based on a synchronous approach to language. However, the possibilities presented by the combination of large scale online social networks with easily affordable GPS enabled devices are so remarkable that might permit us to observe, for the first time, how diatopic differences arise and develop in time.
\section{Acknowledgments} We thank I. Fernández-Ordóñez for useful discussions. This product was made utilizing the LandScan 2007™ High Resolution global Population Data Set copyrighted by UT-Battelle, LLC, operator of Oak Ridge National Laboratory under Contract No. DE-AC05-00OR22725 with the United States Department of Energy. The United States Government has certain rights in this Data Set. Neither UT-BATTELLE, LLC NOR THE UNITED STATES DEPARTMENT OF ENERGY, NOR ANY OF THEIR EMPLOYEES, MAKES ANY WARRANTY, EXPRESS OR IMPLIED, OR ASSUMES ANY LEGAL LIABILITY OR RESPONSIBILITY FOR THE ACCURACY, COMPLETENESS, OR USEFULNESS OF THE DATA SET.
|
\section{Introduction}
Let $\Kn$ be the set of all convex bodies in $\R^n$ having non-empty
interiors,
i.e., $K\in \Kn$ is a convex compact subset of the
$n$-dimensional Euclidean space $\R^n$ with $\inte(K)\ne\emptyset$. As
usual,
we denote by $\langle\cdot,\cdot\rangle$ the inner product on
$\R^n\times\R^n$ with associated Euclidean norm
$\|\cdot\|$. $S^{n-1}\subset\R^n$ denotes the $(n-1)$-dimensional unit
sphere, i.e., $S^{n-1}=\{x\in\R^n : \|x\|=1\}$. The norm associated to
a $o$-symmetric convex body $K\in\Kn$ is denoted by $\|\cdot\|_K$,
i.e., $\|x\|_K=\min\{\lambda\geq 0: x\in\lambda\,K\}$.
For $K\in\Kn$, we write $S_K(\cdot)$ and $h_K(\cdot)$ to denote
its surface area measure and support function, respectively,
and $\nu_K$ to denote the Gau{\ss} map assigning the exterior unit
normal $\nu_K(x)$
to an $x\in \partial_*K$, where $\partial_*K$ consists of all points in the
boundary $\partial K$ of $K$ having an unique outer normal
vector. If the origin $o$ lies in the interior of $K\in\Kn$,
the {\em cone-volume measure of $K$} on $S^{n-1}$ is given by
\begin{equation}
\label{Gauss-cone}
\V_K(\omega)=\int_{\omega}\frac{h_K(u)}n\,dS_K(u)=
\int_{\nu_K^{-1}(\omega)}\frac{\langle x,\nu_K(x)\rangle}n\,d{\mathcal H}_{n-1}(x),
\end{equation}
where $\omega\subset S^{n-1}$ is a Borel set and, in general,
${\mathcal H}_{k}(x)$ denotes the $k$-dimensional
Hausdorff-measure. Instead of ${\mathcal H}_n(\cdot)$, we also
write $\V(\cdot)$ for the $n$-dimensional volume.
The name cone-volume measure stems from the fact that if $K$ is a polytope with facets $F_1,\ldots,F_m$ and corresponding exterior unit normals $u_1,\ldots,u_m$, then
\begin{equation}
\V_K(\omega)=\sum_{i=1}^m\V([o,F_i])\delta_{u_i} (\omega).
\label{eq:cvm_polytope}
\end{equation}
Here $\delta_u$ is the Dirac delta measure on
$S^{n-1}$ at $u\in S^{n-1}$, and for $x_1,\dots,x_m\in\R^n$ and
subsets $S_1,\dots,S_L\subseteq \R^n$ we denote the convex hull of the
set $\{x_1,\dots,x_m, S_1,\dots,S_l\}$ by $[x_1,\dots,x_m,
S_1,\dots,S_l]$. With this notation $[o,F_i]$ is the cone with apex $o$ and
basis $F_i$.
In recent years, cone-volume measures have
appeared and were studied in various contexts, see, e.g.,
F. Barthe, O. Guedon, S. Mendelson and A. Naor \cite{BGMN05},
K.J. B\"or\"oczky, E. Lutwak, D. Yang and G. Zhang \cite{BLYZ12,
BLYZ13},
M. Gromov and V.D. Milman \cite{GrM87},
M. Ludwig \cite{Lud10},
M. Ludwig and M. Reitzner \cite{LuR10},
E. Lutwak, D. Yang and G. Zhang \cite{LYZ05},
A. Naor \cite{Nao07},
A. Naor and D. Romik \cite{NaR03},
G. Paouris and E. Werner \cite{PaW12},
A. Stancu \cite{Sta12},
G.~Zhu \cite{Zhu14a, Zhu14b}.
In particular, cone-volume measure are the subject of
the {\em logarithmic Minkowski problem}, which is the
particular interesting limiting case $p=0$ of the general $L_p$-Minkowski problem -- one
of the central problems in convex geometric analysis. It is the task:
\smallskip \noindent
{\em Find necessary and sufficient conditions for a Borel measure
$\mu$ on $S^{n-1}$ to be the cone-volume measure $\V_K$ of
$K\in\Kn$ (with $o$ in its interior).}
\smallskip
In the recent paper \cite{BLYZ13},
K.J. B\"or\"oczky, E. Lutwak, D. Yang and G. Zhang characterize the cone-volume
measures of origin-symmetric convex bodies. In order to state their
result we have to introduce
the subspace concentration
condition. We say that a Borel measure $\mu$ on $S^{n-1}$ satisfies
the {\em subspace concentration condition} if
for any linear subspace $L\subset \R^n$, we have
\begin{equation}
\label{scc}
\mu(L\cap S^{n-1})\leq \frac{\dim L}{n}\,\mu(S^{n-1}),
\end{equation}
and equality in (\ref{scc}) for some $L$ implies the existence of a complementary linear subspace $\widetilde{L}$ such that
\begin{equation}
\mu(\widetilde{L}\cap S^{n-1})=\frac{\dim\widetilde{L}}{n}\,\mu(S^{n-1}),
\label{eq:sccequality}
\end{equation}
and hence ${\rm supp}\,\mu\subset L\cup \widetilde{L}$, i.e., the
support of the measure ``lives'' in $L\cup \widetilde{L}$.
Via the subspace concentration condition, the logarithmic Minkowski
problem was settled in \cite{BLYZ13} in the
symmetric case.
\begin{theo}[\cite{BLYZ13}] A non-zero finite even Borel measure on
$S^{n-1}$ is the cone-volume measure of an origin-symmetric convex
bodies if and only if it satisfies the subspace concentration
condition.
\label{thm:cvm_symmetric}
\end{theo}
This result was proved earlier for discrete measures on $S^1$, i.e.,
for polygons, by A. Stancu \cite{Sta02, Sta03}.
For cone-volume measures of origin-symmetric polytopes (cf.~\eqref{eq:cvm_polytope}) the necessity
of \eqref{scc} was independently shown by M. Henk, A. Sch\"urmann
and J.M.Wills \cite{HSW05} and
B. He, G. Leng and K. Li \cite{HLL06}.
We recall that the centroid of a $k$-dimensional convex compact set
$M\subset \R^n$ is defined as
$$
c(M)={\mathcal H}_k(M)^{-1}\int_Mx\,d{\mathcal H}_k(x).
$$
The centroid seems also be the right and natural position of the origin
in order to extend Theorem \ref{thm:cvm_symmetric} to arbitrary
convex bodies. In fact, in \cite{HeL14} it was shown by M. Henk and E. Linke
that the necessity part of Theorem \ref{thm:cvm_symmetric} also holds for
polytopes with centroid at the origin, i.e.,
\begin{theo}[\cite{HeL14}]
\label{Henk-Linke-polytope}
Let $K\in \Kn$ be a polytope with centroid at the origin. Then its cone-volume
measure $\V_K$ satisfies the subspace concentration condition.
\end{theo}
Our first result is an extension of Theorem \ref{Henk-Linke-polytope}
to convex bodies.
\begin{theo}
\label{Henk-Linke-body}
Let $K\in\Kn$ with centroid at the origin. Then its cone-volume
measure satisfies the subspace concentration condition.
\end{theo}
While the subspace concentration condition is also the sufficiency
property to characterize cone-volume measures among even non-trivial Borel
measures, the cone-volume measure of a convex body $K\in \Kn$ whose
centroid is the origin should satisfy some extra properties. For
example, in Proposition~\ref{hemi-sphere}
we prove that the measure of any open hemisphere is at least
$\frac1{2n}$.
If the origin is the not the cetroid of the convex body, then the subspace
concentration condition may not hold anymore.
In fact, it was recently shown by G.~Zhu
\cite{Zhu14a} that for unit vectors $u_1,\dots,u_m\in S^{n-1}$ in general position, $m\geq
n+1$, and arbitrary positive numbers
$\gamma_1,\dots,\gamma_m$ there exists a polytope $P$ with outer unit
normals $u_i$ with $\V_P(\{u_i\})=\gamma_i$, $1\leq i\leq m$. In other
words, Zhu settled the logarithmic Minkowski problem for discrete
measures whose support is in general position. In general,
the centroid of such a polytope $P$ is not the origin, and a full characterization of
cone-volume measures of arbitrary polytopes/bodies is still a
challenging and important problem.
We note that \eqref{eq:sccequality} is a kind of condition
on the cone-volume measure which is
independent of the choice of the origin.
\begin{lemma}
\label{general-split}
If $K\in\Kn$ with $o\in\inte K$, and
${\rm supp}\,\V_K\subset L\cup\widetilde{L}$ for the proper complementary linear subspaces
$L,\widetilde{L}\subset\R^n$, then
$$
\V_K(L\cap S^{n-1})= \frac{\dim L}{n}\,\mu(S^{n-1}).
$$
\end{lemma}
Let us provide the simple argument leading to Lemma~\ref{general-split}. It follows from Minkowski's uniqueness theorem that $K=M+\widetilde{M}$ where
$M$, $\widetilde{M}$ are
contained in affine spaces orthogonal to $L$, $\widetilde{L}$, respectively. By
Fubini's theorem, we conclude \eqref{eq:sccequality} for $\V_K$ and
the subspaces $L,\widetilde{L}$.
For a convex body $K$ containing the origin in its interior,
E. Lutwak, D. Yang and G. Zhang \cite{LYZ01} defined the $\SL(n)$
invariant quantity $\U(K)$ as an integral over subsets
$(u_1,\ldots,u_n)\in S^{n-1}\times \cdots\times S^{n-1}$, by
$$
\U(K)=\left(\int_{u_1\wedge\ldots\wedge u_n\neq 0}d\V_K(u_1)\cdots d\V_K(u_n)\right)^\frac{1}{n},
$$
where $u_1\wedge\ldots\wedge u_n\neq 0$ means that the vectors
$u_1,\ldots,u_n$ are linearly independent. The $U$-functional has been proved useful in
obtaining strong
inequalities for the volume of projection bodies \cite{LYZ01}. For
information on projection bodies we refer to the books by Gardner
\cite{Gar95} and Schneider \cite{Sch93}, and for more
information on the importance of centro-affine functionals we refer to
C. Haberl and L. Parapatits \cite{HaberlParapatits:2013, LuR10} and the references within.
We readily have $\U(K)\leq V(K)$, and equality holds if and only if
$V_K(L\cap S^{n-1})=0$
for any non-trivial subspace of $\R^n$
according to K.J. B\"or\"oczky, E. Lutwak, D. Yang and G. Zhang
\cite{BLYZ14}. As a consequence of Theorem \ref{Henk-Linke-body} we prove here a
lower bound on $\U(K)$ in terms of $V(K)$ which was conjectured in
\cite{BLYZ14}.
\begin{theo}
\label{Henk-Linke-UK}
Let $K\in\Kn$ with centroid at the origin. Then
$$
U(K)\geq \frac{(n!)^{1/n}}{n}\,V(K),
$$
with equality if and only if $K$ is a parallepiped.
\end{theo}
In particular, $\U(K)>(1/\mathrm{e})V(K)$. For polytopes,
Theorem~\ref{Henk-Linke-UK} was shown in \cite{HeL14}, where the
special cases if $K$ is an origin-symmetric polytope, or if $n=2,3$ were verified by
B. He, G. Leng and K. Li \cite{HLL06}, and G. Xiong \cite{Xio10}, respectively.
In order to state another consequence of Theorem~\ref{Henk-Linke-UK}
we need the notation of an {\em isotropic measure}, going back to
K.M. Ball's reformulation of the Brascamp-Lieb inequality in
\cite{Bal91}. A Borel measure $\mu$ on $S^{n-1}$ is called {\it
isotropic} if
\begin{equation*}
\label{isotropic}
{\rm Id}_{n}=\int_{S^{n-1}}u\otimes u\,d\mu(u),
\end{equation*}
where ${\rm Id}_{n}$ is the $n\times n$-identity matrix and $u\otimes
u$ the standard tensor product, i.e., $u\otimes
u = u\,u^\intercal$.
Equating traces shows that for an isotropic measure $\mu(S^{n-1})=n$.
The subspace concentration condition of a Borel measure $\mu$ on
$S^{n-1}$ is equivalent to have an {\it isotropic} normalized linear
image of $\mu$, i.e., that is, there exists a $\Phi\in{\rm GL}(n)$ such that
\begin{equation}
\label{lineariso}
{\rm Id}_{n}=\frac{n}{\mu(S^{n-1})}\int_{S^{n-1}}\frac{\Phi u}{\|\Phi u\|}\otimes \frac{\Phi u}{\|\Phi u\|}\,d\mu(u).
\end{equation}
The equivalence in this general form is due to K.J. B\"or\"oczky,
E. Lutwak, D. Yang and G. Zhang \cite{BLYZ14}, while the
discrete case was earlier handled by E. A. Carlen,
and D. Cordero-Erausquin \cite{CCE09}, and J. Bennett, A. Carbery,
M. Christ and T. Tao \cite{BCCT07} in their study of the
Brascamp-Lieb inequality. Moreover, the case of a measure $\mu$ when strict inequality holds for all subspaces in \eqref{scc} is due to
B. Klartag \cite{Kla10}. Isotropic measures on $S^{n-1}$ are discussed also e.g. in
F. Barthe \cite{Bar98,Bar04}, E. Lutwak, D. Yang and G. Zhang \cite{LYZ04, LYZ07}.
We note that isotropic measures on $\R^n$ play a central role in the
KLS conjecture by R. Kannan, L. Lov\'asz and M. Simonovits
\cite{KLM95}, see, e.g.,
F. Barthe and D. Cordero-Erausquin \cite{BCE13}, O. Guedon and E. Milman \cite{GuM11} and B. Klartag \cite{Kla09}.
Now from Theorem \ref{Henk-Linke-UK} and by the equivalence \eqref{lineariso} we
immediately conclude
\begin{coro} Every convex body $K\in\Kn$ has an affine image,
whose cone-volume measure is isotropic.
\end{coro}
This, in particular, answers a question posed by
E. Lutwak, D. Yang and G. Zhang \cite{LYZ05}.
In order to present stronger stability versions of Theorem
\ref{Henk-Linke-body} and Theorem \ref{Henk-Linke-UK}
we need two notions of distance
between the "shapes" of two convex bodies. Let $K,M\in\Kn$, and let
$K'=K-c(K)$, $M'=M-c(M)$
be their translates whose centroids are the origin. Then we define
\begin{eqnarray*}
\label{delta-hom-def}
\delta_{\rm hom}(K,M)&=&\min\{\lambda\geq 0:\exists t>0,
\;M'\subset t\, K'\subset \mathrm{e}^\lambda M'\},\\
\label{delta-vol-def}
\delta_{\rm vol}(K,M)&=&\frac{\V\big(M'\Delta t\,K'\big)}{\V(M)},\; t=\frac{\V(M)^{1/n}}{\V(K)^{1/n}},
\end{eqnarray*}
where $A\Delta B$ denotes the symmetric difference of two sets, i.e.,
$A\Delta B = A\setminus B\, \cup\, B\setminus A$.
Then both $\delta_{\rm hom}$ and $\delta_{\rm vol}$ are metrics on the
space of convex bodies in $\R^n$ whose volumes are $1$, and centroids are the origin.
\begin{theo}
\label{Henk-Linke-stab}
Let $K\in\Kn$ with centroid at the origin, and let
$$
\V_K(L\cap S^{n-1})> \frac{d-\varepsilon}{n}\,\V(K)
$$
for a non-trivial linear subspace
$L$ with $\dim L=d$ and $\varepsilon\in(0,\varepsilon_0)$. Then there exist
an $(n-d)$-dimensional compact convex set $C\subset L^\bot$, and
a complementary $d$-dimensional compact convex set $M$
such that
$$
\delta_{\rm hom}(K,C+M)\leq \gamma_h\varepsilon^{1/(5n)}\mbox{ \ and \ }
\delta_{\rm vol}(K,C+M)\leq \gamma_v\varepsilon^{1/5},
$$
where $\varepsilon_0,\gamma_h,\gamma_v>0$ depend only on $n$.
\end{theo}
Here $L^\bot$ denotes the orthogonal complement of $L$, and $M$ is
called a complementary compact convex set of $C$, if the linear spaces
generated by $M$ and $C$ are complementary.
Observe that the range of $\varepsilon$, i.e., $\varepsilon_0$, in Theorem \ref{Henk-Linke-stab} has
to depend on the dimension.
For if, let $K\in\Kn$ be a simplex whose centroid is the origin,
and let $L$ be generated by $d$ outer normals of the simplex,
$d\in\{1,\ldots,n-1\}$. Then we have $\V_K(L\cap S^{n-1})=\frac{d}{n+1}\,\V(K)$.
Actually, if $L$ is $1$-dimensional,
then a more precise version of Theorem~\ref{Henk-Linke-stab} holds.
\begin{theo}
\label{Henk-Linke-stab1}
Let $K\in\Kn$ with centroid at the origin, and let
$$
\V_K(L\cap S^{n-1})> \frac{1-\varepsilon}{n}\,\V(K)
$$
for a linear subspace
$L$ with $\dim L=1$ and $\varepsilon\in(0,\tilde{\varepsilon}_0)$. Then there exist
$(n-1)$-dimensional compact convex set $C\subset L^\bot$ with
$c(C)=o$,
and $x,y\in \partial K$ such that $y=-\mathrm{e}^sx$ where $|s|<\tilde{\gamma_v}\varepsilon^{\frac1{6}}$,
$[x,y]+C\subset K$, and
$$
K\subset [x,y]+(1+\tilde{\gamma_h}\varepsilon^{\frac1{6n}})C
\mbox{ \ and \ }
\V(K) \leq (1+\tilde{\gamma_v}\varepsilon^{\frac1{6}})\V([x,y]+C),
$$
where $\tilde{\varepsilon}_0,\tilde{\gamma}_h,\tilde{\gamma}_v>0$
depend only on $n$.
\end{theo}
We use this theorem in order to deduce the following stability version of Theorem~\ref{Henk-Linke-UK}.
\begin{theo}
\label{U(K)stab}
Let $K\in \Kn$ with centroid at the origin, and let
$$
\U(K)\leq (1+\varepsilon) \frac{(n!)^{1/n}}{n}\,\V(K)
$$
for $\varepsilon\in(0,\varepsilon_*)$. Then there exists
a $K$ containing parallepiped $P$, such that for any facet $F$ of $P$, we have
$$
{\mathcal H}_{n-1}(F\cap K)\geq (1-\gamma_*\varepsilon^{\frac1{6}}){\mathcal H}_{n-1}(F),
$$
where $\varepsilon_*,\gamma_*>0$ depend only on $n$. In particular, we
have
$$
(1-\gamma \varepsilon^{\frac1{6n}})P\subset K
\mbox{ \ and \ }
\V(P\backslash K)\leq \gamma \varepsilon^{\frac1{6}}\V(K).
$$
\end{theo}
The paper is organized as follows. In the next section we collect some
basic facts and notations from convexity which will be used later on.
The third section is devoted to the proof of Theorem \ref{Henk-Linke-body}.
In Section \ref{another_property} we show another characetristic
property of cone-volume measures of
convex bodies with centroid at the origin.
The proofs of Theorem \ref{Henk-Linke-stab}, \ref{Henk-Linke-stab1}
are given in Section 8 and are prepared in Sections 5--7.
Finally, in Section
\ref{secU(K)} we prove Theorem \ref{Henk-Linke-UK}.
\smallskip
{\it Acknowledgements.} We are grateful to Rolf Schneider for various
ideas shaping this paper. We also acknowledge fruitful discussions with
Daniel Hug and David Preiss about the Gau\ss-Green theorem.
\section{Preliminaries\label{pre}}
Good general references for the theory of convex bodies are provided
by the books of Gardner\cite{Gar95}, Gruber\cite{Gruberbook}, Schneider\cite{Sch93} and
Thompson\cite{Tho96}.
The support function $h_K : \R^n\to \R$ of convex body $K\in\Kn$ is
defined, for $x\in\R^n$, by
\begin{equation*}
h_K(x)=\max\{\langle x,y\rangle : y\in K\}.
\end{equation*}
A boundary point $x\in\partial K$ is said to have a unit outer normal
(vector) $u\in S^{n-1}$ provided $\langle x,u\rangle= h_K
(u)$. $x\in\partial K$ is called singular if it has more than one unit
outer normal, and $\partial_* K$ is the set of all non-singular
boundary points. It is well known that the set of singular boundary points of a convex body has $\mathcal{H}_{n-1}$-measure equal to $0$.
For each Borel set $\omega\subset S^{n-1}$, the inverse spherical
image of $\omega$ is the set of all
points of $\partial K$ which have an outer unit normal belonging to $\omega$. Since the
inverse spherical image of $\omega$ differs from
$\nu_K^{-1}(\omega)\subseteq\partial_* K$ by a set of $\mathcal{H}_{n-1}$-measure equal to $0$, we will often make no distinction between the two sets.
For $K\in \Kn$ the Borel measure $S_K$ on $S^{n-1}$ given by
\begin{equation*}
S_K(\omega)=\mathcal{H}_{n-1}(\nu_K^{-1}(\omega))
\end{equation*}
is called the (Aleksandrov-Fenchel-Jessen) surface area
measure. Observe that
\begin{equation*}
\V(K)=\V_K(S^{n-1})= \int_{S^{n-1}}\frac{h_K(u)}{n}d\,S_K(u).
\end{equation*}
As usual, for two subsets $C, D\subseteq \R^n$ and reals
$\nu,\mu \geq 0$ the Minkowski combination is defined by
\begin{equation*}
\nu\,C+\mu\,D = \{\nu\,c+\mu\,d : c\in C,\,d\in D\}.
\end{equation*}
By the celebrated Brunn-Minkowski inequality we know that the $n$-th root
of the volume of the Minkowski combination is a concave function.
More precisely, for two convex compact sets
$K_0,K_1\subset\R^n$ and for $\lambda\in[0,1]$ we have
\begin{equation}
\V((1- \lambda)\,K_0+\lambda\, K_1)^{1/n}\geq
(1-\lambda)\,\V(K_0)^{1/n}+ \lambda\,\V(K_1)^{1/n}
\label{eq:brunn_minkowski}
\end{equation}
with equality for some $0<\lambda<1$ if and only if $K_0$ and $K_1$ lie
in parallel hyperplanes or are homothetic, i.e., there exist
$t\in\R^n$ and $\mu\geq 0$ such that $K_1=t+\mu\,K_0$ (see also
\cite{Gardnersurvey}).
Let $f:C\to \R_{> 0}$ be a positive function on an open convex subset
$C\subset\R^n$ with the property that
there exists a $k\in\N$ such that $f^{1/k}$ is concave. Then by the
(weighted) arithmetic-geometric mean inequality
\begin{equation*}
\begin{split}
f((1-\lambda)\, x+\lambda\, y) & =
\left(f^{1/k}((1-\lambda)\, x+\lambda\, y)\right)^k \\ &\geq \left((1-\lambda)
f^{1/k}(x)+\lambda f^{1/k}(y)\right)^k\\ &\geq f^{1-\lambda}(x)\cdot f^{\lambda}(y).
\end{split}
\end{equation*}
This means that $f$ belongs to the class of log-concave functions
which by the positivity of $f$ is equivalent to
\begin{equation*}
\ln f((1-\lambda)\, x+\lambda\, y)\geq (1-\lambda)\ln
f(x)+\lambda\ln f(y)
\end{equation*}
for $\lambda\in [0,1]$. Hence, for all $x,y\in C$ there exists a
subgradient $g(y)\in\R^n$ such that (cf., e.g., \cite[Sect. 23]{Rockafellar:1997ww})
\begin{equation}
\ln f(x)-\ln f(y)\leq\langle g(y), x-y\rangle.
\label{eq:gradient_concave}
\end{equation}
If $f$ is differentiable at $y$, the subgradient is the gradient
of $\ln f$ at $y$, i.e., $g(y)=\nabla \ln f = \frac{1}{f(y)}\nabla f(y)$.
For a
subspace $L\subseteq\R^n$, let $L^\perp$ be its orthogonal
complement,
and for $X\subseteq\R^n$ we denote by $X|L$ its orthogonal
projection onto $L$, i.e., the image of $X$ under the linear map
forgetting the part of $X$ belonging to $L^\perp$.
Here, for a convex body $K\in\Kn$ and a $d$-dimensional subspace $L$,
$1\leq d\leq n-1$, we are interested in the function measuring the volume of
$K$ intersected with planes parallel to $L^\perp$, i.e., in the function
\begin{equation}
f_{K,L}: L \to \R_{\geq 0}\text{ with } x\mapsto \mathcal{H}_k(K\cap
(x+L^\perp)),
\label{eq:function}
\end{equation}
where $k=n-d$ is the dimension of $L^\perp$. By
the Brunn-Minkowski inequality and the remark above, $f_{K,L}$ is a log-concave
on function on $K|L$ which is positive at least in the relative interior of
$K|L$
(cf.~\cite{Ball:1988vo}). $f_{K,L}$ is also called the $k$-dimensional X-ray of $K$
parallel to $L^\perp$ (cf.~\cite{Gar95}).
By well-known properties of concave functions we also know
\begin{prop}\hfill
\begin{enumerate}
\item[{\rm i)}] $\ln f_{K,L}$ -- and thus $f_{K,L}$ -- is continuous on
$\inte(K)|L$. Moreover, $\ln f_{K,L}$ -- and thus $f_{K,L}$ -- are Lipschitzian on
any compact subset of $(\inte K)|L$.
\item[{\rm ii)}] $\ln f_{K,L}$ -- and thus $f_{K,L}$ -- is on $\inte(K)|L$
almost everywhere differentiable, i.e.,
there exists a dense subset $D\subseteq \inte(K)|L$, where $\nabla
f_{K,L}$ exists.
\end{enumerate}
\label{rem:basicfacts_concave}
\end{prop}
\begin{proof} For i) see, e.g., \cite[Theorem
1.5.3]{Sch93}, and for ii) see, e.g., \cite[Theorem
25.5]{Rockafellar:1997ww}.
\end{proof}
Now for $K\in\Kn$ with centroid at $0$, i.e., $c(K)=0$, we have by Fubini's theorem with respect to the decomposition $L\oplus L^{\perp}$
\begin{equation*}
\begin{split}
0 & =\int_K x\, \mathrm{d}\ha_n(x) \\ &= \int_{K|L}\left(
\int_{(\hat{x}+L^\perp)\cap K} \tilde{x}\,
\mathrm{d}\ha_k(\tilde{x})\right) \mathrm{d}\ha_d(\hat{x})\\ & = \int_{K|L}
f_{K,L}(\hat{x})\, c((\hat{x}+L^\perp)\cap K)\,\mathrm{d}\ha_d(\hat{x}).
\end{split}
\end{equation*}
Writing $c((\hat{x}+L^\perp)\cap K)=\hat{x}+\tilde{y}$ with
$\tilde{y}\in L^\perp$ gives
\begin{equation}
\int_{K|L} f_{K,L}(\hat{x})\,\hat{x}\,\mathrm{d}\ha_d(\hat{x}) =0.
\label{eq:centroid}
\end{equation}
\section{Proof of Theorem~\ref{Henk-Linke-body}\label{proof1}}
For the proof of Theorem \ref{Henk-Linke-body} we first establish some
more properties of the function $f_{K,L}$, where we always assume that
$L\subset\R^n$ is a $d$-dimensional linear subspace, $1\leq d\leq
n-1$, with
$k$-dimensional orthogonal complement $L^\perp$.
We recall that
a function $f$ is said to be upper semicontinuous on $K|L$ if whenever
$x,y_m\in K|L$ for $m\in\N$ and $y_m$ tends to $x$, then
\begin{equation*}
f(x)\geq \limsup_{m\to\infty} f(y_m).
\end{equation*}
\begin{lemma}
\label{upper-semicont}
The function $f_{K,L}$ is upper semicontinuous on $K|L$.
\end{lemma}
\begin{proof} Let $x,y_m\in K|L$ for $m\in\N$ be such that
$\lim_{m\to\infty}y_m=x$. According to the Blaschke selection
principle (cf., e.g., \cite{Sch93}), we may assume that the sequence of compact convex sets
$$
C_m=[(y_m+L^\bot)\cap K]-y_m\subset L^\bot
$$
tends to a compact convex set $C\subset L^\bot$ in the Hausdorff
topology. Since the $k$-volume of a compact convex set in $L^\bot$ is
a continuous functional, we have ${\mathcal
H}_k(C)=\lim_{m\to\infty}f_{K,L}(y_m)$. However, $x+C\subset K$, and
therefore
$f_{K,L}(x)\geq {\mathcal H}_k(C)$.
\end{proof}
An immediate consequence is that for sequences from the relative interior of
$K|L$,
$f_{K,L}$ behaves ``continiuously'', i.e.,
\begin{coro}
\label{ray-limit} Let $o\in\inte K$ and $x\in K|L$. Then $\lim_{m\to\infty}f_{K,L}(e^{\frac{-1}m}x)=f_{K,L}(x)$.
\end{coro}
\begin{proof} Since $o\in\inte K$, we get by the concavity of
$f_{K,L}^{1/k}$ that
\begin{equation*}
f_{K,L}(e^{\frac{-1}m}x)\geq e^{\frac{-k}m}f_{K,L}(x).
\end{equation*}
Since $f_{K,L}$ is also upper semicontinuous on $K|L$ by Lemma \ref{upper-semicont}, we conclude the corollary.
\end{proof}
Although the gradient $\nabla f_{K,L}$ might not be bounded, its norm belongs
to the space $L^1(K|L)$ of absolute integrable functions.
\begin{lemma}
\label{nablafL1}
$\|\nabla f_{K,L}\|\in L^1(K|L)$, and thus the function $x\mapsto \langle \nabla f_{K,L}(x),x\rangle$ is in $L^1(K|L)$, as well.
\end{lemma}
\begin{proof} Let $f=f_{K,L}$. Since $\nabla f(x) = \nabla
(f^{\frac1k})^k(x) = kf^{\frac{k-1}k}(x)\nabla f^{\frac1k}(x)$ for
almost all $x\in K|L$, it is sufficient to prove
$\|\nabla h\|\in L^1(K|L)$ for the concave function
$h=f^{\frac1k}$. However, by the Brunn-Minkowski theorem, the graph
$X$ of the function $h$ over $K|L$ is part of the boundary of a
$(d+1)$-dimensional compact convex set. Thus
$$
\int_{K|L}\|\nabla h\|\,d{\mathcal H}_d(x) \leq \int_{K|L}\sqrt{1+\|\nabla h\|^2}\,d{\mathcal H}_d(x)={\mathcal H}_d(X)<\infty. \mbox{ \
$$
\end{proof}
The next two statements, which are the core ingredients of the proof of
Theorem~\ref{Henk-Linke-body} have been proved in the special case of polytopes in
\cite{HeL14}.
\begin{prop}
\label{VKf}
If $o\in\inte K$, then
$$
n\,\V_K(L\cap S^{n-1})=d\,\V(K)+\int_{K|L}\langle \nabla f_{K,L}(x),x\rangle\,d{\mathcal H}_d(x).
$$
\end{prop}
\begin{proof} Let $f=f_{K,L}$, and let $F(x)=f(x)x$ for $x\in K|L$, which is
a Lipschitz vector field on any compact subset of $(\inte K)|L$
(cf.~\eqref{rem:basicfacts_concave} i)). To state the
Gau\ss-Green divergence theorem for Lipschitz vector fields on
Lipschitz domains, we follow W.F. Pfeffer \cite{Pfe12}. Naturally,
$$
E_m=e^{\frac{-1}m}K|L\subset (\inte K)|L
$$
is a compact Lipschitz domain for $m\geq 1$, and hence $\op_\star
E_m=\op E_m$ according to Proposition~4.1.2 in \cite{Pfe12},
where
$\op(E_m)$ denotes the (relative) boundary with respect to the linear space $L$.
Therefore Theorem~6.5.4 in \cite{Pfe12} (going back to H. Federer \cite{Fed45}) yields that
\begin{equation}
\label{divergence-theorem}
\int_{\op E_m}\langle F(x),\nu_{E_m}(x)\rangle\,d{\mathcal H}_{d-1}(x)=
\int_{E_m}{\rm div}F(x)\,d{\mathcal H}_d(x).
\end{equation}
If $y\in \op (K|L)$ then $\nu_{K|L}(y)=\nu_{E_m}(e^{\frac{-1}m}y)$; thus the left hand side of
(\ref{divergence-theorem}) is
\begin{align*}
\int_{\op E_m}\langle F(x),\nu_{E_m}(x)\rangle\,d{\mathcal H}_{d-1}(x)=&
e^{\frac{-(d-1)}m}
\int_{\op (K|L)}\langle F(e^{\frac{-1}m}y),\nu_{K|L}(y)\rangle\,d{\mathcal H}_{d-1}(y)\\
=& e^{\frac{-d}m}
\int_{\op(K|L)}f(e^{\frac{-1}m}y)\langle y,\nu_{K|L}(y)\rangle\,d{\mathcal H}_{d-1}(y).
\end{align*}
Therefore, Corollary~\ref{ray-limit} and the Lebesgue dominated convergence theorem yield
\begin{equation}
\lim_{m\to\infty}\int_{\op E_m}\langle F(x),\nu_{E_m}(x)\rangle\,d{\mathcal H}_{d-1}(x)
=\int_{\op (K|L)}f(y)\langle y,\nu_{K|L}(y)\rangle\,d{\mathcal H}_{d-1}(y).
\label{eq:stepone}
\end{equation}
Now, in order to evaluate the right hand side
let $X=\partial K\cap (L^\bot+\op (K|L))$. Then the set of
smooth points of $\partial K$ in $X$, i.e., $\partial_\star K\cap X$
coincides with the set of points in $\nu_K^{-1}(L\cap S^{n-1})$. In addition,
if $z\in X\cap \partial_\star K$, then $\nu_{K|L}(y)=\nu_K(z)$ for
$y=z|L$, and thus \eqref{Gauss-cone} and \eqref{eq:stepone} give
\begin{eqnarray}
\nonumber
\lim_{m\to\infty}\int_{\op E_m}\langle F(x),\nu_{E_m}(x)\rangle\,d{\mathcal H}_{d-1}(x)
&=&\int_{X}\langle z,\nu_{K}(z)\rangle\,d{\mathcal H}_{n-1}(z)\\
\label{left-div-theo}
&=&n\,\V_K(L\cap S^{n-1}).
\end{eqnarray}
Next, if $\nabla f(x)$ exists at $x\in\inte(K)|L$, then
$$
{\rm div}F(x)=d\,f(x)+\langle x,\nabla f(x)\rangle.
$$
Therefore the right hand side of \eqref{divergence-theorem} is
(cf.~Proposition \ref{rem:basicfacts_concave} ii), Lemma \ref{nablafL1})
$$
\int_{E_m}{\rm div}F(x)\,d{\mathcal H}_d(x)=d\int_{E_m}f(x)\,d{\mathcal H}_d(x)+
\int_{E_m}\langle x,\nabla f(x)\rangle\,d{\mathcal H}_d(x).
$$
Since $\int_{K|L}f(x)\,d{\mathcal H}_d(x)=\V(K)$, we deduce that
\begin{equation}
\label{right-div-theo}
\lim_{m\to\infty}\int_{E_m}{\rm div}F(x)\,d{\mathcal H}_d(x)=dV(K)+
\int_{K|L}\langle x,\nabla f(x)\rangle\,d{\mathcal H}_d(x).
\end{equation}
Combining \eqref{divergence-theorem}, \eqref{left-div-theo} and
\eqref{right-div-theo} completes the proof of the proposition.
\end{proof}
If $K$ is an $o$-symmetric convex body, we know by the Brunn-Minkowski
inequality \eqref{eq:brunn_minkowski} that $f_{K,L}(x)$ attains its maximum at the origin
$0$. Hence, in view of \eqref{eq:gradient_concave} we know that
$\langle \nabla f_{K,L}(x),x\rangle\leq 0$ for almost every $x\in
K|L$. Although, this is no longer true for bodies with centroid at
$0$, the next proposition shows that it is true in the average.
\begin{prop}
\label{fcentroid}
If $c(K)=o$, then
$$
\int_{K|L}\langle \nabla f_{K,L}(x),x\rangle\,d{\mathcal H}_d(x)\leq 0,
$$
with equality if and only if $f_{K,L}$ is constant on $K|L$.
\end{prop}
\begin{proof} Again, let $f=f_{K,L}$ and let $g:K|L\to L$ be a subgradient of $f$.
For $z\in{(\rm int}K)|L$, applying \eqref{eq:gradient_concave} to
$y=o$ and $x=z$ first, and next to $y=z$ and $x=o$, we deduce that
\begin{equation}
\label{hprimef}
\langle g(z),z\rangle\leq \ln f(z)-\ln f(0)\leq \langle g(o),z\rangle,
\end{equation}
where $g$ is a subgradient of $f$. In particular, if $\nabla f$ exists
at $z\in{(\rm int}K)|L$, then $\langle \nabla f(z),z\rangle\leq
\langle g(o),zf(z)\rangle$. Together with the property $c(K)=0$ we get
from \eqref{eq:centroid} that
\begin{equation}
\label{fcentroid0}
\int_{K|L}\langle \nabla f(z),z\rangle\,d{\mathcal H}_d(z)\leq
\int_{K|L}\langle g(o),zf(z)\rangle\,d{\mathcal H}_d(z)=0.
\end{equation}
Let us assume that equality holds in (\ref{fcentroid0}), and hence for
almost all $z\in (\inte K)|L$ in
(\ref{hprimef}). In particular, we have
$\ln f(x)-\ln f(0)=\langle g(o),x\rangle$, and in turn $f(x)=f(0)e^{\langle g(o),x\rangle}$
for almost all $x\in (\inte K)|L$. Since $f$ is continuous on $(\inte K)|L$,
Corollary~\ref{ray-limit} yields that
$f(x)=f(0)e^{\langle g(o),x\rangle}$ for all $x\in K|L$. However $f^{\frac1k}$ is concave, therefore
$g(o)=o$, or in other words, $f$ is constant.
\end{proof}
Now, we are ready to give the proof of Theorem~\ref{Henk-Linke-body}.
\begin{proof}[\bf Proof of Theorem~\ref{Henk-Linke-body}] Combining Propositions~\ref{VKf} and \ref{fcentroid} yields that
$$
V_K(L\cap S^{n-1})=\frac{d}n\,V(K)+\frac{1}n\,\int_{K|L}\langle \nabla f_{K,L}(x),x\rangle\,d{\mathcal H}_d(x)\leq
\frac{d}n\,V(K).
$$
Let us assume that equality holds, and hence $f_{K,L}(x)=f_{K,L}(o)$ for $x\in K|L$ according to Proposition~\ref{fcentroid}.
Let $C(x)=K\cap (x+L^\bot)$ for $x\in K|L$.
For any $x\in K|L$, there exists $\eta>0$ such that $-\eta x\in K|L$, and hence
$$
\mbox{$\frac{\eta}{1+\eta}\,C(x)+\frac{1}{1+\eta}\,C(-\eta x)\subset C(o).$}
$$
Therefore $f_{K,L}(x)=f_{K,L}(-\eta x)=f_{K,L}(o)$ and the equality
characterization of the
Brunn-Minkowski inequality \eqref{eq:brunn_minkowski} implies that $C(x)$ is a translate
of $C(o)$.
Choose linearly independent $v_1,\ldots,v_d\in K|L$ such that
$v_0=-v_1-\ldots-v_d\in K|L$, as well. By $\sum_{i=0}^dv_i=o$ we have
$\sum_{i=0}^d \frac1{d+1}\,C(v_i)\subset C(o)$, and we deduce that
$\sum_{i=0}^d \frac1{d+1}\,c(C(v_i))=c(C(o))$. In particular,
$$
c(C(o))\in\Pi=\aff\{c(C(v_0)),\ldots,c(C(v_d))\},
$$
where $\aff\{\}$ denotes the affine hull.
Next let $x\in K|L$. There exists $\eta>0$ such that $-\eta x\in[v_0,\ldots,v_d]$, and so
$\lambda x+\sum_{i=0}^d\lambda_i v_i=o$, where $\lambda+\sum_{i=0}^d\lambda_i=1$ and
$\lambda,\lambda_i\geq 0$ for $i=0,\ldots,d$.
It follows as above that $\lambda\,C(x)+\sum_{i=0}^d\lambda_iC(v_i)=C(o)$,
and hence $c(C(x))\in\Pi$, as well. Therefore, writing
$M=\Pi \cap (K+L^\bot)$ and $C=C(o)-c(C(o))$, we get $K=C+M$. In particular,
${\rm supp}\,V_K\subset L\cup \Pi^\bot$ and $L\cap\Pi^\perp=\{0\}$.
\end{proof}
\section{Another property of the cone-volume measure
if the centroid is the origin \label{another_property}}
Let us recall two basic notions about convex bodies.
Firstly, a convex body in $\R^n$ is called a cylinder
if it is of the form $[p,q]+C$ for $p,q\in\R^n$ and an
$(n-1)$-dimensional convex compact set $C$; $p+C$ and $q+C$ are
called bases of the cylinder.
Secondly, let $v\in S^{n-1}$, and let $M$ be a convex body in $\R^n$.
For any $t$ with $-h_M(-v)<t<h_M(v)$,
we replace the section $M\cap(tv+v^\bot)$ with the $(n-1)$-ball
of the same $(n-1)$-measure, centered at $tv$ in $tv+v^\bot$. Here,
$v^\bot$ is the abbreviation for the linear space orthogonal to $v$.
The closure $\widetilde{M}$ of the union
of these $(n-1)$-balls is called the Schwarz rounding of $M$
with respect to $\R v$. It is a convex body
by the Brunn-Minkowski theorem, and readily satisfies $\V(\widetilde{M})=\V(M)$.
If $\widetilde{M}$ is a cylinder, then all sections of the form $M\cap (tv+v^\bot)$
are of the same $(n-1)$-measure, and hence the equality
case of the Brunn-Minkowski theorem yields that $M$ is a cylinder, as
well. For more on Schwarz rounding we refer to \cite{Gruberbook}.
\begin{prop}
\label{hemi-sphere} Let $K\in\Kn$ with $c(K)=o$ and $V(K)=1$. Then
\begin{equation*}
\V_K(\Omega)\geq \frac{1}{2n},
\end{equation*}
for any open hemisphere $\Omega\subset S^{n-1}$. Equality holds if and
only if $K$
is a cylinder whose generating segment is orthogonal
to the linear $(n-1)$-space bounding the hemisphere $S$.
\end{prop}
\begin{proof} Let $\Omega\subset S^{n-1}$ be an open hemisphere, and
let $v\in S^{n-1}$ such that
$$
\Omega=\left\{u\in S^{n-1}:\,\langle u,v\rangle>0\right\}.
$$
For any convex body $M\in\Kn$ with $o\in\inte M$ and
$x\in M|v^\bot$, let
$$
f_M(x)=\max\{t\in\R\,:x+tv\in M\},
$$
and let $\varphi_M(x)=x+f_M(x)\,v$.
In particular the points of $\partial M$ where
all exterior normals have acute angle with $v$ are of the form $\varphi_M(x)$
for $x\in \inte M|v^\bot$. Therefore
$$
\V_M(\Omega)=\V(\Xi_M)
\mbox{ \ for }\Xi_M=\bigcup_{x\in M|v^\bot}[o,\varphi_M(x)].
$$
For $x\in(\inte M|v^\bot)\backslash\{o\}$, let
$z=\theta^{-1} x\in\partial M|v^\bot$
for some $\theta\in(0,1)$. Since $[\varphi_M(z),o,\varphi_M(o)]\subset \Xi_M$, we have
\begin{equation}
\begin{minipage}{0.9\hsize}
$x+\R v$ intersects $\Xi_M$
in a segment of length at least $(1-\theta)\|\varphi_M(o)\|$,
with equality if and only if $[\varphi_M(z), \varphi_M(o)]\subset\partial M$.
\end{minipage}
\label{eq:(i)}
\end{equation}
Now, let $\lambda=f_K(o)$, and hence $\lambda v\in\partial K$. After a
linear transformation
we may assume that
the tangent hyperplane $H$ at $\lambda v$ is given by $H=\lambda v+v^\bot$.
We shake $K$ down to $H$, i.e., for each $x\in K|v^\bot$,
we translate the section $(x+\R v)\cap K$ by
$(\lambda-f_K(x))v$
and hence one endpoint lands in $H$. We write $K'$ to
denote the resulting convex body, which satisfies
$$
K'|v^\bot=K|v^\bot=C-\lambda\,v
\mbox{ \ for $C=K'\cap H$}.
$$
In addition $\V(K')=\V(K)=1$, and $\Xi_{K'}$ is the cone $[o,C]$.
For $x\in(\inte K|v^\bot)\backslash\{o\}$, it follows
by \eqref{eq:(i)} that $x+\R v$ intersects $\Xi_K$
in a segment of length at least the length of
$\Xi_{K'}\cap(x+\R v)$. Therefore, Fubini's theorem yields
\begin{equation}
\label{K'K}
\V(\Xi_K)\geq \V(\Xi_{K'}).
\end{equation}
Furthermore, Fubini's theorem implies that
\begin{eqnarray*}
\langle c(K'),u\rangle&=&\langle c(K),u\rangle=0
\mbox{ \ for $u\in v^\bot$};\\
\langle c(K'),v\rangle&\geq&\langle c(K),v\rangle=0
\mbox{ \ with equality if and only if $K'=K$}.
\end{eqnarray*}
We deduce
\begin{equation}
\label{K'c}
c(K')=\eta v \mbox{ \ for $\eta\geq 0$, with $\eta=0$ if and only if $K'=K$}.
\end{equation}
Next let $\widetilde{K}$ be the
Schwarz rounding of $K'$ with respect to $\R v$.
It follows from the rotational symmetry
of $\widetilde{K}$ that $\langle c(\widetilde{K}),u\rangle=0$
for $u\in v^\bot$, and by Fubini's theorem that
$\langle c(\widetilde{K}),v\rangle=\langle c(K'),v\rangle$,
which in turn yield by (\ref{K'c}) and $\V(\widetilde{K})=\V(K')=1$ that
\begin{equation}
\label{tildeKc}
c(\widetilde{K})=c(K')=\eta v
\mbox{ \ for $\eta\geq 0$, with $\eta=0$ if and only if $K'=K$}.
\end{equation}
We conclude by (\ref{K'K}) and (\ref{tildeKc}) that
\begin{equation}
\label{tildeK}
\V(\Xi_K)\geq \V(\Xi_{\widetilde{K}-c(\widetilde{K})})
\mbox{ \ with equality if and only if $K'=K$}.
\end{equation}
Finally we compare $\widetilde{K}$ to the
cylinder $Z$
over the $(n-1)$-ball $H\cap \widetilde{K}$, where $\V(Z)=\V(\widetilde{K})=1$
and $Z$ and $K$ lie on the same side of $H$. We deduce from the
rotational symmetry of $Z$ that $\langle c(Z),u\rangle=0$
for $u\in v^\bot$. On the other hand, the rotational symmetry
of $\widetilde{K}$ and $\widetilde{K}|v^\bot=(H\cap \widetilde{K})-\lambda v$
yield that
$$
\langle x,v\rangle>-h_Z(-v)>\langle y,v\rangle
\mbox{ \ for all $x\in {\rm int}Z\backslash \widetilde{K}$
and $y\in \widetilde{K}\backslash Z$}.
$$
Therefore,
$$
c(Z)=\tau v
\mbox{ \ for $\tau\geq\eta$, with $\tau=\eta$ if and only if $Z=\widetilde{K}$}.
$$
We conclude by (\ref{tildeKc}) and (\ref{tildeK}) that
$$
\V(\Xi_K)\geq \V(\Xi_{Z-c(Z)})=1/(2n)
\mbox{ \ with equality iff $K'=K$ and $Z=\widetilde{K}$}.
$$
In turn, we get Proposition \ref{hemi-sphere}.
\end{proof}
\section{Some properties of the symmetric volume distance\label{prop_vol_distance}}
First we show that the distance $\delta_{\rm hom}$ can be estimated in terms of $\delta_{\rm vol}$.
These types of estimates have been around, only we were not able to locate them in the form we need.
\begin{lemma}
\label{centroidhom}
Let $K\in\Kn$ with $c(K)=o$.
\begin{enumerate}
\item[{(i)}] If $Q\subset K$ is a convex body with $\V(K\backslash Q)\leq t\,\V(K)$ for $t\in(0,\frac1e)$, then
$(1-(et)^{1/n})K\subset Q$.
\item[{(ii)}] If $Q$ is a convex body with $\V(K\Delta Q)\leq t\V(K)$ for $t\in(0,\frac1{4^ne})$, then
$(1-(et)^{1/n})K\subset Q\subset (1+4(et)^{1/n})K$.
\end{enumerate}
\end{lemma}
\begin{proof} The main tool is the following result due to B. Gr\"unbaum \cite{Gru60}. If $M\in\Kn$,
and $H^+$ is a half space containing $c(M)$, then
\begin{equation}
\label{Grun}
\V(M\cap H^+)\geq \V(M)/e.
\end{equation}
To prove (i), let $\lambda=0$ if $o\not\in \inte Q$, and let $\lambda>0$ be maximal
with the property that $\lambda K\subset Q$ otherwise. In addition, let
$x=o$ if $o\not\in \inte Q$, and let $x$ be a common boundary point of $Q$ and $\lambda K$ otherwise.
Therefore, there exists a half space $H^+_1$ such that
$x$ lies on its boundary, and $H^+_1\cap \inte Q=\emptyset$. Now there exists a $y\in K$ such that
$x=\lambda y$, and hence $x$ is the centroid of $x+(1-\lambda)K=\lambda y+(1-\lambda)K\subset K$.
It follows from (\ref{Grun}) that
$$
t\V(K)\geq \V(H^+_1\cap K)\geq \V(H^+_1\cap (x+(1-\lambda)K))\geq
\V((1-\lambda)K)/e,
$$
and thus $t\geq \frac{(1-\lambda)^n}e$.
To prove (ii), we observe that $\lambda K\subset Q$ for $\lambda=1-(et)^{1/n}$
by (i). We may assume that $Q\backslash K\neq \emptyset$, and
let $\mu>1$ be minimal with the property that $Q\subset \mu K$. For a common boundary point $z$
of $Q$ and $\mu K$, let $w\in K$ such that $z= \mu w$. In particular,
$w$ is the centroid of
$$
w+\frac{\lambda(\mu-1)}{\mu}\,K\subset \frac{1}{\mu}\,z+\frac{\mu-1}{\mu}\,Q\subset Q.
$$
In addition there exists a half space $H^+_2$ such that
$w$ lies on its boundary, and $H^+_2\cap \inte K=\emptyset$. We
deduce again from (\ref{Grun}) that
$$
t\V(K)\geq \V(H^+_2\cap Q)\geq V\left(H^+_2\cap\left (w+\frac{\lambda(\mu-1)}{\mu}\,K\right)\right)\geq
\frac{\lambda^n(\mu-1)^n}{\mu^ne}\,\V(K).
$$
Now $t<\frac1{4^ne}$ yields that $\lambda>\frac12$ and $2(e\,t)^{1/n}<\frac12$, which in turn implies that
$\mu\leq (1-2(et)^{1/n})^{-1}<1+4(et)^{1/n}$.
\end{proof}
\begin{coro}
\label{deltahomvol}
Let $K, Q\in\Kn$. Then
\begin{align*}
\delta_{\rm hom}(K,Q)&\leq 12\,\delta_{\rm vol}(K,Q)^{1/n}& \mbox{if $\delta_{\rm vol}(K,Q)<\frac1{4^ne}$,}\\
\delta_{\rm vol}(K,Q)&\leq 3n\,\delta_{\rm hom}(K,Q)&\mbox{if $\delta_{\rm hom}(K,Q)<\frac1{2n}$.}
\end{align*}
\end{coro}
\begin{proof} We use that $1+s<e^s<1+2s$ and $1-s<e^{-s}<1-\frac{s}2$ if $s\in(0,1)$.
We may assume that $c(K)=c(Q)=o$, and $\V(K)=\V(Q)=1$. In particular, $\V(K\Delta Q)=\delta_{\rm vol}(K,Q)$,
and hence the estimates for the exponential function and Lemma~\ref{centroidhom} yield with $s=\delta_{\rm vol}(K,Q)$ that
$$
e^{-2e^{1/n}s^{1/n}}K\subset(1-(se)^{\frac1n})K\subset Q\cap K \subset Q.
$$
Using the analogous formula $e^{-2e^{1/n}s^{1/n}} Q\subset K$,
we conclude the first estimate.
For the second estimate, let $t= \delta_{\rm hom}(K,Q)$. It follows that
$e^{-t}K\subset Q\subset e^tK$, thus $\V(K\Delta Q)\leq
e^{nt}-e^{-nt}<3nt$.
\end{proof}
Our next goal is Lemma~\ref{nocentroid} stating
that one does not need to insist on the common centroid in the definition of $\delta_{\rm vol}$. We prepare the argument
by the following observation.
\begin{lemma}
\label{translate-stab}
Let $K\in\Kn$ and $x\in\R^n$. Then
\begin{equation*}
\V(K\Delta (x+K))\leq
2n\|x\|_{K-K}\V(K).
\end{equation*}
\end{lemma}
\begin{proof} We may assume that $x\neq o$. Let $y,z\in K$ such that $x=\|x\|_{K-K}(y-z)$, and hence
$$
\|x\|_{K-K}=\|x\|/\|y-z\|.
$$
Applying Steiner symmetrization with respect to the hyperplane $x^\bot$ shows that
$$
\V(K)\geq \frac{\|y-z\|}n\,{\mathcal H}_{n-1}(K|x^\bot).
$$
We deduce by Fubini's theorem that
$$
\V(K\Delta (x+K))\leq 2\|x\|{\mathcal H}_{n-1}(K|x^\bot)\leq 2n\|x\|_{K-K}\V(K).
$$
\end{proof}
\begin{lemma}
\label{nocentroid}
Let $K,Q\in\Kn$ with $c(K)=o$ and $\V(K\Delta Q)\leq t\V(K)$
for $t\in(0,\frac1{4^ne})$. Then
\begin{equation*}
\|c(Q)\|_{K-K}\leq 4n t \quad\text{ and }\quad \delta_{\rm vol}(K,Q)\leq 9n^2 t.
\end{equation*}
\end{lemma}
\begin{proof} We may assume that $\V(K)=1$, and the minimal volume so
called L\"owner ellipsoid $E$ containing $K-K$ is a ball (see, e.g.,
\cite{Gruberbook}).
In particular, $n^{-1/2}E\subset K-K\subset E$, and the Brunn-Minkowski and Rogers-Shephard theorems yield that
$2^n\leq \V(K-K)\leq{2n \choose n}$. Since the volume of a centrally convex body over the volume of its Loewner ellipsoid
is at least $2^n/(n!\V(B^n))$ according to K. Ball \cite{Bal91}, we have
$$
2^n\leq \V(E)\leq {2n \choose n} \frac{n!}{2^n}\, \V(B^n)<\sqrt{3}\cdot\frac{2^nn^n}{e^n}\, \V(B^n).
$$
It follows that
\begin{equation}
\label{K-Knorm}
\mbox{$\frac{2}{\sqrt{e\pi}}\,B^n\subset K-K\subset n B^n$ \ and \
$\frac1n\,\|x\|\leq\|x\|_{K-K}\leq 2\|x\|$.}
\end{equation}
Therefore, to prove Lemma~\ref{nocentroid}, it is sufficient to verify the corresponding estimate for $\|c(Q)\|$.
If $c(Q)=o$, then we are done, otherwise let $u=c(Q)/\|c(Q)\|$. We have $Q\subset 2K\subset 2n B^n$ by Lemma~\ref{centroidhom} and (\ref{K-Knorm}), and $\V(Q)\geq 1-t$ implies
$\V(Q)^{-1}<2$. By \eqref{K-Knorm} we also have
\begin{equation*}
\|c(Q)\|_{K-K}\leq 2\|c(Q)\|=2\V(Q)^{-1}\langle u,c(Q)\rangle=2\V(Q)^{-1}\left\|\int_Q\langle u,x\rangle\,dx\right\|,
\end{equation*}
and since $c(K)=o$ we get
\begin{equation}
\begin{split}
\|c(Q)\|_{K-K}&\leq 2\V(Q)^{-1}\left\|\int_Q\langle
u,x\rangle\,dx\right\|\\
&= 2\V(Q)^{-1}\left\|\int_{Q\backslash K}\langle
u,x\rangle\,dx-\int_{K\backslash Q}\langle u,x\rangle\,dx\right\|\\
&4\int_{K\Delta Q} |\langle u,x\rangle|\,dx
\leq 4n t.
\end{split}
\label{c(Q)norm}
\end{equation}
Let $K'=K+c(Q)$, thus Lemma~\ref{translate-stab} and (\ref{c(Q)norm}) imply that
$\V(K\Delta K')\leq 8n^2t$. We observe that $Q'=c(Q)+\V(Q)^{-1/n}(Q-c(Q))$ satisfies $c(Q')=c(Q)$,
$\V(Q')=1$, and $\V(Q'\Delta Q)\leq t$ by $1-t\leq \V(Q)\leq 1+t$
(cf.~Lemma \ref{centroidhom}).
Therefore
$$
\delta_{\rm vol}(K,Q)=\V(K'\Delta Q')\leq \V(K'\Delta K) +\V(K\Delta Q)+\V(Q\Delta Q')<9n^2t.
$$
\end{proof}
\section{Some consequences of the stability of the Brunn-Minkowski inequality\label{sec_BM}}
Concerning the Brunn-Minkowski theory, including the properties of mixed volumes, the main reference is
R.~Schneider \cite{Sch93}. We use the Brunn-Minkowski theory in $L^\bot$ in the terminology of Theorem~\ref{Henk-Linke-stab},
whose dimension is $k=n-d$.
For $k,m\geq 1$, let
$$
I^k_m=\{(i_1,\ldots,i_m):\, i_j\in\N,\;j=1,\ldots,m\mbox{ and }i_1+\ldots+i_m=k\}.
$$
For compact convex sets $C_1,\ldots,C_m$ in $\R^k$ and $(i_1,\ldots,i_m)\in I^k_m$, the non-negative mixed volumes
$\V(C_1,i_1;\ldots;C_m,i_m)$ were defined by H. Minkowski in a way such that if $\alpha_1,\ldots,\alpha_m\geq 0$, then
\begin{equation}
\label{mixed-volume}
{\mathcal H}_k\left(\sum_{j=1}^m\alpha_j C_j\right)=\sum_{(i_1,\ldots,i_m)\in I^k_m}
\V(C_1,i_1;\ldots;C_m,i_m)\alpha_1^{i_1}\cdot\ldots \cdot \alpha_m^{i_m}.
\end{equation}
The mixed volume $\V(C_1,i_1;\ldots;C_m,i_m)$ actually depends only on the $C_j$ with $i_j>0$, does not depend on the
order how the pairs $C_j,i_j$ are indexed, and we frequently ignore the pairs $C_j,i_j$ with $i_j=0$.
We have $\V(C_1,k)={\mathcal H}_k(C_1)$, and $\V(C_1,i_1;\ldots;C_m,i_m)>0$ if each $C_j$ is $k$-dimensional. It follows by the Alexandrov-Fenchel inequality that
\begin{equation}
\label{Alexandrov-Fenchel}
\V(C_1,i_1;\ldots;C_m,i_m)^k\geq \prod_{j=1}^m{\mathcal H}_k(C_j)^{i_j}.
\end{equation}
An important special case of (\ref{Alexandrov-Fenchel}) is the classical Minkowski inequality, which says
\begin{equation}
\label{Minkowski-inequality}
\V(C_1,1;C_2,k-1)^k\geq {\mathcal H}_k(C_1){\mathcal H}_k(C_2)^{k-1}.
\end{equation}
Equality holds for $k$-dimensional $C_1$ and $C_2$ in the Minkowski inequality (\ref{Minkowski-inequality}) if and only if
$C_1$ and $C_2$ are homothetic. We remark that the equality conditions in the Alexandrov-Fenchel inequality
(\ref{Alexandrov-Fenchel}) are not yet clarified in general.
Now the Alexandrov-Fenchel inequality (\ref{Alexandrov-Fenchel}), and actually already the Minkowski inequality
(\ref{Minkowski-inequality}) yields the classsical (general) Brunn-Minkowski theorem stating that
if $C_1,\ldots,C_m$ are compact convex sets in $\R^k$, and
$\alpha_1,\ldots,\alpha_m\geq 0$, then (cf.~\eqref{eq:brunn_minkowski})
\begin{equation}
\label{BrunnMink}
{\mathcal H}_k\left(\sum_{j=1}^m\alpha_j C_j\right)^{1/k}\geq
\sum_{j=1}^m\alpha_i{\mathcal H}_k(C_i)^{1/k}.
\end{equation}
Equality holds for $k$-dimensional $C_1,\ldots,C_m$ and positive $\alpha_1,\ldots,\alpha_m$ in the Brunn-Minkowski inequality (\ref{BrunnMink}) if and only if
$C_1$ and $C_j$ are homothetic for $j=2,\ldots,m$.
We need the following stability version of the Minkowski inequality (\ref{Minkowski-inequality}) due to
A. Figalli, F. Maggi and A. Pratelli \cite{FMP10}.
If $C_1,C_2$ are $k$-dimensional compact convex sets in $\R^k$, and
\begin{equation}
\label{Minkowski-stab-cond}
\V(C_1,1;C_2,k-1)^k\leq (1+\varepsilon) {\mathcal H}_k(C_1){\mathcal H}_k(C_2)^{k-1}
\end{equation}
for small $\varepsilon\geq 0$, then \cite{FMP10} proves that
\begin{equation}
\label{Minkowski-vol-stab}
\delta_{\rm vol}(C_1,C_2)\leq\tilde{\gamma}_v\varepsilon^{1/2}
\end{equation}
where the explicit $\tilde{\gamma}_v>0$ depends only on the dimension $k$.
We remark that here we only work out the estimate with respect
to the symmetric volume distance $\delta_{\rm vol}$, and then just use
Corollary~\ref{deltahomvol} for $\delta_{\rm hom}$. Actually, V.I. Diskant \cite{Diskant} proved that (\ref{Minkowski-stab-cond}) implies
\begin{equation}
\label{Minkowski-hom-stab}
\delta_{\rm hom}(C_1,C_2)\leq\tilde{\gamma}_h\varepsilon^{1/k}
\end{equation}
for an unknown $\tilde{\gamma}_h>0$ depending only on $k$.
We note that (\ref{Minkowski-vol-stab}) and Corollary~\ref{deltahomvol} readily yields a version of (\ref{Minkowski-hom-stab})
with exponent $\frac1{2k}$ instead of $\frac1k$.\\
Combining the stability versions (\ref{Minkowski-vol-stab}) and (\ref{Minkowski-hom-stab}) with
Lemma~\ref{translate-stab} and Lemma~\ref{nocentroid} leads to the following stability version of the Brunn-Minkowski inequality.
\begin{lemma}
\label{BrunnMinkowski-stab}
For any $k\geq 1$, $m\geq 2$ and $\omega\in(0,1]$, there exist positive $\varepsilon_0(k,m,\omega)$ and $\gamma(k,m,\omega)$
depending on $k$, $m$ and $\omega$ such that
if $k$-dimensional compact convex sets $C_0,C_1,\ldots,C_m$ in $\R^k$, and $\alpha_1,\ldots,\alpha_m>0$
satisfy that $\alpha_i/\alpha_j\geq\omega$ and
${\mathcal H}_k(C_i)=V$ for $i,j=1,\ldots, m$, and
$$
\alpha_1C_1+\ldots+\alpha_mC_m\subset C_0
\mbox{ \ and \ }{\mathcal H}_k(C_0)\leq e^\varepsilon
(\alpha_1+\ldots+\alpha_m)^kV
$$
for $\varepsilon\in(0,\varepsilon_0(k,m,\omega))$, then for $i=1,\ldots, m$, we have
\begin{eqnarray*}
\delta_{\rm vol}(C_i,C_0)&\leq&\gamma(k,m,\omega)\varepsilon^{1/2},\\
\left\|c(C_0)-\sum_{i=1}^m\alpha_ic(C_i) \right\|_{C_0-C_0}&\leq &(\alpha_1+\ldots+\alpha_m)
\gamma(k,m,\omega)\varepsilon^{1/2}.
\end{eqnarray*}
\end{lemma}
\begin{proof} Since ${\mathcal H}_k(\alpha_1C_1+\ldots+\alpha_mC_m)\geq
(\alpha_1+\ldots+\alpha_m)^kV$ according to the Brunn-Minkowski inequality,
we may assume that
$\alpha_1C_1+\ldots+\alpha_mC_m=C_0$ by Lemma~\ref{nocentroid}.
For $1\leq i<j\leq m$, we apply the Alexandrov-Fenchel inequality (\ref{Alexandrov-Fenchel})
to each term in (\ref{mixed-volume}) except for $ k\alpha_i\alpha_j^{k-1}\V(C_i,1;C_j,k-1)$ and deduce that
$$
k\alpha_i\alpha_j^{k-1}\V(C_i,1;C_j,k-1)\leq k\alpha_i\alpha_j^{k-1}V+(e^\varepsilon-1)(\alpha_1+\ldots+\alpha_m)^kV.
$$
Here $(\alpha_1+\ldots+\alpha_m)^k\leq (\frac{m}{\omega})^k\alpha_i\alpha_j^{k-1}$, and hence
$$
\V(C_i,1;C_j,k-1)\leq\left(1+ \frac2k\left(\frac{m}{\omega}\right)^k \varepsilon\right) V.
$$
Thus (\ref{Minkowski-vol-stab}) yield
\begin{equation}
\label{vol-est}
\delta_{\rm vol}(C_i,C_j)\leq\bar{\gamma}(k,m,\omega)\varepsilon^{1/2}
\end{equation}
for $\bar{\gamma}(k,m,\omega)$
depending only on $k$, $m$ and $\omega$. To compare to $C_0$, we may assume that $V=1$,
$\alpha_1+\ldots+\alpha_m=1$ and
$c(C_i)=o$ for $i=1,\ldots,m$. Let $M=C_1\cap\ldots \cap C_m$.
It follows from (\ref{vol-est}) that
$$
{\mathcal H}_k(C_i\backslash M)\leq m\cdot \bar{\gamma}(k,m,\omega)\varepsilon^{1/2},\;i=1,\ldots,m,
$$
and hence ${\mathcal H}_k(M)\geq 1-m\cdot \bar{\gamma}(k,m,\omega)\varepsilon^{1/2}$.
Since $M\subset C_i$ for $i=1,\ldots,m$ yields $M\subset C_0=\sum_{i=1}^m\alpha_iC_i$, and ${\mathcal H}_k(C_0)\leq e^\varepsilon$, we deduce
$$
{\mathcal H}_k(C_0\Delta C_i)\leq 2\bar{\gamma}(k,m,\omega)\varepsilon^{1/2},\;i=1,\ldots,m.
$$
Therefore Lemma~\ref{translate-stab} and Lemma~\ref{nocentroid} imply
the required estimates for $\delta_{\rm vol}(C_i,C_0)$ and $c(C_0)$.
\end{proof}
To prove the next Proposition~\ref{fBrunnMinkowski-stab}, we need the following observation.
\begin{lemma}
\label{simplex-inK}
If $M$ is a convex body in $\R^d$ such that $-M\subset \eta M$ for $\eta\geq 1$, then there exists an $d$-simplex
$T\subset M$ whose centroid is the origin such that $M\subset \eta d^{3/2}T$.
\end{lemma}
\begin{proof} We may assume that the John ellipsoid $E$ of maximal
volume contained in $M\cap(-M)$ is Euclidean ball, and let
$T\subset M\cap (-M)$ be an inscribed regular simplex. Then
$\eta^{-1}M\subset M\cap(-M)\subset\sqrt{d} E\subset d^{3/2}T$.
\end{proof}
For Proposition~\ref{fBrunnMinkowski-stab} we use the notation of the
previous sections, i.e., $K\in\Kn$ is a convex body with $c(K)=o$,
$d,k\in\{1,\ldots,n-1\}$ with $d+k=n$,
and $L$ is a $d$-dimensional linear subspace.
For $x\in K|L$, we set
$$
f(x)=f_{K,L}(x)={\mathcal H}_k(K\cap(x+L^\bot)).
$$
\begin{prop}
\label{fBrunnMinkowski-stab}
There exist $t_0,\gamma>0$ depending on $n$ with the following properties.
Let $t\in(0,t_0)$, let $M_*\subset K|L$ be a $d$-dimensional convex compact set, and let
$K_*=K\cap(M_*+L^\bot)$. If $e^{-t}\leq f(x)/f(o)\leq e^t$ holds for any $x\in M_*$, then
there exist a
$k$-dimensional compact convex set $C\subset L^\bot$, and a complementary $d$-dimensional compact convex set $M$ such that
$$
\delta_{\rm vol}(K,C+M)\leq \gamma\max\left\{\frac{\V(K\backslash K_*)}{\V(K)},t^{1/2}\right\}.
$$
\end{prop}
\begin{proof}
Since $c(K)=o$ we have $-K\subset nK$. Hence $-K|L\subset
nK|L$ and we may choose, according to Lemma~\ref{simplex-inK},
$v_0,\ldots,v_d\in e^{-s}K|L$, for some $s>0$,
such that $v_0+\ldots+v_d=o$, and
\begin{equation}
\label{simplex-inKL}
e^{-s}K|L\subset n^{5/2}[v_0,\ldots,v_d].
\end{equation}
For $x\in e^{-s}K|L$, let $K(x)=K\cap(x+L^\bot)$, and let
\begin{equation}
\label{tildeKdef}
\widetilde{K}(x)=\frac{f(o)^{1/k}}{f(x)^{1/k}}\,K(x),\mbox{ \ and hence ${\mathcal H}_k(\widetilde{K}(x))=f(o)$.}
\end{equation}
We define
\begin{eqnarray*}
A&=&{\rm aff}\{c(K(v_0)),\ldots,c(K(v_d))\},\\
M&=&\{y\in A: (y+L^\bot)\cap e^{-s}K\neq \emptyset\},\\
C&=&K(o)-c(K(o)).
\end{eqnarray*}
We compare $K_*$ to $M+C$. To this end we consider the affine bijection
$\varphi:L\to A$ defined by the correspondance
$\{\varphi(x)\}=A\cap(x+L^\bot)$ for $x\in L$. In particular,
\begin{equation}
\label{phivio}
\varphi(v_i)=c(K(v_i)),\;i=0,\ldots,d\mbox{ \ and \ }\varphi(o)=\frac1{d+1}\sum_{i=0}^dc(K(v_i)).
\end{equation}
Let $x\in e^{-s}K|L$. We have $\frac{-1}{2n^{5/2}}x\in\frac12[v_0,\ldots,v_d]$ according to (\ref{simplex-inKL}), thus
$$
\frac{-1}{2n^{5/2}}\,x=\sum_{i=0}^d\alpha_iv_i\mbox{ \ where }
\sum_{i=0}^d\alpha_i=1\mbox{ and }\alpha_i\geq\frac1{2(d+1)},\;i=0,\ldots,d.
$$
We define
\begin{eqnarray*}
\tilde{\beta}&=&\frac{\beta f(x)^{1/k}}{f(o)^{1/k}} \mbox{ \ where \ }
\beta=\frac{1}{1+2n^{5/2}};\\
\tilde{\beta}_i&=&\frac{\beta_if(v_i)^{1/k}}{f(o)^{1/k}}
\mbox{ \ where \ }\beta_i= \frac{\alpha_i2n^{5/2}}{1+2n^{5/2}},\;i=0,\ldots,d,
\end{eqnarray*}
and hence $\beta+\sum_{i=0}^d\beta_i=1$ and
$\beta x+\sum_{i=0}^d\beta_iv_i=o$. The condition on the function $f$ yields that
$$
e^{-t/k}\leq \tilde{\beta}+\tilde{\beta}_0+\ldots+\tilde{\beta}_d\leq e^{t/k},
$$
and the ratio of any two of $\tilde{\beta},\tilde{\beta}_0,\ldots,\tilde{\beta}_d$ is at least $1/(4n^{5/2})$.
In particular,
$$
e^t(\tilde{\beta}+\tilde{\beta}_0+\ldots+\tilde{\beta}_d)^k f(o)\geq{\mathcal H}_k(K(o)),
$$
and the convexity of $K$ implies (cf.~\eqref{tildeKdef})
\begin{equation*}
\label{beta-x-vi}
\tilde{\beta}\widetilde{K}(x)+\sum_{i=0}^d\tilde{\beta}_i\widetilde{K}(v_i)=
\beta K(x)+
\sum_{i=0}^d\beta_iK(v_i)\subset K(o).
\end{equation*}
We deduce from Lemma~\ref{BrunnMinkowski-stab}, the stability version
of the Brunn-Minkowski inequality, that
there exists $\gamma^*>0$ depending on $n$ such that
for $i=0,\ldots,d$, we have
\begin{eqnarray}
\label{o-vi-vol}
\delta_{\rm vol}(K(v_i),K(o)),\;\delta_{\rm vol}(K(x),K(o))&\leq&\gamma^*t^{1/2},\\
\label{o-vi-c}
\left\|c(K(o))-\beta c(K(x))-\sum_{i=1}^d\beta_ic(K(v_i)) \right\|_{K(o)-K(o)}&\leq &
\gamma^*t^{1/2}.
\end{eqnarray}
Naturally, if $k=1$, then even $t$ can be written instead of $t^{1/2}$ on the right hand side of (\ref{o-vi-vol}) and
(\ref{o-vi-c}), but we ignore this possibility.
First we asssume that $x=o$. In this case (\ref{phivio}) and (\ref{o-vi-c}) yield
\begin{equation}
\label{c(K(o))}
\left\|c(K(o))-\varphi(o) \right\|_{K(o)-K(o)}\leq
\gamma^*t^{1/2}.
\end{equation}
Next let $x\in e^{-s}K|L$ be arbitrary. We have
$\beta \varphi(x)+\sum_{i=0}^d\beta_i\varphi(v_i)=\varphi(o)$ because $\varphi$ is affine.
We recall that $C=K(o)-c(K(o))$.
Let
$$
w=c(K(o))-\beta c(K(x))-\sum_{i=1}^d\beta_ic(K(v_i)).
$$
Since $\beta \varphi(x)=\varphi(o)-\sum_{i=0}^d\beta_i\varphi(v_i)$, we have
\begin{equation*}
\begin{split}
\big\|c(K(x)) & -\varphi(x) \big\|_{C-C} = \frac{\left\|\beta c(K(x))-\beta \varphi(x)) \right\|_{C-C}}{\beta}\\
&\leq \frac{\|\beta c(K(x))+w-\beta \varphi(x)) \|_{C-C}}{\beta}
+\frac{\|-w\|_{C-C}}{\beta}\\
&= \frac{\|c(K(o))-\varphi(o) -\sum_{i=0}^d\beta_i(c(K(v_i))-\varphi(v_i))\|_{C-C}}{\beta}\\
&\,+\frac{\|c(K(o))-\beta
c(K(x))-\sum_{i=1}^d\beta_ic(K(v_i))\|_{C-C}}{\beta}.
\end{split}
\end{equation*}
As $\varphi(v_i)=c(K(v_i))$ according to (\ref{phivio}), it follows by (\ref{o-vi-c}) and (\ref{c(K(o))}) that
\begin{equation}
\label{cK(x)}
\left\|c(K(x))-\varphi(x) \right\|_{C-C}\leq
\frac{2\gamma^*}{\beta}\cdot t^{1/2}<6n^{5/2}\gamma^* t^{1/2}.
\end{equation}
For $x\in e^{-s}K|L$, we deduce in order from (\ref{cK(x)}), (\ref{o-vi-vol}) and (\ref{tildeKdef}) that
\begin{equation*}
\begin{split}
{\mathcal H}_k\big((C+&\varphi(x))\Delta K(x) \big)\leq
{\mathcal H}_k\big((C+\varphi(x))\Delta (C+c(K(x))) \big)\\
&+{\mathcal H}_k\big( (C+c(K(x)))\Delta(\widetilde{K}(x)-c(\widetilde{K}(x))+c(K(x))) \big)\\
&+{\mathcal H}_k\big( (\widetilde{K}(x)-c(\widetilde{K}(x))+c(K(x))) \Delta K(x)\big)\\
&< 9n^{5/2}\gamma^* t^{1/2} {\mathcal H}_k(C).
\end{split}
\end{equation*}
Hence, by Fubini's theorem we get
$$
\V(K_*\Delta (M+C))<9n^{5/2}\gamma^* t^{1/2}\V(M+C)
$$
and Lemma~\ref{nocentroid} yields the required estimate for
$\delta_{\rm vol}$.
\end{proof}
\section{Some more properties of $f_{K,L}(x)$\label{prop_f}}
Here we establish some more properties of the log-concave function
(cf.~\eqref{eq:function})
\begin{equation*}
f_{K,L}: L \to \R_{\geq 0}\text{ with } x\mapsto \mathcal{H}_k(K\cap
(x+L^\perp)),
\end{equation*}
and use the notation as introduced in Section 2, i.e., $K\in\Kn$ is
an $n$-dimensional convex body with
$c(K)=0$, $L$ is a $d$-dimensional subspace $L$,
$1\leq d\leq n-1$, and we set $k=n-d$. Since we will keep $K$ and $L$
fixed, we just write $f(x)$ instead of $f_{K,L}(x)$. As in Section 2
let $g(x)$ be the subgradient of $f(x)$, and we recall that
$g(x)=\nabla f(x)/f(x)$ almost everywhere on $\inte(K)|L$.
For $\eta\geq 0$, we set
\begin{eqnarray*}
M_\eta&=&\{x\in K|L: \ln f(x)-\ln f(o)\geq \langle g(o),x\rangle-\eta\},\\
K_\eta&= & K\cap(M_\eta+L^\bot).
\end{eqnarray*
Since $\ln f$ is concave, both $M_\eta$ and $K_\eta$ are compact and convex.
\begin{lemma}
\label{nablafstab} Let $\eta\geq 0$. Then
$$
\int_{K|L}\langle \nabla f(x),x\rangle d{\mathcal H}_d(x)\leq -\eta \V(K\backslash K_\eta).
$$
\end{lemma}
\begin{proof} Let $x\in(\inte K)|L$ and $\eta\geq 0$, and let us assume
$\ln f(x)-\ln f(o)\leq \langle g(o),x\rangle-\eta$. Then by \eqref{eq:gradient_concave}
we have $\langle g(x),x\rangle\leq \langle g(o),x\rangle-\eta$.
Hence if $\nabla f$ exists at $x\in(\inte K)|L$, then
\begin{eqnarray*}
\langle \nabla f(x),x\rangle&\leq &0 \mbox{ \ provided that $x\in M_\eta$,}\\
\langle \nabla f(x),x\rangle&\leq &\langle g(o),f(x)x\rangle-f(x)\eta
\mbox{ \ provided that $x\not\in M_\eta$}.
\end{eqnarray*}
We conclude the lemma by \eqref{eq:centroid}
and $\V(K\backslash K_\eta)=\int_{(K|L)\backslash M_\eta}f(x)\,dx$.
\end{proof}
\begin{lemma}
\label{fxf0eta}
Let $\eta\in[0,1]$. If $\V(K\backslash K_\eta)\leq \V(K)/(2^ne)$, then
$$
e^{-\tau}\leq \frac{f(x)}{f(o)}\leq e^{\tau}
\mbox{ \ for $\tau=7n^{3/2}\eta^{1/2}$ and $x\in M_\eta$.}
$$
\end{lemma}
\begin{proof} By Lemma~\ref{centroidhom} we have $\frac12\,K\subset
K_\eta$, and $f(x)\geq f(o)e^{\langle g(o),x\rangle-\eta}$ for $x\in K_\eta$.
We claim that for $\pm y\in K_\eta$
\begin{equation}
\label{g(o)small}
|\langle g(o),y\rangle|\leq 3\sqrt{k\eta}.
\end{equation}
The concavity of $f^{1/k}$ yields that
\begin{eqnarray*}
f(o)^{1/k}&\geq& \frac{f(y)^{1/k}+f(-y)^{1/k}}2
\geq f(o)^{1/k}e^{-\eta/k}\frac{e^{\langle g(o),y\rangle/k}+e^{\langle g(o),-y\rangle/k}}2 \\
&\geq & f(o)^{1/k}e^{-\eta/k} \left(1+\left(\frac{\langle g(o),y\rangle}{2k}\right)^{2} \right).
\end{eqnarray*}
Since $e^t<1+2t$ for $t\in[0,1]$, we conclude (\ref{g(o)small}).
It follows from $\frac12\,K\subset K_\eta$ and $-K\subset nK$ that
$\frac12\,(K|L)\subset M_\eta$ and $-(K|L)\subset n(K|L)$.
In particular, if $x\in M_\eta$ is arbitrary, then
$\pm y\in M_\eta$ for $y=\frac1{2n}\,x$. We deduce from (\ref{g(o)small})
that $|\langle g(o),x\rangle|=2n|\langle g(o),y\rangle|\leq 6n\sqrt{k\eta}$.
Therefore, the lemma follows from
$f(o)e^{\langle g(o),x\rangle-\eta} \leq f(x)\leq f(o)e^{\langle g(o),x\rangle}$.
\end{proof}
\section{Proofs of Theorem~\ref{Henk-Linke-stab} and Theorem~\ref{Henk-Linke-stab1}\label{proof23}}
For the proofs of the two stability theorems \ref{Henk-Linke-stab} and
\ref{Henk-Linke-stab1}, let $K\in\Kn$ with $c(K)=o$, and let
$$
\V_K(L\cap S^{n-1})> \frac{d-\varepsilon}{n}\,\V(K)
$$
for a non-trivial linear subspace
$L$ with $\dim L=d$ and $\varepsilon\in(0,(2^ne)^{-5})$. As before,
for $x\in K|L$ let
$$
f(x)={\mathcal H}_k(K\cap(x+L^\bot)).
$$
According to Proposition~\ref{VKf}, the condition on $\V_K(L\cap S^{n-1})$ is equivalent with
\begin{equation}
\label{HenkLinke}
\int_{K|L}\langle \nabla f(x),x\rangle d{\mathcal H}_d(x)>-\varepsilon \V(K).
\end{equation}
\begin{proof}[Proof of Theorem~\ref{Henk-Linke-stab}] We set $\eta=\varepsilon^{4/5}$, and use the notation of Lemma~\ref{nablafstab}.
It follows from (\ref{HenkLinke}) and Lemma~\ref{nablafstab} that
\begin{equation*}
\label{Ketacomplement}
\V(K\backslash K_\eta)<\varepsilon^{1/5} \V(K)< \V(K)/(2^ne),
\end{equation*}
and from Lemma~\ref{fxf0eta} that
\begin{equation*}
\label{Metaf}
e^{-t}\leq \frac{f(x)}{f(o)}\leq e^{t}
\mbox{ \ for $t=7n^{3/2}\varepsilon^{2/5}$ and $x\in M_\eta$.}
\end{equation*}
We assume that $\varepsilon$ is small enough in order to
apply Proposition~\ref{fBrunnMinkowski-stab} with $M_*=M_\eta$ and $t=7n^{3/2}\varepsilon^{2/5}$.
We deduce the existence of an
$(n-d)$-dimensional compact convex set $C\subset L^\bot$, and complementary $d$-dimensional compact convex set $M$
such that
$$
\delta_{\rm vol}(K,C+M)\leq \gamma_v\varepsilon^{1/5}.
$$
In turn Corollary~\ref{deltahomvol} implies that
$$
\delta_{\rm hom}(K,C+M)\leq \gamma_h\varepsilon^{1/(5n)},
$$
completing the proof of Theorem~\ref{Henk-Linke-stab}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{Henk-Linke-stab1}] Here we have $d=1$.
We may assume that $L=\R$, and
$K|L=[-a,b]$ where $0<a\leq b$. Since $c(K)=o$ implies $-K\subset nK$ according to B. Gr\"unbaum \cite{Gru60},
we have $b\leq na$.
We set $\eta=\varepsilon^{2/3}$, and use again the notation of Lemma~\ref{nablafstab}. We deduce from (\ref{HenkLinke}) and Lemma~\ref{nablafstab} that
\begin{equation}
\label{Ketacomplement1}
\V(K\backslash K_\eta)<\varepsilon^{1/3} \V(K)< \V(K)/(2^ne),
\end{equation}
and from Lemma~\ref{fxf0eta} that
\begin{equation}
\label{Metaf1}
e^{-t}\leq \frac{f(x)}{f(o)}\leq e^{t}
\mbox{ \ for $t=7n^{3/2}\varepsilon^{1/3}$ and $x\in M_\eta$.}
\end{equation}
It follows from
Lemma~\ref{centroidhom} and (\ref{Ketacomplement1}) that
$\frac12[-a,b]\subset M_\eta$, therefore the concavity of $\ln f$ and (\ref{Metaf1}) yield that
\begin{equation}
\label{abfup}
f(x)\leq e^{2t}f(o) \mbox{ \ for $x\in[-a,b]$.}
\end{equation}
Let $M_\eta=[-a_\eta,b_\eta]$ for $a_\eta,b_\eta>0$.
Since $K\backslash K_\eta$ contains two cones, one with base $K(-a_\eta)$ and height $a-a_\eta$, and
one with base $K(b_\eta)$ and height $b-b_\eta$, we get by
(\ref{Metaf1}), (\ref{Ketacomplement1}) and (\ref{abfup}) that
\begin{eqnarray*}
\frac{a-a_\eta+b-b_\eta}n\,e^{-t}f(o)&\leq &\frac{a-a_\eta+b-b_\eta}n(f(-a_\eta)+f(b_\eta))\\
&\leq& \V(K\backslash K_\eta)<\varepsilon^{\frac13} \V(K)\leq\varepsilon^{\frac13}e^{2t}f(o)(a+b).
\end{eqnarray*}
In particular,
$$
{\mathcal H}_1(M_\eta)=a_\eta+b_\eta>(1-2n\varepsilon^{\frac13})(a+b).
$$
Here and below $\gamma_1,\gamma_2,\ldots$ denote positive constants depending on $n$.
We deduce by (\ref{Metaf1}) that if $\varepsilon$ is small enough, then
\begin{eqnarray*}
af(-a)+bf(b)&=&n\V_K(L\cap S^{n-1})>(1-\varepsilon)\V(K)>(1-\varepsilon){\mathcal H}_1(M_\eta)e^{-t}f(o)\\
&>& (1-\gamma_1\varepsilon^{\frac13})(a+b)f(o).
\end{eqnarray*}
Since $b\geq a$ and $\frac{a}{a+b}\geq\frac1{n+1}$ by $b\leq na$, \eqref{abfup} implies that if $\varepsilon$ is small enough, then
$$
f(-a),f(b)\geq (1-\gamma_2\varepsilon^{\frac13})f(o).
$$
As $\ln f$ is concave, we have
\begin{equation*}
\label{fablow}
f(x)\geq (1-\gamma_2\varepsilon^{\frac13})f(o) \mbox{ \ for $x\in[-a,b]$.}
\end{equation*}
However $\frac{a}{a+b}\,C(b)+\frac{b}{a+b}\,C(-a)\subset C(o)$, where
$C(x)=K\cap(x+L^\perp)$. Thus Lemma~\ref{BrunnMinkowski-stab} yields that
\begin{equation}
\label{KoKend}
\delta_{\rm vol}(C(o),C(-a))\leq \gamma_3\varepsilon^{\frac1{6}}\mbox{ \ and \ }
\delta_{\rm vol}(C(o),C(b))\leq \gamma_3\varepsilon^{\frac1{6}}.
\end{equation}
Hence, with
$$
\widetilde{C}=(C(-a)-\tilde{x})\cap (C(b)-\tilde{y}) \mbox{ \ for $\tilde{x}=c(C(-a))$ and $\tilde{y}=c(C(b))$.}
$$
It follows from (\ref{abfup}) and (\ref{KoKend}) that
$$
[\tilde{x},\tilde{y}]+\widetilde{C}\subset K\mbox{ \ and \ }\V(K)\leq
(1+\gamma_4\varepsilon^{\frac1{6}})\V([\tilde{x},\tilde{y}]+\widetilde{C}).
$$
Using Lemma~\ref{nocentroid}, we replace $\widetilde{C}$ by a suitably smaller homothetic copy $C$ such that $c(C)=o$, and obtain that
there exist $x\in \tilde{x}+\widetilde{C}$ and $y\in \tilde{y}+\widetilde{C}$ satisfying $o\in [x,y]$,
$e^{-s}\|x\|\leq \|y\|\leq e^{s}\|x\|$ for $s=\gamma_5\varepsilon^{\frac1{6}}$, and
$$
[x,y]+C\subset K\mbox{ \ and \ }\V(K)\leq
(1+\gamma_6\varepsilon^{\frac1{6}})\V([x,y]+C).
$$
Finally, if $z\in [-a,b]$, then $-z/n\in[-a,b]$ and
$\frac{1}{n+1}\,C(z)+\frac{n}{n+1}\,C(-z/n)\subset C(o)$. Therefore Lemma~\ref{centroidhom},
Lemma~\ref{BrunnMinkowski-stab} and the estimates above imply
$$
K\subset [x,y]+(1+\gamma_5\varepsilon^{\frac1{6n}})C,
$$
completing the proof of Theorem~\ref{Henk-Linke-stab1}.
\end{proof}
\section{Stability of the $\U$-functional $\U(K)$ \label{secU(K)}}
Let $m\in\{1,\ldots,n\}$. In this section, a finite sequence
$u_1,\ldots,u_m$ always denote points of $S^{n-1}$, and by $\lin\{X\}$
we denote the linear hull of a set $X$.
As in \cite{HeL14}, we define $\sigma_m(K)>0$ by
$$
\sigma_m(K)^m=\int_{u_1\wedge\ldots\wedge u_m\neq 0}1\,d\V_K(u_1)\cdots d\V_K(u_m).
$$
In particular, $\sigma_1(K)=\V(K)$, $\sigma_n(K)=U(K)$, and for $m<n$, we have
\begin{equation}
\begin{split}
\label{sigmainduction}
&\sigma_{m+1}(K)^{m+1}=\\
&\,\int_{u_1\wedge\ldots\wedge u_m\neq 0}
\left(\V(K)-\V_K(S^{n-1}\cap{\rm lin}\{u_1,\ldots,u_m\})\right)
d\V_K(u_1)\cdots d\V_K(u_m).
\end{split}
\end{equation}
As $\V_K(S^{n-1}\cap{\rm lin}\{u_1,\ldots,u_m\})\leq \frac{m}n\,\V(K)$
for linearly independent $u_1,\ldots,u_m$ according to Theorem~\ref{Henk-Linke-body}, we deduce that
\begin{equation}
\label{sigmainduction0}
\sigma_{m+1}(K)^{m+1}\geq \left(1-\frac{m}n\right)\V(K)\sigma_m(K)^m.
\end{equation}
Therefore the inequality of Theorem~\ref{Henk-Linke-UK} follows from
$$
\U(K)^n\geq \frac1n\,\V(K)\sigma_{n-1}(K)^{n-1}\geq \ldots\geq \frac{(n-1)!}{n^{n-1}}\,\V(K)^{n-1}\sigma_1= \frac{n!}{n^n}\,\V(K)^n.
$$
Now we assume that
$$
\U(K)\leq (1+\varepsilon)\frac{(n!)^{1/n}}{n}\,\V(K)
$$
where $\varepsilon>0$ is small enough to satify all estimates below. In particular,
$\varepsilon<\frac1{4n^3}\,\tilde{\varepsilon}_0$, where
$\tilde{\varepsilon}_0$ comes from Theorem~\ref{Henk-Linke-stab1}.
Applying \eqref{sigmainduction} for $m=1$, \eqref{sigmainduction0} for $m\geq 2$,
and using $(1+\varepsilon)^n\frac{n-1}{n}<\frac{n-1}{n}+2n\varepsilon$
gives
\begin{equation}
\label{sigmainduction1}
\int_{S^{n-1}}(\V(K)-\V_K(S^{n-1}\cap{\rm lin}\{u\}))\,d\V_K(u)\leq \left(\frac{n-1}{n}+2n\varepsilon\right)\V(K)^2.
\end{equation}
For any $X\subset S^{n-1}$, there exists $u\in X$ maximizing $\V_K(S^{n-1}\cap{\rm lin}\{u\})$ because different $1$-dimensional subspaces have disjoint intersections with $S^{n-1}$.
We consider linearly independent $v_1,\ldots,v_n\in S^{n-1}$ such that $v_1$ maximizes $\V_K(S^{n-1}\cap{\rm lin}\{u\})$
for $u\in S^{n-1}$, and $v_i$ maximizes $\V_K(S^{n-1}\cap{\rm lin}\{u\})$
for all $u\in S^{n-1}\backslash{\rm lin}\{v_1,\ldots,v_{i-1}\}$ if $i=2,\ldots,n$. Let $L={\rm lin}\{v_1,\ldots,v_{n-1}\}$,
and let $\V_K(S^{n-1}\cap{\rm lin}\{v_n\})=(\frac1n-t)\V(K)$, and hence
$t\in[0,\frac1n]$ (cf.~\eqref{scc}). Thus we have
\begin{eqnarray}
\label{VKvi}
\V_K(S^{n-1}\cap{\rm lin}\{v_i\})&\geq &\mbox{$(\frac1n-t)\V(K)$ \ for $i=1,\ldots,n$},\\
\label{VKvnu}
\V_K(S^{n-1}\cap{\rm lin}\{u\})&\leq &\mbox{$(\frac1n-t)\V(K)$ \ for $u\in S^{n-1}\backslash L$.}
\end{eqnarray}
We deduce from (\ref{sigmainduction1}), (\ref{VKvnu}) and
$\V_K(S^{n-1}\cap{\rm lin}\{u\})\leq \frac1n\,\V(K)$ for $u\in S^{n-1}\cap L$ that
$$
(\mbox{$\frac{n-1}n$}+t)\V(K)\V_K(S^{n-1}\backslash L)+
\mbox{$\frac{n-1}n$}\,\V(K)\V_K(S^{n-1}\cap L)\leq \left(\mbox{$\frac{n-1}{n}$}+2n\varepsilon\right)\V(K)^2.
$$
Since $\V_K(S^{n-1}\backslash L)\geq\frac1n\,\V(K)$ according to Theorem~\ref{Henk-Linke-body}, we conclude that
$t\leq 2n^2\varepsilon$. In particular, $\V_K(S^{n-1}\cap{\rm
lin}\{v_i\})\geq (\frac1n-2n^2\varepsilon)\V(K)$
for $i=1,\ldots,n$ by (\ref{VKvi}).
From Theorem~\ref{Henk-Linke-stab1} we find for $i=1,\ldots,n$,
that there exist an
$(n-1)$-dimensional compact convex set $C_i\subset v_i^\bot$ with $c(C_i)=o$, and $x_i,y_i\in \partial K$
such that $y_i=-e^{s_i}x$, where $|s_i|<n\tilde{\gamma}_v\varepsilon^{\frac1{6}}$, and for $i=1,\ldots,n$, we have
\begin{eqnarray}
\label{CixyinK}
[x_i,y_i]+C_i&\subset &K \\
\label{CixyKvol}
\V(K\backslash([x_i,y_i]+C_i))&\leq &n\tilde{\gamma}_v\varepsilon^{\frac1{6}}\V(K)\\
\label{CixyKhom}
K&\subset& [x_i,y_i]+(1+2\tilde{\gamma_h}\varepsilon^{\frac1{6n}})C_i.
\end{eqnarray}
We may assume that $v_i$ is an exterior normal at $x_i$, $i=1,\ldots,n$. After a linear transformation of $K$, we may also assume that $v_1,\ldots,v_n$ form and orthonormal system, and $\langle v_i,x_i-y_i\rangle=2$. In particular,
\begin{equation}
\label{xynorm}
e^{-\tau}<\langle v_i,x_i\rangle,\langle -v_i,y_i\rangle< e^{\tau},\;\tau=n\tilde{\gamma}_v\varepsilon^{\frac1{6}}.
\end{equation}
In what follows, we write $\gamma_1,\gamma_2,\ldots$ for positive constants depending on $n$ only.
It follows from combining (\ref{CixyinK}), (\ref{CixyKvol}) and (\ref{xynorm}) that
\begin{equation}
\label{CiCjarea}
1-\gamma_1\varepsilon^{\frac1{6}}<{\mathcal H}_{n-1}(C_i)/{\mathcal H}_{n-1}(C_j)<
1+\gamma_1\varepsilon^{\frac1{6}}\mbox{ \ for $i,j\in\{1,\ldots,n\}$}.
\end{equation}
For any $i\neq j\in\{1,\ldots,n\}$, we write
\begin{eqnarray*}
w_i(v_j)&=&h_{C_i}(v_j)+h_{C_i}(-v_j),\\
a_i(v_j)&=&\max\left\{ {\mathcal H}_{n-2}(C_i\cap (tv_j+v_j^\bot)):\,-h_{C_i}(-v_j)\leq t\leq h_{C_i}(v_j)\right\},
\end{eqnarray*}
and recall that $h_{C_i}(x)$ denotes the support function.
Hence $w_i(v_j)$ is the width of $C_i$ in the direction of $v_j$. Calculating
${\mathcal H}_{n-1}(C_i)$ by integrating along $\R v_j$ leads to
\begin{equation}
\label{Ciwjajarea}
\mbox{$\frac1{n-1}$}\,w_i(v_j)a_i(v_j)\leq {\mathcal H}_{n-1}(C_i)\leq w_i(v_j)a_i(v_j)
\mbox{ \ for $i\neq j\in\{1,\ldots,n\}$}.
\end{equation}
Let $p\neq q\in\{1,\ldots,n\}$. We choose $t_1\geq t_*\geq t_0$ such that
\begin{eqnarray*}
\langle v_p,t_1x_p\rangle &=& h_{C_q}(v_p)\\
\langle -v_p,t_0x_p\rangle &=& h_{C_q}(-v_p)\\
{\mathcal H}_{n-2}(C_q\cap (t_*x_p+v_p^\bot))&=&a_q(v_p).
\end{eqnarray*}
It follows from (\ref{xynorm}) and (\ref{CixyKhom}) that
\begin{eqnarray}
\label{t1t0diff}
t_1-t_0& > &w_q(v_p)/2,\\
\nonumber
C_q\cap (t_*x_p+v_p^\bot)&\subset & t_*x_p+(1+2\tilde{\gamma_h}\varepsilon^{\frac1{6n}})C_p.
\end{eqnarray}
Therefore $a_p(v_q)\geq (1+2\tilde{\gamma_h}\varepsilon^{\frac1{6n}})^{-(n-2)}a_q(v_p)$,
and hence interchanging the role of $p$ and $q$ leads to
$$
1-\gamma_2\varepsilon^{\frac1{6n}}<a_q(v_p)/a_q(v_p)<
1+\gamma_2\varepsilon^{\frac1{6n}}.
$$
We deduce from (\ref{CiCjarea}) and (\ref{Ciwjajarea}) that
\begin{equation}
\label{wpwq}
\frac1{2n}<\frac{w_p(v_q)}{w_q(v_p)}<2n.
\end{equation}
Now combining (\ref{CixyinK}) and (\ref{CixyKhom}) shows that
\begin{equation}
\label{hpvq}
h_{C_p}(v_q)\leq \langle x_q-t_mx_p,v_q\rangle \leq (1+2\tilde{\gamma_h}\varepsilon^{\frac1{6n}})h_{C_p}(v_q)
\mbox{ \ for $m=0,1$},
\end{equation}
and hence
$$
|\langle(t_1-t_0)x_p,v_q\rangle|\leq 2\tilde{\gamma_h}\varepsilon^{\frac1{6n}} h_{C_p}(v_q)
<2\tilde{\gamma_h}\varepsilon^{\frac1{6n}} w_p(v_q).
$$
Applying \eqref{t1t0diff}, \eqref{wpwq}, and the analoguous argument to $y_q$ implies that
\begin{equation}
\label{xpvq}
|\langle x_p,v_q\rangle|,|\langle y_p,v_q\rangle|\leq \gamma_3\varepsilon^{\frac1{6n}}.
\end{equation}
Let $P$ be the parallepiped
$$
P=\{x\in\R^n:\,\langle x,v_i\rangle\leq \langle x_i,v_i\rangle,\;\langle x,-v_i\rangle\leq \langle y_i,-v_i\rangle,\;
i=1,\ldots,n\},
$$
and hence each facet of $P$ contains one of $x_i+C_i$, $y_i+C_i$, $i=1,\ldots,n$.
We claim that
\begin{equation}
\label{Pclaim}
\mbox{$\frac1{4n}$}\,P\subset K.
\end{equation}
We suppose that (\ref{Pclaim}) does not hold, and seek a contradiction. Possibly reversing the orientation of some of the $v_i$, we may asssume that
\begin{equation}
\label{Pclaimcontra}
z=\frac1{4n}\sum_{i=1}^n\langle x_i,v_i\rangle\, v_i\not\in K.
\end{equation}
In particular, $\|z\|\leq \frac1{2\sqrt{n}}$ by (\ref{xynorm}), and there exists $u\in S^{n-1}$ such that
\begin{equation}
\label{Pclaimu}
\langle u,z\rangle>\langle u,x\rangle\mbox{ \ for $x\in K$.}
\end{equation}
There exists $v_p$ such that $|\langle u,v_p\rangle|\geq 1/\sqrt{n}$, and hence (\ref{xynorm}) and (\ref{xpvq}) yield that
$\langle u,x_p\rangle\geq\frac1{\sqrt{n}}-\gamma_4\varepsilon^{\frac1{6n}}$ if $\langle u,v_p\rangle\geq 1/\sqrt{n}$,
and $\langle u,y_p\rangle\geq\frac1{\sqrt{n}}-\gamma_4\varepsilon^{\frac1{6n}}$ if $\langle u,v_p\rangle\leq -1/\sqrt{n}$.
However $\langle u,z\rangle\leq \|z\|\leq\frac1{2\sqrt{n}}$, contradicting (\ref{Pclaimcontra}). Therefore we conclude
(\ref{Pclaim}).
For $i=1,\ldots,n$, let
$$
\Xi_{2i-1}=[o,x_i+C_i]\mbox{ \ and \ } \Xi_{2i}=[o,y_i+C_i].
$$
Since the basis of the cones $\Xi_1,\ldots,\Xi_{2n}$ lie in different
facets of $P$, the interiors of $\Xi_1,\ldots,\Xi_{2n}$ are pairwise
disjoint. By (\ref{CixyKvol}) and (\ref{xynorm}) we know
$\V(\Xi_j)\geq (\frac1{2n}-\gamma_5\varepsilon^{\frac1{6}})\V(K)$, and
so we get
$$
\V(\Xi)>(1-2n\gamma_5\varepsilon^{\frac1{6}})\V(K)\mbox{ \ for $\Xi=\bigcup_{j=1}^{2n}\Xi_j\subset K$}.
$$
We conclude from (\ref{Pclaim}) that
$$
\V(P\backslash K)\leq \V(P\backslash \Xi)=(4n)^n V\left((\mbox{$\frac1{4n}$}\,P)\backslash \Xi\right)
\leq (4n)^n V\left(K\backslash \Xi\right)\leq \gamma_6\varepsilon^{\frac1{6}}\V(K).
$$
Therefore (\ref{CixyKvol}) yields
$$
{\mathcal H}_{n-1}((P\cap v_i^\bot)\backslash C_i)\leq \gamma_7\varepsilon^{\frac1{6}}{\mathcal H}_{n-1}(C_i),\;
i=1,\ldots,n,
$$
and Lemma~\ref{centroidhom} implies that $(1-\gamma_8\varepsilon^{\frac1{6n}})P\subset K$,
completing the proof of Theorem~\ref{U(K)stab}.
|
\section{Introduction}
In his general theory of relativity, Albert Einstein predicted the
existence of gravitational waves (GWs) \cite{einstein:1916} ---
ripples in the fabric of spacetime caused by asymmetries in
catastrophic and highly accelerated events throughout the cosmos.
Though confirmed indirectly by observations of the binary pulsar PSR
$1913 + 16$ \cite{taylor:1989}, GWs have not been directly detected by
the global network of first generation detectors, such as Initial LIGO
in the United States \cite{abramovici:1992,ligo:2009}, Virgo in Italy
\cite{caron:1997,virgo:2012}, GEO 600 in Germany \cite{grote:2010},
and TAMA 300 in Japan \cite{ando:2005}.
The second generation of LIGO detectors, Advanced
LIGO~\cite{harry:2010}, are currently under construction and will
likely operate as early as 2015~\cite{AdvDet}. Advanced
Virgo~\cite{AdvVirgo} should come on-line in 2016, while the Japanese
KAGRA~\cite{KAGRA} detector will join the world-wide network later in
the decade. The ten-fold improvement in sensitivity of these
detectors \cite{harry:2010, AdvDet, smith:2009}, along with coherent
analysis between observatories, will significantly improve the chances
of detecting GWs from an astrophysical event in the Milky Way and
neighbouring galaxies. It is therefore likely that direct detection
of GWs will occur in the near future.
Potential sources of GWs include the inspiral of compact binary star
systems (of neutron stars or black holes) followed by black hole
formation \cite{thorne:1987}, pulsars \cite{culter:2002}, rotating
core collapse supernovae (CCSN) followed by protoneutron star
formation \cite{ott:2004}, gamma-ray bursts \cite{meszaros:2006}, and
cosmic string cusps \cite{damour:2005}.
Rotating CCSN are of particular interest in this paper. Like
neutrinos, GWs are emitted deep in the core of a progenitor and
propagate through the universe mostly unobscured by astrophysical
objects between the source and a detector on Earth. GWs act like
messengers, providing primary observables about the multi-dimensional
core collapse dynamics and emission mechanisms. It is in this way
that GW astronomy will open a new set of eyes to view the universe,
complementing the conventional electromagnetic-type observations.
Coalescing binary star systems are the most promising source of
detectable GWs \cite{thorne:1987}, with an expected observation rate
that could be as large as a few hundred events per year for Advanced
LIGO \cite{AdvDet,abbott:2005}. In contrast, the expected rate of
CCSN in the Milky Way is around three per century \cite{adams:2013}.
It is of great importance that appropriate data analysis techniques
are in place so we do not miss an opportunity to detect these rare
CCSN events.
The Bayesian statistical framework has proven to be a powerful tool
for parameter estimation in astrophysical and cosmological settings
\cite{loredo:1992}. Bayesian data analysis was first introduced to
the GW community by Christensen and Meyer \cite{christensen:1998}.
Christensen and Meyer \cite{christensen:2001} then demonstrated the
usefulness of the Gibbs sampler \cite{gelman:2013, geman:1984} for
estimating five physical parameters of coalescing binary signals.
Christensen, Meyer, and Libson \cite{christensen:2004a} then went on
to show how a custom-built Metropolis-Hastings algorithm
\cite{gelman:2013, metropolis:1953, hastings:1970}, a generalization
of the Gibbs sampler, was a superior and more suited routine for
eventual implementation into the LIGO Scientific Collaboration (LSC)
algorithm library (LAL). Parameter estimation for compact binary
inspirals has subsequently become more sophisticated in recent years
(see for example \cite{roever:2006, roever:2007a, roever:2007b,
raymon:2009, vdsluys:2008, veitch:2010, CBC-param}). Markov chain
Monte Carlo (MCMC) routines for inferring the physical parameters of
pulsars have also been developed \cite{christensen:2004b,
umstatter:2004, clark:2007}.
Due to the analytical intractability and complex multi-dimensional
nature of rotating core collapse stellar events, a significant amount
of computational time must go into numerically simulating the
gravitational waveforms. Unlike binary inspiral events, one cannot
simply use template search methods for supernova burst events as it is
computationally impossible to cover the entire signal parameter space.
It is therefore important to find alternative parameter estimation
techniques.
Summerscales \etal \cite{summerscales:2008} utilized the maximum
entropy framework to deconvolve noisy data from multiple (coherent)
detectors, with the goal of extracting a CCSN GW signal. Inference on
amplitude and phase parameters was conducted using cross correlation
between the recovered waveform and the set of competing waveforms from
the Ott \etal \cite{ott:2004} catalogue. A match was defined as the
model with the maximum cross correlation to the recovered waveform.
Heng \cite{heng:2009} first proposed a principal component analysis
(PCA) approach to simplify the problem by reducing a given supernova
waveform catalogue space down to a small number of basis vectors.
R\"{o}ver \etal \cite{roever:2009} extended this approach and created
a novel Metropolis-within-Gibbs sampler \cite{gelman:2013} to
reconstruct test signals from the Dimmelmeier \etal catalogue
\cite{dimmelmeier:2008} in noisy data using a principal component
regression (PCR) model with random effects and unknown signal arrival
time. They then attempted to exploit the structure of the posterior
principal component (PC) coefficients with a simple $\chi^2$ measure
of distance to determine which catalogue waveform best matched the
injected test signal. Although the Bayesian reconstruction method
showed much promise, extraction of the underlying physical parameters
had limited success.
Logue \etal \cite{logue:2012} used nested sampling
\cite{skilling:2006} to compute Bayesian evidence for PCR models under
three competing supernova mechanisms --- neutrino, magnetorotational,
and acoustic mechanisms. Each supernova mechanism has a noticeably
distinct gravitational waveform morphology, and the method was
successful at correctly inferring a large majority of injected
signals. They found that for signals embedded in simulated Advanced
LIGO noise, the magnetorotational mechanism could be distinguished to
a distance of up to 10 kpc, and the neutrino and acoustic mechanisms
up to 2 kpc.
Abdikamalov \etal \cite{abdikamalov:2013} generated a new
CCSN waveform catalogue and applied matched
filtering \cite{turin:1960} to infer total angular momentum to within $\pm
20\%$ for rapidly rotating cores. Slowly rotating cores had errors
up to $\pm 35\%$. Along with matched filtering, they employed
the Bayesian model selection method presented in \cite{logue:2012} to
illustrate that under certain assumptions of the rotation law, the
next generation of GW detectors (Advanced LIGO, Advanced Virgo, and
KAGRA), could also extract information about the degree of precollapse
differential rotation. The two methods worked particularly well for
rapidly rotating cores.
In this paper we present an alternative approach to parameter
estimation for rotating CCSN. Using the Abdikamalov \etal catalogue
\cite{abdikamalov:2013}, we fit a Bayesian PCR model to reconstruct a
GW signal in noisy data. Initially, the signal arrival time is
assumed to be known, and PC coefficients are sampled directly from the
posterior distribution. We extend the model to incorporate an unknown
signal arrival time and construct a Metropolis-within-Gibbs MCMC
sampler (as done in \cite{roever:2009}). We then use the posterior
means of the PC coefficients to fit the known physical parameters on
(using a linear model), and sample from the posterior predictive
distribution to make probabilistic statements about the ratio of
rotational kinetic energy to gravitational energy of the inner core at
bounce $\beta_{ic,b}$. We apply two supervised learning algorithms
--- na\"{i}ve Bayes classifier (NBC) and $k$-nearest neighbour
($k$-NN) --- to classify the closest level of precollapse differential
rotation $A$. We also introduce a constrained optimization approach
to model selection and attempt to find an optimal number of PCs for
the Bayesian PCR model.
The paper is organized as follows: in section 2 we describe the
simulated GW data catalogue used in our analysis;
section 3 introduces the statistical models and methods applied;
results of our analysis are presented in section 4; and a discussion
of our findings and future directions are provided in section 5.
\section{Gravitational wave data}
The waveforms used in this paper are the two-dimensional numerical
axisymmetric general-relativistic hydrodynamic rotating core collapse
and bounce supernova simulations generated by Abdikamalov \etal
\cite{abdikamalov:2013}. Based on findings that GW signals are
essentially independent of the progenitor zero age main sequence (ZAMS)
mass by Ott \etal \cite{ott:2012}, a single presupernova progenitor
model (the 12-$M_{\odot}$ at ZAMS solar-metallicity progenitor model
from \cite{woosley:2007}) was adopted. The cylindrical rotation law
from \cite{ott:2004} was also assumed.
The GW catalogue was partitioned into a base catalogue (BC), and a
test catalogue (TC). The BC contains $l = 92$ signals with five
levels of precollapse differential rotation $A$ (where higher values
of $A$ represent weaker differential rotation), a grid of values for
initial central angular velocity $\Omega_c$, and a grid of values for
the ratio of rotational kinetic energy to gravitational energy of the
inner core at bounce $\beta_{ic,b}$ (since $\beta_{ic,b}$ is a
function of $\Omega_c$ for a fixed progenitor structure). Each signal
in the BC was generated using the microphysical finite-temperature
Lattimer-Swesty (LS) equation of state (EOS) \cite{lattimer:1991},
parametrized deleptonization scheme from \cite{dimmelmeier:2008}, and
neutrino leakage scheme from \cite{ott:2012}. As well as varying $A$,
$\Omega_c$, and $\beta_{ic,b}$, the TC contains 47 signals with
differing EOS and deleptonization parametrizations $Y_e(\rho)$.
Specifically, some test signals were generated using the Shen \etal
EOS \cite{shen:1998}, or an increase/decrease in $Y_e(\rho)$
parametrization by $\sim 5\%$. The values of $\Omega_c$ and
$\beta_{ic,b}$ in the TC are in the same parameter space as those in
the BC, but with an alternative grid. The object of our analysis is to
predict the physical parameters ($\beta_{ic,b}$ and $A$) of the
signals in the TC using information gleaned about signals in the BC.
The signals were initially sampled at 100 kHz and subsequently
downsampled by a rational factor to 16384 Hz --- the sampling rate of
the Advanced LIGO detectors. Downsampling by a rational factor
essentially involved two steps: upsampling by an integer factor via
interpolation and then applying a low-pass filter to eliminate the
high frequency components necessary to avoid aliasing at lower
sampling rates; and downsampling by an integer factor to achieve the
desired sampling rate \cite{oppenheim:1999}. The resampled data was
zero-bufferred to ensure each signal was the same length, $N = 16384$,
which corresponded to 1 s of data at the Advanced LIGO sampling rate.
Each signal was then aligned so that the first negative peak (not
necessarily the global minimum), corresponding to the time of core
bounce, occurred halfway through the time series.
In this analysis, the source of a GW emission is assumed to be
optimally oriented (perpendicular) to a single interferometer. Each
signal is linearly polarized with zero cross-polarization.
\begin{center}
\includegraphics[width=1\linewidth]{gwPlot.pdf}
\captionof{figure}{A snapshot of the Abdikamalov \etal
\cite{abdikamalov:2013} catalogue. The top panel shows the GW
strain (scaled by source distance) for five models with different
levels of precollapse differential rotation (from strongest
differential rotation $A1$ to weakest $A5$), each with
$\beta_{ic,b} \sim 0.03$ (i.e., slowly rotating progenitors). The
bottom panel is the same, but for rapidly rotating progenitors
with $\beta_{ic,b} \sim 0.09$.}
\label{fig:gw}
\end{center}
We can see a general waveform morphology in figure~\ref{fig:gw}.
During core collapse, there is a slow increase in GW strain until the
first local maximum is reached (before 0.5 s). This is followed by
core bounce, where the strain rapidly decreases towards a local
minimum (at 0.5 s). This corresponds to the time when the inner core
expands at bounce. After this, there is a period of ring-down
oscillations. For slowly rotating progenitors, we see in the top
panel of figure~\ref{fig:gw} that the GW strain is essentially the
same during collapse and bounce and only differs during ring-down. For
the rapidly rotating progenitors presented in the bottom panel of
figure~\ref{fig:gw}, larger precollapse differential rotation results
in: a smaller local maximum during core collapse; a more negative
local minimum during core bounce; and a larger first ring-down peak.
Because of these patterns, Abdikamalov \etal \cite{abdikamalov:2013}
concluded that inferences about precollapse differential rotation
could in principal be made for rapidly rotating cores.
The data analyzed are CCSN GW signals injected in coloured Gaussian
noise using the Advanced LIGO noise curve with one-sided power
spectral density (PSD), $S_1(f)$. The data is then Tukey windowed to
mitigate spectral leakage. Rather than fixing source distance
to 10 kpc (as done in \cite{abdikamalov:2013}), this analysis assumes
a fixed signal-to-noise ratio (SNR) of $\rho = 20$. SNR is defined as
\begin{equation}
\label{eq:SNR}
\rho = \sqrt{4 \sum_j \frac{\frac{\Delta_t}{N}|\tilde{y_j}|^2}{S_1(f_j)}} ,
\end{equation}
where $\tilde{y_j}, j = 1, 2, \ldots, N$, are the Fourier transformed
data, $\Delta_t$ is the distance between two consecutive time points,
and $f_j, j = 1, 2, \ldots, N$, are the Fourier frequencies. As done
in \cite{roever:2009}, $S_1(f)$ is estimated \textit{a priori} by
averaging 1000 empirical periodograms from identically simulated
Advanced LIGO noise. This corresponds to a realistic scenario where
the noise spectrum must be estimated as well.
Although supernovae from the Milky Way will not produce SNRs as small
as $\rho = 20$, we choose this value to illustrate that our methods
are robust at lower SNRs.
\section{Methods and models}
\subsection{Bayesian inference}
Bayesian inference requires three pivotal quantities.
The \textit{likelihood} function $p(\mathbf{z} | \boldsymbol\theta)$
is the probability density function (PDF) of the data $\mathbf{z}$,
conditional on the random vector of model parameters
$\boldsymbol\theta$. The \textit{prior} $p(\boldsymbol\theta)$ is
the PDF of the model parameters, that takes into account all of the
information known about $\boldsymbol\theta$ before the data is
observed. The \textit{posterior} $p(\boldsymbol\theta |
\mathbf{z})$ is the updated PDF of model parameters after the data is
observed. These quantities are related via Bayes' theorem
\begin{eqnarray}
\label{eq:bayes}
p(\boldsymbol\theta | \mathbf{z})
& = & \frac{p(\boldsymbol\theta)p(\mathbf{z}|\boldsymbol\theta)}{m(\mathbf{z})} \\
& \propto & p(\boldsymbol\theta)p(\mathbf{z} | \boldsymbol\theta) ,
\end{eqnarray}
where $m(\mathbf{z}) = \int p(\boldsymbol\theta)p(\mathbf{z} |
\boldsymbol\theta) \mathrm{d}\boldsymbol\theta$ is the
\textit{marginal likelihood} and is treated as a normalizing constant
since it is independent of $\boldsymbol\theta$. That is, the
posterior is proportional to the prior multiplied by the likelihood.
Posterior sampling can be performed directly if the posterior PDF has
a closed analytical form. Otherwise, MCMC
techniques are a useful work-around. The key building blocks in MCMC
simulations are the Gibbs sampler \cite{geman:1984} and the
Metropolis-Hastings algorithm \cite{metropolis:1953, hastings:1970}.
We use a combination of the two --- the so-called
Metropolis-within-Gibbs sampler --- in this study. For a detailed
account of Bayesian inference and MCMC algorithms, refer to
\cite{gelman:2013}.
\subsection{Model 1: Bayesian PCR with known signal arrival time}
We aim to first reduce the dimension of the BC by a PCA, or
equivalently a singular value decomposition (SVD) as suggested by Heng
\cite{heng:2009}. Each BC waveform is represented as a linear
combination of orthonormal basis vectors, where the projection of the
data onto the first basis vector has maximum variance, the projection
onto the second basis vector has second highest variance, and so on.
By considering only projections on the first $d < l$ basis vectors,
the so-called $d$ PCs, a parsimonious representation of the catalogue
signals in $d$ dimensions is achieved that preserves as much of the
information of the original BC as possible.
Once PCA is conducted, the first $d$ PCs are treated as the
explanatory variables of a linear model. The data analyzed is a time
series vector $\mathbf{y}$ of length $N$ and decomposes into additive
signal and noise components. Let $\tilde{\mathbf{y}}$ be the Fourier
transformed data vector of length $N$ and let $\tilde{\mathbf{X}}$ be
the $N \times d$ design matrix, whose columns are the Fourier
transformed mean-centered PC vectors from the BC. The frequency
domain linear model is
\begin{equation}
\label{eq:freqLM}
\tilde{\mathbf{y}} = \tilde{\mathbf{X}} \boldsymbol\alpha + \tilde{\boldsymbol\epsilon},
\end{equation}
where $\boldsymbol\alpha$ is the vector of PCR
coefficients and $\tilde{\boldsymbol\epsilon}$ is the Fourier
transformed coloured zero-mean Gaussian noise vector whose variance
terms are proportional to the \textit{a priori known} one-sided
power spectral density $S_1(f)$. That is,
\begin{equation}
\label{eq:variances}
\sigma^2_{f_j} = \frac{N}{4 \Delta_t} S_1(f_j).
\end{equation}
Due to Hermitian symmetry, the frequency domain data vector
$\tilde{\mathbf{y}}$ contains only the non-redundant real and
imaginary components and is therefore the same length as the time
domain vector $\mathbf{y}$. Conversion between time and frequency
domains is conducted using a fast Fourier transform (FFT).
The likelihood for the Bayesian PCR model with known signal arrival
time is
\begin{equation}
\label{eq:likelihood1}
p(\tilde{\mathbf{y}} | \boldsymbol\alpha) \propto \exp \left( -2 \sum_{j = 1}^{N} \frac{\frac{\Delta_t}{N}\left( \tilde{y}_j - \left( \tilde{\mathbf{X}} \boldsymbol\alpha \right)_j \right)^2}{S_1 \left(f_j\right)} \right).
\end{equation}
Assuming flat ($\mathrm{Uniform}(-\infty, \infty)$) priors on
$\boldsymbol\alpha$, the posterior distribution for the PC
coefficients is
\begin{equation}
\label{eq:condpost1}
\mathrm{P}(\boldsymbol\alpha | \tilde{\mathbf{y}}) = \mathrm{N}(\boldsymbol\mu, \boldsymbol\Sigma),
\end{equation}
where
\begin{eqnarray}
\boldsymbol\Sigma &=& (\tilde{\mathbf{X}}^{'} \mathbf{D}^{-1} \tilde{\mathbf{X}})^{-1}, \\
\boldsymbol\mu &=& \boldsymbol\Sigma \tilde{\mathbf{X}}^{'} \mathbf{D}^{-1} \tilde{\mathbf{y}},
\end{eqnarray}
and $\mathbf{D} = \mathrm{diag}(\sigma^2_{f_j})$ is the diagonal
covariance matrix of individual variances for the noise component.
This multivariate normal distribution can be sampled directly with no
MCMC required.
Noninformative priors were chosen for this model. It was important to
keep the data and prior knowledge separate and distinct, and to avoid
using information from the waveform catalogue for both purposes. As
the only data available for analysis were the generated GWs, we
assumed complete prior ignorance on all model parameters.
\subsection{Model 2: Bayesian PCR with unknown signal arrival time}
The Bayesian PCR model presented in the previous section assumed a
known signal arrival time. The precise arrival time of a GW signal to
an interferometer will generally not be known in practice,
and must therefore be included as an additional unknown parameter in
the statistical model.
Let $T$ be a cyclical time shift representing
the unknown signal arrival time, and let $\tilde{\mathbf{X}}_T$ be the
Fourier transformed design matrix $\tilde{\mathbf{X}}$ shifted by lag
$T$, such that the Fourier transformed PCs are aligned with the
Fourier transformed data vector, $\tilde{\mathbf{y}}$. This
transformation can be done directly in the frequency domain as a phase
shift by multiplying the columns of $\tilde{\mathbf{X}}$ by $\exp(-2
\pi \mathrm{i} f T)$.
We build on the Bayesian signal reconstruction model presented in
\cite{roever:2009}, although our primary goal is inferring the
physical parameters of a supernova progenitor and not signal
reconstruction.
Using the same reasoning described in the previous section, we assume
flat priors on $\boldsymbol\alpha$ and $T$. The likelihood for the
Bayesian PCR model with unknown signal arrival time is
\begin{equation}
\label{eq:likelihood2}
p(\tilde{\mathbf{y}} | \boldsymbol\alpha, T) \propto \exp \left( -2 \sum_{j = 1}^{N} \frac{\frac{\Delta_t}{N}\left( \tilde{y}_j - \left( \tilde{\mathbf{X}}_T \boldsymbol\alpha \right)_j \right)^2}{S_1 \left(f_j\right)} \right).
\end{equation}
For a given time shift $T$, the conditional posterior
distribution for the PC coefficients $\boldsymbol\alpha | T$ is
\begin{equation}
\label{eq:condpost2}
\mathrm{P}(\boldsymbol\alpha | T, \tilde{\mathbf{y}}) = \mathrm{N}(\boldsymbol\mu_T, \boldsymbol\Sigma_T),
\end{equation}
where
\begin{eqnarray}
\boldsymbol\Sigma_T &=& (\tilde{\mathbf{X}}^{'}_T \mathbf{D}^{-1} \tilde{\mathbf{X}}_T)^{-1}, \\
\boldsymbol\mu_T &=& \boldsymbol\Sigma_T \tilde{\mathbf{X}}^{'}_T \mathbf{D}^{-1} \tilde{\mathbf{y}} .
\end{eqnarray}
To estimate $\boldsymbol\alpha$ and $T$, we construct a Markov chain
whose stationary distribution is the posterior distribution of
interest using Metropolis-within-Gibbs sampler \cite{gelman:2013}.
This is essentially a Gibbs sampler that alternates between the full
set of conditional posterior distributions $P(\boldsymbol\alpha | T,
\tilde{\mathbf{y}})$ and $P(T | \boldsymbol\alpha,
\tilde{\mathbf{y}})$. The former can be sampled directly using
equation~(\ref{eq:condpost2}), and the latter requires a random walk
Metropolis step, hence the name Metropolis-within-Gibbs.
After initialization, step $i + 1$ in the Metropolis-within-Gibbs algorithm is:
\begin{enumerate}
\item Directly sample the conditional posterior of $\alpha_{i+1} |
T_i$ using equation~(\ref{eq:condpost2});
\item Propose $T_{*}$ from $t_{\nu}(T_i, \zeta^2)$ and accept $T_{i+1}
= T_{*}$ with the Metropolis acceptance probability
\begin{equation}
\label{eq:metropolis}
r = \min\left(1, \frac{p(T_{*} | \boldsymbol\alpha, \tilde{\mathbf{y}})}{p(T_i | \boldsymbol\alpha, \tilde{\mathbf{y}})}\right).
\end{equation}
Otherwise reject and set $T_{i+1} = T_i$.
\end{enumerate}
A $t$-distribution was chosen as the proposal distribution for the
algorithm. It has a similar (symmetrical) shape to the normal
distribution but has heavier tails and an additional
degrees-of-freedom parameter, $\nu$. The heavier tails of the
$t$-distribution results in bolder and more robust proposals than the
normal distribution, ensuring the algorithm does not get stuck in
local modes \cite{gelman:2013}. The degrees-of-freedom parameter was
set to $\nu = 3$, which is the smallest integer that yields a
distribution with finite variance. The proposal for $T_{i+1}$ is
centered on $T_i$, and has scale parameter $\zeta^2$ that is initially
and arbitrarily set to 0.05, and subsequently automatically tuned
during the algorithm to ensure good mixing and acceptance rates.
\subsection{Posterior predictive distribution}
For each of the $l = 92$ signals in the BC and $m = 47$ signals in the
TC, we fit both Bayesian PCR models, with $d$ PCs (where the choice of
$d$ is explained below). We then construct an $l \times (d + 1)$
design matrix $\mathbf{A}$ whose rows are the \textit{posterior means}
of the $d$ PC coefficients, plus an intercept term, for each of the
$l$ signals in the BC. The primary goal is to exploit the posterior
PC coefficient space to make inferences on the physical parameters of
rotating core collapse stellar events in the TC. We accomplish this
by fitting a linear model with the known physical parameters from the
BC as the response variable on the design matrix $\mathbf{A}$ using
\begin{equation}
\label{eq:LM2}
\boldsymbol\xi = \mathbf{A} \boldsymbol\gamma + \boldsymbol\delta,
\end{equation}
where $\boldsymbol\xi$ is the vector of known continuous physical
parameters, $\boldsymbol\gamma$ is the vector of regression
coefficients, and $\boldsymbol\delta$ is an error term. The error
term is assumed to come from an independent and identically
distributed normal distribution with zero mean and variance
$\sigma^2$. Predictions using the \textit{posterior predictive
distribution} are the primary interest in this analysis, and not the
model parameters themselves.
Assuming the convenient noninformative prior distribution that is
uniform on $(\boldsymbol\gamma, \log \sigma)$, the posterior
predictive distribution for a normal linear model is a multivariate
$t$-distribution and can be sampled from directly with no MCMC
\cite{gelman:2013}. The formula is
\begin{equation}
\mathrm{P}(\check{\boldsymbol\xi} | \boldsymbol\xi) = t_{l - d}\left(\check{\mathbf{A}}\hat{\boldsymbol\gamma}, s^2\left(I + \check{\mathbf{A}} \mathbf{V}_{\boldsymbol\xi} \check{\mathbf{A}}^{'}\right)\right) ,
\end{equation}
where $\check{\boldsymbol\xi}$ is the vector of outcomes we wish to
predict (i.e., the physical parameters from signals in the TC),
$\check{\mathbf{A}}$ is the $m \times (d + 1)$ matrix whose rows are
the posterior means of the signals in the TC (and an intercept term)
from the Bayesian PCR step, $I$ is the $m \times m$ identity matrix,
and
\begin{eqnarray}
\mathbf{V}_{\boldsymbol\xi} &=& (\mathbf{A}^{'}\mathbf{A})^{-1}, \\
\hat{\boldsymbol\gamma} &=& \mathbf{V}_{\boldsymbol\xi} \mathbf{A}^{'}\boldsymbol\xi, \\
s^2 &=& \frac{1}{l - d} (\boldsymbol\xi - \mathbf{A}\hat{\boldsymbol\gamma})^{'}(\boldsymbol\xi - \mathbf{A}\hat{\boldsymbol\gamma}).
\end{eqnarray}
\subsection{Deviance information criterion and constrained optimization}
The \textit{deviance} is defined as $D = -2 \log p(\mathbf{z} |
\boldsymbol\theta)$ where $p(\mathbf{z} | \boldsymbol\theta)$ is the
likelihood of a statistical model, and $\boldsymbol\theta$ is the
vector of model parameters. The \textit{deviance information
criterion} (DIC) is a Bayesian model comparison technique and a
generalization of Akaike information criterion (AIC) for hierarchical
models \cite{spiegelhalter:2002}. DIC is defined as
\begin{eqnarray}
\mathrm{DIC} & = & \bar{D} + p_D \label{eqn:DIC1} \\
& = & 2\bar{D} - D(\bar{\boldsymbol\theta}) , \label{eqn:DIC2}
\end{eqnarray}
where $\bar{D}$ is the mean deviance from posterior samples, $p_D$ is
the effective number of parameters, and $D(\bar{\boldsymbol\theta})$
is the deviance evaluated at the posterior means of the
parameters. When comparing competing statistical models, the lowest
DIC is preferred. $\bar{D}$ is a measure of fit, and $p_D$ is a
measure of model complexity used to penalize models with too many
parameters. Equation~(\ref{eqn:DIC1}) therefore illustrates how DIC
incorporates Occam's razor, allowing one to select a parsimonious
model, balancing between fit and complexity. Equation~(\ref{eqn:DIC2}),
on the other hand, provides a simple method for computing DIC.
$\bar{D}$ is calculated by evaluating the deviance for each of the
stored model parameters $\boldsymbol\theta$ that have been sampled
from their joint posterior PDF, and then taking the average.
$D(\bar{\boldsymbol\theta})$ is calculated by finding the posterior
mean of each of the model parameters $\bar{\boldsymbol\theta}$ and
then evaluating the deviance.
DIC is the preferred model comparison technique in this analysis. A
popular alternative, Bayes factors, would require computing the
marginal likelihood from equation~(\ref{eq:bayes}), which involves
multi-dimensional integration over a large number of parameters.
Numerical techniques such as nested sampling \cite{skilling:2006} can
be used to derive the marginal likelihood but these methods require
significant computational time and power. On the other hand, DIC is
easily computed from posterior samples. Another benefit of using DIC
over Bayes factors is that improper priors (which we have assumed in
this analysis) do not violate any conditions of use. Bayes factors,
on the other hand, are no longer applicable when improper priors are
used.
The choice of the number of PCs has been arbitrary in most of the
supernova GW parameter estimation literature and this number has
usually been $d = 10$ (see for example \cite{roever:2009,
abdikamalov:2013}). We propose a method for selecting the optimal
choice of $d$ based on careful analysis of the DIC for competing
models and constrained optimization. Since PCs are ordered by the
total amount of variation they make up in the data set, PCA provides a
convenient ordering system for nested modelling. Let $M_d, d \in \{1,
2, \ldots, 92\}$, represent the set of possible PCR models,
where $d$ is the number of explanatory variables. The models are
nested such that $M_1$ has one explanatory variable (PC1), $M_2$ has
two explanatory variables (PC1 and PC2), and so on.
For each of the $l = 92$ signals in the BC (injected in Advanced LIGO
noise), all of the models $M_d, d \in \{1, 2, \ldots, 92\}$, are
fitted and then compared using DIC. The model with the lowest DIC is
the best fit to the data. However, models with an absolute
difference in DIC of $\lesssim 5$ are generally taken to be
indistinguishable from one another \cite{spiegelhalter:2002} and so to
prevent over-fitting, we propose a constrained optimization routine,
where we select the smallest $d$ such that the difference in DIC
between $M_d$ and the model with the minimum DIC is less than 5. More
specifically, let $M_{\min}$ be the model with the minimum DIC, then
find $d$ such that
\begin{equation}
\label{eq:optimization}
\argmin_d \bigg\{ \mathrm{DIC}(M_d) - \mathrm{DIC}(M_{\min}) < 5 \bigg\} .
\end{equation}
We employ this routine for each of the $l = 92$ BC signals, and look
at the distribution of $M_d$'s over all signals. The median of this
distribution seems a prudent choice for a general-purpose number of
PCs since these distributions tend to be skewed. It is
important to note here that we cannot choose a different value for $d$
for each signal when implementing these models as this would lead to a
very sparse design matrix $\mathbf{A}$ when sampling from the
posterior predictive distribution.
We refer the reader to figure~\ref{fig:occam} in the results section
of this paper for an example of this method in action.
\subsection{Na\"{i}ve Bayes classifier}
The NBC \cite{ripley:1996} is a common supervised learning
algorithm and discriminant method used to group objects into a
discrete set of classes based on a set of features. The algorithm
requires a \textit{training set} of objects with known groupings and
observed features. Once the algorithm has learnt from the training
set, each object in a \textit{test set} (containing a set of observed
features and potentially unknown classes) is assigned to the group
that it has the highest probability of belonging to.
The ``Bayes'' component of the method refers to Bayes' theorem
\begin{equation}
\label{eq:naive}
p(c|\mathbf{u}) \propto p(c) p(\mathbf{u} | c)
\end{equation}
where $c \in C$ is the class that an object could belong to, and
$\mathbf{u}$ are the features we wish to exploit to classify the
object. That is, given some observed features $\mathbf{u}$, what is
the posterior probability of an object belonging to class $c$?
The ``na\"{i}ve'' component refers to the assumption of conditional
independence of the model features $\mathbf{u} = (u_1, u_2, \ldots,
u_d)$. This assumption implies the joint PDF $p(\mathbf{u} | c)$ can
be factorized as the product of marginal distributions
\begin{equation}
\label{eq:factorize}
p(\mathbf{u} | c) = \prod_{i = 1}^d p(u_i | c),
\end{equation}
and so equation~(\ref{eq:naive}) becomes
\begin{equation}
\label{eq:naive2}
p(c | \mathbf{u}) = p(c) \prod_{i = 1}^d p(u_i | c).
\end{equation}
Given class $c$, each feature $(u_1, u_2, \ldots, u_d)$ is assumed to
be independently normally distributed. The model parameters are
approximated using the relative frequencies from the training set.
The class prior probabilities $p(c)$ are specified as the number of
objects in class $c$ in the training set divided by the total number
of objects. Objects are grouped into the class that yields the
highest posterior probability. This is known as the maximum \textit{a
posteriori} (MAP) decision rule.
\subsection{$k$-nearest neighbour}
An alternative machine learning algorithm to the NBC is the $k$-NN
\cite{ripley:1996}, which uses a measure of ``closeness'' between
objects rather than a probabilistic framework. We choose $k = 1$,
meaning that an object in the test set is assigned to the class of its
single nearest neighbour in the training set. Ties in distance are
settled at random.
The definition of closeness in this context depends on the choice
of metric. As commonly used in the literature \cite{ripley:1996}, a
Euclidean distance is assumed. For any object with features
$\mathbf{u} = (u_1, u_2, \ldots, u_d)$ in the test set, the $k$-NN
algorithm finds the object with features $\mathbf{v} = (v_1, v_2,
\ldots, v_d)$ in the training set that minimizes the Euclidean
distance
\begin{equation}
\label{eq:euclidean}
\mathrm{distance}(\mathbf{u}, \mathbf{v}) = \sqrt{\sum_{i = 1}^d (u_i - v_i)^2},
\end{equation}
and then assigns $\mathbf{u}$ to the class of $\mathbf{v}$.
\section{Results}
\subsection{Model selection}
An important statistical task is to select a prudent number of model
dimensions whilst incorporating Occam's razor into the decision making
process. More specifically, one needs to balance model fit against
complexity to ensure there is no over-fitting. In the context of PCA,
the decision is usually made based on the amount of variation the
first $d$ PCs contribute to the data set (i.e., analyzing Scree
plots). This approach is arbitrary and deals specifically with
dimension reduction, but not Occam's razor. We propose an alternative
approach, involving DIC and constrained optimization.
We analyze the change in DIC as model dimensionality
increases. Figure~\ref{fig:occam} illustrates DIC as a function of
model dimensionality for signal $A1O2.5$ from the Abdikamalov \etal
catalogue \cite{abdikamalov:2013}. This is the typical shape of the
DIC curve for all signals in the BC and a good visual aid of Occam's
razor in action. There tends to be a sharp decrease in DIC as the
model dimension increases at the beginning, where model fit is
improving. DIC flattens out and then reaches a minimum, where there
is the best balance between fit against complexity. After this, there
is a slow rise in DIC as the model dimension increases and becomes too
complex.
\begin{center}
\includegraphics[width=0.9\linewidth]{typicalDICplot.pdf}
\captionof{figure}{DIC as a function of model dimensionality for
model $A1O2.5$ from the Abdikamalov \etal catalogue
\cite{abdikamalov:2013}. The dashed vertical line to the right
represents the model with the minimum DIC ($M_{\min} = M_{22}$).
The dotted vertical line to the left represents the model
dimension after constrained optimization ($M_d = M_{13}$).}
\label{fig:occam}
\end{center}
The flat basin around the global minimum in figure~\ref{fig:occam} is
of particular interest. Since models with an absolute difference in
DIC of less than 5 are essentially indistinguishable, it is sensible
to select the model with the smallest number of dimensions in this
region to prevent over-fitting. For signal $A1O2.5$, we see a
significant decrease in model dimensionality from $M_{\min} = M_{22}$
to $M_d = M_{13}$. The choice of $d$ for this particular signal is $d
= 13$.
It is important to note that $d$ will differ between GW signals but we
must only choose one general-purpose value of this. We therefore
conduct the proposed constrained optimization model selection method
on all of the $l = 92$ BC signals and take the median of the
distribution of $d$'s as the general-purpose $d$. We prefer the
median to the mean as our central measure as it is more robust against
outliers.
\begin{center}
\includegraphics[width=1\linewidth]{dicOptimHistBin1.pdf}
\captionof{figure}{Distribution of model dimensionality for all $l =
92$ signals in the BC under our constrained optimization routine.}
\label{fig:dicHist}
\end{center}
The histogram in figure~\ref{fig:dicHist} shows the distribution of
$d$ for all $l = 92$ signals in the BC. It is highly skewed to the
right, with a median (and mode) of 14 PCs and mean of 17 PCs. We
choose $d = 14$ based on the median of this distribution, and use this
number of explanatory variables in both Bayesian PCR models. We
choose this as the model that minimizes the risks of both
over-fitting and under-fitting.
\subsection{Inferring the ratio of rotational kinetic energy to
gravitational energy of the inner core at bounce, $\beta_{ic,b}$}
We injected each of the $l = 92$ BC and $m = 47$ TC signals in
Advanced LIGO noise (SNR $\rho = 20$) and fitted the two Bayesian PCR
models with $d = 14$ PCs. We then regressed the known value of
$\beta_{ic,b}$ on the posterior means of the BC signals from these
models and sampled from the posterior predictive distribution of the
TC signals.
Figures~\ref{fig:betaKnownT}--\ref{fig:betaUnknownT_Other} show these
predictions of $\beta_{ic,b}$. The true value from the TC (red
triangle) is compared with the predicted value (blue circle) and
uncertainty is measured using 90\% credible intervals (black lines).
Figures~\ref{fig:betaKnownT}~and~\ref{fig:betaKnownT_Other} assume a
\textit{known} signal arrival time. $T$ is \textit{unknown} for
figures~\ref{fig:betaUnknownT}~and~\ref{fig:betaUnknownT_Other}. The
change in background gradient for
figures~\ref{fig:betaKnownT}~and~\ref{fig:betaUnknownT} represents the
varying precollapse differential rotation model $A$ for signals with
LS EOS and standard $Y_e(\rho)$ parametrization. For
figures~\ref{fig:betaKnownT_Other}~and~\ref{fig:betaUnknownT_Other},
the background shade represents GW signals (from a precollapse
differential rotation model $A1$) with a Shen EOS, or
increase/decrease in $Y_e(\rho)$ of $\sim 5\%$. $\beta_{ic,b}$ is
scaled up by a factor of 100 in these plots.
\begin{center}
\includegraphics[width=1\linewidth]{betaKnownT.pdf}
\captionof{figure}{90\% credible intervals of $\beta_{ic,b}$ for the
29 test signals with the LS EOS and standard $Y_e(\rho)$
parametrization. $T$ is \textit{known}.}
\label{fig:betaKnownT}
\end{center}
\begin{center}
\includegraphics[width=1\linewidth]{betaKnownT_Other.pdf}
\captionof{figure}{90\% credible intervals of $\beta_{ic,b}$ for the
18 test signals with varying EOS and $Y_e(\rho)$
parametrization. Note that $m$ refers to an increase in
$Y_e(\rho)$ of 5\%, $p$ refers to a decrease in $Y_e(\rho)$ of
5\%, and $s$ refers to the Shen EOS. $T$ is \textit{known}.}
\label{fig:betaKnownT_Other}
\end{center}
\begin{center}
\includegraphics[width=1\linewidth]{betaUnknownT.pdf}
\captionof{figure}{90\% credible intervals of $\beta_{ic,b}$ for the
29 test signals with the LS EOS and standard $Y_e(\rho)$
parametrization. $T$ is \textit{unknown}.}
\label{fig:betaUnknownT}
\end{center}
\begin{center}
\includegraphics[width=1\linewidth]{betaUnknownT_Other.pdf}
\captionof{figure}{90\% credible intervals of $\beta_{ic,b}$ for the
18 test signals with varying EOS and $Y_e(\rho)$
parametrization. Note that $m$ refers to an increase in
$Y_e(\rho)$ of 5\%, $p$ refers to a decrease in $Y_e(\rho)$ of
5\%, and $s$ refers to the Shen EOS. $T$ is \textit{unknown}.}
\label{fig:betaUnknownT_Other}
\end{center}
We yield accurate predictions of $\beta_{ic,b}$ for most of the test
signals in figure~\ref{fig:betaKnownT}. Signal 27 ($A5O3.25$ from the
catalogue) is an outlier and comes from a slowly rotating core with
uniform rotation. It is likely an outlier due to the strong
stochastic components in the GW signal from prompt postbounce
convection \cite{abdikamalov:2013}. The true values of $\beta_{ic,b}$
are on the boundary of the 90\% credible intervals for signals 3
($A1O10.25$), 9 ($A2O6.25$), 19 ($A3O5.25$), and 23 ($A4O3.25$), but
there is no distinguishable pattern between these signals. The
credible intervals are relatively small, at approximately four units
(times $10^{-2}$) long. This means that it is particularly easy to
distinguish $\beta_{ic,b}$ between GW signals.
The length of credible interval widens by a factor of $\sim 1.5$ when
changing from known to unknown signal arrival time. Incorporating an
unknown time shift increases the uncertainty of the PC coefficients
since the MCMC algorithm draws $\boldsymbol\alpha | T$. That is,
conditioning on an uncertain $T$ creates additional uncertainty for
$\boldsymbol\alpha$. However, predictions are still accurate in most
cases. We see in figure~\ref{fig:betaUnknownT}, that signal 27
($A5O3.25$) is an outlier again. Signal 23 ($A4O3.25$) is another
outlier with credible interval on the negative side of the number
line. This is an absurd and physically impossible range for a
strictly positive variable, and is a consequence of the fact that the
priors can only constrain the linear model parameters
$(\boldsymbol\gamma, \sigma^2)$. More specifically, we could not put
priors on the response variable of physical parameters
$\boldsymbol\xi$ to constrain the predicted physical parameters
$\check{\boldsymbol\xi}$. A similarity that this signal has with the
other outlier is that it comes from a slowly rotating core with weak
differential rotation.
Our methods work reasonably well when varying the EOS and
deleptonization parametrization, although we underestimate some
signals with moderate rotation in figure~\ref{fig:betaKnownT_Other}.
Three of these signals come from an increase of $Y_e(\rho)$
parametrization, one from a decrease of $Y_e(\rho)$ parametrization,
and two from the Shen EOS. When incorporating an unknown time shift
in figure~\ref{fig:betaUnknownT_Other}, the uncertainty of $T$
increases and covers the true parameters. The increase in the width
of credible interval makes it more difficult to distinguish
$\beta_{ic,b}$ between signals.
We can conclude that the methods employed in this study are moderately
sensitive to uncertainties in $Y_e(\rho)$ and EOS. It was found that
a GW signal has relatively weak dependence on the nuclear EOS by
\cite{dimmelmeier:2008}. We showed in an unpublished study
\cite{edwards:2013} that we could correctly identify between the LS
and Shen EOS for 50\% of the signals in the Dimmelmeier \etal
\cite{dimmelmeier:2008} waveform catalogue using model comparison
techniques. Note that 21\% were incorrectly identified and 29\%
unidentified. It could therefore be useful to incorporate EOS as an
additional unknown that we wish to infer in future statistical
analyses.
The results presented assume a SNR of $\rho = 20$. To test
robustness, we trialled the analysis on SNRs of $\rho = 50$ and $\rho
= 100$, which are more realistic levels for detecting CCSN events in
the Milky Way. Our predictions and credible intervals of
$\beta_{ic,b}$ were the same, regardless of the SNR. This can be
attributed to using only the posterior means of the PC coefficients in
constructing design matrix $\mathbf{A}$, and not the full spread of
the posterior distributions. This therefore removes uncertainty due
to LIGO noise and signal reconstruction when predicting $\beta_{ic,b}$
from the posterior predictive distribution.
\subsection{Classifying the precollapse differential rotation, $A$}
Precollapse differential rotation is treated as a categorical variable
with five different levels in this analysis. We define the set of
classes $C = \{A1, A2, A3, A4, A5\}$ and apply the NBC and $k$-NN
supervised learning algorithms to extract precollapse differential
rotation from each of the signals in the TC. The model features
$\mathbf{u}$ are the posterior means of the PC coefficients from the
Bayesian PCR models ($\bar{\boldsymbol\alpha}$ for the training set
and $\bar{\check{\boldsymbol\alpha}}$ for the test set). The goal of
this analysis is to let both algorithms learn from the training set to
discriminate GW signals in the test set.
\begin{table}[!h]
\caption{\label{tab:diffRot} Percentage of signals in the TC with correctly identified precollapse differential rotation $A$ using NBC and $k$-NN.}
\begin{indented}
\item[] \begin{tabular}{lcccc}
\br
&\multicolumn{2}{c}{Known $T$ (\%)}&\multicolumn{2}{c}{Unknown $T$ (\%)}\\
\cline{2-5}\noalign{\smallskip}
Differential Rotation, $A$&NBC&$k$-NN&NBC&$k$-NN\\
\mr
$A1$&&&&\\
-- Standard&83&83&83&83\\
-- $\uparrow Y_e(\rho)$&67&50&67&50\\
-- $\downarrow Y_e(\rho)$&67&83&83&100\\
-- Shen EOS&33&17&0&17\\
$A2$&50&75&50&50\\
$A3$&43&57&29&57\\
$A4$&0&80&20&80\\
$A5$&33&33&0&33\\
\br
\end{tabular}
\end{indented}
\end{table}
Table~\ref{tab:diffRot} shows the percentage of signals in the TC that
have a correctly identified level of $A$ using NBC and $k$-NN. We
compare how the methods work when using $\bar{\boldsymbol\alpha}$ and
$\bar{\check{\boldsymbol\alpha}}$ from data with known and unknown
signal arrival times.
The results between models with known and unknown signal arrival times
are quite similar. The standard GWs from class $A1$ are discriminated
well by both algorithms. The decrease (and to some degree, the
increase) in $Y_e(\rho)$ parametrization did not affect the
algorithms' abilities to discriminate. Both algorithms performed
particularly poorly for the Shen EOS test signals, which illustrates
that $A$ is sensitive to the EOS. This is in line with the findings
from \cite{abdikamalov:2013}.
The $k$-NN generally performs better than the NBC for GW signals with
weak to moderate differential rotation ($A3, A4, A5$). This could be
attributed to our choice in prior classes for the NB method. Since
models with stronger differential rotation are more populated in the
BC, they have a higher prior probability than those with weaker
differential rotation.
\section{Discussion}
We have presented a Bayesian framework for inferring the physical
parameters of CCSN from GW data. We have shown
that with a SNR of $\rho = 20$ and optimal orientation of detector to
source, we can extract $\beta_{ic,b}$ with reasonable levels of
uncertainty for the majority of injected test signals. Both of the
Bayesian PCR models presented in this paper worked well. The level
of uncertainty increased when incorporating an unknown signal arrival
time into the model, but this is no surprise as PC coefficients are
conditioned on the signal arrival time for that model. Further, we
found that our methods were moderately sensitive to varying EOS and
$Y_e(\rho)$ parametrizations, and predictions are generally good.
The chosen measure of uncertainty in this analysis was the 90\%
credible interval. A great benefit of the Bayesian framework is the
probabilistic interpretation of credible intervals, enabling one to
make statements such as, ``with probability 0.9, $\beta_{ic,b}$ is
between $2.5 \times 10^{-2}$ and $6.5 \times 10^{-2}$.''
A true strength of the methods presented in this paper is their
generality. We initially applied these techniques to the Dimmelmeier
\etal catalogue \cite{dimmelmeier:2008} as a proof of concept and then
easily transferred to the Abdikamalov \etal catalogue
\cite{abdikamalov:2013}. In this study we sampled $\beta_{ic,b}$ from
its posterior predictive distribution. This method could however be
conducted on any continuous variable of physical interest. Although
not presented here, predictions of the initial central angular
velocity $\Omega_c$ were comparable to what we found with
$\beta_{ic,b}$.
Choosing to only use the posterior means of the PC coefficients
$\bar{\boldsymbol\alpha}$ in the construction of the design matrix
$\mathbf{A}$ removed some of the variability due to LIGO noise and
signal reconstruction. The uncertainty from the Bayesian PCR
modelling step therefore does not flow onto the posterior predictive
sampling step. A more realistic case would be to incorporate this
uncertainty through an errors-in-variables model, which is commonly
used when there are measurement errors in the explanatory variables of
a regression model. We plan to explore this in a future study.
However, a benefit of our method was that predictions were essentially
independent of SNR (at least for $\rho \geq 20$).
An important task in Bayesian analysis is specifying the prior PDF to
describe our beliefs about model parameters before observing the data.
We wanted to avoid using information from the waveform catalogue as
both data and prior knowledge. It is in this light that we believe the
waveform catalogue should be used only as data, and assume complete
prior ignorance on all of the model parameters.
We applied the NBC and $k$-NN algorithm to extract precollapse
differential rotation. We found that results were comparable between
known and unknown signal arrival times. The $k$-NN algorithm
generally performed better than the NBC under the assumptions made.
In future work, we plan to investigate how the choice of prior for the
NBC affects classification, as well as exploring different metrics
such as the Mahalanobis distance (which takes correlations of the data
into account) for the $k$-NN. We are also investigating an
alternative classification routine, Bayesian ordinal probit
regression.
We introduced a constrained optimization approach to model selection
that allowed us to select an appropriate number of PCs for the
Bayesian PCR models. To our knowledge, this is the first attempt at
doing so. Techniques such as reversible jump MCMC (RJMCMC)
\cite{green:1995} have been utilized in GW data analysis contexts
\cite{umstatter:2005}. RJMCMC could prove to be a useful and more
sophisticated approach than the method presented in the current study.
Although our method required a lot of parallel computing, we found it
to be a novel solution to the model selection problem.
Our analysis assumed optimal orientation of a GW source to a single
interferometer. As presented in \cite{roever:2007a, roever:2007b} for
compact binary inspiral signals, we plan to extend the methods
presented in this study to a network of detectors. This is an important
generalization as one can triangulate the position of a GW source
using coherent data from multiple detectors. The ability to locate a
GW source would allow astronomers to compare and verify whether there
was a true astrophysical event or a glitch with electromagnetic
observations.
\ack
We thank Ik Siong Heng for a thorough reading of the script, Ernazar
Abdikamalov for providing us with the waveform catalogue and
supplementary materials, Christian D. Ott for helpful discussions, and
the New Zealand eScience Infrastructure (NeSI) for their high
performance computing facilities and support. NC's work is supported
by NSF grant PHY-1204371. This paper has been given LIGO Document
Number P1400034.
\section*{References}
\bibliographystyle{unsrt}
|
\section{Multicut requests as partially ordered sets}
The problem of finding a smallest edge cut separating vertices in a graph has received much attention over the past 50 years. \dmulticut, one of the more general forms of this problem, encompasses numerous applications in algorithmic graph theory.
\begin{framed}
\begin{description}
\item [Name:] \dmulticut.
\item[Instance:] A directed graph $G$ and pairs of terminal vertices $\{(s_1, t_1), \dotsc, (s_k, t_k)\}$ from~$G$.
\item[Problem:] Find a smallest set of edges in $G$ whose deletion eliminates all paths $s_i \to t_i$.
\end{description}
\end{framed}
Special cases of the \dmulticut problem have been met with success, although the general problem has no polynomial-time solution unless $\PT = \NP$. The classical and efficient Ford-Fulkerson algorithm \cite{FF56} solves \dmulticut for the case of a single pair of terminal vertices, yet deciding whether there exists a minimum edge cut of a given size separating both $s$ from $t$ and $t$ from $s$ in a directed graph is $\NP$-complete \cite{GVY94} as is deciding the size of a minimum edge cut separating three vertices in an undirected graph \cite{DJPSY92}.
While \dmulticut appears intractable from the perspective of $\NP$-completeness, it remains an open problem to determine whether we can find an efficient parametrized solution for \dmulticut. In practice we can optimize our solution based on other parameters besides the input length. In the case of \dmulticut, the relevant parameters are the number of (sets of) terminal vertices $k$ and and the size of the smallest solution, or \emph{cutset}, $p$. Formally a problem is \emph{fixed parameter tractable (FPT) in parameters $k$ and~$p$} if there exists an algorithm which, on input $x$, either gives a solution consistent with parameters~$k$ and~$p$ or correctly decides that no such solution exists in at most $f(k,p) \cdot \poly(\size{x})$ steps for some computable bound~$f$.
Some subcases of \dmulticut already have FPT solutions within the realm of fixed-parameter tractability. Recently Kratsch, Pilipczuk, Pilipczuk, and Wahlstr\"{o}m \cite{KPPW12} showed that \dmulticut restricted to acyclic graphs is fixed-parameter tractable when parameterized in both the size of the cutset and the number of terminals. Chitnis, Hajiaghayi, and Marx \cite{CHM12}, on the other hand, investigated \dmulticut with restrictions of the terminal pairs. They showed that \dmultiwaycut, the special case of \dmulticut where all pairs of terminal vertices must be separated in both directions, is FPT when parametrized in just the size of the cutset. In the negative direction, Marx and Razgon \cite{MR11} showed that \dmulticut is $\W[1]$-hard when parameterized the size of the cutset. Thus an FPT solution for \dmulticut, if such an algorithm exists, most likely requires parameterization in the number of terminals in addition to the size of the cutset. We remark that in this same paper \cite{MR11} Marx and Razgon also showed that the undirected \umulticut problem is FPT when parametrized in the size of the cutset. Bousquet, Daligault, and Thomass{\'e} independently achieved this same result \cite{BDT11}.
We now formalize the \latticecut problem, a subject which derives from a network security framework \cite{Pik09}. We shall show that \latticecut is equivalent to \dmulticut with respect to fixed parameter tractability.
\begin{framed}
\begin{description}
\item [Name:] \latticecut
\item[Instance:] A directed graph $G=(V,E)$ with terminal vertices $T \subseteq V$, a partially ordered set $P$, and a surjective map $\ell : T \to P$.
\item[Problem:] Find a minimum set of edges $S \subseteq E$ so that for all terminal vertices $x,y \in T$, if there is a path from $x$ to $y$ in $(V, E \setminus S)$ then $\ell(x) \geq_P \ell(y)$.
\end{description}
\vspace{-1ex}
\end{framed}
The \latticecut problem is immediately a special case of \dmulticut. Indeed, given an instance of \latticecut, we can read off from the poset $P$ and mapping $\ell: T \to P$ those pairs of terminals which must be separated in the \latticecut solution. These pairs together with the original input graph give us an instance of \dmulticut such that an edge cut is a solution to the \latticecut instance if and only if it is a solution to the \dmulticut instance. Thus if \dmulticut is fixed-parameter tractable, then so is \latticecut. We now show that the reverse is also true.
\begin{thm} \label{thm: hop}
If \latticecut is $\FPT$, then so is \dmulticut. In particular, given an instance of \dmulticut with $k$ terminal pairs and a permitted maximum of $p$ cuts, we can efficiently find an instance of \latticecut with at most $2k$ terminal nodes and a permitted maximum of $p$ cuts such that the \latticecut instance has a solution iff the \dmulticut instance does.
\end{thm}
\begin{proof}
Consider an instance of \dmulticut consisting of a graph $G$, forbidden terminal pairs $s_1 \not\to t_1, \dotsc, s_k \not\to t_k$, and a cutsize parameter $p$. We define the corresponding \latticecut instance as follows. The graph $G'$ will consist of all the nodes and edges in $G$ plus some extra nodes and edges. For each terminal node $s_i$, add a node $a_i$ and enough paths from $a_i$ to $s_i$ so that $a_i$ and $s_i$ remain connected in any solution for the \latticecut instance. In more detail
\begin{itemize}
\item add $p+1$ nodes $c_{i,1}, \dotsc, c_{i,p+1}$,
\item add an edge from $a_i$ to each $c_{i,j}$, and
\item add a further edge from each $c_{i,j}$ to $s_i$.
\end{itemize}
Similarly for each terminal node $t_i$, we add a node $b_i$ and connect $t_i$ to $b_i$ with many paths: make $p+1$ new nodes $d_{i,1}, \dotsc d_{i,p+1}$, add an edge from $t_i$ to each $d_{i,j}$, and add an edge from each $d_{i,j}$ to $b_i$. We define the poset for this \latticecut instance as follows: set $a_i$ to be greater than $b_j$ for all $i \neq j$, and all other pairs of terminal nodes are designated as incomparable.
By construction, there is a path $a_i \to b_i$ iff there is a path $s_i \to t_i$, and this condition holds even when up to $p$ edges are deleted from $G'$. If there is a \latticecut solution on $G'$ under the given poset with at most $p$ cuts, there is a further solution which is identical but avoids cutting any paths between $a_i$ and $s_i$ or $t_i$ and $b_i$. Hence we may assume that the solution has all its cuts inside the embedding of $G$ within $G'$. Transferring these cuts back to the original graph $G$ gives a solution for the \dmulticut instance. On the other hand, any solution for \dmulticut in $G$ will also be a solution for \latticecut in $G'$ because the only paths between pairs of terminal vertices in the \latticecut instance start at some $a_i$ and end at some $b_j$.
\end{proof}
Edwards, Jaeger, Muthukmaran, Rueda, Talele, Teutsch, Vijayakumar \cite{MRTVJTE12} and Jaeger, Teutsch, Talele, Erbacher \cite{EJTT13a} distilled the placement of host security mediators on a distributed system to a solution for the \latticecut problem. They interpreted the components of a distributed system as nodes in a directed graph with edges indicating which components can communicate directly with others. Some information traveling through a network will have high integrity, and other information will have lower integrity, and security is achieved by blocking all flows from lower integrity to higher integrity nodes. Terminal nodes represent both the possible attack surfaces and higher integrity entities in the system, and each terminal corresponds to a specific integrity level as measured by the poset. In this context, we can interpret \latticecut as a search for minimum intervention which mediates between all illegal information flows.
For the remainder of this paper, we will focus on the subcase of \latticecut where the poset is a chain.
\begin{framed}
\begin{description}
\item[Name:] \linearcut
\item[Instance:] A directed graph $(V,E)$ and a tuple of \emph{terminal} sets $\pair{T_1, \dotsc, T_k}$ which are subsets of $V$.
\item[Problem:] Find a smallest set of edges $S \subseteq E$ such that for any $s \in T_i$ and $t \in T_j$, if there is a path from $s$ to $t$ in $(V, E \setminus S)$, then $i \geq j$.
\end{description}
\end{framed}
That is, \linearcut wants to find a smallest edge cut which prevents every terminal set $T_i$ from flowing to $T_j$ whenever $j > i$. We shall show that \linearcut, which is $\NP$-hard in the sense of Proposition~\ref{prop: lcnp}, is FPT when parameterized in the size of the cutset. Rephrased in terms of posets, Chitnis, Hajiaghayi, and Marx's algorithm \cite{CHM12} for \dmultiwaycut shows that \latticecut is FPT parametrized in the cutset size when the underlying poset is an antichain.
\section{A parameterized algorithm for Linear Cut}
We shall show that \linearcut is FPT when parametrized in the size of cutset. Before presenting our parametrized algorithm, we first analyze the following example which illustrates why the na\"{i}ve greedy cut does not yield an optimal solution. The graph given in Figure~\ref{fig: greedy} has three terminal vertices $t_0$, $t_1$, and $t_2$, and we would like to find a small set of edges whose removal eliminates all paths from $t_0$ to either $t_1$ or $t_2$ as well as all paths from $t_1$ to $t_2$. Consider the greedy algorithm which uses the Ford-Fulkerson algorithm to first eliminate all paths from $t_0$ to the other terminal vertices and then again to extinguish the paths from $t_1$ to $t_2$. A minimal edge cut from $t_0$ to the set $\{t_1, t_2\}$ has size~3, so let us assume that the algorithm chooses edges $\{a,b,c\}$. Now a minimal edge cut from $t_1$ to $t_2$ has size~2, for example $\{h,i\}$. Thus this greedy algorithm solves the \linearcut instance with a cut of size~5. On the other hand, $\{d,e,f,g\}$ is a solution of size~4.
\begin{figure}[h]
\caption{The greedy algorithm is not optimal.}
\begin{center}
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,node distance=2cm,
thick,terminal/.style={circle,fill=white!20,draw}, regular/.style ={circle,fill=white!20,draw},scale = 1]
\node[terminal] (t_0){$t_0$};
\node[regular] (t) [below of =t_0, yshift = 0.5cm] {\phantom{$t_0$}};
\node[regular] (l) [below left of=t, yshift = 0.5cm] {\phantom{$t_0$}};
\node[regular] (r) [below right of=t, yshift = 0.5cm] {\phantom{$t_0$}};
\node[regular] (b) [below right of=l, yshift = 0.5cm] {\phantom{$t_0$}};
\node[terminal] (t_1)[below left of=l, xshift = 0.5cm, yshift = -0.5cm]{$t_1$};
\node[terminal] (t_2) [below right of=r, xshift = -0.5cm, yshift = -0.5cm] {$t_2$};
\path[every node/.style={font=\sffamily\small}]
(t_0) edge [bend right] node [left] {a} (l)
edge node [left] {b} (t)
edge [bend left] node [right] {c} (r)
(t) edge (l)
edge (r)
(l) edge (b)
(r) edge [bend right=10] (b)
(b) edge [bend right=10] (r)
(l) edge [bend right=10] node [left] {d} (t_1)
(t_1) edge [bend right=7] node [right] {h} (l)
(t_1) edge [bend left=7] node [above] {i} (b)
(b) edge [bend left=10] node [below] {e} (t_1)
(r) edge node [right] {g} (t_2)
(b) edge node [below] {f} (t_2);
\end{tikzpicture}
\end{center}
\label{fig: greedy}
\end{figure}
We now describe our parametrized solution for \linearcut. Our algorithm either outputs a solution cut of size less $\leq p$ or returns \NO if no such cut exists. Our construction exploits a technique used in Chen, Liu, and Lu's fixed-parameter solution \cite{CLL09} to the \umultiwaycut problem in undirected graphs which improved a result of Marx \cite{Mar06}. A similar idea appeared earlier in Chen, Liu, Lu, O'Sullivan, and Razgon's algorithm \cite{CLLOR08} for \skewmulticut, a key step in their parametrized solution for \dfvs. We remark that the pushing of important separators technique along the lines of \cite[Theorem~3.7]{Mar06} gives a parameterized solution for \linearcut in time $O(4^{p^3} n^{O(1)})$, and using a reduction to the \skewmulticut algorithm in \cite{CLLOR08} one can also show that \linearcut has a solution which runs in the same time as the algorithm given below, namely $O(4^p p n^4)$.
An ($X$,$Y$)-\emph{separator} is a set of edges such that any path from $X$ to $Y$ passes through one of its members. Our solution, Algorithm~\ref{alg: lc} proceeds in two phases. First we handle the trivial cases where $\T = \pair{}$, $p=0$, or $T_1$ is either already separated from the other terminals or can't be separated with $p$ edge cuts (lines 1--12). The second phase picks an edge pointing out of the $T_1$ region and checks whether making it undeleteable hurts the min size of a $(T_1, T_2 \cup \dotsb \cup T_k)$-separator. If not we add the edge to the list of undeleteable edges, and if so we branch on the only two possibilities: either the edge belongs in the \linearcut solution or it doesn't.
The following theorem gives the main justification for this algorithm. A set of edges is a \emph{linear cut} with respect to the $k$-tuple of terminals $\pair{T_1, \dotsc, T_k}$ if there is no path from $T_i$ to $T_j$ whenever $i<j$ once these edges have been removed.
\newpage
\ \vspace{.3in}
\begin{algorithm}[h!]
\algnewcommand\algorithmicinput{\textbf{Input:}}
\algnewcommand\Input{\item[\algorithmicinput]}
\algnewcommand\algorithmicoutput{\textbf{Output:}}
\algnewcommand\Output{\item[\algorithmicoutput]}
\algnewcommand\llet{\textbf{let}\xspace}
\algnewcommand\LC{\mathsf{LC}\xspace}
\caption{FPT algorithm for \linearcut parameterized in cutset size.}
\label{alg: lc}
\begin{algorithmic}[1]
\Input A graph $G = (V, E)$, a $k$-tuple of terminal sets $\pair{T_1, T_2, \dotsc T_k}$ which are subsets of $V$, some undeletable edges $F \subseteq E$, and a parameter $p$.
\Statex
\Output A set of $\leq p$ edges in $E \sm F$ such that when these edges are deleted from $G$ there is no path from $T_i$ to $T_j$ for any $i<j$, if such a set of edges exists, otherwise return \NO.
\Statex
\Function{$\LC$}{$(V,E), \pair{T_1, \dotsc, T_k}, F, p$}
\State For ease of reading, let $\T = \pair{T_1, \dotsc, T_k}$.
\If {$\T = \pair{}$} \Return $\emptyset$;
\ElsIf {$p \leq 0$}
\If {for all $i <j$, $T_j$ is not reachable from $T_i$ in $G$} \Return $\emptyset$;
\Else\ \Return \NO;
\EndIf
\EndIf
\State
\llet $m$ be the size of a minimum $(T_1, T_2 \cup \dotsb \cup T_k)$-separator which does not include edges from $F$.
\If {$m > p$ or no separator exists due to undeleteable edges} \Return \NO;
\ElsIf {$m = 0$} \Return $\LC((V,E), \pair{T_2, \dotsc, T_k}, F, p)$;
\Else
\State \llet $e \in E \sm F$ be an edge with a tail reachable from $T_1$ via undeleteable edges.
\If {the size of a minimum $(T_1, T_2 \cup \dotsb \cup T_k)$-separator which does not include edges from $F \cup \{e\}$ exists and is equal to $m$,}
\State \Return $\LC ((V,E), \T , F \cup \{e\}, p)$;
\ElsIf {$\{e\} \cup \LC ((V,E \sm \{e\}), \T, F, p-1)$ or $\LC ((V,E), \T, F \cup \{e\}, p)$ is not \NO,}
\State \Return the first of these two found to have a solution;
\Else
\State \Return NO;
\EndIf
\EndIf
\EndFunction
\end{algorithmic}
\end{algorithm}
\newpage
\begin{thm} \label{thm: bobo}
Let $\pair{(V,E), \T, F, p}$ be an input to Algorithm~1, where $\T$ is an abbreviation for $\pair{T_1, \dotsc, T_k}$, and let $e$ be an edge pointing from some node reachable from $T_1$ via undeleteable edges to a node outside $T_1 \cup F$. Suppose that the smallest $(T_1, T_2 \cup \dotsb \cup T_k)$-separator with undeletable edges $F$ is the same size as the smallest $(T_1, T_2 \cup \dotsb \cup T_k)$-separator with undeletable edges $F \cup \{e\}$ and has cardinality at most $p$. Then the smallest linear cut among the terminal sets $\pair{T_1, \dotsc, T_k}$ with undeletable edges $F$ in $(V,E)$ has the same size as the smallest linear cut among these same terminals with undeletable edges $F \cup \{e\}$.
\end{thm}
\begin{proof}
First note that making edges undeleteable can only increase the size of the smallest cut. Hence it suffices to show, under the hypothesis of the theorem, that the smallest linear cut with forbidden edges $F \cup \{e\}$ is no bigger than a minimal linear cut with forbidden edges $F$.
Let $S$ be a minimal $(T_1, T_2 \cup \dotsb \cup T_k)$-separator with undeletable edges $F \cup \{e\}$. Then $S$ is also a separator between these same sets with undeletable edges $F$, and by the assumption of the theorem $S$ is also a minimal such separator. Let $W$ be a minimal linear cut in $G = (V,E)$ for $\T$ with undeletable edges $F$, and let $R$ denote the set of edges that are reachable from $T_1$ in $(V, E \sm S)$. We shall show that $W' = (W \cup S) \sm R$ is a linear cut in $G$ for $\T$ with undeletable edges $F \cup \{e\}$ which is no larger than $W$. Since making edges undeletable can only increase the size of a smallest solution, $W'$ will indeed be minimal.
For clarity, we reformulate the problem instance without undeletable edges. We replace each undeletable edge $(x,y) \in F \cup \{e\}$ with $p+1$ new, regular edges from $x$ to $y$, whereby transforming the graph into a multigraph without any undeletable edges. Now any linear cut (resp.\ $(T_1, T_2 \cup \dotsb \cup T_k)$-separator) consisting of at most $p$ edges will be a solution in the transformed multigraph if and only if it is a solution in the original graph. The reason is that there are not enough total cuts in the instance to sever connectivity between any vertices with $p+1$ multiedges. Thus these edges are effectively undeletable, and of course cuts not involving undeletable edges or multiedges will work the same in both the original and transformed instance.
First we argue that $W'$ is not larger than $W$ by proving $\size{S \sm W} \leq \size{W \cap R}$. Since $S$ does not contain any of the undeletable, multiedge parts of $G$, by Menger's Theorem \cite[Theorem~7.45]{KT06}, or more precisely its generalization to sets of vertices \cite[Lemma~1]{CLL09}, there are $\size{S}$ disjoint edge paths from $T_1$ to $\bigcup_{j > 1} T_j$, each containing an edge in $S$. It follows that there are $\size{S \sm W}$ disjoint edge paths from $T_1$ to $S \sm W$. Now suppose that $\size{W \cap R} < \size{S \sm W}$. Then there must be a path from $T_1$ to some edge $x \in S \sm W$ which avoids $W \cap R$. Furthermore, by minimality of $S$, there is a path from $x$ to some terminal set $T_j$ with $j > 1$. But now there is a path from $T_1$ to some $T_j$ which avoids $W$, contradicting that $W$ is a linear cut.
It remains to show that $W'$ is in fact a linear cut in $G$ for $\T$ with undeletable edges $F \cup \{e\}$. Let $Q$ be a forbidden path. If $Q$ does not intersect $R$, then it must pass through $W \sm R$ and hence through $W'$. On the other hand, suppose that $Q$ does pass through $R$. Since $T_1$ is the least-indexed terminal set, $Q$ must end at $T_j$ for some $j > 1$, and therefore $Q$ must pass through $S \subseteq W'$. In either case, removing $W'$ eliminates the forbidden path $Q$.
\end{proof}
\begin{thm} \label{thm: ewok}
Algorithm~\ref{alg: lc} finds a solution in time $O[4^p p\cdot (\size{V} + \size{E}) \cdot \size{E}]$, if one exists, and outputs NO otherwise.
\end{thm}
\begin{proof}
Line~13 of Algorithm~\ref{alg: lc} selects an edge $e \in E \sm F$ for consideration. If the condition for edge~$e$ in line~14 holds, then preserving $e$ does not hurt the $(T_1, T_2 \cup \dotsb \cup T_k)$-separator, and therefore by Theorem~\ref{thm: bobo} no harm comes to the \linearcut instance by adding $e$ to the list of undeletable edges. If this condition is not satisfied, then the algorithm exhaustively searches both for a solution containing the edge~$e$ (Option~1) and for a solution not containing~$e$ (Option~2). In Option~1, the algorithm searches for a solution of size $p-1$ containing $e$, and in Option~2, the size of the smallest$(T_1, T_2 \cup \dotsb \cup T_k)$-separator increases by~1. Along any branch of the algorithm, either of these two Options can occur at most $p$ times for each terminal before the algorithm returns \NO, and the latter happens only when exhaustive search fails to find a solution. Hence the algorithm eventually terminates with the correct answer.
We can refine our analysis further to show that there are at most $4^p$ possible branches in the algorithm. We argue that any branch of the algorithm witnesses at most $2p$ branching splits. Suppose that the initial input parameter is $p$ and that the smallest $(T_1, T_2 \cup \dotsb \cup T_k)$-separator has size $m$. Since each iteration of Option~1 decreases the size of the minimal $(T_1, T_2 \cup \dotsb \cup T_k)$-separator by~1, the path which always chooses Option~1 will witness exactly~$m$ branches up to the point where Line~11 of Algorithm~\ref{alg: lc} recognizes that $T_1$ has been separated and removes it from further consideration. Each time Option~2 is chosen along the path, the size of the smallest$(T_1, T_2 \cup \dotsb \cup T_k)$-separator increases by at least~1, so if Option~2 happens $r$ times, then Option~1 must happen a total of at least $m + r$ times before $T_1$ is separated. Thus the size of the cutset size parameter when $T_1$ becomes separated is at most $p - m - r$, the initial parameter value minus the number of times Option~1 was chosen, and the total number of splits witnessed is $(m+r) + r$, which is at most twice the number of edges added to the cutset. The same counting argument holds for separators for successive $T_i$'s and it follows that each search path can witness at most $2p$ splits in case the algorithm succeeds.
The number of steps between each encounter with an Option is essentially the time required to check whether a separator size~$p$ exists, which is $O[p (\size{V} + \size{E})]$ by the argument in \cite[Lemma~2]{CLL09}, times the number edges. The multiplicative factor of $\size{E}$ comes from the potential recursion in line~15. Hence the total runtime is $O[2^{2p} p \cdot (\size{V} + \size{E}) \cdot \size{E}]$.
\end{proof}
\begin{cor}
\linearcut is fixed-parameter tractable when parameterized in the size of the cutset.
\end{cor}
\section{Hardness result}
Marx and Razgon \cite{MR11} showed that \dmulticut parameterized in the size of the cutset is $\W[1]$-hard by reducing this problem to the known $\W[1]$-hard problem $\clique$. Therefore the following is immediate from Theorem~\ref{thm: hop}.
\begin{cor}
\latticecut is $\W[1]$-hard when parameterized in the size of the cutset.
\end{cor}
Whether $\dmulticut$ is fixed-parameter tractable when parameterized in both the size of the cutset and the number of terminals remains an open problem, even in the case where we fix the number of terminal pairs at $k=3$ \cite{CHM12, MR11}. $\linearcut$ for $k=2$ is possible via the Ford-Fulkerson algorithm, however for longer chains the problem also becomes $\NP$-hard.
\begin{prop} \label{prop: lcnp}
Deciding whether a \linearcut instance has a solution of size $p$ is $\NP$-complete for $k=3$ terminals.
\end{prop}
\begin{proof}
\linearcut is trivially in $\NP$ as one can easily check by breadth-first search whether a given set of edges is a solution.
We reduce the undirected \umultiwaycut problem for $k=3$, which is $\NP$-hard \cite{DJPSY92}, to the \linearcut problem for $k=3$. Let $G$ be an undirected graph with terminal nodes $s$, $t$ and $u$ be an instance of \umultiwaycut, the problem of finding a smallest edge cut which separates $s$, $t$, and $u$. Construct a new directed graph $G'$ which has the same vertices as $G$ except for each edge $e=\{x,y\}$ in $G$ we also add two new vertices $a_e$ and $b_e$. The edges from $G$ do not carry over to $G'$, and instead we add directed edges $(x, a_e)$, $(y, a_e)$, $(a_e, b_e)$, $(b_e, y)$, and $(b_e , x)$. We call this collection of edges the \emph{gadget} for $e$. Our \linearcut instance consists of the graph $G'$ together with the embedded terminals nodes $s$, $t$, and $u$ from $G$ with the (arbitrary) tuple ordering $\pair{s,t,u}$. Technically we treat the terminal nodes here as singleton sets when formulating this instance of \linearcut.
Assume $C = \{e_1, \dotsc, e_p\}$ is a \umultiwaycut solution for $G$. We claim that $C' = \{(a_{e_1}, b_{e_1}), \dotsc, (a_{e_p}, b_{e_p})\}$ is then a \linearcut solution for $G'$. Suppose there were some prohibited path in $G'$ between two terminals, say $s$ and $t$, which avoids $C'$. This path must have the form
\[
s \to a_{(s,x_1)} \to b_{(s,x_1)} \to x_1 \to a_{(x_1,x_2)} \to b_{(x_1,x_2)} \to x_2 \to \dotsb \to t
\]
for some vertices $x_1, x_2, \dotsc$ in $G$. Contracting all the $a_i$'s and $b_i$'s from this path yields a path from $s$ to $t$ in $G$ which avoids $C$, which is impossible.
Conversely, assume that $C' = \{d_1, \dotsc, d_p\}$ is a \linearcut solution for $G$. For each $i \leq p$, let $e_i$ be the gadget for the edge in $G$ which $d_i$ belongs to. Then $C = \{e_1, \dotsc, e_p\}$ is a \umultiwaycut solution for $G$ as any path $x_1 \to \dotsb \to x_k$ between terminals in $G$ avoiding $C$ gives rise to a path between the same terminals in $G'$ which avoids $C'$, namely
\[
x_1 \to a_{(x_1,x_2)} \to b_{(x_1,x_2)} \to x_2 \to a_{(x_2, x_3)} \to b_{(x_2,x_3)} \to x_3 \to \dotsb \to x_k,
\]
which cannot exist. Thus \umultiwaycut is polynomial-time reducible to \linearcut.
\end{proof}
\section{Approximation}
It seems difficult to efficiently approximate \dmulticut \cite{AACM07, CK06, Gup03}, which indicates that \latticecut may not have a good approximation algorithm either. The best known polynomial-time approximation algorithm for \dmulticut is just under $O(\sqrt n)$ \cite{AACM07}. We wonder whether \linearcut may be easier to approximate.
Recall that $\dmultiwaycut$ is the problem of \latticecut restricted to the instances where the underlying poset is an antichain.
\begin{framed}
\begin{description}
\item[Name:] \dmultiwaycut
\item[Instance:] A directed graph $(V,E)$ and a tuple of \emph{terminal} sets $T_1, \dotsc, T_k$ which are subsets of $V$.
\item[Problem:] Find a smallest set of edges $S \subseteq E$ such that there is no path from $T_i$ to $T_j$ in $(V, E\sm S)$ for all $i \neq j$.
\end{description}
\end{framed}
Garg, Vazirani, and Yannakakis \cite{GVY94} gave a $2 \log n$ approximation for \dmultiwaycut, later improved to a factor of 2 by Naor and Zosin \cite{NZ97} using an LP relaxation. The undirected \umultiwaycut problem for $k$~terminals has a simple $2 - 2/k$ approximation algorithm using isolated cuts \cite{DJPSY92} and even a $1.5 - 2/k$ approximation using LP relaxation \cite{CKR98} (see also \cite{Vaz03}). By making two calls to Algorithm~\ref{alg: lc}, we can obtain a simple approximation to \dmultiwaycut which runs faster than Chitnis, Hajiaghayi, and Marx's $2^{2^{O(p)}}n^{O(1)}$-time exact solution \cite{CHM12} but does not beat Naor and Zosin's polynomial-time 2-approximation \cite{NZ97}.
\begin{cor} \label{cor: 2chm}
One can find a solution for \dmultiwaycut of instance size~$n$ in time $O(4^p p n^4)$ which is within a factor of two of optimal whenever a solution of size~p exists.
\end{cor}
\begin{proof}
Assume that $T_1, \dotsc, T_k$ are the terminal sets which need to be separated in the directed graph $(V,E)$. Using Algorithm~\ref{alg: lc}, make one \linearcut which cuts using the terminal sets $\pair{T_1,\dotsc, T_k}$ and another which uses this $k$-tuple reversed, $\pair{T_k, \dotsc, T_1}$. The union of these two cuts is a solution to the \dmultiwaycut instance, when both exist, and neither cut is larger than the smallest possible solution.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:introduction}
Jupiter Trojans and Hilda asteroids are small primitive bodies located beyond the snow line, around the L4 and L5 Lagrange points of Jupiter at $\sim$\,5.2 AU (Trojans) and in the 2:3 mean-motion resonance with Jupiter near 3.9\,AU (Hilda asteroids). Their origin remains a major challenge to current theories of the formation of the solar system.
There are two current models, the Grand Tack (\citealt{Walsh:2011co}) and the Nice model (\citealt{Morbidelli:2005dr, Levison:2009jta}), each one addressing distinct epochs of the early dynamical evolution of the solar system, and which - when taken together - make key predictions on the origin of both the Hilda and the Jupiter Trojan populations. The Grand Tack model addresses the early dynamical evolution of the solar system 3 to 5 Myr after its formation; it suggests that Jupiter roamed inward as close as the present location of Mars ($\sim$1.5\,AU) and then outward. This migration profoundly influenced the solar system, causing a substantial radial mixing of planetesimals throughout the solar system and making Mars smaller than it should have been. The Nice model addresses the late dynamical evolution of the solar system $\sim$700 Myr after its formation; it suggests that a large portion of both the Trojans and the Hilda asteroids formed in more distant regions - typically in the primordial transneptunian disk, the precursor of the Kuiper Belt - and subsequently chaotically migrated towards the inner solar system as a consequence of the outward migration of all four giant planets.
Taken together, these models suggest that the immediate precursors of both the Hilda asteroids and Jupiter Trojans are Kuiper Belt objects (Nice model) and that the latter are not a homogeneous population (Grand Tack model), comprising a large portion of objects formed at large heliocentric distances ($\geq$\,10\,AU) and a minor fraction of planetesimals formed closer to the Sun ($\leq$\,3\,AU). The Trojans and Hilda asteroids could therefore represent a condensed or mixed version of the primordial solar system, very much like the asteroid belt.
Until recently, telescopic observations of both populations had only focused on the brightest, that is, the largest, objects. Their spectroscopic observations in the visible \citep{Dahlgren:1995tp, Dahlgren:1997uta, Bus:2002bea, Lazzaro:2004ja, Fornasier:2004if, Fornasier:2007kf, Dotto:2006kx, Roig:2008jx} and near-infrared \citep{Dumas:1998jn, Dotto:2006kx, Yang:2011fl, Emery:2011kr} have revealed uniformly red and featureless spectra, and their albedo measurements \citep{Fernandez:2003kl} have indicated low values ($\sim$0.07) for both populations.
However, more recent measurements with the Spitzer Space Telescope (SST, \citealt{Fernandez:2009joa}) and with WISE \citep{Grav:2011eqa, Grav:2012ho, Grav:2012cl} have revealed a small group ($\sim$1.7$\%$ of the total population) of presumably bright (${\rm p}_{\rm v}$\,$\geq$\,0.15) objects among both Hilda asteroids and Jupiter Trojans, which might indicate a minor fraction of interlopers - that is, objects not originating from the Kuiper Belt region, with a reflectance spectrum different from the typical X\,/\,T\,/\,D-class asteroids - among these two populations.
In this paper, we report the very first spectroscopic characterization of a sample of high-albedo Hilda asteroids and Trojans.
\section{Observations and data reduction}
\label{sec:observations}
The spectra presented in this work were collected between April and December 2013 using two different instruments, X-SHOOTER (ESO/VLT, Chile) and SpeX (NASA/IRTF, Hawaii). A total of 17 spectra for 12-high-albedo Hilda asteroids and Trojans (with geometric albedo ${\rm p}_{\rm v}$\,$\ge$\,0.15) were obtained; in particular, a few objects were observed twice with both instruments. All observations were performed under clear sky conditions with the seeing being predominantly in the $0.6''$ to $1.2''$ range. Observation circumstances are summarized in Table \ref{tab:obs}.
\subsection{VLT/X-SHOOTER}
\label{sec:xsh}
Observations with X-SHOOTER were performed in visitor mode. X-SHOOTER is a medium-resolution {\it Echelle} spectrograph that covers the 0.3$-$2.5\,{\rm$\mu$m} spectral range in one shot by splitting the incoming light into three beams, each sent to a different arm (UVB, VIS, NIR) of the instrument \citep{Vernet:2011bya}. We used the $1.6''$, $1.5''$ and $1.2''$ slits for the UVB, VIS and NIR arms respectively, each arm has a resolving power in the 3000$-$5500 range. To monitor the high luminosity and variability of the sky in the near-IR, the telescope was moved along the slit during the acquisition of the data so as to obtain a sequence of spectra located at two different positions on the array. These paired observations provided near-simultaneous sky and detector bias measurements.
The asteroid observations were interspaced with those of the standard stars SA\,102-1081, SA\,98-978, SA\,93-101, and Hyades\,64 \citep{Landolt:1973br}. These solar analogs were observed at similar airmass with respect to our asteroid targets to ensure a good correction for atmospheric extinction.
Note that the slit was aligned with the parallactic angle for all asteroid and star observations to limit the loss of flux at short wavelengths due to differential atmospheric refraction.
Data reduction - which included bad-pixel mapping and correction, flat-fielding, wavelength calibration, merging of the different orders, and subtraction of dithered pairs - was performed with the X-SHOOTER ESO pipeline (v2.3.0) using the ESO reflex environment \citep{Freudling:2013cp}. After they were extracted, the asteroid spectra were divided by those of the solar analogues to produce reflectance spectra that were smoothed with a median-filter technique, using a box of 7 pixels in the spectral direction for each point of the spectrum.
Note that the near-IR data were obtained using the non-destructive readout mode of the detector, in which each independent pixel has a threshold limit of about 42\,kADUs that - if reached before the end of the integration time - causes the ADU level to be extrapolated from values below the threshold. This mode avoids saturation but can return an incorrect extrapolation of the signal, especially if the observing conditions are variable. An analysis of our raw spectra revealed that, during some observations, the {\it K}-band signal did reach the ADU threshold limit because of the contribution of the sky background, which resulted in a potentially incorrect extrapolation (in our case a drop in flux) of $\sim$\,5 to 10\% of the signal in the 2.0$-$2.5\,$\mu$m range. The affected spectra are flagged in Table \ref{tab:obs} to prevent alteration of our final analysis (see Section \ref{sec:results}).
\subsection{IRTF/SpeX}
\label{sec:spex}
IRTF observations of some objects in our sample (Table \ref{tab:obs}) were carried out remotely from Europe. The low- to medium-resolution near-IR spectrograph and imager SpeX \citep{Rayner:2003hfa} combined with a 0.8\,$\times$\,15 arcsec slit was used in the low-resolution prism mode to acquire the spectra in the 0.7$-$2.5 \,{\rm$\mu$m} wavelength range. Our solar analogue stars were SA\,93-101, Hyades\,64, and SA\,98-978 \citep{Landolt:1973br}. Standard techniques for near-IR spectroscopy data reduction (see above) were used with the software Spextool \citep{Cushing:2004bq} to extract the asteroid and solar analogue spectra. Finally, the asteroid reflectance was obtained by dividing each asteroid average spectrum by the average solar star spectrum.
\begin{table*}[tbh]
\scriptsize
\begin{center}
\caption[]{\em{\small Observation circumstances}}
\label{tab:obs}
\begin{tabular}{llccccccccc}
\hline \noalign {\smallskip}
No. & Name & Group & Date & Instrument & AM & Exp. time (s) & {\it D} (km) & ${\rm p}_{\rm v}$ & Ref. & \\
\hline \noalign {\smallskip}
1162 & 1930 AC & Hilda & 2013-08-29 & X-SHOOTER & 1.464$-$1.501 & 2$\times$880/900/900 & 40.38 $\pm$ 0.30 & 0.186 $\pm$ 0.037 & 1 & \\
& & & 2013-09-01 & SpeX & 1.038$-$1.074 & 10$\times$120 & & & & \\
3290 & 1973 SZ$_1$ & Hilda & 2013-08-28 & X-SHOOTER & 1.201$-$1.277 & 3$\times$880/900/900 & 10.18 $\pm$ 0.30 & 0.324 $\pm$ 0.055 & 1 & \\
& & & 2013-08-29 & X-SHOOTER & 1.316$-$1.422 & 880/900/900 & & & & \\
5023 & 1985 TG$_3$ & Trojan & 2013-04-13 & X-SHOOTER & 1.400$-$1.716 & 3$\times$880/900/900 & 27.85 $\pm$ 3.51 & 0.173 $\pm$ 0.093 & 2 & * \\
9713 & 1973 SP$_1$ & Trojan & 2013-04-14 & X-SHOOTER & 1.370$-$1.652 & 3$\times$880/900/900 & 18.65 $\pm$ 3.38 & 0.168 $\pm$ 0.086 & 2 & * \\
11249 & 1971 FD & Hilda & 2013-08-28 & X-SHOOTER & 1.065$-$1.070 & 2$\times$880/900/900 & 9.97 $\pm$ 0.88 & 0.371 $\pm$ 0.097 & 1 & \\
& & & 2013-08-29 & X-SHOOTER & 1.067$-$1.077 & 880/900/900 & & & & \\
& & & 2013-09-01 & SpeX & 1.288$-$1.631 & 26$\times$120 & & & & \\
13331 & 1998 SU$_{52}$ & Trojan & 2013-04-15 & X-SHOOTER & 1.313$-$1.460 & 3$\times$880/900/900 & 17.68 $\pm$ 1.54 & 0.171 $\pm$ 0.033 & 2 & * \\
14669 & 1999 DC & Hilda & 2013-12-13 & SpeX & 1.002$-$1.023 & 30$\times$120 & 16.11 $\pm$ 0.66 & 0.206 $\pm$ 0.038 & 1 & \\
15529 & 2000 AA$_{80}$ & Trojan & 2013-04-16 & X-SHOOTER & 1.295$-$1.431 & 3$\times$880/900/900 & 16.43$\pm$ 1.33 & 0.198 $\pm$ 0.093 & 2 & * \\
24452 & 2000 QU$_{167}$ & Trojan & 2013-08-28 & X-SHOOTER & 1.402$-$1.426 & 880/900/900 & 18.69 $\pm$ 0.99 & 0.184 $\pm$ 0.029 & 2 & \\
& & & 2013-08-29 & X-SHOOTER & 1.326$-$1.333 & 880/900/900 & & & & \\
32430 & 2000 RQ$_{83}$ & Trojan & 2013-08-29 & X-SHOOTER & 1.447$-$1.520 & 2$\times$880/900/900 & 13.37 $\pm$ 0.55 & 0.157 $\pm$ 0.007 & 2 & \\
63284 & 2001 DM$_{46}$ & Trojan & 2013-04-13 & X-SHOOTER & 1.267$-$1.548 & 3$\times$800/900/900 & 12.27 $\pm$ 0.68\,/\,10.26 $\pm$ 1.32 & 0.129 $\pm$ 0.026\,/\,0.252 $\pm$ 0.050 & 1\,/\,3 & * \\
& & & 2013-04-15 & X-SHOOTER & 1.275$-$1.593 & 3$\times$880/900/900 & & & & * \\
65227 & 2002 ES$_{46}$ & Trojan & 2013-04-14 & X-SHOOTER & 1.335$-$1.626 & 3$\times$880/900/900 & 13.58 $\pm$ 1.51\,/\,14.04 $\pm$ 0.10 & 0.126 $\pm$ 0.035\,/\,0.179 $\pm$ 0.003 & 1\,/\,3 & * \\
& & & 2013-04-16 & X-SHOOTER & 1.341$-$1.487 & 3$\times$880/900/900 & & & & * \\
\hline \noalign {\smallskip}
\end{tabular}
\begin{flushleft}
{\bf Notes.} {\it D}: effective diameter; ${\rm p}_{\rm v}$: optical albedo. Exp. times for X-SHOOTER are given for the UVB/ VIS/NIR arms. The asterisk indicates objects for which the signal has been partially extrapolated in the K-band. \\
{\bf References.} 1: \citealt{Grav:2012ho}; 2: \citealt{Grav:2011eqa, Grav:2012cl}; 3: \citealt{Fernandez:2009joa}. \\
\end{flushleft}
\end{center}
\end{table*}
\section{Results}
\label{sec:results}
The X-SHOOTER and SpeX reflectance spectra were normalized to unity at 1\,{\rm$\mu$m} and are shown in Fig.\,\ref{fig:trojans} for Trojans and in Fig.\,\ref{fig:hildas} for Hilda asteroids. A comparison of our spectra with those of low-albedo Hilda asteroids and Jovian Trojans \citep{Dahlgren:1995tp, Dahlgren:1997uta, Fornasier:2004if, Fornasier:2007kf, Roig:2008jx, Emery:2011kr} suggests that both populations have very similar spectral properties (Figure\,\ref{fig:mean}). This spectral similarity is confirmed by placing the objects into the taxonomic system of \citet{Bus:1999tw} for the X-SHOOTER spectra (using only the visible portion of our spectra to avoid the saturation problem in the K band) and in the \citet{DeMeo:2009gz} system for the near-IR (SpeX) spectra. Whereas all Trojans appear to be D-types based on the visible range alone, a comparison of their complete visible and near-IR spectra with the average \citet{DeMeo:2009gz} spectra for X-, T- and D-types reveals that their average spectrum is closer to the T-type. The only exception is (24452) 2000 QU$_{167}$, which has a bluer slope that is consistent with X-types. Hilda asteroids have slightly redder slopes than Trojans, with values in between those of the T- and D-type spectra.\footnote{Note that most authors preferentially use a broader definition of the X- and D- classes instead of using the T-class (e.g., \citealt{GilHutton:2008de, DeMeo:2014hz}), which explains the scarcity of T-type classification of asteroids in the literature.}
\begin{figure*}
\centering
\includegraphics[angle=0,width=0.8\linewidth, trim=0cm 8cm 0cm 1.6cm, clip]{all_spt_trojans.eps}
\caption{X-SHOOTER spectra of Jupiter Trojans normalized to unity at 1\,$\mu$m. Second spectrum from multiple observations of the same asteroid are overplotted in red. The grey bands indicate possible saturation in the X-SHOOTER data (see Section\,\ref{sec:xsh}).
}
\label{fig:trojans}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[angle=0,width=0.8\linewidth, trim=0cm 11.5cm 0cm 1.6cm, clip]{all_spt_hildas.eps}
\caption{X-SHOOTER and SpeX spectra of Hilda asteroids normalized to unity at 1\,$\mu$m.
}
\label{fig:hildas}
\end{figure*}
A careful inspection of the spectra furthermore reveals that none of them displays any apparent absorption feature within the level of noise of our data set. The absence of absorption bands in the spectra of Trojans was also reported by \citet{Emery:2011kr} for low-albedo objects and is consistent with the few spectra that have been acquired for low-albedo Hilda asteroids \citep{Dahlgren:1995tp, Dahlgren:1997uta, Dumas:1998jn, Bus:2002bea, Takir:2012cza}.
As a last step, we performed a more detailed comparison of our dataset with previously published spectra of low-albedo Trojans (this is not possible for Hilda asteroids as both visible and near-IR spectra have not yet been collected for a statistically significant sample). Specifically, we calculated the spectral slope over the broadest possible wavelength range (0.55$-$1.8\,${\rm \mu}$m) both for our Trojan sample and for the low-albedo objects measured by \citet{Emery:2011kr}. The near-IR spectra of these authors were complemented by the visible spectra of \citet{Vilas:1993ij}, \citet{Xu:1995gy}, \citet{Bus:2002bea}, \citet{Lazzaro:2004ja}, \citet{Fornasier:2004if}, and \citet{Fornasier:2007kf}. Figure\,\ref{fig:slopes} shows that our spectra fall within the range of the blue Trojans, although they are on average redder. Note that they are not sufficiently red to fill the gap between the blue and the red Trojans that results from the bimodality reported by \citet{Emery:2011kr}, yet they narrow the gap to some extent. The redder colours of our sample with respect to the blue Trojans from \citet{Emery:2011kr} are essentially due to steeper slopes in the UVB/VIS range.
\begin{figure}[ht]
\centering
\includegraphics[trim=0mm 2mm 0mm 5mm, clip, width=0.43\textwidth]{mean_spectra_all.eps}
\caption{All spectra are normalized to unity at 0.55\,$\mu$m. (Top) Mean spectra of the two groups identified by \citet{Emery:2011kr} among low-albedo Trojans compared with the mean spectrum of bright Trojans (from this study). (Bottom) Visible spectra of Hilda asteroids from \citet{Bus:2002bea}, \citet{Lazzaro:2004ja}, and the VNIR spectrum of (361) Bononia \citep{Bus:2002bea, Takir:2012cza} compared with the mean spectrum of bright Hilda asteroids (from this study).}
\label{fig:mean}
\end{figure}
\begin{figure}
\centering
\includegraphics[trim=0mm 0mm 0mm 15mm, clip, angle=0,width=0.8\linewidth]{histo_trojans_2.eps}
\caption{Histogram of VNIR (0.55-1.80\,$\mu$m) spectral slopes for Trojans. Black corresponds to data from previous studies and grey to new data from this paper. The {\bf 1\,$\sigma$} bounds of the X-, T- and D- classes from \citet{DeMeo:2009gz} are shown.
}
\label{fig:slopes}
\end{figure}
\section{Discussion}
\label{sec:discussion}
Our study showed that the observed twelve high-albedo Hilda asteroids and Jupiter Trojans have the same spectral properties in the visible and near-IR ranges as the low-albedo objects. The lack of spectral differences in our sample may suggest that the high albedos were incorrectly derived by SST and WISE. This would not be surprising considering the natural broadening of the Gaussian error with the increasing number of objects at smaller sizes (for a more detailed discussion, see \citealt{Grav:2011eqa, Grav:2012ho}). The few high-albedo objects would therefore correspond to the outlier measurements for which the statistical error pulled towards higher signal.
Alternatively, if the albedo measurements are valid, we propose as a first suggestion that they result from a collision-induced resurfacing process that exposed bright material at the surface. Because an origin in the primordial transneptunian disk is predicted for these objects \citep{Morbidelli:2005dr, Levison:2009jta}, this bright material could include water ice. However, our spectra do not show any evidence for this material at the surface of these objects. More studies in the 3\,$\mu$m region may shed light on the nature of this putative bright material.
The lack of spectral differences in our sample allows setting an upper limit on the fraction of potential interlopers within Jupiter Trojan and Hilda populations (i.e., inner solar system small bodies such as S-type asteroids) and thus constraints on migration models. If the high albedos result from statistical measurement errors, both our study and the WISE survey indicate that there are no interlopers among Trojans and Hilda asteroids. Conversely, if the high albedos are valid, we find that among the fifty bright Jupiter Trojans and Hilda asteroids detected by SST and WISE, and considering that twelve of them are X-, T- or D-types, at most $\sim$\,22\% of them can be interlopers at the 99.7\% confidence level (this result was derived from the hypergeometric probability distribution). Therefore, at most $\sim$0.4\% of the $\sim$2900 objects surveyed by SST and WISE may be interlopers at the 99.7\% confidence level. The paucity or absence of interlopers among Jupiter Trojans and Hilda asteroids revealed by our study is consistent with the results of \citet{DeMeo:2013hw}, who found no evidence of S-types or other unexpected classes (e.g., A-, V-) among them based on the SDSS survey.
\section{Conclusions}
\label{sec:conclusions}
We have measured the visible and\,/\,or near-IR spectral properties for twelve high-albedo (${\rm p}_{\rm v}$\,$\geq$\,0.15) Hilda asteroids and Jupiter Trojans that are part of the fifty high-albedo objects detected by SST and WISE out of a total of 2875 objects surveyed in both populations. These twelve objects have the same spectral properties as their low-albedo counterparts, which suggests a similar origin for low- and high-albedo Hilda asteroids and Jupiter Trojans. The lack of interlopers in our sample allowed us to place an upper limit of 0.4\% at a confidence level of 99.7\% on the fraction of interlopers (e.g., A-, S-, V-type asteroids) among these two populations. This result is based on the assumption that albedos were accurately determined by SST and WISE. If this were not the case, both Hilda asteroids and Jupiter Trojans probably comprise no interlopers. Future work should attempt to determine the cause of the high albedos.\\
\begin{acknowledgements}
We are grateful to Josh Emery for providing his observations of 66 Trojans and to Driss Takir for providing his observations of (153) Hilda, (190) Ismene, and (361) Bononia. We also thank the referee, Tommy Grav, for his very valuable comments, and Francesca DeMeo for helping us clarify some ambiguities. Finally, we acknowledge the ESO VLT and NASA IRTF staff for their assistance with asteroid observations.
\end{acknowledgements}
|
\section{Introduction and summary of results}
In recent years great attention has been devoted to the properties of the entanglement and R\'enyi entropies of quantum systems in their ground states. In a system with a local Hamiltonian, the Hilbert space is split between the degrees of freedom of a spatial region $V$ and its complement $\overline{V}$, and the entanglement entropy (EE) is defined as the von Neumann entropy of the reduced density matrix of one of the subsystems:
\es{EEDef}{
S_E = -{\Tr}_V \rho_V \log \rho_V\,, \qquad \rho_V = {\Tr}_{\overline{V}} \ket{\mathrm{gd}}\bra{\mathrm{gd}}\,.
}
The R\'enyi entropy is a generalization of EE defined by
\es{Renyi}{
S_q=-{1\ov q-1}\, \log \Tr \rho_V^q \ .
}
For a normalized density matrix $\Tr \rho_V = 1$, the EE \eqref{EEDef} is obtained in the limit $q\to1$.
By now several properties of these quantities have been uncovered for various classes of systems, and this has led to substantial progress in disparate fields, from numerical methods to the classification of phases of matter. See~\cite{Ryu:2006ef,Amico:2007ag,Calabrese:2009qy,Casini:2009sr} for reviews from different viewpoints.
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{wiggly_region.pdf}
\end{center}
\caption{\label{fig:wiggly_region}Perturbed circle (\ref{garesult}) as entangling region $V$, $\Sigma = \partial V$. The correction to the universal coefficient $s_{3}$ is of order $\epsilon^2$ and it is given by \eqref{garesult} in a holographic CFT$_3$.}
\end{figure}
Here we consider the ground state EE of conformal field theories (CFTs), and we focus on their dependence on the shape of the entangling region $V$. More precisely, we consider in 3 and 4 spacetime dimensions the shape dependence of the universal coefficients $s_3$, $s_4$ that appear in the well known expansion
\es{Divergences}{
&S = \# \frac{R}{\delta} - s_3 + \mathcal O\left(\frac{\delta}{R}\right)\quad \text{for $d = 3$}\,, \\
&S = \# \frac{R^2}{\delta^2} - s_4 \log \frac{R}{\delta} + \mathcal O\left(\frac{\delta^2}{R^2}\right) \quad \text{for $d = 4$}\,,
}
where $\delta$ is a short distance regulator, $R$ is the linear size of the entangling region $V$, and $\#$ stands for non-universal coefficients.
A similar formula with $q$-dependent coefficients applies to R\'enyi entropies.
While in even dimensions the universal coefficient multiplies a logarithmic divergence, and hence its shape dependence is given by a local functional of the geometric invariants of $\Sigma = \partial V$,\footnote{See~(\ref{genes}) for the case of EE in $d=4$ ~\cite{Solodukhin:2008dh}.} in odd dimensions it is a fully non-local functional of the entangling region, whose computation is rather challenging.
Here we show that, for a general CFT in any dimension, the variation of the universal term for a perturbed sphere $r(\Omega) = R_0(1 + \epsilon f(\Omega))$ is second order in the perturbation:
\begin{equation}\label{eq:second_order}
s_d = s_d^{(0)} + \epsilon^2 s_d^{(2)} + \mathcal{O}(\epsilon^3)\,.
\end{equation}
Based on ideas of the work~\cite{aitor}, we generalize these results for R\'enyi entropies.
We obtain this second order variation for the EE of a perturbed circle (Fig.~\ref{fig:wiggly_region}) in a $d = 3$ CFT with a gravity dual:
\es{garesult}{
&r(\theta) = R_0\left[1 + \epsilon \sum_{n} \left(a_n \cos n\theta +b_n \sin n\theta \right)\right]\,, \\
&\tilde s_3=2\pi\left[1+\ep^2\,\sum_{n} {n(n^2-1)\ov4} \le(a_n^2+b_n^2\ri)\right]\,,
}
where $\tilde s_d \equiv (4G_N / L^{d-1}) s_d$, with $L$ the radius of the dual AdS$_{d+1}$ and $G_N$ Newton's constant.
We also consider an ellipse of semi-axes $a$, $b$ as entangling region, still in a $d = 3$ CFT with a gravity dual. Fig.~\ref{fig:ellipse} displays various analytic and numerical lower bounds on $\tilde s_3$, as a function of the aspect ratio $b/a$. We find that $\tilde s_3$ smoothly interpolates between the value for a circle and the value for an infintely long strip, as the aspect ratio goes to zero. In particular we have:
\begin{align}\label{eq:ellipse_bounds}
&\tilde s_3 \geq 2 \pi\,, && \frac{b}{a}\tilde s_3 \geq \frac{2\pi^2\Gamma\left(\frac34\right)^2}{\Gamma\left(\frac14\right)^2} \equiv \frac{\pi}{2} \tilde s_3^{\mathrm{(strip)}}\,,
\end{align}
where the first bound is saturated at $b/a = 1$ and the second at $b/a = 0$. The very tight lower bound shown in Fig.~\ref{fig:ellipse} with a blue solid line is obtained numerically. It is basically saturated for $b/a \gtrsim 0.1$. From~\eqref{garesult} we can also determine the approach to $b/a = 1$:
\es{gaEllipse}{
\tilde s_3 \sim 2\pi\left[1 + {3\ov8}\left(1 - \frac{b}{a}\right)^2\right] \quad \text{for $b/a \to 1$} .
}
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{ellipse_bounds.pdf}
\end{center}
\caption{\label{fig:ellipse} Universal coefficient $\tilde s_3$ for an elliptic entangling region with semi-axes $a$, $b$, in a $d = 3$ holographic CFT. The blue, solid curve is a tight lower bound obtained numerically. The red, dashed curve $\tilde s_3=2 \pi$ is a lower bound set by the area of an ellipsoid (\ref{eq:ellipsoid_trial}). The yellow, dotted curve is a lower bound set by the area of a deformed strip \eqref{strip}. The green, dash-dot curve $\tilde s_3 = 2 \pi \left[1 + {3\ov8}(1 - b/a)^2\right]$ is an approximation obtained by considering perturbations of a circle~\eqref{gaEllipse}. It is not a bound.}
\end{figure}
From~\eqref{eq:second_order} it clear that the sphere is a stationary point for the universal term in EE among all shapes. From~\eqref{garesult} we conclude that in holographic theories in $d=3$ it is a local minimum, while the numerical results for an ellipse (see Fig.~\ref{fig:ellipse}) hint at it being a global minimum.
In $d=4$ the sphere is a global minimum in the universal term for EE~\cite{Astaneh:2014uba}. Let us repeat the analysis here. Solodukhin's formula~\eqref{genes} determines the universal piece for all CFTs~\cite{Ryu:2006ef,Solodukhin:2008dh}:
\es{genes}{
s_4 &= { a_4\ov 180} \int_\Sig d^2 \sig \sqrt{\ga} \, E_2 + {c_4\ov 240\pi} \int_\Sig d^2 \sig \sqrt{\ga} \, I_2\,,\\
I_2 &= K_{ab} K^{ab} - \ha K^2 \,,
}
where $a_4$ and $c_4$ are coefficients of the trace anomaly,\footnote{We normalize $a_4$ and $c_4$ so that they both equal one for a real scalar field.} $E_2$ is the Euler density normalized such that $\int_{S^2} d^2 \sig \sqrt{\ga} \, E_2=2$, $\ga$ is the induced metric, and $K$ the extrinsic curvature. Because the first term in~\eqref{genes} is topological, shapes continuously connected to $S^2$ give the same contribution. It is easy to see that $I_2$ is nonnegative and vanishes only for the sphere. Thus, we showed that the sphere minimizes the universal term in EE.
The above evidence lead us to conjecture that, in a CFT, {\it the sphere minimizes the universal contribution to EE in all dimensions} among shapes continuously connected to it.\footnote{We thank Hong Liu for crucial discussions on this topic and Eric Perlmutter for discussions on the topology of $\Sig$.} It is then natural to use the EE across a sphere as a c-function~\cite{Casini:2004bw,Myers:2010tj,Komargodski:2011vj,Casini:2012ei}.
It would be nice to provide more checks for this conjecture; one could investigate e.g.~higher dimensional cases, second order perturbations around in the CFT setup, and whether the sphere still minimizes the universal term to EE away from a CFT fixed point.\footnote{See~\cite{Klebanov:2012yf} for the shape dependence in gapped theories.} We believe these are fascinating topics to explore.
The rest of this paper presents a derivation of these results, organized as follows. In section \ref{sec:perturbed_sphere} we derive (\ref{eq:second_order}) using CFT techniques. In section \ref{sec:perturbed_circle} we derive (\ref{garesult}) and an analogous result for $d = 4$. Section \ref{sec:ellipse} derives the analytic bounds (\ref{eq:ellipse_bounds}) for the elliptic entangling region in a holographic CFT$_3$. Section \ref{sec:generic_surface} describes how to establish tight numerical bounds on $s_3$ for a generic entangling region in a holographic CFT$_3$, and in particular the numerical bound in Fig.~\ref{fig:ellipse}.
{\bf Note added:} During the finalization of this paper, we learned about the work~\cite{aitor}, which will appear soon.
\section{Perturbed sphere in a generic CFT}
\label{sec:perturbed_sphere}
\subsection{First order corrections to entanglement entropy}
In this section, using general conformal field theory arguments we investigate EE of a deformed sphere. The parameter of the deformation will be denoted by $\ep$. We show that the contribution to the universal term in the entropy linear in $\ep$ vanishes.
In polar coordinates
\begin{equation}
\dd s^2= -\dd t^2+\dd r^2+r^2\dd \Omega^2\,,
\end{equation}
we take the entangling surface $\Sigma$ to be $r=R \le[1+\epsilon f(\Omega)\ri]$. By changing coordinates we can think about the family of these surfaces as being spheres ($r'=R$), and the field theory living in curved space~\cite{Banerjee:2011mg}:
\begin{align}\label{NewCoord1}
\begin{split}
\dd s^2& =-\dd t^2+\le(\le[1+\epsilon f(\Omega)\ri]\dd r'+ \ep r' \p_\Omega f(\Omega)\, \dd \Omega \ri)^2\\
&\phantom{=}\ +r'^2\le[1+\epsilon f(\Omega)\ri]^2\dd \Omega^2\\
&=-\dd t^2+\le(\delta_{ij} +\ep h_{ij}+\sO(\epsilon^2)\ri) \dd x^i \dd x^j \,,
\end{split}
\end{align}
with
\es{hExplicit}{
h_{ij}\dd x^i \dd x^j &= 2 \le(f \dd r'^2 + r' \p_\Omega f \, \dd r' \dd \Omega+ r'^2 f \dd \Omega^2\ri) \,,
}
where $i,j$ indices run over spatial directions, while $\mu,\nu$ will run over all spacetime directions. Soon we will introduce a mapping to $\sH=S^1\times \HH^{d-1}$; there we will use $\al,\beta$ as spacetime indices.
From now on we drop the prime from $r'$. An important thing to note is that $h$ is pure gauge:
\es{PureGauge}{
h_{\mu\nu}=2\nabla_{(\mu}\, \xi_{\nu)} \qquad \xi_\nu=(0, r f(\Omega), 0, \dots, 0) \,,
}
where $\nabla_{\mu}$ is the covariant derivative in polar coordinates.
The reduced density matrix in curved space differs from the flat space one. To linear order in the perturbations:
\es{PerturbedEE}{
\rho&=\rho_0+\ep\delta \rho\\
\delta S_E&=-{\ep}\, {\Tr \le(\delta \rho\, \log \rho_0\ri)} \,,
}
where the subscript $E$ stands for entanglement. To arrive at this formula we used the cyclicity of the trace and the normalization condition $\Tr\rho_0=1$.
The reduced density matrix for a spherical entangling surface is
given by~\cite{Casini:2011kv}\footnote{An explicit expression for the reduced density matrix is only know in the case of planar and spherical entangling surface. These are the known cases, where there the entanglement Hamiltonian $K_0$ generates a symmetry around $\Sig$.}
\es{EntanglementH}{
\rho_0&=\,\frac{e^{-K_0}}{\Tr e^{-K_0}}\,,\\
K_0&=2\pi \int_{r<R}d\vec{x}\, {R^2-r^2\ov 2R} T^{00}(\vec{x}) \ .
}
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{MBdy_alt}
\end{center}
\caption{\label{fig:MBdy} Geometry of the manifold $\sM$ of~\eqref{deltaRho}. Lines of constant $(u,\Omega)$ are drawn in purple. The entangling surface $\Sig$ is marked by a blue line, and sits at $u=\infty$. We use a regularization procedure with a cutoff $\delta$ that cuts out a tube centered around $\Sig$ from $\sM$. $\p\sM$ is at constant $u=u_m$, and it has topology $S^1\times S^{d-2}$. It is drawn in yellow. When we map to hyperbolic space the Hamiltonian generates time evolution along the purple lines. $\p\sM$ maps to the boundary of hyperbolic space.}
\end{figure}
Recently,~\cite{Rosenhaus:2014woa} gave an elegant formula for $\de \rho$:
\es{deltaRho}{
\rho_0^{-1}\, \delta \rho&={1 \ov 2 } \int_\sM h_{\mu\nu} \, \le(T_\sI^{\mu\nu}-\Tr \le(T^{\mu\nu} \rho_0\ri) \ri) \,, \\
T_\sI^{\mu\nu}(v,u,\Om)&\equiv\rho_0^{v/2\pi} \, T^{\mu\nu}(v,u,\Om) \, \rho_0^{-v/2\pi} \,,
}
where $\sM$ is (Euclidean) $\R^d$, with a tube of size $\delta$ around the entangling surface $\Sigma$ cut out (see fig.~\ref{fig:MBdy}). The removal of this tube serves as a short distance regulator. We like to think of $T_\sI^{\al\beta}(v,u,\Om)$ as the analog of an operator in the interaction picture in weak coupling perturbation theory. Because instead of a time-ordered exponential $\rho_0=e^{-K_0}/\Tr e^{-K_0}$, its fractional powers indeed generate the appropriate time evolution.
Plugging into~\eqref{PerturbedEE} and using the cyclicity of the trace we get:
\es{PerturbedEE2}{
\delta S_E&={\ep\ov 2}\, \int_\sM h_{\mu\nu}\le[\Tr \le(T^{\mu\nu} K_0\, \rho_0\ri)-\Tr \le(T^{\mu\nu}\rho_0\ri)\Tr\le(K_0\,\rho_0\ri)\ri] \,.
}
To lighten the notation we introduce the ``connected'' trace:
\es{ConnectedTrace}{
\Tr \le(T^{\mu\nu} K_0\, \rho_0\ri)_c\equiv\Tr \le(T^{\mu\nu} K_0\, \rho_0\ri)-\Tr \le(T^{\mu\nu}\rho_0\ri)\Tr\le(K_0\,\rho_0\ri) \,.
}
\eqref{PerturbedEE2} can then be written as
\es{PerturbedRenyiLimit}{
\delta S_E&={\ep \ov 2}\, \int_\sM h_{\mu\nu}\,\Tr \le(T^{\mu\nu} K_0\, \rho_0\ri)_c \,.
}
where we used $\Tr\rho_0=1$.\footnote{In $\sM$ the ''connectedness'' of the correlator does not matter, as one point functions in the ground state vanish, hence $\Tr \le(T^{\mu\nu}\, \rho_0\ri)=0$. Nevertheless we kept the disconnected terms to facilitate the transformation to $\sH$, where in even $d$ the transformation rule of these terms cancel the anomalous term coming from the transformation rule of the stress tensor~\cite{Hung:2011nu,Perlmutter:2013gua,Rosenhaus:2014woa}.}
Now we want to make use of the fact that $h$ is pure gauge. We partially integrate to get:
\begin{equation}\label{PartInt}
\begin{split}
\int_\sM \nabla_{\mu}\, \xi_{\nu}\, \Tr\le(T^{\mu\nu} K_0\, \rho_0\ri)&=\int_{\p\sM} n_{\mu}\xi_{\nu}\, \Tr\le(T^{\mu\nu} K_0\, \rho_0\ri)\\
&-\int_\sM \xi_{\nu}\, \Tr\le(\le(\nabla_{\mu}\,T^{\mu\nu}\ri)K_0\, \rho_0\ri) \,,
\end{split}
\end{equation}
where $\p \sM$ is shown on Fig.~\ref{fig:MBdy}.
The Ward identity from the conservation of the energy momentum tensor is:
\es{Ward}{
\Tr\le(\le(\nabla_{\mu}\,T^{\mu\nu}\ri)K_0\, \rho_0\ri)&=0\,,
}
hence only the boundary term remains on the right-hand side of~\eqref{PartInt}.\footnote{We have to show that the counter term contributions vanish. From~\eqref{EntanglementH} we see that $K_0$ is an integral of $T^{00}$, and the one point function of $T^{00}$ vanishes by conformal invariance, and~\eqref{Ward} follows.}
Finally, we are left with
\es{PT4}{
\delta S_E&=\ep\, \int_{\p\sM} n_{\mu}\xi_{\nu}\,\Tr \le(T^{\mu\nu}\, \rho_0^{q}\ri)_c \,.
}
The non-universal contributions to the entropy come from the entanglement of degrees of freedom on the cutoff scale $\delta$. According to~\eqref{PT4} the only contribution to $\delta S_E$ comes from the cutoff-size region $\p\sM$, hence we already anticipate that $\delta S_E$ will not contain a universal piece. In the following, we confirm this intuition by explicit calculation.
We make a cautionary remark about boundary contributions here. We have not been careful about imposing boundary conditions on $\p\sM$. In the context to mapping to hyperbolic space, it is known that these boundary terms contribute to the thermal entropy~\cite{Casini:2011kv}. See~\cite{Lewkowycz:2013laa} for additional discussions. We leave the analysis of this subtlety to future work.
\subsection{Calculation in hyperbolic space}
The calculation of $\delta S_E$ is most easily done by going to hyperbolic space, $\sH=S^1\times \HH^{d-1}$. We work in Euclidean signature, and set the radius of hyperbolic space, or equivalently of the entangling surface to 1.
The change of coordinates
\es{MapToH}{
\tau&={\sin (v)\ov \cosh u + \cos(v)}\,,\\
r&={\sinh u \ov \cosh u + \cos(v)}\,,
}
leads to the unperturbed metric
\es{ConfEqMetric}{
ds^2_0&=\omega^2\le[dv^2+du^2+\sinh^2u \,d\Omega_{d-2}^2\ri]\\
\omega&={1\ov \cosh u + \cos(v)} \,.
}
The range of the coordinates are $u\in[0,\infty),\, v\in[0,2\pi)$.
We can get rid of the conformal factor $\omega^2$ by a Weyl scaling, and the remaining line element is $\sH$. Actually, there is a conformal transformation relating the operators on $\sM$ to those on $\sH$, and implementing this transformation on the entanglement Hamiltonian $K_0$ we obtain the Hamitonian on $\HH^{d-1}$, $2\pi H$~\cite{Casini:2011kv}. Going through these steps we obtain~\eqref{PerturbedEE2} in $\sH$~\cite{Rosenhaus:2014woa}:
\es{PT5}{
\delta S_E&={\ep\ov 2}\, \int_{\p\sH} \omega^{-2}\, h_{\al\beta}\, \Tr \le[T^{\al\beta}\le(2\pi H\, e^{-2\pi q H}\ri)\ri]_c \,.
}
where we used the conformal transformation rule of the stress tensor.\footnote{We emphasize that $h_{\al\beta}$ is what we get by the coordinate transformation~\eqref{MapToH} and does not change under Weyl scaling. Alternatively, we could also use that $\delta S_E$ should be given by a Weyl invariant expression, and under a Weyl scaling the metric deformation also changes. In the latter way of thinking the $\om^{-2}$ factor comes from the transformation of $h$.}
We can make the next step in two equivalent ways; we can either perform a conformal transformation on~\eqref{PT4} or we can use that $h$ is pure gauge,\footnote{Note that~\eqref{PureGauge} only holds in flat space, in the conformally related $\sH$ there are additional terms (due to the change of the covariant derivative under Weyl scalings). They conspire to yield and integrand which is again a total divergence.} and integrate partially to obtain
\es{PT6}{
\delta S_E&=\ep\, \int_{\p\sH} \omega^{-2}\, n_{\al}\xi_{\beta}\, \Tr \le[T^{\al\beta}\le(2\pi H\, e^{-2\pi q H}\ri)\ri]_c\, ,
}
where $\p \sH$ is $S^1\times S^{d-2}$ at constant radial coordinate $u=u_m$.
The integrals can now be evaluated. From the tracelessness and conservation of the stress tensor it follows that $\Tr \le[T^{\al}_{\,\,\beta}\le(2\pi H\, e^{-2\pi q H}\ri)\ri]_c$ is position independent, and the non-zero elements are~\cite{Rosenhaus:2014woa,Perlmutter:2013gua}:
\es{TTCorrelator2}{
\Tr \le[T^{v}_{\,\,v}\le(2\pi H\, e^{-2\pi H}\ri)\ri]&=-{(d-1)\om_{d+2}\ov 2^{d+1} \pi \, d}\, C_T\\
\Tr \le[T^{I}_{\,\,J}\le(2\pi H\, e^{-2\pi H}\ri)\ri]&={\om_{d+2}\ov 2^{d+1} \pi \, d}\, C_T \, \delta^I_J \,,
}
where $I,\, J$ run over $\HH^{d-1}$ and $C_T$ is the coefficient of the stress tensor two-point function:
\es{TT2Point}{
\<\, T_{\mu\nu}(x)\, T_{\rho\lambda}(0)\>_{\R^d}&={C_T \ov x^{2d}}\,\le[\frac12 \le(I_{\mu\rho} I_{\nu\lambda}+I_{\mu\lam} I_{\nu\rho}\ri)-{\delta_{\mu\nu} \delta_{\rho\lambda} \ov d}\ri]\,, \\
I_{\mu\nu}&\equiv \delta_{\mu\nu} -2 {x_\mu x_\nu\ov x^2} \,,
}
and
\es{omd}{
\om_d={2\pi^{(d+1)/2}\ov\Gamma\le(d+1\ov2\ri)}
}
is the volume of $S^d$.
In $\sH$ coordinates $n=(0,1,0,\dots)$ and $\xi=f\, \omega^{3}\sinh u$ $\le(\sin v \sinh u,\, 1+\cos v \cosh u,\, 0, \dots\ri)$. Plugging these formulae in~\eqref{PT6}, and not forgetting about the volume element on $\p \sH$ that we omitted in the above formulae, we get:
\begin{equation}\label{BdyTerms2}
\begin{split}
\delta S_E=&{\ep\, \om_{d+2}\ov 2^{d+1} \pi\, d}\, C_T\,\, \le[\int_{S^{d-2}} d\Omega_{d-2}\ f \ri]\\\times &\int_{S^1}dv \ \om(v,u_m)\, (1+\cos v \cosh u_m) \, \sinh^{d-1}u_m \\
=&{\ep\, \om_{d+2}\ov 2^{d+1}\, d}\, C_T\, \le(2 e^{-u_m}\sinh^{d-1}u_m \ri) \le[\int_{S^{d-2}} d\Omega_{d-2}\ f \ri] \,.
\end{split}
\end{equation}
It can be checked that the original expression~\eqref{PT5} evaluates to the same answer, if we plug in the explicit form of $h$.
\subsection{Generalization to R\'enyi entropies}
The above discussion can be generalized to the case of R\'enyi entropies. The upcoming paper~\cite{aitor} develops perturbation theory for R\'enyi entropies well beyond what we consider here, and its authors suggested to us to generalize the EE results to R\'enyi entropies. The formulae below have some overlap with~\cite{aitor}, but were obtained independently.
The change in the reduced density matrix~\eqref{PerturbedEE} induces a change in the R\'enyi entropies. To linear order we get:
\es{PerturbedRenyi}{
\delta S_q&=-{\ep q\ov q-1}\, {\Tr \le(\delta \rho\, \rho_0^{q-1}\ri)\ov \Tr\rho_0^q} \ .
}
The $q\to1$ limit of this formula gives~\eqref{PerturbedEE}. As we have a formula for the operator $\delta\rho$~\eqref{deltaRho}, we can follow the same steps as in previous subsections to arrive at
\es{PT7}{
\delta S_q&=-{\ep q\ov (q-1)\Tr\rho_0^q}\, \int_{\p\sH} \omega^{-2}\, n_{\al}\xi_{\beta}\, \Tr \le(T^{\al\beta}\, e^{-2\pi q H}\ri)_c\,.
}
The traces that we need can be argued to have the same form as~\eqref{TTCorrelator2}, except that the overall constant is not known:
\es{TTCorrelator3}{
\Tr \le(T^{v}_{\,\,v}\, e^{-2\pi q H}\ri)_c&=-(d-1)\al_q\,,\\
\Tr \le(T^I_{\,\,J}\,e^{-2\pi q H}\ri)_c&=\al_q\, \delta^I_J \,,
}
where $\al_q$ is a $q$-dependent undetermined constant. Finally, we obtain
\begin{equation}\label{BdyTermsRenyi}
\begin{split}
\delta S_q=&-{\ep q\, \al_q \pi \ov (q-1)\Tr\rho_0^q}\, \le(2 e^{-u_m}\sinh^{d-1}u_m \ri) \\
\times& \le[\int_{S^{d-2}} d\Omega_{d-2}\ f \ri] \,.
\end{split}
\end{equation}
It can be checked that the $q\to1$ limit of these expressions gives back the EE results.
\subsection{Analysis of the results}
Let us analyze the result. $\int_{S^{d-2}} d\Omega_{d-2}\ f $ picks out the constant piece from the spherical harmonic decomposition of $f(\Om)$, which is just a change of radius $R$. Changing the radius does not result in the change of the universal piece in a CFT, however it changes the divergent pieces.\footnote{See~\eqref{Divergences} for the divergence structure and the universal pieces in $d=3,4$. } We conclude that $\delta S_q$ {\it does not contain a universal piece}.
To see this more explicitly from~\eqref{BdyTerms2}, we have to express $u_m$ in terms of the field theory cutoff $\delta$. We do not know what the exact relation between these two quantities is, only
its leading behavior
\es{CutoffRelation}{
{\delta\ov 2R}=e^{-u_m}+\dots \ ,
}
where we have reintroduced the radius of $\mathbb{H}^{d-1}$, or equivalently the size of the entangling region.
The relation~\eqref{CutoffRelation} can be motivated from the from the coordinate transformation~\eqref{MapToH} by setting $\tau=0$, going to $\delta$ distance to the entangling region at $r=R-\delta$, and relating this to the boundary of $\mathbb{H}^{d-1}$, See Fig.~\ref{fig:MBdy}. However, we should not take this argument literally, as it would give the relation~\cite{Casini:2011kv}
\es{AltCutoffRelation}{
{\delta\ov R}=1-{\sinh u_m\ov \cosh u_m+1} \ .
}
Although this expression gives the same leading behavior as~\eqref{CutoffRelation}, it contains all both even and odd powers of $e^{-u_m}$. However, even powers would result in the change of universal terms through~\eqref{BdyTerms2}, hence they are not allowed.\footnote{Even powers of $e^{-u_m}$ in~\eqref{CutoffRelation} would also invalidate the results of~\cite{Klebanov:2011uf}.}
We conclude that~\eqref{CutoffRelation} can only involve odd powers of $e^{-u_m}$, so going through the explicit analysis of~\eqref{BdyTerms2} we learned something about~\eqref{CutoffRelation}.
In the regularization scheme we defined by cutting out a tube around $\Sig$, we obtain
\es{AreaCorrection}{
\delta S\propto&\,{\ep}f_0\, \le[\le(R\ov \delta\ri)^{d-2}+\#\le(R\ov \delta\ri)^{d-4}+\dots\ri] \,,
}
where there the powers of $R/\delta$ decrease in steps of 2, and hence no universal terms occur. We introduced $f_0={1\ov \om_{d-2}}\,\int_{S^{d-2}} d\Omega_{d-2}\ f $ for the constant mode of $f$.
In the EE case, we determined the prefactor as well:
\es{AreaCorrectionEE}{
\delta S_E=&{\ep}f_0\,{\om_{d+2}\om_{d-2}\ov 2^{d+1} \, d}\, C_T\, \le[\le(R\ov \delta\ri)^{d-2}+\#\le(R\ov \delta\ri)^{d-4}+\dots\ri] \,,
}
Our calculation then makes it possible to determine the coefficient of the area law term as follows. We obtained that changing the radius $R\to R(1+\ep f_0)$ introduces a change in the EE~\eqref{AreaCorrectionEE}.
We can then reconstruct the coefficient of the area law term
\es{AreaLaw}{
S&_E={\om_{d+2}\om_{d-2}\ov 2^{d+1} \, d(d-2)}\, C_T\, \le(R\ov \delta\ri)^{d-2}+\dots \ .
}
Of course, this result only applies in the particular regularization scheme that we used in this calculation.
We repeat that we have not been careful about boundary conditions in this calculation. They could potentially give additional contributions to the result~\eqref{AreaLaw}.
In summary, we found that to linear order in the deformation parameter $\ep$ there is no change in the universal term in the entropies. The only change is in the divergent terms, and they are all proportional to the spherical average of the deformation. In particular the entropies do not change, if this average vanishes. In the next section we calculate the $\sO(\ep^2)$ pieces for EE in the holographic setup.
\section{Perturbed circle in a holographic CFT}
\label{sec:perturbed_circle}
According to the Ryu-Takayanagi formula~\cite{Ryu:2006bv,Lewkowycz:2013nqa}, in a QFT with a gravity dual, the EE of a region $\Sigma$ is proportional to the area of the minimal surface in the dual geometry that is homologous to $\Sigma$. For a CFT$_3$ in its ground state, the dual geometry is AdS$_4$:
\begin{equation}\label{eq:AdS_coordinates}
g = \frac{L^2}{z^2} \left(-\dd t^2 + \dd z^2 + \dd r^2 + r^2 \dd \theta^2\right)\,.
\end{equation}
We take the entangling region to be a perturbed circle $r = R(\theta)$ with
\begin{align}\label{eq:perturbed_circle1}
R(\theta)=R_0+\ep \, A \, \cos\le(n\theta\ri)\,,
\end{align}
and parametrize the surface inside AdS as
\begin{align}
r = R(\theta, z)\,, && R(\theta, 0) = R(\theta)\,.
\end{align}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.5]{sample_surface1_cut.jpg}
\end{center}
\caption{\label{fig:perturbed_circle_surface} Actual minimal surface for a perturbed circle (\ref{eq:perturbed_circle1}) with $n = 5$ and $A/R_0 = 0.12$, obtained numerically.}
\end{figure}
We organize the perturbation theory in $\ep$ so that the tip of the minimal surface is at $z=1$ to all orders. Then both $R_0$ and $A$ are non-trivial series in $\ep$, which we evaluate to the order needed to obtain the leading correction to $s_3$:
\begin{align}
R_0=1+{\ep^2\ov 4}+\mathcal{O}(\epsilon^4)\,, &&
A=1+\mathcal{O}(\epsilon) \,.
\end{align}
The minimal surface to $O(\ep^2)$ is:
\es{}{
R(z,\theta)&=\sqrt{1-z^2}\big[1+\ep\, R_1(z)\,\cos\le(n\theta\ri)\\
&+\ep^2\,\le(R_{20}(z)+R_{22}(z)\,\cos\le(2n\theta\ri)\ri)+\dots\big]\\
R_1(z)&=\le(1-z\ov 1+z\ri)^{n/2}\, {1+nz\ov 1-z^2}\\
R_{20}(z)&=\le(1-z\ov 1+z\ri)^{n}\frac{1}{4(1-z^2)^2}\big[1 + 2 n z \\
& + (3 n^2-2) z^2 + 2n (n^2-1) z^3\big] \,.
}
(To obtain the leading result we do not need the explicit form of $R_{22}$). See~ \cite{Hubeny:2012ry} for the same solution in different coordinates. Plugging into the area functional we obtain:
\be
{4G_N\ov L^{d-1}}\, S={2\pi\le(1+\ep^2\, {n^2+1\ov4}\ri)\ov \delta}-2\pi\le(1+\ep^2\, {n(n^2-1)\ov4}\ri) \,,
\ee
where $L$ is the radius of AdS$_4$ and $G_N$ is Newton's constant.
The first term is just the area law term; the length of the entangling region appears expanded to $O(\ep^2)$. From the constant term we read off:
\es{gaResult}{
\tilde s_3&=2\pi\le(1+\ep^2\, {n(n^2-1)\ov4}\ri)\,,\\
\tilde s_3&\equiv {4G_N\ov L^{d-1}}\, s_3\,.
}
As shown in Fig.~\ref{fig:wiggly_plot}, we verified this result numerically, using the methods outlined in section \ref{sec:generic_surface}, and found excellent agreement.
We can easily convince ourselves that for a generic perturbation of the form
\es{GenericPert}{
r&=R_0\le(1+\ep f(\theta)\ri) \,,\\
f(\theta)&=\sum_{n} \le[a_n \cos(n\theta)+b_n \sin(n\theta) \ri]\,,
}
the result is a sum of contributions from different Fourier components:
\es{FourierSum}{
\tilde s_3&=2\pi\le(1+\ep^2\,\sum_{n} {n(n^2-1)\ov4} \le(a_n^2+b_n^2\ri)\ri) \,.
}
Of course at higher orders in $\ep$ different harmonics mix, and hence the result would no longer be a sum of their individual contributions.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.5]{wiggly_3d.pdf}
\end{center}
\caption{\label{fig:wiggly_plot} Universal contribution to the EE for a perturbed circle (\ref{eq:perturbed_circle1}) in a $d = 3$ holographic CFT$_3$. The blue dots are a tight lower bound obtained numerically, the blue line is the analytic result (\ref{gaResult}).}
\end{figure}
It would be very interesting to perform the CFT calculation of section \ref{sec:perturbed_sphere} to second order in $\ep^2$. It would reveal to what extent~\eqref{FourierSum} is universal. We leave this problem to future work.
For completeness, let us derive the same result for a perturbed sphere in $d = 4$. We can either carry out a perturbative holographic computation, or directly use the result~\eqref{genes}.
Considering an entangling region in the form of a perturbed sphere
\be
R(\theta)=R\le[1+\ep \, f\le(\theta\ri) \ri]\,,
\ee
we obtain
\es{4dResult}{
s_4&={a_4\ov 90}+ \ep^2 \, c_4\, g[f]\,,\\
g[f]&\equiv{1\ov 240}\int_0^\pi d\theta\ \sin\theta \le[f''\le(\theta\ri) -\cot \theta\, f'\le(\theta\ri)\ri]^2 \,.
}
If we further specialize to the case $f_n\le(\theta\ri)=\cos\le(n\theta\ri)$, then $g[f_n]$ can be calculated explicitly
\begin{equation}\label{4dResultCos}
\begin{split}
g[f_n]&={n^2\ov 480}\Bigg[\psi\le(n+\ha\ri)+\psi\le(-n+\ha\ri)\\
&+2\le(\ga+\log 4\ri)+4n^2{2n^2-5\ov 4n^2-1}\Bigg] \,,
\end{split}
\end{equation}
where $\psi(x)$ is the digamma function. We list the first few values of $g(f_n)$ in Table~\ref{tab:gValues}. As in $d=3$~\eqref{FourierSum}, we can verify that the quadratic functional~\eqref{4dResult} does not mix different harmonics. A crucial difference from~\eqref{gaResult} is that~\eqref{4dResultCos} is an even function of $n$. The reason for this is that~\eqref{genes},~\eqref{4dResult} are local functionals, while $s_3$ is expected to be a nonlocal functional of $f(\theta)$.
\renewcommand\arraystretch{2}
\renewcommand\tabcolsep{7pt}
\begin{table}[!h]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}
\hline
n & $1$ & $2$ &$3$ &$4$ &$5$& $6$\\
\hline
\hline
$g[f_n]$& $ 0$ & $32\ov45$ & $ 128\ov35$ & $ 512\ov45$ & $ 56960\ov2079$&$ 844384\ov15015$\\
\hline
\end{tabular}
\end{center}
\caption{The values of $g[f_n]$.}\label{tab:gValues}
\end{table}%
\section{Ellipse in a holographic CFT$_3$}
\label{sec:ellipse}
In this section we derive two analytic lower bounds on the universal coefficient $\tilde s_3$ for an elliptic entangling region in a holographic CFT$_3$. Similarly to the perturbed circle case, it would be interesting to also carry out the computation in a general CFT, or at least in free theories.
Still with reference to the coordinates (\ref{eq:AdS_coordinates}), the entangling region is
\begin{equation}\label{eq:ellipse_R}
r = R(\theta) \equiv \left(\frac{\cos^2\theta}{a^2}+\frac{\sin^2\theta}{b^2}\right)^{-\frac{1}{2}}\,,
\end{equation}
where $a$, $b$ are the semi-axes of the ellipse. A trial surface that satisfies the boundary conditions is the squashed hemisphere
\begin{equation}\label{eq:ellipsoid_trial}
z(r,\theta)=a \sqrt{1-\frac{r^2}{R^2(\theta)}}\,.
\end{equation}
Plugging this expression into the area functional we obtain a bound on $\tilde s_3$, which turns out not to depend on the aspect ratio $b/a$ of the ellipse
\be\label{squashed}
\tilde s_3 \geq 2\pi \,.
\ee
It is well known that the surface (\ref{eq:ellipsoid_trial}) is minimal in the case of a circle ($a = b$), and hence the bound is saturated at this point.
To tackle the opposite limit, $b/a\ll1$, of a very thin ellipse we start from the minimal surface determining the EE of an infinite strip:
\es{}{
y(z)&=-z_t\int_1^{z/z_t}du\, {u^2\ov \sqrt{1-u^4}}\\
z_t&={\Ga\le(\frac14\ri)\ov \sqrt{\pi}\, \Ga\le(\frac34\ri)}\, {l\ov 2} \,.
}
From this we construct the following trial surface satisfying the boundary conditions:
\es{eq:deformed_strip}{
y(z,x)&=-z_t(x)\int_1^{z/z_t(x)}du\, {u^2\ov \sqrt{1-u^4}}\\
&\equiv z_t(x)\, Y\le(z\ov z_t(x)\ri)\\
z_t(x)&={\Ga\le(\frac14\ri)\ov \sqrt{\pi}\, \Ga\le(\frac34\ri)}\, b {\sqrt{1-x^2 / a^2}} \,.
}
By plugging into the area functional
\be
A=2\int_{-a}^a dx\, \int_{\delta}^{z_t(x)} dz\, {1\ov z^2}\sqrt{1+(\p_x y)^2 +(\p_z y)^2}
\ee
we can obtain a bound on $\tilde s_3$. (We have to keep in mind that we only integrate for $x$'s for which $z_t(x)>\delta$. We did not display this to avoid clutter.) Let us use that the $x$ dependence only appears through $z_t(x)$; we rescale $z\equiv z_t \, Z$ to obtain:
\begin{equation}
\begin{split}
A=&2\int_{-a}^a dx\, \int_{\delta/z_t(x)}^{1} dZ\, {1\ov z_t(x) Z^2}\times\\
&\times\sqrt{{1\ov 1-Z^4}+z_t'(x)^2\le(Y(Z)-{Z^3\ov \sqrt{1-Z^4}}\ri)^2 } \end{split}
\end{equation}
By subtracting the diverging piece of the integrand and adding it back we can get an expression for $\tilde s_3$
\be
\begin{split}
A=&2\int_{-a}^a dx\, \int_{\delta/z_t(x)}^{1} dZ\,\Bigg[ {1\ov z_t(x) Z^2}\times\\
&\times\sqrt{{1\ov 1-Z^4}+z_t'(x)^2\le(Y(Z)-{Z^3\ov \sqrt{1-Z^4}}\ri)^2 }\\
&-{1+Z^2\ov Z^2}\,{ \sqrt{\pi}\, \Ga\le(\frac34\ri)\ov \Ga\le(\frac14\ri) }\,{\sqrt{a^2(a^2-x^2) + b^2 x^2}\ov b(a^2-x^2) }\Bigg]\\
&+{4aE(1-b^2/a^2)\ov \delta}-{ 8\sqrt{\pi}\, \Ga\le(\frac34\ri)\ov \Ga\le(\frac14\ri) } \,,
\end{split}
\ee
where $4aE(1-b^2/a^2)$ is the perimeter of the ellipse,\footnote{$E(x)$ is complete elliptic integral of the second kind.} and thus the divergent term reproduces the area law.
Note that the minimal subtraction $\propto 1/Z^2$ would not have cancelled all divergences. By setting $\delta\to 0$ in the integral we only introduce an $O(\delta)$ error, hence:
\begin{equation}
\label{strip}
\begin{split}
\tilde s_3\geq&{ 8\sqrt{\pi}\, \Ga\le(\frac34\ri)\ov \Ga\le(\frac14\ri) }-2\int_{-a}^a dx\, \int_{0}^{1} dZ\,\Bigg[ {1\ov z_t(x) Z^2}\times\\ &\times\sqrt{{1\ov 1-Z^4}+z_t'(x)^2\le(Y(Z)-{Z^3\ov \sqrt{1-Z^4}}\ri)^2 }\\
&-{1+Z^2\ov Z^2}\,{ \sqrt{\pi}\, \Ga\le(\frac34\ri)\ov \Ga\le(\frac14\ri) }\,{\sqrt{a^2(a^2-x^2) + b^2 x^2}\ov b(a^2-x^2) }\Bigg]\,.
\end{split}
\end{equation}
Now we can do the logarithmically divergent $x$ integral first. The coefficient of the logarithmic divergent term vanishes when integrated over $Z$. Then we are left with the finite $Z$ integral that depends on $b/a$. We do not know how to evaluate this integral. However, we can calculate the limit
\be\label{striplimit}
\frac{b}{a}\, \tilde s_3\Big\vert_{e\to 1} ={\pi\ov 2}\, \tilde s_3^{\mathrm{(strip)}} \qquad \tilde s_3^{\mathrm{(strip)}} ={4\pi\, \Ga\le(\frac34\ri)^2\ov \Ga\le(\frac14\ri)^2}
\ee
by expanding the integrand around $b/a= 0$.
Intuitively, this result comes from decomposing the surface into strips of different lengths and adding up their contributions. Because the EE of a strip is proportional to its length and inversely proportional to its width we get the answer by integrating $\tilde s_3^{\mathrm{(strip)}}{dx/ 2y(x)}$:
\be
\tilde s_3\to \tilde s_3^{\mathrm{(strip)}}\int_{-a}^{a}{dx\ov 2b\sqrt{1-x^2/a^2}}=\frac{a}{b} \,{\pi\ov 2}\tilde s_3^{\mathrm{(strip)}}\,.
\ee
Thus we provided a lower bound on $\tilde s_3$ which is saturated for $b/a\to 0$.
\section{Generic region in a holographic CFT$_3$}
\label{sec:generic_surface}
In this section we describe how to compute numerically the universal coefficient $\tilde s_3$ for a generic entangling region $\Sigma$ in a holographic CFT$_3$.
We consider a surface embedded in AdS$_4$ that is topologically equivalent to a disk. With reference to the coordinates (\ref{eq:AdS_coordinates}), we parametrize the surface as:
\begin{equation}\label{embedding}
\left\{
\begin{aligned}
&r = R(\rho, \theta)\\
&z = Z(\rho, \theta)
\end{aligned}
\right.\,,
\quad \rho \in [0, 1]\,,\quad \theta \in [0, 2\pi]\,.
\end{equation}
Smoothness of the surface constrains the functions $R$ and $Z$ to have the following small $r$ behavior:
\es{}{
&R(\rho, \theta) \sim \sum_\ell R_\ell\, \rho^{|\ell| + 1} e^{\ii \ell \theta}\,,\\
&Z(\rho, \theta) \sim \sum_\ell Z_\ell\, \rho^{|\ell|} e^{\ii \ell \theta}\,.
}
The above result is most easily seen by going to Cartesian coordinates. At $\rho = 1$ we impose the boundary condition
\begin{align}
Z(1, \theta) = 0 &&R(1, \theta) = R(\theta)\,,
\end{align}
where $R(\theta)$ defines the entangling region. For an ellipse, it is given by (\ref{eq:ellipse_R}).
The function $R$ can be chosen arbitrarily within the constraints above, and we take it to be
\es{}{
&R(\rho, \theta) \equiv \sum_\ell R_\ell\, \rho^{|\ell| + 1} e^{\ii \ell \theta}\,,\\
&R_\ell = \frac{1}{2\pi} \int_0^{2\pi} \dd \theta\ e^{-\ii \ell \theta} R(\theta)\,.
}
Expanding the equations of motion for $Z$ near $\rho = 1$ reveals that the $Z$ of the minimal surface is not analytic at $\rho = 1$. Instead, it can be written as
\begin{equation}\label{z asymptotics}
Z(\rho, \theta) = \sqrt{1 - \rho^2}\, \Xi(\rho, \theta)\,,
\end{equation}
with $\Xi$ analytic at $\rho = 1$. We represent $\Xi$ by expanding over a basis of functions:
\begin{equation}
\Xi(\rho, \theta) = \sum_{n,\ell} \Xi_{n\ell}\ e^{i \ell \theta} \rho^{|\ell|} P_n^{(0, |\ell|)}(2 \rho^2 - 1)\,,
\end{equation}
where $P_n^{(a, b)}$ are the Jacobi polynomials. This choice of basis is known to be good, because its elements have little linear dependence on each other, and span the space of analytic functions on the disk quite efficiently. We have to truncate the expansion at a finite value of $n$ and $\ell$, and, strictly speaking, our numerical result for $\tilde s_3$ will be a lower bound. However, if the minimal surface is sufficiently regular, the bound obtained will be very tight.
In order to find the minimal surface, it is necessary to evaluate efficiently and accurately both the area of a generic surface and its variation with respect to an infinitesimal change of the surface. This problem is complicated by the fact that the area is a divergent quantity because of the diverging conformal factor at the boundary of AdS. We now describe how this can be accomplished.
The embedding (\ref{embedding}) induces the following metric on the disk:
\begin{equation}
\begin{split}
g_2 &= \frac{1}{Z^2}\Big[\left(R_{,\rho}^2 + Z_{,\rho}^2\right)\dd \rho^2\\
&+2\left(R_{,\rho} R_{,\theta} + Z_{,\rho}Z_{,\theta}\right)\dd \rho \dd \theta+\left(R^2 + R_{,\theta}^2 + Z_{,\theta}^2\right)\dd\theta^2\Big]\,,
\end{split}
\end{equation}
and hence the area functional is
\begin{align}
\begin{split}
A[Z] &= \int_0^{2\pi} \dd \theta \int_0^1\dd \rho\ \sqrt{g_2}\\
& = \int \frac{\dd\theta \dd \rho}{Z^2}\sqrt{\left(Z_{,\theta} R_{,\rho} - Z_{,\rho} R_{,\theta}\right)^2 + R^2\left(R_{,\rho}^2 + Z_{,\rho}^2\right)}\,.
\end{split}
\end{align}
Varying this functional with respect to $Z$ it is possible to derive a rather involved equation of motion, which is best handled with the aid of computer algebra. By expanding it about $\rho = 1$ we obtain the asymptotic behavior (\ref{z asymptotics}) for $Z$. Because of the same asymptotics we find it convenient to write
\begin{equation}
\sqrt{g_2} = \frac{\rho}{(1-\rho^2)^{\frac{3}{2}}} a(\rho,\theta)\,,
\end{equation}
so that $a(\rho, \theta)$ is smooth on the disk and
\begin{equation}
a(1, \theta) = \frac{\sqrt{R^2(1,\theta) + R_{,\theta}^2(1,\theta)}}{\Xi(1,\theta)}\,.
\end{equation}
The area integral is divergent, and needs to be regulated. The prescription is to cut off the integral at a fixed height $Z(\rho, \theta) = \delta$:
\begin{equation}
A[Z, \delta ] = \int \dd \theta \dd \rho\ \sqrt{g_2}\ \Theta[Z(\rho,\theta) - \delta]\,,
\end{equation}
where $\Theta$ is the unit step function.
Because the divergence is associated with the boundary of the surface, it is proportional to the perimeter $P$ of the entangling region, and we have
\begin{align}\label{regulated area}
A[Z, \delta] \sim \frac{P}{\delta} - \tilde s_3[Z] + \mathcal{O}(\delta)\,.
\end{align}
The divergence does not depend on $Z$ (provided the boundary conditions are satisfied), and hence does not enter the minimization procedure. In order to compute the finite quantity $\tilde s_3[Z]$ we introduce a reference integrand $a_{\text{ref}}$, which agrees with $a$ at the boundary, but whose regulated integral can be computed analytically. The \emph{finite} difference between the integral of $a$ and the integral of $a_{\text{ref}}$ can then be computed numerically with good accuracy.
We define
\es{alDef}{
&a_\text{ref}(\rho, \theta) \equiv \sum_\ell a_\ell\, \rho^{|\ell|} e^{i \ell \theta}\,, \\
&a_\ell = \frac{1}{2\pi} \int_0^{2\pi} \dd \theta\ a(1,\theta) e^{-\ii \ell \theta}\,,
}
so that $a_\text{ref}$ is a smooth function on the disk and $a_\text{ref}(1,\theta) = a(1, \theta)$. Then we write
\begin{equation}
A[Z,\delta] = \Delta A[Z, \delta] + A_{\text{ref}}[Z, \delta]\,,
\end{equation}
with
\es{}{
& \Delta A[Z,\delta] = \int \frac{\dd \theta \dd \rho\,\rho}{(1-\rho^2)^{\frac{3}{2}}} \left(a - a_{\text{ref}}\right) \Theta[Z - \delta]\,,\\
& A_\text{ref}[Z,\delta] = \int \frac{\dd \theta \dd \rho\,\rho}{(1-\rho^2)^{\frac{3}{2}}}\,a_{\text{ref}}\,\Theta[Z - \delta]\,.
}
Then we have
\begin{equation}
\Delta A[Z,\delta] = \int \frac{\dd \theta \dd \rho\,\rho}{(1-\rho^2)^{\frac{3}{2}}} (a - a_{\text{ref}}) + \mathcal O(\delta)\,,
\end{equation}
where the integral is finite. The integral $A_{\mathrm{ref}}$ can be done analytically, and we have
\begin{equation}
\begin{split}
A_\text{ref}[Z, \delta] &= \frac{1}{\delta} \int_0^{2\pi} \dd \theta\ \sqrt{R^2 + R_{,\theta}^2} - 2 \pi a_{ 0} + \mathcal O(\delta)\, ,
\end{split}
\end{equation}
where $a_0$ is given by~\eqref{alDef}.
Collecting all the pieces we have
\begin{align}
\begin{split}
A[Z,\delta]&=\frac{1}{\delta} \int_0^{2\pi} \dd \theta\ \sqrt{R^2 + R_{,\theta}^2}+
\int \frac{\dd \theta \dd \rho\,\rho}{(1-\rho^2)^{\frac{3}{2}}}\left[a - a_{\text{ref}}\right] \\&- \int_0^{2\pi} \dd \theta\ a(1,\theta) + \mathcal{O}(\delta)=\frac{P}{\delta} - \tilde s_3[Z] + \mathcal{O}(\delta)\,,
\end{split}
\end{align}
where we recognized in
\begin{equation}
P \equiv \int_0^{2\pi} \dd \theta\ \sqrt{R^2 + R_{,\theta}^2}
\end{equation}
the perimeter of the entangling region. Using computer algebra, it is possible to compute explicitly also the variation of $A$ with respect to a change in $Z$, without forgetting that both $a$ and $a_\text{ref}$ depend on $Z$.
The integrals are conveniently evaluated using Gaussian quadrature. In particular, the double integrals have the form
\begin{equation}
I = \int_0^{2\pi} \dd \theta \int_0^1 \dd \rho\ \frac{\rho}{(1-\rho^2)^{\frac{1}{2}}} f(\rho, \theta)\,,
\end{equation}
with $f$ smooth on the disk. We evaluate them as
\begin{equation}
I = \frac{\pi}{n}\sum_{i = 1}^{m}\sum_{j=1}^{n} w_i f(\rho_i, \theta_j)\,,
\end{equation}
where $\theta_j = 2 \pi j / n$ and $w_i$, $\rho_i$ are the weights and collocation points of the Gaussian quadrature associated with the measure
\begin{equation}
\int_{-1}^1 \dd \rho\ \frac{|\rho|}{(1-\rho^2)^{\frac{1}{2}}}\,.
\end{equation}
We have now cast the problem to the numerical minimization of a real function of many variables (the coefficients $\Xi_{n\ell}$), of which we know explicitly the gradient. Therefore, we search for a minimum using the conjugate gradient method.
\section*{Acknowledgments}
We thank A.~Lewkowycz, E.~Dyer, J.~Lee, E.~Perlmutter, S.~Pufu, and V.~Rosenhaus, and especially H.~Liu for useful discussions. We thank A.~Lewkowycz and E.~Perlmutter for proposing the generalization to R\'enyi entropies and sharing~\cite{aitor} with us before publication. AA was suppported by DE-FG02-05ER41360, DE-FG03-97ER40546, by the Alfred P. Sloan Foundation, and by the Templeton Foundation. MM was supported by the U.S. Department of Energy under cooperative research agreement Contract Number DE-FG02-05ER41360\@. Simulations were done on the MIT LNS Tier 2 cluster, using the Armadillo C++ library.
\bibliographystyle{ssg}
|
\section{Introduction}
\label{sec:1}
The $B_c$ meson is a doubly heavy quark-antiquark system and carries flavors explicitly. It has been pointed out that sizable $B_c$ meson events can be produced at the hadronic colliders~\cite{bc1,bc2,bc3}. Thus, it provides a useful laboratory for studying both the Quantum Chromodynamics (QCD) and the weak interaction theories.
\begin{table}[htb]
\begin{tabular}{c|c}
\hline\hline
& ~~$\Re (J/\psi \ell^+ \nu)$~~\\
\hline
$p_{\rm T}>6{\rm GeV}$~\cite{Bc98} & $0.132_{-0.037}^{+0.041} ({\rm st}) \pm 0.031({\rm sy}) _{-0.020}^{+0.032}({\rm lf})$ \\
$p_{\rm T}>4{\rm GeV}$~\cite{Ru05} & $0.249 \pm 0.045 ({\rm st}) \pm 0.069 ({\rm sy}) _{ - 0.033}^{ + 0.082}({\rm lf})$ \\
$p_{\rm T}>6{\rm GeV}$~\cite{Ru05} & $0.245 \pm 0.045 ({\rm st}) \pm 0.066 ({\rm sy}) _{ - 0.032}^{ + 0.080}({\rm lf})$ \\
$p_{\rm T}>4{\rm GeV}$~\cite{Re05} & $0.282 \pm 0.038 ({\rm st}) \pm 0.035 ({\rm y}) \pm 0.065 ({\rm a})$ \\
$p_{\rm T}>6{\rm GeV}$~\cite{Re05} & $0.242 \pm 0.036 ({\rm st}) \pm 0.031 ({\rm y}) \pm 0.051 ({\rm a})$ \\
$p_{\rm T}>4{\rm GeV}$~\cite{Ru09} & $0.295 \pm 0.040 ({\rm st}) _{ - 0.026}^{ + 0.033} ({\rm sy}) \pm 0.036 ({\rm sp})$ \\
$p_{\rm T}>6{\rm GeV}$~\cite{Ru09} & $0.227 \pm 0.033 ({\rm st}) _{ - 0.017}^{ + 0.024} ({\rm sy}) \pm 0.014 ({\rm sp})$ \\
$p_{\rm T}>6{\rm GeV}$~\cite{Ru14} & $0.211 \pm 0.012 ({\rm st}) _{ - 0.020}^{ + 0.021} ({\rm sy})$ \\
\hline\hline
\end{tabular}
\caption{The $\Re (J/\psi \ell^+ \nu)$ measured by CDF collaboration under two $p_T$ cuts. The symbols ``st'', ``sy'', ``lf'', ``y'', ``a'' and ``sp'' stand for the statistical error, the systematic error, the $B_c$ lifetime error, the systematic error on the yield, the systematic error on acceptance ratio and the $B_c$ spectrum error, respectively. The second line is for $\Re (J/\psi e^+ \nu)+\Re (J/\psi \mu^+ \nu)$, the fifth and sixth lines are for $\Re (J/\psi e^+ \nu)$, and the remaining lines are for $\Re (J/\psi \mu^+ \nu)$.}
\label{tab:Rexpe}
\end{table}
Experimentally, the Collider Detector at Fermilab (CDF) collaboration discovered the $B_c$ meson in the year 1998 via the semi-leptonic decay channel $B_c^+ \to J/\psi \ell^+ \nu$~\cite{Bc98}. It also predicted the value of the production cross section times branching ratio fraction between the $B_c^+ \to J/\psi \ell^+ \nu$ and $B^+ \to J/\psi K^+$, i.e.
\begin{equation}
\Re(J/\psi \ell^+ \nu) = \frac{\sigma(B_c^+) BR(B_c^+ \to J/\psi \ell^+ \nu)}{\sigma(B^+) BR(B^+ \to J/\psi K^+)}. \label{ration}
\end{equation}
Later on, more measurements for the $B_c$ meson properties have been done at both the Tevatron and the LHC colliders, cf.Refs.\cite{CDFexp08, D0exp08, D0exp09, LHCbexp1,LHCbexp2,LHCbexp3,LHCbexp4,LHCbexp5,LHCbexp6, Ru05,Re05, Re06, RuRe05, Ru09, Ru14}. For our present purpose, we put the values of $\Re(J/\psi \ell^+ \nu)$ predicted by the CDF collaboration in Table \ref{tab:Rexpe}, in which two $B_c$ meson $p_T$ cuts have been adopted.
Theoretically, the predicted values for $\Re(J/\psi{\ell^+}\nu)$ are always smaller than the experimental measurements. Such a comparison has firstly been done by Ref.\cite{Bc98}, whose Fig.(3) shows that the theoretical predictions are well below the CDF prediction. As shown by Table~\ref{tab:Rexpe}, the updated Tevatron Run II measurements for $\Re(J/\psi{\ell^+}\nu)$~\cite{Ru05,Re05,Re06,RuRe05,Ru09,Ru14} are almost doubled in comparison to the previous one~\cite{Bc98}, then the discrepancy becomes even worse. This discrepancy arouses people's great interests, many attempts have been tried to solve the puzzle. It has been argued that, by including the next-to-leading order (NLO) QCD correction to the $B_c\to\eta_c(J/\psi)$ semi-leptonic decays, the prediction on $\Re(J/\psi{\ell^+}\nu)$ can be consistent with the experimental results~\cite{qiao12,qiao13}. However, in Ref.\cite{qiao12} the NLO estimation is done for a fixed $\alpha_s\simeq0.2$ and in Ref.\cite{qiao13}, a quite small $\Lambda_{\rm QCD}\simeq 0.1$ GeV has been adopted. Even worse, a large renormalization scale uncertainty and a large NLO contributions (or a large $K$ factor) make the pQCD prediction questionable.
It is noted that the pQCD series in Refs.~\cite{qiao12,qiao13} is derived under the conventional scale setting, in which the renormalization scale is set to be the typical momentum flow ($Q\sim m_b$) of the process and an arbitrary region as $[Q/2,2Q]$ is adopted for estimating the scale error. The conventional scale setting assigns an arbitrary range and an arbitrary systematic error to the fixed-order pQCD prediction~\cite{PMC3}, thus, it is natural to assume that the present puzzling situation may not be the question of the pQCD theory but the question of the conventional scale setting. As an attempt to improve the pQCD predictions, we shall apply the principle of maximum conformality (PMC)~\cite{PMC1,PMC2,PMC3,PMC4} to deal with $B_c\to\eta_c(J/\psi)$ semi-leptonic decays up to NLO level.
The PMC provides a systematic procedure to set the renormalization scale for any QCD processes. It is well-known that the running behavior of the coupling constant is governed by the renormalization group equation (RGE). Following such principal, the PMC is to use the $\beta$-terms in the perturbative series to determine the optimal behavior of coupling constant, or equivalently, to determine the optimal scale for the coupling constant~\cite{PMC4}. One can start the pQCD calculation with an arbitrary but hard enough initial scale. Then, by finding out all the related $\beta$-terms that rightly determine the running behavior of coupling constant at each perturbative order and resuming them into the coupling constant, the argument of the coupling constant at each perturbative order shall be shifted from its initial value to its optimal one. The resultant PMC expressions being free of $\beta$-terms are thus independent of the renormalization scheme, as required by the renormalization group invariance~\cite{PMC5}. In the present paper, we shall show that after applying the PMC scale setting, an improved QCD estimation for $B_c \to \eta_c(J/\psi) \ell \nu$ can be achieved, i.e. the scale uncertainty can be greatly suppressed and a more reasonable central value for the decay width of $B_c \to \eta_c(J/\psi) \ell \nu$ can be achieved.
As the key components for the $B_c \to \eta_c(J/\psi) \ell \nu$ semi-leptonic decays, the $B_c \to \eta_c(J/\psi)$ transition form factors (TFFs) have been calculated up to NLO level. It is noted that those TFFs are pQCD calculable only in large recoil region with $q^2\sim0$. Thus, one needs to extrapolate them to all their allowable physical region so as to estimate the total decay widths (or the branching ratios) of $B_c \to \eta_c(J/\psi) \ell \nu$. Several extrapolation approaches have been suggested in the literature, cf. Refs.\cite{Kiselev00,Kiselev02, Huang07,Lu09,BSWFF85,monopole}. In the present work, we shall adopt the monopole form to do the extrapolation, which has been firstly suggested in Ref.\cite{monopole}. Then, we shall reestimate the value of $\Re(J/\psi \ell^+ \nu)$ and make a comparison with the CDF predictions.
The remaining parts of the paper are organized as follows. In Sec.\ref{sec:2}, we present the calculation technology for dealing with the $B_c \to \eta_c(J/\psi)$ TFFs up to NLO level. The $B_c \to \eta_c(J/\psi)$ TFFs and their relations to the $B_c$ meson semi-leptonic decay widths are presented. The PMC treatment and the treatment of the low-energy running coupling are also presented here. In Sec.\ref{sec:3}, we present the numerical results for the TFFs at the large recoil region, and the $\Re(J/\psi{\ell^+}\nu)$ is recalculated and compared with the experimental predictions. The last section is reserved for a summary.
\section{Calculation technology}
\label{sec:2}
After integrating the phase-space, the differential decay width over $q^2$ for the semi-leptonic decay $B_c \to \eta_c \ell \nu$ or $B_c \to J/\psi \ell \nu$ can be formulated as
\begin{widetext}
\begin{eqnarray}
\frac{{\rm d}\Gamma(B_c\to\eta_c\ell\nu)}{{\rm d}q^2} &=& \bigg(\frac{q^2-m_{\ell}^2}{q^2}\bigg)^2 \frac{\sqrt{\lambda_1(q^2)} G_F^2 |V_{\rm cb}|^2}{384m_{B_c}^3\pi^3 q^2} \left[ (m_{\ell}^2+2q^2) \lambda_1(q^2) F_1^2(q^2)+3m_{\ell}^2(m_{B_c}^2-m_{\eta_c}^2)^2 F_0^2(q^2) \right], \\
\frac{{\rm d}\Gamma_{\rm L}(B_c\to J/\psi\ell\nu)}{{\rm d}q^2} &=& \bigg(\frac{q^2-m_{\ell}^2}{q^2}\bigg)^2 \frac{\sqrt{\lambda_2(q^2)} G_F^2 |V_{\rm cb}|^2} {384m_{B_c}^3\pi^3} \left[\frac{3m_{\ell}^2}{q^2} \lambda_2(q^2) A_0^2(q^2)+(m_{\ell}^2+2q^2)|h_0(q^2)|^2 \right], \\
\frac{{\rm d}\Gamma_{\rm T}(B_c\to J/\psi\ell\nu)}{{\rm d}q^2} &=& \bigg(\frac{q^2-m_{\ell}^2}{q^2}\bigg)^2 \frac{\sqrt{\lambda_2(q^2)} G_F^2 |V_{\rm cb}|^2} {384m_{B_c}^3\pi^3} \times (m_{\ell}^2+2q^2) \left[|h_+(q^2)|^2 + |h_-(q^2)|^2 \right],
\end{eqnarray}
\end{widetext}
where we have separated the total decay for the $J/\psi$ case as $\Gamma=\Gamma_{\rm L}+ \Gamma_{\rm T}$, the lepton $\ell=e,\mu,\tau$ and the Fermi constant $G_F=1.16638\times10^{-5}$~\cite{RPP}. $q=P-p$ is the momentum transfer, $P$ is momentum of $B_c$ meson and $p$ is momentum of $\eta_c$ or $J/\psi$. The phase-space factors $\lambda_1(q^2)=(m_{B_c}^2+m_{\eta_c}^2-q^2)^2- 4 m_{B_c}^2 m_{\eta_c}^2$ and $\lambda_2(q^2)=(m_{B_c}^2+m_{J/\psi}^2-q^2)^2-4m_{B_c}^2m_{J/\psi}^2$. The longitudinal and transverse helicity amplitudes for $\Gamma_{\rm L}$ and $\Gamma_{\rm T}$ are expressed as:
\begin{widetext}
\begin{eqnarray}
h_{\pm}(q^2) &=& \frac{\sqrt{\lambda_2(q^2)}}{m_{B_c}+m_{J/\psi}}\Big[V(q^2) \mp \frac{(m_{B_c}+m_{J/\psi})^2}{\sqrt{\lambda_2(q^2)}}A_1(q^2)\Big], \\
h_0(q^2) &=& \frac{1}{2m_{J/\psi}\sqrt{q^2}} \Big[-\frac{\lambda_2(q^2)}{m_{B_c}+m_{J/\psi}}A_2(q^2) +(m_{B_c}+m_{J/\psi})(m_{B_c}^2-m_{J/\psi}^2-q^2)A_1(q^2)\Big].
\end{eqnarray}
\end{widetext}
The two $B_c\to\eta_c$ TFFs $F_0(q^2)$ and $F_1(q^2)$, and the four $B_c\to J/\psi$ TFFs $V(q^2)$, $A_0(q^2)$, $A_1(q^2)$ and $A_2(q^2)$, are defined as follows~\cite{BSWFF85}:
\begin{widetext}
\begin{eqnarray}
\langle \eta_c(p) |\bar c \gamma^{\mu} b| B_c(P)\rangle &=& F_0(q^2)\frac{m_{B_c}^2-m_{\eta_c}^2}{q^2}q^{\mu} + F_1(q^2) \bigg(P^{\mu}+p^{\mu}-\frac{m_{B_c}^2-m_{\eta_c}^2}{q^2}q^{\mu} \bigg), \\
\langle J/\psi(p,\varepsilon^*) |\bar c \gamma^{\mu}b| B_c(P)\rangle
&=& \frac{2 i V(q^2)}{m_{B_c}+m_{J/\psi}} \epsilon^{\mu\nu\rho\sigma} \varepsilon_{\nu}^* p_{\rho}P_{\sigma}, \\
\langle J/\psi(p,\varepsilon^*) |\bar c \gamma^{\mu}{\gamma_5}b| B_c(P)\rangle &=& 2 m_{J/\psi}\frac{{\varepsilon^*} \cdot q}{q^2} q^{\mu} A_0(q^2) + (m_{B_c}+m_{J/\psi}) \bigg( \varepsilon^{*\mu}-\frac{\varepsilon^* \cdot q}{q^2} q^{\mu} \bigg) A_1(q^2) \nonumber\\
&& - \bigg({P^\mu} + {p^\mu} - \frac{m_{B_c}^2 - m_{J/\psi}^2}{q^2}q^{\mu}\bigg)\frac{{\varepsilon^*} \cdot q}{m_{B_c}+m_{J/\psi}} A_2(q^2).
\end{eqnarray}
\end{widetext}
For the case of $\ell=e$ or $\mu$, the lepton mass $m_\ell$ tends to zero, the contributions from $F_0$ and $A_0$ can be safely neglected due to the chiral suppression. All those TFFs are key components for the $B_c$ meson decays to charmonia. In the large recoil region, they are pQCD calculable and have been calculated up to NLO level~\cite{Bell06,Bell07,qiao12,qiao13}. Up to NLO level, we can schematically write the TFFs in the following form:
\begin{widetext}
\begin{equation}
f_i(q^2)=C^{f_i}(q^2)\alpha_s(\mu_r^{\rm init}) \left[1 + B^{f_i}(q^2,\mu_r^{\rm init}) \frac{\alpha_s(\mu_r^{\rm init}) }{\pi} \right], \label{starting}
\end{equation}
\end{widetext}
where $\mu_r^{\rm init}$ stands for some arbitrary initial renormalization scale, which should be large enough to ensure the pQCD calculation. For example, it can be chosen as the typical momentum flow of the process, i.e. $\mu_r^{\rm init}=m_b$. Under the conventional scale setting, the renormalization scale is fixed to be $\mu_r^{\rm init}$. On the other hand, for a certain scale setting, its resultant optimal scale depends on how we deal with the perturbative series and is usually different from $\mu_r^{\rm init}$. As for the PMC scale setting, its optimal scale is determined by the $\{\beta_i\}$-terms that rightly governs the running of the coupling constant via RGE. The function $f_i$ represents any one of the TFFs, $F_1(q^2)$, $F_0(q^2)$, $V(q^2)$, $A_0(q^2)$, $A_1(q^2)$ and $A_2(q^2)$, respectively. The tree-level coefficients $C^{f_i}(q^2)$ are put in the appendix and the NLO coefficients $B^{f_i}(q^2,\mu_r^{\rm init})$ can be read from Refs.\cite{Bell06,Bell07,qiao12,qiao13}. In those references, only the asymptotic expressions under the limit $m_b\to\infty$ have been given. Fortunately, however, as pointed out by Ref.\cite{qiao13}, those approximate expressions are of high precision in comparison to the full expressions~\footnote{We thank the authors of Ref.\cite{qiao13} for helpful discussions on this point and for kindly supplying us their Mathematical program for calculating the TFFs up to NLO level.}.
\subsection{The PMC scale setting for the TFFs}
\begin{figure}
\includegraphics[width=0.45\textwidth]{selfenergy.eps}
\caption{Two NLO diagrams (together with their counter-terms) that contribute to the $\beta_0$-terms for the $B_c\to \eta_c (J/\psi)$ TFFs. The `cross' symbol means the weak interaction vertex and the circle stands for the light-quark loop. } \label{Feyn}
\end{figure}
To set the PMC scales for the TFFs, $F_1(q^2)$, $F_0(q^2)$, $V(q^2)$, $A_0(q^2)$, $A_1(q^2)$ and $A_2(q^2)$, we first decompose the NLO coefficients $B^{f_i}(q^2,\mu_r^{\rm init})$ into the non-conformal $\{\beta_i\}$-dependent part and the conformal $\{\beta_i\}$-independent part, i.e.
\begin{equation}
B^{f_i}(q^2,\mu_r^{\rm init}) = B^{f_i}_{(\beta)}(q^2,\mu_r^{\rm init}) \beta_0 + B^{f_i}_{\rm conf}(q^2,\mu_r^{\rm init}),
\end{equation}
where $\beta_0=11-\frac{2}{3}{n_f}$. The conformal part $B^{f_i}_{\rm conf}(q^2,\mu_r^{\rm init})$ are in general complex and their analytic expressions at $q^2=0$ are put in the Appendix. The coefficients of non-conformal part are the same for all TFFs, i.e.
\begin{eqnarray}
B^{f_i}_{(\beta)}(q^2,\mu_r^{\rm init}) = \frac{5}{12} + \frac{1}{4} \ln \left( \frac{(\mu_r^{\rm init})^2}{2\gamma m_c^2} \right), \label{betaterm}
\end{eqnarray}
where $\gamma=\frac{m_b^2-q^2}{4 m_b m_c}$ is the relativistic boost. One may observe that at the NLO level, all the TFFs have the same $n_f$-terms and hence the same $\beta_0$-terms. As shown by Fig.(\ref{Feyn}), this is due to the same one-loop gluon self-energy diagrams (together with their counter-term ones) for all the TFFs. Such $\beta_0$-terms rightly determine the running behavior of the LO coupling constant, thus, they should be absorbed into the coupling constant following the RGE~\cite{PMC1,PMC2,PMC3,PMC4}.
More specifically, after applying the PMC scale setting, the TFFs (\ref{starting}) shall be transformed as
\begin{widetext}
\begin{eqnarray}
f_i^{\rm PMC}(q^2) = C^{f_i}(q^2) \alpha_s(\mu_r^{\rm PMC})\left[ 1 + \frac{\alpha_s(\mu_r^{\rm PMC})}{\pi}B_{\rm conf}^{f_i}(q^2,\mu_r^{\rm init}) \right], \label{startingPMC}
\end{eqnarray}
\end{widetext}
where to eliminate the non-conformal $\beta_0$-terms, the renormalization scale has been transformed from its initial value $\mu^{\rm init}_r$ to the LO PMC scale $\mu_r^{\rm PMC}$, i.e.
\begin{eqnarray}
\mu_r^{\rm PMC} &=& \mu_r^{\rm init} \exp{\left(-2B_{(\beta)}^{f_i}(q^2,\mu_r^{\rm init})\right)} .
\end{eqnarray}
It is noted that the scale displacement contains a similar function as the simplest scale displacement $e^{-{5}/{6}}$ that ensures the scheme invariance between the $\overline{\rm MS}$ scheme and the Gell-Mann-Low scheme~\cite{gml}. More explicitly, with the help of Eq.(\ref{betaterm}), the LO PMC scale can be simplified as
\begin{equation}
\mu_r^{\rm PMC} = e^{-\frac{5}{6}}\sqrt{\frac{m_c}{2m_b}(m_b^2-q^2)}, \label{pmcscale}
\end{equation}
The interesting point is that the LO PMC scale is independent of $\mu^{\rm init}_r$. As discussed above, since all the TFFs has the same $\beta_0$-term $\left[B^{f_i}_{(\beta)}(q^2,\mu_r^{\rm init}) \beta_0\right]$, all of them shall have the same PMC scale. We have no NNLO $\{\beta_i\}$-terms to set the NLO PMC scale for $\alpha^2_s$-terms, and we have implicitly set $\mu_r^{\rm PMC;NLO}=\mu_r^{\rm PMC;LO}=\mu_r^{\rm PMC}$, since $\mu_r^{\rm PMC;LO}$ is last known PMC scale. This treatment will lead to some residual scale dependence, which, however, shall be highly exponentially suppressed.
\subsection{The running coupling in the low-energy region}
Eq.(\ref{pmcscale}) indicates that, at the maximum recoil region $q^2=0$, the optimal PMC scale for the $B_c\to \eta_c (J/\psi)$ TFFs is $\mu_r^{\rm PMC}=e^{-\frac{5}{6}}\sqrt{m_b m_c /2}$. If setting $m_b=4.9{\rm GeV}$ and $m_c=1.4{\rm GeV}$, we obtain $\mu_r^{\rm PMC} \approx 0.8{\rm GeV}$, which is close to the low energy (LE) region. In the LE region, the conventional running behavior of the coupling constant may overestimate the pQCD predictions. Several LE effective models have been suggested in the literature~\cite{sjb,Webber,MPT1,MPT2,BPT,CON,GI,AnaMod}. A comparison of six typical LE coupling constant models can be found in Ref.\cite{HQdecay}. The MPT model is phenomenologically successful, i.e. the moment of the spin-dependent structure function calculated within the MPT model is consistent with the experimental data down to a few hundreds of MeV~\cite{MPT2}. For clarity, we shall adopt the MPT model to do our discussion.
The MPT model~\cite{AnaMod,MPT1,MPT2}, based on the massive analytic pQCD theory, provides a convenient way for analyzing the data below $1{\rm GeV}$. It is designed to ensure the nonsingular behavior in the infra-red (IR) region and to eliminate the Landau pole. On the basis of the mass-dependent (massive) Bogoliubov RGE~\cite{mRG}, by introducing the effective gluonic mass $m_{\rm gl} = \sqrt{\xi} \Lambda_{\rm QCD}$ as the IR regulator, $\ln \frac{\mu_r^2}{\Lambda^2_{\rm QCD}} \to \ln \frac{\mu_r^2 + m_{\rm gl}^2}{\Lambda^2_{\rm QCD}}=\ln \left( \xi+\frac{\mu_r^2}{\Lambda^2_{\rm QCD}}\right)$, one can disentangle the unwanted singularity in the IR limit from the usual ultra-violet logs. More explicitly, up to two-loop level, the MPT model suggests
\begin{widetext}
\begin{eqnarray}\label{alphasMPT}
\alpha_{s;\rm MPT}(\mu_r)=\alpha_{\rm crit}\left\{1+\alpha_{\rm crit}\frac{\beta_0}{4\pi} \ln \left(1+\frac{\mu_r^2}{\xi \Lambda^2_{\rm QCD}} \right) + {\alpha_{\rm crit}}\frac{\beta_1}{4\pi \beta_0} \ln \left[1 + \alpha_{\rm crit} \frac{\beta_0}{4\pi} \ln \left(1+\frac{\mu_r^2}
{\xi \Lambda^2_{\rm QCD}} \right) \right] \right\}^{-1},
\end{eqnarray}
\end{widetext}
where $\beta_0=11-\frac{2}{3}{n_f}$ and $\beta_1 = 102 - \frac{38}{3}{n_f}$. The critical running coupling, $\alpha_{\rm crit}=\alpha_{s;\rm MPT}(0)$, is determined via the relation, $\alpha_{\rm crit} ={4\pi}/{({\beta_0}\ln \xi)}$. If setting $\xi=10\pm2$~\cite{MPT2}, we obtain $\alpha_{\rm crit} = 0.606_{-0.044}^{+0.065}$.
\begin{table}[bt]
\centering
\begin{tabular}{c |c|c|c}
\hline
& $n_f=3$ & $n_f=4$ & $n_f=5$ \\
\hline
~Conv.~ & $0.388\pm0.007$ & $0.338\pm0.007$ & $0.233\pm0.005$ \\
\hline
MPT & $0.260\pm0.005$ & $0.235\pm0.005$ & $0.186\pm0.004$ \\
\hline
\end{tabular}
\caption{The weighted averages of $\Lambda_{\rm QCD}$ (in unit: GeV) based on the conventional and the MPT $\alpha_s$-running together with the measurements $\alpha_s(M_Z)=0.1185\pm0.0006$ and $\alpha_s(m_\tau)=0.330\pm0.014$~\cite{RPP}. `Conv.' stands for the conventional two-loop $\alpha_s$ running. } \label{Lambda}
\end{table}
Using the two-loop $\alpha_s$-running together with the measurements $\alpha_s(M_Z)=0.1185\pm0.0006$ and $\alpha_s(m_\tau)=0.330\pm0.014$~\cite{RPP}, the resultant weighted averages for the $\Lambda_{\rm QCD}$ under the conventional and the MPT two-loop $\alpha_s$-running are shown in Table~\ref{Lambda}.
\begin{figure}[hbt]
\includegraphics[width=0.45\textwidth]{alphas.eps}
\caption{A comparison of the strong running coupling $\alpha_s(\mu_r)$ up to two-loop level, where $\alpha_{s;{\rm QCD}}$ stands for the conventional $\alpha_s$ running and $\alpha_{s;{\rm MPT}}$ the MPT-model (\ref{alphasMPT}) with $\xi=10$. } \label{alphas}
\end{figure}
A comparison of the strong running coupling for the conventional behavior and the MPT-model with $\xi=10$ is put in Fig.(\ref{alphas}). In drawing the curves, the values of $\Lambda_{\rm QCD}$ are taken as their central values shown in Table \ref{Lambda}. It is noted that in the large scale region, both are consistent with each other.
\subsection{Extrapolation of the $B_c \to \eta_c(J/\psi)$ TFFs}
The pQCD predictions for the $B_c \to \eta_c(J/\psi)$ TFFs are reliable in the large recoil region with small $q^2$. In order to explain the semi-leptonic decay, we need the TFFs in all their allowable $q^2$ region. For the purpose, several extrapolations have been suggested, cf.Refs.\cite{Kiselev00,Kiselev02, Huang07,Lu09,BSWFF85,monopole}. In the present paper, we adopt the monopole form, which has been suggested in Ref.\cite{monopole}, to do the extrapolation, i.e.
\begin{eqnarray}
{f_i}(q^2)={f_i}(0)\bigg(1-\frac{q^2}{m_{\rm pole}^2}\bigg)^{-1}, \label{pole}
\end{eqnarray}
where $m_{\rm pole}$ stands for the mass of the lowest-lying resonance. The first derivative of the TFFs over $q^2$ at $q^2=0$, $f'_{i}(0)=f_{i}(0)/m_{\rm pole}^2$, can be calculated within the framework of the QCD sum rules, which inversely can be adopted for determining $m_{\rm pole}$. We take $m_{\rm pole}=4.50$ GeV~\cite{Kiselev00,Kiselev02,qiao13} to do our analysis.
\section{Numerical results and discussions}
\label{sec:3}
We set the $c$- and $b$-quark pole masses as: $m_c = 1.4 \pm 0.1$ GeV and $m_b = 4.9 \pm 0.1$ GeV. And the following PDG values are adopted~\cite{RPP}: $|V_{\rm cb}|=0.0409\pm0.0011$, $m_{\tau}=1.777$ GeV, $m_{Bc}=6.2745$ GeV, $m_{\eta_c}=2.9837$ GeV and $m_{J/\psi}=3.0969$ GeV. We ignore the spin effect for determining the wavefunction at the origin for $\eta_c$ and $J/\psi$, i.e. we adopt $|\psi_{\eta_c}(0)|= |\psi_{J/\psi}(0)|$. The value of $|\psi_{J/\psi}(0)|$ can be determined from the $J/\psi$ leptonic decay width with a relatively high precision. By taking $\Gamma_{J/\psi \to e^+e^-}=5.55\pm0.16$ keV~\cite{RPP} and following the idea of Ref.\cite{wangnpb}, we obtain $|\psi_{J/\psi}(0)|= (0.257_{-0.006}^{+0.010}){\rm GeV}^{3/2}$. As for the wavefunction at the origin for the $B_c$ meson, it can be related with the decay constant via the relation~\cite{psif}, $f_{B_c}^2={12{|\psi_{B_c}(0)|^2}}/{m_{B_c}}$. Using the lattice QCD estimation, $f_{B_c}=(0.489\pm0.005)$ GeV~\cite{fBc}, we obtain $|\psi_{B_c}(0)|=(0.354\pm0.004)\;{\rm GeV}^{3/2}$. As a cross check of our present calculation, when taking the same input parameters as those of Refs.\cite{qiao12,qiao13} , we recover the same numerical results for the $B\to\eta_c(J/\psi)$ TFFs.
\subsection{The TFFs at the maximum recoil region $q^2=0$}
\begin{table}[htb]
\begin{tabular}{c|c|c|c|c}
\hline
& \multicolumn{3}{c|}{Conventional} & PMC \\
\hline
~~~$\mu_r^{\rm init}$~~~ & ~~$m_b/2$~~ & ~~$m_b$~~ & ~~$2m_b$~~ & ~~$m_b/2;m_b;2m_b$~~ \\
\hline
$F_1^{B_c\to \eta_c}$ & 1.50 & 1.28 & 1.13 & 1.65 \\
\hline
$F_0^{B_c\to \eta_c}$ & 1.50 & 1.28 & 1.13 & 1.65 \\
\hline
$A_0^{B_c\to J/\psi}$ & 1.08 & 0.97 & 0.89 & 0.87 \\
\hline
$A_1^{B_c\to J/\psi}$ & 1.20 & 1.06 & 0.96 & 1.07 \\
\hline
$A_2^{B_c\to J/\psi}$ & 1.28 & 1.14 & 1.03 & 1.15 \\
\hline
$V^{B_c\to J/\psi}$ & 1.65 & 1.46 & 1.32 & 1.47 \\
\hline
\end{tabular}
\caption{The $B_c \to \eta_c (J/\psi)$ TFFs at $q^2=0$ under the conventional scale setting and the PMC scale setting, in which three typical initial scales are adopted. The PMC scale (\ref{pmcscale}) and hence the PMC predictions are independent of $\mu_r^{\rm init}$. } \label{tab:ffscales}
\end{table}
We put the numerical results for the $B_c \to \eta_c(J/\psi)$ TFFs at the maximum recoil region $q^2=0$ in Table~\ref{tab:ffscales}, where three typical initial scales, $\mu_r^{\rm init}=m_b/2$, $m_b$ and $2m_b$, are adopted. The TFFs decrease with a larger value for $\mu_r^{\rm init}$. Under the conventional scale setting, $\mu_r \equiv \mu_r^{\rm init}$, all the TFFs show a strong scale dependence, i.e. they change by about $[-10\%,+17\%]$ for $\mu_r^{\rm init}\in[m_b/2,2m_b]$. After applying the PMC scale setting, the PMC scale is fixed via Eq.(\ref{pmcscale}), thus, the PMC predictions are independent of the choice of $\mu_r^{\rm init}$.
The usual assumption that the renormalization scale depends on $m_b$ does not have a clear justification. As a byproduct, it is noted that our present scale invariant PMC prediction inversely provides us a chance to set the typical scale for the TFFs. That is, the typical renormalization scale $\mu_r=\mu^{\rm ty}_r$ for the conventional scale setting can be predicted such that it leads to the same TFFs as that of the PMC predictions. Following such argument, we obtain ${\mu_r^{\rm ty}}\simeq 0.3{m_b}$ for $F_{1,0}^{B_c\to \eta_c}$, ${\mu_r^{\rm ty}}\simeq {m_b}$ for $A_{1,2}^{B_c\to J/\psi}$ and $V^{B_c J/\psi}$, and ${\mu_r^{\rm ty}}\simeq 2.3{m_b}$ for $A_{0}^{B_c\to J/\psi}$. This shows that not all of the TFFs have the usual typical scale $m_b$.
\begin{table}[htb]
\begin{tabular}{c|c|c|c|c|c|c}
\hline
& \multicolumn{3}{c|}{Conventional} & \multicolumn{3}{c}{PMC} \\
\hline
~~ & ~~LO~~ & ~~NLO~~ & ~~sum~~ & ~~LO~~ & ~~NLO~~ & ~~sum~~ \\
\hline
$F_1^{B_c\to \eta_c}$ & 0.86 & 0.42 & 1.28 & 1.75 & -0.10 & 1.65 \\
\hline
$F_0^{B_c\to \eta_c}$ & 0.86 & 0.42 & 1.28 & 1.75 & -0.10 & 1.65 \\
\hline
$A_0^{B_c\to J/\psi}$ & 0.75 & 0.23 & 0.98 & 1.51 & -0.64 & 0.87 \\
\hline
$A_1^{B_c\to J/\psi}$ & 0.78 & 0.28 & 1.06 & 1.59 & -0.52 & 1.07 \\
\hline
$A_2^{B_c\to J/\psi}$ & 0.84 & 0.30 & 1.14 & 1.70 & -0.55 & 1.15 \\
\hline
$V^{B_c\to J/\psi}$ & 1.08 & 0.38 & 1.46 & 2.19 & -0.72 & 1.47 \\
\hline
\end{tabular}
\caption{The LO and NLO terms for the $B_c\to \eta_c (J/\psi)$ TFFs at $q^2=0$ under the conventional scale setting and the PMC scale setting. $\mu^{\rm init}_r=m_b$. } \label{tab:fflonlo}
\end{table}
After applying the PMC scale setting, due to the elimination of the divergent renormalon terms as $n!\beta^n\alpha_s^n$ with $n$ being the $n$-loop correction, the pQCD convergence can be greatly improved in principle. To show how the pQCD convergence behaves for the $B_c\to \eta_c (J/\psi)$ TFFs, we present the LO and NLO terms for those TFFs at $q^2=0$ before and after the PMC scale setting in Table~\ref{tab:fflonlo}. For clarity, we define a $K$ factor that equals to the magnitude of the ratio between NLO-term and the LO-term, i.e. $K_i=|f_i^{\rm NLO}|/|f_i^{\rm LO}|$. Under the conventional scale setting, for the case of $\mu^{\rm init}_r=m_b$, we obtain
\begin{eqnarray}
&& K_{F_{0}^{B_c\to \eta_c}}=49\%\;,\; K_{F_{1}^{B_c\to \eta_c}}=49\% \;, \nonumber\\
&& K_{A_0^{B_c\to J/\psi}}=31\%\;,\; K_{A_1^{B_c\to J/\psi}}=36\% \;,\nonumber\\
&& K_{A_2^{B_c\to J/\psi}}=36\%\;,\; K_{V^{B_c\to J/\psi}}=35\%\;,
\end{eqnarray}
and after the PMC scale setting, we have
\begin{eqnarray}
&& K_{F_{0}^{B_c\to \eta_c}}=6\%\;,\; K_{F_{1}^{B_c\to \eta_c}}=6\%\;, \nonumber\\
&&K_{A_0^{B_c\to J/\psi}}=42\%\;,\; K_{A_1^{B_c\to J/\psi}}=33\% \;, \nonumber\\
&& K_{A_2^{B_c\to J/\psi}}=32\%\;,\; K_{V^{B_c\to J/\psi}}=33\%\;.
\end{eqnarray}
After the PMC scale setting, the pQCD convergence for the $B_c \to \eta_c$ TFFs $F_{0,1}^{B_c\to \eta_c}$ can be greatly improved, while the $K$ factors for the $B_c \to J/\psi$ TFFs $A_{0,1,2}^{B_c\to J/\psi}$ and $V^{B_c\to J/\psi}$ are still large. The large $K$ factors for the $B_c \to J/\psi$ TFFs indicate that the unknown even higher-order pQCD corrections shall give sizable contributions to the TFFs, which are important either for fixing more precise lower-order PMC scales or for estimating the sizable higher-order conformal contributions. As a minor point, from Table.~\ref{tab:fflonlo}, one may observe that all the NLO corrections to TFFs change from positive values to negative ones after the PMC scale setting.
\begin{table}[htb]
\begin{tabular}{c|c|c}
\hline
& ~~Conventional~~ & ~~PMC~~ \\
\hline
$\Delta(F_1^{B_c\to \eta_c})$ & $\pm0.42$ & $\pm0.10$ \\
\hline
$\Delta(F_0^{B_c\to \eta_c})$ & $\pm0.42$ & $\pm0.10$ \\
\hline
$\Delta(A_0^{B_c\to J/\psi})$ & $\pm0.26$ & $\pm0.64$ \\
\hline
$\Delta(A_1^{B_c\to J/\psi})$ & $\pm0.30$ & $\pm0.52$ \\
\hline
$\Delta(A_2^{B_c\to J/\psi})$ & $\pm0.33$ & $\pm0.55$ \\
\hline
$\Delta(V^{B_c\to J/\psi})$ & $\pm0.42$ & $\pm0.72$ \\
\hline
\end{tabular}
\caption{An estimation of the unknown even higher-order pQCD corrections, $\Delta=\pm |{\tilde{\cal C}} \alpha^{2}_s|_{\rm MAX}$, for $B_c\to \eta_c (J/\psi)$ TFFs at $q^2=0$ under the conventional and PMC scale settings. } \label{tab:higherOrder}
\end{table}
For a pQCD estimation, it is helpful to predict what's the ``unknown" QCD corrections could be. The conventional estimation done by varying the scale over a certain range is not proper, since it can only estimate the non-conformal contribution but not the conformal one. To achieve an estimation of how the ``unknown" QCD corrections could be from the ``known" QCD corrections, a more conservative method for the scale error analysis has been suggested in Ref.\cite{PMC4}; i.e. to take the scale uncertainty as the last known perturbative order. More explicitly, for the present NLO estimation, the pQCD scale uncertainty $\Delta=\pm |{\tilde{\cal C}} \alpha^{2}_s|_{\rm MAX}$, where both ${\tilde{\cal C}}$ and $\alpha_s$ are calculated by varying $\mu^{\rm init}_{r}\in[m_b/2,2m_b]$ and the symbol ``MAX'' stands for the maximum value of $|{\tilde{\cal C}} \alpha^{2}_s|$ within this region. The expression of ${\tilde{\cal C}}$ can be read from Eqs.(\ref{starting},\ref{startingPMC}). We put the $\Delta$ uncertainty for various TFFs in TABLE~\ref{tab:higherOrder}. The large $\Delta$ values also confirm the importance of a next-to-next-to-leading order correction for the $J/\psi$ case. As examples, some PMC analysis up to two-loop, three-loop and four-loop QCD corrections have been done in Refs.\cite{PMC1,PMC2,PMC4,pmcapp1,pmcapp2,pmcapp3,pmcapp4}, which show exactly that the pQCD convergence and the pQCD prediction can be greatly improved after the PMC scale setting.
\subsection{The $B_c\to \eta_c (J/\psi)$ semi-leptonic decays}
\begin{table*}
\begin{tabular}{c |c|c|c|c|c|c|c|c|c|c}
\hline
& PMC & Conventional & \cite{qiao13} & \cite{nrcqm06} & \cite{rcqm05} & \cite{Huang07,Huang08} & \cite{Chang94,Chang02} & \cite{QCDrpm} & \cite{Kiselev00} & \cite{LOpQCD}\\
\hline
$B_c \to \eta_c \ell \nu$ & $71.53_{-8.90}^{+11.27}$ & $43.36_{-5.75}^{+7.17}$ & $30.50^{+9.74}_{-4,82}$ & $6.95^{+0.29}$ & 10.7 & 23.98 & 14.2 & 11.1
& $11\pm1$ & $6.45^{+1.78}_{-1.59}$ \\
\hline
$B_c \to \eta_c \tau \nu$ & $27.14_{-4.33}^{+5.93}$ & $16.46_{-2.73}^{+3.69}$ & $9.29^{+2.70}_{-1.62}$ & $2.46^{+0.07}$ & 3.52 & 7.16 & $\sim$ & $\sim$ & $3.3\pm0.9$ & $2.00_{+0.54}^{-0.50}$ \\
\hline
$B_c \to J/\psi \ell \nu$ & $106.31_{-14.01}^{+18.59}$ & $104.74_{-15.49}^{+20.08}$ & $97.30^{+36.22}_{-20.33}$ & $21.9^{+1.2}$ & 28.2 & 34.69 & 34.4 & 30.2 & $28\pm5$ & $14.7^{+1.94}_{-1.73}$ \\
\hline
$B_c \to J/\psi \tau \nu$ & $28.25_{-4.35}^{+6.02}$ & $28.12_{-4.72}^{+6.38}$ & $7.55^{+2.85}_{-1.56}$ & $5.86^{+0.23}_{-0.03}$ & 7.82 & 9.50 & $\sim$ & $\sim$ & $7\pm2$ & $4.27^{+0.58}_{-0.50}$ \\
\hline
\end{tabular}
\caption{The decay widths (in unit: $10^{-15}{\rm GeV}$) for the $B_c \to \eta_c (J/\psi) \ell \nu (\ell=e,\mu)$ and $B_c \to \eta_c (J/\psi) \tau \nu$ under the conventional and the PMC scale settings. As a comparison, we also present the results derived from the NLO pQCD factorization~\cite{qiao13}, the constituent quark model~\cite{rcqm05,nrcqm06}, the Bethe-Salpeter equation~\cite{Chang94,Chang02}, the QCD sum rules~\cite{Kiselev00,Huang07,Huang08}, the QCD relativistic potential model~\cite{QCDrpm}, and the LO pQCD~\cite{LOpQCD}.} \label{tab:width}
\end{table*}
\begin{figure}[htb]
\includegraphics[width=0.5\textwidth]{GammaEtaclv.eps}
\includegraphics[width=0.5\textwidth]{GammaJpsilv.eps}
\caption{The decay widths for $B_c \to \eta_c(J/\psi) \ell \nu$ up to NLO level versus the initial scale $\mu_r^{\rm init}$ under the conventional and the PMC scale settings, respectively.}
\label{fig:decaywidth}
\end{figure}
By using the monopole extrapolation (\ref{pole}) for the $B_c\to \eta_c (J/\psi)$ TFFs, we are ready to predict the $B_c\to \eta_c (J/\psi)$ semi-leptonic decay widths. The results are presented in Table~\ref{tab:width}, where the errors are the squared average of the mentioned error sources. As a comparison, we present the results before and after the PMC scale setting simultaneously, and we also present the results derived from the NLO pQCD factorization~\cite{qiao13}, the constituent quark model~\cite{rcqm05,nrcqm06}, the Bethe-Salpeter equation~\cite{Chang94,Chang02}, the QCD sum rules~\cite{Kiselev00,Huang07,Huang08}, the QCD relativistic potential model~\cite{QCDrpm}, and the LO pQCD~\cite{LOpQCD}. The renormalization scale dependence of $B_c \to \eta_c(J/\psi) \ell \nu$ decay widths are shown in Fig.(\ref{fig:decaywidth}). After the PMC scale setting, the decay widths are also independent to the choice of $\mu^{\rm init}_r$, which are consistent with the above discussions on the TFFs. Due to the $\tau$ mass suppression, the decay widths for the $\tau$-lepton pair production are smaller than the those of $e$-lepton pair or $\mu$-lepton pair. More specifically, after the PMC scale setting, we have
\begin{eqnarray}
R(\eta_c)&=&\frac{Br(B_c \to \eta_c \ell \nu)}{Br(B_c \to \eta_c \tau \nu)}\simeq2.6, \;{\rm for}\; \ell=e,\mu, \\
R(J/\psi)&=&\frac{Br(B_c \to J/\psi \ell \nu)}{Br(B_c \to J/\psi \tau \nu)}\simeq3.8, \;{\rm for}\; \ell=e,\mu.
\end{eqnarray}
\begin{figure}[htb]
\includegraphics[width=0.45\textwidth]{Rcompare.eps}
\caption{The value of $\Re(J/\psi{\ell^+}\nu)$ after the PMC scale setting, which is shown by the upper shaded band. The CDF predictions~\cite{Ru05,Ru09,Ru14,Re05}, the QCD sum rule (SR) prediction~\cite{Kiselev00} and the LO pQCD prediction~\cite{LOpQCD} are presented as a comparison. The middle shaded band represents the QCD sum rule prediction and the lower shaded band is the LO pQCD prediction under the conventional scale setting.}
\label{fig:Rcompare}
\end{figure}
By taking the hadronization fractions $f_{\bar b \to B_c^+}=(1.3\pm0.2)\times10^{-3}$~\cite{Bcprod}, $f_{\bar b \to B^+}=0.401\pm0.008$ and $BR(B^+ \to J/\psi K^+)=(1.028 \pm 0.031)\times 10^{-3}$~\cite{RPP}, we can predict the $\sigma \cdot BR$ ratio $\Re(J/\psi{\ell^+}\nu)$. Our results for $\Re(J/\psi{\ell^+}\nu)$ as a function of the $B_c$ meson lifetime $\tau_{B_c}$ are presented in Fig.(\ref{fig:Rcompare}). As a comparison, the CDF measurements~\cite{Bc98,Ru05,Ru09,Ru14,Re05} as shown in Table~\ref{tab:Rexpe}, where all the errors are added in quadrature. As a comparison, the estimations based on QCD sum rule~\cite{Kiselev00} and LO pQCD prediction~\cite{LOpQCD} are also presented. All those predictions on $\Re(J/\psi{\ell^+}\nu)$ are close in shape, all of which increase with the increment of $\tau_{B_c}$. However our estimation of $\Re(J/\psi{\ell^+}\nu)$ shows a better agreement with the CDF measurements, which indicates the importance of the NLO calculations and also the importance of a correct scale setting.
\subsection{A detailed discussion on the uncertainties of the decay widths}
Table \ref{tab:width} shows the squared average of all errors. In the present subsection, we present a detailed discussion on the dominant error sources. There are many error sources for determining the decay widths, such as the $|V_{cb}|$, $m_{\rm pole}$, $\xi$, $\Lambda_{\rm QCD}$, the bound-state parameters $m_c$, $m_b$, $|\psi_{B_c}(0)|$, $|\psi_{\eta_c}(0)|$ and $|\psi_{J/\psi}(0)|$. Taking the dominant error sources into consideration, we obtain
\begin{eqnarray}
\Gamma_{B_c \to \eta_c \ell \nu} &=&(71.53_{-6.16-3.55-3.80-3.72-0.62}^{+8.19+4.87+3.90+4.55+0.62}) 10^{-15} {\rm GeV} \nonumber\\
&=& (71.53_{-8.90}^{+11.27}) 10^{-15} {\rm GeV},\\
\Gamma_{B_c \to \eta_c \tau \nu} &=&(27.14_{-2.34-3.02-1.44-1.41-0.24}^{+3.11+4.50+1.48+1.73+0.23}) 10^{-15} {\rm GeV} \nonumber\\
&=& (27.14_{-4.33}^{+5.93}) 10^{-15} {\rm GeV},\\
\Gamma_{B_c \to J/\psi \ell \nu} &=&(106.31_{-8.76-8.78-5.64-3.21-0.51}^{+11.74+12.66+5.80+3.68+0.50}) 10^{-15} {\rm GeV} \nonumber\\
&=& (106.31_{-14.01}^{+18.59}) 10^{-15} {\rm GeV},\\
\Gamma_{B_c \to J/\psi \tau \nu} &=&(28.25_{-2.33-3.24-1.50-0.84-0.13}^{+3.12+4.81+1.54+0.96+0.13}) 10^{-15} {\rm GeV} \nonumber\\
&=& (28.25_{-4.35}^{+6.02}) 10^{-15} {\rm GeV},
\end{eqnarray}
where $\ell$ stands for the light leptons $e$ and $\mu$, the uncertainties from the left to right are for a combined effect of the bound state parameters, $m_{\rm pole}$, $|V_{cb}|$, $\xi$, $\Lambda_{\rm QCD}$, respectively. More specifically,
\begin{itemize}
\item The squared average of the uncertainties from the bound state parameters $m_c$, $m_b$, $|\psi_{B_c}(0)|$, $|\psi_{\eta_c}(0)|$ and $|\psi_{J/\psi}(0)|$ are $[-9\%,+11\%]$ for both $B_c \to \eta_c \ell \nu$ and $B_c \to \eta_c \tau \nu$; and $[-8\%,+11\%]$ for both $B_c \to J/\psi \ell \nu$ and $B_c \to J/\psi \tau \nu$.
\item The value of $m_{\rm pole}$ determines the extrapolated shape of the $B_c \to \eta_c(J/\psi)$ TFFs, and we adopt $m_{\rm pole}=(4.50\pm0.25){\rm GeV}$ to do the estimation. The error is $[-5\%,+7\%]$ for $\Gamma_{B_c \to \eta_c \ell \nu}$ and $[-11\%,+17\%]$ for $\Gamma_{B_c \to \eta_c \tau \nu}$; and $[-8\%,+12\%]$ for $\Gamma_{B_c \to J/\psi \ell \nu}$ and $[-11\%,+17\%]$ for $\Gamma_{B_c \to J/\psi \tau \nu}$.
\item The $|V_{cb}|$ being the overall factor for all the $B_c \to \eta_c(J/\psi)$ TFFs, then its error to the decay widths are the same for the channels, which reads $[-5\%,+5\%]$ for $|V_{\rm cb}|=0.0409\pm0.0011$.
\item The errors caused by the MPT parameter $\xi=10\pm2$ are within the region of $[-5\%,+6\%]$ for $\Gamma_{B_c \to \eta_c \ell \nu}$ and $\Gamma_{B_c \to \eta_c \tau \nu}$, and $[-3\%,+3\%]$ for $\Gamma_{B_c \to J/\psi \ell \nu}$ and $\Gamma_{B_c \to J/\psi \tau \nu}$, respectively.
\item By using the values listed in Table \ref{Lambda}, we show that the $\Lambda_{\rm QCD}$ shall cause small errors, i.e. less than $\pm1\%$, for all the decay channels.
\end{itemize}
\begin{figure}[htb]
\includegraphics[width=0.48\textwidth]{dGammaEtaclv.eps}
\includegraphics[width=0.48\textwidth]{dGammaJpsilv.eps} \caption{Differential decay widths for $B_c \to \eta_c(J/\psi) \ell \nu$ $(\ell=e,\mu,\tau)$ versus $q^2$ under the PMC scale setting. }
\label{fig:dgamma}
\end{figure}
It is also helpful to show how those error sources affect the differential decay widths. Also by taking the squared average of those errors, we draw the differential decay widths of $B_c \to \eta_c(J/\psi) \ell \nu(\ell=e,\mu,\tau)$ for $m_{\ell}^2 \leq q^2 \leq (m_{B_c}-m_{\eta_c(J/\psi)})^2$ in Fig.(\ref{fig:dgamma}).
\section{Summary}
\label{sec:4}
The PMC provides a systematic and unambiguous way to set the renormalization scale for any high-energy QCD processes. In the present paper, we have studied the NLO QCD corrections for the $B_c \to \eta_c(J/\psi)$ TFFs by adopting the PMC scale setting. As a further step, we have calculated the $B_c \to \eta_c(J/\psi) \ell \nu$ semi-leptonic decay widths and estimated the measurable parameter $\Re(J/\psi{\ell^+}\nu)$. We have found that
\begin{itemize}
\item After applying the PMC scale setting, all the same type higher-order $\beta_0$-terms have been resummed into the running coupling, which rightly determines the optimal renormalization scale for the $B_c \to \eta_c(J/\psi)$ TFFs. All the $B_{c} \to \eta_c (J/\psi)$ TFFs have the same PMC scale due to the same $\beta_0$-terms at the NLO level. Thus, the large renormalization scale uncertainty for all the TFFs under the conventional scale setting, which is about $[-10\%,+17\%]$ for $\mu_r\in[m_b/2,2m_b]$, are strongly suppressed.
After applying the PMC scale setting, the pQCD convergence can be improved in principle due to the elimination of the renormalon terms. We have seen an obvious improvement on the pQCD convergence for the $B_c \to \eta_c$ TFFs. However for $B_c \to J/\psi$ TFFs, the $K$ factor is still large, which indicates a NNLO calculation is necessary before an obvious pQCD convergence can be achieved.
\item After applying the PMC scale setting, we obtain larger decay widths for the decays $B_c \to \eta_c(J/\psi) \ell \nu$ in comparison to values under the conventional scale setting, i.e.
\begin{eqnarray}
\Gamma_{B_c \to \eta_c \ell \nu}&=&(71.53^{+11.27}_{-8.90})\times 10^{-15} {\rm GeV},\\
\Gamma_{B_c \to \eta_c \tau \nu}&=&(27.14^{+5.93}_{-4.33})\times 10^{-15} {\rm GeV},\\
\Gamma_{B_c \to J/\psi \ell \nu}&=&(106.31^{+18.59}_{-14.01})\times 10^{-15} {\rm GeV},\\
\Gamma_{B_c \to J/\psi \tau \nu}&=&(28.25^{+6.02}_{-4.35})\times 10^{-15} {\rm GeV},
\end{eqnarray}
where $\ell=e,\mu$, and the errors are squared averages of those from the dominant sources as $m_c$, $m_b$, $|\psi_{B_c}(0)|$, $|\psi_{J/\psi}(0)|$, $|\psi_{\eta_c}(0)|$, $m_{\rm pole}$, $\xi$, $|V_{cb}|$ and etc..
\item The PMC scale for the decays $B_c \to \eta_c(J/\psi) \ell \nu$ is $\mu_r^{\rm PMC} \approx 0.8{\rm GeV}$, which is in the low-energy region. To provide a reliable pQCD estimation, we have adopted the MPT running coupling model to do the calculation. By taking its input parameter $\xi=10\pm2$, we obtain $[-5\%,+6\%]$ uncertainty for $\Gamma_{B_c \to \eta_c \ell \nu}$ and $\Gamma_{B_c \to \eta_c \tau \nu}$, and $[-3\%,+3\%]$ uncertainty for $\Gamma_{B_c \to J/\psi \ell \nu}$ and $\Gamma_{B_c \to J/\psi \tau \nu}$, respectively.
\item We have estimated the value of $\Re(J/\psi{\ell^+}\nu)$, the production cross section times branching ratio fraction between $B_c^+ \to J/\psi \ell^+ \nu$ and $B^+\to J/\psi K^+$. Our present estimation, as shown in Fig.(\ref{fig:Rcompare}), shows a good agreement with CDF measurements.
\end{itemize}
\hspace{2cm}
{\bf Acknowledgement:} This work was supported in part by the Fundamental Research Funds for the Central Universities under Grant No.CQDXWL-2012-Z002, by Natural Science Foundation of China under Grant No.11275280.
\hspace{1cm}
|
\section{Introduction\label{sec:intro}}
Computational simulations have opened a new avenue for the
exploration and prediction of ``\`a la carte'' molecular complexes and materials,
i.e., with tailored properties and functionality, due to the development of
powerful algorithms and an increase in computational power.
Generally, an increase in complexity or system size goes hand-in-hand with a loss of accuracy or an increase in numerical cost.
Highly accurate wave function methods, which have recently become available also for solids \cite{BGKA2013}, are limited in the complexity of systems
they can handle.
This hinders the application of these methods
to complex molecular structures, e.g.\ nanostructures and biomolecules.
However, now, more than ever, there is a need for methods that are able to
handle large-scale systems with high precision in order to assess the challenges
in material, bio-, and nanosciences.
Density functional theory (DFT)\cite{HK1964, KS1965}, on the
other hand, comes with a low computational cost, which allows its application to
rather large systems. The lack of a
systematic way to improve functionals is hampering the progress towards a more accurate
and efficient DFT first-principles scheme.
Currently, several properties are still difficult to predict within standard DFT, with the band gap and
molecular dissociation being prominent examples.
The latter is described correctly with a standard local density approximation (LDA)
or generalized gradient approximation functional only if the spin symmetry is artificially broken.
The recently introduced strictly correlated functional\cite{MG2012} rectifies this problem,
however, at the expense of a wrong description of the equilibrium properties.
Although the exact exchange-correlation (xc) functional of DFT is universal,
one could of course use different approximations depending on the physical situation
or quantity of interest. To this end, large databases collecting
information from different functionals are compiled\cite{PT2013}.
The lack of a proper description of static correlation
is at the heart of many of the failures of present xc approximations of DFT
for describing strongly correlated systems and molecular bond breaking and formation.
This deficiency can be
traced back to the fact that within Kohn-Sham (KS) DFT the density is reproduced via a
single Slater determinant. When the true wave function has multireference character the KS kinetic energy is a poor approximation to the true kinetic energy. The difference has to be compensated by the xc energy which for many approximations is not done very successfully, leading to a wrong total energy in the case of molecular dissociation.
In electronic structure theory, a desirable feature of independent particle models,
like the KS scheme in DFT, is the direct
prediction of single-electron properties, like ionization potentials (IPs), from the eigenvalues
of the single-particle Hamiltonians.
Although the question about the physical content of the KS orbital energies raised
a scientific debate~\cite{PY1989}, theoretical justification for this result
was given by Baerends and co-workers~\cite{CGB2002,GBB2003} and by Bartlett and co-workers~\cite{BLS2005,VB2012}.
Unfortunately, orbital energies from approximate functionals tend to underestimate substantially the
IPs of molecular systems~\cite{PN1982,PA1998}.
Reduced-density-matrix-functional theory (RDMFT) \cite{G1975}
is an alternative to DFT to approximate the many-electron problem. It is based
on approximating the total energy of an electronic system in terms of the one-body reduced
density matrix (1RDM). A
main difference from DFT is the introduction of fractional occupation numbers which
allows the exact treatment of the kinetic energy and potentially leads to improved
accuracy whenever the
ground-state many-body wave function is far from a single Slater determinant.
Most approximations in RDMFT are explicit functionals of -- and are minimized in
terms of -- the natural orbitals (NOs), $\phi_j(\v r)$, and their occupation
numbers, $n_j$. So far,
various different approximations for the total energy functional have become
available\cite{M1984,GU1998,bb0,GPB2005,power_finite,pernal2010,AC3,pade,pnof1,pnof2,PNOF5,sharma08}
which have proven to describe correctly such diverse properties as molecular
dissociation\cite{bb0,GPB2005,power_finite,pernal2010,AC3}
or band gaps\cite{sharma08,helbig09,dfg10,SDSG2013}.
However,
applications have been restricted to small molecules
due to the computational cost to determine the orbitals.
Although the optimization of the occupation numbers is a relatively inexpensive task,
orbital minimization is complicated: it does not reduce to an iterative eigenvalue problem
(as in DFT) and requires numerically expensive procedures.
Significant effort has been devoted to devising effective
Hamiltonians\cite{pernal_epot,piris_jcc,Baldsiefen2013114} to improve the efficiency, but with
limited success so far.
If we were able to incorporate all the advantages of RDMFT functionals into DFT while
keeping the cost of standard DFT functionals we would make a big step forward in constructing
an efficient and accurate scheme able to cope with the challenge of describing
structural and dynamical properties of many-electron systems including bond breaking and
bond formation. In this paper, we propose such a framework that combines the best of
both DFT and RDMFT. One can regard this approach as either an extension of DFT,
where fractional occupations for the orbitals are introduced using an
approximation for the xc energy functional borrowed from
RDMFT, or, alternatively, as a constrained RDMFT calculation. In either case, we
incorporate the proper nonidempotent nature of the density matrix in the
calculation of the kinetic energy that is fundamental to the success of
RDMFT approximations.
The central idea in our proposed framework, which we call
{\it local RDMFT}, is to restrict the minimization with respect to the NOs
to a domain where these orbitals are eigenfunctions of a single-particle Hamiltonian with a local potential.
The best possible Hamiltonian is the one whose eigenorbitals minimize the total energy.
The resulting equations are similar to the optimized effective potential (OEP)
equations\cite{sharp,talman,kummel2008,BL2009}. The OEP improves the accuracy in DFT
and significant expertise has been
developed in the past two decades\cite{kummel2008} for its efficient implementation.
Hence, our method can be implemented directly
in existing DFT codes with only small modifications to address fractional occupation numbers.
Fractional occupations, as in standard RDMFT, are provided by the minimization of the
energy functional under the appropriate conditions. Our approach has some similarity with
the idea explored by Gr\"uning et al.~\cite{GGB2003}, where the common energy denominator approximation
is used together with the M\"uller functional. In that approach, occupancies are obtained in
an empirical way and not through optimization.
The local RDMFT framework provides an energy eigenvalue spectrum connected to the NOs and as we show,
single electron properties, like the IPs of small molecules, are well reproduced by the
energies of the highest occupied molecular orbitals (HOMOs).
In Sec.~\ref{sec:theory},
we describe in detail the formalism of local RDMFT. Then in Sec.~\ref{sec:applications}, we show that
the restriction to a local effective potential has little effect on the
dissociation of dimers; hence, the accuracy of RDMFT is
retained by the proposed method. In Sec.~\ref{sec:applications} we also illustrate the quality of
the energy spectrum provided by the effective Hamiltonian by comparing the obtained IPs with
experiment. In the Appendix we show that pure density xc functionals are
not adequate for the present scheme since they cannot lead to fractional occupations
in a minimization procedure and we need to employ functionals of the 1RDM as we do in the present work.
\section{Local RDMFT \label{sec:theory}}
Clearly, the integration of DFT with RDMFT is desirable as it could combine the best of
both worlds. With fractional occupation numbers, static correlation would become accessible,
while use of a common local potential to yield the NOs would improve
dramatically the efficiency of the method.
However, this target is not straightforward.
A natural way to incorporate a local potential
for the orbitals, is to consider the approximate RDMFT xc energy to be a density functional,
so that its functional derivative would give the local potential.
Unfortunately, such an approach leads in general to an idempotent
density matrix (see Appendix) and we are back to square one.
Hence, we abandon
the requirement that the xc energy must be a functional of the density and that the potential
must arise as a functional derivative with respect to the density.
Instead, we consider the total energy as a functional of the one-body-reduced-density matrix (1RDM),
$\gamma ( {\bf r} , {\bf r}' ) $,
\begin{multline} \label{eq:functional}
E_v^{\rm RDM} [ \gamma ({\bf r}, {\bf r}' ) ] = T [ \gamma ({\bf r}, {\bf r}' ) ] + \int d^3 r \, \gamma ({\bf r}, {\bf r} ) v ( {\bf r}) \\ +
\frac{1}{2} \iint d^3r d^3 r' \frac{ \gamma ({\bf r}, {\bf r} ) \gamma ({\bf r}' , {\bf r}' )}{ | {\bf r - r'} | }
+ E_{\rm xc} ^ {\rm RDM}[ \gamma ({\bf r}, {\bf r}' )] \,,
\end{multline}
where $v( {\bf r})$ is the external local potential and $T [ \gamma ({\bf r}, {\bf r}' ) ]$ is the
interacting kinetic energy which is a functional of the 1RDM. The
electron-electron interaction energy can be cast into the last two terms in Eq.~(\ref{eq:functional})
where $E_{\rm xc} ^ {\rm RDM}[ \gamma ({\bf r}, {\bf r}' )]$ needs to be approximated.
$E_{\rm xc} ^ {\rm RDM}[ \gamma ({\bf r}, {\bf r}' )]$ can be considered a functional of the
occupation numbers $n_j$ and the natural orbitals $\phi_j$, i.e., the eigenvalues and the
eigenvectors of $ \gamma$:
\begin{equation}
\gamma ({\bf r}, {\bf r}' ) = \sum_j n_j \: \phi_j^* ({\bf r}') \: \phi_j ({\bf r})\,.
\end{equation}
The central idea in our proposed local RDMFT scheme is to restrict the search for the
optimal $\phi_j$'s within a domain where they are also
eigenfunctions of a single-particle Hamiltonian with a local potential, $v_{\rm rep}(\v r)$,
\begin{equation}
\label{eq:locham}
\left[ -\frac{\nabla^2}{2} + v_{\rm ext} (\v r) + v_{\rm rep}(\v r) \right] \phi_j (\v r) =
\epsilon_j \phi_j(\v r)\,,
\end{equation}
where $v_{\rm rep}$ is the effective repulsive potential acting on any electron in the system,
caused by the effective
repulsion of the remaining $N-1$ electrons (atomic units are used throughout the paper).
Fractional occupations, $n_j$, as in standard RDMFT, are provided by the minimization of the
energy functional of Eq.~(\ref{eq:functional}) under the appropriate $N$-representability
conditions.
The natural orbitals in the exact theory, as well as the minimizing orbitals in
RDMFT approximations, are typically satisfying a Schr\"odinger equation
with a nonlocal effective potential. The local potential constraint leads to
approximate natural orbitals (ANOs), $\phi_j$, which cannot become equal to the true natural orbitals.
By enforcing Eq.~(\ref{eq:locham}), the total energy becomes a functional of the
local effective potential and of the occupation numbers, $E=E[v_{\rm rep},\{n_j\}]$.
In the same way as in the OEP
method \cite{sharp,talman,kummel2008,BL2009}, the optimal local potential is obtained by
solving the integral equation
\begin{equation} \label{oep_0}
\int d^3r' \chi (\v r , \v r') \, v_{\rm rep} (\v r') = b(\v r)\,,
\end{equation}
where $\chi (\v r, \v r')$, a generalized density-density response function, and $b (\v r)$ are given by
\begin{equation}
\label{eq:A}
\chi (\v r , \v r') = {\sum_{j \ne k}}^\prime \phi_j^*(\v r)\,\phi_k(\v r)\,\phi_k^*(\v r')\,\phi_j(\v r')\frac{n_j-n_k}{\epsilon_j -\epsilon_k}\,,
\end{equation}
\begin{equation}
b(\v r) = {\sum_{j \ne k}}^\prime \langle \phi_j | \frac{F_{\rm Hxc}^{(j)}-F_{\rm Hxc}^{(k)}}{\epsilon_j - \epsilon_k}
| \phi_k \rangle\, \phi_k^*(\v r)\, \phi_j(\v r)\,,
\label{eq:B}
\end{equation}
with $F_{\rm Hxc}^{(j)}$ defined by
\begin{equation}
\label{eq:Liapis}
\frac{\delta E_{\rm Hxc}}{\delta \phi_j^*(\v r)} \doteq
\int d^3r' \, F_{\rm Hxc}^{(j)}(\v r, \v r') \, \phi_j(\v r')\,.
\end{equation}
Here $E_{\rm Hxc}$ is the approximation for the electron-electron interaction
energy, i.e., the last two terms in Eq.~(\ref{eq:functional}).
In the summations of Eqs.~(\ref{eq:A}) and (\ref{eq:B}),
we have excluded terms over pairs of orbitals differing in occupation
by less than a cutoff value $\Delta n_c$. This choice excludes pairs of weakly occupied orbitals whose
energies $\epsilon_j$ are not accurate for finite localized basis sets;
e.g., occasionally they violate the Aufbau principle
and the negative definiteness of $\chi$. By excluding these terms, we also partly alleviate a common
inaccuracy of most RDMFT functionals to show a spurious excess of total occupation in
weakly occupied orbitals \cite{L2013}.
In that way, we have examined the dependence of our solution on $\Delta n_c$. Typically, for very small
values of $\Delta n_c$ we run into convergence issues and IPs vary substantially as a function of $\Delta n_c$. As
$\Delta n_c$ increases to a typical value of 0.1 convergence improves dramatically and IPs remain unchanged
as a function of $\Delta n_c$ for a broad range of values. Even for values of $\Delta n_c$ larger than the HOMO-lowest occupied molecular orbital (LUMO) occupation difference, we do not find any noticeable change in the results.
Further increase deteriorates the accuracy since fewer terms are included in the summations.
A choice for $\Delta n_c\sim 0.1-0.3$ usually leads to a converged solution.
Accidental exclusions of strongly-weakly occupied pairs have a negligible effect on the results.
To ensure a physical asymptotic decay of the effective repulsive potential, we do not solve Eq. (\ref{oep_0}) directly.
Instead, we follow the methodology in Ref.~\cite{GL2012} and express $v_{\rm rep} (\v r)$ as the electrostatic potential
of an effective repulsive density, $\rho_{\rm rep} (\v r)$:
\begin{equation}
v_{\rm rep}(\v r) = \int d^3r' \frac{\rho_{\rm rep}(\v r')}{|\v r-\v r'|}\,.
\end{equation}
The requirement~\cite{GL2012} that the effective repulsive density corresponds to a (fictitious)
system of $N-1$ electrons repelling the electron at $\v r$ yields the following two constraints:
\begin{eqnarray}
Q_{\rm rep}\doteq \int d^3r \: \rho_{\rm rep}(\v r) &=& N-1, \label{eq:asympt}\\
\rho_{\rm rep}(\v r) & \geq & 0\,. \label{eq:pos}
\end{eqnarray}
The first condition is necessary for the asymptotic decay of the effective repulsive potential as $(N-1)/r$, which is a property of
the exact Hxc potential\cite{PhysRevLett.83.5459}.
The two conditions together become sufficient (although probably not necessary anymore) to guarantee the correct asymptotic
behavior and a well-posed mathematical problem.
Minimization of the total energy leads to an integral equation for the effective repulsive density:
\begin{equation}
\int d^3 r' \, {\tilde \chi} (\v r , \v r') \, \rho_{\rm rep} (\v r') = {\tilde b}( \v r),
\end{equation}
with
\begin{eqnarray}
\label{eq:oep2}
{\tilde \chi} ( \v r , \v r')\! \doteq\!\! \iint \!\! d^3 x \, d^3 y \, \frac{ \chi ( \v x , \v y) }{ | \v x - \v r| | \v y - \v r'| }\,,\nonumber \\
{\tilde b} ( \v r ) \! \doteq \int \!\! d^3 x \, \frac{ b ( \v x ) }{ | \v x - \v r| }.
\end{eqnarray}
The two constraints can be incorporated with a
Lagrange multiplier (\ref{eq:asympt}) and a penalty term (\ref{eq:pos}) that introduces an energy
cost for every point where $ \rho_{\rm rep}$ becomes negative.
We expand the ANOs in a basis set (orbital basis) and the effective repulsive density
(rather than the potential) in another (auxiliary) basis and
we obtain, similarly to Ref.~\cite{GL2012}, a linear system of equations
for the expansion coefficients of the repulsive density.
The inversion of the linear equations is complicated
by the fact that the matrix of the response function $\tilde \chi$ becomes
singular when the auxiliary basis is large compared to the orbital
basis\cite{SSD2006,HIGBBT2001,GL2012b,BFBG2011}.
As a result, the
effective repulsive density becomes indeterminate in the null space of the
(finite-orbital-basis) response function. This indeterminacy is substantially
suppressed by the two constraints, Eqs.~(\ref{eq:asympt}) and (\ref{eq:pos}), that reduce drastically the
form of the admissible effective repulsive densities and effective repulsive potentials.
We have found that with a singular value decomposition, we obtain systematically smooth
and physical densities and potentials (see Ref.~\onlinecite{GL2012}).
A consequence of the local-potential approximation is that the asymptotic decay of the
ANOs depends on the energy
eigenvalue $\epsilon_j$ and hence, differs from the (necessarily uniform) asymptotic decay
of exact NOs with fractional occupancy\cite{MPL1975,Davidson1976}.
As a result, the asymptotic exponential decay of the density is related to the
highest energy eigenvalue
with nonzero occupation and not to the IP.
The effect in the total energy from the different asymptotics - compared with the behavior of the exact NOs - is negligible
since the energy contribution of the asymptotic region is insignificant.
Similarly to standard RDMFT, nominal scaling of local RDMFT is $N^4\times M$, where $M$
is the total number of different generalized Fock matrices [defined in Eq.~(\ref{eq:Liapis})] that
need to be evaluated and then transformed to obtain the matrix elements in Eq.~(\ref{eq:B}). For most approximations equal occupations
result in equal generalized Fock matrices. Hence,
$M$ is typically equal to the number of orbitals with fractional occupations
plus an additional one common for all the fully occupied orbitals. Since $M$ can be fixed,
the nominal scaling of local RDMFT reduces to that of Hartree-Fock (HF) or hybrid DFT calculations.
The number of different $F_{\rm Hxc}^{(j)}(\v r, \v r')$ that need to be evaluated, and the additional
cycle of convergence of the occupations apart from the ANOs add to the overall computational cost
of local-RDMFT as compared to standard DFT and Hartree-Fock calculations.
The present scheme would benefit from the recent developments in reducing
the scaling of wave-function-based schemes~\cite{MPT2013}. Significant numerical cost saving
is also expected by improving the enforcement of the positivity condition of Eq.~(\ref{eq:pos}).
However, the search for the optimal
repulsive density through an iterative eigenvalue equation is still much more efficient than
full minimization for the whole set of natural orbitals. This computational efficiency
represents a
big advantage of local RDMFT over standard RDMFT allowing for the first time
relatively large systems to be in the capability of RDMFT.
The application of local RDMFT to larger systems gives very promising results\cite{our_2}.
Finally, one advantage of this method is that it can
be easily implemented in standard electronic structure codes
due to its similarity with the OEP method.
Finally, with regard to
the efficiency of our scheme, our target
is not to replace well-established methods in routine calculations where
these methods are known to be accurate but to complement them
and improve over their results where they deviate from experiment.
Such cases are for instance bond breaking and highly correlated systems where static
correlations are important. At the same time, in our scheme, the orbital energies from the
effective Hamiltonian offer improved spectral properties.
\section{Applications \label{sec:applications}}
In Fig.~\ref{fig:allpots}, we show the effective local potential for a Ne
atom employing two
RDMFT approximations, the M\"uller\cite{M1984,bb0} and
BBC3\cite{GPB2005} functionals. As we see, the optimal potentials are similar
to the exact-exchange OEP (x-OEP), especially for BBC3 while local M\"uller
is closer to the exact KS\cite{exact_KS} potential.
The comparison with the local potentials
of different theories and the exact KS scheme are useful since
a reasonable approximate potential resembling the exact KS potential will
hopefully lead to a reasonable single-particle spectrum.
\begin{figure*}
\vspace{0.4cm}
\begin{center}
\begin{tabular}{cc}
\includegraphics [width=8cm]{Ne_pot.jpg} \ \ \ \ \ & \ \ \ \ \
\includegraphics [width=8cm]{acetylene_RDMFT_1.jpg} \\
\end{tabular}
\end{center}
\caption{The xc part of the optimal local potential, $v_{\rm rep}$
for the Ne atom (left) and
acetylene molecule (right)
using local RDMFT. For Ne,
the exact KS potential\cite{exact_KS}
as well as the LDA and x-OEP potentials are also shown.
\label{fig:allpots}}
\end{figure*}
\begin{figure}
\vspace{0.4cm}
\begin{center}
\includegraphics[width=8.0cm, clip]{H2_N2.jpg}
\end{center}
\caption{(Color online) Total energy versus internuclear separation for H$_2$ and N$_2$ molecules.
The HOMO-LUMO energy difference, $\Delta E$, for H$_2$, is also shown in the inset.
Energies for N$_2$ are
shifted vertically to match at the equilibrium distance. CI and experimental\cite{N2dis} results were used as
reference for H$_2$ and N$_2$, respectively.}
\label{fig:dis}
\end{figure}
\begin{table}
\setlength{\tabcolsep}{8pt}
\begin{tabular}{lrccc}
& Method & $R_0$ & $D_e$ (Ha) & $\omega_0$ (cm$^{-1}$) \\ \hline
H$_2$ & Local-M\"uller & 0.76 & 0.134 & 4344 \\
& M\"uller & 0.76 & 0.143 & 4134 \\
& Local power & 0.74 & 0.197 & 4646 \\
& Power & 0.75 & 0.200 & 4277 \\
& Local BBC3 & 0.75 & 0.155 & 4328 \\
& BBC3 & 0.75 & 0.156 & 4432 \\
& CI & 0.74 & 0.172 & 4374 \\ \hline
N$_2$ & Local power & 1.10 & 0.200 & 2150 \\
& Local BBC3 & 1.08 & 0.436 & 2500 \\
& Expt. & 1.09 & 0.36 & 2360 \\ \hline
\end{tabular}
\caption{\label{tab:spec}
Equilibrium distances $R_0$, binding energies $D_e$, and vibrational frequencies $\omega_0$,
for H$_2$ and N$_2$ molecules
compared with CI calculations (using the same basis set), and experiment~\cite{N2dis} respectively.
For H$_2$, the full-minimization RDMFT results are also given for comparison.}
\end{table}
An important advantage of many RDMFT functionals is the qualitatively correct dissociation
of small molecules, in contrast to available xc density functionals.
One example is the H$_2$ molecule and, as we show in
Fig.~\ref{fig:dis}(top), this property holds for local RDMFT as well:
for all three functionals, the M\"uller, BBC3, and the power functionals,
local RDMFT reproduces the correct dissociation. The
description at the equilibrium distance also agrees very well with the
configuration interaction (CI) results~\cite{refdata,g09}
both for the position of the minimum and
the curvature as can be seen in Table~\ref{tab:spec}.
In the inset of Fig.~\ref{fig:dis}(top), we show
the difference between the energy eigenvalues corresponding to the HOMO and the LUMO,
as they are defined by the number of electrons and the assumption of a
single-electron picture. We find the same behavior as the one found for
various DFT functionals in Ref.~\cite{CRHRRRS2013}, i.e., the energy difference
is approaching zero at small distances between the two hydrogen atoms.
In Fig.~\ref{fig:dis}(bottom), the dissociation curves for the triple bond of the N$_2$
molecule are shown for BBC3 and power functionals in comparison to HF and a DFT
calculation~\cite{refdata,g09} using the B3LYP functional~\cite{B1993}. Unfortunately, the M\"uller
functional fails to describe the dissociation of N$_2$. The binding energy
as well as the vibrational frequency obtained with BBC3 are closer
to the experimental values than those given by the power functional. Interestingly local RDMFT results are
qualitatively much better than many density functional
approximations including B3LYP, as seen in Fig.~\ref{fig:dis}(bottom). It is worth mentioning
that the correct dissociation of N$_2$ is not reproduced even at the level of accurate quantum
chemistry methods like MP2 and single-reference coupled cluster.
\begin{table*}
\begin{center}
\begin{tabular}{ll|ccccccccc}
System & & HF & M\"uller & GU & BBC3 & AC3 & Power & PNOF1 & ML & Expt. IP \\
\hline\hline
He & IP & 24.970 & 24.69 & 24.87 & 24.57 & 24.84 & 24.84 & 24.88 & 25.15 & 24.59 \\
& $E_c$ & 0 & 0.77 & 0.75 & 0.74 & 0.75 & 0.68 & 0.75 & 0.79 & \\
\hline
Be & IP & 5.60 & 9.51 & 8.46 & 8.73 & 8.46 & 8.58 & 8.44 & 8.55 & 9.32 \\
&$E_c$ & 0 & 0.87 & 0.89 & 0.90 & 0.90 & 0.91 & 0.90 & 0.94 & \\
\hline
Ne & IP1 & 23.01 & 22.90 & 21.32 & 20.92 & 20.88 & 21.65 & 20.91 & 21.32 & 21.60 \\
& IP2 & 52.45 & 46.52 & 45.02 & 44.67 & 44.59 & 45.38 & 44.62 & 45.04 & 48.47 \\
&$E_c$ & 0 & 0.75 & 0.73 & 0.70 & 0.72 & 0.65 & 0.73 & 0.84 & \\
\hline
H$_2$ & IP & 16.17 & 16.24 & 16.19 & 16.15 & 16.15 & 16.13 & 16.18 & 16.28 & 15.43 \\
&$E_c$ & 0 & 0.68 & 0.67 & 0.62 & 0.65 & 0.57 & 0.65 & 0.73 & \\
\hline
H$_2$O &IP1 & 13.73 & 12.59 & 12.03 & 12.35 & 12.06 & 12.10 & 12.12 & 12.64 & 12.78 \\
& IP2 & 15.71 & 14.21 & 14.09 & 14.42 & 14.16 & 14.06 & 14.20 & 14.75 & 14.83 \\
& IP3 & 19.15 & 17.52 & 17.57 & 17.88 & 17.63 & 17.45 & 17.67 & 18.23 & 18.72 \\
& $E_c$ & 0 & 0.66 & 0.61 & 0.61 & 0.56 & 0.53 & 0.57 & 0.73 & \\
\hline
NH$_3$ & IP1 & 11.64 & 11.03 & 10.52 & 10.65 & 10.53 & 10.74 & 10.56 & 10.95 & 10.80 \\
& IP2 & 16.93 & 15.22 & 15.36 & 15.48 & 15.42 & 15.39 & 15.45 & 15.88 & 16.80 \\
& $E_c$ & 0 & 0.65 & 0.59 & 0.59 & 0.54 & 0.53 & 0.55 & 0.72 & \\
\hline
CH$_4$ & IP1 & 14.82 & 13.55 & 13.47 & 13.72 & 13.41 & 13.43 & 13.54 & 13.84 & 13.60,14.40 \\
& IP2 & 25.65 & 21.34 & 21.16 & 21.52 & 21.20 & 21.21 & 21.32 & 21.62 & 23.00 \\
& $E_c$ & 0 & 0.62 & 0.64 & 0.55 & 0.55 & 0.46 & 0.55 & 0.71 & \\
\hline
C$_2$H$_2$& IP1 & 11.07 & 11.67 & 11.01 & 11.12 & 11.20 & 11.46 & 11.31 & 11.59 & 11.49 \\
& IP2 & 18.47 & 16.15 & 15.98 & 16.25 & 16.29 & 16.37 & 16.39 & 16.78 & 16.70 \\
& IP3 & 20.88 & 17.88 & 17.76 & 18.02 & 18.07 & 18.13 & 18.19 & 18.52 & 18.70 \\
& $E_c$ & 0 & 0.72 & 0.66 & 0.64 & 0.64 & 0.62 & 0.65 & 0.79 & \\
\hline
C$_2$H$_4$&IP1 & 10.24 & 10.68 & 10.43 & 10.45 & 10.61 & 10.47 & 10.59 & 10.90 & 10.68 \\
& IP2 & 13.76 & 12.07 & 12.11 & 12.41 & 12.47 & 12.15 & 12.61 & 12.87 & 12.80 \\
& $E_c$ & 0 & 0.68 & 0.63 & 0.55 & 0.59 & 0.57 & 0.60 & 0.75 & \\
\hline
CO$_2$ & IP1 & 14.74 & 13.81 & 13.24 & 13.67 & 13.41 & 13.30 & 13.89 & 14.42 & 13.78 \\
& IP2 & 19.21903 & 16.95 & 16.48 & 16.93 & 16.84 & 16.51 & 17.16 & 17.57 & 18.30 \\
& $E_c$ & 0 & 0.78 & 0.74 & 0.69 & 0.64 & 0.72 & 0.73 & 0.82 & \\
\hline
\multicolumn{2}{l|}{$\Delta_{\rm all} $}& 7.85 & 4.06 & 5.34 & 4.01 & 4.73 & 4.56 & 4.25 & 3.11 & \\
\multicolumn{2}{l|}{$\Delta_{\rm 1st}$}& 8.57 & 2.50 & 4.20 & 2.98 & 3.91 & 3.21 & 3.52 & 3.14 & \\
\multicolumn{2}{l|}{ $\bar E_c$ } & - & 0.72 & 0.69 & 0.66 & 0.65 & 0.63 & 0.67 & 0.78 & \\
\hline
\end{tabular}
\end{center}
\caption{\label{tab:IPs} Ionization potentials (in eV) for atoms and small molecules compared to experiment,
and the ratio, $E_c$, of correlation energy captured by the local approximation over that of
a full RDMFT minimization for several functionals. $\Delta_{\rm all} $, $\Delta_{\rm 1st}$, are the average absolute errors for all IPs and for
the first IP, respectively, defined as $\Delta =100\times(1/N)\sum_i |(x_i - x_i^{\rm ref})/x_i^{\rm ref}|$,
and $\bar E_c$ is the average of $E_c$ over the whole set of systems. Vertical experimental IPs in the last column are obtained from the
NIST Chemistry WebBook\cite{NIST} and references in Ref.~\cite{ZSDGL2012}.}
\end{table*}
We now focus on
the single-electron spectrum of
the local potential Hamiltonian. For this
purpose, we obtained IPs as the corresponding eigenvalues of the local potential
Hamiltonian for a test set of atomic and molecular systems and basis sets.
This set comprises small atomic and molecular systems
using the cc-pVTZ and uncontracted cc-pVTZ basis sets for the orbital and
the auxiliary basis respectively, and we obtained IPs up to the third one.
Our numerical results for several RDMFT functionals are shown in Table~\ref{tab:IPs} where
on the bottom we also show the average,
absolute, percentage error of IPs.
The same errors for the
IPs obtained as the energy eigenvalues of Hartree-Fock (Koopmans' theorem) are also included for comparison.
We find a remarkable agreement between the energy eigenvalues and the
experimental IPs for the functionals we tested.
All errors are below or around 5\%, with the Marques-Lathiotakis (ML) functional\cite{pade}
going as low as $\sim$3\%.
The agreement with experiment is even better for the first IPs
with the M\"uller functional being the most accurate with an error of only $\sim$2\%.
Overall the agreement with experimental values is very good and substantially
better than the HF Koopmans' theorem.
To estimate the effect of the local potential approximation we include in Table~\ref{tab:IPs}
the ratio $E_c$ of the correlation energies (defined as the energy differences from Hartree-Fock)
with local RDMFT over those provided by the full RDMFT minimization.
The average $E_c$ over the whole set of systems, included in the last row of Table~\ref{tab:IPs},
is in the range 0.6-0.8\% for all functionals considered.
However, as in DFT calculations, the comparison of the obtained correlation energy
to the exact one
is not the only decisive factor to assess the accuracy of an approximation
in reproducing many properties.
For the M\"uller functional, the local RDMFT
recovers on average 71\% of the full minimization correlation energy.
The M\"uller functional generally
overestimates the correlation energy substantially in the full minimization and
the constraint of the local potential
offers an improvement. For more accurate functionals, however, this is not always the case.
\section{Conclusion\label{conclusion}}
In conclusion, we presented a method on how to incorporate static correlation into
KS-like equations by
employing xc functionals from RDMFT.
We have shown that when the xc energy is a density functional then the total energy minimization
leads to an idempotent solution (Appendix). Consequently, we relaxed the requirement that the
potential must be the functional derivative of the energy with respect to the density and
decided to minimize the total energy with respect to the
occupation numbers and the ANOs generated by a local effective potential.
In this way,
we manage to describe the dissociation of diatomic molecules accurately.
In addition, our approach allows us to connect a single-particle
energy spectrum to the ANOs. This spectrum is in good agreement with experimental IPs and
photoelectron spectra for molecules.
The proposed method provides a powerful tool which opens a new avenue:
physically motivated approximations in density-matrix based
schemes, able to cope with strongly correlated systems\cite{sharma08} and static correlation, can
now be brought to the realm of DFT. The resulting KS-like approach is able to
simultaneously describe ground-state properties, bond breaking, and single-electron spectral properties.
The scaling of our method can be easily reduced to that of hybrid DFT methods using standard
techniques. The improved computational efficiency compared to full RDMFT minimization allows
for the application
to large systems opening the road for an improved description of electronic
correlations in technologically important molecular systems.
\begin{acknowledgments}
N.N.L. acknowledges financial support from the GSRT, Greece, Polynano-Kripis project
(Grant No. 447963), N.H. from a DFG Emmy-Noether grant, and AR from the European Community’s
FP7 through the CRONOS project, grant agree- ment no. 280879; the European Research Council
Advanced Grant DYNamo (ERC-2010-AdG-267374); Grupos Consolidados UPV/EHU del Gobierno Vasco
(Grant: IT578-13). N.I.G. thanks Professor Mel Levy for helpful comments.
\end{acknowledgments}
|
\section{Introduction}
Generalised complex geometry~\cite{Hitchin,Gualtieri} provides a geometric description of string theory that makes its symmetries manifest. The generalised tangent bundle incorporates information about the space-time geometry and the topology of the $B$-field and so is well-suited to describe the on-shell supersymmetry~\cite{GMPT} and give a handle on understanding non-local symmetries such as T-duality~\cite{GMPW}. In fact generalised geometry provides a compact description of the full off-shell type II supergravity theories using a generalised version of the Levi--Civita connection that encodes the geometry of NSNS fields. The geometrical description of the bosonic theory was first considered in~\cite{Siegel,HK,JLP} using the language of Double Field Theory (DFT)~\cite{dft}. The full theory with fermions was given using generalised geometry in~\cite{CSW1} (see also~\cite{Jeon:2012hp} for a subsequent DFT version, given to all orders in fermions). Suitable generalisations of the basic construction, involving the exceptional groups $\mathit{E}_{d(d)}$, describe all massless fields in type II string theory and M-theory in $d-1$ and $d$ dimensions respectively~\cite{Hull:2007zu,Pacheco:2008ps}. The dynamics are again encoded by the corresponding generalised Levi--Civita connections~\cite{CSW-11d}. That this geometry also describes the full ten- or eleven-dimensional theory was recently shown in~\cite{GGHNS} using the ideas of~\cite{dWN} and~\cite{EFT}. So far the generalised geometrical constructions have been classical.
In this paper we shall address two natural questions concerning the extensions of the formalism. The first question is whether it is possible to incorporate the $\alpha'$ expansion in the framework of generalised geometry.\footnote{We shall not speak about the string loops and the $g_s$ expansion here.} If so, does the result agree with the known string theoretic corrections? The second question may be thought of as a particular case of the first. Is there a generalised geometric description of the heterotic strings? The problem here is that the closure of the three-form $H$ which, as will be reviewed in sec.~\ref{sec:GG}, is essential for the standard construction of \cite{Hitchin,Gualtieri} is no longer satisfied.\footnote{In type II theories we can think of $H$ as a curvature and of the $B$-field as a connection on a gerbe. For the heterotic string this is not longer true and leads to a number of subtleties explored in particular in~\cite{Witten:1999eg}.}
Our conclusion will be that to correctly describe corrections generically one needs to extend the generalised tangent space. There can be important topological restrictions arising from the construction of the extended tangent space $E$. The underlying supergravity degrees of freedom and dynamics arise from choosing the appropriate sub-structures on $E$ and considering compatible connections. The action takes a universal form which arises as a generalised Lichnerowitz--Bismut theorem~\cite{BisTor}. A key new point here is that we naturally obtain the correct gravitational connection in the Bianchi identity. The $\alpha'$-corrected generalised geometry in the heterotic string context has previously been considered in~\cite{GF,Baraglia}. A similar construction describing heterotic $\alpha'$-corrections in the DFT formalism was presented in recent work~\cite{Bedoya:2014pma}. These results can also be compared with~\cite{Hohm:2013jaa}, which takes a different approach in the context of DFT by perturbing the generalised Lie derivative and the $O(n,n)$ metric to include higher-derivative corrections. However, in the latter approach the extra terms are not manifestly diffeomorphism invariant and, in addition, different generalised tensors receive corrections at different orders in $\alpha'$. As we discuss in section~\ref{sec:qb!}, the correct geometrical framework to understand such corrections is as lifts into the generalised tangent spaces described here.\footnote{While this paper was being completed a discussion of further corrections to the generalised Lie derivative for bosonic string theory appeared~\cite{HZ-new}, now also depending on the dynamical fields. We will not comment on the relation to the bosonic theory here.}
The heterotic Bianchi identity provides a natural starting point. It reads schematically
\begin{equation} \label{eq:hetBianchi}
\mathrm{d} H = \tfrac{1}{4 } \alpha' (\tr F\wedge F - \tr R\wedge R) \ ,
\end{equation}
where the first term is built out of the curvature two-form $R$ and the second of the gauge field strength $F$
of either the gauge group $G = E_8 \times E_8$ or $G= SO(32)$. The word schematic refers to the fact that we have not specified yet what connections are used in \eqref{eq:hetBianchi}. A change of connection will result in a shift on the right-hand side by an exact form. In a number of applications, e.g.\ when determining the consistency of heterotic compactifications, one is interested in the integrated form of \eqref{eq:hetBianchi}, where the choice of connection is immaterial. On the other hand, the manifest $(0,1)$ world-sheet supersymmerty together with the choice of a hermitian covariant metric in space-time singles out the (torsionful) connection
\begin{equation}\label{eq:Omegaminus}
\Omega^{-} = \omega^{{\rm LC}} - \tfrac12 H \ .
\end{equation}
Here $\omega^{{\rm LC}}$ is the Levi-Civita connection and $H$ is understood as the $\mathfrak{o}(n)$-valued one-form $H_{ab} = \iota_{\hat e^a} \iota_{\hat e^b} H$. A priori, there is a freedom in choosing the sign of $H$, but this choice is correlated with the signs in the supersymmetry transformations.
The two terms in \eqref{eq:hetBianchi} present one with different types of problems. The gauge part comes with new degrees of freedom and an enlarged local symmetry. So it is clear from the beginning that new ingredients need to be added. For an Abelian gauge group the situation is well-understood. As we shall review in detail, by considering a backgrounds with a set of commuting isometries and dimensional reduction of an $SO(n+d,n+d)$ structure, one gets the Abelian version of~\eqref{eq:hetBianchi} with two set of $U(1)^d$ fields.\footnote{We shall denote the dimension of the space-time by $n$ (typically $n=10$ or rather $n=(1,9)$). When we talk about the backgrounds with isometries or dimensional reductions the number of compact dimensions will be denoted by $d$.} More generally one can reduce on a general group manifold and obtain the non-Abelian versions of~\eqref{eq:hetBianchi}. A closely related notion, $B_n$ generalised geometry on a bundle $TM \oplus \mathbb{R}\oplus T^*M$ is discussed in~\cite{Hitchin:Bn,Bn,Rubio}. Incorporation of the non-Abelian degrees of freedom requires a new extension of the generalised tangent bundle. Locally $E\simeq TM\oplus \adj{P_G} \oplus T^*M$, where $\adj{P_G}$ is the vector bundle for the adjoint representation of a Lie group $G$, that is, with fibres that are the Lie algebra $\mathfrak{g}$ of $G$. Such structures are known in the mathematics literature as transitive Courant algebroids~\cite{severa,csx}, and can always be constructed as dimensional reductions of conventional generalised geometry on a group manifold.\footnote{Heterotic flux (torsionful) backgrounds provide a natural context for applying these constructions \cite{ags, xos} The role of torsionful connections \eqref{eq:Omegaminus} in heterotic backgrounds has been discussed in \cite{mms}.}. We refer to section \ref{sec:nonAb} for explanations and definitions. These structures were used in~\cite{GF,Baraglia} to construct heterotic actions. On the physics side, constructing heterotic theories by reduction has been discussed in~\cite{dnpw,David}, the form of the generalised Lie derivative and the local generalised geometry in the DFT context is given in~\cite{DFT-het}, and in~\cite{GM} it is shown how such generalised structures via dimensional reduction. There are global constraints associated with the extensions of the generalised tangent bundle (to be reviewed shortly) but as far as the consistency of the construction goes there are no restrictions on the choice of connection in the $\frac{1}{4} \alpha' \tr F\wedge F$ factor.\footnote{As will be discussed in subsection \ref{sec:redgauge} the connection may be restricted by extra requirements, such as T-duality covariance.}
On the contrary, the gravitational contribution to \eqref{eq:hetBianchi} does not enlarge the symmetry group. It is tempting and in many ways natural to consider $GL(n,\mathbb{R})$ as a special case of the gauge group (see e.g. \cite{dnpw}), so that together with the gauge part one is considering a group $G\times\GL(n,\mathbb{R})$. By doing so one can correctly reproduce the heterotic Bianchi identity. Such an approach was used to reproduce the heterotic dynamics using generalised geometry in~\cite{GF,Baraglia}. The problem however is that the new extended generalised tangent bundle is too large and there will be unphysical degrees of freedom. In particular, the $\GL(n,\mathbb{R})$ connection in~\cite{GF,Baraglia} has to be put in by hand. This is reflecting the fact that the $\GL(n,\mathbb{R})$ symmetry is not independent but is the group that rotates bases for the tangent and cotangent spaces. The solution is that when defining the dynamical degrees of freedom we specify a more refined structure. In particular we identify an $O(n)$ sub-bundle of $\GL(n,\mathbb{R})$ and then identify this $O(n)$ structure with one of the $O(n)$ structures defined on the $TM\oplus T^*M$ part of the tangent space. Thus the overall structure group is simply $O(n)\times G \times O(n) $. We then require that the generalised connections which define the dynamics are compatible with the refined structure. We find we must slightly relax the constraint that the generalised connection is torsion-free, and demand instead that all the connections that appear in supergravity equations be unique. Remarkably, this procedure by construction gives the full Bianchi identity \eqref{eq:hetBianchi} with precisely the connection \eqref{eq:Omegaminus}. The generalised geometric construction moreover reproduces the known fact that the supersymmetry transformation of the gravitino should contain a covariant derivative using $ \Omega^{+} = \omega^{\rm LC} + \tfrac12 H$. All this is in accordance with the known string-theoretic calculations, but the discussion of the way the local data enters in the construction of the extended generalised tangent bundle is new.
Finally, we may turn to the effective actions. As shown in \cite{Siegel,HK,CSW1} even though a generalised Riemann tensor is not unambiguously defined by vanishing (generalised) torsion conditions, the generalised Ricci tensor is. Moreover, the Ricci scalar gives the full NSNS part of the effective action. The Ricci scalar can be constructed using the action of the generalised connection on spinor representations~\cite{CSW1}, in a way that is closely related to a generalisation of Lichnerowitz theorem due to Bismut~\cite{BisTor}, which states that there exists a pair of first-order operators, a covariant derivative and a Dirac operator respectively,\footnote{Note that the Dirac operator here is {\sl not} the trace of the covariant derivative. In general, the torsion parts in the respective operators may involve different physical degrees of freedom.} such that the difference of their squares is tensorial. The construction of the extended generalised bundle provides us with the needed first-order operators (corrected up to order $\sim \alpha'$). The order $\alpha'$ effective action $S$ is schematically written as
\begin{equation}
\label{eq:newBismut*}
( D^aD_a + D^{\alpha}D_{\alpha} - (\slash\!\!\!\! D)^2) \epsilon
= \tfrac14 S \epsilon + \gamma^{abcd} I_{abcd} \epsilon \ .
\end{equation}
It will not come as a surprise that $\slash\!\!\!\! D$, $D_a$ ($a=1, \cdots, n$) and $D_{\alpha}$ ($\alpha = 1, \cdots, \mbox{dim}(G)$) are the operators appearing in the supersymmetry variations of the dilatino, gravitino and gaugini respectively. Naively the right-hand side of \eqref{eq:newBismut*} contains not only a non-trivial scalar $S$ but also four-form $I$. However the latter is equal to the Bianchi identity and thus vanishes. In section~\ref{sec:higher} we show that after adding an $\GL(n,\mathbb{R})$ factor and identifying the $O(n)\times G\times O(n)$ structure, we naturally find a composite fermion that corresponds to the gravitino curvature.
Making the variations of the composite field compatible with the variations of the gravitino while preserving the condition that $I$ above vanishes actually induces a whole hierarchy of higher order $\alpha'$ corrections. This is the generalised geometric version of analysis of~\cite{BdR}.
The appearance of a generalised Lichnerowitz--Bismut theorem is not an accident. As we discuss in section~\ref{sec:gen}, it is a generic consequence of supersymmetry. We show that one can construct a general formalism by which generalised geometry can be viewed as an infinite-dimensional version of the embedding tensor formalism~\cite{embed}. The generalised Lichnerowitz--Bismut theorem is then simply a result of a supersymmetric Ward identity~\cite{s-ward,embed-lecture}.
Interestingly, if one tries to apply this same set of ideas to type II theories (this suggestion goes back to \cite{dnpw}) and consider an extended generalised tangent bundle with $G=O(n)\times O(n)$ and then try and identify these with the $O(d)_+$ and $O(d)_-$ acting on $TM\oplus T^*M$, one finds one cannot consistently choose a generalised connection compatible with $O(n) \times O(n)$. As will be explained in detail in Appendix~\ref{sec:typeII} depending on the allowed choices, one either ends up with ambiguities in the effective action or in the Bianchi identity. Hence $\mathrm{d} H=0$ is the only consistent choice. The absence of corrections linear in $\alpha'$ in theories with 32 supercharges is well known, but the exercise is instructive and shows that the outlined procedure is restrictive.
In a slightly different direction, one can take the extended generalised generalised tangent space $E\simeq TM\oplus \adj{P_G} \oplus \adj{P_{\GL(n,\mathbb{R})}}\oplus T^*M$, where we now take $P_{\GL(n,\mathbb{R})}$ to be the frame bundle and try to restrict the symmetry parameters to elements of $TM \oplus \adj{P_G}\oplus T^*M$, that is diffeomorphisms, $B$-field gauge transformations and gauge transformations of the $G$-connections. To do so we need to determine a $\GL(n,\mathbb{R})$ transformation in terms of a diffeomorphism. This requires a lift of vectors, essentially defining the $\GL(n,\mathbb{R})$ transformation via
\begin{equation} \label{eq:relphys}
{\cal L}_v e^a = \tilde \Lambda^a{}_b e^b \ ,
\end{equation}
where $v \in \Gs{TM}$ is the generator of a diffeomorphisms, $\tilde \Lambda \in \Gs{\adj{P_{\GL(n,\mathbb{R})}}}$ and $e^a$ is an arbitrary section of the frame bundle.\footnote{Changing the choice of section $e^a$ corresponds to a gauge transformation in $\tilde G$. Therefore, the choice of $e^a$ fixes a gauge in $\tilde G$.} This states that the $\GL(n,\mathbb{R})$ transformations are generated by the action of diffeomorphisms on the auxiliary frame $e^a$. Note that due to the appearance of the Lie derivative ${\cal L}_v$, \eqref{eq:relphys} does not define a subbundle of the generalised tangent bundle. Rather, it defines a subspace in the space of sections on the original bundle $E$. In section~\ref{sec:qb!} we show that inserting \eqref{eq:relphys} into the natural metric on $E$ and the corresponding generalised Lie derivative leads to terms that look like higher-derivative corrections to the conventional expressions in the unextended case where $E\simeq TM\oplus T^*M$. In fact, when restricted to coordinate frames (and hence written in non-manifestly covariant form), these corrections match the $\alpha'$ corrections recently computed in DFT~\cite{Hohm:2013jaa}. However, in our context, viewing them as corrections is somewhat misleading, and not only because manifest diffeomorphism invariance is then lost. Globally the generalised vectors are sections of the extended generalised tangent space and not the conventional one: the only way to make sense of these expression is in terms of the extended space. Note that related discussions to this point recently appeared in~\cite{Bedoya:2014pma,HZ-het}.
As we see the complete construction requires both global and local considerations. As mentioned, some of the former have appeared in the maths literature, in the context of transitive Courant algebroids, as have aspects of the latter, though without giving a natural explanation for the appearance of $\Omega^-$ in the Bianchi identity. The extension of the construction needed to correctly cover the gravitational contributions to the Bianchi identity and the important appearance of the local data is new. Our results for the effective action reproduce the known $\alpha'$-corrected supersymmetric actions, but the ``all-at-once" construction, based on the generalised Lichnerowitz-Bismut theorem, presented here is also original.
So far we have mostly concentrated on the heterotic Bianchi identity, with type II just providing an illustrative example of how a complete construction may rule out certain corrections. However we think that the main principles/ingredients should apply in a more general discussion as well:
\vskip 0.3cm
\noindent {\bf Extensions $\,\,\,$} The first step is always to extend the generalised tangent bundle. We have discussed here the most intuitive extension using the local gauge symmetries of the theory. However one might expect generically a series of extensions, perhaps most naturally mirroring the local symmetries of the hierarchy of massive states in the corresponding string theory. The extended space always admits a conventional generalised Lie derivative that is first-order in the derivative. In the general formalism this can be regarded as identifying the infinite-dimensional analogue of the embedding tensor. These objects are then the main building blocks of the construction. They in particular determine the possible corrections to the Courant bracket.
\vskip 0.3cm
\noindent {\bf Obstructions $\,\,\,$} The extensions are obstructed. The triviality of the {\sl total} (gauge and tangent bundle) Pontryagin class is well known and has been much used in heterotic compactifications. Yet crucially, $H$ is not a curvature of a gerbe. In the type II case all possible extensions preserve the gerbe nature of the $B$-field, and eventually result in the absence of corrections linear in $\alpha'$. This happens due to subtle cancellations and there is no reason that further (e.g. $(\alpha')^3$) corrections should not appear.
\vskip 0.3cm
\noindent {\bf Connections $\,\,\,$} One may allow for the possibility that the extended generalised tangent bundle may have more symmetry than needed. It is important that the connections are compatible with the correct symmetries. This compatibility may in turn impose restrictions.
\vskip 0.3cm
\noindent {\bf Local actions $\,\,\,$} The generalised Lichnerowitz-Bismut theorem allows the computation of effective actions starting from the extended generalised tangent bundle. The procedure is somewhat different from computing supersymmetric completions. It is important to note that a particular correction to a given order in $\alpha'$ may induce all order corrections in the effective action (as is the case with the heterotic Bianchi identity \cite{BdR}). On the other hand, a family of corrections (a priori all orders in $\alpha'$) in the effective action should trace back to a single extension of the generalised tangent bundle. A particular higher-derivative coupling may belong to different families and be derived from different extensions.
\vskip 0.3cm
\noindent
The paper is organised as follows. After reviewing some facts on generalised geometry in the beginning of section \ref{sec:GG}, we show in subsection \ref{sec:GGBianchi} how its dimensional reductions give rise to a non-trivial Bianchi identity.
Section \ref{sec:nonAb} is concerned with the incorporation of non-Abelian gauge groups in generalised geometry and the construction of the extended generalised tangent bundle. In particular, in subsection \ref{sec:nonAbelian} we discuss how a non-trivial Bianchi identity arises in generalised geometry, and discuss the symmetries in subsection \ref{sec:nonAbelian1}. In subsection \ref{sec:nonAbelian2} we then discuss an analogous construction for the gravitational part of the Bianchi identity and in subsection \ref{sec:redgauge} we discuss the importance of choice of connections for symmetry properties, such as T-duality. The significance of such a choice is further elaborated on in our discussion of the effective actions in section \ref{sec:gravconnection} where we consider the coupling of the $N=1$ ten-dimensional supergravity to Yang-Mills. We start by identifying the physical degrees of freedom and by constructing the generalised connections in subsection \ref{sec:localO}, which are then used to write down the action and equations of motion in subsection \ref{sec:sugraEQ}. Section \ref{sec:het-corr} is devoted to the $\alpha'$ corrections in heterotic strings. In particular, in subsection \ref{sec:het} we explain why $\Omega^-$ is the necessary (composite) connection for the gravitational part of the Bianchi identity. The equations of motion up to order $\sim \alpha'$ are discussed in subsection \ref{sec:hetEQ}, while the inclusion of higher order $\alpha'$ corrections in the effective action is the subject of subsection \ref{sec:higher}. We also show how the consistency of the extended generalised tangent bundle prevents the appearance of corrections linear in $\alpha'$ in type II theories. For the fluency of presentation, we have placed this discussion in Appendix~\ref{sec:typeII}. Section \ref{sec:qb!} discusses how the construction of the extended generalised tangent bundle reflects on corrections to the Courant bracket. In section \ref{sec:gen} we show how to construct a general formalism by which generalised geometry can be viewed as an infinite-dimensional version of the embedding tensor formalism. Finally we conclude with a speculative section \ref{sec:NS5}, where the modifications of the Bianchi identity due to the presence of NS5 sources are discussed.
\section{Basics of generalised geometry}
\label{sec:GG}
We first very briefly review some basic notions in generalised
geometry \cite{Hitchin,Gualtieri}, which can be thought of as giving
a geometric formulation of the NSNS sector of type II theories, namely
the metric $g$, B-field $B$ and dilaton $\phi$. The bosonic symmetries
of the theory are
\begin{equation*}
\begin{aligned}
\text{1. } & \text{diffeomorphisms} \quad
(g,B,\phi) \mapsto (g+{\cal L}_v g, B+{\cal L}_v B, \phi+
{\cal L}_v\phi)\ ,\\
\text{2. } & \text{gauge transformations}
\quad (g,B,\phi) \mapsto (g,B - \mathrm{d}\lambda,\phi) . \ \\
\end{aligned}
\end{equation*}
where ${\cal L}$ is the Lie derivative. The symmetries are therefore
generated by vectors $v$ and one-forms $\lambda$. These are combined
in the generalised tangent bundle into generalised vectors. More
precisely, the generalised tangent
bundle $E$ on a manifold $M$ is defined as a particular exact extension of
$TM$ by $T^*M$
\begin{equation}
\label{eq:Edef}
0 \longrightarrow T^*M \longrightarrow E
\stackrel{\pi}{\longrightarrow} TM \longrightarrow 0 .
\end{equation}
Generalised vectors are sections of $E$. Locally they can be written
as $V=v+\lambda$ or $V=(v,\lambda)$ where $v\in \Gs{TM}$ and $\lambda \in
\Gs{T^*M}$. In going from one coordinate patch $U_i$ to another $U_j$, we
have to first make the usual patching of vectors and one-forms, and
then give a further patching describing how $T^*M$ is fibered over
$TM$ in $E$. The latter is given by
\begin{equation}
\label{eq:EpatchGG}
\begin{aligned}
v_{(i)} &= v_{(j)}\ , \\
\lambda_{(i)} &= \lambda_{(j)}
- \iota_{v_{(j)}}\mathrm{d}\lambda_{(ij)} \ .
\end{aligned}
\end{equation}
This corresponds to the patching
\begin{equation}
\label{eq:Bpatch}
B_{(i)}=B_{(j)} - \mathrm{d} \lambda_{(ij)} \ .
\end{equation}
There is a natural $O(n,n)$ pairing on $E$, given by
\begin{equation} \label{eta}
\langle V , W \rangle
= \tfrac{1}{2}\iota_v \rho + \tfrac{1}{2}\iota_w \lambda \ ,
\quad {\rm where} \ V=(v,\lambda)\ , \ W = (w, \rho) \ .
\end{equation}
The differentiable structure on $E$ is encoded in the generalised Lie
derivative defined as
\begin{equation}\label{eq:genLie}
\mathbb{L}_{V} W = [v,w] + {\cal L}_v \rho - \iota_w \mathrm{d} \lambda \
\end{equation}
This has the Leibniz property
\begin{equation}\label{eq:relLieCourant}
\comm{\mathbb{L}_V}{\mathbb{L}_W}
= \mathbb{L}_{\mathbb{L}_V W} = \mathbb{L}_{\Bgen{V}{W}} \ ,
\end{equation}
where the Courant bracket is defined as the antisymmetrization
\begin{equation}\label{eq:Courant}
\Bgen{V}{W}
= \tfrac{1}{2}\left( \mathbb{L}_V W - \mathbb{L}_W V \right)
= [v,w] + {\cal L}_v \rho - {\cal L}_w \lambda
- \tfrac12 \mathrm{d} (\iota_v \rho -\iota_w \lambda ) \ .
\end{equation}
The inner product (\ref{eta}) is invariant under the generalised Lie derivative, i.e.\
\begin{equation}\label{eq:scalarLie}
\langle \mathbb{L}_{V} W , U \rangle + \langle W , \mathbb{L}_{V} U \rangle = {\cal L}_v \langle W , U \rangle \ .
\end{equation}
In addition to diffeomorphisms, the action $ \lambda \to \lambda -
\iota_v B$ is an automorphism of the generalised Lie derivative if
$\mathrm{d} B = 0$. If however $H=\mathrm{d} B \ne 0$, it transforms
as
\begin{equation} \label{eq:genLietrafo}
\mathbb{L}_{\e^B V} (\e^B W) = \e^B ( \mathbb{L}_{V} W + \iota_v \iota_w H) \ .
\end{equation}
Note that here $H=\mathrm{d} B$, and correspondingly
\begin{equation} \label{BianchitypeII}
\mathrm{d} H=0 \ .
\end{equation}
The NSNS fields define a generalised metric $G$. To include the
dilaton, one considers the weighted generalised tangent space
$\tilde{E}=\det T^*M\otimes E$ which admits a natural action of
$O(n,n)\times\mathbb{R}^+$, the $\mathbb{R}^+$ corresponding to a rescaling of the
weighting. A generalised metric $G$ is an $O(n)\times O(n)$
structure. It defines an isomorphism $\tilde{E}\simeq E$ together with
a splitting $E=C_+\oplus C_-$, with the orthonormal generalised
vielbeins on $C_\pm$ given by
\begin{equation}
\begin{aligned}
\label{eq:OdOdframe}
\hat{E}^+_a &=
\mathrm{e}^{-2\phi}\sqrt{-g}\,\big(
\hat{e}^+_a + e^+_a + i_{\hat{e}^+_a}B \big) , \\
\hat{E}^-_{\bar{a}} &=
\mathrm{e}^{-2\phi}\sqrt{-g}\,\big(
\hat{e}^-_{\bar{a}} -e^-_{\bar{a}} +
i_{\hat{e}^-_{\bar{a}}}B \big) ,
\end{aligned}
\end{equation}
where $\hat{e}^{\pm}$ are orthonormal bases for $g$ in $TM$ and $e^{\pm}$
their duals. The two $O(n)$ groups act separately on $\hat{E}^+$ and
$\hat{E}^-$ and
\begin{align}
\GM{\hat{E}^+_a}{\hat{E}^+_b} &= \norm{\vol_G}^2\delta_{ab} , &
\GM{\hat{E}^-_{\bar{a}}}{\hat{E}^-_{\bar{b}}}
&= -\norm{\vol_G}^2\delta_{\bar{a}\bar{b}} , &
\GM{\hat{E}^+_a}{\hat{E}^-_{\bar{b}}} &= 0 ,
\end{align}
where $\norm{\vol_G}=\mathrm{e}^{-2\phi}\sqrt{-g}$. In the special case where the
two frames are aligned, so $e^+_a=e^-_a=e_a$, it is conventional to
also define generalised vielbeins of the form
\begin{equation} \label{genviel}
\hat{E}_{A} = \mathrm{e}^{-2\phi}\sqrt{-g}\left( \begin{aligned}
e^a \\ \hat e_a + \iota_{e_a} B
\end{aligned} \right) \ .
\end{equation}
such that
\begin{equation}
\GM{\hat{E}_A}{\hat{E}_B}
= \frac12 \norm{\vol_G}^2
\begin{pmatrix} 0 & \mathbf{1} \\ \mathbf{1} & 0 \end{pmatrix}
\end{equation}
In the next section we will see how by dimensional reduction we obtain
a non-zero right-hand side to (\ref{BianchitypeII}). We will do the
simple case of a circle reduction, which is enough to understand the
idea, and can be somewhat easily generalised to tori reductions.
\subsection{Circle reductions and Bianchi identities}
\label{sec:GGBianchi}
From a dimensional reduction we can obtain a right-hand side to the
Bianchi identity of the form given by the contribution from (an
Abelian) gauge bundle in the heterotic theory. This will allow us to
get some intuition on the concepts introduced in the next section.
Consider an $(n+1)$-dimensional manifold $\tilde{M}$ that is a circle
fibration with base $M$. The generalised geometry on $\tilde{M}$ has
an $O(n+1,n+1)$ structure invariant under a $U(1)$ action generated by a
vector $\partial/\partial_{\varphi}$, where locally $\varphi$ parametrises the
circle fibration. If we require all the fields to be independent of
$\varphi$, the metric on such a background takes the form
\begin{equation}
\mathrm{d} s^2 = \mathrm{d} s^2_{n} + \e^\rho (\mathrm{d} \varphi + C_1)^2 \ ,
\end{equation}
where
\begin{equation}
G_2 = \mathrm{d} C_1
\end{equation}
gives the field strength of the fibration. Moreover, we can also
decompose the field strength $H$ of the $B$-field as
\begin{equation}\label{eq:Hdecomp}
H = H_3 + H_2 \wedge (\mathrm{d} \varphi + C_1) \ ,
\end{equation}
where $\iota_{\partial \varphi} H_3=0$.
If we started with a trivial Bianchi identity (\ref{BianchitypeII}) in
$n+1$ dimensions, we find now for the $n$-dimensional fields
\begin{equation}
\mathrm{d} H_3 = - H_2 \wedge G_2 \ , \qquad \mathrm{d} H_2 = 0 \ ,
\end{equation}
i.e.\ a non-trivial Bianchi identity. If we define
\begin{equation}\label{eq:redefpm}
F^\pm = \tfrac12 (H_2 \pm G_2) \ ,
\end{equation}
we end up with the Bianchi identity
\begin{equation}
\label{eq:modBI}
\mathrm{d} H_3 = F^- \wedge F^- - F^+ \wedge F^+ \ ,
\end{equation}
which can be viewed as an Abelian version of \eqref{eq:hetBianchi}.
If we reduce the two-form gauge field $B$ itself, using
\begin{equation}
B = \tilde B_2 + B_1 \wedge (\mathrm{d} \varphi + C_1) \
\end{equation}
we find
\begin{equation}\label{Habelian}
H_3 = \mathrm{d} \tilde B_2 - B_1 \wedge G_2 \ , \qquad H_2=\mathrm{d} B_1 \ .
\end{equation}
Under the redefinition \eqref{eq:redefpm} the first equality becomes
\begin{equation} \label{H3}
H_3 = \mathrm{d} B_2 + A^- \wedge \mathrm{d} A^- - A^+ \wedge \mathrm{d} A^+ \ ,
\end{equation}
where we defined
\begin{equation}
B_2 = \tilde B_2 + A^- \wedge A^+ \ , \qquad A^\pm = \tfrac12 (B_1 \pm C_1)
\end{equation}
and we have $F^\pm = \mathrm{d} A^\pm$.
The generalised vielbein (\ref{genviel}) now reads
\begin{equation}
E_{A} = \left( \begin{aligned}
e^a \\ \e^\rho (\mathrm{d} \varphi + C_1) \\ \hat e_a + \iota_{e_a} \tilde B_2 \\ \e^{-\rho} (\partial_\varphi + B_1)
\end{aligned} \right) \ .
\end{equation}
Note that T-duality along $\varphi$ just amounts to an exchange of
$G_2$ and $H_2$ (and of course an invertion $\rho\to - \rho$). In
other words it maps $A^\pm \to \pm A^\pm$, while $H_3$ and $B_2$ are
invariant.
Recall that the $B$-field must be patched as in~\eqref{eq:Bpatch}. On
triple overlaps one has the gerbe structure
\begin{equation}
\Lambda_{(ij)} + \Lambda_{(jk)} + \Lambda_{(ki)}
= \mathrm{d} \Lambda_{ijk}
\end{equation}
with $g_{ijk}=\exp(4\pi\alpha'\mathrm{i}\Lambda_{ijk})$ satisfying
$g_{jkl}g^{-1}_{ikl}g_{ijl}g_{ijk}^{-1}=1$ so that $H$ is
quantised. Assuming all the gauge transformation parameters are
independent of $\varphi$ it is easy to see that this implies that the
$B_1$ bundle is trivial, and hence the Pontryagin classes $[F^-\wedge
F^-]$ and $[F^+\wedge F^+]$ are equal in cohomology, as required by the form
of~\eqref{eq:modBI}.
Forgetting about the circle as a physical direction, we can view the
above construction as $n$-dimensional generalised geometry on $M$ with two
$U(1)$ gauge groups. The symmetries are now
\begin{equation*}
\begin{aligned}
\text{1. } & \text{diffeomorphisms} \ ,\\
\text{2. } & \text{one-form transformations}
\quad B \mapsto B - \mathrm{d}\lambda \ ,\\
\text{3. } & \text{gauge transformations} \quad (A^+,A^-) \mapsto (A^+ + \mathrm{d}\Lambda^+,A^- + \mathrm{d}\Lambda^-) \ .
\end{aligned}
\end{equation*}
Geometrically, the bundle in consideration becomes $TM \oplus
\mathbb{R} \oplus \mathbb{R}\oplus T^*M$, the two $\mathbb{R}$ factors corresponding to
the adjoint representation of the $U(1)$ groups, and there is a
scalar product
\begin{equation}
\label{eq:modscalar}
\langle V, W \rangle = \tfrac12\iota_v \sigma + \tfrac12\iota_w \lambda + \Lambda^+ \Sigma^+ - \Lambda^- \Sigma^- \ ,
\end{equation}
where we defined $V= (v, \Lambda^+, \Lambda^-,\lambda)$ and $W = (w, \Sigma^+, \Sigma^-, \sigma)$.
Furthermore, the generalised vielbein is given by
\begin{equation}
E_A^M = \left( \begin{aligned}
e^a_m \\ f^+ + f^+ A^+_m \\ f^- + f^- A^-_m \\ \hat e^m_a + (\iota_{e_a} \tilde B_2)_m
\end{aligned} \right) \ .
\end{equation}
where $f^\pm$ are some non-vanishing functions giving a basis for the
two $\mathbb{R}$ bundles.
We find a generalised geometry whose Bianchi identity is sourced by the first Pontryagin class of two $U(1)$ bundles. Similarly, one can also discuss the Bianchi identity with only one $U(1)$ bundle, giving rise to $B_n$ generalised geometry on a bundle $TM \oplus \mathbb{R}\oplus T^*M$ \cite{Hitchin:Bn,Bn,Rubio}. In section~\ref{sec:nonAbelian} we will discuss non-Abelian gauge groups in generalised geometry by generalizing the bundle construction employed here to the non-Abelian case.
Let us now consider the generalised Lie derivative \eqref{eq:genLie}. After the circle reduction (assuming that the generalised vectors do not depend on $\varphi$) we find that the generalised Lie derivative splits into
\begin{equation}\label{eq:genLieAbel}
\mathbb{L}_V W = {\cal L}_v w + {\cal L}_v \rho - \iota_w \mathrm{d} \lambda + 2 \Sigma^+ \mathrm{d} \Lambda^+ - 2 \Sigma^- \mathrm{d} \Lambda^- + {\cal L}_v \Sigma^+ - {\cal L}_w \Lambda^+ + {\cal L}_v \Sigma^- - {\cal L}_w \Lambda^- \ .
\end{equation}
Notably, the one-form transformation has a non-trivial contribution from gauge transformations.
Note that the scalar product \eqref{eq:modscalar} obeys the relation
\eqref{eq:scalarLie} with this generalised Lie derivative
\eqref{eq:genLieAbel}. The Courant bracket then gets modified to
\begin{equation}\begin{aligned}
\Bgen{V}{W} = & [v,w] + {\cal L}_v \rho - {\cal L}_w \lambda - \tfrac12 \mathrm{d} (\iota_v \rho -\iota_w \lambda) + \Sigma^+ \mathrm{d} \Lambda^+ - \Lambda^+ \mathrm{d} \Sigma^+ - \Sigma^- \mathrm{d} \Lambda^- + \Lambda^- \mathrm{d} \Sigma^-\\&
+ {\cal L}_v \Sigma^+ - {\cal L}_w \Lambda^+ + {\cal L}_v \Sigma^- - {\cal L}_w \Lambda^- \ .
\end{aligned}\end{equation}
Before dimensional reduction, shifts by $B$ were automorphisms of the Courant bracket if $\mathrm{d} B=0$. This together with shifts in $A_1$ such that $F_2= \mathrm{d} A_1 = 0$ give the automorphisms of \eqref{eq:genLieAbel}, that are transformations induced by $B_2$ and $A^\pm$,
\begin{equation} \label{patchingabelian} \begin{aligned}
B_2\ :\ & (v,\lambda) \to (v,\lambda - \iota_v B_2) \ ,\\
& \Lambda^\pm \to \Lambda^\pm , \\
A^\pm \ : \ & (v,\lambda) \to (v,\lambda \pm 2 \Lambda^\pm A^\pm \mp (\iota_v A^\pm) A^\pm) \ , \\
& \Lambda^\pm \to \Lambda^\pm - \iota_v A^\pm \ .
\end{aligned}\end{equation}
From \eqref{eq:genLietrafo} we can also compute the transformation behavior if $B_2$ and $A$ are not closed, that is
\begin{equation} \label{twisting1} \begin{aligned}
\mathbb{L}_{\e^{B_2} V} (\e^{B_2} W) = & \e^{B_2} \left(\mathbb{L}_{V} W + \iota_v \iota_w \mathrm{d} B_2 \right) \ , \\
\mathbb{L}_{\e^{A^+}V} (\e^{A^+} W) = & \e^{A^+} \left( \mathbb{L}_{V} W + 2 (\Sigma^+ \iota_v - \Lambda^+ \iota_w)F^+ -\iota_v \iota_w (A^+ \wedge \mathrm{d} A^+) + \iota_v \iota_w (\mathrm{d} A^+)\right) \ , \\
\mathbb{L}_{\e^{A^-}V} (\e^{A^-} W) = & \e^{A^-} \left( \mathbb{L}_{V} W - 2 (\Sigma^- \iota_v - \Lambda^- \iota_w)F^- +\iota_v \iota_w (A^- \wedge \mathrm{d} A^-)+ \iota_v \iota_w (\mathrm{d} A^-) \right) \ .
\end{aligned}\end{equation}
Puting everything together we can defined the twisted generalised Lie derivative
\begin{equation}
\hat \mathbb{L}_{V} W = \mathbb{L}_{V} W + 2 (\Sigma^+ \iota_v - \Lambda^+ \iota_w)F^+ - 2 (\Sigma^- \iota_v - \Lambda^- \iota_w)F^- + \iota_v \iota_w H_3 + \iota_v \iota_w F^+ + \iota_v \iota_w F^-) \ .
\end{equation}
with the correction terms combining into the gauge-invariant field strengths $H_3$ and $F^\pm$.
The simple calculation presented here exemplifies many of the points we will see in the more general set-up.
We shall turn now to considering a non-Abelian gauge group $G$. Formally one may take $G$ to be a product group and have an $GL(n,\mathbb{R})$ or $O(n)$ factor. Trying to identify this $GL(n,\mathbb{R})$ factor with gravitational contributions to the Bianchi identity is associated to a number of complications, the discussion of which is the essential part of this paper.
\section{Non-Abelian gauge groups in generalised geometry}
\label{sec:nonAb}
In this section we will discuss different aspects of incorporating non-Abelian gauge symmetries in generalised geometry. This gives the geometrical basis for describing $N=1$ supergravity coupled to super Yang--Mills theory, and also $\alpha'$ corrections.
We have already seen how Abelian groups can appear from a reduction on a circle. For non-Abelian groups the basic construction is by reduction on a general group manifold. In the mathematics literature, this is known as a ``generalised reduction''~\cite{reduction}. The resulting extended generalised tangent space with the corresponding generalised Lie derivative is known as a ``transitive Courant algebroid''. That such algebroids can be constructed by reduction was first observed by Severa~\cite{severa} (see also~\cite{csx} for a discussion) and first discussed in the generalised geometric context in~\cite{Hitchin:Bn,Bn} and~\cite{Rubio}. It was specifically applied to the heterotic theory in~\cite{GF,Baraglia}. We note that a Courant algebroid is a vector bundle $E$ with a map $\pi:E\to TM$, a Dorfman bracket (or generalised Lie derivative) $\mathbb{L}_VW\in \Gs{E}$ and an (indefinite) metric $\met{U}{V}$, with the properties
\begin{equation}
\label{eq:CA-def}
\begin{aligned}
(1) && & \comm{\mathbb{L}_U}{\mathbb{L}_V}W = \mathbb{L}_{\mathbb{L}_UV}W , \\
(2) && & \met{\mathbb{L}_UV}{W}+\met{V}{\mathbb{L}_UW}
= \iota_{\pi(U)} \mathrm{d}\met{V}{W} , \\
(3) && & \mathbb{L}_VV = 2 \mathrm{d} \met{V}{V} .
\end{aligned}
\end{equation}
Using the metric one can define the dual map $\frac{1}{2}\pi^*:T^*M\to E$, and in the last equation this is used to interpret the one-form on the right-hand side as an element of $E$. The Courant algebroid is transitive if the map $\pi$ is surjective, and is exact if it is transitive and the sequence $T^*M\to E\to TM$ defined by $\pi$ and $\frac{1}{2}\pi^*$ is exact. It is crucially the third condition that means that transitive Courant algebroids encode a non-trivial Bianchi identity for $H$.
In the physics literature, the reduction picture was discussed in~\cite{dnpw,David}, the form of the generalised Lie derivative and the local generalised geometry in the DFT context is given in~\cite{DFT-het}, and in~\cite{GM} it is shown how such structures appear by effectively a generalised reduction.
\subsection{Symmetry algebra and patchings}
\label{sec:nonAbelian}
We saw in the previous section how a reduction on $S^1$ led to a theory with two $U(1)$ gauge fields, which could be described by an extension of conventional generalised geometry, with a generalised Lie derivative given by~\eqref{eq:genLieAbel}. Necessarily the total Pontryagin class of the bundle was zero. We now review how this extends to the non-Abelian case with group $G$. One way to view the structure is again as a reduction from higher dimensions, this time choosing a space which is a fibration with fibre $G$, that is a principle bundle.\footnote{The $S^1$ reduction gave $U(1)^2$ with a trivial $U(1)$ factor from the reduction of the $B$-field gerbe. In the following, we will simply keep a single $G$ factor.} However, one can also just construct the generalised geometry from first principles using the non-Abelian analogues of the objects defined in the previous section, or equivalently, the symmetries of $N=1$ supergravity coupled to non-Abelian gauge theory. For simplicity this is the approach we will use. Again we will see that the total Pontryagin class of the gauge bundle has to vanish. This is encoded in the Bianchi identity of the form
\begin{equation} \label{eq:Bianchi_gauge}
\mathrm{d} H_3 = \tr F \wedge F\ ,
\end{equation}
where $F=\mathrm{d} A + A \wedge A$ is now the field strength of some non-Abelian gauge field $A$. This will be the starting point of the construction.
The non-Abelian version of (\ref{H3}) realizing \eqref{eq:Bianchi_gauge} is
\begin{equation}
H_3 = \mathrm{d} B_2 + \omega_3(A) \ ,
\end{equation}
where the Chern--Simons three-form $\omega_3(A)$ is given by
\begin{equation}\label{eq:CS}
\omega_3(A) = \tr\left( A\wedge F - \tfrac{1}{3}A^3 \right) \ .
\end{equation}
Under a finite gauge transformation
\begin{equation} \label{finitegaugetr}
A\mapsto A'=gAg^{-1}+ g\mathrm{d} g^{-1} \ ,
\end{equation}
one finds
\begin{equation}
\omega_3(A') = \omega_3(A) + \mathrm{d}\tr\left(g^{-1}\mathrm{d} g\wedge A\right)
- \tfrac{1}{3}\tr\left(g\mathrm{d} g^{-1}\right)^3 \ .
\end{equation}
The final term is closed,\footnote{Actually it must be since
$\mathrm{d}\omega_3(A)=\tr F\wedge F$ is a gauge invariant four-form.}
thus locally we can write
\begin{equation}
\tfrac{1}{3}\tr\left(g\mathrm{d} g^{-1}\right)^3 = \mathrm{d} \mu_2(g) \ ,
\end{equation}
and hence to keep $H$ invariant, one actually transforms
\begin{equation} \label{Bdef}
B \mapsto B' = B - \tr\left(g^{-1}\mathrm{d} g\wedge A\right)
+ \mu_2(g)\ .
\end{equation}
Infinitesimally the three transformations are generated by vectors $v$
for diffeomorphisms, one-forms $\lambda$ and Lie algebra elements
$\Lambda$.
As before, we can combine these
into a single object $V=v+\Lambda+\lambda$ such that
\begin{equation}
\label{eq:variations}
\begin{aligned}
\delta_V A_{(i)} &= \mathcal{L}_{v_{(i)}}A_{(i)}
- \mathrm{d}\Lambda_{(i)} + \comm{\Lambda_{(i)}}{A_{(i)}}\ , \\
\delta_V B_{(i)} &= \mathcal{L}_{v_{(i)}}B_{(i)}
- \mathrm{d} \lambda_{(i)}
- \tr\left(\mathrm{d}\Lambda_{(i)}\wedge A_{(i)}\right)\ .
\end{aligned}
\end{equation}
Eqs. (\ref{finitegaugetr}) and (\ref{Bdef}) imply that the patching (\ref{eq:EpatchGG}) is now of the form (cf. (\ref{patchingabelian}) for the Abelian case, without diffeomorphisms)
\begin{equation}
\begin{aligned}
v_{(i)} = & v_{(j)}\ , \\
\mathrm{d}\lambda_{(i)} = & \mathrm{d}\lambda_{(j)}
- \mathcal{L}_{v_{(j)}}\mathrm{d}\lambda_{(ij)}
+2 \mathrm{d}\tr\left(\Lambda_{(j)}g^{-1}_{(ij)}\mathrm{d} g_{(ij)}\right)
+ \mathrm{d}\tr\left[(\mathcal{L}_vg_{(ij)})\mathrm{d} g_{(ij)}^{-1}\right]\\ & + \mathrm{d} (\iota_{v_{(j)}} \mu_2(g_{(ij)}) )\ , \\
\Lambda_{(i)} = & g_{(ij)}\Lambda_{(j)} g_{(ij)}^{-1}
+ g_{(ij)}\mathcal{L}_{v_{(j)}}g_{(ij)}^{-1}\ .
\end{aligned}
\end{equation}
There is an ambiguity in integrating the second relation but we can
\emph{choose} it such that
\begin{equation}
\label{eq:Epatch}
\begin{aligned}
v_{(i)} &= v_{(j)}\ , \\
\lambda_{(i)} &= \lambda_{(j)}
- \iota_{v_{(j)}}\mathrm{d}\lambda_{(ij)}
+2 \tr\left(\Lambda_{(j)} g^{-1}_{(ij)}\mathrm{d} g_{(ij)}\right)
+ \tr\left[(\mathcal{L}_vg_{(ij)})\mathrm{d} g_{(ij)}^{-1}\right]+ \iota_{v_{(j)}} \mu_2(g_{(ij)})\ , \\
\Lambda_{(i)} & = g_{(ij)}\Lambda_{(j)} g_{(ij)}^{-1}
+ g_{(ij)}\mathcal{L}_{v_{(j)}} g_{(ij)}^{-1} \ .
\end{aligned}
\end{equation}
The first two terms in the second line are the usual
patching one takes for the generalised tangent space, eq. (\ref{eq:EpatchGG}). The last three
terms are corrections which describe how the $B$-patching is twisted
by the gauge bundle.
We can view~\eqref{eq:Epatch} as defining a generalised heterotic
tangent space
\begin{equation}
E \simeq TM\oplus \adj P_G \oplus T^*M
\end{equation}
where $\adj P_G$ is the adjoint bundle with fibres in the Lie algebra $\mathfrak{g}$ of $G$. The bundle is naturally defined as an extension using the patchings~\eqref{eq:Epatch}. Crucially, we have seen that such an extension is only possible if the gauge bundle has trivial first Pontryagin class.
This space has the natural metric, invariant across patches, given by the non-abelian generalisation of~\eqref{eq:modscalar}
\begin{equation}
\label{eq:metricgauge}
\met{V}{W} = \tfrac12\iota_v \sigma+\tfrac12\iota_w \lambda
+ \tr \Lambda \Sigma \ .
\end{equation}
We note that this gives $E$ the structure of a $O(n,n+\dim{G})$ bundle, where $\dim{G}$ is the dimension of $G$. The differential structure is given by the generalised Lie derivative (\ref{eq:genLieAbel}) where the last term gets an extra contribution with repect to the non-Abelian version, i.e. we have
\begin{equation} \label{eq:genLiegauge}
\mathbb{L}_V W = {\cal L}_v w + {\cal L}_v \Sigma - {\cal L}_w \Lambda +
\comm{\Lambda}{\Sigma} + {\cal L}_v \rho - \iota_w \mathrm{d} \lambda + 2 \tr \Sigma \mathrm{d} \Lambda \ ,
\end{equation}
where $V=v+\Lambda+\lambda$. Similarly, the Courant bracket is given by the antisymmetrisation $\Bgen{V}{W} = \tfrac12 \mathbb{L}_V W - \tfrac12 \mathbb{L}_W V$. It takes the form
\begin{equation}
\label{eq:Bgengauge}
\Bgen{V}{W} = [v,w] + {\cal L}_v \Sigma - {\cal L}_w \Lambda + \comm{\Lambda}{\Sigma} +
{\cal L}_v \rho - {\cal L}_w \lambda - \tfrac12 \mathrm{d} (\iota_v \rho -\iota_w \lambda) + \tr (\Sigma \mathrm{d} \Lambda - \Lambda \mathrm{d} \Sigma) \ .
\end{equation}
and since $ \comm{\mathbb{L}_U}{\mathbb{L}_V} = \mathbb{L}_{\Bgen{U}{V}}$ it reproduces the algebra of the
variations~\eqref{eq:variations} by
$\comm{\delta_V}{\delta_{W}}=\delta_{\Bgen{V}{W}}$. It is easy to check that $\mathbb{L}_VW$ satisfies the conditions~\eqref{eq:CA-def} and so defines a transitive Courant algebroid.
\subsection{Splittings and automorphisms}
\label{sec:nonAbelian1}
Given the symmetries of the underlying supergravity we expect that the construction is invariant under under combinations of diffeomorphisms, closed $B$-shifts, pure-gauge ``$A$-shifts'' and global gauge transformations
\begin{equation} \label{eq:automorphisms}
\begin{aligned}
\text{$B$-shifts:}& \quad \lambda \mapsto \lambda - \iota_v B &&
\qquad \mathrm{d} B=0 \ , \\
\text{$A$-shifts:}& \quad
\begin{cases}
\lambda &\mapsto \lambda + 2 \tr(\Lambda A) - \tr[(\iota_v A)A] \ , \\
\Lambda &\mapsto \Lambda - \iota_vA \ ,
\end{cases}
&& \qquad \mathrm{d} A + A\wedge A =0 , \\
\text{global gauge:}& \quad
\Lambda \mapsto g \Lambda g^{-1}
&& \qquad \mathrm{d} g = 0 \ .
\end{aligned}
\end{equation}
These preserve the metric~\eqref{eq:metricgauge} and furthermore are automorphisms of the generalised Lie derivative~\eqref{eq:genLiegauge}.
More generally, we can view generic choice of $A$ and $B$ as providing a splitting of the generalised tangent space, that is, an explicit isomorphism $E\simeq TM\oplus \adj{P_G}\oplus T^*M$, which allows us to take a combination of vectors, gauge parameters and one-forms and lift them into a section of $E$. Writing this map as $\mathrm{e}^B\mathrm{e}^A$ we find, for generic $A$ and $B$ shifts, the generalised Lie derivative transforms as
\begin{equation} \label{eq:genLiegauge_shifts}
\mathbb{L}_{\e^B \e^A V} \e^B \e^A V' = \e^B \e^A \hat \mathbb{L}_{V} V' \ ,
\end{equation}
where
\begin{equation}\begin{aligned}\label{eq:genLietwist}
\hat \mathbb{L}_{V} V' = & \comm{v}{v'} + \mathcal{L}_v \lambda'
- \iota_{v'} \mathrm{d}\lambda +2 \tr (\Lambda' D \Lambda) + 2 \tr (\Lambda' \iota_v F) - 2 \tr (\Lambda \iota_{v'} F) \\ &
+ \iota_v \iota_{v'} H + \comm{\Lambda}{\Lambda'} + \iota_v D \Lambda' - \iota_{v'}D\Lambda + \iota_v \iota_w F \ ,
\end{aligned} \end{equation}
and where we introduced the gauge covariant derivative $D \Lambda = \mathrm{d} \Lambda +[A,\Lambda]$ and the field strength $F = \mathrm{d} A + A\wedge A$.
Note also, that there are vector duals of $A$-shifts, which we might call
$\alpha$-shifts
\begin{equation}
\text{$\alpha$-shifts:} \quad
\begin{cases}
v \mapsto v + 2 \tr(\Lambda\alpha) - \tr[(\iota_\alpha
\lambda)\alpha]\ , \\
\Lambda \mapsto \Lambda - \iota_\alpha \lambda \ ,
\end{cases}
\end{equation}
which are the analogues of $\beta$-transformations (and similarly fail
to preserve the bracket).
These are the complementary elements in $O(n,n+\dim G)$.
\subsection{Extended generalised tangent space for $G \times \tilde{G}$}
\label{sec:nonAbelian2}
As the first step towards understanding the gravitational contribution $ \tfrac{1}{4 } \alpha' \tr R\wedge R$ to the heterotic Bianchi identity \eqref{eq:hetBianchi} let us consider an extension by a product group $G\times \tilde{G}$, where $\tilde{G}=GL(n,\mathbb{R})$. We will also now include explicit factors of $\alpha'$. Of course by treating the gravitational connection $\omega = \tilde A$ as independent $\tilde G$ connection, we introduce unphysical degrees of freedom. For now, let us not worry about that and simply take advantage of the fact and in the ``new" Bianchi identity
\begin{equation}
\label{eq:fakeBI}
\mathrm{d} H = \tfrac14 \alpha' \left[\tr F \wedge F - \tr {\tilde F}\wedge {\tilde F}\right] \ ,
\end{equation}
the gravitational contribution takes the familiar form and can be treated like a gauge one. In particular
\begin{equation}
H = \mathrm{d} B + \tfrac14 \alpha' (\omega_3( A) - \omega_3(\tilde A)) ,
\end{equation}
where $\omega_3$ is given by \eqref{eq:CS}.
The gauge fields $A$ and $\tilde A$ transform under gauge transformations as
\begin{equation}
A\mapsto gAg^{-1} + g\mathrm{d} g^{-1} \ , \qquad \tilde A \mapsto \tilde g \tilde A \tilde g^{-1}+\tilde g\mathrm{d} \tilde g^{-1} \ .
\end{equation}
This implies that $B$ transforms as
\begin{equation} \label{eq:gaugetrafoB}
B \to B - \tfrac14 \alpha' \left(\tr\left(g^{-1}\mathrm{d} g\wedge A\right) - \tr\left(\tilde g^{-1}\mathrm{d} \tilde g\wedge \tilde A\right) \right)
+ \tfrac14 \alpha' \left(\mu_2( g) - \mu_2(\tilde g)\right) \ .
\end{equation}
For the patching of the generalised tangent bundle we find (similar to \eqref{eq:Epatch})
\begin{equation}\label{eq:Epatch2gauge}
\begin{aligned}
v_{(i)} =& v_{(j)} \ , \\
\lambda_{(i)} =& \lambda_{(j)}
- \iota_{v_{(j)}}\mathrm{d}\lambda_{(ij)}
+ \tfrac12 \alpha' \left( \tr\left(\Lambda_{(j)} g^{-1}_{(ij)}\mathrm{d} g_{(ij)}\right) - \tr\left(\tilde \Lambda_{(j)} \tilde g^{-1}_{(ij)}\mathrm{d} \tilde g_{(ij)}\right) \right) \\ &
+ \tfrac14 \alpha' \left( \tr\left((\mathcal{L}_vg_{(ij)})\mathrm{d} g_{(ij)}^{-1}\right) - \tr\left((\mathcal{L}_v \tilde g_{(ij)})\mathrm{d} \tilde g_{(ij)}^{-1}\right) \right) + \tfrac14 \alpha' \iota_{v_{(j)}} (\mu_2( g_{(ij)}) - \mu_2(\tilde g_{(ij)}) )\ , \\
\tilde \Lambda_{(i)} =& g_{(ij)}\Lambda_{(j)} g_{(ij)}^{-1}
+ g_{(ij)}\mathcal{L}_{v_{(j)}} g_{(ij)}^{-1} \ ,\\
\Lambda_{(i)} =& g_{(ij)}\Lambda_{(j)} g_{(ij)}^{-1}
+ g_{(ij)}\mathcal{L}_{v_{(j)}} g_{(ij)}^{-1} \ .
\end{aligned}
\end{equation}
The scalar product reads
\begin{equation}
\label{eq:metric2gauge}
\met{V}{V'} = \tfrac12\iota_v \lambda' + \tfrac12\iota_{v'} \lambda + \tfrac14 \alpha' (\tr ( \Lambda \Lambda') - \tr (\tilde \Lambda \tilde \Lambda')) \ ,
\end{equation}
and is invariant under the patching given in \eqref{eq:Epatch2gauge}. The generalised Lie
derivative is given by
\begin{equation} \label{eq:genLie2gauge}\begin{aligned}
\mathbb{L}_V V' = & \comm{v}{v'} + \mathcal{L}_v \lambda'
- \iota_{v'} \mathrm{d}\lambda + \tfrac14 \alpha' (\tr \Lambda' \mathrm{d} \Lambda - \tr \tilde\Lambda' \mathrm{d}\tilde\Lambda ) \\ &
+ \comm{\Lambda}{\Lambda'} + \iota_v \mathrm{d} \Lambda' - \iota_{v'}\mathrm{d}\Lambda + \comm{\tilde \Lambda}{\tilde \Lambda'} + \iota_v \mathrm{d} \tilde \Lambda' - \iota_{v'}\mathrm{d}\tilde \Lambda\ .
\end{aligned}\end{equation}
and the Courant bracket by
\begin{equation}
\label{eq:Bgen2gauge}
\begin{aligned}
\Bgen{V}{V'} = & \comm{v}{v'}
+ \mathcal{L}_v\lambda' - \mathcal{L}_{v'}\lambda
- \tfrac{1}{2}\mathrm{d}\left(\iota_v\lambda' - \iota_{v'}\lambda\right)
\\ &
+ \tfrac18 \alpha' (\tr ( \Lambda' \mathrm{d} \Lambda) - \tr ( \Lambda \mathrm{d} \Lambda') - \tr (\tilde\Lambda' \mathrm{d}\tilde\Lambda) + \tr (\tilde\Lambda \mathrm{d}\tilde\Lambda'))\\ &
+ \comm{\Lambda}{\Lambda'}
+ \left(\iota_{v}\mathrm{d}\Lambda' - \iota_{v'}\mathrm{d}\Lambda \right)
+ \comm{\tilde \Lambda}{\tilde \Lambda'} + \left( \iota_v \mathrm{d} \tilde \Lambda' - \iota_{v'}\mathrm{d}\tilde \Lambda \right) \ .
\end{aligned}
\end{equation}
The relative minus sign between the two terms in \eqref{eq:fakeBI} is a matter of trace conventions for $G$ and $\tilde G$ respectively and is reflected in our construction of $E$. However, we shall see that once chosen this relative sign propagates to the the effective action and the $F^2$ and ${\tilde F}^2$ terms will also have opposite signs. This will be addressed in sections~\ref{sec:gravconnection} and~\ref{sec:het-corr}.
We shall now have a preliminary discussion of the effect of choosing $\tilde{G}=GL(n,\mathbb{R})$ on the Courant brackets and how imposing extra symmetry properties, such as T-duality, covariance may restrict the so-far unconstrained connections.
\subsection{Local data: isometries and T-duality}
\label{sec:redgauge}
In section~\ref{sec:GGBianchi} we reviewed the circle reduction in generalised geometry. We will now redo the analysis in the extended case, and study the effect of the presence of a gauge group in the higher-dimensional theory. Once again, the Bianchi identity \eqref{eq:Bianchi_gauge} will be our starting point and we introduce a shorthand for the right-hand side
\begin{equation}\label{eq:Bianchi_general}
\mathrm{d} H = X (F) \ .
\end{equation}
We have already seen how the generalised geometry restricts the topology of the gauge bundle. The reason for redoing this exercise is to study the role of the connections themselves, and see how they can be restricted by the constraint of realising of T-duality. This leads us in to the local constructions of sections~\ref{sec:gravconnection} and~\ref{sec:het-corr}.
As already explained, $X(F)$ is exact, i.e.\
\begin{equation}
X(F) = \kappa \mathrm{d} \omega(A) \ ,
\end{equation}
where $\omega(A)$ in the case of $X = \kappa \tr(F^2)$ is given by the Chern-Simons term \eqref{eq:CS}.
Using the decomposition \eqref{eq:Hdecomp} and
\begin{equation}
X(F)= X_4(F) + X_3(F) \wedge (\mathrm{d} \varphi + C_1) \ ,
\end{equation}
the Bianchi identity \eqref{eq:Bianchi_general} translates into
\begin{equation}
\mathrm{d} H_3 = X_4 - H_2 \wedge G_2 \ , \qquad \mathrm{d} H_2 = X_3 \ .
\end{equation}
We can replace $X_4$ and $X_3$ by the components $\omega_3$ and $\omega_2$ of the Chern-Simons term $\omega(A)$, with
\begin{equation}\label{eq:dimdecomp_CS}
\omega(A)= \omega_3 + \omega_2 \wedge (\mathrm{d} \varphi + C_1) \ .
\end{equation}
We find that
\begin{equation}\label{eq:Bianchi_red}
\mathrm{d} H_3 = \kappa \mathrm{d} \omega_3 - \tilde G_2 \wedge G_2 \ ,
\end{equation}
and
\begin{equation}
\mathrm{d} \tilde G_2 = 0 \ ,
\end{equation}
where we defined
\begin{equation}\label{eq:tildeG2}
\tilde G_2= H_2 - \kappa \omega_2 \ .
\end{equation}
In order to understand the properties of $\omega_3$ and $\omega_2$, we dimensionally reduce the gauge connection $A$, i.e.\
\begin{equation}
A = A_1 + a (\mathrm{d} \varphi + C_1) \ ,
\end{equation}
where $A_1$ is the lower-dimensional connection while the Wilson lines $a$ transform tensorially under lower-dimensional gauge transformations, i.e.\
\begin{equation}
A_1 \to g A_1 g^{-1} + g \mathrm{d} g^{-1} \ , \qquad a \to g a g^{-1} \ ,
\end{equation}
where $g$ does not depend on $\varphi$.
From \eqref{eq:dimdecomp_CS} and \eqref{eq:CS} we find
\begin{equation}\begin{aligned}
\omega_3 = & \tr( A_1 \wedge \mathrm{d} A_1 + \tfrac23 A_1^3) + \tr(a A_1) \wedge G_2 \ , \\
\omega_2 = & \tr(A_1 \wedge \mathrm{d} a + a \mathrm{d} A_1 + 2 a A_1 \wedge A_1) + \tr(a^2) G_2 \ .
\end{aligned}\end{equation}
By defining the gauge-covariant derivative $D$ on $a$ as
\begin{equation}
Da = \mathrm{d} a + [A_1,a] \ ,
\end{equation}
we can rewrite $\omega_2$ in a tensorial way as
\begin{equation}
\omega_2 = \tr(DDa) + \tr(a^2) G_2 \ ,
\end{equation}
which means $\omega_2$ is gauge invariant and therefore globally defined, which in turn means that $\tilde G_2$ in \eqref{eq:tildeG2} is globally defined. Moreover, defining
\begin{equation}
\tilde H_3 = H_3 - \tfrac12 \tr(Da) \wedge G_2 \ ,
\end{equation}
we can solve \eqref{eq:Bianchi_red} by
\begin{equation}
\tilde H_3 = \mathrm{d} \tilde B_2 + \kappa \omega(A_1) + \omega(A^-) - \omega(A^+) \ ,
\end{equation}
where $A^\pm$ are the gauge connections for $F^\pm = \tfrac12 (\tilde G_2 \pm G_2)$. Furthermore, the two-form gauge field $\tilde B_2$ appears in the decomposition of $B$ as
\begin{equation}
B = (\tilde B_2 + \tfrac12 \tr(a) G_2) + B_1 \wedge (\mathrm{d} \varphi + C_1) \ .
\end{equation}
In total, the Bianchi identity of the lower-dimensional three-form field strength $\tilde H_3$ reads
\begin{equation}
\mathrm{d} \tilde H_3 = \kappa \tr (F_2^2) + \tr ((G^-)^2) - \tr ((G^+)^2) \ .
\end{equation}
We can now consider T-duality on the circle. This will exchange $G_2$ and $\tilde G_2$, and we see that the three-form $\tilde H_3$ has been constructed such that it remains invariant. Furthermore, also the gauge fields $A_1$ must transform such that $\tr (F_2^2)$ is T-duality invariant. Examples are $A_1$ itself being invariant or anti-invariant. The Wilson lines $a$ on the other hand can have a very non-trivial transformation behaviour.
If $A$ is simply a connection in $G$ these constraints are rather mild. However, let us now turn to the gravitational connection $\tilde A$ in $\tilde{G}=GL(n,\mathbb{R})$. The invariance of the horizontal part of $\tr { \tilde F} \wedge {\tilde F}$ immediately implies that $\tilde A$ cannot be the Levi-Civita connection but must involve $H$, since T-duality exchanges metric and $B$-field components. In fact as shown in \cite{LM} this is satisfied provided $\tilde F$ is computed using the torsionful connections
\begin{equation}
\Omega^\pm = \omega^{\tiny{\rm LC}} \pm \tfrac12 H \ ,
\end{equation}
where $\omega^{\tiny{\rm LC}}$ is the Levi-Civita connection. The torsion of $\Omega^\pm$ is just $\mp H$.
Note that the T-duality considerations do not fix the relative sign in $\Omega^\pm$ but do imply that the connection cannot be generic. As already mentioned, this is also the connection that is naturally chosen by world-sheet supersymmetry. The goal of sections~\ref{sec:gravconnection} and~\ref{sec:het-corr} is to show how the generalised geometry singles out the choice of $\Omega^\pm$.
\section{Action, connections and more}
\label{sec:gravconnection}
Thus far we have discussed how the extension of conventional
generalised geometry by a non-Abelian gauge group captures the global
topology that encodes the non-trivial Bianchi identity for $H$. To
connect to the actual supergravity theories, we need to construct the
local dynamics.
In this section we shall present an ``all-at-once'' construction of
effective actions based on the generalisations of the Lichnerowitz
theorem. This will reproduce the standard ten-dimensional, $N=1$
supergravity coupled to super Yang--Mills with gauge group
$G$~\cite{CM}, which has particle content
\begin{equation}
\begin{aligned}
\text{gravity multiplet:} \quad &
\left\{ g, B, \phi ; \psi_m, \rho \right\} \\
\text{gauge multiplet:} \quad &
\left\{ A ; \zeta \right\} .
\end{aligned}
\end{equation}
As usual in generalised geometry, the construction is to consider a
refined structure on $E$ and then define the appropriate compatible
generalised connection.
In the next section, we shall adapt this general construction to the
specific examples of lowest order $\alpha'$ corrections in heterotic
string and see the importance of the local data. A related
construction based on~\cite{CSW1} was given
in~\cite{GF}, see also~\cite{Baraglia}. However, when applied to the heterotic theory,
this required the $\Omega^-$ connection in the Bianchi
identity~\eqref{eq:hetBianchi} to be put in by hand. Using the
construction here, as we will see in the next section, $\Omega^-$
will appear naturally.
\subsection{Local structure group}
\label{sec:localO}
As in the previous section, we identify a generalised tangent space
$E\simeq TM\oplus \adj P_G \oplus T^*M$, for a generic gauge group
$G$ whose Lie algebra is $\mathfrak{g}$. Given an invariant metric on
$\mathfrak{g}$, denoted by the trace, there is a natural interior
product on $E$
\begin{equation}
\GM{v+\Lambda+\lambda}{v'+\Lambda'+\lambda'}
=\tfrac12\left(\iota_v \lambda' + \iota_{v'} \lambda\right)
+\tfrac{\alpha'}{4}\tr\Lambda\Lambda' ,
\end{equation}
providing an isomorphism $E^*\simeq E$. For the weighted generalised
vector bundle $\tilde{E}=(\det T^*M)\otimes E$ one can then define a
corresponding bundle $\tilde{F}$ of frames, orthonormal up to a
conformal factor. By definition, this is an
$O(n+\dim G,n)\times\mathbb{R}^+$ principal bundle.
Given the fields $(g,B,A,\phi)$ one can define the set of
frames\footnote{Nearly all notation and conventions follow~\cite{CSW1}.}
\begin{equation}
\begin{aligned}
\label{eq:splitframe}
\hat{E}^+_a &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\big(\hat{e}^+_a + e^+_a + \iota_{\hat{e}^+_a}B +
\iota_{\hat{e}^+_a}A - \tfrac{\alpha'}{4}\tr A\,\iota_{\hat{e}^+_a}A \big) ,\\
\hat{E}^+_{\alpha} &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\big(\sqrt{\tfrac{4}{\alpha'}}t_{\alpha}
- \sqrt{\alpha'}\,\tr t_{\alpha}A\big) ,\\
\hat{E}^-_{\bar{a}} &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\big(\hat{e}^-_{\bar{a}}
-e^-_{\bar{a}} + \iota_{\hat{e}^-_{\bar{a}}}B
+ \iota_{\hat{e}^-_{\bar{a}}}A
- \tfrac{\alpha'}{4} \tr A\,\iota_{\hat{e}^-_{\bar{a}}}A\big) ,
\end{aligned}
\end{equation}
where $\hat{e}^{\pm}$ are orthonormal bases in $TM$, $e^{\pm}$ their
dual, and
$[t_{\alpha},t_{\beta}]=f_{\alpha\beta}{}^{\gamma}t_{\gamma}$ give a
basis for $\mathfrak{g}$, with $\tr t_{\alpha}t_{\beta}=
\delta_{\alpha\beta}$. We have
\begin{equation}
\begin{aligned}
\GM{\hat{E}^+_a}{\hat{E}^+_b} &= \norm{\vol_G}^2\delta_{ab} , \\
\GM{\hat{E}^+_\alpha}{\hat{E}^-_{\beta}}
&= \norm{\vol_G}^2\delta_{\alpha\beta} , \\
\GM{\hat{E}^-_{\bar{a}}}{\hat{E}^-_{\bar{b}}}
&= -\norm{\vol_G}^2\delta_{\bar{a}\bar{b}} ,
\end{aligned}
\end{equation}
where $\norm{\vol_G}=\mathrm{e}^{-2\phi}\sqrt{-g}$, and all other inner
products vanishing. Following~\cite{GF, Baraglia}, one can view these
frames as defining a generalised metric which splits $E=\tilde{C}_+\oplus
\tilde{C}_-$, and reduces the structure group to $O(n+\dim G)
\times O(n)$, with the respective frames\footnote{Note that given the
$O(n+\dim G)$ frame is not generic: one could also take
rotated combinations of $\hat{E}^+_a$ and $\hat{E}^+_{\alpha}$.}
\begin{equation}
\hat{E}_A =
\begin{cases}
\left(\hat{E}^+_a,\hat{E}^+_{\alpha}\right) &
O(n+\dim G)_+ \\
\hat{E}^-_{\bar{a}} & O(n)_-
\end{cases} .
\end{equation}
We actually want to reduce the structure group on $E$ further to
disentangle $G$ from the $O(n)$ inside $\tilde{C}_+$. In other words, we pick
a substructure $\tilde{P}\subset\tilde{F}$ of the generalised frame
bundle which is an $O(n)\times G \times O(n)$ principal bundle. This
further splits $E = C_+\oplus C_{\mathfrak{g}}\oplus C_-$ with the
respective frames
\begin{equation}
\hat{E}_A =
\begin{cases}
\hat{E}^+_a & O(n)_+ \\
\hat{E}^+_{\alpha} & G \\
\hat{E}^-_{\bar{a}} & O(n)_-
\end{cases} .
\end{equation}
It is in this setting that we would like to present the
supergravity equations in terms of generalised connections, much in
the same way as for the type II case in~\cite{CSW1}.
In~\cite{CSW1} it was shown that all supergravity equations can be
written in terms of generalised connections which are compatible with
the local $O(n)\times O(n)$ structures and are torsion-free, and that
even though there are several such connections, the supergravity
equations are uniquely determined, as they should
be. In~\cite{GF} a similar approach was taken for the
heterotic theory, using the local $O(n+\dim G)\times O(n)$
structure discussed above. The supergravity equations could again be
formulated using torsion-free, compatible generalised connections,
however, the form of the $R$-curvature in the Bianchi
identity~\eqref{eq:hetBianchi} has to be put in by hand. A priori, it
is the curvature of an arbitrary connection on a generic $O(n)$
bundle. The condition that the bundle is that coming from $TM$ and
that the relevant connection is $\Omega^-$ as given
in~\eqref{eq:Omegaminus} does not come from the construction.
To overcome this problem we will instead consider generalised
connections that are compatible with the $O(n)\times G \times O(n)$
structure $\tilde{P}=\tilde{P}_+\oplus \tilde{P}_G\oplus
\tilde{P}_-$. The consequence is that we can no longer
require that generalised connections compatible with the reduced
structure be torsion-free, as for a generic choice of $\tilde{P}$
(i.e. a generic field configuration) there exists a non-vanishing
intrinsic torsion. To see this, let $D$ and $D'$ be two arbitrary
compatible generalised connections. Fix $D$ and define the
tensor $\Sigma = D' - D$. Then by varying $D'$ we have that $\Sigma$
spans
\begin{equation}
K = E \otimes
(\Lambda^2C_+\oplus \adj\tilde{P}_G\oplus\Lambda^2C_-),
\end{equation}
where $\adj\tilde{P}_G$ is the adjoint bundle with fibres
in the Lie algebra $\mathfrak{g}$ and we have used we used
$\adj\tilde{P}_\pm\simeq \Lambda^2 C_{\pm}$. Now the torsion of a
connection is defined by the bracket~\eqref{eq:genLiegauge} and is
an element~\cite{CSW1}
\begin{equation}
T(D) \in \Gs{W}, \quad \text{where } W= \Lambda^3E\oplus E .
\end{equation}
Therefore we can define the associated map for
the difference of the torsions, $\tau: K \rightarrow W$,
such that
\begin{equation}
\tau(\Sigma) = T(D') - T(D) .
\end{equation}
It is easy to check that the intrinsic torsion space
$W_{\text{int}}=\coker\tau$ is non-trivial,
\begin{equation}
W_{\text{int}} = C_+\otimes C_{\mathfrak{g}}\otimes C_-.
\end{equation}
We thus have that, for a generic structure, every compatible
connection will have some torsion. Using the
frame~\eqref{eq:splitframe}, we can calculate the intrinsic torsion
explicitly and find it takes the form
\begin{equation}
\label{eq:int-tor}
\hat{T}_{\bar{a}b\gamma}=
-\tfrac12 \sqrt{\alpha'}F_{\bar{a}b\gamma} .
\end{equation}
Thus we see that requiring a torsion-free connection sets $F=0$, which
is too strong a condition.
Instead of requiring the full torsion-free condition, we can take the
weaker physical conditions that the connections that appear in the
supergravity equations are uniquely determined. We consider first the
fermionic fields. Using the $O(n)\times G \times O(n)$ structure we
identify the fields\footnote{Note that we are using
the redefined dilatino $\rho=\gamma^m\psi_m-\lambda$
where $\lambda$ is the conventional dilatino field. Also, for the ten-dimensional theory one would need to correctly identify the chiralities of the fermion fields but in order to keep the discussion completely general we will not be explicit about this here.}
\begin{equation}\label{eq:FermionFields}
\begin{aligned}
\text{gravitino:} & && \psi_a \in \Gs{C_+\otimes S(C_-)} , \\
\text{gaugino:} & &&
\zeta_\alpha \in \Gs{C_\mathfrak{g}\otimes S(C_-)} , \\
\text{dilatino:} & && \rho \in \Gs{S(C_-)} ,
\end{aligned}
\end{equation}
where $S(C_-)$ is the spin-bundle for the $\Spin(n)$ group on $C_-$
(we assume throughout that our manifold is spin). The supersymmetry
parameter $\epsilon$ is similarly a section of $S(C_-)$. Comparing
with the structure of type II theories discussed in~\cite{CSW1},
we require that the following operators are uniquely determined, since
they will appear in either the supersymmetry transformations or the
fermion equations of motion
\begin{equation}
\begin{aligned}
\label{eq:SugraOps}
\gamma^{\bar{a}}D_{\bar{a}}\rho, \quad &
D^a\psi_a, &
D^{\alpha}\zeta_{\alpha},\quad & \in \Gs{S(C_-)},\\
\phantom{\gamma^{\bar{a}}D_{\bar{a}}\rho,\quad} & D_a \rho, &
\gamma^{\bar{a}}D_{\bar{a}}\psi_a, \quad
& \in \Gs{C_+\otimes S(C_-)},\\
\phantom{\gamma^{\bar{a}}D_{\bar{a}}\rho,\quad} & D_{\alpha}\rho, &
\gamma^{\bar{a}}D_{\bar{a}}\zeta_{\alpha}, \quad
& \in \Gs{C_{\mathfrak{g}}\otimes S(C_-)}.
\end{aligned}
\end{equation}
Solving for the compatible connection we find that this implies
that the torsion of $T(D)$ is restricted to lie in the subspace
\begin{equation}
T(D) \in \Gs{W_{\text{restr}}},
\end{equation}
where
\begin{equation*}
W_{\text{restr}} = \Lambda^3C_+ \oplus \Lambda^3C_{\mathfrak{g}}
\oplus \big(C_+\otimes \adj \tilde{P}_G \big)
\oplus \big(C_{\mathfrak{g}}\otimes\Lambda^2C_+\big)
\oplus \big(C_+\otimes C_{\mathfrak{g}}\otimes C_- \big)\subset \Lambda^3 E \oplus E,
\end{equation*}
with the last term being the intrinsic torsion, which is independent of the
choice of $D$. Using the frames~\eqref{eq:splitframe}, we can solve
explicitly for $D$. If $\nabla=\partial+\omega+A$ is the covariant derivative associated with the combined
Levi--Civita and gauge connection, then acting on a generalised vector $W =
w_+^a\hat{E}^+_a + w_{\mathfrak{g}}^{\alpha}\hat{E}^+_{\alpha} +
w_-^{\bar{a}}\hat{E}^-_{\bar{a}} \in\Gs{E}$, we have
\begin{subequations}
\label{eq:OdddLC}
\begin{align}
D_a w_+^b
&= \nabla_a w_+^b
- \tfrac{2}{n-1}\big(
\delta_a{}^b \partial_c\phi
-\eta_{ac}\partial^b\phi \big)w_+^c
+ Q_a{}^b{}_c w_+^c ,\label{eq:OdddLC1} \\
D_{\alpha} w_+^b
&= Q_{\alpha}{}^b{}_a w_+^{a},\label{eq:OdddLC2} \\
D_{\bar{a}} w_+^b
&= \nabla_{\bar{a}} w_+^b
- \tfrac{1}{2}H_{\bar{a}}{}^b{}_cw_+^c,
\label{eq:OdddLC3}\\
D_a w_{\mathfrak{g}}^{\beta}
&= \nabla_a w_{\mathfrak{g}}^{\beta}
+ Q_a{}^{\beta}{}_{\gamma}w_{\mathfrak{g}}^{\gamma},
\label{eq:OdddLC4}\\
D_{\alpha} w_{\mathfrak{g}}^{\beta}
&= Q_{\alpha}{}^{\beta}{}_{\gamma}w_{\mathfrak{g}}^{\gamma},
\label{eq:OdddLC5}\\
D_{\bar{a}} w_{\mathfrak{g}}^{\beta}
&= \nabla_{\bar{a}} w_{\mathfrak{g}}^{\beta},
\label{eq:OdddLC6}\\
D_a w_-^{\bar{b}}
&= \nabla_a w_-^{\bar{b}}
+ \tfrac{1}{2}H_a{}^{\bar{b}}{}_{\bar{c}}w_-^{\bar{c}} ,
\label{eq:OdddLC7}\\
D_{\alpha} w_-^{\bar{b}}
&= -\tfrac{1}{2}\sqrt{\alpha'}
F^{\bar{b}}{}_{\bar{a}}{}_{\alpha}w_-^{\bar{a}} ,
\label{eq:OdddLC8}\\
D_{\bar{a}} w_-^{\bar{b}}
&= \nabla_{\bar{a}} w_-^{\bar{b}}
+ \tfrac{1}{6}H_{\bar{a}}{}^{\bar{b}}{}_{\bar{c}}w_-^{\bar{c}}
- \tfrac{2}{n-1}\big(
\delta_{\bar{a}}{}^{\bar{b}} \partial_{\bar{c}}\phi
- \eta_{\bar{a}\bar{c}}\partial^{\bar{b}}\phi
\big)w_-^{\bar{c}}
+ Q_{\bar{a}}{}^{\bar{b}}{}_{\bar{c}} w_-^{\bar{c}} ,
\label{eq:OdddLC9}
\end{align}
\end{subequations}
for any $Q\in\Gs{E\otimes\adj\tilde{P}}$ satisfying
\begin{equation}
\begin{aligned}
Q_{a\bar{b}\bar{c}}&=0, & &&
Q_{\alpha\bar{b}\bar{c}}&=0, & &&
Q_{\bar{a}bc}&=0, & &&
Q_{\bar{a}\beta\gamma}&=0, \\
Q_{[\bar{a}\bar{b}\bar{c}]}&=0, & &&
Q_{\bar{a}}{}^{\bar{a}}{}_{\bar{b}}&= 0, & &&
Q_{\alpha}{}^{\alpha}{}_{\beta}&=0, & &&
Q_{a}{}^{a}{}_b&= 0,
\end{aligned}
\end{equation}
and otherwise arbitrary. This means that $Q$ will in general
contribute to the torsion of $D$, for instance the fully antisymmetric
component $Q_{[abc]}$ will be in the torsion if it is non-vanishing.
In addition, again note that irrespective of the particular choice of
$Q$, all connections $D$ will have an intrinsic torsion given by
$\hat{T}_{\bar{a}b\gamma}= -\tfrac12 \sqrt{\alpha'}F_{\bar{a}b\gamma}$,
simply due to the fact that they are compatible with the reduced
structure $\tilde{P}$. Crucially, however, the supergravity
operators~\eqref{eq:SugraOps} are independent of the particular choice
of $D$, since all $Q$ contributions will drop out
of~\eqref{eq:SugraOps}.\footnote{As an alternative way of stating this,
while we have been thinking of the unconstrained $Q$ as an element
of $U\subset K = E\otimes \text{ad} \tilde{P}$, we could instead
consider it as a map $\mathcal{Q}: S\oplus J\oplus J_{\mathfrak{g}}
\rightarrow E\otimes (S\oplus J\oplus J_{\mathfrak{g}})$ where
$S=S(C_-)$, $J=C_+\otimes S(C_-)$ and $J_{\mathfrak{g}}=
C_{\mathfrak{g}}\otimes S(C_-)$.
The linear projections in~\eqref{eq:SugraOps} on the other hand are
a map $\mathcal{P}: E\otimes (S\oplus J\oplus J_{\mathfrak{g}})
\rightarrow S\oplus J\oplus J_{\mathfrak{g}}$. The map
$\mathcal{Q}$ is then such that precisely $\text{ker}\mathcal{P} =
\text{Im} \mathcal{Q}$.
The space $U$ spanned by $Q$ is thus set by the kernel of the
projection to $S\oplus J\oplus J_{\mathfrak{g}}$. For the type II
case, this space coincided with the kernel of the torsion map
within the metric connections space. In this case, since
the fermionic representations are smaller, the kernel of the
projection is larger and includes some of the torsion.}
\subsection{Supergravity equations}
\label{sec:sugraEQ}
We next see how the generalised connection $D$ encodes the
equations of $N=1$ ten-dimensional supergravity coupled to super
Yang-Mills~\cite{CM}.
We start with the bosonic action. For any spinor
$\epsilon\in\Gs{S(C_-)}$, we can consider the following Bismut-type
equation:
\begin{equation}
\label{eq:newBismut}
\gamma^{\bar{a}}D_{\bar{a}}\gamma^{\bar{b}}D_{\bar{b}}\epsilon
= D^aD_a\epsilon
+ D^{\alpha}D_{\alpha}\epsilon
- \tfrac14 S^-\epsilon ,
\end{equation}
Generically one might expect the right-hand side to contain scalar,
two-form and four-form terms of the form $T_{\bar{a}_a\dots
\bar{a}_{2n}}\gamma^{\bar{a}_a\dots \bar{a}_{2n}}$ (since the
spinors are chiral the six-form terms can be written as
four-forms). However, remarkably, provided the Bianchi
identity~\eqref{eq:Bianchi_gauge} holds, only the scalar term $S^-$
survives. This scalar gives the dilaton equation of motion, or equivalently, the
bosonic action
\begin{equation}
\label{eq:BosonicAction}
S^- = s +4 \nabla^2 \phi -4 (\partial \phi)^2
- \tfrac{1}{12} H^2 -\tfrac{\alpha'}{8}\tr F^2 ,
\end{equation}
where $s$ is the Ricci scalar for the metric $g$.
The supersymmetry variations of the fermion fields are given by
\begin{equation}
\label{eq:FermVar}
\begin{aligned}
\delta\psi_a &= D_a \epsilon
= \nabla_a\epsilon+ \tfrac{1}{8}H_{a\bar{b}\bar{c}}
\gamma^{\bar{b}\bar{c}} \epsilon , \\
\delta\zeta_{\alpha} &= D_{\alpha}\epsilon
= -\tfrac18 \sqrt{\alpha'}F_{\bar{a}\bar{b}\alpha}
\gamma^{\bar{a}\bar{b}}\epsilon , \\
\delta\rho &= \gamma^{\bar{a}}D_{\bar{a}}\epsilon
= \gamma^{\bar{a}}\nabla_{\bar{a}} \epsilon
+ \tfrac{1}{24}H_{\bar{a}\bar{b}\bar{c}}
\gamma^{\bar{a}\bar{b}\bar{c}}\epsilon
- (\partial_{\bar{a}}\phi) \gamma^{\bar{a}}\epsilon ,
\end{aligned}
\end{equation}
while the fermionic action has the form
\begin{equation}
\begin{aligned}
\label{eq:FermionicAction}
S_F = -\frac{1}{2\kappa^2}\int 2 \vol_G &\Big[
\bar\psi^{a} \gamma^{\bar{b}} D_{\bar{b}} \psi_{a}
+ \bar\psi^{a} \gamma^{\bar{b}}
\hat{T}_{\bar{b}}{}^{\alpha}{}_a \zeta_{\alpha}
+ 2 \bar\rho D_a \psi^{a}
- \bar\rho \gamma^{\bar{a}} D_{\bar{a}} \rho \\&
+ \bar\zeta^{\alpha} \gamma^{\bar{b}}
D_{\bar{b}}\zeta_{\alpha}
+ \bar\zeta^{\alpha} \gamma^{\bar{b}}
\hat{T}_{\bar{b}}{}^{a}{}_{\alpha} \psi_{a}
+ 2 \bar\rho D_{\alpha} \zeta^{\alpha}
\Big] ,
\end{aligned}
\end{equation}
where $\hat{T}$ is the intrinsic torsion~\eqref{eq:int-tor} which is
fixed by the choice of local structure $\tilde{P}$ (and independent of
choice of compatible $D$).
The Bismut-type formula~\eqref{eq:newBismut} can then be understood as
a consequence of the dilatino equations of motion closing into the
dilaton equation of motion under susy. The bosonic equations of motion
for $g,B,A$ will similarly contain torsion terms, with the
generalised Ricci given by
\begin{equation}
\begin{aligned}
\label{eq:Bosoniceoms2}
\tfrac12 R_{\bar{a}b}\gamma^{\bar{a}}\epsilon &= [\gamma^{\bar{a}}D_{\bar{a}}, D_b]\epsilon+\gamma^{\bar{a}}\hat{T}_{\bar{a}}{}^{\beta}{}_b D_{\beta} \epsilon ,\\
\tfrac12 R_{\bar{a}\beta}\gamma^{\bar{a}}\epsilon &= [\gamma^{\bar{a}}D_{\bar{a}}, D_{\beta}]\epsilon+\gamma^{\bar{a}}\hat{T}_{\bar{a}}{}^{b}{}_{\beta} D_{b} \epsilon ,
\end{aligned}
\end{equation}
and the $g$ and $B$ equations and $A$ equation corresponding to the
generalised Ricci flat conditions
\begin{equation}
R_{\bar{a}a} = 0 \qquad \text{and} \qquad R_{\bar{a}\alpha} = 0
\end{equation}
respectively. For the bosonic supersymmetry transformations we have
\begin{equation}
\label{eq:Evar}
\begin{aligned}
\delta\hat{E}^+_a
&= \bar{\epsilon} \rho \hat{E}^+_a
- \bar{\epsilon} \gamma_{\bar{a}} \psi_a \hat{E}^{-\bar{a}}, \\
\delta\hat{E}^+_{\alpha}
&= \bar{\epsilon} \rho \hat{E}^+_{\alpha}
- \bar{\epsilon} \gamma_{\bar{a}} \zeta_{\alpha} \hat{E}^{-\bar{a}} , \\
\delta\hat{E}^-_{\bar{a}}
&= \bar{\epsilon}\rho \hat{E}^-_{\bar{a}}
- \bar{\epsilon} \gamma_{\bar{a}} \psi_a \hat{E}^{+a}
- \bar{\epsilon} \gamma_{\bar{a}} \zeta_{\alpha} \hat{E}^{+\alpha} ,
\end{aligned}
\end{equation}
Note that all these equations are independent of $Q$ as required.
\section{$\alpha'$ corrections for heterotic strings}
\label{sec:het-corr}
We shall now turn to our principal example to see how the framework
introduced in the previous section can be used to describe the $\alpha'$
corrections in heterotic strings. A key point is that we will see the
Bianchi identity~\eqref{eq:hetBianchi} appears naturally with the
correct $\Omega^-$ connection. The same formalism can be used to
analyse type II theory and in particular to show why linear $\alpha'$
corrections are ruled out in this case. However we put this discussion
into an appendix~\ref{sec:typeII} for sake of fluency of the presentation.
\subsection{Structures and connections}
\label{sec:het}
The first step is to consider the construction of the previous section
with a product gauge group $G_1\times G_2$. We will eventually
identify $G_1$ with the gauge group $G$ and $G_2$ with $O(n)$ or $GL(n,\mathbb{R})$ as in
section~\ref{sec:nonAbelian2}, but, for the moment, we will keep them
general. We will write their respective adjoint
indices as $\alpha = (\alpha^1,\alpha^2)$, with corresponding
generators $\{t_{\alpha^i}\}$. For the metric on the sum of the Lie
algebras we take the indefinite form $\tr = \tr_1 - \tr_2$ where
$\tr_i t_{\alpha^i}t_{\beta^i}= \eta^i_{\alpha^i\beta^i}$.
The generalised frame defining an $O(n)_+\times G_1\times G_2\times
O(n)_-$ structure and corresponding connection follow from the
expressions in the previous section. The generalised tangent space
decomposes as
\begin{equation}
E = C_+ \oplus C_{\mathfrak{g}_1} \oplus C_{\mathfrak{g}_2} \oplus C_- ,
\end{equation}
We find that~\eqref{eq:splitframe} becomes
\begin{equation}
\begin{aligned}
\label{eq:splitframe2}
\hat{E}^+_a &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\Big(\hat{e}^+_a
+ e^+_a + \iota_{\hat{e}^+_a}B + \iota_{\hat{e}^+_a}A^1
+ \iota_{\hat{e}^+_a}A^2
\\ & \qquad \qquad \qquad \qquad \qquad
- \tfrac{1}{4}\alpha'\tr_1 A^1\,\iota_{\hat{e}^+_a}A^1
+ \tfrac{1}{4}\alpha'\tr_2 A^2\,\iota_{\hat{e}^+_a}A^2 \Big) , \\
\hat{E}^1_{\alpha^1} &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\Big(
\sqrt{\tfrac{4}{\alpha'}}t_{\alpha^1}
- \sqrt{\alpha'}\,\tr_1 t_{\alpha^1}A^1 \Big) , \\
\hat{E}^2_{\alpha^2} &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\Big(
\sqrt{\tfrac{4}{\alpha'}}t_{\alpha^2}
+ \sqrt{\alpha'}\,\tr_2 t_{\alpha^2}A^2 \Big) , \\
\hat{E}^-_{\bar{a}} &= \mathrm{e}^{-2\phi}\sqrt{-g}\,\Big(
\hat{e}^-_{\bar{a}} - e^-_{\bar{a}}
+ \iota_{\hat{e}^-_{\bar{a}}}B + \iota_{\hat{e}^-_{\bar{a}}}A^1
+ \iota_{\hat{e}^-_{\bar{a}}}A^2
\\ & \qquad \qquad \qquad \qquad \qquad
-\tfrac{1}{4}\alpha' \tr_1 A^1\,\iota_{\hat{e}^-_{\bar{a}}}A^1
+ \tfrac{1}{4}\alpha' \tr_2 A^2\,\iota_{\hat{e}^-_{\bar{a}}}A^2
\Big) ,
\end{aligned}
\end{equation}
where $A^i$ is a connection for $G_i$. The generalised connection becomes
\begin{subequations}\label{eq:OdggdLC}
\begin{align}
D_a w_+^b
&= \nabla_a w_+^b
- \tfrac{2}{n-1}\big(
\delta_a{}^b \partial_c\phi-\eta_{ac}\partial^b\phi \big)w_+^c
+ Q_a{}^b{}_c w_+^c ,\label{eq:OdddLC12} \\
D_{\alpha^i} w_+^b &=
Q_{\alpha^i}{}^b{}_a w_+^{a},\label{eq:OdddLC22} \\
D_{\bar{a}} w_+^b &= \nabla_{\bar{a}} w_+^b -
\tfrac{1}{2}H_{\bar{a}}{}^b{}_cw_+^c,
\label{eq:OdddLC32}\\
D_a w_{\mathfrak{g}_i}^{\beta^i}
&= \nabla^i_a w_{\mathfrak{g}_i}^{\beta^i}
+Q_a{}^{\beta^i}{}_{\gamma^i}w_{\mathfrak{g}_i}^{\gamma^i},
\label{eq:OdddLC42}\\
D_{\alpha^i} w_{\mathfrak{g}_j}^{\beta^j} &=
Q_{\alpha^i}{}^{\beta^j}{}_{\gamma^j}w_{\mathfrak{g}_j}^{\gamma^j},
\label{eq:OdddLC52}\\
D_{\bar{a}} w_{\mathfrak{g}_i}^{\beta^i} &=
\nabla^i_{\bar{a}} w_{\mathfrak{g}_i}^{\beta^i},
\label{eq:OdddLC62}\\
D_a w_-^{\bar{b}}
&= \nabla_a w_-^{\bar{b}}
+ \tfrac{1}{2}H_a{}^{\bar{b}}{}_{\bar{c}}w_-^{\bar{c}} ,
\label{eq:OdddLC72}\\
D_{\alpha^i} w_-^{\bar{b}} &= \begin{cases}
-\tfrac{1}{2}\sqrt{\alpha'}F^{1\bar{b}}{}_{\bar{a}}{}_{\alpha^1}
w_-^{\bar{a}} & \text{if $i=1$} \\
\tfrac{1}{2}\sqrt{\alpha'}F^{2\bar{b}}{}_{\bar{a}}{}_{\alpha^2}
w_-^{\bar{a}} & \text{if $i=2$}
\end{cases} , \label{eq:OdddLC82}\\
D_{\bar{a}} w_-^{\bar{b}} &= \nabla_{\bar{a}} w_-^{\bar{b}}
+ \tfrac{1}{6}H_{\bar{a}}{}^{\bar{b}}{}_{\bar{c}}w_-^{\bar{c}}
- \tfrac{2}{n-1}\big(
\delta_{\bar{a}}{}^{\bar{b}} \partial_{\bar{c}}\phi
- \eta_{\bar{a}\bar{c}}\partial^{\bar{b}}\phi \big)w_-^{\bar{c}}
+ Q_{\bar{a}}{}^{\bar{b}}{}_{\bar{c}} w_-^{\bar{c}} ,\label{eq:OdddLC92}
\end{align}
\end{subequations}
In order to match to the $\alpha'$ theory, we need to make a further
reduction of the structure group. The crucial step is that we identify the $G_2$ structure group with the $O(n)_+$ structure group. In the case were $G_2=O(n)$ we can just directly identity the two. If $G_2 = GL(n,\mathbb{R})$, one can think of first reducing to $O(n)$ and then identifying with the $O(n)_+$ structure group. In this latter case, the reduction $O(n)\subset GL(n,\mathbb{R})$ specifies a ``metric'', which is to be identified with the gravitational metric $g$. In either case we end up with a refined structure
\begin{equation}
O(n)_+ \times G_1 \times G_2 \times O(n)_-
\supset O(n)_+ \times G \times O(n)_- ,
\end{equation}
where now we identify $G_1=G$, the heterotic gauge group. Equivalently we identify
$C_{\mathfrak{g}_2}\simeq \Lambda^2C_+$, that is we consider
\begin{equation}
E = C_+ \oplus C_{\mathfrak{g}} \oplus \Lambda^2C_+ \oplus C_- ,
\end{equation}
Concretely the $\alpha^2$ index now corresponds to an antisymmetric
pair of vector indices $[ab]$, and writing just $\alpha$ for the
$G_1=G$ index we
have
\begin{equation}
\alpha^1 = \alpha , \qquad
\alpha^2 = [ab] ,
\end{equation}
so, writing $w_{\mathfrak{g}_1}^{\alpha^1}=w^\alpha_{\mathfrak{g}}$ and
$w^{\alpha^2}_{\mathfrak{g}_2}=w_+^{ab}$, the components of the
generalised vector are $W=(w_+^a,w_{\mathfrak{g}}^\alpha,w_+^{ab},w_-^{\bar{a}})$.
We can then rewrite~\eqref{eq:OdddLC32} and~\eqref{eq:OdddLC62} as
\begin{equation}
\label{eq:matchD}
\begin{aligned}
D_{\bar{a}} w_+^b
&= \partial_{\bar{a}} w_+^b + \Omega^-_{\bar{a}}{}^b{}_cw_+^c, \\
D_{\bar{a}}w_+^{ab}
&= \partial_{\bar{a}}w_+^{ab}
+ A^2_{\bar{a}}{}^{a}{}_{c}w_+^{cb}
+ A^2_{\bar{a}}{}^{b}{}_{c}w_+^{ac} ,
\end{aligned}
\end{equation}
where $\Omega^-=\omega^{\text{LC}}-\frac{1}{2}H$, with
$\omega^{\text{LC}}$ the Levi--Civita connection.
We then make the further natural requirement that the generalised
connection is compatible with the $O(n)_+\times G\times O(n)_-$
structure. This requires that two derivatives in~\eqref{eq:matchD}
agree, that is
\begin{equation}
A^2 = \Omega^- = \omega^{\text{LC}} - \tfrac{1}{2}H .
\end{equation}
With this identification the generalised derivative is given by
\begin{equation}
\label{eq:hetD}
\begin{aligned}
D_a w_+^b
&= \nabla_a w_+^b - \tfrac{2}{n-1}\big(
\delta_a{}^b \partial_c\phi-\eta_{ac}\partial^b\phi \big)w_+^c
+ Q_a{}^b{}_c w_+^c ,\\
D_{\alpha} w_+^b &= Q_{\alpha}{}^b{}_a w_+^{a}, \\
D_{aa'} w_+^b &= Q_{aa'}{}^b{}_c w_+^c, \\
D_{\bar{a}} w_+^b &= \nabla_{\bar{a}} w_+^b -
\tfrac{1}{2}H_{\bar{a}}{}^b{}_cw_+^c, \\
D_a w_{\mathfrak{g}}^{\beta}
&= \nabla_a w_{\mathfrak{g}}^{\beta}
+Q_a{}^{\beta}{}_{\gamma}w_{\mathfrak{g}}^{\gamma}, \\
D_{\alpha} w_{\mathfrak{g}}^{\beta} &=
Q_{\alpha}{}^{\beta}{}_{\gamma}w_{\mathfrak{g}}^{\gamma}, \\
D_{aa'} w_{\mathfrak{g}}^{\beta} &=
Q_{aa'}{}^{\beta}{}_{\gamma}w_{\mathfrak{g}}^{\gamma}, \\
D_{\bar{a}} w_{\mathfrak{g}}^{\beta} &=
\nabla_{\bar{a}} w_{\mathfrak{g}}^{\beta}, \\
D_a w_-^{\bar{b}}
&= \nabla_a w_-^{\bar{b}}
+ \tfrac{1}{2}H_a{}^{\bar{b}}{}_{\bar{c}}w_-^{\bar{c}} , \\
D_{\alpha} w_-^{\bar{b}} &=
-\tfrac{1}{2}\sqrt{\alpha'}F^{\bar{b}}{}_{\bar{c}}{}_{\alpha}
w_-^{\bar{c}} , \\
D_{aa'} w_-^{\bar{b}} &=
\tfrac{1}{2}\sqrt{\alpha'}R(\Omega^-)^{\bar{b}}{}_{\bar{c}aa'}
w_-^{\bar{c}} \\
D_{\bar{a}} w_-^{\bar{b}} &= \nabla_{\bar{a}} w_-^{\bar{b}}
+ \tfrac{1}{6}H_{\bar{a}}{}^{\bar{b}}{}_{\bar{c}}w_-^{\bar{c}}
- \tfrac{2}{n-1}\big(
\delta_{\bar{a}}{}^{\bar{b}} \partial_{\bar{c}}\phi
- \eta_{\bar{a}\bar{c}}\partial^{\bar{b}}\phi \big)w_-^{\bar{c}}
+ Q_{\bar{a}}{}^{\bar{b}}{}_{\bar{c}} w_-^{\bar{c}} ,
\end{aligned}
\end{equation}
where $R(\Omega^-)$ is the curvature of the $\Omega^-$
connection. Compatibility means that the $D_Mw_+^{bb'}$ expressions
follow directly from those for $D_Mw^b$, namely
\begin{equation}
\begin{aligned}
D_a w_+^{bb'}
&= \nabla_a w_+^{bb'} - \tfrac{2}{n-1}\big(
\delta_a{}^b \partial_c\phi-\eta_{ac}\partial^b\phi \big)w_+^{cb'}
+ Q_a{}^b{}_c w_+^{cb'} \\ & \qquad
- \tfrac{2}{n-1}\big(
\delta_a{}^{b'} \partial_c\phi-\eta_{ac}\partial^{b'}\phi \big)w_+^{bc}
+ Q_a{}^{b'}{}_c w_+^{bc} ,\\
D_{\alpha} w_+^{bb'} &= Q_{\alpha}{}^b{}_c w_+^{cb'}
+ Q_{\alpha}{}^{b'}{}_c w_+^{bc}, \\
D_{aa'} w_+^{bb'} &= Q_{aa'}{}^b{}_c w_+^{cb'}
+ Q_{aa'}{}^{b'}{}_c w_+^{bc} , \\
D_{\bar{a}} w_+^{bb'} &= \nabla_{\bar{a}} w_+^{bb'}
- \tfrac{1}{2}H_{\bar{a}}{}^b{}_cw_+^{cb'}
- \tfrac{1}{2}H_{\bar{a}}{}^{b'}{}_cw_+^{bc} .
\end{aligned}
\end{equation}
As before the undetermined parts of that connection satisfy
\begin{equation}
\begin{aligned}
Q_{a}{}^{a}{}_b&= 0 , & &&
Q_{\bar{a}}{}^{\bar{a}}{}_{\bar{b}}&= 0, & &&
Q_{\alpha}{}^{\alpha}{}_{\beta}&=0, & &&
Q_{ac}{}^c{}_b &=0 .
\end{aligned}
\end{equation}
\subsection{Supergravity equations}
\label{sec:hetEQ}
As before the supergravity equations of motion and supersymmetry
variations all follow from the generalised connection $D$ defined
in~\eqref{eq:hetD}. The action comes from the Bismut-type
equation~\eqref{eq:newBismut}. Specialising to the case in hand of a
product group $G_1\times G_2=G\times O(n)$ with $\tr=\tr_1-\tr_2$, gives
\begin{equation}
\label{eq:newBismut2}
\gamma^{\bar{a}}D_{\bar{a}}\gamma^{\bar{b}}D_{\bar{b}}\epsilon
= D^aD_a\epsilon+D^{\alpha}D_{\alpha}\epsilon
- D^{ab}D_{ab}\epsilon
- \tfrac14 S^-\epsilon ,
\end{equation}
where we have fixed the normalisation of
the $O(n)$ trace $\tr t_{aa'}t_{bb'}=\delta_{ab}\delta_{a'b'}$. As
before, this equation only defines a scalar $S^-$ if the corresponding
Bianchi condition holds. Here this reads
\begin{equation}
\label{eq:hetBianchi*}
\mathrm{d} H = \tfrac{1}{4} \alpha' \left[\tr F\wedge F -
\tr R(\Omega^-)\wedge R(\Omega^-)\right] ,
\end{equation}
reproducing the $\alpha'$ corrected condition with the correct
connection $\Omega^-$. Calculating the scalar $S^-$, one finds the
$\alpha'$ corrected bosonic action (or equivalently the dilaton
equation of motion)
\begin{equation}
\label{eq:BosonicAction22}
S^- = s +4 \nabla^2 \phi -4 (\partial \phi)^2
- \tfrac{1}{12} H^2 -\tfrac{1}{8}\alpha'\tr F^2
+ \tfrac{1}{8}\alpha'\tr R(\Omega^-)^2 ,
\end{equation}
with again the appropriate curvature-squared term.
The supersymmetry variation of the fermion fields
follow as before from the set of uniquely determined operators
\begin{equation}
\label{eq:FermVar2}
\begin{aligned}
\delta\psi_a &= D_a \epsilon
= \nabla_a\epsilon + \tfrac{1}{8}H_{a\bar{b}\bar{c}}
\gamma^{\bar{b}\bar{c}} \epsilon , \\
\delta\zeta_{\alpha} &= D_{\alpha}\epsilon
= -\tfrac18 \sqrt{\alpha'}F_{\bar{a}\bar{b}\alpha}
\gamma^{\bar{a}\bar{b}}\epsilon , \\
\delta\rho &= \gamma^{\bar{a}}D_{\bar{a}}\epsilon
= \gamma^{\bar{a}}\nabla_{\bar{a}} \epsilon
+ \tfrac{1}{24}H_{\bar{a}\bar{b}\bar{c}}
\gamma^{\bar{a}\bar{b}\bar{c}}\epsilon
-\gamma^{\bar{a}}(\partial_{\bar{a}}\phi)\epsilon .
\end{aligned}
\end{equation}
The formalism naturally has, in addition, a ``gaugino''
$\psi_{ab}\in\Gs{\Lambda^2C_+\otimes S(C_-)}$ for the $G_2=O(n)_+$
gauge group. The corresponding variation is
\begin{equation}
\delta\psi_{ab} = D_{ab}\epsilon
= \tfrac18
\sqrt{\alpha'}R(\Omega^-)_{\bar{a}\bar{b}ab}
\gamma^{\bar{a}\bar{b}}\epsilon .
\end{equation}
The obvious interpretation is that $\psi_{ab}$ is the standard
composite ``gravitino curvature''~\cite{BdR}, given to leading order
in the fermions by
\begin{equation}\label{eq:gravitino-curv-def}
\psi_{mn} = \tfrac12 \sqrt{\alpha'}\left(\partial_m \psi_n - \partial_n \psi_m
+ \tfrac{1}{4} \Omega^+_{mpq} \gamma^{pq} \psi_{n}
- \tfrac{1}{4} \Omega^+_{npq} \gamma^{pq} \psi_{m} \right).
\end{equation}
Calculating the corresponding variation one finds~\cite{BdR}
\begin{equation}
\label{eq:gravitino-curv}
\delta\psi_{ab} = \tfrac18\sqrt{\alpha'}
R(\Omega^+)_{ab\bar{a}\bar{b}}
\gamma^{\bar{a}\bar{b}}\epsilon
=D_{ab}\epsilon + \mathcal{O}(\alpha^{\prime }) .
\end{equation}
as required, where we have used the fact that
\begin{equation}
R_{mnpq}(\Omega^-) = R_{pqmn}(\Omega^+) - \tfrac12 \mathrm{d} H_{mnpq}
= R_{pqmn}(\Omega^+) + \mathcal{O}(\alpha') .
\end{equation}
We will discuss briefly the question of the higher order $\alpha'$
corrections to~\eqref{eq:gravitino-curv} in the next section.
For the bosonic variations we have
\begin{equation}
\label{eq:Evar2}
\begin{aligned}
\delta\hat{E}^+_a
&= \bar{\epsilon} \rho \hat{E}^+_a
- \bar{\epsilon} \gamma_{\bar{a}} \psi_a \hat{E}^{-\bar{a}}, \\
\delta\hat{E}_{\alpha}
&= \bar{\epsilon} \rho \hat{E}^1_{\alpha}
- \bar{\epsilon} \gamma_{\bar{a}} \zeta_{\alpha} \hat{E}^{-\bar{a}} , \\
\delta\hat{E}^-_{\bar{a}}
&= \bar{\epsilon}\rho \hat{E}^-_{\bar{a}}
- \bar{\epsilon} \gamma_{\bar{a}} \psi_a \hat{E}^{+a}
- \bar{\epsilon} \gamma_{\bar{a}} \zeta_{\alpha} \hat{E}^{\alpha}
+ \bar{\epsilon} \gamma_{\bar{a}} \psi_{ab} \hat{E}^{ab} ,
\end{aligned}
\end{equation}
and for the $O(n)$ basis
\begin{equation}
\delta\hat{E}_{ab}
= \bar{\epsilon} \rho \hat{E}_{ab}
- \bar{\epsilon} \gamma_{\bar{a}} \psi_{ab} \hat{E}^{-\bar{a}} ,
\end{equation}
which is equivalent to variation of the composite object
\begin{equation}
\label{eq:Omega-var}
\tfrac12 \sqrt{\alpha'}\,\delta \Omega^-_{mab}
= -\bar{\epsilon} \gamma_{m} \psi_{ab} + \mathcal{O}(\alpha^{\prime}),
\end{equation}
which indeed follows from the other variations. We are reproducing the
standard result~\cite{BdR} that $(\psi_{mn}, \Omega^-_m)$ indeed
transform as a gauge multiplet.
Finally we can also write the fermionic action including higher
derivative terms for $\psi_m$ following~\eqref{eq:FermionicAction}, as
\begin{equation}
\begin{aligned}
\label{eq:FermionicAction2}
S_F = -\frac{1}{2\kappa^2}\int 2 \vol_G &\Big[
\bar\psi^{a} \gamma^{\bar{b}} D_{\bar{b}} \psi_{a}
+ 2 \bar\rho D_a \psi^{a}
- \bar\rho \gamma^{\bar{a}} D_{\bar{a}} \rho
+ 2 \bar\rho D_{\alpha} \zeta^{\alpha}
- 2 \bar\rho D_{ab} \psi^{ab} \\
&+ \bar\zeta^{\alpha} \gamma^{\bar{b}} D_{\bar{b}}
\zeta^1_{\alpha}
- \bar\psi^{ab} \gamma^{\bar{b}} D_{\bar{b}} \psi_{ab} \\
&+ \bar\zeta^{\alpha} \gamma^{\bar{b}}
\hat{T}_{\bar{b}}{}^{a}{}_{\alpha} \psi_{a}
- \bar\psi^{ab}\gamma^{\bar{b}}
\hat{T}_{\bar{b}}{}^{c}{}_{ab} \psi_{c}
+\bar\psi^{a} \gamma^{\bar{b}}
\hat{T}_{\bar{b}}{}^{\alpha}{}_a \zeta_{\alpha}
-\bar\psi^{a} \gamma^{\bar{b}}
\hat{T}_{\bar{b}}{}^{bc}{}_a \psi_{bc}
\Big] ,
\end{aligned}
\end{equation}
where the intrinsic torsion terms are given by
\begin{equation}
\hat{T}_{\bar{a}b\gamma}
= -\tfrac{1}{2}\sqrt{\alpha'} F_{\bar{a}b\gamma} , \qquad
\hat{T}_{\bar{a}bcc'}
= \tfrac{1}{2}\sqrt{\alpha'} R(\Omega^-)_{\bar{a}bcc'} .
\end{equation}
The bosonic equations of motion then come from varying the
corresponding fermionic equations and are given by vanishing of
the components of the corresponding generalised Ricci tensor, defined
by
\begin{equation}
\begin{aligned}
\label{eq:Bosoniceoms3}
\tfrac12 R_{\bar{a}b}\gamma^{\bar{a}}\epsilon
&= [\gamma^{\bar{a}}D_{\bar{a}}, D_b]\epsilon
+ \gamma^{\bar{a}}\hat{T}_{\bar{a}}{}^{\beta}{}_b D_{\beta}
\epsilon
- \gamma^{\bar{a}}\hat{T}_{\bar{a}}{}^{cc'}{}_b D_{cc'}
\epsilon , \\
\tfrac12 R_{\bar{a}\beta}\gamma^{\bar{a}}\epsilon
&= [\gamma^{\bar{a}}D_{\bar{a}}, D_{\beta}]\epsilon
+ \gamma^{\bar{a}}\hat{T}_{\bar{a}}{}^{b}{}_{\beta} D_{b}
\epsilon , \\
\tfrac12 R_{\bar{a},bb'}\gamma^{\bar{a}}\epsilon
&= [\gamma^{\bar{a}}D_{\bar{a}}, D_{bb'}]\epsilon
- \gamma^{\bar{a}}\hat{T}_{\bar{a}}{}^{c}{}_{bb'} D_{c}
\epsilon .
\end{aligned}
\end{equation}
The full set of independent bosonic field equations are then
\begin{equation}
S^- = R_{\bar{a}b} = R_{\bar{a}\beta} = 0 ,
\end{equation}
while the composite equation $R_{\bar{a},bb'}=0$ follows identically from the previous
to zeroth order in $\alpha'$, a well-known result from~\cite{BdR}.
\subsection{Higher orders in $\alpha'$ }
\label{sec:higher}
In~\cite{BdR} a quartic action with $(\alpha')^3$ corrections was derived, and in fact an iterative procedure that can in principle be used to generate an entire family of corrections with arbitrarily high powers of $\alpha'$ was pointed out. The procedure is based on the supersymmetric completion of the heterotic Bianchi identity \eqref{eq:hetBianchi}, and the key observation is that an identification of the auxiliary $O(n)$ Yang-Mills gauge fields with the gravitational degrees of freedom is possible.\footnote{Note that only the corrections to the Bianchi identity considered in \cite{BdR} are linear in $\alpha'$. The Bianchi identity itself does not change and is used in the process of iteration.} However, in order to make such an identification one has to equate the ``gaugino'' with the gravitino curvature~\eqref{eq:gravitino-curv-def}. On the other hand, as we have already seen, this identification is not quite consistent with supersymmetry variation we had for the gaugino so far -- in fact it only works to order $\alpha'$
\begin{equation}
\delta\psi_{ab} = D_{ab}\epsilon + \mathcal{O}(\alpha^{\prime }) .
\end{equation}
This is a consequence of the heterotic Bianchi identity \eqref{eq:hetBianchi}, since in order to make the identification of the gaugino one has to use the relation $R_{mnpq}(\Omega^-) - R_{pqmn}(\Omega^+) = -12 \mathrm{d} H_{mnpq}$. Explicitly, the correction to the variation of the gravitino curvature is then
\begin{equation}\label{eq:gravitino-curv-def2}
\delta\psi_{ab} = D_{ab}\epsilon +\tfrac18\sqrt{\alpha'} \left( \tfrac18\alpha'[\tr F\wedge F-\tr R(\Omega^-)\wedge R(\Omega^-)]_{ab\bar{a}\bar{b}}\right) \gamma^{\bar{a}\bar{b}}\epsilon.
\end{equation}
With this change, though, supersymmetry no longer closes on the action. Indeed, in~\cite{BdR} it was found that in order to restore supersymmetry it is necessary to introduce corrections to the supersymmetry transformations of the ``fundamental'' (i.e. non-composite) fields at order $(\alpha')^2$, and add new quartic terms to the action at order $(\alpha')^3$. The resulting theory is then supersymmetric, but only up to order $(\alpha')^3$ since the corrections to the gravitino variation mean that~\eqref{eq:gravitino-curv-def2} would get new $(\alpha')^3$ corrections -- which would then break supersymmetry again and require new higher-order corrections to the variations and the action, and so on. Clearly this is a process that can, in theory, be repeated to generate ever higher $\alpha'$ corrections.
In our context these issues can be viewed in a slightly different manner. Let us denote the corrected variation $\delta\psi_{ab} = D_{ab}\epsilon + \mathcal{O}(\alpha^{\prime }) = \hat{D}_{ab}\epsilon$. Our guiding principle for writing bosonic actions is the Bismut formula
\begin{equation}
\gamma^{\bar{a}}D_{\bar{a}}\gamma^{\bar{b}}D_{\bar{b}}\epsilon
- D^aD_a\epsilon+D^{\alpha}D_{\alpha}\epsilon
+ D^{ab}D_{ab}\epsilon
=- \tfrac14 S^-\epsilon.
\end{equation}
However, if we were to simply replace $D_{ab}$ with $\hat{D}_{ab}$, the equation will no longer define just a ($(\alpha')^3$-corrected) scalar $S^-$, but it would also generate on the right-hand side a two-form and a four-form at order $(\alpha')^2$. (This is just an equivalent, and somewhat more straightforward way of seeing that the action is no longer supersymmetric.) We would then be naturally led to also modify the supersymmetry rules~\eqref{eq:FermVar2} at order $(\alpha')^2$ and introduce corrected operators $\hat{D}_a\epsilon$ and $\gamma^{\bar{a}}\hat{D}_{\bar{a}}\epsilon$ such that they precisely cancel the two-forms and four-forms in the Bismut formula, at least to order $(\alpha')^2$. As in~\cite{BdR}, this procedure would give rise to an iterative rule for generating higher $\alpha'$ corrections. We hope to provide some results in this direction in upcoming work.
We would like to emphasise that these all-orders-in-$\alpha'$ corrections to the effective action originate from a single extension of the generalised tangent bundle, that captures the contribution of the heterotic Bianchi identity \eqref{eq:hetBianchi*} linear in $\alpha'$. This aspect of our construction is fully concordant with the construction of \cite{BdR}.
\section{$\alpha'$ corrections and brackets}
\label{sec:qb!}
Our starting point for deriving the possible $\alpha'$ gravitation corrections has been to view $O(n)$ or $\GL(n,\mathbb{R})$ as a ``gauge group'', build an appropriate extension of the generalised tangent bundle, and hence enlarge the set of symmetries parametrised by the generalised tangent space. The additional symmetries are simply local changes of frame.
However, one could also consider a different parametrisation, where the additional frame rotation symmetries are induced by diffeomorphisms, and so are not independent parameters. This is particularly interesting when flows and the related Lie algebroids are considered.
Vector fields generate diffeomorphisms by generating flows on manifolds. These diffeomorphisms in turn induce a $\GL(n,\mathbb{R})$ action on the frame bundle of the manifold $F$. To see this, choose a frame $e^a$ and its dual $\hat e_b$. Under the flow generated by some vector field $v$ the frame $e^a$ will be rotated by an element in $\GL(n,\mathbb{R})$, which to leading order is just given by its Lie derivative along $v$. Therefore, if we call the generator of this rotation $Lv$, we find for its components
\begin{equation}\label{eq:rotmatrix_explicit}
(Lv)^a{}_b = \iota_{\hat e_b} {\cal L}_v e^a \ ,
\end{equation}
in terms of a basis of $\GL(n,\mathbb{R})$ generators $\hat{u}^{b}{}_a$ which are given in terms of the frame $\hat{e}_a$ by $\hat{u}^b{}_a =e^b\otimes \hat{e}_a$.
Note that \eqref{eq:rotmatrix_explicit} depends on the choice of a particular frame $e^a$. However, changing to a different frame corresponds to a global $\GL(n,\mathbb{R})$ transformation which just transforms the generators $\hat{u}^{b}{}_a$ into each other. Therefore, the choice of frame is part of the gauge choice, and the construction will not depend on it.
The idea now is to take the gravitational gauge bundle $\tilde{P}$ to have gauge group $\tilde{G}=GL(n,\mathbb{R})$ and identify it with the frame bundle. We can then restrict the corresponding gauge parameter in the generalised tangent space to being of the form $\tilde \Lambda = Lv$.
Let us then discuss how this can be accomplished in more detail. For simplicity we will ignore for the moment the gauge group $G$, which we already know how to handle, and keep the focus on just the frame bundle. Generic infinitesimal $\tilde{G}$-equivariant automorphisms of $F$ can be understood in terms of the bundle $\mathcal{A}=TF/\tilde{G}\simeq TM\oplus\adj{F}$ (that is, the Atiyah algebroid for $F$) which inherits a Lie algebroid structure from the Lie bracket algebra on $TF$ itself. Elements of $\mathcal{A}$ are simply combinations of vectors $v$ and gauge parameters $\tilde{\Lambda}$ patched just as in~\eqref{eq:Epatch}. In fact, the extended generalised vector space $E$ is often thought of as an extension $T^*M\to E\to \mathcal{A}$.
Given a vector $v\in \Gs{TM}$ there is then a natural lift to an element $\tilde{v}\in\Gs{\mathcal{A}}$ given locally by
\begin{equation}
\tilde{v} = v + Lv ,
\end{equation}
where the components of the gauge part $(Lv)^a{}_b$ are precisely the ones given in \eqref{eq:rotmatrix_explicit}. This has the property that $\widetilde{[v,w]}=[\tilde{v},\tilde{w}]$, where $[\tilde{v},\tilde{w}]$ is the bracket on the Lie algebroid $\mathcal{A}$. One can check that the gauge part $\tilde\Lambda=Lv$ indeed patches correctly.
Note that one cannot think of $\tilde{v}$ as being a section of some sub-bundle of $\mathcal{A}$, since its definition involves derivatives of $v$ itself. It is merely an element of the subspace of sections of $\mathcal{A}$ defined to consist of those $(v, \tilde \Lambda)$ for which the condition $\tilde \Lambda = Lv$ with \eqref{eq:rotmatrix_explicit} holds.
As always, one needs a splitting to identify the isomorphism $\mathcal{A}\simeq TM\oplus\adj{F}$, which here is given by an arbitrary choice of $\GL(n,\mathbb{R})$-connection $\omega$ for the frame with the associated covariant derivative $\nabla$. Using the identity $\iota_{\hat e_b} {\cal L}_v e^a = (\nabla_b v^a) - \iota_v\omega^a{}_b$ we write
\begin{equation}
\tilde{v} = (v - \iota_v\omega) + (\nabla_b v^a)\hat{u}^b{}_a ,
\end{equation}
and identify $\nabla_b v^a$ as the element of $\adj{F}$. This is just an example of the gauge shifts we observed in \eqref{eq:automorphisms}.
One might wonder whether the same reparametrisation can be achieved for a choice $\tilde{G}=O(n)$, that is considering instead the orthogonal frame bundle $P\subset F$ with fibre $O(n)$. However this does not work since $P$ does not admit natural lifts of vectors on $M$. The closest related notion, the Kosmann lift~\cite{Kosmann} $v^K$ corresponding to lowering an index of~\eqref{eq:rotmatrix_explicit} and taking the antisymmetric part, does not satisfy $[v,w]^K = [v^K,w^K]$, where $v^K,w^K \in \Gs{TP/O(n)}$ are the Kosmann-lifted vectors, unless either $v$ or $w$ are Killing (i.e. the Kosmann lift actually provides an homomorphism from the Lie algebra of infinitesimal isometries of $M$ to infinitesimal automorphisms of $P$). For a discussion see for example~\cite{LieSpinor}.
The construction that we have described so far can be easily embedded into the construction of $E$ as described in section \ref{sec:nonAb}. In particular, we can reproduce the inner product \eqref{eq:metric2gauge} and the generalised Lie derivative \eqref{eq:Bgen2gauge} after lifting to $\mathcal{A}$ and subsequently to $E$. From the identification $\tilde\Lambda=Lv$, we take generalised vectors of the form $V = v+ Lv + \lambda$ and find for the scalar product
\begin{equation}
\label{eq:metricgravity}
\met{V}{V'}
= \tfrac12\iota_v \lambda' + \tfrac12\iota_{v'} \lambda
+ \tr (Lv Lv') ,
\end{equation}
while the generalised Lie derivative (or equivalently the Courant bracket) becomes
\begin{equation}
\label{eq:genLiegaugediff}
\begin{aligned}
\mathbb{L}_V V =
& \comm{v}{v'} + \mathcal{L}_v\lambda'
- \iota_{v'}\mathrm{d}\lambda
+ \tfrac14 \alpha'\tr (Lv') \mathrm{d} (Lv)
+ \comm{Lv}{Lv'} .
\end{aligned}
\end{equation}
Naively the $Lv$ terms in~\eqref{eq:metricgravity} and~\eqref{eq:genLiegaugediff} look like higher-derivative corrections to the inner product and generalised Lie derivative on the conventional, unextended generalised tangent space $E'\simeq TM\oplus T^*M$. However, this interpretation is misleading precisely because $V=v+Lv+\lambda$ is really part of a larger generalised tangent space. Crucially, the one-form components $\lambda$ are patched in a way that depends non-trivially on the $\tilde{G}$-bundle patching. They are not the same objects that appear in the conventional generalised tangent space $E'$.\footnote{Of course we have the isomorphisms $E'\simeq TM\oplus T^*M$ and $E\simeq TM\oplus T^*M\oplus\adj{F}$. Thus given a choice of splitting for each bundle, that is a $B$-fields $B'$ for $E'$ and $B$ for $E$ (with different patching properties and Bianchi identities!) it is possible to map between objects in $E'$ and $E$, but there is no such natural choice.} We see that the only way to make sense of such higher-derivative corrections, both from a diffeomorphism-invariant perspective and from the patching, is to realise that they really come from the extended bundle $E$.
One should note that working in a local coordinate frame where $\hat{e}_a=\delta_a^\mu\partial_\mu$ we have $Lv^a{}_b=\partial_bv^a$ and strikingly~\eqref{eq:metricgravity} and the vector and one-form part of ~\eqref{eq:genLiegaugediff} become exactly the DFT $\alpha'$-corrected expressions given in~\cite{Hohm:2013jaa}. The above construction gives a description of how to realize these $\alpha'$-corrected generalised Lie derivatives and inner products on a general curved manifold, and suggests that corrections linear in $\alpha'$ have to be of the described kind.
Another variation is to consider taking $\tilde{G}$ to be the full $O(n,n)$ group and lift not vectors $v$ but conventional generalised vectors $V=v+\lambda\in \Gs{E'}$ into $O(n,n)$ gauge transformations. This possibility was first pointed out in section 4.2 of~\cite{Bedoya:2014pma} in a DFT context. Taking $E\simeq TM\oplus T^*M\oplus \adj\tilde{F}$, where $\tilde{F}$ with fibre $O(n,n)$ is the generalised frame bundle, they defined the lifted vector $V+\Lambda\in \Gs{E}$ with $\Lambda^M{}_N = \tfrac12 (\partial_N V^M-\partial^M V_N)$. Again this is a coordinate-dependent expression, but nonetheless substituting into, for instance, the definition of the scalar product, they obtained terms like
\begin{equation}
\tfrac14(\partial_N V^M-\partial^M V_N)(\partial_M V'^N-\partial^N V'_M)=\tfrac12(\partial_N V^M-\partial^M V_N)(\partial_M V'^N)=\tfrac12 (\partial_n v^m)(\partial_m v'^n) ,
\end{equation}
where the last equality follows from the ``strong constraint'' of DFT. This term then matches the $GL(n,\mathbb{R})$ transformation we considered and hence also gives the $\alpha'$ correction discussed in~\cite{Hohm:2013jaa}.
Again we can try and make this description properly covariant by introducing the generalised geometric analogue of the $Lv$ lift defined above. If $\hat{U}^{AB}=\hat{E}^A\otimes\hat{E}^B$ is a basis for $\adj\tilde{F}$ defined by the generalised frame $\hat{E}_A$, then we define the lifted object
\begin{equation}
\tilde{V} = V + LV,
\end{equation}
where $LV=(LV)_{AB}\hat{U}^{AB}$ with\footnote{Note that the form of the generalised Lie derivative means that $LV$ is actually only an element of the Lie algebra of the geometric subgroup $G_{\text{geom}}\subset O(n,n)$, corresponding to $\mathfrak{gl}(n,\mathbb{R})$ transformations combined with $B$-shifts.}
\begin{equation}
(LV)_{AB} = \met{\hat E_B}{ \mathbb{L}_V \hat E_A}.
\end{equation}
One then finds, for example, that there is a generalised Lie derivative on $E$ induced from that on $E'$, given by
\begin{equation}
\label{eq:tilde-LD}
\mathbb{L}_{\tilde{V}}\tilde{W}
:= \widetilde{\mathbb{L}_V W}
= \mathbb{L}_V W + [LV,LW]+ \met{V}{\mathrm{d} (LW)}-\met{W}{\mathrm{d} (LV)} .
\end{equation}
Although one still has $[\mathbb{L}_{\tilde{V}},\mathbb{L}_{\tilde{W}}]=\mathbb{L}_{\mathbb{L}_{\tilde{V}}\tilde{W}}$ and $\mathbb{L}_{\tilde{U}}\met{\tilde{V}}{\tilde{W}}=\met{\mathbb{L}_{\tilde{U}} \tilde{V}}{\tilde{W}}+\met{\tilde{V}}{\mathbb{L}_{\tilde{U}} \tilde{W}}$, crucially the third condition in~\eqref{eq:CA-def} required for~\eqref{eq:tilde-LD} to define a Courant algebroid is not satisfied. This is precisely because the terms of the form $2\tr LW \mathrm{d}(LV)$ are missing. This is the same problem mentioned above: there is no natural map from $E'\simeq TM\oplus T^*M$ to $E$ respecting the generalised Lie derivative structures on each. Nonetheless the existence of the lift $LV$ from $E$ to the generalised frame bundle $\tilde{F}$ is perhaps worthy of further study. In addition, it is interesting to note that the $E_{7(7)}$ generalised Lie derivative~\cite{CSW-11d} also fails to satisfy the corresponding condition, but nonetheless defines the correct structure to describe the relevant truncations of eleven-dimensional and type II supergravities.
Returning to our original reparametrisation, we again stress that the relation between gauge transformations $\tilde{\Lambda}^a{}_b=(Lv)^a{}_b$ of $\tilde G$ and diffeomorphisms does not define a subbundle of the bundle defined by~\eqref{eq:Epatch2gauge}, as it involves derivatives of the section. Note also that all the discussion in this section refers to the underlying differential structure before one introduces any of the dynamical degrees of freedom. For this reason it gives no information relating to the particular choice of connection $\Omega^-$ that appears in the gravitational part of the Bianchi identity.
\section{General formalism}
\label{sec:gen}
Our general philosophy has been that $\alpha'$ corrections are naturally encoded by first considering further extensions of the generalised tangent space with a corresponding generalised Lie derivative, then defining a particular class of generalised connections, and finally using a Bismut-type formula to write the action. A test of the construction would be its usefulness in ordering higher $\alpha'$ terms. The first observation is that at the very first step of defining the extended generalised tangent space there can be obstructions. To leading order in the heterotic theory these were that the Pontryagin classes of the tangent and gauge bundle cancelled. At higher $(\alpha')^n$ order, simply on dimensional ground one would expect obstructions involving classes in $n+1$ powers of the curvature. Since $\mathrm{d} H$ is a four-form it is hard to envisage any higher order obstructions involving the Bianchi identity. Interestingly, at order $(\alpha')^3$ though one might though have an obstruction encoding the $B\wedge X_8$ terms.
More generally one might ask whether there is a general framework for
describing such extensions and in particular if the Bismut-type
formulae~\eqref{eq:newBismut} and~\eqref{eq:newBismut2} are generic
features. One approach is motivated by the structure of gauged
supergravity theories. Generically the gauging modifies the
supersymmetry transformations, adding new fermionic variations of the
form
\begin{equation}
\label{eq:susy-mod}
\delta'_\epsilon\Psi = A \cdot \epsilon , \qquad \qquad
\delta'_\epsilon\chi = B \cdot \epsilon ,
\end{equation}
where $\Psi_\mu$ are the gravitinos in the theory and $\chi$ the other
spin-$\frac12$ fermions, and $A$ and $B$ are generic matrices. There
is then a supersymmetric ``Ward identity''~\cite{s-ward,embed-lecture}
that relates these variations to the potential $V$, namely
\begin{equation}
\label{eq:sward}
B^\dag B - A^\dag A = V \, \mathbf{1} ,
\end{equation}
which reminiscent of the Bismut-type formulae. These relations are
actually best described using the embedding tensor
formalism~\cite{embed} (for a review see~\cite{embed-lecture}).
As we will now show, all generalised geometric constructions can
actually be rephrased as very particular infinite-dimensional versions
of the embedding tensor formalism. In this framework, the Bismut-type
relation follows directly from the supersymmetric Ward
identity~\eqref{eq:sward}. It also provides a formalism for addressing
how these objects might be generalised to describe higher-order corrections.
That there is a relation between generalised geometry and the
embedding tensor was already observed in~\cite{CSW-11d}, and the
formalism also played a crucial r\^ole in the reformulations given
in~\cite{EFT}, and in the discussions in~\cite{ext-geom}. Here,
though, we go a step further and show that generalised geometry can be
viewed precisely as an infinite-dimensional version of the embedding
tensor construction.
We start by recalling the basic ingredients of the formalism. One
begins with an ungauged supergravity theory in $n$ dimensions with a
global symmetry group $\mathcal{G}$ and an R-symmetry $\mathcal{H}$,
and (at least) the following content
\begin{equation}
\begin{aligned}
\text{scalars:} & \qquad &
M &\in \mathcal{G}/\mathcal{H} , \\
\text{Abelian gauge fields:} & \qquad &
A_\mu &\in \mathcal{E} , \\
\text{gravitinos:} & \qquad &
\Psi_\mu & \in \mathcal{S} , \\
\text{spin-$\tfrac12$ fields:} & \qquad &
\chi & \in \mathcal{J} ,
\end{aligned}
\end{equation}
where $\mathcal{E}$ is a $\mathcal{G}$-representation and $\mathcal{J}$ and $\mathcal{S}$ are $\mathcal{H}$-representations. In general there may also be higher-rank $p$-form fields. The gauged theory is determined by the embedding tensor $X$. This is a map
\begin{equation}
X : \mathcal{E} \to \adj{\mathcal{G}} ,
\end{equation}
where $\adj{\mathcal{G}}$ is the adjoint representation of $\mathcal{G}$. Supersymmetry requires that the map satisfies
\begin{align}
\label{eq:X-Leibniz}
(1) \quad & \comm{X(U)}{X(V)} = X(X(U)\cdot V) , \\
\label{eq:X-rep}
(2) \quad & \text{a particular restriction on $\mathcal{G}$-reps
appearing in $X$} ,
\end{align}
where $X(U)\cdot V$ is the adjoint action of $X(U)$ on $V$. Note that
the first condition means that $X$ makes $\mathcal{E}$ into a Leibniz
algebra.\footnote{The relation~\eqref{eq:X-Leibniz} is more usually
written in explicit indices: viewing $\adj{\mathcal{G}}\subset
\mathcal{E}^*\otimes \mathcal{E}$, if elements of $\mathcal{E}$
have components $V^M$ then the embedding tensor can be written as
$X_{MN}{}^P$, with~\eqref{eq:X-Leibniz} taking the form
$\comm{X_M}{X_N}=-X_{MN}{}^PX_P$.}
The tensor $X$ completely determines the gauged theory. It generates a
potential for the scalars and mass terms for the fermions, and it
describes the gauging of the kinetic terms and also the corrections to
the supersymmetry transformations. In particular, it determines the
matrices $A\in\mathcal{S}\otimes\mathcal{S}^*$ and
$B\in\mathcal{J}\otimes\mathcal{S}^*$ appearing
in~\eqref{eq:susy-mod}. In fact, for maximal supergravity theories the
representation constraint on $X$ is that, decomposing under
$\mathcal{H}$, only the $A$ and $B$ representations appear. In other
words, $A$ and $B$ are uniquely determined by $X$ and (for the maximal
case) vice versa. The potential then follows from~\eqref{eq:sward}.
Let is now see how generalised geometry falls within this
framework. For concreteness we will consider $O(d,d)\times\mathbb{R}^+$
formulation of type II theories but the same ideas work
equally well for the full $\mathit{E}_{d(d)}\times\mathbb{R}^+$
formulations~\cite{CSW-11d} of $d$-dimensional truncations of type II
or eleven-dimensional supergravity. The idea is to imagine
``dimensionally reducing'' the theory on a $d$-dimensional manifold
$M$ without actually truncating any of the modes, following the
original idea of de~Wit and Nicolai~\cite{dWN} (see also~\cite{GLW} in the context of generalised geometry). Thus our set of ``moduli'' is an infinite-dimensional space formed of arbitrary choices of metric, $B$-field and dilaton tensor fields on $M$, since these all transform as scalars from the point of view of the lower-dimensional theory.\footnote{Note that one is also keeping an infinite set of spin-two, and spin-$\frac{3}{2}$ fields in the lower-dimensional theory, so this framework describes the gauging of a very unconventional lower-dimensional supergravity theory.}
We can formally describe this moduli space as a coset
$\mathcal{G}/\mathcal{H}$ where we define
\begin{equation}
\begin{aligned}
\mathcal{G} &= \text{group of diffeomorphisms and
local $O(d,d)\times\mathbb{R}^+$ gauge transf} , \\
\mathcal{H} &= \text{subgroup of local $O(d)\times O(d)$ gauge transf} .
\end{aligned}
\end{equation}
Mathematically we consider an (arbitrary) $O(d,d)\times\mathbb{R}^+$
principle bundle $\tilde{F}$ over $M$. The group $\mathcal{G}$ is then
the group of bundle isomorphisms, that is equivariant
diffeomorphisms, those which preserve the group action on the
fibre. Identifying the subgroup $\mathcal{H}$ is equivalent to
identifying an $O(d)\times O(d)$ sub-bundle $P\subset\tilde{F}$. The
group $\mathcal{H}$ is then the group of bundle automorphisms of
$P$. Each choice of $P\subset\tilde{E}$ is equivalent to giving the
generalised metric function $G(x)$ (more precisely this is a section of the
corresponding vector bundle), which is equivalent to specifying the
functions $g(x)$, $B(x)$ and $\phi(x)$. Thus $\mathcal{G}/\mathcal{H}$
is an infinite-dimensional space of sections.
To define $\mathcal{E}$, we start with $E$ the vector bundle over $M$
defined by the $2d$-dimensional fundamental representation of $O(d,d)$
with zero weight under the $\mathbb{R}^+$ factor. Then the space of sections
of $E$ forms a representation of $\mathcal{G}$, since any section is
mapped to another under diffeomorphisms and gauge
transformations. Thus we define $\mathcal{E}$
\begin{equation}
\mathcal{E} = \Gs{E}, \qquad \qquad \text{space of sections of
$E$}.
\end{equation}
Note that the $O(d,d)$ metric $\met{U}{V}$ can be thought of as
defining an equivariant map between representations of
$\mathcal{G}$. Specifically it defines a map
$\mathcal{E}\otimes\mathcal{E}\to C^\infty(M)$ where $C^\infty(M)$ is
the space of smooth functions on $M$.
It is important to note that thus far there is no requirement that
$E$ or $\tilde{F}$ have anything to do with the conventional
generalised tangent space $TM\oplus T^*M$, for the moment they are
completely general $O(d,d)\times\mathbb{R}^+$ bundles. However, let us now
turn to the embedding tensor. Remarkably, this is none other than the
generalised Lie derivative
\begin{equation}
X(V) = \mathbb{L}_V ,
\end{equation}
with the understanding that by definition $X$ relates $E$ to the
generalised tangent space $E\simeq TM\oplus T^*M$ and $\tilde{F}$ to
the corresponding generalised frame bundle. From the explicit
expression~\eqref{eq:genLie} we see that $\mathbb{L}_V$ is indeed the combination of an infinitesimal diffeomorphism and an infinitesimal
$O(d,d)\times\mathbb{R}^+$ gauge transformation as required. From the
Leibniz property~\eqref{eq:relLieCourant} we see that the
condition~\eqref{eq:X-Leibniz} is identically satisfied.
The representation constraint connects $X$ to modifications to the supersymmetry variations~\eqref{eq:susy-mod}. For $O(d,d)\times\mathbb{R}^+$ generalised geometry the spin-$\frac12$ fermions $\chi$ in the ``dimensionally reduced'' theory come from the internal components of the gravitinos $\psi^\pm_m$, while the variation of the non-compact gravitinos $\Psi_\mu$ actually gets related to the variation of the internal dilatinos $\rho^\pm$. Thus we can identify the corresponding infinite-dimensional spaces
\begin{equation}
\mathcal{S}^\pm = \Gs{S(C_\pm)} , \qquad \qquad
\mathcal{J}^\pm = \Gs{C_\mp\otimes S(C_\pm)} .
\end{equation}
If $|\vol_G|$ is the density defined by the generalised metric $G$, then we have the inner products on $\mathcal{S}^\pm$ and $\mathcal{J}^\pm$
\begin{equation}
\label{eq:SJnorm}
\Met{\epsilon}{\epsilon^{\prime\pm}}
= \int_M \norm{\vol_G} \,
\bar{\epsilon}^\pm\epsilon^{\prime\pm} ,
\qquad \qquad
\Met{\psi^\pm}{\psi^{\prime\pm}}
= \int_M \norm{\vol_G} \,
\bar{\psi}^\pm\cdot\psi^{\prime\pm} ,
\end{equation}
where $\bar{\psi}^+\cdot\psi^{\prime+}
=\bar{\psi}^{+\bar{a}}\psi^{\prime+}_{\bar{a}}$ and $\bar{\psi}^-\cdot\psi^{\prime-}
=\bar{\psi}^{-a}\psi^{\prime-}_{a}$. From the generalised geometry construction we can then read off the operators $A$ and $B$ that appear in~\eqref{eq:susy-mod}. We find
\begin{equation}
\begin{aligned}
A^+ &= \gamma^aD_a , & & \qquad \qquad &
B_{\bar{a}}^+ &= D_{\bar{a}} , & \\
A^- &= \gamma^{\bar{a}}D_{\bar{a}} , & &&
B_a^- &= D_a .
\end{aligned}
\end{equation}
Thus these four unique operators can be viewed as the decomposition of $X=\mathbb{L}$ into $\mathcal{H}$ representations. To see this one replaces $\mathbb{L}_V=\mathbb{L}^D_V$, which is possible because $D$ is torsion-free, and then, because $D$ is compatible with the generalised metric, one can decompose into $\mathcal{H}$ representations. Put another way, given the data of $\mathbb{L}$ (to define a notion of torsion) and $G$, we know that the operators $A^\pm$ and $B^\pm$ are unique. Finally, in the type II theory we have the two Bismut relations~\cite{CSW1}
\begin{equation}
\label{eq:newBismut4}
\begin{aligned}
D^aD_a\epsilon^-
- \gamma^{\bar{a}}D_{\bar{a}}
\gamma^{\bar{b}}D_{\bar{b}}\epsilon^-
= \tfrac14 S^-\epsilon^-, \\
D^{\bar{a}}D_{\bar{a}}\epsilon^+
- \gamma^{a}D_{a}\gamma^{b}D_{b}\epsilon^+
= \tfrac14 S^+\epsilon^+,
\end{aligned}
\end{equation}
where $S^+=S^-$ is the Lagrangian for the NSNS fields. We see that these are precisely the supersymmetric Ward identities~\eqref{eq:sward}, with the appropriate conjugate operators $(A^{\pm})^\dag$ and $(B^{\pm})^\dag$ defined using the norms~\eqref{eq:SJnorm}.
Formally, one can consider the full ten-dimensional theory as a reduction to zero dimensions on a ten-dimensional manifold $M$. The ungauged theory is completely trivial since there are no kinetic terms in zero dimensions. The whole theory appears via the gauging, so that, for example, the ``potential on the internal space'' actually gives the full NSNS Lagrangian. But this immediately raises the point that this gives a general formalism for describing general supersymmetric theories, and in particular the $\alpha'$ corrections. In fact, the description of the heterotic theory given in section~\ref{sec:het-corr} is exactly of this type. The corrections appeared as extension of the generalised tangent space (and how the coset structure appeared). One might, of course, also consider modifications of the embedding tensor -- that is, in the this context the generalised Lie derivative. This might be the way one can relate to the construction from the previous section.
\section{Comment on NS5-branes}
\label{sec:NS5}
We would like to conclude with some brief comments about how NS5-brane sources might be treated within the generalised geometry framework.
The modification of the Bianchi identity due to the fivebranes is another instance where $H$ is no longer a curvature of a connection on a gerbe.\footnote{Indeed, as explained in \cite{fhmm} the global treatment of the branes requires nontrivial transition functions that are a combination of antisymmetric tensor gauge transformations and diffeomorphisms. } In the presence of fivebranes the Bianchi identity \eqref{eq:hetBianchi*} changes to
\begin{equation}
\label{eq:NS5}
\mathrm{d} H = \tfrac{1}{4} \alpha' \left[\tr F\wedge F -
\tr R(\Omega^-)\wedge R(\Omega^-) \right]+ 2\pi \delta ,
\end{equation}
where $\delta$ is a four-form which integrates to one in the directions transverse to the fivebrane and has delta function support on the fivebrane. This means in particular that in compact geometries one does not need to balance the gravitational contribution to~\eqref{eq:NS5} by instantons only but may use NS5-branes. Topologically we can think of $\delta$ as representing the cohomology class $[W]$ that is Poincar\'e dual to the collection $W=W_1+\dots + W_n$ of the individual cycles $W_i$ on which each fivebrane is wrapped. In cohomology the Bianchi identity implies
\begin{equation}
\label{eq:NS5int}
\tfrac{1}{2}p_1(M) = c_2 + [W] ,
\end{equation}
where $c_2$ is the second Chern class of the gauge bundle.
A central point of the discussion in this paper is that the standard heterotic Bianchi identity is precisely captured by the generalised geometry on the extended tangent space $E$. In particular, it is encoded in the transitive Courant algebroid structure defined by the generalised Lie derivative. It is then very interesting to understand if the same structures can be used to encode the fivebrane terms as well. One might simply add an additional auxiliary gauge bundle with second Chern class equal to $[W]$. This would at least capture the correct topology, but one then requires local data that there are no corresponding gauge fields, and a singular delta function source.
There are two different simple situations were aspects of the source contribution to \eqref{eq:NS5} are particularly interesting. The first is when the fivebrane embedding (i.e. the normal bundle) is nontrivial. The local $n$-dimensional geometry near the brane can be viewed as a transverse space given by a ball with a sphere boundary $S^{n-7}$ fibered over the six-dimensional world volume $W$ of the fivebrane. The second case is when the manifold is a product $M\times K$ such that $K$ is a compact space transverse to the fivebranes. The source contribution is cohomologically equivalent to the (non-trivial) volume element on $K$. In both cases, one can consider decomposing the $SO(n)$ symmetry to $SO(6) \times SO(n-6)$. This breaking can introduce some novelties into our previous construction. We shall not try to address the general case but consider the simplest interesting example where $K$ is a $K3$ compactification of heterotic string. The fivebranes are points on $K$ and have a trivial normal bundle. It is well-known that, while in absence of the branes, the cohomology condition~\eqref{eq:NS5int} for the $K3$ background requires the instanton number for the internal part of the gauge field to be 24, in the presence of $Q_5$ fivebranes the required instanton number becomes $24-Q_5$.
We can then make a minimal identification of the relevant gauge bundles and obstructions involved. Over $K$ we have a non-trivial gauge theory with group $H\subset E_8 \times E_8$ or $SO(32)$ and corresponding bundle $P_H$. We also have the frame bundle which reduces to an $\SU(2)$-bundle $P_{\SU(2)}$ since $K3$ has $\SU(2)$ holonomy. Over $M$, there is a bundle $P_G$ for the ``unbroken'' part of the gauge group (that is the commutant of $H$ in $E_8$ or $SO(32)$) and the frame bundle $F_M$. Over $M$ the extension gives a non-trivial Bianchi identity in six dimensions, dependent on $P_G$ and $F_M$, the construction of which has already been explained. Over $K$, in the absence of fivebranes, one finds a topologically-trivial internal part to the Bianchi identity, which depends on the $P_H$ and $P_{\SU(2)}$ bundles, and whose integrated version is the $c_2(P_H)=24$ condition. One could have come up with a more elaborate version of the construction, but in the absence of NS5-branes this minimal version captures the essential ingredients.
If we include fivebranes we have a new ingredient, a four-form $V(K) \simeq \Lambda^4 T^*K$, cohomologically equivalent to a volume form on $K3$ and denoting the fivebrane charge. The construction over $M$ is unchanged, but we now have an obstruction on $K3$, given by cohomology condition~\eqref{eq:NS5int} which integrates to $c_2(P_G) + Q_5 = 24$ and the generalised tangent space and generalised Lie derivative on $K$ is determined by $P_H$, $P_{\SU(2)}$ and $V(K)$. One possibility is to represent $V(K)$ by another bundle $P_5$ with $c_2(P_5)=[V(K)]$, but somehow restrict this to be non-dynamic and have a localised field strength. Alternatively, one could relax the third condition~\eqref{eq:CA-def} so that the Courant algebroid is no longer transitive, and the original Bianchi identity need not hold. However, this relaxation must be in a controlled way that depends on the class $[V(K)]$. One might imagine for example further extending the generalised tangent space by $\Lambda^4T^*K$, however $V(K)$ plays the role of charge rather than a gauge parameter and so this is not the most natural approach to take. Geometrically the quantised fivebrane is captured by a sort of singular 2-gerbe (the charge is a four-form rather than a three-form) perhaps best described using sheaves.
The discussion here is rather sketchy and speculative, and it would be interesting to have a more complete description of NS5-branes in generalised geometry, notably away from the small-instanton limit. It also makes no attempt to capture aspects of the theory on the fivebrane itself. It may also be of some interest to work out a more complete description of six-dimensional $(0,1)$ theories and the higher-derivative $\alpha'$ couplings in this formalism.
\section*{Acknowledgments}
We would like to thank Mariana Gra\~na for collaboration at the initial stage of the project, and Marco Gualtieri for useful discussions. This work was supported in part by the ERC Starting Grant 259133 -- ObservableString, the German Research Foundation DFG within the Cluster of Excellence ``QUEST'' (AC), the Agence Nationale de la Recherche under the grant 12-BS05-003-01 (RM), the EPSRC Programme Grant ``New Geometric Structures from String Theory'' EP/K034456/1 (DW) and the STFC Consolidated Grant ST/J0003533/1 (DW). AC also thanks CEA Saclay for hospitality during the completion of this work.
|
\section{Introduction}\label{S-Int}
In the realm of finite dimensional algebras, pure injective modules may be defined as direct summands of direct
products of finite dimensional modules. Even when the structure of finite dimensional modules is known (say, the
algebra is tame) the effect of this construction is understood in just a few cases.
For instance (see \cite{Pun04}) superdecomposable pure injective modules can occur over some tame (string,
non-domestic) algebras. If we are interested in indecomposable pure injective modules only, then a complete
classification is known for tame hereditary finite dimensional algebras (see \cite{Pre98} or \cite{Rin98}) but in
just a few other examples (see \cite{B-P}).
In 1995 Ringel \cite{Rin95} constructed examples of pure injective modules over string algebras corresponding to
some infinite strings and (in \cite{Rin00}) conjectured that these modules, the finite dimensional modules, together
with the infinite dimensional band modules (i.e. Pr\"ufer, adic and generic) is a complete list of the
indecomposable pure injective modules over domestic string algebras. Over the past 15 years there have been
persistent attempts, starting from \cite{B-P}, to settle this conjecture, most recently in the (as yet largely
unpublished) thesis of Richard Harland \cite{Har} and by the first author \cite{Pun14}, who completed the case of
1-domestic string algebras.
In this paper we will complete the proof of Ringel's conjecture by reducing it to the (already known) 1-domestic
case. It was the first author who saw how to use a portion of Harland's thesis to effect this reduction, and we will
use this opportunity to include some of Harland's arguments, though his proof of one particularly important theorem
is not included because of its highly combinatorial nature and strong dependence on the approach and details of the
whole of his text. We will, however, give a direct proof of that result in the particular case of domestic string
algebras, leaving that to the end of the paper so as not to break the flow of the argument. We will provide each
statement borrowed from \cite{Har} with a precise reference.
The main drawback of the current situation is that the interested reader
should go through a couple of hundred pages of journal papers, preprints and a thesis, in order to recover the whole
logic of the proof. Of course there is need for a unified text with more accessible and conceptual proofs, but for
now we are satisfied with obtaining the result. It may be that the combinatorial nature of string algebras prevents
there being a proof which is significantly easier and shorter.
We will also prove the non-existence of superdecomposable pure injective
modules over domestic string algebras as a consequence of our main result: every pure injective module over such an
algebra contains an indecomposable direct summand. We will postpone a further consequence, the finiteness of
Krull--Gabriel dimension of domestic string algebras to a forthcoming paper.
\section{Basics}\label{S-bas}
We will be very brief on basic definitions, relying mostly on illustrative diagrams. One can find rigorous
definitions in \cite{B-R}, with almost no diagrams, or in Schr\"oer's thesis \cite{Sch97}, with a lot of them.
In this paper $A$ will denote a finite dimensional string algebra of infinite representation type over an
algebraically closed field $K$. The assumption that $K$ be algebraically closed is for convenience and simplicity of
arguments when treating band modules; it is almost certainly not essential for any result. A \emph{string} algebra is
a special kind of a bound quiver algebra $KQ/I$ with monomial relations, in particular there are at most two ingoing
and two outgoing arrows at each vertex of $Q$ and also, for each ingoing arrow, there is at most one nonzero
composition with an outgoing arrow, and {\it vice versa}. For instance, the Kronecker algebra
$$
\wt A_1 \hspace{1cm}
\vcenter
\xymatrix@R=8pt@C=20pt
*+={\circ}\ar@/^.5pc/[r]^{\al}\ar@/_.5pc/[r]_{\be}&*+={\circ}\\
}
}
$$\label{A1}
\vspace{2mm}
\noindent is a string algebra which is hereditary. A non-hereditary example is the Gelfand--Ponomarev algebra
$G_{2,3}$, which is the path algebra of the following quiver
$$
G_{2,3} \hspace{1cm}
\vcenter{
\xymatrix@R=16pt@C=16pt
*+={\circ}\ar@(ul,dl)_{\alpha}\ar@(ur,dr)^{\beta} } }
$$
\vspace{2mm}
\noindent with relations $\al^2= \be^3= \al\be= \be\al=0$.
For another example one could take the algebra $R_1$ with the same quiver but with noncommutative relations:
$\al^2= \be^2= \al\be= 0$. Finally, let $\Lam_2$ be the following string algebra:
$$
\Lam_2 \hspace{1cm}
\vcenter{
\xymatrix@R=20pt@C=20pt
*+={\circ}&*+={\circ}\ar@/_0.5pc/[l]_{\del}="del"\ar@/^0.5pc/[l]^{\eps}="eps"&*+={\circ}\ar[l]^{\ga}="ga"&
*+={\circ}\ar@/_0.5pc/[l]_{\al}="al"\ar@/^0.5pc/[l]^{\be}="be"\\
\ar@/_0.2pc/@{-}"del"+<6pt,-4pt>;"ga"+<-4pt,12pt>
\ar@/^0.2pc/@{-}"ga"+<8pt,4pt>;"be"+<-6pt,4pt>
}
}
$$
\vspace{2mm}
\noindent whose relations $\del\ga= \ga\be= 0$ are shown by solid curves. Note that we apply arrows from right to left
- in the path $\del\ga$ one goes first by $\ga$ and then by $\del$ - and our modules will be left modules over the
path algebra.
A \emph{letter} is an arrow (a \emph{direct} letter) or the formal inverse of an arrow (an \emph{inverse} letter). A (finite) \emph{string} over $A$ is a walk through the quiver of $A$ (hence a finite word in direct and inverse
letters) such that there can be no cancellation (of a letter and its inverse) nor can any relation (or its inverse)
be met on the way. For instance $\al\be^{-1}$ is a string over each of the above algebras, which we will illustrate by
the following diagram:
$$
\vcenter{
\xymatrix@R=12pt@C=8pt
&*+={\circ}\ar[dl]_{\al}\ar[dr]^{\be}&\\
*+={\circ}&&*+={\circ}
}}
$$
\vspace{2mm}
\noindent (note that we draw direct arrows from the upper right to the lower left, and inverse arrows from the upper
left to the lower right). By the definition of string algebra, there is at most one way to extend a string of length
$\geq 1$ to the right by a direct arrow, to the right by an inverse arrow, and similarly on the left.
In fact the above diagram represents not a string, but rather the corresponding string module (the circles correspond
to basis elements), which is finite dimensional and indecomposable.
If $u$ is a string then the corresponding string module, denoted $M(u)$, is indecomposable. According to \cite{B-R} all the
other indecomposable finite dimensional $A$-modules are \emph{band modules}, where a band is a walk which returns to
its starting point, contains both a direct and an inverse arrow, can be repeated twice and is not a proper power of another
walk. A typical example, see Figure \ref{figband}, is the (two layer) band module $M= M(C,\lam, 2)$ corresponding to
the band $C= \al\be^{-1}$ over $\widetilde{A_1}$ and $0\neq \lam\in K$. For instance $\be(z_2^2)= \lam z_1^2+ z_1^1$.
\begin{figure}[b]
$$
\vcenter{
\xymatrix@R=24pt@C=30pt
*+={\circ}\ar@{}+<-10pt,10pt>*{z_1^2}\ar@{.}[d]&
*+={\circ}\ar@{.}[d]\ar@/^0.9pc/[l]_{\alpha}\ar@{.}[d]\ar@{-->}@/_0.3pc/[dl]^(.3){\be}\ar@/_0.9pc/[l]_{\beta=\lambda}
\ar@{}+<10pt,10pt>*{z_2^2}\\
*+={\circ}\ar@{}+<-10pt,-10pt>*{z_1^1}&
*+={\circ}\ar@/^0.9pc/[l]^{\alpha}\ar@/_0.9pc/[l]^{\beta=\lambda}\ar@{}+<10pt,-10pt>*{z_2^1}
}
}
$$
\caption{}\label{figband}
\end{figure}
\vspace{2mm}
It is technically convenient to insist that a band should start at a point in the socle of the corresponding module:
so a \emph{band} will mean any string $C$ of the form $\al u \be^{-1}$, where $\al$ and $\be$ are different arrows
with target the same vertex of $Q$, such that $C$ is \emph{primitive}, i.e. $C\neq v^k$ for $k\geq 2$ and any string
$v$. Of course if $C$ is a band then so also is $C^{-1}= \be u^{-1}\al^{-1}$, as well as some cyclic permutations of
$C$. We assume that $A$ is of infinite representation type, so there are some
bands over $A$.
For example the algebras $\wt A_1$ and $R_1$ have essentially one band $\al\be^{-1}$, but $\Lam_2$ has two bands
$\al\be^{-1}$ and $\eps\del^{-1}$. There are infinitely many bands over $G_{2,3}$, for which
$\al\be^{-2}\al\be^{-2}\al\be^{-1}$ is one example.
One can extend the notion of a finite string to (1-sided or 2-sided) infinite string. For instance the 1-sided string
$(\al\be^{-1})^{\fty}$ is periodic, whereas the string $\be(\al\be^{-1})^{\fty}$ over $R_1$ is \emph{almost periodic},
but not periodic. Furthermore, the 2-sided string $^{\fty} (\be\al^{-1})\be (\al\be^{-1})^{\fty}$
(see Figure \ref{fig1}) over $R_1$ is \emph{biperiodic} (i.e. almost periodic on the right and on the left but not
periodic).
\begin{figure}[b]
$$
\vcenter{
\xymatrix@R=14pt@C=10pt
&&&&&&&*+={\circ}\ar[dl]_{\al}\ar[dr]^{\be}&&*+={\circ}\ar[dl]^{\al}\ar[dr]^{\be}&&\\
&&*+={\circ}\ar[dl]_{\be}\ar[dr]^{\al}&&*+={\circ}\ar[dl]^{\be}\ar[dr]^{\al}&&*+={\circ}\ar[dl]^{\be}&&*+={\circ}&&
*+={\circ}&\dots\\
\dots&*+={\circ}&&*+={\circ}&&*+={\bullet}&&&&&&
}
}
$$
\caption{}\label{fig1}
\end{figure}
\vspace{2mm}
A string algebra $A$ is said to be \emph{domestic} if, for every arrow $\al$, there exists at most one band over $A$
starting with $\al$, equivalently, \cite[Prop.~2]{Rin95}, there are just finitely many bands. (In fact this is a
specialization for string algebras (see \cite{Rin95}) of the general notion of domesticity for tame finite
dimensional algebras.) For example, $\al\be^{-1}$ is the unique (up to a cyclic permutation and inversion) band for
the algebras $\widetilde{A_1}$ and $R_1$, therefore we say that these algebras are 1-domestic. For $\Lam_2$ we have essentially
two bands, therefore this algebra is 2-domestic. Finally $G_{2,3}$ has infinitely many bands starting with $\al$,
hence this string algebra is not domestic.
By \cite[Prop.~2]{Rin95} a string algebra $A$ is domestic if and only if every 1-sided string over $A$ is almost
periodic. It follows that every 2-sided string over $A$ is either biperiodic or periodic.
Suppose that $u$ is an infinite string which is either 1-sided almost periodic on the right, or 2-sided almost
periodic on the right, but not totally periodic. It can be uniquely written in the form $v l D^{\fty}$, where $D$ is
a primitive cycle and $l$ is a (direct or inverse) letter such that the string $lD^{\fty}$ is no longer periodic; we allow $l$ to be empty, in which case $u$ is 1-sided periodic. Following Ringel \cite{Rin95} we say that this string is \emph{expanding} if the last (meaning right-most) letter of $D$ is inverse. From the definition of string algebra it follows that $l$ is direct (or empty) and there is a repeatable ``shift" to the right of the word away from the middle - meaning the non-periodic part - which corresponds to an endomorphism of the
associated infinite-dimensional module. For instance the string shown on Figure \ref{fig1} is expanding with
$l= \be$ and $D= \al\be^{-1}$.
We say that $u$ is \emph{contracting} if the last letter of $D$ is direct, hence $l$ is an inverse letter (or empty) and there
is a corresponding shift/endomorphism towards the middle of the word. Corresponding `left' notions are defined by
considering the string $u^{-1}$. For instance the string in Figure \ref{fig1} is contracting on the left: one should
first flip it over and check this condition on the right.
To each 1-sided almost periodic (including periodic) string, and to each 2-sided biperiodic (that is, almost periodic
in each direction but not (totally) periodic) string Ringel \cite{Rin95} assigned an indecomposable pure injective
module which is a direct sum, direct product, or mixed module, depending on the shape (expanding or contracting) of
its ends. For instance, for the string $u$ shown in Figure \ref{fig1} the corresponding mixed module $M^+(u)$ is a
direct product module on the expanding end (on the right), and a direct sum module on the contracting end (on the
left). Precisely, we take a basis element at each node of the string, form the product of the 1-dimensional
$K$-vector spaces they generate, then take the subspace of sequences which are eventually $0$ to the left; the
action of the algebra is as given by the labelling of the arrows of the string.
For each band $C$ over (any) string algebra $A$ and for a fixed $0\neq \lam\in K$ the finite dimensional band modules
$R_k= M(C,\lam, k)$ form a ray of irreducible monomorphisms, $R_1\to R_2\to \dots$, in the Auslander--Reiten quiver.
The direct limit along this ray is an indecomposable pure injective \emph{Pr\"ufer} module. Similarly, there is a
coray of irreducible epimorphisms $R_1\xl{} R_2\xl{} \dots$ whose inverse limit is the \emph{adic} module and which
is also pure injective and indecomposable. We will refer to Pr\"ufer and adic modules, as well as the generic modules
associated to bands (one to each band, see, e.g., \cite[\S 8.1.2]{Preb2}), as \emph{infinite dimensional band modules}.
The following conjecture is due to Ringel \cite{Rin00}.
\begin{conj}\label{r-conj}
Let $M$ be an infinite dimensional indecomposable pure injective module over a domestic string algebra $A$. Then $M$
is either an infinite dimensional band module or a 1-sided or 2-sided direct sum, direct product, or mixed module
corresponding to a 1-sided almost periodic or 2-sided biperiodic string.
\end{conj}
We will denote by $M_w$ the module, described above, corresponding to the string $w$ in this conjecture, and say that
$M_w$ is on Ringel's list. For 1-domestic string algebras this conjecture was verified in \cite{Pun14}.
\section{Pure injective modules}\label{S-pi}
A module $M$ over a finite dimensional algebra $A$ is said to be \emph{pure injective} (or algebraically compact) if
it is a direct summand of a direct product of finite dimensional modules. In particular, every finite dimensional
module is pure injective. From now on $A$ will denote a (usually domestic) string algebra with a fixed presentation
by a quiver with relations.
Recall (see \cite{B-R}) that for each vertex $S$ of $Q$ one can partition the strings entering $S$, including extra
strings $1_{S,\pm 1}$, into two sets $H_{\pm 1}= H_{\pm 1}(S)$. Namely we require that $1_{S,i}\in H_i$ and each
$H_i$ contains at most one direct (and inverse) arrow ending in $S$. Furthermore for each direct arrow $\al\in H_i$
and inverse arrow $\be^{-1}\in H_i$, we insist that $\be\al$ is a relation in $A$. If $C$ is a string of length
$\geq 1$ then we put it in $H_i(S)$ if the first (that is, furthest to the left) letter of $C$ is there. Observe that this partition is often
non-unique.
For instance over $R_1$ we can choose $\be, \be^{-1}\in H_1$ and $\al, \al^{-1}\in H_{-1}$ and then take a string $u$
in $H_i(S)$ if the first letter of $u$ is there. Thus for this example $H_1(S)$ consists of $1_{S,1}$ and strings that
start with either $\be$ or $\be^{-1}$.
The strings in each $H_i(S)$ are ordered in a natural way (with $\del^{-1} < 1_{S,i} < \ga$) and each of these sets
forms a chain with respect to this ordering. For instance, with the above choice for $R_1$, we get that
$\be\al^{-1}< \be $ in $H_1$ and $\al^{-1}< 1_{S,-1}< \al$ in $H_{-1}$. When fixing a node in a diagram such as
Figure \ref{fig1} and considering related diagrams, these two sets will distinguish between the possible strings
extending to the right and to the left from that node.
One can extend the partition $H_{\pm 1}$ to include 1-sided infinite strings: we take a 1-sided string $u$ in
$\wh H_i= \wh H_i(S)$ if the first letter of $u$ is there. The ordering on $H_i(S)$ extends naturally to a linear
ordering on the set $\wh H_i(S)$ of 1-sided (finite or infinite) strings. For instance, over $R_1$, the infinite
string $(\al\be^{-1})^{\fty}$ belongs to $\wh H_{-1}$ and is larger than each finite string $(\al\be^{-1})^n$. In
general (say, over $G_{2,3}$) the structure of this ordering is quite complicated. Over domestic algebras it is not
so complicated and the following fact follows from \cite[4.10 (and 3.2)]{Sch97}.
\begin{fact}\label{non-dense}
Let $A$ be a domestic string algebra. Then each chain $\wh H_i$ contains no subchain isomorphic to the ordering
of the rationals.
\end{fact}
Suppose that $M$ is an $A$-module, $N$ is a subspace of $M$ and $\al$ is an arrow. Then $\al N$ will denote the image
of $N$ with respect to $\al$, and $\al^{-1} N$ is the preimage of $N$. If $D\in H_1(S)$ is a string of length $k$
then define $DM$ by induction on $k$, by setting $1_{S,1} M= M$ and $DM= l (D' M)$, if $D= l D'$ for a letter $l$.
Now for each $D\in H_1(S)$ we define the pp formula $(.D)(x)$. Namely if there is an arrow $\ga$ such that $D\ga^{-1}$
is a string then this formula states that $x= e_S\, x$ and $x\in D\ga^{-1}(0)$. If no such arrow exists then we
require just $x= e_S x$ and $x\in D M$. Informally we will consider this formula as asserting divisibility (in fact a bit more)
by $D$ on the right.
For instance if we choose $\be\in H_1(S)$ over $\wt A_1$ and $D= \be\al^{-1}$ then $m\in (.D)(M)$ iff $m\in \be M$
(in this case there is no $\ga$ as above). But over $R_1$ for the same formula we obtain $m\in (.D)(M)$ iff there
exists $n\in M$ such that $m= \be n$ and $\be\al n= 0$ (in this case $\ga$ in the definition above is $\be$).
$$
\vcenter
\xymatrix@R=14pt@C=10pt
&*+={\circ}\ar[ld]_{\be}\ar[rd]^{\al}\ar@{}+<0pt,10pt>*{_n}&&\\
*+={\circ}\ar@{}+<-10pt,-6pt>*{_m}&&*+={\circ}\ar[rd]^{\be}&\\
&&&*+={\circ}\ar@{}+<8pt,-2pt>*{_0}
}}
$$
\vspace{2mm}
Observe that, if $D, D'\in H_1(S)$, then $D\leq D'$ if and only if $(.D')$ implies $(.D)$ (hence the ordering on
pp formulas inverts the ordering on strings).
The left-handed formulas $(C^{-1}.)$ for $C\in H_{-1}(S)$ are defined similarly. Finally let $(C^{-1}.D)$ denote the
conjunction of $(C^{-1}.)$ and $(.D)$.
Suppose that $M$ is an $A$-module and $m$ is a nonzero element in $e_S M$ for some primitive idempotent $e_S$. There
is standard way (see \cite{P-P04} or \cite[Sect. 5.8]{Har}) to assign to $m$ a (finite, 1-sided, or 2-sided) string
$w= u^{-1}.v$, where $u\in \wh H_{-1}(S)$ and $v\in\wh H_1(S)$.
Namely define $v(m)$, the \emph{right handed string}, as the supremum of strings $D\in H_1(S)$ such that the formula
$(.D)(x)$ holds for $m$ in $M$. For instance, if $M= M(u)$ is a finite dimensional string module, $m$ is its
leftmost basis element and $u\in H_1(S)$ then $v(m)= u$. Furthermore if $M= M^+$ is as in Figure \ref{fig1} and
$m$ is the marked element from the socle of $M$, then it is easily checked that $v(m)= \be (\al\be^{-1})^{\fty}$.
Note that if $v= v(m)$ is infinite then one can choose an ascending sequence of substrings $D_1< D_2< \dots$ of $v$
such that $v= D_k v_k$ where $D_k$ ends with an inverse arrow, $v_k$ starts with a direct arrow and $v$ is the
supremum of the $D_k$. Furthermore if $M$ is pure injective, we can always divide $m$ in $M$ by $v(m)$, meaning
that there is a sequence $m_k$ of elements in the socle of $M$ which together
witness each of the conditions $m\in (.D_k)(M)$.
The \emph{left handed string of $m$}, $u(m)$, is defined similarly. For instance in our running example we obtain
$u= (\al\be^{-1})^{\fty}$, or rather $u^{-1}= {}^{\fty} (\be\al^{-1})$ - which is the way it is shown on the diagram.
Finally we will set $w(m)= u^{-1}.v$. For instance (see \cite[L. 111]{Har}) if $M= M_w$ is from Ringel's list and
$m$ is an element from a standard basis placed between the strings $u^{-1}$ and $v$ (so $w= u^{-1}v$), then
$u= u(m)$ and $v= v(m)$. Note also that to every element in a standard basis of a finite dimensional band module we
assign in this way a 2-sided periodic string $^{\fty}D. D^{\fty}$ for some primitive cycle $D$ (we have included
among the string, rather than band, modules those `band' modules, such as that with parameter $\lambda =0$ in
Figure \ref{figband}, which can be described in both ways).
Let us make the following trivial, but important, remark that will be used frequently. Suppose that $m= \al n$
for some nonzero $m\in e_S M$, $n\in e_T M$ and $\al^{-1}\in H_{-1}(T)$. Then $u(n)= \al^{-1} u(m)$ and
$v(m)\geq \al v(n)$. Namely if $u(n)> \al^{-1} u(m)$ then there is $\ga\in H_{-1}(T)$ such that $n$ is divisible
by $\ga$, hence $\al\ga= 0$ implies $m=0$, a contradiction.
$$
\vcenter
\xymatrix@C=10pt@R=10pt
*+={\circ}\ar[rd]^{\ga}&\\
&*+={\circ}\ar[ld]^{\al}\ar@{}+<10pt,0pt>*{_n}\\
*+={\circ}\ar@{}+<-10pt,-4pt>*{_m}&
}}
$$
\vspace{2mm}
Suppose that $M$ is a pure injective module pointed at a (nonzero) element $m\in e_S M$ and let $w(m)= u^{-1}.v$
(a \emph{pointed module} is a module with a specified element or tuple). Say that $m$ is \emph{homogeneous} if there is no
$x\in e_S M$ such that the left string of $x$ is greater than $u$ and the right string of $m-x$ is greater than $v$.
Note (by Lemma \ref{triang} below) that if an element $m$ could be written in this way then the right strings of $m$
and $x$ would be equal, as would be the left strings of $m-x$ and $m$.
$$
\vcenter{
\xymatrix@R=20pt@C=14pt
&*+={\circ}\ar@{--}[dl]\ar@{--}[dr]\ar@{}+<0pt,10pt>*{_m}&\\
*+={\circ}\ar@{}+<-1pt,-10pt>*{_x}\ar@{}+<-30pt,6pt>*{_{u(x)> u(m)}}&&
*+={\circ}\ar@{}+<0pt,-10pt>*{_{m-x}}\ar@{}+<38pt,6pt>*{_{v(m-x)> v(m)}}
}
}
$$
\vspace{2mm}
For instance it follows from \cite[L. 156]{Har} that every element in the canonical basis of a module $M= M_w$ from
Ringel's list is homogeneous. Note that Harland uses `fundamental' instead of `homogeneous' (and the terminology in
\cite{BarPre} was `maximal').
We will discuss this notion in detail in the next section. The following result is crucial for our considerations.
Including the proof would make the paper considerably longer so we have given a different but direct proof of this
for the case of domestic string algebras. We have placed that at the end of the paper so as not to break the flow
towards our main result.
\begin{fact}(see \cite[Thm. 51, p. 268]{Har} and its proof)\label{har}
Suppose that $A$ is an arbitrary string algebra and let $w= u^{-1}.v$ be either a 1-sided string or a 2-sided
\textbf{non-periodic} string. Then there exists a unique indecomposable pure injective module $N_w$ containing a
homogeneous element $m$ whose string $w(m)$ equals $u^{-1}.v$.
Furthermore if $M$ is an arbitrary pure injective module with a homogeneous element $m$ with $w(m) = u^{-1}.v$, then
$N_w$ is a direct summand of $M$.
\end{fact}
Observe that over a domestic string algebra every string $w$ is almost periodic (in one or both directions), hence
$N_w$ is isomorphic to $M_w$.
\section{Homogeneous elements}\label{S-hom}
In this section we will develop the machinery of homogeneous elements, closely following Harland's thesis.
We start with a straightforward lemma, a `triangle inequality'.
\begin{lemma}\label{triang}(see \cite[L. 155]{Har})
Suppose that $m_1, m_2\in e_S M$ have right strings $v_1$ and $v_2$. Then the element $m= m_1+ m_2$ has
right string greater than or equal to $\min(v_1,v_2)$. Furthermore if $v_1\neq v_2$ then the right string
of $m$ equals $\min(v_1,v_2)$.
\end{lemma}
\begin{proof}
Suppose that $C$ is a finite string such that $C\leq v_1, v_2$, therefore $m_1$ and $m_2$ are divisible by $C$ on
the right. It follows easily by induction on the length of $C$ that $m= m_1+ m_2$ is divisible by $C$ on the right,
hence $C\leq v(m)$, and the first claim follows.
For the second claim, by symmetry we may assume that $v_1> v_2$. Suppose that $v=v(m) > \min(v_1, v_2)= v_2$ hence
$\min(v, v_1)> v_2$. Writing $m_2= m - m_1$, by the first statement of the lemma we derive a contradiction.
\end{proof}
The following proposition says that when multiplying or dividing a homogeneous element by an arrow, we can choose the
resulting element to be homogeneous.
\begin{lemma}\label{div}(see \cite[L. 158]{Har})
Suppose that $m$ is a homogeneous element in a pure injective module $M$ with $w(m)= u^{-1}.v$, where $v= lv'$ for a
direct or inverse arrow $l$. Choose an element $n\in e_T M$ such that $m= \al n$ and $v(n)= v'$ if $l=\al$ is direct,
and set $n= \be m$ if $l= \be^{-1}$ is inverse. Assume without loss of generality that $v'$ is in $\wh H_1(T)$ and hence $l^{-1}u\in \wh H_{-1}(T)$. Then $n$ is homogeneous, with $w(n)= u^{-1}l.v'$.
More generally, if $v(m)=Cv''$ and, working along $C$, we choose a sequence of elements as above, obtaining $m=Cn$,
then $n$ will be homogeneous and, with a suitable choice of $H_{\pm 1}$, $w(n)$ is equal to $u^{-1}C.v''$.
\end{lemma}
By symmetry all this holds also for left handed strings.
\begin{proof}
Suppose first that $l=\al$ is a direct arrow, hence $m= \al n$ and $v(n)= v'$.
As we have already mentioned $w(n)= u^{-1}\al. v'$. It remains to show that $n$ is homogeneous.
Suppose for a contradiction that $n= n_1+ n_2$ such that $u(n_1)> u(n)$ and $v(n_2)> v(n)$. Set $m_1= \al n_1$ and
$m_2= \al n_2$, so $m= m_1+ m_2$.
$$
\vcenter{
\xymatrix@R=24pt@C=18pt
&&&*+={\circ}\ar@{}+<10pt,4pt>*{_n}\ar@{--}[dll]\ar@{--}[d]\ar[dl]^{\al}\\
&*+={\circ}\ar@{}+<-10pt,4pt>*{_{n_1}}\ar[dl]_{\al}&
*+={\circ}\ar@{}+<-4pt,-10pt>*{_m}\ar@{--}[dll]\ar@{--}[d]&
*+={\circ}\ar@{}+<12pt,0pt>*{_{n_2}}\ar[dl]^{\al}\\
*+={\circ}\ar@{}+<0pt,-10pt>*{_{m_1}}&&*+={\circ}\ar@{}+<0pt,-10pt>*{_{m_2}}&
}}
$$
\vspace{2mm}
Because $v(n_2)> v(n)$ it follows that $v(m_2)\geq \al v(n_2)> \al v(n)= v(m)$, in particular $m_1\neq 0$. Consequently
$\al^{-1} u(m_1) = u(n_1) > u(n)=\al^{-1} u$, therefore $u(m_1)> u= u(m)$. But $m= m_1 + m_2$ and $m$ is
homogeneous, a contradiction.
Now consider the case when $l= \be^{-1}$ is inverse, so $\be m= n$; note that $n\neq 0$. Then, by definition, $v(n)= v'$ and
$u(n)\geq \be u$. Suppose, by way of contradiction, that $u(n)> \be u$. Therefore we can divide $n$ by $\be$, finding
an element $m'$ such that $\be m'= n$ and $u(m')> u$.
$$
\vcenter{
\xymatrix@R=10pt@C=10pt
&*+={\circ}\ar[rdd]^{\be}\ar@{}+<-5pt,8pt>*{_{m'}}&\\
*+={\circ}\ar[rrd]_{\be}\ar@{}+<-10pt,0pt>*{_m}&&\\
&&*+={\circ}\ar@{}+<8pt,-2pt>*{_n}
}}
$$
\vspace{2mm}
Then $\be(m-m')=0$ yields $v(m-m')> v(m)$. It follows that $m= m'+ (m-m')$ with $u(m')> u(m)$ and $v(m-m')> v(m)$, a
contradiction to homogeneity of $m$. So $u(n)=\be u$.
It remains to show that $n$ is homogeneous. Otherwise $n= n_1+ n_2$ such that $u(n_1)> u(n)$ and $v(n_2)> v(n)$. Since
$u(n_1)> u(n)= \be u$, $n_1$ is divisible by $\be$: there exists $m_1$ such that $\be m_1= n_1$ and $u(m_1)> u$.
$$
\vcenter{
\xymatrix@R=14pt@C=14pt
*+={\circ}\ar@{}+<0pt,10pt>*{_{m_1}}\ar[rd]_{\be}&&
*+={\circ}\ar@{}+<0pt,10pt>*{_n}\ar@{--}[ld]\ar@{--}[rd]&\\
&*+={\circ}\ar@{}+<0pt,-10pt>*{_{n_1}}&&*+={\circ}\ar@{}+<0pt,-10pt>*{_{n_2}}
}}
$$
\vspace{2mm}
Because $m$ is homogeneous, from the decomposition $m= m_1 + (m-m_1)$ it follows that $v(m-m_1)\leq v= \be^{-1}v'$,
therefore $v(\be (m-m_1))\leq v'$. But $\be(m-m_1)= n_2$ and $v(n_2)> v(n)= v'$, a contradiction.
The statement for general finite strings $C$ follows by induction.
\end{proof}
We say that a pure injective module $M$ is \emph{1-sided}, if it contains a nonzero element $m\in e_S M$ whose
string $w(m)$ is 1-sided (or finite), and $M$ is \emph{2-sided} otherwise. For instance it can be checked that the module $M_w$
corresponding to a string $w$ on Ringel's list is 1-sided if and only if $w$ is 1-sided. Furthermore each finite
dimensional band module is 2-sided.
One-sided indecomposable pure injective modules over (any!) string algebra $A$ were classified in \cite{P-P04}: they
one-to-one correspond to the 1-sided strings. In particular, if $A$ is a domestic string algebra then each 1-sided
module is on Ringel's list. Thus in classifying indecomposable pure injective modules it suffices to consider the
2-sided case. Furthermore, by Fact \ref{har}, if $M$ contains a homogeneous element whose string is non-periodic,
then we know the structure of $M$, in particular it is on Ringel's list.
From the next proposition it follows that, in the domestic case, every pure injective module contains a homogeneous
element. That will leave us with the periodic 2-sided case to deal with in the next section.
\begin{prop}\label{dense}(see \cite[L. 168]{Har})
Let $A$ be an arbitrary string algebra and let $M$ be a 2-sided pure injective $A$-module containing no homogeneous
elements. Take any $0\neq m\in e_S M$ and set $w(m)= u^{-1}.v$. Then for any string $u'> u$ there exists $x\in e_S M$
such that $u'> u(x)> u$ and $v(m-x)> v$.
\end{prop}
It follows from Lemma \ref{triang} that $v(x)= v$ and $u(m-x)= u$ in this case.
\begin{proof}
Let $X$ be the set of finite strings $D\in H_1$, $D> v$ such that $m$ can be written as a sum $m_1+ m_2$, where $m_1$
is divisible by $E$ for some finite string $E\in H_{-1}$ greater than $u$, and $m_2$ is divisible by $D$. Because $m$
is not homogeneous this set is not empty. Because $M$ is 2-sided it follows immediately that the supremum $v'$ of $X$
is an infinite (1-sided) string larger than $v$.
Because $v'$ is infinite it can be written as $D_i l_i t_i$ for some direct arrows $l_i$ and finite strings $D_i\in X$
of strictly increasing length, each ending with an inverse arrow; in particular $D_1< D_2< \dots < v'$ in
$\wh H_1$ and $D_i$ is a \emph{presubstring} of $v'$ in the terminology of \cite{Har} (described as ``closed under
predecessors" in \cite{CB}).
For each positive $k\in \N$ consider the set of strings $F> u$ such that $m$ can be written as a sum $m= m_1+ m_2$,
where $m_1$ is divisible by $F$ and $m_2$ is divisible by $D_k$. Because $D_k\in X$, this set is non-empty; let $u(k)$
denote the supremum of strings in this set. Then $u(k)> u$ and clearly $u(k)\geq u(k+1)$ for each $k$. We claim that
the infinum of the $u(k)$ equals $u$, from which the result follows.
Suppose for a contradiction that this is not the case, so choose a finite string $C$ with $u< C< \inf_k u(k)$. Consider
the (infinite in $k$) set of (pp) conditions in a variable $x$ saying that $m$ can be written as a sum $x+ (m-x)$
such that $x$ is divisible by $C$ and $m-x$ is divisible by $D_k$. By assumption and construction, any finite subset
of this set is satisfied by some value of $x$ in $M$.
Since $M$ is pure injective ($=$ algebraically compact) there is a simultaneous solution for this set of conditions.
Namely there exists $x\in M$ such that $x$ is divisible by $C$ and $m-x$ is divisible by $D_k$ for each $k$. It
follows that the right string of $m-x$ is at least $v'= \sup X$, and therefore equals $v'$.
Because $m-x$ is not homogeneous, there exists $y$ such that $u(y)> u(m-x)= u$ and $v(m-x-y)> v(m-x)= v'$.
$$
\vcenter{
\xymatrix@R=14pt@C=12pt
&*+={\circ}\ar@{--}[ld]\ar@{--}[rd]\ar@{}+<10pt,4pt>*{_m}&&\\
*+={\circ}\ar@{}+<-9pt,4pt>*{_x}&&*+={\circ}\ar@{--}[ld]\ar@{--}[rd]\ar@{}+<19pt,4pt>*{_{m-x}}&\\
&*+={\circ}\ar@{}+<-10pt,-4pt>*{_y}&&*+={\circ}\ar@{}+<22pt,-4pt>*{_{m-x-y}}
}}
$$
\vspace{2mm}
Now, $m= x+ y+ (m-x-y)$. However $x+y$ has left string greater than $u$ (because $x$ and $y$ do). Thus there is a
finite string $C'> u$ such that $x+y$ is divisible by $C'$. Furthermore the right string of $m-x-y$ is greater than
$v'= \sup X$, so there is a finite string $D> v'$ such that this element is divisible by $D$. It follows that
$D\in X$, a contradiction.
\end{proof}
The following is derived immediately.
\begin{cor}\label{hom-dom}
Let $M$ be any nonzero pure injective module over a domestic string algebra. Then $M$ contains a homogeneous element.
\end{cor}
\begin{proof}
This follows from Proposition \ref{dense} and Fact \ref{non-dense}.
\end{proof}
\section{Main result}\label{S-main}
Before proving the main result we have to introduce the \emph{bridge quiver} of a domestic string algebra $A$. It will
be more convenient for this paper to define it as a poset, rather than as a directed graph (see \cite{Sch97}) (though
this might lose information - for instance there can be more than one path between points in the bridge quiver).
We fix a set $\mB$ of representatives of bands over $A$ up to a cyclic permutation (but a band and its inverse are
represented by different elements in $\mB$). For instance, for $\wt A_1$ or $R_1$, this set consists of two elements
$C= \al\be^{-1}$ and $C^{-1}= \be\al^{-1}$.
We define $C\preccurlyeq D$ for bands $C, D\in \mB$ if there is a finite string of the form $CuD$ over $A$. It
follows from the combinatorics of domestic string algebras (for example, from Fact \ref{band} (ii), (iii) below) that
$C\preccurlyeq D\preccurlyeq C$ implies $C= D$, hence $\mB$ is a poset. For instance, for $A= \Lam_2$ we can choose
$C= \eps\del^{-1}$, $D= \al\be^{-1}$ together with $C^{-1}, D^{-1}$ as the elements of $\mB$. Then the bridge quiver
of $\Lam_2$ is a union of two chains, $C\prec D$ and $D^{-1} \prec C^{-1}$, the comparison $C\prec D$ being realized
by the string $C\eps\ga D$.
It follows from \cite{Sch97} that for domestic string algebras $\mB$ is a finite poset.
Now we are in a position to prove the main result of the paper.
\begin{theorem}\label{main}
Every indecomposable pure injective module over a domestic string algebra $A$ is on Ringel's list.
\end{theorem}
The proof of this result will need a few lemmas. Let $M$ be an indecomposable pure injective $A$-module. As we already
mentioned, if $M$ is 1-sided, then it is on Ringel's list. Thus we may assume that $M$ is 2-sided. Furthermore by
Corollary \ref{hom-dom} we know that $M$ contains a homogeneous element. If the string of this element is non-periodic
then, by Fact \ref{har}, $M$ is on Ringel's list.
Thus we may assume for the rest of the proof that $M$ is \textbf{2-sided, contains a homogeneous element, and any
homogeneous element in $M$ is periodic} (i.e.\ its string is 2-sided and periodic).
We recall the following facts about the combinatorics of bands over domestic string algebras.
\begin{fact}(see \cite{Rin95})\label{band}
Suppose that $E, F$ are bands over a domestic string algebra.
(i) If $\al^{-1}\be$ occurs in both $E$ and $F$ then these bands are equivalent up to cyclic permutation.
(ii) If both $E$ and $F$ begin with $\al$ then $E=F$.
(iii) If $E$ begins with $\al$, ends with $\be^{-1}$ and the same is true for a string $G$ then $G$ is a power of $E$.
\end{fact}
Let us say that a band $C= \ga\dots \del^{-1}$ is \emph{realized} in $M$ if there is a nonzero element $n\in \ga M\cap \del M$ (we do not assume $n$ to be homogeneous), in particular $n$ belongs to the socle of $M$.
Choose $D= \al\dots \be^{-1}$ to be any band which is minimal in the ordering on the bridge quiver $\mB$ among all
bands realized in $M$; we may assume that $\al\in H_1$ and $\be\in H_{-1}$.
\begin{lemma}\label{D}
Let $D$ be as above. Suppose that $ m$ is any nonzero element in $ \be M$. Then $m$ is homogeneous with
$w(m)= {}^{\fty} D.D^{\fty}$.
\end{lemma}
\begin{proof}
First we will check that $u(m)^{-1}= {}^{\fty} D$. Now, $u(m)^{-1}$ is an infinite string which ends with $\be^{-1}$;
since $A$ is domestic, this string is almost periodic: $u(m)^{-1}= {}^{\fty} C u' \be^{-1}$ for some band $C$ and
some finite string $u'$. If $C$ is a cyclic permutation of $D$ then, changing $u'$ if necessary, we may assume that
$C=D$, hence $u(m)^{-1}= {}^{\fty} D$ by Fact \ref{band} (iii).
Otherwise $C$ is not a cyclic permutation of $D$. Because $u(m)^{-1} D$ is, clearly, a string it follows that
$C\prec D$ in the bridge quiver. Since $C$ is obviously realized in $M$, we get a contradiction to minimality of $D$.
Second, we show that $m$ is homogeneous. If not, then $m= x+ (m-x)$ where $u(x)> u(m)$ in $\wh H_{-1}$ and
$v(m-x)> v(m)$ in $\wh H_1$, in particular $x\neq 0$. From $u(x)\geq u(m)$ it follows that $x\in \be M$, hence
$u(x)^{-1}= {}^{\fty} D$ by the first part of the proof. But then $u(x)= u(m)$, a contradiction.
Finally, $v(m)= D^{\fty}$ (in particular $m$ is divisible by $\al$) since we are assuming that homogeneous elements
have periodic words.
\end{proof}
For the remaining portion of the proof we will assume some familiarity with free realizations of pp formulas
(see \cite[\S 1.2.2]{Preb2}). Namely (in our context), we say that a pointed finite dimensional module $(N,n)$ is a
\emph{free realization} of a pp formula $\phi(x)$ if $n$ satisfies $\phi$ in $N$ and, for any element $l\in L$
satisfying $\phi$ in a module $L$, there exists a morphism $f: N\to L$ sending $n$ to $l$.
Every formula has a finite dimensional free realization (see \cite[1.2.14]{Preb2}). Recall that if $a\in A$ then the
\emph{divisibility formula} $a\mid x$ claims that there exists $y$ such that $x= ay$, hence it defines in each left
$A$-module $M$ the subspace $aM$. For instance this divisibility formula has $(A,a)$ as a free realization; furthermore
if $a\in Ae$ for some idempotent $e$, then the pointed module $(Ae,a)$ is another free realization of $a\mid x$.
For another example, the string module $M(\be\al^{-1}\be^{-1})$ pointed at the left end is a free realization of the
divisibility formula $\be\mid x$ over $R_1$ which therefore is equivalent to the formula $(.\be\al^{-1}\be^{-1})$.
$$
\vcenter{
\xymatrix@R=14pt@C=8pt
&*+={\circ}\ar[dl]_{\be}\ar[dr]^{\al}&&\\
*+={\bullet}&&*+={\circ}\ar[dr]^{\be}&\\
&&&*+={\circ}
}}
$$
\vspace{2mm}
Observe that over $\wt A_1$ the same formula $\be\mid x$ has $M(\be\al^{-1})$ as a free realization, hence is
equivalent to the formula $(.\be\al^{-1})$.
$$
\vcenter{
\xymatrix@R=14pt@C=8pt
&*+={\circ}\ar[dl]_{\be}\ar[dr]^{\al}&&\\
*+={\bullet}&&*+={\circ}&\\
}}
$$
\vspace{2mm}
The following lemma says that $D$ covers the support of $M$.
\begin{lemma}\label{ga}
If $\ga$ is an arrow not occurring in $D$ or $D^{-1}$ then $\ga M= 0$.
\end{lemma}
After having proved this lemma, we will be able to assume that every arrow occurs in $D$ or $D^{-1}$.
\begin{proof}
Otherwise choose $0\neq n\in \ga M$. Take $m$ to be a homogeneous element of $M$ whose string is\
${}^{\fty} D. D^{\fty}$. Since $M$ is pure injective and indecomposable, by \cite[4.3.72]{Preb2} there is a pp formula
$\phi(x,y)$ such that $M\ms \phi(m,n)\wg \neg \phi(m,0)$ (``$M\ms \phi(m,n)\wg \neg \phi(m,0)$" may be read as
``the pair $(m,n)$ satisfies the condition $\phi$ in $M$ but $(m,0)$ does not"; the fact that $M$ is indecomposable
pure injective means that there is a non-trivial link, expressed by a pp formula, between any nonzero elements).
We will show that this leads to a contradiction.
We may assume that $\phi(x,y)$ implies $(D.D)(x)$ and implies $\ga\mid y$ (both these are pp conditions, so may be
incorporated into $\phi$). We may further suppose, possibly changing $m$, that $\phi$ has an indecomposable free
realization. To see this, decompose a free realization, $N$, of $\phi$ as a direct sum of indecomposables,
$N=\bigoplus N_i$, and write $\phi$ as a corresponding sum of formulas $\phi_i$ with indecomposable free realizations;
then decompose $(m,n)= (\sum m_i, \sum n_i)$ accordingly, so $(m_i, n_i)$ realizes $\phi_i$.
Since $\phi$ implies $ (D.D)(x)$ so does each $\phi_i$, and similarly $\phi_i$ implies $\ga\mid y$. It follows from
Lemma \ref{D} that either $m_i= 0$ or $m_i$ is a homogeneous element with string $^{\fty} D. D^{\fty}$. Furthermore
$n_i\in \ga M$.
For each $j$ we have $M\ms \phi(m_j,n_j)$ and there is $i$ such that $M\ms \neg \phi_i(m_i,0)$ (in particular both
$m_i, n_i$ are nonzero). Otherwise, for each $i$ we would have $M\ms \phi_i(m_i, 0)$ and, adding up, would get
$M\ms \phi(m,0)$, a contradiction. Now replace $\phi$ with $\phi_i$, $m$ with $m_i$ and $n$ with $n_i$.
Therefore we can choose $m$, $n$ and $\phi$ so that $\phi$ has a finite dimensional indecomposable free realization,
say $(c,d)\in N$. There is a morphism $f: N\to M$ such that $f(c)= m$ and $f(d)= n$.
$$
\vcenter
\xymatrix@C=20pt@R=30pt
*+={}\ar@{-}@/^.5pc/[rrr]\ar@{-}@/_.5pc/[rrr]&*+={\bullet}\ar[d]_f\ar@{}+<0pt,12pt>*{_c}&
*+={\bullet}\ar[d]^f\ar@{}+<0pt,12pt>*{_d}&*+={}\ar@{}+<10pt,0pt>*{_N}\\
*+={}\ar@{-}@/^.5pc/[rrr]\ar@{-}@/_.5pc/[rrr]&*+={\bullet}\ar@{}+<0pt,-12pt>*{_m}&
*+={\bullet}\ar@{}+<0pt,-12pt>*{_n}&*+={}\ar@{}+<10pt,0pt>*{_M}
}}
$$
\vspace{2mm}
Suppose first that $N$ is a string module, $N=M(G)$ say, where $G$ is a finite string. Choose a standard basis of $N$
and write $c= \sum_i \lam_i c_i$ with each $c_i$ in this basis and each $\lam_i\neq 0$. By \cite[4.2]{P-P04}, $c_i$
is divisible by $D$ to both the right and the left, in particular $c_i\in \al N\cap \be N$ so, if we set $m_i=f(c_i)$
then, by Lemma \ref{D}, $m_i=0$ or $w(m_i)= {}^\infty D.D^\fty$.
The reader could have in mind the following diagram for $A= R_1$, where $D= \be\al^{-1}$ and $N$ is the following
string module:
$$
\vcenter
\xymatrix@C=10pt@R=16pt
&&&&&&*+={\circ}\ar[dl]_{\al}\ar[dr]^{\be}&&*+={\circ}\ar[dl]^{\al}\ar[dr]^{\be}&&
*+={\circ}\ar[dl]^{\al}\ar[dr]^{\be}&\\
&*+={\circ}\ar[dl]_{\be}\ar[dr]_{\al}&&*+={\circ}\ar[dl]_{\be}\ar[dr]_{\al}&&
*+={\circ}\ar[dl]_{\be}&&*+={\circ}&&*+={\bullet}\ar@{}+<0pt,-10pt>*{_{c_2}}&&*+={\circ}\\
*+={\circ}&&*+={\bullet}\ar@{}+<0pt,-10pt>*{_{c_1}}&&*+={\circ}&&&&&&&
}}
$$
\vspace{3mm}
However this diagram is misleading. For instance $v(c_1)= \be \al^{-1}\be\al\dots$ is larger than
$D^{\fty}= \be\al^{-1}\be\al^{-1}\dots$ which cannot happen (see arguments below).
Write $G=G_i.H_i$ where $c_i$ lies at the cut point and set $H_i=H_i'H_i''$ where the length of $H_i'$ is equal to
the length of $D$; then we must have $H_i'=D$: for $H_i$ must be at least $D$ since $c_i$ is divisible by $D$ to
the right, but it cannot be larger than $D$ since then the right word of $m_i$, the image of $c_i$, would begin with
a word larger than $D$ (unless $m_i= 0$). It could be that $c_i$ is so far to the right in $G$ that the length
of $H_i$ is less than that of $D$, in which case, by the same reasoning, $H_i$ must be an initial substring of $D$
(and will be a presubstring of $D$).
Similarly write $d= \sum_j \mu_j d_j$ such that each $d_j$ is an element of the chosen basis and each $\mu_j\neq 0$.
Again by choice of $\phi$ and \cite[4.2]{P-P04}, $d_j \in \ga N$, in particular $c_i\neq d_j$ for all $i, j$. We will
show that, in fact, the $c_i$ and $d_j$ are sufficiently separated in $G$. Namely we can find a submodule $L$ of $N$
which contains all the $d_j$ but no $c_i$ (with $m_i\neq 0$) and such that, if we factor out $L$, by
$\pi_L: N\la N'= N/L$, we can then embed $N'$ into $M$ so that the image of $c$ under the composite map is exactly
$m=f(c)$.
$$
\vcenter
\xymatrix@C=20pt@R=18pt
*+{N}\ar[r]^{\pi_L}\ar[d]_{f}&*+{N'}\ar[dl]^{f'}\\
*+{M}&
}}
$$
\vspace{2mm}
In order to embed $N'$ in $M$ we have to ensure that each segment of the string $G$ which remains after factoring out
$L$ is a presubstring of $^\fty D.D^\fty$. We will then be able to complete the argument since, from $N\ms \phi(c,d)$
we will obtain $N'\ms \phi(\pi_L(c),0)$ and hence $M\ms \phi(m,0)$, which is a contradiction.
We must define $L$. Fix $d_j$; we will find a presubstring of $G$ which contains $d_j$ but no $c_i$, with $m_i\neq 0$
(terminologically, we will confuse strings and their realizations). Suppose that there is some $c_k$ to the left of
$d_j$; choose the nearest one, say $c_i$. Consider the portion of $G$ between $c_i$ and $d_j$; as seen above, it has
the form $DH$ for some (at first sight possibly empty) string $H$.
$$
\vcenter
\xymatrix@C=20pt@R=14pt
&&&&*+={\circ}\ar[dl]_{\tau}&\\
*+={\circ}\ar@{}+<0pt,-10pt>*{_{c_i}}\ar@{-}@/^1pc/[rr]^{D}&&*+={\circ}\ar@{-}@/^1pc/[rrr]^(.8){H}&
*+={\circ}\ar[dr]_{\varepsilon}&&*+={\circ}\ar@{}+<0pt,-10pt>*{_{d_j}}\\
&&&&*+={\bullet}&
}}
$$
\vspace{2mm}
Note that either $H$ ends with $\ga^{-1}$ or $DH$ continues in $G$ as $DH\ga$. Working along $DH$ from $c_i$,
consider the first letter where $DH$ (or $DH\ga$) differs from, and hence, as argued above, is strictly less than,
$D^\fty$; say $D^\fty$ has $\tau$ at that point and $DH$ has $\eps^{-1}$. Clearly this alternation of letters
occurs strictly on the left of $d_j$ (otherwise $\ga$ will be a part of $D$, a contradiction). We put the (marked)
image of $\eps$ into $L$ along with all subsequent basis elements up to and including $d_j$.
On the other hand, since the corresponding letter, $\tau$, in $D^\fty$ is direct, the remaining string will, at that
point, be closed under predecessors in $D^\infty$, allowing us to use the restriction of $f$ to define the embedding
of $N'$ into $M$. We use the same procedure to the right of $d_j$ and note that what we have put into $L$ is indeed
given by a presubstring, hence is a submodule of $N$.
Now assume that $N= M(E, \lam, k)$ is a band module for some band $E$. If $E$ is equivalent to $D$ then, since $\ga$
does not appear in $D$, $\ga N= 0$, hence $d=0$. Applying $f$ we obtain $M\ms \phi(m,0)$, the required contradiction.
Suppose, then, that $E$ is not equivalent to $D$. Since $\phi$ implies $ (D.D)(x)$, there is a morphism
$g: M(D.D)\to N$ sending the basis element $z\in M(D.D)$, between two copies of $D$, to $c$. The description of
morphisms between string and band modules (see \cite[Sec. 6.3.2]{Har}) says that $g$ is a linear combination
$\sum_i \mu_i g_i$ of simple string maps (also called graph maps in \cite{Kra91}). Consider the maps $g_i$.
By the description of simple string maps, $g_i$ is given by first possibly removing arrows at either or both ends of
$D.D$ so that it equals a presubstring of\, ${}^{\fty} E^{\fty}$ - as in the argument above, this corresponds to mapping
$M(D.D)$ to a factor module and then embedding that into the direct sum module $M({}^{\fty} E^{\fty})$ - and then
applying a canonical morphism from $M(^{\fty} E^{\fty})$ to the band module $N$. Because, by Fact \ref{band}(ii) $D$
is not a presubstring of\, $^{\fty} E^{\fty}$, it follows that proper factoring must occur at each end of $D$. When
factoring we will properly increase the left and the right string of $z$ making each definitely larger than $D^{\fty}$.
Now, applying $g$ to $z$ then $f$ to $c$, we see that $m$ is a sum of elements $m_i$ such that the left and right
string of each $m_i$ exceeds $D^{\fty}$. Since $m$ is homogeneous this is not possible.
\end{proof}
Replace $A$ by the string algebra obtained by factoring out the ideal generated by all arrows that do not occur in $D$.
By the result we have just proved, this ideal is contained in the annihilator of $M$, so we may assume that each arrow
in the quiver defining $A$ occurs in $D$ or $D^{-1}$. From this it now follows that $A$ is 1-domestic.
\begin{lemma}\label{1-dom}
Every band $E$ of $A$ is a cyclic permutation of $D$ or $D^{-1}$, therefore $A$ is 1-domestic.
\end{lemma}
\begin{proof}
Suppose that $E= \eps\dots \pi^{-1}$. By assumption $D$ contains $\eps$ or $\eps^{-1}$, and also $D$ contains $\pi$ or
$\pi^{-1}$. It follows from \cite[Prop. 5.3]{Pun07} that either $D$ or $D^{-1}$ is a cyclic permutation of some band
$F= \eps \dots \pi^{-1}$. Since $A$ is domestic, we conclude that $F= E$.
\end{proof}
Thus we may assume that $A$ is 1-domestic and therefore, in view of \cite[Thm. 9.1]{Pun14}, $M$ is on Ringel's list.
This completes the proof of Theorem \ref{main}.
Recall that a (nonzero) module $M$ is said to be \emph{superdecomposable} if $M$ contains no (nonzero) indecomposable
direct summand. It is known that many (conjecturally all - see \cite{Pun04} for a precise statement) non-domestic
string algebras posses a superdecomposable pure injective module. This never happens for domestic string algebras, as
the following result shows.
\begin{theorem}\label{sup}
Every pure injective module over a domestic string algebra contains an indecomposable direct summand.
\end{theorem}
\begin{proof}
Let $M$ be a pure injective $A$-module. By Corollary \ref{hom-dom} $M$ contains a homogeneous element $m$. If the
string $w(m)$ is not periodic, it follows from Fact \ref{har} that the indecomposable pure injective module $M_w$
from Ringel's list is a direct summand of $M$.
Thus we may assume that each homogeneous element in $M$ has a periodic string. As in the proof of Theorem \ref{main}
choose a homogeneous element $m\in M$ whose string $w(m)= {}^{\fty} D. D^{\fty}$ is minimal. Let $M'$ be the hull of
$m$ in $M$: this is a direct summand of $M$ which contains $m$ and is minimal such (see, e.g., \cite[\S 4.3.5]{Preb2});
it has the property (see \cite[4.3.74]{Preb2}) that every nonzero element of $M'$ is related to $m$ by a pp formula,
as at the start of the proof of Lemma \ref{ga} (this is the only point in that proof where we used indecomposability
of $M$). Then, just as above, we obtain that $M'$ is a module over a 1-domestic algebra. But, by
\cite[Cor. 6.7]{Pun12}, there is no superdecomposable pure injective module over any 1-domestic string algebra so
$M'$, and hence $M$, has an indecomposable direct summand.
\end{proof}
\section{Reproving Harland's theorem for domestic string algebras}\label{S-repr}
In this section we develop a different approach to the proof of Fact \ref{har} that we need from Harland.
\begin{theorem} (see \cite[Thm. 51]{Har})\label{unique}
Suppose that $A$ is a domestic string algebra. Let $w= u^{-1}.v$ be a 2-sided non-periodic string. Then $M_w$
is the unique indecomposable pure injective module which contains a homogeneous element $m$ with string
$w(m)= u(m)^{-1}.v(m)$ equal to $w$.
\end{theorem}
The proof given in \cite{Har} works for arbitrary string algebras. Our approach will give the result just for domestic
string algebras.
\begin{proof}
Suppose that $M$ is an indecomposable pure injective module containing a homogeneous element $m$ with $w(m)=u^{-1}.v$
and let $p$ be the pp type of $m$ in $M$. Write $p^-$ for the set of pp formulae not in $p$ and, for emphasis, $p^+$ for the set of those which are in $p$. Using Lemma \ref{div}, if necessary, we may assume that $m$ is in the socle
of $M$.
We will choose a pair of pp formulae $\psi< \phi$ with $\phi\in p$, $\psi\notin p$ and show that the restriction
of $p$ to the interval $[\psi,\phi]$ is independent of $M$ and $m$. By Ziegler \cite[L. 7.10]{Zie} it will follow that the
isomorphism type of $M$ is uniquely determined, and hence $M$ is isomorphic to $M_w$.
Since $A$ is domestic it follows that $u^{-1}= {}^{\fty} E u'$ and $v= v' F^{\fty}$, where $E= \al \dots \be^{-1}$
and $F= \eps\dots \pi^{-1}$ are bands which are not equal up to cyclic permutation. We
set $C^{-1}= E u'$, $D= v'F$, so $\phi= (C^{-1}.D)\in p$. Also write $C^{-1}= \al E'$
and $D= F'\pi^{-1}$ and consider the formula $\psi= (E'.D)+ (C^{-1}.F')\in p^-$. Since $m$ is in the socle of $M$,
it follows that $C$ and $D$ begin with direct arrows.
By the construction of $u(m)$ and $v(m)$, the element $m$ will satisfy each formula $(C_1^{-1}.D_1)$, where
$C_1\in H_{-1}$ is such that $C_1\leq u$ and $D_1\in H_1$ is such that $D_1\leq v$. Furthermore, the set of such formulae
in $p$ is clearly closed with respect to finite conjunctions.
Furthermore, because $m\in M$ is homogeneous, $m$ satisfies no formula $(E_2^{-1}.D)+ (C^{-1}.F_2)$ for $E_2>u$ in $H_{-1}$
and $F_2>v$ in $H_1$, and the set of such formulae in $p^-$ is closed with respect to finite sums.
Each formula in the interval $[\psi, \phi]$, except $\phi$, is obtained as follows: take any formula $\chi$ strictly
below $\phi$ and add it to $\psi$. We will show that exactly one of the following holds.
\vspace{1mm}
\noindent 1) $\chi$ is implied by a formula $(C_1^{-1}.D_1)\in p$ as above, and hence
$\chi+ \psi$ must be in $p^+$;
or
\noindent 2) $\chi$ implies a formula $(E_2^{-1}.D)+ (C^{-1}.F_2)\in p^-$, therefore $\chi+ \psi$ implies a formula of similar shape and hence must be in $p^-$.
\vspace{1mm}
Note that this is independent of $M$.
Fix $\phi$ and $\psi$ as above. Denote by $n\in M(C^{-1}.D)$ the element in a standard basis of $M(C^{-1}.D)$
between $C^{-1}$ and $D$; so $n\in M(C^{-1}.D)$ is a free realization of $\phi$.
Choosing a formula $\chi$ strictly below $\phi$ is equivalent to picking a morphism $f$ from $(M(C^{-1}.D), n)$ to a
finite dimensional pointed module $(L,l)$ (such that $l\in L$ is a free realization of $\chi$) such that $f$ is not
a split embedding. Because the sets of formulae in 2) are closed with respect to finite sums, we may assume that $L$ is indecomposable, hence either a string or a band module.
If $L$ is a band module then, from the description of morphisms between string and band modules, it follows, as in
the proof of Lemma \ref{ga}, that, in applying $f$, the string $C^{-1}D$ is shortened at one or both ends and hence $l\in L$
satisfies $\psi$. Then $\chi$ implies $\psi$, hence we are in case 2) and $\chi \in p^-$.
Therefore we may assume that $L$ is a string module. Write $l$ as a linear combination of basis elements $l_i$ in
$L$, each $l_i$ corresponding to a simple string map from $M(C^{-1}.D)$ to $L$. Since $C$ and $D$ start with direct
arrows each $l_i$ lies in the socle of $L$.
As in the case that $L$ is a band module, any $l_i$ arising from a simple string map which first involves a proper
factorization of $C^{-1}.D$ on one (or both) ends is a free realization of a pp formula below $\psi$, so we can
ignore these elements for the purpose of deciding whether or not $\chi \in p$. Thus we may assume that the $H_{-1}$
string, $C_i$, of each $l_i$ in $L$ is an extension of $C$ (i.e.~$C$ is a presubstring of $C_i$); and the $H_1$
string, $D_i$, of each $l_i$ in $L$ is an extension of $D$.
If $l_i$ and $l_j$ (or rather their strings) for $i\neq j$ are embedded in $L$ with the same orientation (say, with $H_{-1}$ to
the left) then we obtain a contradiction as follows. We can suppose, without loss of generality, that $l_i$ lies
`to the left' of $l_j$:
$$
\vcenter
\xymatrix@C=12pt@R=10pt
&*+={\circ}\ar[dl]_{\al}&&&&&*+={\circ}\ar[dr]_(.3){\be}&&*+={\circ}\ar[dl]_{\al}&\\
*+={\circ}\ar@{-}@/_.7pc/[rrr]_{C^{-1}}&&&*+={\circ}\ar@{}+<0pt,-10pt>*{_{l_i}}\ar@{--}@/_.9pc/[rrrr]&&&&
*+={\circ}\ar@{-}@/_.7pc/[rrr]_{C^{-1}}&&&
*+={\circ}\ar@{}+<0pt,-10pt>*{_{l_j}}&
}}
$$
\vspace{2mm}
The string between two occurrences of $\al$ on this diagram, by Lemma \ref{band} (iii), is a power of $E$. But that is
impossible by the non-periodicity of $C^{-1}.D$.
If there is just one $l_i$ (that is, $l$ is a standard basis element), then the formula $\chi$ is equivalent to a
formula of the form $(G^{-1}.H)$ generating the pp type of $l_i$ in $L$. If $G\leq u$ and $H\leq v$, then this formula
is in $p^+$, hence we are in case 1) and $\chi \in p^+$. Otherwise, say $G> u$, therefore $(G^{-1}.H)$, hence $\chi$, already is in
$p^-$.
Thus we have reduced to the case that $l= l_1+ l_2$, with $l_1$ embedded in $L$ as $C_1^{-1}.D_1$ (with left to right
orientation), and $l_2$ embedded somewhere to the right of $l_1$ with inverse orientation $D_2^{-1}.C_2$ of strings
(the case with $l_1$ to the right of $l_2$ is treated just like this one).
Because $C$ is a presubstring of $C_1$ and $C_2$, we have $C\leq C_1, C_2$ in $H_{-1}$ and similarly $D\leq D_1, D_2$
in $H_1$.
We claim that we can assume that $C_1\leq u$ and $D_1\leq v$ (or a similar assertion for $C_2$ and $D_2$). Otherwise,
by symmetry we may suppose that $D_1> v$. If $D_2> v$ then, by Lemma \ref{triang}, $l$ is divisible by
$\min(D_1,D_2)> v$ and hence (since $l$ freely realizes $\chi$ in $L$) $\chi$ implies divisibility by $\min(D_1,D_2)$,
therefore we are in case 2) and $\chi\in p^-$.
Thus we may assume that $D_2\leq v< D_1$. We may further suppose that the strings $C_1^{-1}.D_1$ and $C_2^{-1}.D_2$
are incomparable. For otherwise the pp type of $l_2$ in $L$ is strictly less than the pp type of $l_1$, therefore we
can get rid of $l_1$ without affecting $\chi$ (the pp types of $l_2$ and $l_1+l_2$ will be equal).
It follows that $C_1< C_2$. If $C_2> u$ then we have that $(C_1.D_1)+(C_2.D_2) \in p^-$ so $\chi\in p^-$.
Otherwise $C_1< C_2\leq u$, hence we have obtained the desired conclusion for $C_2$ and $D_2$, proving the claim.
Thus we may assume that $C\leq C_1\leq u$ and $D\leq D_1\leq v$. Now, $D$ is embedded as the start of $D_1$ on the
right of $l_1$ and $D^{-1}$ is embedded as the start of $D_2^{-1}$ on the left of $l_2$; these two copies of $D$
cannot overlap each other (or even touch), otherwise we would obtain a configuration $\tau \tau^{-1}$ for some
arrow $\tau$ or its inverse, which is not possible.
Thus we have obtained the following configuration in $L$:
$$
\vcenter
\xymatrix@R=12pt@C=10pt
&&&&*+={\circ}\ar[dl]_{\eps}\ar@{--}[r]&*+={\circ}\ar[dr]^{\pi}&&*+={\circ}\ar[dl]^{\eps}\ar@{--}[r]&
*+={\circ}\ar[dr]_{\eps}&&*+={\circ}\ar[dl]_{\pi}\ar@{--}[r]&*+={\circ}\ar[dr]^{\eps}&&\\
*+={\circ}\ar@{-}@/_1pc/_{C_1^{-1}}[rr]&&*+={\circ}\ar@{}+<3pt,-10pt>*{_{l_1}}\ar@{-}@/_1pc/_{D}[rrrr]&
*+={\circ}&&&*+={\circ}\ar@{-}@/_1pc/_{t}[rrr]&&&*+={\circ}&&&
*+={\bullet}&*+={\circ}\ar@{}+<1pt,-10pt>*{_{l_2}}\ar@{-}@/^1pc/^{D^{-1}}[llll]&&
*+={\circ}\ar@{-}@/^1pc/^{C_2}[ll]
}
}
$$
\vspace{2mm}
We know that $D_1= DtD^{-1}C_2$ is less or equal to $v$. Note also that $v= D F^{\fty}$ cannot coincide with $D_1$ for
the whole length of the latter - for instance $v$ cannot include $\eps^{-1}\pi$ at the right hand end of $t$ on the
above diagram, for otherwise $\eps^{-1}\pi$ would be a substring of $F$, which contradicts Fact \ref{band}(i).
It follows that $D_1$ is different from $v$ (and hence less than $v$) somewhere before the first letter $\pi$ of
$D^{-1}$.
Consider the string module $L'= M(C_1^{-1}.D')$ which is obtained from $L$ by factoring out the (marked) image of the
rightmost $\eps$ in the diagram and everything to the right of that, and let $l'_1$ be the image of $l_1$ in this
factor module. From what we have said above it follows that $D'\leq v$ and therefore the formula $(C_1^{-1}.D')$ is
in $p^+$. Because (by the construction) this formula implies $\chi$, it follows that we are in case 1) and $\chi \in p^+$.
\end{proof}
It follows from the proof of this theorem that, if $M$ is any pure injective module with a homogeneous element $m$
with word $w$, then $M_w$ is a direct summand of $M$. Thus the second part of Harland's theorem, Fact \ref{har},
has now been justified and we have a self-contained proof of Theorem \ref{sup}.
We draw one more corollary from the proof of Theorem \ref{unique}.
\begin{cor}\label{basis}
Suppose that $w=u^{-1}v$ is a 2-sided non-periodic string over a domestic string algebra and $M_w$ is the corresponding
indecomposable pure injective module. A basis of open sets in the Ziegler topology for $M_w$ is given by the
pairs $(C^{-1}.D) \,\, / \,\, (E^{-1}.D)+ (C^{-1}.F)$, where $C\leq u< E$ in $\wh H_{-1}$ and $D\leq v< F$ in $\wh H_1$.
\end{cor}
|
\section{Introduction}
A recurrent theme in mathematics is the classification of structures up to isomorphism. In logic, it is natural to use a sentence in some formal language as a description of the structure. A countable, infinite model $S$ in a countable language $\mcL$ is usually not determined up to isomorphism by its first-order theory, even within the class of countable structures. In contrast, Scott showed that such a description for $S$ always exists when one uses the infinitary language $L_{\omega_1\omega}(\mcL)$ in which countable conjunctions and disjunctions of formulas are allowed, provided that these formulas share a finite set of free variables. A sentence whose unique countable model is the structure $S$ is called a \df{Scott sentence} for $S$.
If one only considers structures with domain $\N$, the set of isomorphic copies of a structure $S$ can be seen as the orbit of $S$ under the natural action of $S_\infty$, and this orbit is Borel in the space of countable models in the language $\mcL$ (see \cite{Kechris1995}*{16.C} for details). Lopez-Escobar~\cite{Lopez-Escobar1965} generalised Scott's result by showing the following definability theorem: every isomorphism-invariant Borel set of $\mcL$-structures with domain $\N$ can be described by an $L_{\omega_1\omega}(\mcL)$ sentence; the converse of this assertion is also true and easy to show.
In his paper, the proof theorist Lopez-Escobar gave a syntactic proof of a version of the Craig interpolation theorem in the setting of $L_{\omega_1\omega}(\mcL)$ and then derived the definability theorem as a corollary. Somewhat later, Vaught~\cite{Vaught1974} gave a direct proof of the theorem, exploiting the action $S_\infty$ on the space of models. As a main tool he used the \df{Vaught transforms}. Since then they have played a major role in studying the actions of Polish groups.
A \df{metric structure} is a metric space together with some additional real valued predicates denoted by symbols in a language $\mcL$. In this note, we prove a generalization of the Lopez-Escobar theorem to the setting of separable metric structures and continuous logic. The language always contains a distance symbol $d$, which is a real-valued binary predicate intepreted as the metric. This corresponds to equality in first-order structures. For metric structures, instead of first-order logic, it is natural to use \df{continuous logic} \cite{BenYaacov2008}. In this setting, a sentence $\phi$ evaluates in a metric $\mcL$-structure $p$ to a real number $\phi^p$.
Similarly, instead of isomorphism invariant Borel relations on the class of countable structures, we consider isomorphism invariant Borel functions defined on the class of Polish metric structures. We only consider real-valued bounded functions. A simple example of such a function is $A(p) = \min \{1, \text{diam}(p)\}$ where $p$ is a code for Polish metric space.
We use the analogue of $L_{\omega_1\omega}$-logic for metric structures, which was introduced in \cite{BenYaacov2009a}. In this language, countable suprema $\bigvee_n \phi_n(\bar x)$ and infima $\bigwedge_n \phi_n(\bar x)$ are allowed, provided that the formulas $\phi_n$ \emph{have the same modulus of continuity}. This condition replaces the classical one that $\phi_n(\bar x)$ have the same \emph{finite} set of free variables and ensures that a formula $\phi(\bar x)$ defines a uniformly continuous function on every model with a predefined modulus of continuity that depends only on the language $\mcL$. We denote this language by $L_{\omega_1\omega}(\mcL)$ as well; the context yields the intended meaning. We refer the reader to \cite{BenYaacov2009a} for the exact description of the logic.
Let $\mcL$ be a countable language for continuous logic. We will encode a separable $\mcL$-structure by the values of the predicates on a dense subsequence thereof. More precisely, the space $\mcM(\mcL)$ (denoted only by $\mcM$ if the language $\mcL$ is understood) of (codes for) Polish metric $\mcL$-structures is the subspace of
the product $\prod_{B \in \mcL} \R^{\N^{|B|}}$ defined by: $p \in \mcM$ iff $d^p \colon \N^2 \to \R$ is a pseudo-distance and $B^p \colon \N^{|B|} \to \R$ has the continuity modulus specified by $\mcL$ for every $B \in \mcL$. A function symbol $f$ is represented by the predicate $B_f$ defined by
\[
B_f(i_0, \ldots, i_{k-1}, r) = d(f(i_1, \ldots, i_{k-1}), r).
\]
We do not assume that the predicates are bounded. It is easy to see that $\mcM$ is a $G_\delta$ subset of $\prod_{B \in \mcL} \R^{\N^{|B|}}$.
The following is the main theorem of this note.
\begin{theorem}
\label{th:Vaught-simple}
For every isomorphism-invariant bounded Borel function $A \colon \mcM \to \R$, there exists a $L_{\omega_1\omega}$-sentence $\phi$ such that
\[
A(p) = \phi^p \quad \text{for all } p \in \mcM.
\]
\end{theorem}
For example if $A(p) = \min \{1, \text{diam}(p)\}$, as a sentence $\phi$ we can take $\sup_{x,y} \widehat d(x,y)$, where $\widehat d = \min (d,1)$. This is in fact a sentence of continuous first-order logic.
Lupini and Coskey (\url{http://arxiv.org/abs/1405.2859}, Theorem 1.1) have independently obtained a related result. They prove a Lopez-Escobar type theorem for $\text{Iso}(\mathbb U)$ invariant Borel functions on a suitable variant of Urysohn space $\mathbb U$ for Polish structures. Unlike us, they use a logic with category quantifiers.
\section{Proof of the main theorem using Vaught transforms}
\label{sec:vaught}
Let $\mcL$ be a fixed countable language for continuous logic, and denote by $\mcM$ the class of Polish metric $\mcL$-structures.
Let $E$ be the (analytic) equivalence relation of isomorphism on $\mcM$. If $p \in \mcM$ is a model, we denote by $Y(p)$ the underlying (Polish) metric space and by $d^p$ its metric. Let
\[
D(p) = \set{y \in Y(p)^\N : \set{y(n) : n \in \N} \text{ is dense in } Y(p)}.
\]
Note that $D(p)$ is a dense $G_\delta$ set in $Y(p)^\N$, and therefore a Polish space \cite{Kechris1995}*{Theorem~3.11}.
For each $p \in \mcM$, there is a natural continuous surjection $\pi_p \colon D(p) \to [p]_E$ defined by
\[
B^{\pi_p(y)}(i_0, \ldots, i_{n-1}) = B^p(y(i_0), \ldots, y(i_{n-1}))
\]
for all predicates $B \in \mcL$ of arity $n$ and all $i_0, \ldots, i_{n-1} \in \N$.
Here and in what follows, we suppress from the notation the map $\N \to Y(p)$ given by the coding of $p$ (that is, we treat elements of $\N$ as names for elements of $Y(p)$).
If $X$ is a Baire space and $f \colon X \to \R$ is a Baire measurable function, define the operators $\supstar$ and $\infstar$ as follows:
\[ \begin{split}
\supstar_{x \in X} f(x) > t &\iff \exists^* x \in X \ f(x) > t \\
\infstar_{x \in X} f(x) < t &\iff \exists^* x \in X \ f(x) < t.
\end{split} \]
Here $\exists^*$ is the category quantifier ``for a non-meagre set of''. (Similarly, we define $\forall^*$ to mean ``for a comeagre set of''.) Note that if $f$ is continuous, then $\supstar f(x) = \sup f(x)$ and $\infstar f(x) = \inf f(x)$. The operators $\supstar$ and $\infstar$ are just the regular essential supremum and infimum with respect to the $\sigma$-ideal of meagre sets on~$X$. For instance, we have $\supstar f(x) = \inf \{r \colon \{ y \colon \, f(y) \ge r\} \ \text{is meagre}\}\}$.
If $y$ and $z$ are finite or infinite sequences of elements of a metric space $(Y, d)$ at least one of which is finite, define
\[
d(y, z) = \max_{i < \min(|z|, |y|)} d(y(i), z(i)).
\]
In particular, if $y$ or $z$ is the empty sequence, then $d(y, z) = 0$.
The usual development of the Vaught transform (see \cite{Vaught1974} or \cite{Kechris1995}*{16.B}) requires the presence of a group action. In our context, there is no such group action present, but it turns out that this is not essential; what is important is that on every equivalence class $[p]_E$, we have the $\sigma$-ideal of meagre sets on $D(p)$ pushed forward by the map $\pi_p$.
If $A \colon \mcM \to \R$ is a bounded Borel function and $k \in \N$, define the function $A^{*k} \colon \mcM \times \N^k \to \R$ as follows:
\[
A^{*k}(p, u) = \supstar_{y \in D(p)} A(\pi_p(y)) - k \widehat d^p(y, u),
\]
where $\widehat d^p = \min \set{d^p, 1}$. Note that $A^{*k}(p, \cdot)$ is a $k$-Lipschitz function in $\widehat d^p$. If $A$ is $E$-invariant, then $A^{*k}(p, u) = A(p)$ for all $k, u$. To see this, note that $A(\pi_p(y)) $ is constant for $y \in D(p)$, so we can replace $\supstar$ by $\sup$. The second term $- k \widehat d^p(y, u)$ vanishes with an appropriate choice of $y$.
The following result therefore implies the main theorem. We need the more complicated form to be able to carry out induction over the Borel rank of the function $A$.
\begin{theorem}
\label{th:Vaught}
For every bounded Borel function $A \colon \mcM \to \R$ and for every $k \in \N$, there exists a $L_{\omega_1\omega}$-formula $\phi_{A,k}(x_0, \ldots, x_{k-1})$ (with Lipschitz constant $k$) such that
\[
A^{*k}(p, u) = \phi_{A,k}^p(u(0), \ldots, u(k-1)) \quad \text{for all } (p, u) \in \mcM \times \N^k.
\]
\end{theorem}
\begin{proof}
It is enough to check that the class of functions $A$ that satisfy the theorem contains the basic continuous functions and is closed under pointwise limits of sequences \cite{Kechris1995}*{24.3,24.10}. Since $\lim_n f_n = \inf_n \sup_{m \geq n} f_m$, this is equivalent to showing that this class is closed under countable sups and negation.
First suppose that $A$ is a basic continuous function, that is, a bounded continuous function of finitely many $B_i(a_1, \ldots, a_{n_i})$ with $a_1, \ldots, a_{n_i} \in \set{0, \ldots, N-1}$. There exists a quantifier-free formula $\theta(z_0, \ldots, z_{N-1})$ such that
\[
A(p) = \theta^p(0, \ldots, N-1).
\]
As $A$ is continuous, $\sup$ and $\supstar$ coincide, and we can write
\[
\phi_{A, k}(\bar x) = \sup_{\bar z} \theta(\bar z) - k \widehat d(\bar x, \bar z),
\]
where $\widehat d = \min \set{d, 1}$.
Suppose now that $A = \sup_n A_n$ and that the formulas $\phi_{{A_n}, k}$ have already been constructed.
We verify that
\[
\phi_{A, k}(\bar x) = \bigvee_n \phi_{A_n, k}(\bar x).
\]
works, using that $\supstar$ commutes with taking countable suprema. By inductive hypothesis, %
\begin{eqnarray*}
A^{*k}(p, u) & = & \supstar_{y \in D(p)} \sup_n A_n (\pi_p(y)) - k \widehat d^p(y, u) \\
& =& \sup_n \supstar_{y \in D(p)}A_n (\pi_p(y)) - k \widehat d^p(y, u) \\
& = & \sup_n \phi_{A_n,k}^p(u(0), \ldots, u(k-1)).
\end{eqnarray*}
Finally, suppose that $A = -B$ and $\phi_{B, m}$ have already been constructed for every $m$. We show that
\[
\phi_{A,k}(\bar x) = \bigvee_m \sup_{z_0, \ldots, z_{m-1}} - kd(\bar x, \bar z) - \phi_{B, m}(\bar z)
\]
works. Fix $p$ and $u$, let $d = \min (d^p,1)$, and, suppressing $\pi_p$ from the notation, denote
\[ \begin{split}
L &= A^{*k}(p, u) = \supstar_{y \in D(p)} A(y) - kd(y, u) \text{ and} \\
R &= \phi_{A, k}^p(u) = \sup_m \sup_{z \in Y(p)^m} \infstar_{v \in D(p)} -kd(z, u) + A(v) + m d(v, z).
\end{split} \]
(The last equality holds by the inductive hypothesis that \begin{center} $\phi_{B, m}^p(u) = \supstar_{v \in D(p)} B(v) -m d(v,z)$.) \end{center}
Our goal is to show that $L = R$.
For $R \le L$, suppose that $L < r$ for a real $r$. Then there exists $\eps > 0$ such that \[ \forall^* y \in D(p)\ A(y) - kd(y, u) < r - \eps. \] Let $m$ and $z \in Y(p)^m$ be arbitrary (by increasing $m$ if necessary, we can assume that $m \geq k$). Let $U = \set{v \in D(p) : d(v, z) < \eps/(2m)}$. Then for every $v \in U$,
\[
|\big(-k d(z, u) + A(v) + m d(v, z)\big) - \big(A(v) - k d(v, u)\big)| < \eps
\]
(to see this note that $A(v)$ cancels and $k( d(v,u ) - d(z,u)) \le k d(v,z) \le \eps/2$ because $m \ge k$). For a non-meagre set of $v \in U$, the second term in the last displayed inequality is less than $r - \eps$, so $R < r$.
To show that $L \le R$, let $M$ be an upper bound for $|A|$. Suppose that $L > r$ for a real $r$. Then there exists $\eps > 0$ and an open set $W \sub D(p)$ such that \[ \forall^* y \in W \ A(y) - k d(y, u) > r + \eps. \] Choose $\delta > 0$, $m \in \N$ and $z \in Y(p)^m$ such that
\begin{itemize}
\item $B_\delta(z) = \set{v \in D(p) : d(v, z) < \delta} \sub W$;
\item $k \delta < \eps$;
\item $m \delta > M + r + k d(u, z)$.
\end{itemize}
To do this, first choose $z$ and $\delta$ satisfying the first two conditions and then choose $m$ (and prolong $z$ arbitrarily if necessary) to satisfy the third. Now observe that if $v \notin B_\delta(z)$, then the term $m d(v, z)$ is sufficiently large to ensure that $R > r$. On the other hand, for a comeager set of $v \in B_\delta(z)$, $A(v) - k d(v, u) > r + \eps$ and for those $v$,
\[
-k d(z, u) + A(v) + m d(v, z) \geq -k d(z, u) + A(v) \geq (r + \eps) - k \delta > r,
\]
thus proving that $R > r$.
\end{proof}
Note that the complexity of the formulas $\phi_A$ is controlled by the complexity of~$A$.
Theorem~\ref{th:Vaught-simple} now follows quickly. Suppose $A \colon \mcM \to \R$ is an isomorphism-invariant, bounded Borel function. In Theorem~\ref{th:Vaught} take $u = \emptyset$ and observe that $A^{*0}(p, \emptyset) = A(p)$ for every $p$. Thus, $\phi= \phi_{A,0} $ is an $L_{\omega_1\omega}(\mcL)$-sentence such that $A(p) = \phi^p$ for every $p$.
The equivalence relation $E$ on $\mcM$ is reducible to the orbit equivalence relation of an action of a Polish group by Elliott et al. \cite{Elliott2013}*{Theorem~2.2 and Section 4}. So the $E$-classes are Borel and Theorem~\ref{th:Vaught-simple} yields the following.
\begin{cor}[Continuous Scott sentences]
For each Polish structure $S$ in $\mcL$, there is a sentence $\phi \in L_{\omega_1\omega}(\mcL)$ such that
\[
\phi^q = \begin{cases}
0, &\text{if } q \cong S, \\
1, &\text{otherwise}
\end{cases}
\qquad \text{for every } q \in \mcM.
\]
\end{cor}
As a further example, there is a sentence $\gamma \in L_{\omega_1\omega}(\mcL)$ such that $\gamma^p = 0$ if $p$ is compact and $\gamma^p = 1 $ otherwise.
Finally, we obtain the infinitary version of continuous Craig interpolation.
\begin{cor}[Interpolation]
\label{c:interpolation}
Suppose that $\mcL_1$ and $\mcL_2$ are languages and $\phi_1$ and $\phi_2$ are $L_{\omega_1\omega}(\mcL_1)$ and $L_{\omega_1\omega}(\mcL_2)$ sentences respectively. Suppose that $\phi^p_1 \leq \phi^p_2$ for every model $p$ for $L_{\omega_1 \omega}(\mcL_1 \cup \mcL_2)$. Then there is an interpolating sentence $\theta$ in $L_{\omega_1\omega}(\mcL_1 \cap \mcL_2)$ such that $\models \phi_1 \leq \theta \leq \phi_2$.
\end{cor}
By $\models \phi_1 \leq \theta$ we mean that for each $\mcL_1$-structure $p_1$ and its restriction $p_0$ to $\mcL_1 \cap \mcL_2$, we have $\phi^{p_1} \le \theta^{p_0}$. We interpret $\models \theta \le \phi_2$ in a similar way.
\begin{proof}
Let $\mcL_0 = \mcL_1 \cap \mcL_2$ and let $\pi_1 \colon \mcM(\mcL_1) \to \mcM(\mcL_0)$ and $\pi_2 \colon \mcM(\mcL_2) \to \mcM(\mcL_0)$ be the natural restriction maps. Let $E$ be the analytic equivalence relation of isomorphism on $\mcM(\mcL_0)$. For every $r \in \Q$, consider the two analytic sets
\[ \begin{split}
\set{p \in \mcM(\mcL_0) &: \exists q \in \mcM(\mcL_1) \ \pi_1(q) = p \And \phi_1^q > r} \text{ and} \\
\set{p \in \mcM(\mcL_0) &: \exists q \in \mcM(\mcL_2) \ \pi_2(q) = p \And \phi_2^q < r}.
\end{split} \]
They are $E$-invariant and by hypothesis, they are disjoint. By \cite{Kechris1995}*{Exercise~14.14}, there exists an invariant Borel set $C_r$ that separates them. Define $A \colon \mcM(\mcL_0) \to \R$ by
\[
A(p) = \sup \set{r : p \in C_r}.
\]
The function $A$ is $E$-invariant and Borel. So by Theorem~\ref{th:Vaught}, there exists a sentence $\theta$ such that $A(p) = \theta^p$ and then $\models \phi_1 \leq \theta \leq \phi_2$.
\end{proof}
\iffalse
\section{The model-theoretic approach}
\label{sec:model-theoretic}
We give a direct proof of Cor.\ \ref{c:interpolation} and derive Theorem~\ref{th:Vaught-simple} from it.
Let $\mcL$ be a language for continuous logic.
Recall that a \df{fragment} of $L_{\omega_1 \omega}(\mcL)$ is a separable subspace\marginpar{What's that mean?} of $L_{\omega_1 \omega}(\mcL)$ containing all quantifier-free formulas and closed under continuous combinations, quantifiers, subformulas, and substituting terms for variables. Every countable subset of $L_{\omega_1 \omega}(\mcL)$ generates a fragment; if $\Phi_1$ and $\Phi_2$ are fragments, then $\Phi_1 \cap \Phi_2$ is one too; the union $\Phi_1 \cup \Phi_2$ also generates a fragment that we will denote by $\Phi_1 \vee \Phi_2$. The following distributivity law holds: for all fragments $\Phi$, $\Psi$, and $\Xi$,
\begin{equation} \label{eq:1}
(\Phi \vee \Psi) \cap \Xi = (\Phi \cap \Xi) \vee (\Psi \cap \Xi).
\end{equation}
If $\Phi$ is a fragment of $L_{\omega_1 \omega}(\mcL)$, let $\mcL^\Phi$ be the language that contains a relation symbol $P_\phi(\bar x)$ for every formula $\phi(\bar x) \in \Phi$ (with the same arity and continuity modulus). Every formula $\phi$ in $\mcL^\Phi$ gives rise to an $L_{\omega_1 \omega}(\mcL)$-formula $\hat \phi$, where each predicate symbol $P_\psi$ is replaced by the corresponding formula $\psi$. Every model in $p \in \mcM(\mcL)$ expands naturally to a model $p^\Phi$ in $\mcL^\Phi$ by interpreting $P_\phi$ as $\phi$; let $\mcM^\Phi = \set{p^\Phi : p \in \mcM}$ and let $T^\Phi$ denote the theory of all models in $\mcM^\Phi$:
\[
T^\Phi = \set{\phi = 0 : \phi \text{ is a sentence in } \mcL^\Phi \And \hat \phi^p = 0
\text{ for all } p \in \mcM}.
\]
We have the following basic fact about the theories $T^\Phi$. Let $\mcL_1$, $\mcL_2$ be languages.
\begin{lemma}
\label{l:union}
If $\Phi$ and $\Psi$ are fragments in $L_{\omega_1 \omega}(\mcL_1)$ and $L_{\omega_1 \omega}(\mcL_2)$, respectively, then $T^{\Phi \vee \Psi}$ is equivalent to $T^\Phi \cup T^\Psi$.
\end{lemma}
Let $\tS_\omega(\Phi)$ denote the space of types over $T^\Phi$ in $\omega$ variables. There is a natural map $\iota \colon \mcM \to \tS_\omega(\Phi)$ defined by evaluation. Then we have the following simple lemma.
\begin{lemma}
\label{l:models-Gdelta}
The set $\mcM^\Phi$ is dense $G_\delta$ in $\tS_\omega(\Phi)$.
\end{lemma}
Another way to view $S_\omega(\Phi)$ is as follows: define $\iota \colon \mcM \to \R^\Phi$ by $\iota(p)(\phi) = \phi^p(0, \ldots, n-1)$, where $\phi$ is a formula with $n$ free variables, and let $\tS_\omega(\Phi)$ be the closure of $\iota(\mcM)$ in $\R^\Phi$.
If $X$ and $Y$ are compact spaces, $\pi \colon X \to Y$ is a surjective continuous map, and $\phi$ is a real-valued continuous function on $X$, define $\pi_*\phi \colon Y \to \R$ by
\[
(\pi_*\phi)(y) = \inf \set{\phi(x) : \pi(x) = y}
\]
and note that $\pi_* \phi$ is lower semicontinuous.
The following proposition goes back to Robinson (for classical logic) and is well-known to be equivalent to the classical Craig interpolation theorem for first-order logic.
\begin{prop}[Robinson consistency lemma]
\label{p:Robinson}
Let $T_1$ and $T_2$ be two theories in the languages $\mcL_1$ and $\mcL_2$ respectively. If there exists a complete theory $T_0$ in the intersection language $\mcL_1 \cap \mcL_2$ that is consistent with both $T_1$ and $T_2$, then $T_1 \cup T_2$ is consistent.
\end{prop}
\begin{cor}
\label{c:amalgamation}
Let $\Phi$ and $\Psi$ be fragments (perhaps in different languages) and consider the commutative diagram
\[ \xymatrix{
& \tS_\omega(\Phi \vee \Psi) \ar[ld]_{\rho_1} \ar[rd]^{\rho_2} & \\
\tS_\omega(\Phi) \ar[dr]^{\pi_1} & & \tS_\omega(\Psi) \ar[dl]_{\pi_2} \\
& \tS_\omega(\Phi \cap \Psi), &
} \]
where the maps are the natural projections. Let $\phi \in \Phi$ be any formula. Then
\[
(\rho_2)_*(\phi \circ \rho_1) = ((\pi_1)_*\phi) \circ \pi_2.
\]
\end{cor}
\begin{proof}
By definition,
\[
\begin{split}
(\rho_2)_*(\phi \circ \rho_1)(q) &= \inf \set{\phi(\rho_1(p)) : \rho_2(p) = q} \text{ and}\\
((\pi_1)_*\phi) \circ \pi_2(q) &= \inf \set{\phi(s) : \pi_1(s) = \pi_2(q)}.
\end{split}
\]
It is obvious that the second $\inf$ is not larger than the first (one can always take $s = \rho_1(p)$); for the other direction, given $s \in S_\omega(\Phi)$ and $q \in S_\omega(\Psi)$ such that $\pi_1(s) = \pi_2(q)$, use Proposition~\ref{p:Robinson} to find $p \in S_\omega(\Phi \vee \Psi)$ such that $\rho_1(p) = s$ and $\rho_2(p) = q$.
\end{proof}
\begin{theorem}
\label{th:open-fragment}
Let $\mcL \sub \mcL'$ be two languages in continuous logic and $\Phi$ be an $L_{\omega_1 \omega}(\mcL')$-fragment. Then there exists an $L_{\omega_1 \omega}(\mcL)$-fragment $\Psi$ such that the natural projection map $\pi \colon \tS_\omega(\Phi \vee \Psi) \to \tS_\omega(\Psi)$ is open.
\end{theorem}
\begin{proof}
Let $\Psi_0 = \Phi \cap L_{\omega_1 \omega}(\mcL)$. We construct by induction fragments $\Psi_0 \sub \Psi_1 \sub \cdots$ of $L_{\omega_1 \omega}(\mcL)$ such that if $\pi_{k+1} \colon S_\omega(\Phi \vee \Psi_{k+1}) \to S_\omega(\Psi_{k+1})$ denotes the natural projection map, for every $\phi \in \Phi \vee \Psi_k$, we have that $(\pi_{k+1})_*\phi$ is continuous. Once the construction is completed, note
\end{proof}
\begin{cor}[Beth definability]
\label{c:Beth}
Let $\mcL \sub \mcL'$ be two languages and $\phi$ an $L_{\omega_1 \omega}(\mcL')$-sentence with the property that for every two models $p_1, p_2 \in \mcM(\mcL')$, if $p_1|_\mcL = p_2|_\mcL$, then $\phi^{p_1} = \phi^{p_2}$. Then there exists an $L_{\omega_1 \omega}(\mcL)$-sentence $\psi$ such that $\models \phi = \psi$.
\end{cor}
\begin{proof}
Let $\Phi$ be the fragment generated by $\phi$ and let $\Psi$ be as given by Theorem~\ref{th:open-fragment}. Let $\pi$ denote the natural projection $\tS_\omega(L', \Phi \vee \Psi) \to \tS_\omega(L, \Psi)$. As $\pi$ is open, $\pi_*\phi$ is continuous and so agrees with a formula $\psi$. SHOW THAT $\psi$ IS A SENTENCE.
We now show that $\psi^p = \phi^p$ for every $p \in \mcM(\mcL')$. First, from the definition of $\psi$, it is clear that $\models \psi \leq \phi$. Suppose, for contradiction, that the open set $U = \set{p \in \tS_\omega(\mcL', \Phi \vee \Psi) : \psi^p < \phi^p}$ is non-empty.
\end{proof}
NEED TO FINISH THIS
\fi
\iffalse
\section{The Gromov--Hausdorff distance and bi-Katetov functions}
If $X$ and $Y$ are metric spaces, define the Gromov--Hausdorff distance $d_{GH}$ between $X$ and $Y$ to be
\[
\inf \set{d_H(iX, jY) : i \colon X \to Z, j \colon Y \to Z},
\]
where $i$ and $j$ are embeddings in a third metric space $Z$ and $d_H$ is the Hausdorff distance:
\[
d_H(A, B) = \inf \{r : A \sub (B)_r \And B \sub (A)_r\},
\]
where $A$ and $B$ are subsets of a metric space $Z$.
A \df{bi-Katetov function} $f \colon X \times Y \to \R$ is defined as
\[
f(x, y) = d_Z(i(x), j(y)),
\]
where $i, j$ are embeddings into some metric space as above. Equivalently, $f$ is $1$-Lipschitz in both variables and
\begin{align*}
d_A(x, w) &\leq f(x, y) + f(w, y) \\
d_B(y, z) &\leq f(x, y) + f(x, z)
\end{align*}
for all $x, w \in A$ and $y, z \in B$. It follows that if $\diam(X), \diam(Y) < 1$, the space of bounded-by-$1$ bi-Katetov functions on $X \times Y$ is compact. Clearly this space is metrisable because one can view it as a subset of $[0, 1]^{D \times E}$, where $D$ is a countable dense subset of $X$ and $E$ is a countable dense subset of $Y$.
A bi-Katetov function $f$ can be seen as an approximate isometry. Its error $q_f$ is given by
\[ q_f= \max(\sup_x \inf_y f(x, y), \sup_y \inf_x f(x, y)).\]
By definition this equals the Hausdorff distance of the isometric images above.
For instance, if there is an actual onto isometry $\theta: X \to Y$, we can let $f(x,y) = d_Y(\theta(x), y)$ and obtain the optimal error $0$. Conversely, if the spaces are complete and the error is $0$ then there is an onto isometry.
Clearly we have
\[ d_{GH}(X, Y) = \inf_{f } q_f,
\]
where $f$ runs through all the bi-Katetov functions on $X \times Y$.
\begin{remark} \label{rem: extension} f $A \sub X$ and $B \sub Y$, then any bi-Katetov function defined on $A \times B$ extends to one defined on $X \times Y$ by amalgamation:
\[
f'(x, y) = \inf_{a, b} d_X(x, a) + f(a, b) + d_Y(b, y).
\]
\end{remark}
\subsection{Scott analysis of $d_{GH}$}
\
We define approximations to $d_{GH}$ from below by induction on countable ordinals.
\noindent Suppose $\bar a = \seq {a_i}_{i < n}$ and $\bar b = \seq {b_i}_{i < n}$ are enumerated finite metric spaces. Following \cite{Uspenskii2008} define
\[
r_{0, n} (\bar a, \bar b) = \inf_{f \ \text{is bi-Katetov on} \, \bar a \times \bar b } \max_{i< n} f(a_i, b_i).
\]
An explicit expression for this is given by \cite{Uspenskii2008}*{Proposition 7.1}:
\begin{equation} \label{eq:Uspenskii r0}
r_{0, n} (\bar a, \bar b) = \varepsilon /2 \, \text{ where } \, \varepsilon = \max_{i,k < n} | d(a_i, a_k) - d(b_i, b_k)|.
\end{equation}
(In fact, Uspenskii builds a bi-Katetov function such that $f(a_i, b_i) = \varepsilon/2 $ for each~$i$.)
\begin{defn} Suppose $A$ and $B$ are metric spaces and $\bar a \in A^n, \bar b \in B^n$. Define by induction on ordinals $\alpha$:
\begin{align*}
r_{0, n}^{A, B}(\bar a, \bar b) &= r_{0, n}(\bar a, \bar b) \\
r_{\alpha+1, n}^{A, B}(\bar a, \bar b) &= \max \big(
\sup_{x \in A} \inf_{y \in B} r_{\alpha, n+1}^{A, B}(\bar a x, \bar b y),
\sup_{y \in B} \inf_{x \in A} r_{\alpha, n+1}^{A, B}(\bar a x, \bar b y)
\big) \\
r_{\alpha, n}^{A, B}(\bar a, \bar b) &= \sup_{\beta < \alpha} r_{\beta, n}^{A, B}(\bar a, \bar b), \quad \text{for } \alpha \ \text{a limit ordinal}.
\end{align*}
\end{defn}
We will need to verify some properties of these functions. We will use the following well-known fact.
\begin{lemma}
\label{l:incr-fun}
Suppose $X$ is a separable metric space. Let $\seq {f_\alpha}_{\alpha < \omega_1}$ be a sequence of continuous functions $X \to [0, 1]$ such that $f_\alpha(x) \le f_\beta(x)$ for each $x\in X$ and $\alpha < \beta$. Then there exists $\alpha^*$ such that $f_\beta = f_{\alpha^*}$ for each $\beta\ge \alpha*$.
\end{lemma}
\begin{proof}
Let $\seq{B_n}_{n \in \omega}$ be a countable basis of $X$. Suppose $\seq{C_\alpha}_{\alpha < \gamma}$ is an increasing sequence of closed sets in $X$. For each $\alpha < \beta$ let $g({\alpha, \beta}) $ be the least $n$ such that $B_n \cap C_\alpha = \emptyset$ and $B_n \cap C_\beta \neq \emptyset$. Since $g$ is injective, $\beta$ is countable.
Using this fact one sees that for each $q \in \mathbb Q$ the sequences $f^{-1}_\alpha([q, 1])$ of closed subsets of $X$ stabilizes, from which the result follows.
\end{proof}
Given a metric space $(X, d)$ and $n \ge 1$, we equip $X^n$ with the ``maximum'' metric $d (\bar u, \bar v) = \max_{i< n} d(u_i, v_i)$.
\begin{lemma} \label{lem: basic props of r's} Fix separable metric spaces $A, B$ of finite diameter.
\begin{enumerate}
\item \label{i:1} For each $\alpha$ and each $n$, the functions $r_{\alpha, n}^{A, B}(\bar a, \bar b)$ are 1-Lipschitz in $\bar a$ and $\bar b$.
\item \label{i:2} The functions $r_{\alpha, n}^{A, B}(\bar a, \bar b)$ are nondecreasing in $\alpha$.
\item \label{i:3}There is $\alpha < \omega_1$ after which all the $r_{\alpha, n}^{A, B}$ stabilize. \end{enumerate}
\end{lemma}
\begin{proof}
\eqref{i:1} For $\alpha = 0$ this is immediate by Uspenskii's expression (\ref{eq:Uspenskii r0}). The limit case is immediate too. For the successor case, it suffices to observe that for metric spaces $X,Y$ and a bounded function $g \colon X \times Y \to \mathbb R^+_0$ that is 1-Lipschitz in both arguments, the functions $h(x) = \inf_y g(x,y)$ and $\ell (x) = \sup_y g(x,y)$ are 1-Lipschitz. For the infimum case, suppose $x_0,x_1 \in X$. For given $\gamma > 0$, choose $y$ such that $g(x_0, y) - h(x_0) \le \gamma$. Then
\bc $h(x_1) -h(x_0) \le g(x_1, y) - h(x_0) \le g(x_0, y) - h(x_0) \le d(x_0,x_1) + \gamma$. \ec By symmetry and since $\gamma$ was arbitrary, this shows $|h(x_0)- h(x_1) \le d(x_0,x_1)$ as required. The case of supremum is similar.
\medskip
\n \eqref{i:2} Let $n \le k$, $\bar a \in A^n, \bar b \in B^n, \bar c \in A^k, \bar d \in B^k$. We write $(\bar a, \bar b) \preceq ( \bar c , \bar d)$ if there is a 1-1 function $p \colon n \to k$ such that $a_i = c_{p(i)}$ and $b_i = d_{p(i)}$ for each $i< n$. Clearly if $u \in A $ and $v \in B$ then $(\bar a, \bar b) \preceq ( \bar c , \bar d)$ implies $(\bar a u , \bar b v ) \preceq ( \bar c u , \bar d v)$. Then by induction on ordinals $\beta$, using (\ref{eq:Uspenskii r0}) for the initial case, one checks that $(\bar a, \bar b) \preceq ( \bar c , \bar d)$ implies that $r_{\beta, n}^{A, B}(\bar a, \bar b) \le r_{\beta, k}^{A, B}(\bar c, \bar d)$.
In particular, we have $r_{\beta, n}^{A, B}(\bar a, \bar b) \le r_{\beta, n+1}^{A, B}(\bar a x , \bar b y)$ for any $x\in A, y \in B$, so by definition $r_{\beta, n}^{A, B}(\bar a, \bar b) \le r_{\beta +1 , n}^{A, B}(\bar a , \bar b )$.
Now fix $\alpha$. Using the above, by induction on $\beta \ge \alpha $ we can show $r_{\alpha, n}^{A, B}(\bar a, \bar b) \le r_{\beta, n}^{A, B}(\bar a, \bar b)$.
\medskip
\n \eqref{i:3} This follows from Lemma~\ref{l:incr-fun} since all the functions $r^{A,B}_{\alpha, n}$ are continuous.
\end{proof}
\begin{theorem} \label{prop:GH approx} Let $A,B$ be separable metric spaces of finite diameter. Let $\alpha^*$ be such that $r_{\alpha^*+1, n}^{A, B} = r_{\alpha^* , n}^{A, B}$ for each $n$. Then
\[
r_{\alpha^*, 0}^{A, B} = d_{GH}(A,B).
\]
\end{theorem}
\begin{proof} Since $A,B$ are fixed we suppress them in our notations. Variables $a,a_i$ etc range over $A$, and $b_i$ etc.\ range over $B$. For tuples $\bar a, \bar b$ of length $k$, let
\[ \delta_k ( \bar a , \bar b) = \inf_f \{ \max (q_f, \max_{i<k} f(a_i,b_i))\}, \]
where $f$ ranges over bi-Katetov functions on $A \times B$.
We show that for each $n$ and tuples $\bar a, \bar b$ of length $n$,
\[ r_{\alpha^*, n} (\bar a, \bar b) = \delta_n ( \bar a , \bar b).\]
For $n=0$ this establishes the theorem.
Firstly, we show by induction on ordinals $\aaa$ that
\[ r_{\alpha, n} (\bar a, \bar b) \le \delta_n ( \bar a , \bar b)).\]
The cases $\aaa = 0$ and $\aaa $ limit ordinal are immediate. For the successor case, suppose that $ \delta_n ( \bar a , \bar b) < s$ via a bi-Katetov function $f$ on $A \times B$. For each $x \in A$ we can pick $y \in B$ such that $f(x,y) < s$. Then $\delta_{n+1} (\bar a x, \bar b y) < s$ via the same $f$. Inductively we have $r_{\alpha, n+1} (\bar ax , \bar by) < s$. Similarly, for each $y \in B$ we can pick $x \in A$ such that $r_{\alpha, n+1} (\bar ax , \bar by) < s$. This shows that $ r_{\alpha, n} (\bar a, \bar b) \le s$.
Secondly, we verify that
\[ \delta_n ( \bar a , \bar b)) \le r_{\alpha^*, n} (\bar a, \bar b) \]
Let $ r_{\alpha^*, n} (\bar a, \bar b) < t$. We combine a back-and-forth argument with the compactness of the space of bi-Katetov functions in order to build a bi-Katetov function $f$ with $q_f \le t$ and $\max_{i < n} f(a_i, b_i ) \le t$.
To do so we extend $\bar a, \bar b$ to dense sequences in $A,B$ respectively. Let $D \sub A, E \sub B$ be countable dense sets. Let $\bar u ^k$ denote a tuple of length $k$; in particular, we can write $\bar a = \bar a^n$ and $\bar b = \bar b^n$.
We ensure that
\bc $ r_{\alpha^*, k} (\bar a^k, \bar b^k) < t$ for each $k \ge n$. \ec
Suppose $\bar a ^k , \bar b^k$ have been defined. If $k$ is even, let $a_k$ be the next element in $D$. Using $r_{\alpha^*+1, k} (\bar a^k, \bar b^k) = r_{\alpha^*, k} (\bar a^k, \bar b^k)$ we can choose $b_k$ so that $ r_{\alpha^*, k+1} (\bar a^{k+1}, \bar b^{k+1}) < t$. Similarly, if $k$ is odd, let $b_k$ be the next element in $E$ and choose $a_k$ as required.
By Lemma~\ref{lem: basic props of r's}(2) we have $ r_{0, k} (\bar a^k, \bar b^k) < t$ for each $k\ge n$ via some bi-Katetov function $\widetilde f_k$ defined on $\{ a _0, \ldots, a_{k-1}\} \times \{ b_0, \ldots, b_{k-1} \}$. We can extend this to a bi-Katetov function $ f_k$ defined on $A \times B$ as noted in Remark~\ref{rem: extension}. By the compactness of the space of bi-Katetov functions on $A \times B$, viewed as elements of $\mathbb R^{D \times E}$, there is a subsequence $k_0 < k_1 < \ldots$ such that $\seq {f_{k_u}}$ converges pointwise to a bi-Katetov function $f$. Since bi-Katetov functions are 1-Lipschitz in both arguments, this implies $\lim_u f_{k_u} (a_p, b_p) = f(a_p, b_p)$ for each $p$. Therefore $f(a_p,b_p) \le t$. This implies $q_f \le t$ as required.
\end{proof}
\begin{defn} The \emph{continuous Scott rank} of $A$ is least $\alpha$ for which
\[
r_{\alpha, n}^{A, A}(\bar a_1, \bar a_2) = r_{\alpha+1, n}^{A, A}(\bar a_1, \bar a_2), \quad \text{for all } n, \bar a_1, \bar a_2 \in A^n.
\]
\end{defn}
\subsection{The global picture}
\label{sec:global-picture}
Consider the compact Polish space $\mcM$ of all pseudo-metrics on $\N$. ($\mcM$ is a closed subset of $[0, 1]^{\N \times \N}$.) We will identify an $A \in \mcM$ with its completion, so that elements of $\mcM$ are codes for complete separable metric spaces.
$d_{GH}$ is a pseudo-metric on $\mcM$ and by definition, the sets
\[
\set{(A, B) \in \mcM^2 : d_{GH}(A, B) < t}
\]
are $\bSigma^1_1$ for each $t \in [0, 1]$. One can define an equivalence relation $E_{GH}$ on $\mcM$ by
\[
A \eqrel{E_{GH}} B \iff d_{GH}(A, B) = 0.
\]
This equivalence relation is $\bSigma^1_1$ but not Borel (on discrete spaces it coincides with isometry and isometry of graphs (metric spaces with distances $0-1-2$) is $S_\infty$-complete).
$E_{GH}$ does not globally coincide with isometry. (Ita\"\i's example of the Gurarii space with a smooth/non-smooth vector. Find more examples. Phillipp has some nice examples of discrete spaces, based on colorings of $\Z$ (splitting distinguished points in two).)
By induction each $r_{\alpha, n}$ is a Borel function $\mcM \times \mcM \times \N^n \times \N^n \to [0, 1]$. To show that each $E_{GH}$ class is Borel, one needs to prove the following. Fix $A_0 \in \mcM$ and let $\alpha_0 = \rank A_0$.
\begin{itemize}
\item for each $\alpha$, the set $\set{B \in \mcM : \rank(B) = \alpha}$ is Borel;
\item $B \eqrel{E_{GH}} A_0 \iff \rank(B) = \alpha_0 \And r_{\alpha, 0}^{A_0, B} = 0$.
\end{itemize}
Then one would want to show that the function $d_{GH}(A_0, \cdot)$ is Borel on $\mcM$.
\fi
\def$'${$'$}
\begin{bibdiv}
\begin{biblist}
\bib{BenYaacov2008}{incollection}{
author={{Ben Yaacov}, Ita{\"{\i}}},
author={Berenstein, Alexander},
author={Henson, C.~Ward},
author={Usvyatsov, Alexander},
title={Model theory for metric structures},
date={2008},
booktitle={Model theory with applications to algebra and analysis. {V}ol.
2},
series={London Math. Soc. Lecture Note Ser.},
volume={350},
publisher={Cambridge Univ. Press},
address={Cambridge},
pages={315\ndash 427},
url={http://dx.doi.org/10.1017/CBO9780511735219.011},
}
\bib{BenYaacov2009a}{article}{
author={Ben~Yaacov, Ita{\"{\i}}},
author={Iovino, Jos{\'e}},
title={Model theoretic forcing in analysis},
date={2009},
ISSN={0168-0072},
journal={Ann. Pure Appl. Logic},
volume={158},
number={3},
pages={163\ndash 174},
url={http://dx.doi.org/10.1016/j.apal.2007.10.011},
review={\MR{2500090 (2010j:03031)}},
}
\bib{Elliott2013}{article}{
author={Elliott, G.~A.},
author={Farah, I.},
author={Paulsen, V.~I.},
author={Rosendal, C.},
author={Toms, A.~S.},
author={Törnquist, A.},
title={The isomorphism relation for separable c*-algebras},
date={2013},
journal={Math. Res. Lett.},
volume={20},
number={6},
pages={1071\ndash 1080},
}
\bib{Kechris1995}{book}{
author={Kechris, Alexander~S.},
title={Classical descriptive set theory},
series={Graduate Texts in Mathematics},
publisher={Springer-Verlag},
address={New York},
date={1995},
volume={156},
ISBN={0-387-94374-9},
}
\bib{Lopez-Escobar1965}{article}{
author={Lopez-Escobar, E. G.~K.},
title={An interpolation theorem for denumerably long formulas},
date={1965},
ISSN={0016-2736},
journal={Fund. Math.},
volume={57},
pages={253\ndash 272},
}
\bib{Vaught1974}{article}{
author={Vaught, Robert},
title={Invariant sets in topology and logic},
date={1974/75},
ISSN={0016-2736},
journal={Fund. Math.},
volume={82},
pages={269\ndash 294},
note={Collection of articles dedicated to Andrzej Mostowski on his
sixtieth birthday, VII},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\subsubsection{\@startsection{subsubsection}{3}
\newcommand{\bB}{\mathbb{B}}
\newcommand{\conc}{^\frown}
\newcommand{\acts}{\curvearrowright}
\newcommand{\symd}{\triangle}
\newcommand{\Stab}{\mathrm{Stab}}
\newcommand{\Cay}{\mathrm{Cay}}
\newcommand{\dom}{\mathrm{dom}}
\newcommand{\E}{\mathrm{E}}
\newcommand{\sH}{\mathrm{H}}
\newcommand{\cP}{\mathcal{P}}
\newcommand{\pmp}{p{$.$}m{$.$}p{$.$}}
\renewcommand{\:}{\,:\,}
\newcommand{\res}{\restriction}
\theoremstyle{plain}
\newtheorem{theorem1}{Theorem}
\newtheorem{corollary1}[theorem1]{Corollary}
\theoremstyle{definition}
\newtheorem*{definition1}{Definition}
\newtheorem*{example1}{Example}
\theoremstyle{remark}
\newtheorem*{remark1}{Remark}
\renewcommand{\thesection}{{\bf \arabic{section}}}
\renewcommand{\thesubsection}{{\bf \thesection.\Alph{subsection}}}
\renewcommand{\theequation}{{\bf \arabic{section}.\arabic{equation}}}
\renewcommand{\thetheorem}{{\bf \thesection.\arabic{theorem}}}
\begin{document}
\pagestyle{headings}
\title{Invariant means and the structure of inner amenable groups}
\author{Robin D. Tucker-Drob}
\address{Rutgers University, Hill Center for the Mathematical Sciences, 110 Frelinghuysen Rd, Piscataway, NJ 08854-8019}
\email{<EMAIL>}
\subjclass[2010]{Primary 37A20, 43A07; Secondary 20H20}
\begin{abstract}
We study actions of countable discrete groups which are amenable in the sense that there exists a mean on $X$ which is invariant under the action of $G$. Assuming that $G$ is nonamenable, we obtain structural results for the stabilizer subgroups of amenable actions which allow us to relate the first $\ell ^2$-Betti number of $G$ with that of the stabilizer subgroups. An analogous relationship is also shown to hold for cost. This relationship becomes even more pronounced for transitive amenable actions, leading to a simple criterion for vanishing of the first $\ell ^2$-Betti number and triviality of cost. Moreover, for any marked finitely generated nonamenable group $G$ we establish a uniform isoperimetric threshold for Schreier graphs $G/H$ of $G$, beyond which the group $H$ is necessarily weakly normal in $G$.
Even more can be said in the particular case of an atomless mean for the conjugation action -- that is, when $G$ is inner amenable. We show that inner amenable groups have cost 1 and moreover they have fixed price. We establish $\mathscr{U}_{\mathrm{fin}}$-cocycle superrigidity for the Bernoulli shift of any nonamenable inner amenable group. In addition, we provide a concrete structure theorem for inner amenable linear groups over an arbitrary field. We also completely characterize linear groups which are stable in the sense of Jones and Schmidt. Our analysis of stability leads to many new examples of stable groups; notably, all nontrivial countable subgroups of the group $H(\R )$, recently studied by Monod, are stable. This includes nonamenable groups constructed by Monod and by Lodha and Moore, as well as Thompson's group $F$.
\end{abstract}
\maketitle
\section*{Introduction}\label{sec:intro}
\subsection{Amenable actions}\label{subsec:intro1} An action of a discrete group $G$ on a set $X$ is said to be {\bf amenable} if there exists a finitely additive probability measure $\bm{m} : \mathscr{P}(X)\ra [0,1]$, henceforth called a {\bf mean}, defined on the powerset of $X$, which is invariant under the action of $G$. This definition goes back to von Neumann's 1929 memoir on paradoxicality \cite{vN29}, where the notion of amenability of a group \emph{simpliciter} was also introduced: by definition, $G$ is amenable if the left translation action of $G$ on itself is amenable in the above sense.
Every action of an amenable group is amenable, and for a long time this simple observation could account for most of the known examples of amenability in actions.\footnote{There are early examples of actions of nonamenable groups with asymptotic fixed points (e.g., in \cite{MvN43}), although the amenability of such actions was not adduced until much later (e.g., in \cite{Ef75, Pr83, BH86, JS87})} A more systematic study of actions whose amenability could not be traced back to that of some acting group began with van Douwen's constructions of amenable actions of the free group \cite{vD90}, and has continued in recent years with \cite{MP03}, \cite{Pe03}, \cite{GM07}, \cite{GN07}, \cite{Mo10}, \cite{AE11c}, \cite{Mo11}, \cite{Mo11b}, \cite{Fi12}, \cite{JM13}, \cite{JNS13}, \cite{JS13}. A stunning recent application of amenable actions is in the article \cite{JM13} of Juschenko and Monod in which the authors turn the classical implication on its head, deducing amenability of a group from that of an action, thereby providing the first examples of infinite finitely generated simple amenable groups.
If a nonamenable group $G$ acts amenably on $X$ then it is well known that this action is far from being free: any $G$-invariant mean $\bm{m}$ on $X$ must concentrate on the set of points $x\in X$ whose associated stabilizer subgroup $G_x$ is nonamenable. Our first result strengthens this considerably by showing that, on a $\bm{m}$-conull set, the subgroups $G_x$ are in fact so large in $G$ as to be "visible from above." The precise statement uses the following variation of Popa's notions of $q$-normality and $wq$-normality \cite{Pop06b}. A subgroup $H$ of $G$ is said to be {\bf $q^*$-normal} in $G$ if the set $\{ g\in G\csuchthat gHg^{-1}\cap H \text{ is nonamenable}\}$ generates $G$. The subgroup $H$ is {\bf $wq^*$-normal} in $G$ if there exists an ordinal $\lambda$ and an increasing sequence $(H_\alpha )_{\alpha \leq \lambda}$ of subgroups of $G$, with $H_0=H$ and $H_\lambda = G$, such that $\bigcup _{\beta <\alpha} H_\beta$ is $q^*$-normal in $H_\alpha$ for all $\alpha \leq \lambda$. The notions of {\bf $q$-normal} and {\bf $wq$-normal} subgroups are defined in the same way, except with "nonamenable" replaced by "infinite." It is immediate that a $wq$-normal subgroup is necessarily infinite, and a $wq^*$-normal subgroup is necessarily nonamenable.
\begin{theorem1}\label{thm:wqstar}
Let $G$ be a finitely generated nonamenable group. Assume that $G$ acts amenably on $X$ and fix a $G$-invariant mean $\bm{m}$ on $X$. Then $G_x$ is $wq^*$-normal in $G$ for $\bm{m}$-almost every $x\in X$.
\end{theorem1}
\begin{example}\label{ex:fg}
The assumption of finite generation is necessary in the statement of Theorem \ref{thm:wqstar}. Let $G$ be a free group with free generating set $S = \{ s_i \} _{0\leq i<\infty}$, and let $G_n\leq G$ be the subgroup generated by $\{ s_i \} _{0\leq i<n}$. The action $G\cc X=\bigsqcup _{n\geq 0}G/G_n$ is amenable, although the stabilizer of any $x\in X$ is malnormal in $G$.
\end{example}
\begin{remark} Even when $G$ is not finitely generated Theorem \ref{thm:wqstar} can be applied to finitely generated subgroups, as in Corollary \ref{cor:obstr} below, to obtain a statement which holds for all countable groups. Corollary \ref{cor:obstr} also shows that Example \ref{ex:fg} is in fact the prototypical obstruction to $wq^*$-normality of stabilizer subgroups when $G$ is not finitely generated.
\end{remark}
\begin{remark}
Theorem \ref{thm:wqstar} can be strengthened. In \S\ref{sec:normality} we introduce a natural hierarchy of incremental strengthenings of $wq^*$-normality (see Definition \ref{def:nth}). The conclusion of Theorem \ref{thm:wqstar} then remains true when $wq^*$-normality is replaced by any of these strengthenings. In addition, a relativized version of Theorem \ref{thm:wqstar} holds; see Theorem \ref{thm:isoperim}.
\end{remark}
Theorem \ref{thm:wqstar} provides a means of studying measured group theoretical properties of $G$ via its amenable actions, since many such properties are known to be reflected in the structure of $wq$-normal subgroups. For example, Popa has shown that if $G$ contains a $wq$-normal subgroup whose Bernoulli shift is $\mathscr{U}_{\mathrm{fin}}$-cocycle superrigid, then the same holds for the Bernoulli shift of $G$ \cite{Pop07}. In \cite{PT11}, Peterson and Thom show that the first $\ell ^2$-Betti number, $\beta ^{(2)}_1(G)$, of $G$ is bounded above by that of its $wq$-normal subgroups. An analogous statement also holds for the pseudocost, $\mathscr{PC}(G)$, of $G$, see Proposition \ref{prop:Furman2}. We therefore obtain the following corollary, which strengthens a theorem of Promislow \cite{Pr83} concerning actions of free groups.
\begin{corollary1}\label{cor:l2cost}
Let $G$ be a countable nonamenable group. Assume that $G$ acts amenably on $X$ and fix a $G$-invariant mean $\bm{m}$ on $X$.
\begin{enumerate}
\item Suppose that $G$ is finitely generated. Then $\beta ^{(2)}_1(G_x)\geq \beta ^{(2)}_1(G)$ and $\mathscr{PC}(G_x)\geq \mathscr{PC}(G)$ for $\bm{m}$-almost every $x\in X$.
\item In general, if $\beta ^{(2)}_1(G)>r$, then $\beta ^{(2)}_1(G_x)>r$ for $\bm{m}$-almost every $x\in X$. Likewise, if $\mathscr{PC}(G)>r$ then $\mathscr{PC}(G_x)>r$ for $\bm{m}$-almost every $x\in X$.
\end{enumerate}
\end{corollary1}
\begin{proof}[Proof of Corollary \ref{cor:l2cost}]
Part (1) follows from Theorem \ref{thm:wqstar} using Theorem 5.6 of \cite{PT11} in the case of first $\ell ^2$-Betti number, and using Proposition \ref{prop:Furman2} in the case of pseudocost. For part (2), if $\mathscr{PC}(G) >r$ then by Proposition \ref{prop:PC} there exists a finitely generated nonamenable subgroup $H_0\leq G$ such that $\mathscr{PC}(H) >r$ for all $H_0\leq H \leq G$. By Corollary \ref{cor:obstr} there exists a $G$-map $\varphi : X\ra Y$ to a $G$-set $Y$ such that $G_x$ is $wq^*$-normal in $G_{\varphi (x)}$ and $H_0\leq G_{\varphi (x)}$ for $\bm{m}$-almost every $x\in X$. Therefore, by Proposition \ref{prop:Furman2}, $\mathscr{PC}(G_x)\geq \mathscr{PC}(G_{\varphi (x)})>r$ for $\bm{m}$-almost every $x\in X$. An analogous argument goes through for the first $\ell ^2$-Betti number using Corollary 5.13 of \cite{Ga02} in place of Proposition \ref{prop:PC}.
\end{proof}
\subsection{An isoperimetric threshold}\label{subsec:isop} Let $X$ be a $G$-set and for a finite subset $S$ of $G$ denote by $\phi _S(X)$ the {\bf isoperimetric constant} of the Schreier graph with respect to $S$, associated with the action
\begin{equation}\label{eqn:isop}
\phi _{S}(X) = \inf \Big\{ \sum _{s\in S} \frac{|sP\setminus P|}{|P|} \csuchthat P\subseteq X \text{ is finite and nonempty} \Big\} .
\end{equation}
When $S$ generates $G$, the value $\phi _{S}(G)$ is then the isoperimetric constant of the Cayley graph of $G$ with respect to $S$. We always have $\phi _{S}(X)\leq \phi _{S}(G)$, and if $S$ generates $G$ then $X$ is an amenable $G$-set if and only if $\phi _{S}(X) = 0$.
\begin{remark}
There are several variations of the definition \eqref{eqn:isop}. For example, it can often be convenient to work with the {\bf conductance constant} $h_S(X) = \frac{1}{|S|}\phi_S(X)$. See \cite{JN12} for a discussion in the case $X=G$. For our purpose, any fixed multiplicative renormalization of $\phi _S(X)$ would be suitable since our main interest will be in the ratio $\phi _S(X)/\phi _S (G)$.
\end{remark}
Assume now that $S$ generates $G$. Theorem \ref{thm:wqstar} then has a surprising consequence when combined with Kazhdan's trick. Namely, there exists a constant $\epsilon = \epsilon _{G,S} >0$ such that any subgroup $H$ of $G$ satisfying $\phi _{S}(G/H ) <\epsilon$ must be $wq^*$-normal in $G$. Indeed, otherwise, for each $n\geq 0$ there is a subgroup $H_n\leq G$ such that $\phi _{S}(G/H_n ) < 2^{-n}$ but with $H_n$ not $wq^*$-normal in $G$, so the amenable action $G\cc \bigsqcup _n G/H_n$ contradicts Theorem \ref{thm:wqstar}. While this argument does not give any indication about the actual value of $\epsilon _{G,S}$, we obtain a sharp estimate in \S\ref{subsec:isoperim}. (See \S\ref{sec:normality} for the definition of $n$-degree $\mathscr{N}^X$-$wq$-normality.)
\begin{theorem1}\label{thm:isoperim}
Let $G$ be a nonamenable group with finite generating set $S$ and let $H$ be a subgroup of $G$. If $\phi _{S}(G/H ) <\frac{1}{2}\phi _{S}(G)$, then $H$ is $wq^*$-normal in $G$. More generally, for each $G$-set $X$, and for each nonnegative integer $n$ we have the following implication
\[
\phi _S(G/H ) < \tfrac{1}{2^n}\phi _S (X) \ \Ra \ H\text{ is }n\text{-degree }\mathscr{N}^X\text{-}wq\text{-normal in }G .
\]
\end{theorem1}
\begin{example}
Theorem \ref{thm:isoperim} is sharp. Consider the free group $\F _2$ of rank $2$ with free generating set $S= \{ a, b \}$. We have $\phi _{S}(\F _2) = 1$ \cite[Example 47]{C-SGH99}. The subgroup $H=\langle a,bab^{-1}\rangle$ is \emph{not} $wq^*$-normal in $\F _2$ (although it is $q$-normal). An inspection of the Schreier graph of $\F _2 / H$ verifies that $\phi _{S}(\F _2 /H ) =\tfrac{1}{2}$.
\end{example}
\subsection{Transitive actions and a vanishing criterion} In the case of an amenable transitive action we obtain the following strengthening of Corollary \ref{cor:l2cost}.
\begin{theorem1}\label{thm:coamen}
Let $G$ be a countable group. Assume that $G$ acts amenably and transitively on an infinite set $X$ and fix some $x\in X$. If $\beta ^{(2)}_1(G_x)<\infty$ then $\beta ^{(2)}_1(G)=0$. Likewise, if $\mathscr{PC}(G_x )<\infty$ then $\mathscr{PC}(G)=1$.
\end{theorem1}
It follows that if $G$ is a group with $\beta ^{(2)}_1(G)>0$ or with $\mathscr{PC}(G)>1$, then for any finitely generated infinite index subgroup $H\leq G$, the action $G\cc G/H$ is \emph{not} amenable.
Theorem \ref{thm:coamen} is closely related to a vanishing criterion due to Peterson and Thom \cite{PT11}. They define a subgroup $H$ of $G$ to be {\bf $s$-normal} in $G$ if $gHg^{-1}\cap H$ is infinite for every $g\in G$; the notion of {\bf $ws$-normality} is then obtained by iterating $s$-normality transfinitely. Theorem 5.12 of \cite{PT11} states that if $G$ contains a $ws$-normal infinite index subgroup $H$ with $\beta ^{(2)}_1(H)<\infty$ then $\beta ^{(2)}_1(G) =0$. Ioana (unpublished) has shown that the analogous statement also holds for cost (Ioana's argument works for pseudocost as well). Theorem \ref{thm:coamen} would therefore follow from these results if the subgroup $G_x$ were always $ws$-normal in $G$. This turns out not to be the case however, as the following example shows.
\begin{example}\label{ex:ws}
Let $K$ be a group which is isomorphic with one of its proper malnormal subgroups $K_0$ (e.g., any nonabelian free group has this property, see \cite[Example 1]{BMR99}). Fix an isomorphism $\varphi :K\ra K_0$ and let $G=\langle t,K \, | \, tkt^{-1}=\varphi (k) \rangle$ be the associated HNN-extension. By Proposition 2 of \cite{MP03}, the action of $G$ on $G/K$ is amenable. However, $K$ is not $ws$-normal in $G$. To see this note that, from the semidirect product decomposition $G= \big( \bigcup _{n\geq 0}t^{-n}Kt^n \big) \rtimes \langle t \rangle$, it follows that every intermediate subgroup $K\lneq L \leq G$ contains an element of the form $g=t^{-n}kt^m$, where $n, m>0$ and $k\in K-K_0$, and clearly $gKg^{-1}\cap K = 1$.
\end{example}
\subsection{Inner amenability} In their 1943 study of $\mathrm{II}_1$ factors \cite{MvN43}, Murray and von Neumann distinguished the hyperfinite $\mathrm{II}_1$ factor from the free group factor $\mathrm{L}\F _2$ by means of {\bf property Gamma}, that is, the existence of nontrivial asymptotically central sequences. In demonstrating that $\mathrm{L}\F _2$ lacks this property, Murray and von Neumann hinted at a connection with amenability \cite[footnote 71]{MvN43}, remarking that their argument, which makes ancillary use of approximately invariant measures, closely mirrors Hausdorff's famous paradoxical division of the sphere. This connection was not made explicit however until 1975 when Effros \cite{Ef75} introduced the following group theoretic notion:
\begin{definition} A group $G$ is {\bf inner amenable} if the action of $G$ on itself by conjugation admits an atomless invariant mean.
\end{definition}
Effros showed that if a group factor $\mathrm{L}G$ has property Gamma, then $G$ is necessarily inner amenable. An ICC counterexample to the converse statement was found only very recently by Vaes \cite{Va12}.
The proof of Theorem \ref{thm:wqstar} naturally involves exploiting the tension between the nonamenability of $G$ and the amenability of the action. In the case of the conjugation action, this tension leads to remarkably strong consequences for the group theoretic and measured group theoretic structure of $G$. Many of these consequences will in fact be shown to hold in the more general setting of inner amenable pairs: If $H$ is a subgroup of $G$ then we say that {\bf the pair $(G,H)$ is inner amenable} if the conjugation action of $H$ on $G$ admits an atomless invariant mean.\footnote{This is a bit different from the notion, defined by Jolissaint in \cite{Jo13}, of {\bf $H$ being inner amenable relative to $G$}, which amounts to amenability of the conjugation action of $H$ on $G-H$. While Jolissaint's notion does not appear anywhere else in this article, to avoid conflicting terminology we will make sure to refer to inner amenability of the pair $(G,H)$ when referring to the notion defined in the main text.}
\subsection{The cost of inner amenable groups}
We let $\mathscr{C}(G)$ denote the cost of $G$, that is, $\mathscr{C}(G)$ is the infimum of the costs of free probability measure preserving actions of $G$. We let $\mathscr{C}^*(G)$ denote the supremum of the costs of free probability measure preserving actions of $G$. Then $G$ has {\bf fixed price} if $\mathscr{C}(G)=\mathscr{C}^*(G)$.
\begin{theorem1}\label{thm:FP1}
Let $G$ be a countable group.
\begin{enumerate}
\item[(1)] Suppose that $G$ contains a $wq$-normal subgroup $H$ such that $(G,H)$ is inner amenable. Then $\mathscr{C}(G)=1$.
\item[(2)] Suppose that $G$ is inner amenable. Then $\mathscr{C}(G)=1$ and $G$ has fixed price.
\end{enumerate}
\end{theorem1}
\begin{remark}\label{rem:prod}
In part (1) of Theorem \ref{thm:FP1} it would be desirable to additionally obtain that $G$ has fixed price. The proof of part (1) shows that this holds if and only if direct products of infinite groups have fixed price, which is a well known open problem.
\end{remark}
As a consequence of Theorem \ref{thm:FP1} we recover the result of Chifan, Sinclair, and Udrea \cite[Corollary D]{CSU13}, that inner amenable groups have vanishing first $\ell ^2$-Betti number. Moreover, we obtain a strengthening which holds for inner amenable pairs.
\begin{corollary1}\label{cor:l2Betti}
Let $G$ be a countable group and suppose that $G$ contains a wq-normal subgroup $H$ such that the pair $(G,H)$ is inner amenable. Then $\beta ^{(2)}_1(G)=0$. In particular, if $G$ is inner amenable then $\beta ^{(2)}_1(G)=0$.
\end{corollary1}
\begin{proof}
This follows from Theorem \ref{thm:FP1} and the inequality $\beta ^{(2)}_1(G)\leq \mathscr{C}(G)-1$
due to Gaboriau \cite{Ga02}. Alternatively, a direct proof may be obtained by observing that each step of the proof of Theorem \ref{thm:FP1} in \S\ref{sec:FP1} has an analogue for the first $\ell ^2$-Betti number.
\end{proof}
In \cite{AN08}, Ab\'{e}rt and Nikolov show that for a finitely generated, residually finite group $G$, the rank gradient of any Farber chain in $G$ is equal to one less than the cost of the associated boundary action of $G$. We therefore obtain the following corollary.
\begin{corollary1}\label{cor:RG}
Let $G$ be a finitely generated, residually finite group which is inner amenable. Then the rank gradient of any Farber chain in $G$ vanishes. In particular, the absolute rank gradient of $G$ vanishes.
\end{corollary1}
The two main ingredients in the proof of Theorem \ref{thm:FP1}.(2) concern the subgroup structure of nonamenable inner amenable groups.
\begin{theorem1}\label{thm:relmain1}
Let $G$ be a nonamenable inner amenable group. Then every nonamenable subgroup of $G$ is $wq$-normal in $G$.
\end{theorem1}
The next result roughly states that very large portions of $G$ commute. To make the statement somewhat less cumbersome we define $\mathscr{N}$ to be the collection of all nonamenable subgroups of $G$.
\begin{theorem1}\label{thm:subgroup}
Let $G$ be a nonamenable inner amenable group. Then at least one of the following holds:
\begin{enumerate}
\item[(1)] For every finite $\mathscr{F}\subseteq \mathscr{N}$ there exists an infinite amenable subgroup $K$ of $G$ such that $L\cap C_G(K)$ is nonamenable for all $L\in \mathscr{F}$.
\item[(2)] For every finite $\mathscr{F}\subseteq \mathscr{N}$ there exists an increasing sequence $M_0\leq M_1\leq \cdots$ of finite subgroups of $G$, with $\lim _{n\ra\infty}|M_n|=\infty$, such that $L\cap C_G(M_n)$ is nonamenable for all $L\in \mathscr{F}$, $n\in \N$.
\item[(3)] For every finite $\mathscr{F}\subseteq \mathscr{N}$ and every $n\in \N$ there exist pairwise commuting nonamenable subgroups $K_0,K_1,\dots , K_{n-1} \leq G$ such that $L\cap C_G(K_i)$ is nonamenable for all $L\in \mathscr{F}$, $i<n$. Moreover, there exists a sequence $(M_n)_{n\in \N}$ of finite subgroups of $G$ with $\lim _{n\ra\infty}|M_n|=\infty$ such that $L\cap C_G(M_n)$ is nonamenable for all $L\in \mathscr{F}$, $n\in \N$.
\end{enumerate}
\end{theorem1}
\begin{corollary1}\label{cor:subgroup}
Let $G$ be a nonamenable inner amenable group. Then $G$ either contains an infinite amenable subgroup or $G$ contains finite subgroups of arbitrarily large order. In addition, $G$ contains an infinite subgroup $K$ such that $C_G(K)$ is infinite.
\end{corollary1}
\begin{proof}[Proof of Corollary \ref{cor:subgroup}]
The first statement follows easily from Theorem \ref{thm:subgroup}. The second statement is clear if either (1) or (3) of Theorem \ref{thm:subgroup} holds. If (2) holds then $G$ contains an infinite locally finite subgroup, hence by \cite{HK64} $G$ contains an infinite abelian subgroup $A$, so $A\leq C_G(A)$.
\end{proof}
\subsection{Cocycle superrigidity} If $H$ is a subgroup of $G$, then a cocycle $w$ of a probability measure preserving action of $G$ is said to {\bf untwist on $H$} if $w$ is cohomologous to a cocycle $w'$ whose restriction to $H$ is a homomorphism. Following \cite{Pop07, Pop08}, let $\mathscr{U}_{\mathrm{fin}}$ denote the class of all Polish groups which embed as a closed subgroup of the unitary group of a finite von Neumann algebra. A free, probability measure preserving action of $G$ is said to be {\bf $\mathscr{U}_{\mathrm{fin}}$-cocycle superrigid} if every cocycle for the action which takes values in some group in $\mathscr{U}_{\mathrm{fin}}$ untwists on the entire group $G$.
Popa's Second Cocycle Superrigidity Theorem (Theorem 1.1 of \cite{Pop08}) provides general conditions for a cocycle which takes values in some group $L\in \mathscr{U}_{\mathrm{fin}}$, to untwist on the centralizer $C_G(H)$ of a nonamenable subgroup $H$ of $G$. The following theorem, which is joint with Adrian Ioana, strengthens Popa's theorem by showing that, under suitable conditions, the untwisting in fact occurs on the centralizer $C_{\mathscr{M}(G)}(H)$ of $H$ in the semigroup $\mathscr{M}(G)$ of all means on $G$.
\begin{theorem1}[with A. Ioana] \label{thm:superrigid1}
Let $H$ be a nonamenable subgroup of $G$ such that the pair $(G, H)$ is inner amenable. Let $G\cc ^{\sigma} (X,\mu )$ be a p.m.p.\ action of $G$ and assume that
\begin{itemize}
\item $\sigma _{|H}$ has stable spectral gap;
\item $\sigma _{|C_{\mathscr{M}(G)}(H)}$ is weakly mixing (see Definition \ref{def:WM});
\item $\sigma$ is $s$-malleable.
\end{itemize}
Then there exists a subgroup $G_0$ of $G$ with $H\leq G_0\leq G$ such that
\begin{enumerate}
\item Every $H$-conjugation-invariant mean on $G$ concentrates on $G_0$;
\item Every cocycle $w : G\times X \ra L$ which takes values in a group $L \in \mathscr{U}_{\mathrm{fin}}$ untwists on $G_0$.
\end{enumerate}
\begin{comment}
Let $G\cc ^{\sigma _0} (X,\mu )$ be a probability measure preserving action of a countable group $G$. Let $H\leq G$ be a nonamenable subgroup and assume that
\begin{itemize}
\item $\sigma _0{}_{|H}$ has stable spectral gap;
\item $\sigma _0{}_{|C_{\mathscr{M}(G)}(H)}$ is weakly mixing (Definition \ref{def:WM});
\item $\sigma _0$ is $s$-malleable.
\end{itemize}
Let $L$ be a group in $\mathscr{U}_{\text{\emph{fin}}}$ and let $d_L$ denote the (compatible, bi-invariant) metric on $L$ coming from an embedding of $L$ as a closed subgroup of the unitary group of a finite von Neumann algebra. Let $w: G\times X\ra L$ be a measurable cocycle with values in $L$. Then there exists a cocycle $w':G\times X\ra L$ cohomologous to $w$ such that:
\begin{enumerate}
\item The restriction of $w'$ to $H$ is a homomorphism;
\item For any separable subalgebra $\mathscr{A}\subseteq \ell ^\infty (G)$ there exists a closed subsemigroup $\mathscr{M}_0$ of $C_{\mathscr{M}(G)}(H)$, with $\mathscr{M}_0|\mathscr{A} = C_{\mathscr{M}(G)}(H)|\mathscr{A}$, along with a conull set $X_0\subseteq X$ and a map $\rho :G\ra L$ such that
\begin{align*}
\sup _{\bm{m}\in \mathscr{M}_0}\sup _{x\in X_0} &\int _G d_L (w'(g,x) , \rho (g) ) \, d\bm{m} = \sup _{\bm{m}\in \mathscr{M}_0} \int _G \int _X d_L (w'(g,x) , \rho (g) ) \, d\mu \, d\bm{m} = 0.
\end{align*}
\end{enumerate}
\end{comment}
\end{theorem1}
Theorem \ref{thm:superrigid1} applies to the Bernoulli shift of $G$ whenever $H\leq G$ is nonamenable and the pair $(G,H )$ is inner amenable. Applying Lemma 3.5 of \cite{Fu07}, we therefore obtain:
\begin{corollary1}\label{cor:superrigid}
Let $G$ be a countable group containing a $wq$-normal nonamenable subgroup $H$ such that the pair $(G,H)$ is inner amenable. Then the Bernoulli shift of $G$ is $\mathscr{U}_{\mathrm{fin}}$-cocycle superrigid. In particular, the Bernoulli shift of any nonamenable inner amenable group is $\mathscr{U}_{\mathrm{fin}}$-cocycle superrigid.
\end{corollary1}
Corollary \ref{cor:superrigid} strengthens a result of Peterson and Sinclair \cite{PS11}, stating that the Bernoulli shift of $G$ is $\mathscr{U}_{\mathrm{fin}}$-cocycle superrigid provided $G$ is nonamenable and $\mathrm{L}G$ has property Gamma.
The case $H=G$ of Corollary \ref{cor:superrigid} would follow from Popa's theorem combined with Lemma 3.5 of \cite{Fu07} and Theorems \ref{thm:relmain1} and \ref{thm:subgroup} above, provided that alternative (2) could be dropped from the statement of Theorem \ref{thm:subgroup}. However, the following example exhibits an inner amenable group with the property that the centralizer of every nonamenable subgroup is finite; in particular, such a group does not satisfy either of the alternatives (1) or (3) of Theorem \ref{thm:subgroup}.
\begin{example}\label{ex:faction}
Let $\F _2 \cc X$ be a transitive amenable action of the free group $\F _2$ on an infinite set $X$ with the following property: for all $u\in \F _2 - 1$ the set $\{ P\in \mathscr{P}_{\text{f}}(X) \csuchthat u\cdot P = P \}$ is finite, where $\mathscr{P}_{\text{f}}(X)$ denotes the collection of all finite subsets of $X$. Such an action is constructed in Theorem \ref{thm:safree}. The group $G= \mathscr{P}_{\text{f}}(X)\rtimes \F _2 \cong (\bigoplus _{x\in X}\Z _2 ) \rtimes \F _2$ is then inner amenable so Corollary \ref{cor:superrigid} applies to $G$. The group $G$ is also finitely generated and ICC. In addition, the centralizer of any nonamenable subgroup of $G$ is finite, or equivalently, the centralizer of every infinite subgroup $H$ of $G$ is amenable. Indeed, if $(P,u) \in H$ then we have $C_G(H)\leq C_G((P,u)) \leq \mathscr{P}_{\text{f}}(X)\rtimes C_{\F _2}(u)$, which is amenable unless $u = 1$. We may therefore assume that $H\leq \mathscr{P}_{\text{f}}(X)$, in which case, since $H$ is infinite we have
\[
C_G(H) = \mathscr{P}_{\text{f}}(X)\rtimes \{ u\in \F _2 \csuchthat u\cdot P = P \text{ for all }P\in H \} = \mathscr{P}_{\text{f}}(X)\rtimes 1 ,
\]
which is amenable.
\end{example}
\subsection{The structure of inner amenable linear groups}\label{sec:strlinear}
In \S\ref{sec:radicals} we characterize inner amenability for linear groups in terms of a certain amenable characteristic subgroup of $G$. The {\bf AC-center} of a countable group $G$ is the subgroup
\[
\mathscr{AC}(G) = \langle \{ N\leq G\csuchthat N\text{ is normal in }G\text{ and }G/C_G(N)\text{ is amenable}\}\rangle .
\]
The {\bf inner radical} of $G$ is the subgroup
\[
\mathscr{I}(G) = \langle \{ N\leq G\csuchthat N\text{ is normal in }G\text{ and the action }N\rtimes G \cc N\text{ is amenable} \}\rangle .
\]
Here, $N\rtimes G \cc N$ is the action where $N$ acts by left translation and $G$ acts by conjugation. The relevant properties of these subgroups are summarized in the following theorem.
\begin{theorem1}\label{thm:I(G)} Let $G$ be a countable group.
\begin{itemize}
\item[i.] $\mathscr{AC}(G)$ and $\mathscr{I}(G)$ are amenable characteristic subgroups of $G$;
\item[ii.] $\mathscr{AC}(G)\leq \mathscr{I}(G)$;
\item[iii.] The actions $\mathscr{AC}(G)\rtimes G \cc \mathscr{AC}(G)$ and $\mathscr{I}(G)\rtimes G\cc \mathscr{I}(G)$ are amenable;
\item[iv.] $G/C_G(\mathscr{AC}(G))$ is residually amenable;
\item[v.] If $\mathscr{I}(G)$ is infinite then $G$ is inner amenable;
\item[vi.] Let $N$ be a normal subgroup of $G$ with $N\leq \mathscr{I}(G)$. Then $\mathscr{I}(G/N) = \mathscr{I}(G)/N$;
\item[vii.] $\mathscr{I}(G/\mathscr{I}(G))=1$ and $G/\mathscr{I}(G)$ is $\mathrm{ICC}$;
\item[viii.] Every conjugation invariant mean on $G/\mathscr{I}(G)$ is the projection of a conjugation invariant mean on $G$.
\end{itemize}
Moreover, if $G$ is linear then
\begin{itemize}
\item[ix.] $\mathscr{AC}(G)=\mathscr{I}(G)$;
\item[x.] $G/C_G(\mathscr{I}(G))$ is amenable;
\item[xi.] $\mathscr{I}(G)=C_G(C_G(\mathscr{I}(G)))$;
\item[xii.]$G/\mathscr{I}(G)$ is not inner amenable;
\item[xiii.] Every conjugation invariant mean on $G$ concentrates on $\mathscr{I}(G)$;
\item[xiv.] Let $N$ be a normal subgroup of $G$ with $N\leq \mathscr{I}(G)$. Then \emph{ix.}\ through \emph{xiii.}\ all hold with $G/N$ in place of $G$.
\end{itemize}
\end{theorem1}
\begin{remark}
Theorem \ref{thm:I(G)}.xi.\ implies that if $G$ is linear then so are the groups $G/\mathscr{I}(G)$ and $G/C_G(\mathscr{I}(G))$ (see
Theorem 6.2 of \cite{We69}). It then follows from item x.\ and the Tits alternative that if $G$ is additionally finitely generated, then $\mathscr{I}(G)$ is virtually solvable.
\end{remark}
Using Theorem \ref{thm:I(G)} we are able to show that within the class of linear groups, inner amenability occurs only for the most obvious reasons: every linear inner amenable group is an amenable extension either of a group with infinite center or of a near product group in which one of the factors is infinite and amenable. More precisely, we obtain the following structure theorem for inner amenable linear groups.
\begin{theorem1}\label{thm:linear} Let $G$ be a countable linear group. Then the following are equivalent
\begin{itemize}
\item[(1)] $G$ is inner amenable.
\item[(2)] $\mathscr{I}(G)$ is infinite.
\item[(3)] There exists a short exact sequence $1\ra N \ra G \ra K \ra 1$, where $K$ is amenable
and either
\begin{itemize}
\item[$\bullet$] $Z(N)$ is infinite, or
\item[$\bullet$] $N = LM$, where $L$ and $M$ are commuting normal subgroups of $G$ such that $M$ is infinite and amenable, and $L\cap M$ is finite.
\end{itemize}
\end{itemize}
\end{theorem1}
In \cite{Sch86}, Schmidt raises the question of whether every inner amenable group $G$ possesses a free ergodic p.m.p.\ action $G\cc (X,\mu )$ which generates an orbit equivalence relation $\mc{R}^G_X$ for which the outer automorphism group of the full group $[\mc{R}^G_X]$ is not Polish, or equivalently, for which the full group $[\mc{R}^G_X]$ contains an asymptotically central sequence $(T_n)_{n\in \N}$ with $\liminf _n \mu ( \{ x \in X \csuchthat T_n x \neq x \} ) >0$. See also Problem 9.3 of \cite{Ke10}. Using Theorems \ref{thm:I(G)}, \ref{thm:linear}, and Theorem \ref{thm:linearstable} below, we obtain a positive answer to Schmidt's question when $G$ is linear.
\begin{theorem1}\label{thm:Schmidt}
A countable linear group $G$ is inner amenable if and only if there exists a free ergodic p.m.p.\ action $G\cc (X,\mu )$ such that the outer automorphism group of $[\mc{R}^G_X]$ is not Polish.
\end{theorem1}
\subsection{Stability} A discrete probability measure preserving equivalence relation $\mc{R}$ is said to be {\bf stable} if it is isomorphic to its direct product $\mc{R}\times \mc{R}_0$ with the equivalence relation $\mc{R}_0$, of eventual equality on $2^\N$ equipped with the uniform product measure. A countable group $G$ is said to be {\bf stable} if it possesses a free ergodic probability measure preserving action which generates a stable equivalence relation. Stability was introduced by Jones and Schmidt in \cite{JS87}, where it was also shown that stable groups are necessarily inner amenable. The first examples of ICC inner amenable groups which are not stable were recently constructed by Kida \cite{Ki12}; these groups are obtained as HNN extensions of property (T) groups with infinite center. Further results of Kida from \cite{Ki13a} show that if the center $Z(G)$ of a group $G$ is infinite, then the question of whether $G$ is stable is intimately related to the question of whether the pair $(G,Z(G))$ lacks relative property (T). Using Theorem \ref{thm:I(G)}, in \S\ref{sec:stability} we are able to completely characterize stability for linear groups in terms of relative property (T).
\begin{theorem1}\label{thm:linearstable}
Let $G$ be a countable linear group. Then the following are equivalent:
\begin{enumerate}
\item $G$ is stable.
\item The pair $(G, \mathscr{I}(G))$ does not have relative property \emph{(T)}.
\end{enumerate}
\end{theorem1}
\begin{remark}\label{rem:minC}
The hypothesis that $G$ is linear in Theorems \ref{thm:I(G)}.ix.-xiv., \ref{thm:linear}, \ref{thm:Schmidt}, and \ref{thm:linearstable} can be weakened: we only need to assume that $G$ satisfies the {\bf minimal condition on centralizers}, that is, every decreasing sequence $C_G(A_0)\geq C_G(A_1)\geq \cdots$ of centralizers of subsets of $G$ eventually stabilizes. Every linear group satisfies the minimal condition on centralizers, since centralizers of arbitrary subsets of $\mathrm{GL}_n(F)$ are closed in the Zariski topology. \end{remark}
In addition to Theorem \ref{thm:I(G)}, an essential component in the proof of Theorem \ref{thm:linearstable} is the following extension theorem for stability (see \S\ref{sec:Kida} for the definition of stability sequence).
\begin{theorem1}\label{thm:extension}
Let $1\ra N\ra G\ra K \ra 1$ be a short exact sequence of groups in which $K$ is amenable. Assume that there exists a probability measure preserving action $G\cc (X,\mu )$ such that the translation groupoid $N\ltimes (X,\mu )$ admits a stability sequence. Then $G$ is stable.
\end{theorem1}
Theorem \ref{thm:extension} has a variety of applications outside the context of linear groups. Under each of the following hypotheses {\bf (H1)}-{\bf (H6)}, the stability of $G$ will be established in \S\ref{sec:stability} by applying Theorem \ref{thm:extension} to an appropriate input action of $G$. The application of Theorem \ref{thm:extension} to groups satisfying {\bf (H4)} and the ensuing Corollary \ref{cor:gBS} were kindly suggested by Yoshikata Kida (remarking on an earlier draft of this paper), who had obtained stability of $G$ from {\bf (H4)} by different means.
\begin{theorem1}\label{thm:extension2}
Let $1\ra N\ra G\ra K \ra 1$ be a short exact sequence of groups in which $K$ is amenable. Then $G$ is stable provided at least one of the following hypotheses holds:
\begin{enumerate}
\item[{\bf (H1)}] $N = LM$, where $L$ and $M$ are commuting subgroups of $N$ which are normal in $G$, with $M$ amenable and $[N:L]=\infty$.
\item[{\bf (H2)}] There exists a central subgroup $C$ of $N$ such that the pair $(N,C)$ does not have relative property \emph{(T)}.
\item[{\bf (H3)}] There exists a sequence $L_0\leq L_1\leq \cdots$, of subgroups of $N$ with $N=\bigcup _{m\in \N}L_m$, and for each $m\in \N$ there exists a central subgroup $D_m$ of $L_m$ such that the pair $(L_m, D_m )$ does not have relative property \emph{(T)}.
\item[{\bf (H4)}] There exists a commensurated abelian subgroup $A$ of $G$ such that $N$ is the kernel of the modular homomorphism from $G$ into the abstract commensurator of $A$, and the pair $(N,A)$ does not have relative property \emph{(T)}.
\item[{\bf (H5)}] $N$ has the Haagerup property and is asymptotically commutative, i.e., there exists an injective sequence $(c_n)_{n\in \N}$ in $N$ such that each $h\in N$ commutes with $c_n$ for cofinitely many $n\in \N$.
\item[{\bf (H6)}] $N$ is doubly asymptotically commutative, i.e., there exist sequences $(c_n)_{n\in\N}$ and $(d_n)_{n\in \N}$ in $N$ such that $c_nd_n\neq d_nc_n$ for all $n\in \N$, and each $h\in N$ commutes with both $c_n$ and $d_n$ for cofinitely many $n\in \N$.
\end{enumerate}
\end{theorem1}
\begin{corollary1}[Y. Kida] \label{cor:gBS}
Let $G$ be a generalized Baumslag-Solitar group (i.e., the Bass-Serre fundamental group of a finite graph of infinite cyclic groups), or an HNN-extension of $\Z ^n$ relative to an isomorphism between two finite index subgroups. Then $G$ is stable.
\end{corollary1}
\begin{proof}[Proof of Corollary \ref{cor:gBS}]
Suppose first that $G$ is a generalized Baumslag-Solitar group. Then any vertex group $A\leq G$ is commensurated by $G$, and if $N$ denotes the kernel of the modular homomorphism from $G$ into the abstract commensurator $\mathrm{comm}(A)$ of $A$, then $G/N$ is abelian, since $\mathrm{comm}(A)$ is isomorphic to $\Q ^*$. By Corollary 1.7 of \cite{CV12}, $G$ has the Haagerup property, so the pair $(N,A)$ does not have property (T). The hypothesis {\bf (H4)} is therefore satisfied, so $G$ is stable by Theorem \ref{thm:extension2}.
The case where $G$ is an HNN-extension of $\Z ^n$ relative to an isomorphism between two finite index subgroups is similar. The image of $G$ under the modular homomorphism into the abstract commensurator of $\Z ^n$ is cyclic and, letting $N$ denote the corresponding kernel, the pair $(N,\Z ^n )$ does not have property (T) since by Corollary 1.7 of \cite{CV12}, $G$ has the Haagerup property. Hypothesis {\bf (H4)} once again holds, so $G$ is stable by Theorem \ref{thm:extension2}
\end{proof}
\begin{example} (i) Let $K$ be an infinite amenable group acting on a countable set $X$, and let $H$ be any countable group. Then the restricted wreath product $H\wr _X K$ is stable. This is clear if $H$ is amenable, and it follows from Theorem \ref{thm:extension2} via {\bf (H1)} if $X$ is finite. In the remaining case, the group $\bigoplus _X H$ is doubly asymptotically commutative, so Theorem \ref{thm:extension2} applies to $G$ via {\bf (H6)}, using the short exact sequence $1\ra \bigoplus _X H \ra H\wr _X K \ra K \ra 1$.
(ii) Let $H$ be a group which is doubly asymptotically commutative. Let $\varphi : H\ra H$ be an injective homomorphism and let $G = \langle t, H \, | \, tht^{-1} = \varphi (h)\rangle$ be the associated ascending HNN-extension. Theorem \ref{thm:extension2} then shows that $G$ is stable, since we have a short exact sequence $1\ra N \ra G \ra \Z \ra 1$ in which the group $N = \bigcup _{i\in \N} t^{-i}Ht^i$ is doubly asymptotically commutative, and hence the hypothesis {\bf (H6)} holds. Similarly, if we instead assume that $H$ is an increasing union $H=\bigcup _m H_m$, where for each $m\in \N$ the pair $(H_m, Z(H_m))$ does not have property (T), then the any ascending HNN extension of $H$ will be stable via {\bf (H3)}.
\end{example}
Notably, Theorem \ref{thm:extension2} also applies to the group $H(\R )$, recently studied by Monod \cite{Mo13}, consisting of all homeomorphisms of the projective line $\bm{\mathrm{P}}^1$ which fix $\infty$ and are piecewise in $\mathrm{PSL}_2(\R )$ with respect to a finite subdivision of $\bm{\mathrm{P}}^1$. It is shown by Monod in \cite{Mo13} that $H(\R )$ does not contain any nonabelian free subgroups, and Theorem 1 of \cite{Mo13} exhibits a family of countable nonamenable subgroups of $H(\R )$. An explicit finitely presented nonamenable subgroup of $H(\R )$ is constructed by Lodha and Moore in \cite{LM13}. We now have the following.
\begin{theorem1}\label{thm:HR}
Every nontrivial countable subgroup of $H(\R )$ is stable.
\end{theorem1}
Since Thompson's group $F$ is a subgroup of $H(\R )$, Theorem \ref{thm:HR} implies:
\begin{corollary1}\label{cor:F}
Thompson's group $F$ is stable. In particular, $F$ and $F\times A$ are measure equivalent, where $A$ is any amenable group.
\end{corollary1}
Corollary \ref{cor:F} yields a new proof of the fact, due to L\"{u}ck \cite{Lu02} and also proved by Bader, Furman, and Sauer in \cite{BFS12}, that all $\ell ^2$-Betti numbers of $F$ vanish. Indeed, Gaboriau has shown that vanishing of $\ell ^2$-Betti numbers is an invariant of measure equivalence \cite{Ga02}, and by \cite{CG86} all $\ell ^2$-Betti numbers of $F\times \Z$ vanish.
\newpage
\setcounter{tocdepth}{3}
\tableofcontents
\section{Weak forms of normality}\label{sec:normality}
In this section we gather some facts about $wq^*$-normality (defined in \S\ref{subsec:intro1}), and we discuss several related normality conditions. We work in the following general setting. Fix an ambient group $G$ along with a nonempty collection $\mathscr{L}$ of subgroups of $G$ which is upward closed in $G$. Let $H\leq M$ be subgroups of $G$. We say that $H$ is {\bf $\mathscr{L}$-$q$-normal} in $M$, denoted $H\leq ^{\mathscr{L}}_{q} M$, if the set $\{ g\in M\csuchthat gHg^{-1}\cap H \in \mathscr{L}\}$ generates $M$. We say that $H$ is {\bf $\mathscr{L}$-$wq$-normal} in $M$, denoted $H\leq ^{\mathscr{L}}_{wq}M$, if there exists an ordinal $\lambda$ and an increasing sequence $(H_\alpha )_{\alpha \leq \lambda}$ of subgroups of $M$, with $H_0=H$ and $H_\lambda = M$, such that $\bigcup _{\beta <\alpha} H_\beta \, \leq ^{\mathscr{L}}_q H_\alpha$ for all $\alpha \leq \lambda$. The notions of $wq$-normality and $wq^*$-normality then correspond to taking $\mathscr{L}$ to be, respectively, the collection $\mathscr{I}$, of infinite subgroups of $G$, and the collection $\mathscr{N}$, of nonamenable subgroups of $G$. Given a $G$-set $X$, we will be interested in the collection
\[
\mathscr{N}^X = \{ H\leq G \csuchthat H\cc X \text{ is nonamenable}\} .
\]
For $S\subseteq G$ finite and $r>0$, the collection $\mathscr{L}^{S,r} = \{ H\leq G\csuchthat \phi _S (G/H ) < r \}$ (where $\phi _S (G/H )$ is defined by \ref{eqn:isop}) will also be of interest, albeit less directly than $\mathscr{N}^X$. Both of these collections are upward closed in $G$ (for $\mathscr{L}^{S,r}$ this follows from Lemma \ref{lem:push} below) and both are invariant under conjugation by $G$.
The following characterization of $\mathscr{L}$-$wq$-normality, along with its proof, is a straightforward extension of \cite[Lemma 5.2]{PT11}.
\begin{lemma}\label{lem:PT} Let $H\leq M$ be subgroups of $G$. Then $H\leq _{wq}^{\mathscr{L}} M$ if and only if for any intermediate proper subgroup $H\leq K\lneq M$ there exists $g\in M \setminus K$ such that $gKg^{-1}\cap K\in \mathscr{L}$.
\end{lemma}
Let $H\leq M$ be subgroups of $G$ and let $\wh{H}$ denote the union of all subgroups $L\leq M$ with $H\leq _{wq}^{\mathscr{L}}L$. We call $\wh{H}$ the {\bf $\mathscr{L}$-$wq$-closure of $H$ in $M$}.
\begin{proposition}\label{prop:wqclosure}
$\wh{H}$ is a subgroup of $M$. Moreover, $\wh{H}$ is the unique subgroup of $M$ satisfying $(i)$ $H\leq _{wq}^{\mathscr{L}} \wh{H}$ and $(ii)$ $g\wh{H}g^{-1} \cap \wh{H} \not\in \mathscr{L}$ for every $g\in M\setminus \wh{H}$.
\end{proposition}
\begin{proof}
By Zorn's Lemma the set $\{ K\leq M \csuchthat H\leq _{wq}^{\mathscr{L}}K \}$ contains a maximal element $L$. By definition, $L\subseteq \wh{H}$. A consequence of Lemma \ref{lem:PT} is that if $L_0,L_1\leq M$ are two subgroups of $M$ with $H\leq _{wq}^{\mathscr{L}} L_0$ and $H\leq _{wq}^{\mathscr{L}}L_1$, then $H\leq _{wq}^{\mathscr{L}} \langle L_0,L_1 \rangle$. It follows that $L=\wh{H}$. Properties $(i)$ and $(ii)$ are immediate. If $K$ is a subgroup of $M$ satisfying properties $(i)$ and $(ii)$ in place of $\wh{H}$, then $H\leq K\leq \wh{H}$ by property $(i)$ for $K$ and the definition of $\wh{H}$, and since $H\leq _{wq}^{\mathscr{L}}\wh{H}$, Lemma \ref{lem:PT} and property $(ii)$ for $K$ imply that $K=\wh{H}$.
\end{proof}
\begin{definition}\label{def:nth}
Let $\mathscr{L}_0=\mathscr{L}$ and for each $n\geq 0$ define $\mathscr{L}_{n+1} = \{ H \leq G \csuchthat H\leq _{wq}^{\mathscr{L}_n}G \}$, which is upward closed by Lemma \ref{lem:PT} and induction. A subgroup $H$ of $G$ is said to be {\bf $n$-degree $\mathscr{L}$-$wq$-normal} in $G$ if $H\in \mathscr{\mathscr{L}}_n$.
\end{definition}
Note that if the collection $\mathscr{L}$ is invariant under conjugation by $G$, then so are each of the collections $\mathscr{L}_n$, $n\in \N$. By applying this definition to the collection $\mathscr{N}^X$, for a $G$-set $X$, we obtain the sequence $\mathscr{N}^X_n$, $n\in \N$. If $H\in \mathscr{N}_n=\mathscr{N}^G_n$, then we say that $H$ is {\bf $n$-degree $wq^*$-normal in $G$}. Thus, $H$ is $0$-degree $wq^*$-normal in $G$ if and only if $H$ is nonamenable, and $H$ is $1$-degree $wq^*$-normal in $G$ if and only if $H$ is $wq^*$-normal in $G$ in the previously defined sense.
For the next proposition, we equip the space of subgroups of $G$ with the subspace topology inherited from the product topology on $2^G$. Note that for any $G$-set $X$, the collection $\mathscr{N}^X$ is an open set, since for a subgroup $H\leq G$, nonamenability of the action $H\cc X$ is witnessed by a finite subset of $H$. The same holds for the collection $\mathscr{L}^{S,r}$, as well as for each $\mathscr{N}^X_n$, $n\in \N$, when $G$ is finitely generated. Taking $\mathscr{L}=\mathscr{N}^X_n$ in the following proposition then shows that, when $G$ is finitely generated, the transfinite sequence in the definition of $\mathscr{L}$-$wq$-normality can be replaced by a finite sequence.
\begin{proposition}\label{prop:Chaubaty}
Let $\mathscr{L}$ be an upward closed collection of subgroups of the countable group $G$. Suppose in addition that $\mathscr{L}$ is open in the space of subgroups of $G$.
\begin{enumerate}
\item[(i)] Let $H$ and $M$ be subgroups of $G$. Assume that $M$ is finitely generated and $H \leq _{wq}^{\mathscr{L}}M$. Then there exist subgroups $H_0,\dots , H_n$ such that
\begin{equation}\label{eqn:nseq}
H=H_0\leq _q^{\mathscr{L}}H_1\leq _q^{\mathscr{L}}\cdots \leq _q^{\mathscr{L}} H_n= M .
\end{equation}
Moreover, for any such sequence $(H_i)_{i=0}^n$ there exists a sequence $(H_i')_{i=0}^n$ with $H_0'\leq _q^{\mathscr{L}}H_1'\leq _q^{\mathscr{L}}\cdots \leq _q^{\mathscr{L}} H_n'= M$, where $H_i'$ is finitely generated and $H_i'\leq H_i$ for all $i$.
\item[(ii)] If $G$ is finitely generated then $\mathscr{L}_n$ is open for all $n\geq 0$.
\end{enumerate}
\end{proposition}
\begin{proof}
For $M,K\leq G$ define $f_M(K) = \langle \{ g\in M\csuchthat gKg^{-1}\cap K \in \mathscr{L} \}\rangle$. Then $f_M$ is monotone and, since $\mathscr{L}$ is open and upward closed, the function $f_M$ is lower semicontinuous, that is, $f_M(\liminf _i K_i ) \leq \liminf _i f_M(K_i)$ for any sequence $(K_i)_{i\in \N}$ of subgroups of $G$, where $\liminf _i K_i$ denotes the subgroup of elements of $G$ which are in cofinitely many $K_i$. It follows that for any finite sequence $M_0, M_1, \cdots , M_n \leq G$, the function $f_{M_n}\circ\cdots \circ f_{M_1}\circ f_{M_0}$ is lower semicontinuous.
(i): Let $H_{\omega} = \bigcup _{n\in \N}f^n_M (H)$. Then semicontinuity of $f_M$ implies $f_M(H_\omega )=H_\omega$. This shows that $H_\omega$ is the $\mathscr{L}$-$wq$-closure of $H$ in $M$, hence $H_\omega = M$. Since $M$ is finitely generated and the sequence $f^n_M(H)$ is nondecreasing there exists an $n$ with $f^n_M(H) = M$. This shows the first part of (i). Fix now any sequence $(H_i)_{i=0}^{n}$ as in \eqref{eqn:nseq}. Then $f_{H_n}\circ \cdots \circ f_{H_1}(H_0) = M$, so there exists a finitely generated $H_0'\leq H_0$ with $f_{H_n}\circ \cdots \circ f_{H_1}(H_0')=M$. Assume now that $k< n-1$ and $f_{H_n}\circ\cdots \circ f_{H_{k+1}}(H_k') = M$. Let $Q_0\subseteq Q_1\subseteq \cdots$ be a sequence of finite sets which exhaust $\{ g\in H_{k+1}\csuchthat gH_k'g^{-1} \cap H_k'\in \mathscr{L}\}$. Then $\bigcup _i \langle H_k', Q_i \rangle = f_{H_{k+1}}(H_k' )$, so there exists some $i$ such that $f_{H_n}\circ \cdots \circ f_{H_{k+2}}(\langle H_k', Q_i \rangle ) = M$. Take $H_{k+1}'=\langle H_k', Q_i\rangle$. The resulting groups $H_0',H_1',\dots ,H_{n-1}',H_n'=M$ satisfy the conclusion of (ii) by construction.
(ii): It suffices to show $\mathscr{L}_1$ is open. This follows from (i) and semicontinuity of $f_G^n$.
\end{proof}
\begin{remark}
In \cite{BFS12}, Bader, Furman, and Sauer define higher order notions of $s$-normality and establish a connection with higher $\ell ^2$-Betti numbers. It seems reasonable to expect a similar connection to hold between higher degree $wq$-normality (or some variant) and higher $\ell ^2$-Betti numbers, although this is largely speculative.
\end{remark}
\section{Amenable actions}\label{sec:isoperim}
Let $X$ be a $G$-set. Let $S\subseteq G$ be finite and let $\epsilon >0$. A nonempty finite subset $P$ of $X$ is said to be {\bf $(S,\epsilon )$-invariant} if $\sum _{s\in S}|sP\setminus P | < \epsilon |P|$. Equivalently, $P$ is $(S,\epsilon )$-invariant if $\sum _{s\in S}|sP\cap P| >(|S|-\epsilon )|P|$.
\begin{remark}\label{rem:bound}
Assume that $S$ generates $G$ and that every $G$-orbit has cardinality greater than $1/\epsilon$. Then any $(S,\epsilon )$-invariant set $P$ has cardinality $|P|>1/\epsilon$. Otherwise we would have $\sum _{s\in S}|sP\setminus P| < 1$, so $P$ would be a $G$-invariant set of cardinality at most $1/\epsilon$, a contradiction.
\end{remark}
\begin{remark}\label{rem:GM07}
We will make use of the observation \cite[Remark 2.12]{GM07} that if $P\subseteq X$ is $(S,\epsilon )$-invariant, then there exists a single $G$-orbit $X_0\subseteq X$ such that $P\cap X_0$ is $(S,\epsilon )$-invariant.
\end{remark}
\subsection{An estimate with F{\o}lner sets}\label{sec:Folner}
For each $n\geq 1$ we let $X^{\circledast n}$ denote the set of all $n$-tuples of distinct points in $X$ which lie in the same $G$-orbit
\[
X^{\circledast n} = \{ (x_0,\dots , x_{n-1}) \in X^n \csuchthat i\neq j \, \Ra \, G x_i = Gx_j \text{ and } x_i\neq x_j \} .
\]
Then we have a natural action $G\cc X^{\circledast n}$ under which the inclusion map $X^{\circledast n} \hookrightarrow X^n$ is a $G$-map to the diagonal product action. For a subset $P\subseteq X$ let $P^{\circledast n} = P^n\cap X^{\circledast n}$.
\begin{lemma}\label{lem:ndiag}
Let $S$ be a finite subset of $G$. Let $n\geq 1$ and let $\epsilon >0$. Let $P\subseteq X$ be an $(S,\epsilon )$-invariant set which is contained in a single $G$-orbit, and assume $|P|\geq n$. Then $P^{\circledast n}$ is $(S, n\epsilon )$-invariant in $X^{\circledast n}$.
\end{lemma}
\begin{proof}
For each $s\in S$ let $\epsilon _s = \tfrac{|sP\setminus P|}{|P|}$.
Then $\sum _{s\in S}\epsilon _s < \epsilon$, so it suffices to show that for all $k\leq n$ we have
\begin{equation}\label{eqn:inter}
|sP^{\circledast k}\cap P^{\circledast k}|\geq |P^{\circledast k}|(1-k\epsilon _s ) .
\end{equation}
If $k=1$ then we have equality, so assume inductively that \eqref{eqn:inter} holds, where $k<n$, and we will show that it holds with $k+1$ in place of $k$. Note that $\epsilon _s >0$ implies $\epsilon _s \geq 1/|P|$. It follows that $(1- \tfrac{|P|\epsilon _s}{|P|-k} )(1-k\epsilon _s )\geq (1-(k+1)\epsilon _s )$, and hence
\begin{align*}
|s P^{\circledast (k+1)}\cap P^{\circledast (k+1)} | &=|(s P\cap P )^{\circledast (k+1)} | \\
&= (|s P \cap P | - k)|(s P\cap P )^{\circledast k}| \\
&\geq (|P|(1- \epsilon _s ) - k)|P^{\circledast k}|(1-k\epsilon _s ) \\
&= (1- \tfrac{|P|\epsilon _s}{|P|-k} )(|P|-k)|P^{\circledast k}|(1-k\epsilon _s ) \\
&\geq |P^{\circledast (k+1)}|(1-(k+1)\epsilon _s ). \qedhere
\end{align*}
\end{proof}
\subsection{Proof of Theorems \ref{thm:wqstar} and \ref{thm:isoperim}}\label{subsec:isoperim} Assume now that $G$ is a finitely generated by $S$. Let $\phi _S(X)$ denote the isoperimetric constant of $X$ with respect to $S$, defined in \eqref{eqn:isop}.
\begin{lemma}\label{lem:push}
Let $X$ and $Y$ be $G$-sets and assume that there exists a $G$-map $\varphi : X\ra Y$ from $X$ to $Y$. Then, given any $P\subseteq X$ which is $(S,\epsilon )$-invariant in $X$, we may find some $Q\subseteq \varphi (P)$ which is $(S,\epsilon )$-invariant in $Y$. In particular $\phi _S (Y)\leq \phi _S(X)$.
\end{lemma}
\begin{proof}
See the first proof in {\S}1.2 of \cite{Gr08}.
\end{proof}
\begin{lemma}\label{lem:peramen}
Let $X$ be a $G$-set and let $H$ be a subgroup of $G$. Assume that the action $H\cc X$ is amenable. Then $\phi _S(G/H ) \geq \phi _S (X)$.
\end{lemma}
\begin{proof}
Let $Y_0$ denote the $G$-set $G/H \times X$ equipped with the diagonal product action of $G$. Then $\phi _S(Y_0)\geq \phi _S (X)$ by Lemma \ref{lem:push}, so it suffices to show that $\phi _S (G/H ) \geq \phi _S (Y_0)$ (so in fact $\phi _S (G/H ) = \phi _S (Y_0)$ by Lemma \ref{lem:push}). Fix a section $\sigma :G/H \ra G$ for the map $G\ra G/H$, and let $\rho :G\times G/H \ra H$ be the corresponding Schreier cocycle given by $\rho (g,kH)= \sigma (gkH)^{-1}g\sigma (kH)$. Let $Y_1$ denote the $G$-set $G/H \times X$ equipped with the action $g\cdot (kH , x) = (gkH, \rho (g,kH ) \cdot x )$, for $kH\in G/H$, $x\in X$ (the $G$-set $Y_1$ is isomorphic to the $G$-set obtained by inducing from the $H$-set $X$, described in $\S$2.C of \cite{GM07}). The map $\varphi : Y_0 \ra Y_1$ given by $\varphi (kH ,x ) = (kH ,\sigma (kH)^{-1}\cdot x )$ provides an isomorphism between the $G$-sets $Y_0$ and $Y_1$, so it suffices to show that $\phi _S(G/H ) \geq \phi _S (Y_1)$. Let $\emptyset\neq P\subseteq G/H$ be finite. Then, for each finite $\emptyset\neq Q\subseteq X$, we have $P\times Q \subseteq Y_1$, so $\phi _S(Y_1)$ is bounded above by
\begin{align*}
\sum _{s\in S}\frac{|s\cdot (P\times Q) \setminus (P\times Q) |}{|P\times Q |} = \frac{1}{|P|}\sum _{s\in S}\sum _{kH \in sP\cap P}\frac{|\rho (s,s^{-1}kH)\cdot Q\setminus Q |}{|Q|} + \sum _{s\in S}\frac{|sP\setminus P|}{|P|} .
\end{align*}
Since $H\cc X$ is amenable, taking the infimum over all such $Q\subseteq X$ shows that $\phi _S(Y_1) \leq \sum _{s\in S} \frac{|sP\setminus P|}{|P|}$, so taking the infimum over $P$ shows that $\phi _S(Y_1)\leq \phi _S(G/H )$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:isoperim}]
The base case $n=0$ is immediate from Lemma \ref{lem:peramen}. Assume now that $\phi _S (G/H )<\frac{1}{2^n}\phi _S (X)$, where $n>0$, and we will show that $H$ is $n$-degree $\mathscr{N}^X$-$wq$-normal in $G$. By Lemma \ref{lem:PT}, given $H\leq L \lneq G$, it suffices to find some $g\in G\setminus L$ such that $gLg^{-1}\cap L$ is $(n-1)$-degree $\mathscr{N}^X$-$wq$-normal in $G$. The quotient map $G/H \ra G/L$ is a $G$-map, so by Lemma \ref{lem:push} we have $\phi _S(G/L)\leq \phi _S(G/H)< \frac{1}{2^n}\phi _S(X)$. Then $\phi _S((G/L )^{ \circledast 2} ) < \frac{1}{2^{n-1}}\phi _S(X)$ by Lemma \ref{lem:ndiag}, so by Remark \ref{rem:GM07} there is some point $x=(g_0L,g_1L)\in (G/L)^{\circledast 2}$ with $\phi _S (G/G_x ) < \frac{1}{2^{n-1}}\phi _S(X)$. By the induction hypothesis, the group $G_x=g_0Lg_0^{-1}\cap g_1Lg_1^{-1}$ is $(n-1)$-degree $\mathscr{N}^X$-$wq$-normal in $G$. Then $g=g_1^{-1}g_0\in G \setminus L$ and $gLg^{-1}\cap L$ is $(n-1)$-degree $\mathscr{N}^X$-$wq$-normal in $G$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:wqstar}]
This follows immediately from Theorem \ref{thm:isoperim} and the observation that for any $\epsilon >0$, the set $\{ x\in X \csuchthat \phi _S(G/G_x)<\epsilon \phi _S(G) \}$ is $\bm{m}$-conull.
\end{proof}
\subsection{An extension to infinitely generated groups} Example \ref{ex:fg} shows that a direct translation of Theorem \ref{thm:wqstar} does not hold in the general infinitely generated setting. However, a refined version of Theorem \ref{thm:wqstar} still holds in general. In what follows, for each $G$-set $X$ let $X_0 = \{ x\in X\csuchthat G_x \text{ is nonamenable}\}$.
\begin{corollary}\label{cor:obstr}
Let $G$ be a nonamenable group. For each $G$-set $X$ there is a $G$-map $\varphi _X : X\ra \wh{X}$ to a $G$-set $\wh{X}$ with the following properties:
\begin{enumerate}
\item[(i)] $G_x$ is $wq^*$-normal in $G_{\varphi _X(x)}$ for all $x\in X_0$;
\item[(ii)] If $\bm{m}$ is any $G$-invariant mean on $X$, then for any finitely generated subgroup $H$ of $G$ we have $H\leq G_{\varphi _X(x)}$ for $\bm{m}$-almost every $x\in X$.
\end{enumerate}
Moreover, this assignment can be made functorial: if $\psi : X\ra Z$ is a $G$-map then there exists a unique $G$-map $\wh{\psi}: \wh{X} \ra \wh{Z}$ with $\varphi _Z\circ \psi = \wh{\psi}\circ \varphi _X$.
\end{corollary}
\begin{proof}
For each nonamenable subgroup $H\leq G$ let $\wh{H}$ denote the $wq^*$-closure of $H$ in $G$ (see \S\ref{sec:normality}). For $x\in X$ let $O(x)$ denote the $G$-orbit of $X$. Let $G$ act on the set $\wh{X}_0 = \{ (\wh{G}_x , O(x)) \csuchthat x\in X_0 \}$ by conjugating the first coordinate, and let $\wh{X} = \wh{X}_0\sqcup X\setminus X_0$. Define $\varphi _X : X\ra \wh{X}$ by $\varphi _X(x) = (\wh{G}_x , O(x))$ for $x\in X_0$, and $\varphi _X(x)=x$ for $x\in X\setminus X_0$. Then $\varphi _X$ is a $G$-map, and for each $x\in X_0$ we have $G_{\varphi _X(x)} = \wh{G}_x$ since $\wh{G}_x$ is self-normalizing. This verifies (i). For (ii), let $\bm{m}$ be a $G$-invariant mean on $X$ and let $H\leq G$ be finitely generated. After making $H$ larger we may assume that $H$ is nonamenable. By Theorem \ref{thm:wqstar}, $H_x$ is $wq^*$-normal in $H$ for $\bm{m}$-almost every $x\in X$. For each such $x$, since $H_x\leq G_x$, we have $H\leq \wh{G}_x = G_{\varphi _X(x)}$.
If $\psi : X \ra Z$ is a $G$-map, then we must show that $\varphi _Z(\psi (x))$ only depends on $\varphi _X(x)$. This is clear for $x\in X\setminus X_0$. Suppose now that $(\wh{G}_{x_0} , O(x_0) ) = (\wh{G}_{x_1},O(x_1))$ where $x_0 , x_1\in X_0$. Find $g\in G$ with $gx_0 = x_1$. Then $\wh{G}_{x_0}=\wh{G}_{x_1} = g\wh{G}_{x_0}g^{-1}$, so $g\in \wh{G}_{x_0} \leq \wh{G}_{\psi (x_0)}$. It follows that $\wh{G}_{\psi (x_1)} = \wh{G}_{\psi (gx_0)} = g\wh{G}_{\psi (x_0)}g^{-1} = \wh{G}_{\psi (x_0)}$, hence $\varphi _Z(\psi (x_0))=\varphi _Z(\psi (x_1))$.
\end{proof}
\section{Transitive amenable actions}
\subsection{Weak normality for groupoids}\label{sec:groupoid}
To prove Theorem \ref{thm:coamen} we need an extension of the results of \cite{PT11} on weakly normal inclusions of discrete probability measure preserving (p.m.p.) groupoids. We adopt the notation and conventions for discrete p.m.p.\ groupoids from \cite[\S 6]{PT11}, and we will need a few additional definitions.
Let $(\mc{G},\mu )$ be a discrete p.m.p.\ groupoid. We do not distinguish two subgroupoids $\mc{H}$ and $\mc{K}$ of $\mc{G}$ if they agree off of a $\mu$-null set. Recall that a {\bf local section} of $\mc{G}$ is a measurable map $\phi : \dom (\phi ) \ra \mc{G}$, with $\dom (\phi ) \subseteq \mc{G}^0$ and $s(\phi x) = x$ for all $x\in \dom (\phi )$, such that the assignment $\phi ^0 : x\mapsto r(\phi x )$ is injective. We do not distinguish two local sections whose domains and values agree off of a $\mu$-null set. Let $[[\mc{G}]]$ denote the collection of all local sections of $\mc{G}$. The {\bf inverse} of $\phi \in [[\mc{G}]]$ is the local section $\phi ^{-1} :\ran (\phi ^0 ) \ra \mc{G}$ given by $\phi ^{-1}(y) = \phi ((\phi ^0)^{-1}y)^{-1}$. The {\bf composition} of two local sections $\phi , \psi \in [[\mc{G}]]$ is the local section $\phi\circ \psi : (\psi ^0)^{-1}(\ran (\psi ^0 ) \cap \dom (\phi ^0 )) \ra \mc{G}$, $x\mapsto \phi (\psi ^0(x))\psi (x)$.
We equip $[[\mc{G}]]$ with the separable complete metric $d(\phi , \psi ) = \mu (\dom (\phi )\triangle \dom (\psi ) ) + \mu (\{ x\in \dom (\phi ) \cap \dom (\psi ) \csuchthat \phi (x) \neq \psi (y) \} )$. A consequence of separability of the metric $d$ is that if $\Phi$ is any subset of $[[\mc{G}]]$ then up to a $\mu$-null set there is a unique smallest subgroupoid $\mc{K}$ of $\mc{G}$ with $\Phi \subseteq [[\mc{K}]]$; we call $\mc{K}$ the {\bf subgroupoid generated by $\Phi$} and denote it by $\langle \Phi \rangle$.
For measurable subsets $R\subseteq \mc{G}$ and $A\subseteq \mc{G}^0$ we let $R_A = \{ \gamma \in R \csuchthat s(\gamma ),r(\gamma )\in A \}$. For $\phi \in [[\mc{G}]]$ and $\gamma \in \mc{G}_{\ran (\phi ^0)}$ let $\gamma ^{\phi }=\phi ^{-1}(r(\gamma ))\gamma \phi ^{-1}(s(\gamma ))^{-1} \in \mc{G}_{\dom (\phi ^0)}$. The {\bf $q$-normalizer} of $R$ in $\mc{G}$ is the set
\[
Q_{\mc{G}}(R) = \{ \phi \in [[\mc{G}]] \csuchthat (R_A)^{\phi } \cap R_{(\phi ^0)^{-1}A} \text{ has infinite measure for all non-null }A\subseteq\ran (\phi ^0) \} .
\]
A subgroupoid $\mc{H}$ of $\mc{G}$ is said to be {\bf $q$-normal} in $\mc{G}$ if $Q_{\mc{G}}(\mc{H})$ generates $\mc{G}$. As usual, we obtain the corresponding notion of $wq$-normality by iterating $q$-normality transfinitely. Then the analogue of Lemma \ref{lem:PT} holds: $\mc{H}$ is $wq$-normal in $\mc{G}$ if and only if for every intermediate proper subgroupoid $\mc{H}\subseteq \mc{K}\subsetneq \mc{G}$ there exists a local section $\phi \in [[\mc{G}]] \setminus [[\mc{K}]]$ with $\phi \in Q_{\mc{G}}(\mc{K})$. While Theorem 6.9 of \cite{PT11} is stated for $ws$-normal subgroupoids, we note that the proof holds more generally for $wq$-normal subgroupoids.
\begin{theorem}[\emph{cf.} {\cite[Theorem 6.9]{PT11}}]\label{thm:restr}
Let $\mc{H}$ be a subgroupoid of the discrete p.m.p.\ groupoid $(\mc{G},\mu )$. If $\mc{H}$ is $wq$-normal in $\mc{G}$ then the restriction map $H^1(\mc{G},\mc{U}(\mc{G},\mu )) \ra H^1(\mc{H},\mc{U}(\mc{G},\mu ))$ is injective.
\end{theorem}
\begin{proof}
The proof of Theorem 6.9 of \cite{PT11} shows that if $c$ is a $\mc{G}$-cocycle with values in $\mc{U}(\mc{G},\mu )$ which vanishes on $[[\mc{H}]]$, then $c$ vanishes on $Q_{\mc{G}}(\mc{H})$, and therefore on $\langle Q_{\mc{G}}(\mc{H})\rangle$ since the set where $c$ vanishes is closed under compositions and inverses, and $c$ respects countable decompositions. The theorem follows. (We note the following minor correction to the proof of Theorem 6.9 of \cite{PT11}: using the notation from that proof, the fact that $\mc{H}_A$ is $s$-normal in $\mc{G}_A$ is irrelevant to the proof; what is being used is that $(\chi _A \psi )^{-1}\mc{H}_A(\chi _A \psi ) \cap \mc{H}_{(\psi ^0)^{-1}A}$ has infinite measure, which holds since $\mc{H}$ is $s$-normal in $\mc{G}$ and hence $\psi \in Q_{\mc{G}}(\mc{H})$. The rest of the proof remains unchanged after replacing $A$ by $(\psi ^0)^{-1}A$ in the appropriate places.)
\end{proof}
\subsection{Recurrence and normality} Let $G\cc (Y,\nu )$ be a free probability measure preserving action of $G$. We let $\mc{R}^G$ denote the orbit equivalence relation generated by this action. Then $(\mc{R}^G , \nu )$ is a discrete p.m.p.\ groupoid so that the notation and terminology of \S\ref{sec:groupoid} applies. In this case, we will identify each local section $\phi \in [[ \mc{R}^G]]$ with the corresponding partial isomorphism $\phi ^0$ of $(Y,\nu )$, and we identify elements of $G$ with their image in $[[\mc{R}^G]]$.
For each subset $P\subseteq G$ let $R^P\subseteq \mc{R}^G$ denote the graph $R^P = \{ (y,k y ) \csuchthat y\in Y, \ k\in P \}$. Define the sets
\begin{align*}
Q(P) &= \{ g\in G\csuchthat (\forall n )(\exists \text{distinct }k_0,\dots , k_{n-1}\in G )\ ( k_i^{-1}k_j \in P\cap P^g \text{ for all }i<j<n ) \} \\
L(P) &= \{ g\in G\csuchthat (\forall n )(\exists \text{distinct }k_0,\dots , k_{n-1}\in G )\ ( k_i^{-1}k_j \in P\cap g^{-1}P \text{ for all }i<j<n ) \} .
\end{align*}
For a subgroup $H\leq G$ we then have $R^H = \mc{R}^H$ and $Q(H) = \{ g\in G\csuchthat H\cap H^g\text{ is infinite} \}$.
\begin{lemma}\label{lem:thick1}
Let $G\cc (Y,\nu )$ be a free probability measure preserving action of $G$. Let $A\subseteq Y$ be measurable and let $P=P^{-1}$ be a subset of $G$.
\begin{enumerate}
\item[(i)] If $g\in Q(P)$ then $g_{|(A\cap g^{-1}A)} \in Q_{\mc{R}^G}(R^P_A)$. In particular, if $H$ is an infinite subgroup of $G$ then $\mc{R}^H_A$ is $q$-normal in the equivalence relation generated by $R^{Q(H)}_A$.
\item[(ii)] Let $d^P_A,\, d^{L(P)}_A:Y\times Y\ra \N \cup \{ \infty\}$ denote the extended graph metrics on $R^P_A$ and $R^{L(P)}_A$ respectively. Then $d^P_A(x,y) \leq 2d^{L(P)}_A(x,y)$ for almost every $(x,y)\in \mc{R}^G$. In particular, if $L(P) =G$ then $R^P_A$ generates $\mc{R}^G_A$.
\end{enumerate}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:thick1}]
(i) Fix $g\in Q(P)$. It suffices to show that for almost every $y\in A\cap g^{-1}A$, the set $\{ z\in A \csuchthat (z,y), (gz,gy)\in R^{P}_A \}$ is infinite. Suppose toward a contradiction that there exists an $m>0$ such that the set
\[
C = \{ y\in A\cap g^{-1}A \csuchthat |\{ z\in A\csuchthat (z,y), (gz,gy)\in R^P_A \} | < m \}
\]
has positive measure, say $\nu (C)=\epsilon >0$. By the Poincar\'{e} recurrence theorem there exists some $n\in \N$, depending only on $\epsilon$ and $m$, such that if $(C_i)_{i<n}$ is any sequence of measurable sets in $Y$, each with $\nu (C_i)\geq \epsilon$, then there exists $i_0<i_1<\cdots <i_{m}<n$ with $\nu (\bigcap _{j<m}C_{i_j}) >0$. Using this $n$, let $(k_i)_{i<n}$ be a sequence as in the definition of $g\in Q(P)$. By our choice of $n$ there exists $i_0<i_1<\cdots <i_{m}<n$ with $\nu (\bigcap _{j\leq m}k_{i_j}C ) > 0$. For each $0\leq j <m$ let $h_j = k_{i_m}^{-1}k_{i_j}$ so that $h_j^{-1} \in P\cap P^g$ and $\nu (C \cap \bigcap _{j<m} h_j C ) >0$, and the elements $h_0,\dots , h_{m-1}$ are pairwise distinct. Fix $y\in C\cap \bigcap _{j<m} h_jC$ and fix any $j<m$ and put $h=h_j$. Then $y, h^{-1}y\in C \subseteq A\cap g^{-1}A$, so $y, h^{-1}y, gy, gh^{-1}y\in A$. Moreover, $h^{-1}\in P$ and $gh^{-1}g^{-1}\in P$, so it follows that $(h^{-1}y, y ) \in R^P_A$ and $(gh^{-1}y, gy ) = ((gh^{-1}g^{-1})gy, gy )\in R^P_A$. This shows that $\{ h_j ^{-1}y \} _{j=0}^{m-1}\subseteq \{ z\in A \csuchthat (z,y), (gz,gy ) \in R^P_A \}$, which contradicts that $y\in C$.
(ii) It suffices to show that $d^P_A(gy, y) \leq 2$ for all $g\in L(P)$ and almost every $y\in A\cap g^{-1}A$. Suppose toward a contradiction that there exists some $g\in L(P)$ such that the set
\[
D = \{ y\in A\cap g^{-1}A \csuchthat d^P_A(gy, y) > 2 \}
\]
has positive measure. Let $n\in \N$ be so large that $\tfrac{1}{n} < \nu (D)$. Let $(k_i)_{i<n}$ be a sequence as in the definition of $g\in L(P)$. Then $\nu (k_i D ) = \nu (D) >\tfrac{1}{n}$ for all $0\leq i<n$, so there exists $i<j<n$ with $\nu (k_i D \cap k_j D) > 0$. Let $k=k_i^{-1}k_j$, so that $k\in P\cap g^{-1}P$, and the set $D_0 := D\cap kD$ is non-null. Fix $y\in D_0$. Then we have $y, gy ,k^{-1}y \in A$ and $k, gk \in P$, so $(k^{-1}y ,gy ) = (k^{-1}y, (gk)k^{-1}y) \in R^P_A$ and $(k^{-1}y,y)= (k^{-1}y,k(k^{-1}y))\in R^P_A$. This shows that $d^P_A(gy , y ) \leq 2$, which contradicts that $y\in D$.
\end{proof}
Example \ref{ex:ws} shows that there are nonamenable groups having a transitive amenable action $G\cc X$ such that $G_x$ is not $ws$-normal in $G$. The next Lemma shows that $G_x$ is still very close to being $s$-normal in $G$. Recall that a subset $B$ of $G$ is said to be {\bf thick} in $G$ if for every finite subset $F\subseteq G$ the intersection $\bigcap _{g\in F}gB$ is nonempty (equivalently: infinite). Observe that if $B\subseteq G$ is thick then $L(B) = G$ since given $g\in G$ we can define $k_0=1$ and inductively let $k_{n+1}$ be any element of $\big( \bigcap _{i\leq n}k_i(B\cap g^{-1}B)\big)\setminus \{k_0,\dots , k_n \}$, so that $(k_n)_{n\geq 0}$ witnesses that $g\in L(B)$.
\begin{lemma}\label{lem:thick}
Let $G\cc X$ be a transitive amenable action of a nonamenable group $G$. Fix any element $x\in X$ and let $H=G_x$. Then $Q(H)$ is thick in $G$. In particular, $L(Q(H)) = G$.
\end{lemma}
\begin{proof} Since $H$ is a subgroup of $G$ we have $Q(H) = \{ g\in G \csuchthat H\cap H^g \text{ is infinite} \}$. Note that $H$ is nonamenable since $G$ is nonamenable and the action $G\cc G/H$ is amenable. Let $\bm{m}$ be a $G$-invariant mean on $G/H$. Then $\bm{m}$ is also $H$-invariant, so we obtain
\begin{equation}\label{eqn:K0}
\bm{m}(\{ gH\in G/H \csuchthat H\cap gHg^{-1} \text{ is nonamenable} \} )=1 .
\end{equation}
Let $\pi : G\ra G/H$ be the projection map $\pi (g)= gH$. Then \eqref{eqn:K0} implies that $\bm{m}(\pi (Q(H)))= 1$. If $g\in G$ then by $G$-invariance of $\bm{m}$ we have $\bm{m} (\pi (gQ(H))) = \bm{m}(g\pi (Q(H))) = \bm{m}(\pi (Q(H))) = 1$. Therefore, for any finite subset $F$ of $G$ we have $\bm{m}(\pi(\bigcap _{g\in F}gQ(H))) = \bm{m}(\bigcap _{g\in F}\pi (gQ(H))) = 1$, where the equality $\pi (\bigcap _{g\in F}gQ(H)) = \bigcap _{g\in F}\pi (gQ(H))$ follows from $Q(H)$ being a union of left cosets of $H$. In particular, $\bigcap _{g\in F}gQ(H)\neq \emptyset$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:coamen}}
Using Theorem \ref{thm:restr} and Lemmas \ref{lem:thick1} and \ref{lem:thick}, we can now argue as in Theorem 5.12 of \cite{PT11}.
\begin{proof}[Proof of Theorem \ref{thm:coamen}]
We can of course assume that $G$ is nonamenable. Let $H= G_x$ and assume that $\beta ^{(2)}_1(H)<\infty$. Since $H$ is infinite index in $G$ there exists a free ergodic p.m.p.\ action $G\cc (Y,\nu )$ of $G$ whose restriction to $H$ has a continuum of ergodic components. Such an action may be obtained, e.g., by coinducing from any free p.m.p.\ action of $H$ with a continuum of ergodic components. Fix $n\geq 1$ and let $A_0,\dots , A_{n-1}$ be a partition of $Y$ into $H$-invariant sets of equal measure. Let $\mc{R}_n = \bigsqcup _{i<n}\mc{R}^G_{A_i}$. Observe that $\mc{R}^H\subseteq \mc{R}_n \subseteq \mc{R}^G$. For each $i<n$ let $\mc{S}_i$ be the equivalence relation generated by $R^{Q(H)}_{A_i}$. Then by Lemma \ref{lem:thick1}.(i), $\mc{R}^H$ is $q$-normal in $\bigsqcup _{i<n}\mc{S}_i$. By Lemma \ref{lem:thick} we have $L(Q(H))=G$, so Lemma \ref{lem:thick1}.(ii) implies that $\mc{S}_i=\mc{R}^G_{A_i}$ for each $i<n$, and hence $\bigsqcup _{i<n}\mc{S}_i = \mc{R}_n$. This shows that $\mc{R}^H$ is $q$-normal in $\mc{R}_n$ and hence
\[
\beta ^{(2)}_1 (\mc{R}_n) \leq \beta ^{(2)}_1 (\mc{R}^H )
\]
by Theorem \ref{thm:restr}. Since $G\cc (Y,\nu )$ is ergodic, each of the sets $A_i$ is a complete section for $\mc{R}^G$, so $[\mc{R}^G:\mc{R}_n] = n$. By Corollary 3.16 and Proposition 5.11 of \cite{Ga02} we therefore have
\[
\beta ^{(2)}_1(G) = \beta ^{(2)}_1(\mc{R}^G) = \frac{1}{n}\beta ^{(2)}_1(\mc{R}_n) \leq \frac{1}{n}\beta ^{(2)}_1 (\mc{R}^H) = \frac{1}{n}\beta ^{(2)}_1(H).
\]
Since $n\geq 1$ was arbitrary and $\beta ^{(2)}_1(H)<\infty$, we conclude that $\beta ^{(2)}_1(G)=0$.
Assume now that $\mathscr{PC}(H)<\infty$ and let $H\cc (Y_0,\nu _0)$ be a free p.m.p.\ action of $H$ with $\mathscr{PC}(\mc{R}^H_{Y_0}) < \infty$. After taking the product of this action with an identity action of $H$ on an atomless probability space, we may assume that $H\cc (Y_0,\nu _0)$ has continuous ergodic decomposition. Let $G\cc (Y,\nu )$ be the coinduced action, which is free and ergodic. Since $H\cc (Y,\nu )$ factors onto $H\cc (Y_0, \nu _0)$, we have $\mathscr{PC}(\mc{R}^H_Y)\leq \mathscr{PC}(\mc{R}^H_{Y_0}) <\infty$. The proof now proceeds as above, using Proposition \ref{prop:Furman1} in place of Theorem \ref{thm:restr}, and Proposition 25.7 of \cite{KM04} in place of Proposition 5.11 of \cite{Ga02} (the proof of Proposition 25.7 of \cite{KM04} works just as well for pseudocost as for cost).
\end{proof}
\section{The cost of inner amenable groups}
\subsection{Proof of Theorem \ref{thm:relmain1}}
We will often use the following well-known classical fact, which is a weakening of Theorem \ref{thm:wqstar}.
\begin{lemma}\label{lem:null}
Let $G\cc X$ be an amenable action of a nonamenable group $G$ and let $\bm{m}$ be a $G$-invariant mean on $X$. Then $G_x$ is nonamenable for $\bm{m}$-almost every $x\in X$.
\end{lemma}
The following simple consequence will be very useful.
\begin{lemma}\label{lem:conull}
Assume that the pair $(G,H)$ is inner amenable and let $\bm{m}$ be an $H$-conjugation invariant mean on $G$. Let $\mathscr{F}$ be a finite collection of nonamenable subgroups of $H$. Then
\[
\bm{m} ( \{ g\in G\csuchthat L\cap C_G(g) \text{ is nonamenable for all }L\in \mathscr{F} \} ) = 1 .
\]
\end{lemma}
\begin{proof}
It suffices to prove the lemma in the case where $\Es{F}= \{ L \}$ is a singleton. This follows from Lemma \ref{lem:null} by taking $X=G$ along with the conjugation action $L\cc X$.
\end{proof}
Theorem \ref{thm:relmain1} is an immediate consequence of the following more detailed analysis.
\begin{theorem}\label{thm:relmain}
Assume that the pair $(G,H)$ is inner amenable and that $H$ is nonamenable. For each nonamenable subgroup $L\leq H$ let $K_L =\langle \{ g\in G\csuchthat L\cap C_G(g)\text{ is nonamenable} \}\rangle$.
\begin{enumerate}
\item[(i)] Let $L$ be a nonamenable subgroup of $H$. Then $L\leq _{q^*}LK_L\leq _q \langle H,K_L\rangle \leq _{q^*}HK_H$. In particular, $L$ is $wq$-normal in $HK_{H}$, and $H$ is $q^*$-normal in $HK_H$.
\item[(ii)] Every $H$-conjugation invariant mean $\bm{m}$ on $G$ concentrates on $HK_H$. In particular, $\bm{m}$ concentrates on the $wq^*$-closure of $H$ in $G$.
\end{enumerate}
\end{theorem}
\begin{proof} Fix an atomless $H$-conjugation invariant mean $\bm{m}$ on $G$. Let $L\leq H$ be nonamenable. Let $S_L = \{ g\in G\csuchthat L\cap C_G(g)\text{ is nonamenable} \}$. Then $\bm{m}(S_L)=1$ by Lemma \ref{lem:conull}, and $S_L\subseteq \{ g\in G\csuchthat gLg^{-1}\cap L\text{ is nonamenable}\}$ implies that $L\leq _{q^*} LK_L$. Since $S_L\subseteq K_L$ we have $\bm{m}(LK_L) =1$, so $\bm{m}(hLK_Lh^{-1}\cap LK_L ) =1$ for all $h\in H$, and since $\bm{m}$ is atomless this shows that $LK_L\leq _q \langle H, K_L\rangle$. Then $\langle H,K_L\rangle \leq _{q^*}HK_H$ follows from $H\leq _{q^*}HK_H$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:subgroup}}
We begin with the version of Theorem \ref{thm:subgroup} for inner amenable pairs. Let $\mathscr{N}(H)$ denote the collection of all nonamenable subgroups of $H$.
\begin{theorem}\label{thm:pairsubgroup}
Assume that the pair $(G,H)$ is inner amenable and that $H$ is nonamenable. Then at least one of the following holds:
\begin{enumerate}
\item[(1)] For every finite $\mathscr{F}\subseteq \mathscr{N}(H)$ there exists an infinite amenable subgroup $K$ of $G$ such that $L\cap C_G(K)$ is nonamenable for all $L\in \mathscr{F}$.
\item[(2)] For every finite $\mathscr{F}\subseteq \mathscr{N}(H)$ there exists an increasing sequence $M_0\leq M_1\leq \cdots$ of finite subgroups of $G$, with $\lim _{n\ra\infty}|M_n|=\infty$, such that $L\cap C_G(M_n)$ is nonamenable for all $L\in \mathscr{F}$, $n\in \N$.
\item[(3)] For every finite $\mathscr{F}\subseteq \mathscr{N}(H)$ there exists a nonamenable subgroup $K$ of $G$ such that $L\cap C_G(K)$ is nonamenable for all $L\in \mathscr{F}$.
\end{enumerate}
\end{theorem}
\begin{proof}
Assume that neither (1) nor (2) holds, as witnessed by the collections $\mathscr{F}_1$ and $\mathscr{F}_2$ respectively. We will show that (3) holds. Toward this end, fix $\mathscr{F} \subseteq \mathscr{N}(H)$ finite. We may assume that $\mathscr{F}_1\cup \mathscr{F}_2\subseteq \mathscr{F}$. Fix an atomless $H$-conjugation invariant mean $\bm{m}$ on $G$. Let $M_0 = \{ e \}$ and for each $L\in \mathscr{F}$ let $\varphi _0(L) = L$. Assume for induction that $\varphi _0(L)\geq \cdots \geq \varphi _n(L)$, $L\in \mathscr{F}$, and $M_0\lneq \cdots \lneq M_n$ have been defined with $\varphi _n (L)$ nonamenable and commuting with $M_n$ for all $L\in \mathscr{F}$. If $M_n$ is infinite then we stop; otherwise, since $\bm{m}$ is atomless we have $\bm{m}(M_n)=0$, so by Lemma \ref{lem:conull} there exists $g\in G\setminus M_n$ such that $\varphi _{n+1}(L):= \varphi _n(L)\cap C_G(g)$ is nonamenable for all $L\in \mathscr{F}$. The induction continues with $M_{n+1}:= \langle M_n,g\rangle$.
We claim that this process stops at some stage, i.e., there is some $n>0$ such that $M_n$ is infinite. Otherwise, if the process never stops, we would obtain an infinite sequence $M_0\lneq M_1\lneq \cdots$ of finite groups such that $\varphi _n(L)\leq L\cap C_G(M_n)$ for all $L\in \mathscr{F}_2$, contradicting our choice of $\mathscr{F}_2$. Let $n$ be the stage at which the process stops and take $K:=M_n$. Then for each $L\in \mathscr{F}$ we have $\varphi _n (L)\leq L\cap C_G(K)$, so $L\cap C_G(K)$ is nonamenable. Then $K$ must be nonamenable since $\mathscr{F}_1\subseteq \mathscr{F}$.
\end{proof}
We can now prove Theorem \ref{thm:subgroup}.
\begin{proof}[Proof of Theorem \ref{thm:subgroup}]
Assume that neither (1) nor (2) of Theorem \ref{thm:subgroup} holds and fix $\mathscr{F}\subseteq \mathscr{N}$ finite and $n\in \N$ toward the goal of verifying (3). We already know that the pair $(G,G)$ satisfies alternative (3) of Theorem \ref{thm:pairsubgroup}. We may therefore find a nonamenable subgroup $K_0^0\leq G$ such that the group $\psi _0(L):= L\cap C_G(K^0_0)$ is nonamenable for all $L\in \mathscr{F}$. Let $k\geq 0$ and assume for induction that we have defined the nonamenable subgroups $\psi _k(L)$, $L\in \mathscr{F}$, and nonamenable $K^k_0,K^k_1,\dots , K^k_k$, such that
\begin{itemize}
\item $K_i^k$ and $K_j^k$ commute for all $0\leq i<j\leq k$,
\item $\psi _k(L)\leq L$ and $\psi _k(L)$ and $K_i$ commute for all $L\in \mathscr{F}$ and $0\leq i\leq k$.
\end{itemize}
Now apply alternative (3) of Theorem \ref{thm:pairsubgroup} to $(G,G)$, using the finite collection $\{ \psi _k (L)\csuchthat L\in \mathscr{F}\}\cup \{ K^k_i\csuchthat 0\leq i\leq k \}$, to obtain a nonamenable subgroup $K^{k+1}_{k+1}$ of $G$ such that $\psi _{k+1}(L):=\psi _k(L)\cap C_G(K^{k+1}_{k+1})$ is nonamenable for all $L\in \mathscr{F}$, and $K^{k+1}_i := K^k_i\cap C_G(K^{k+1}_{k+1})$ is nonamenable for all $i\leq k$. This continues the induction. For each $i<n$ let $K_i = K^{n-1}_i$. Then $K_0,K_1, \dots , K_{n-1}$ are the desired subgroups for the first part of (3). For the second statement, we may assume that $\mathscr{F}$ contains a subset $\mathscr{F}_1$ witnessing that alternative (1) fails. For each $0\leq i<n$ inductively let $g_i$ be any element of $K_i\setminus \langle g_0,\dots , g_{i-1}\rangle$; this set is nonempty since $K_i$ is nonamenable and $\langle g_0,\dots , g_{i-1}\rangle$ is abelian. The group $M_n := \langle g_0,g_1,\dots ,g_{n-1}\rangle$ is an abelian group which commutes with $\psi (L)$ for all $L\in \mathscr{F}$, hence $L\cap C_G(M_n)$ is nonamenable for all $L\in \mathscr{F}$. Then $M_n$ is finite since $\mathscr{F}_1\subseteq \mathscr{F}$. By construction, $M_n$ has order $|M_n|\geq 2^n$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:FP1}}\label{sec:FP1}
\begin{lemma}\label{lem:fin}
Let $M$ be a finite normal subgroup of $G$. Then $|M|(\mathscr{C}^*(G) -1)\leq \mathscr{C}^*(G/M) - 1$.
\end{lemma}
\begin{proof}
Let $G \cc (X,\mu )$ be a free p.m.p.\ action of $G$. Let $Y\subseteq X$ be a measurable transversal for $\mc{R}^M_X$, so $\mu (Y) =\tfrac{1}{|M|}$. Let $\mu _Y$ be the normalized restriction of $\mu$ to $Y$. Then $G/M$ acts on $(Y,\mu _Y)$ by the rule $gM\cdot y_0 = y_1$ if and only if $gMy_0 = My_1$. This is an action since $M$ is normal in $G$, and it is free and measure preserving. Fix $\epsilon >0$ and let $R$ be a graphing of $\mc{R}^{G/M}_Y$ with $\mathscr{C}_{\mu _Y} (R)<\mathscr{C}^*(G/M)+\epsilon$. Let $T$ be a treeing of $\mc{R}^M_X$. Then $\mathscr{C}_\mu (T)= 1-\tfrac{1}{|M|}$ and $T\cup R$ is a graphing of $\mc{R}^G_X$, hence
\begin{align*}
\mathscr{C}_\mu (\mc{R}^G_X)&\leq \mathscr{C}_\mu (T\cup R) = \mathscr{C}_\mu (T) + \mu (Y)\mathscr{C}_{\mu _Y} (R)\\
&< 1-\frac{1}{|M|} + \frac{1}{|M|}(\mathscr{C}^*(G/M) + \epsilon ) = 1 + \frac{\mathscr{C}^*(G/M) -1}{|M|} + \epsilon /|M| .
\end{align*}
Since $\epsilon >0$ was arbitrary this shows that $\mathscr{C}_\mu (\mc{R}^G_X)\leq 1 + \frac{\mathscr{C}^*(G/M) -1}{|M|}$. Since this holds for all free p.m.p.\ actions $G\cc (X,\mu )$ the proof is complete.
\end{proof}
\begin{lemma}\label{lem:cost2}
Let $G$ be a nonamenable group.
\begin{enumerate}
\item There exists a finitely generated nonamenable subgroup $H\leq G$ with $\mathscr{C}^*(H) \leq 2$.
\item Let $M$ be a finite normal subgroup of $G$. Then there exists a finitely generated nonamenable subgroup $H\leq G$ containing $M$ with $\mathscr{C}^*(H)\leq 1+1/|M|$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1): Let $F$ be a finite subset of $G$ which is minimal (under inclusion) with respect to the property that $\langle F\rangle$ is nonamenable. Take $H= \langle F\rangle$. Let $H\cc (X,\mu )$ be a free p.m.p.\ action of $H$. By minimality of $F$, for any $g\in F$ the group $K= \langle F\setminus \{ g \} \rangle$ is amenable, hence $\mathscr{C}_\mu (\mc{R}^H_X)\leq \mathscr{C}_\mu (\mc{R}^K_X)+\mathscr{C}_\mu (\mc{R}^{\langle g\rangle}_X)\leq 2$.
(2): By applying part (1) to $G/M$ we may find a finitely generated nonamenable subgroup $H\leq G$ containing $M$ such that $\mathscr{C}^*(H/M ) \leq 2$. Then by Lemma \ref{lem:fin} we have
\[
\mathscr{C}^*(H) \leq 1+ \frac{\mathscr{C}^*(H /M) - 1}{|M|} \leq 1+ \frac{1}{|M|} . \qedhere
\]
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:FP1}]
(1): If $H$ is amenable then $\mathscr{C}^*(H)=1$, so $\mathscr{PC}^*(H) =1$, hence $\mathscr{PC}^*(G)=1$ by Proposition \ref{prop:Furman2} and thus $\mathscr{C}^*(G)=1$, so we are done. Assume now that $H$ is nonamenable.
Fix an $H$-conjugation invariant mean $\bm{m}$ on $G$. If $L$ is any subgroup of $G$ such that $L\cap H$ is nonamenable, then Theorem \ref{thm:relmain} implies that $(L\cap H )\leq _{wq} HK_H\leq _{wq}G$, so $L\leq _{wq}G$, and hence $\mathscr{PC}^*(G)\leq \mathscr{PC}^*(L)$ by Proposition \ref{prop:Furman2}. This shows that
\begin{equation}\label{eqn:inf}
\mathscr{PC}^*(G) = \inf \{ \mathscr{PC}^*(L)\csuchthat L\leq G\text{ and }L\cap H\text{ is nonamenable}\} .
\end{equation}
Apply Theorem \ref{thm:pairsubgroup} and take $\mathscr{F}=\{ H\}$. If alternative (1) holds then the subgroup $L= (H\cap C_G(K))K$ has fixed price $1$, and $L\cap H$ is nonamenable, so $\mathscr{PC}^*(G)=1$ by \eqref{eqn:inf}, and hence $\mathscr{C}^*(G)=1$. If alternative (2) holds then by Lemma \ref{lem:cost2} we may find a sequence $(L_n)_{n\in \N}$ of nonamenable subgroups $L_n\leq (H\cap C_G(M_n))M_n$ with $\mathscr{PC}^*(L_n) \leq 1+1/|M_n| \longrightarrow 1$ as $n\ra \infty$. Since $L_n\cap H$ is nonamenable for all $n\in \N$ we conlude once again that $\mathscr{PC}^*(G)=1$ and hence $\mathscr{C}^*(G)=1$. Finally, suppose that alternative (3) holds and let $L=(H\cap C_G(K))K$. Then $\mathscr{C}(L)=1$ since $H\cap C_G(K)$ commutes with $K$ and both groups are infinite \cite{Ga00}. Since $L\leq _{wq}G$ it follows from Proposition \ref{prop:Furman2} that $\mathscr{PC}(G)=1$ and hence $\mathscr{C}(G)=1$.
(2): We may assume that $G$ is nonamenable, and by the proof of part (1) we may assume that alternative (3) of Theorem \ref{thm:subgroup} holds. Then using the sequence $(M_n)_{n\in \N}$ we obtain that $\mathscr{C}^*(G)=1$ as in the proof for alternative (2) above.
\end{proof}
\section{Cocycle superrigidity}\label{sec:superrigid}
\subsection{Amenable actions and stable spectral gap}
Fix a finite subset $S$ of $G$. Given a unitary representation $\pi$ of $G$ on the Hilbert space $\Es{H}$, we define
\begin{align*}
\phi _S (\pi ) = \inf \Big\{ \sum _{s\in S}\frac{\| \pi _s \xi - \xi \|}{\| \xi \| } \csuchthat 0\neq \xi \in \Es{H} \Big\} .
\end{align*}
The representation $\pi$ is said to have {\bf stable spectral gap} if $1\not\prec \pi\otimes \ol{\pi}$, where $\prec$ denotes weak containment of unitary representations. If $S$ generates $G$, then the representation $\pi$ has stable spectral gap if and only if $\phi _S (\pi \otimes \ol{\pi} ) >0$.
\begin{lemma}\label{lem:square}
Let $\kappa$ and $\pi$ be unitary representation of $G$.
\begin{enumerate}
\item If $\kappa \prec \pi$ then $\phi _S (\pi ) \leq \phi _S (\kappa )$.
\item $\phi _S (\pi \otimes \kappa )\geq \tfrac{1}{2}\phi _S (\pi \otimes \ol{\pi} )^2$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) is clear, and (2) follows from the Powers-St{\o}rmer inequality (see \cite[Lemma 3.2]{Pop08}).
\end{proof}
\begin{lemma}\label{lem:ssg}
Assume that $S$ generates $G$. Let $\pi$ be a unitary representation of $G$ with stable spectral gap. Let $H$ be a subgroup of $G$ and let $\lambda _{G/H}$ denotes the quasi-regular representation of $G$ on $\ell ^2(G/H )$. If $\phi _S (\lambda _{G/H})< \tfrac{1}{2}\phi _S (\pi \otimes \ol{\pi} )^2$ then $\pi _{|H}$ has stable spectral gap.
It follows that if $G\cc X$ is an amenable action of a countable group $G$ on a set $X$ with $G$-invariant mean $\bm{m}$, and if $\pi$ is a representation of $G$ with stable spectral gap, then $\pi _{|G_x}$ has stable spectral gap for $\bm{m}$-almost every $x\in X$.
\end{lemma}
\begin{proof}
We prove the contrapositive. Suppose that $\pi _{|H}$ does not have stable spectral gap, i.e., $1_{|H}\prec (\pi \otimes \ol{\pi} )_{|H}$. Then $\lambda _{G/H}\prec \mathrm{Ind}_H^G ((\pi \otimes \ol{\pi}) _{|H}) \cong \pi \otimes \ol{\pi} \otimes \lambda _{G/H}$. Applying (1) and then (2) of Lemma \ref{lem:square}, we obtain $\phi _S(\lambda _{G/H}) \geq \phi _S(\pi \otimes (\ol{\pi} \otimes \lambda _{G/H} ) ) \geq \tfrac{1}{2}\phi _S (\pi \otimes \ol{\pi} )^2$.
For the last statement, note that it is enough to prove this in the case where $G$ is finitely generated, say by $S$. Then for any $\epsilon >0$ we have $\bm{m} (\{ x\in X \csuchthat \phi _S (\lambda _{G/G_x}) <\epsilon \} ) =1$, so we can take $\epsilon = \tfrac{1}{2}\phi _S (\pi \otimes \ol{\pi} )^2$ and apply the first part of the lemma.
\end{proof}
\subsection{The space of means}\label{subsec:means}
Let $\mathscr{M}=\mathscr{M}(G)$ denote the space of means on $G$. Then $\mathscr{M}$ is a weak${}^*$-closed subset of the unit ball of $\ell ^\infty (G)^*$, hence by the Banach-Alaoglu Theorem $\mathscr{M}$ is compact in the weak${}^*$-topology. Let $\Es{P} = \mathscr{M}\cap \ell ^1(G)$ denote the collection of all probability vectors on $G$. Then $\Es{P}$ is a weak${}^*$-dense subset of $\mathscr{M}$.
For $g\in G$ and $\bm{m}\in \mathscr{M}$ we define the means $g\bm{m}$ and $\bm{m}g$ respectively by $(g\bm{m})(A) = \bm{m}(g^{-1}A)$ and $(\bm{m}g)(A) = \bm{m}(Ag^{-1})$ for $A\subseteq G$. The assignments $(g,\bm{m})\mapsto g\bm{m}$ and $(\bm{m},g)\mapsto \bm{m}g$ define left and right actions respectively of $G$ on $\mathscr{M}$ which commute. The {\bf convolution} of two means $\bm{m}$ and $\bm{n}$ on $G$ is defined to be the mean $\bm{m}\ast \bm{n} = \int _{g\in G} g\bm{n}\, d\bm{m}(g)$. This gives $\mathscr{M}$ the structure of a semigroup in which multiplication is weak${}^*$-continuous in the left variable. We identify $G$ with the collection of point masses $\{ \delta _g \} _{g\in G}\subseteq \Es{P}$. Then we have $g\bm{m}= \delta _g\ast \bm{m}$ and $\bm{m}g=\bm{m} \ast \delta _g$.
For a subgroup $H\leq G$, let $C_{\mathscr{M}}(H)= C_{\mathscr{M}(G)}(H)$ denote the subset of $\mathscr{M}$ consisting of all means on $G$ which are invariant under conjugation by $H$. Observe that $C_{\mathscr{M}}(H)$ is a closed, convex, subsemigroup of $\mathscr{M}$ with $C_G(H) \subseteq C_{\mathscr{M}}(H)$.
\begin{lemma}\label{lem:Mazur}
Let $H$ be a subgroup of $G$, let $\bm{m}\in C_{\mathscr{M}}(H)$, and let $\mathscr{A}\subseteq \ell ^\infty (G)$ be a separable subalgebra. Then there exists a sequence $(p_n)_{n\in \N}$ in $\Es{P}$ such that $\lim _n \| hp_nh^{-1} - p_n \| _1 = 0$ for all $h\in H$, and $\lim _n p_n (\phi ) = \bm{m}(\phi )$ for all $\phi \in \mathscr{A}$.
\end{lemma}
\begin{proof}
Fix finite sets $S\subseteq H$ and $\mathscr{A}_0\subseteq \mathscr{A}$, along with $\epsilon >0$. It suffices to show that there exists some $p\in \Es{P}$ with $\sup _{s\in S} \| sps^{-1} - p \| _1 <\epsilon$ and $\sup _{\phi \in \mathscr{A}_0} |p(\phi ) - \bm{m}(\phi ) | <\epsilon$. Since $\Es{P}$ is weak${}^*$-dense in $\mathscr{M}$, the convex set $K_0 = \{ p \in \Es{P} \csuchthat \sup _{\phi \in \mathscr{A}_0} | p (\phi ) - \bm{m}(\phi ) | <\epsilon \}$ contains $\bm{m}$ in its weak${}^*$-closure. Since $\bm{m}\in C_{\mathscr{M}}(H)$, the convex subset $\{ (sps^{-1} - p)_{s\in S} \csuchthat p \in K_0 \}$ of $\ell ^1 (G)^S$ contains $0\in \ell ^1(G)^S$ in its weak closure, hence in its norm closure by Mazur's Theorem. This implies that there exists $p\in K_0$ such that $\sup _{s\in S} \| sps^{-1} - p \| _1 <\epsilon$.
\end{proof}
\subsection{Weak mixing for subsemigroups of $\mathscr{M}$}\label{subsec:WM} Let $\Es{H}$ be a Hilbert space and let $\varphi : G \ra \mathscr{B}(\Es{H})$, $g\mapsto \varphi _g$, be a map from $G$ into the bounded linear operators on $\Es{H}$ whose image is contained in the unit ball of $\mathscr{B}(\Es{H})$. We extend $\varphi$ to a map $\mathscr{M}\ra \mathscr{B}(\Es{H})$, by taking $\varphi _{\bm{m}} = \int _G \varphi _g \, d\bm{m}$, i.e., $\varphi _{\bm{m}}$ is the unique bounded linear operator satisfying $\langle \varphi _{\bm{m}}\xi ,\eta \rangle = \int _{g\in G}\langle \varphi _g\xi , \eta \rangle \, d\bm{m}(g)$ for all $\xi ,\eta \in \Es{H}$. In particular, each unitary representation $\pi : G\ra \mathscr{U}(\Es{H})$ of $G$ extends to a map $\mathscr{M}\ra \mathscr{B}(\Es{H})$ by taking $\pi _{\bm{m}} = \int _G \pi _g \, d\bm{m}$ for each $\bm{m}\in \mathscr{M}$.
\begin{proposition}\label{prop:contract} Let $\pi : G\ra \mathscr{U}(\Es{H})$ be a unitary representation of $G$. Then the extended map $\pi : \mathscr{M}\ra \mathscr{B}(\Es{H})$ has the following properties:
\begin{enumerate}
\item[i.] $\pi$ is an affine semigroup homomorphism.
\item[ii.] For each $\bm{m}\in \mathscr{M}$ we have $\pi _{\bm{m}}^* = \pi _{\check{\bm{m}}}$, where $\check{\bm{m}} (A) = \bm{m}(A^{-1})$ for $A\subseteq G$.
\item[iii.] For each $\bm{m}\in \mathscr{M}$ the operator $\pi _{\bm{m}}$ is a contraction, i.e., $\| \pi _{\bm{m}}\| _{\infty}\leq 1$.
\item[iv.] $\pi$ is continuous when $\mathscr{M}$ is given the weak${}^*$-topology and when $\mathscr{B}(\Es{H})$ is given the weak operator topology.
\item[v.] $\{ \pi _{\bm{m}} \} _{\bm{m}\in \mathscr{M}} \subseteq W^*(\pi (G))$.
\end{enumerate}
\end{proposition}
\begin{proof}
Properties i.\ through iv.\ follow from the definitions, and v.\ follows from iv.
\end{proof}
If $\pi$ and $\kappa$ are unitary representations of $G$ on $\Es{H}$ and $\Es{K}$ respectively, then for $\bm{m}\in \mathscr{M}$, the operators $(\pi \otimes \kappa )_{\bm{m}} = \int _G \pi _g\otimes \kappa _g \, d\bm{m}$ and $\pi _{\bm{m}}\otimes \kappa _{\bm{m}}$ are generally distinct. We will only make use of the operator $(\pi \otimes \kappa )_{\bm{m}}$.
\begin{definition}
Let $\pi$ be a unitary representation of $G$ and let $\mathscr{M}_0$ be a subsemigroup of $\mathscr{M}$. We say that $\pi _{| \mathscr{M}_0}$ is {\bf weakly mixing} if $(\pi \otimes \bar{\pi})_{|\mathscr{M}_0}$ has no nonzero invariant vectors.
\end{definition}
The next proposition extends several well-known characterizations of weak mixing from the group setting to the setting of subsemigroups of $\mathscr{M}$.
\begin{proposition}\label{prop:WM}
Let $\pi :G\ra \mathscr{U}(\Es{H})$ be a unitary representation of $G$ and let $\mathscr{M}_0$ be a subsemigroup of $\mathscr{M}$. Then the following are equivalent:
\begin{enumerate}
\item[i.] $\pi _{|\mathscr{M}_0}$ is weakly mixing.
\item[ii.] $(\pi \otimes \kappa )_{|\mathscr{M}_0}$ has no nonzero invariant vectors for every unitary representation $\kappa$ of $G$.
\item[iii.] For every finite $F\subseteq \Es{H}$ and $\epsilon >0$ there exists $\bm{m}\in \mathscr{M}_0$ such that
\[
\int _{g\in G}|\langle \pi _g\xi , \eta \rangle | ^2 d\bm{m}(g) <\epsilon
\]
for all $\xi ,\eta \in F$.
\item[iv.] There is no nonzero finite dimensional subspace $\Es{L}$ of $\Es{H}$ such that $P_{\Es{L}} = \int _G \pi _g P_{\Es{L}}\pi _g^* \, d\bm{m}$ for all $\bm{m}\in \mathscr{M}_0$. Here, $P_{\Es{L}}$ denotes the orthogonal projection onto $\Es{L}$.
\end{enumerate}
\end{proposition}
\begin{proof}
Using the properties in Proposition \ref{prop:contract}, the proof is a routine extension of the proof for the case $\mathscr{M}_0=G$ (see, e.g., \cite{Pe11}).
\end{proof}
\begin{example}
Let $\lambda :G \ra \mathscr{U}(\ell ^2(G))$ be the left regular representation of $G$. Then $\lambda$ is a mixing representation of $G$, so if $\bm{m}$ is any atomless mean on $G$ then $\lambda _{\bm{m}} = 0$ in $\mathscr{B}(\ell ^2(G))$. It follows that if $\mathscr{M}_0$ is a subsemigroup of $\mathscr{M}$ whose weak${}^*$-closure contains a mean which is atomless, then $\lambda _{|\mathscr{M}_0}$ is weakly mixing.
\end{example}
The next proposition shows that weak mixing for the Koopman representation associated to a p.m.p.\ action of $G$ behaves as expected.
\begin{proposition}\label{prop:ERG} Let $G\cc (X,\mu )$ be a p.m.p.\ action of $G$ and let $\kappa$ denote the associated Koopman representation on $L^2(X,\mu )$. Let $\mathscr{M}_0$ be a subsemigroup of $\mathscr{M}$. Then the collection $\{ A\subseteq X \csuchthat \kappa _{\bm{m}}(1_A)=1_A \text{ for all }\bm{m}\in \mathscr{M}_0 \}$ is a $\| \cdot \| _2$-norm closed sigma subalgebra of the measure algebra of $(X,\mu )$. Furthermore, a function $\xi \in L^2(X,\mu )$ is $\kappa _{|\mathscr{M}_0}$-invariant if and only if $1_A$ is $\kappa _{|\mathscr{M}_0}$-invariant for every $\xi$-measurable set $A\subseteq X$.
\end{proposition}
\begin{proof}
A function $\xi \in L^2(X,\mu )$ is $\kappa _{\bm{m}}$-invariant if and only if $\int _G \| \kappa _g (\xi ) - \xi \| _2 ^2 \, d\bm{m} = 0$. Therefore, if $f_0,f_1\in L^\infty (X,\mu )$ are both $\kappa _{\bm{m}}$-invariant, then
\[
\int _G\| \kappa _g(f_0f_1)- f_0f_1\| _2 ^2 \, d\bm{m} \leq \| f_1 \|_\infty ^2 \int _G \| \kappa _g (f_0) - f_0 \| _2^2 \, d\bm{m} + \| f_0 \| _{\infty} ^2 \int _G \| \kappa _g (f_1)- f_1 \| _2^2 \, d\bm{m} = 0,
\]
hence $f_0f_1$ is also $\kappa _{\bm{m}}$-invariant.
Assume that $\xi \in L^2(X,\mu )$ is $\kappa _{|\mathscr{M}_0}$-invariant. It suffices to show that sets of the form $A_r = \{ x\in X\csuchthat \xi (x)\geq r \}$, $r\in \R$, are $\kappa _{|\mathscr{M}_0}$-invariant. Suppose toward a contradiction that $A_r$ is not $\kappa _{\bm{m}}$-invariant for some $r\in \R$ and $\bm{m}\in \mathscr{M}_0$. Then we have $\int _G \mu (A_r \setminus gA_r ) \, d\bm{m} > 0$, so there is some $\epsilon >0$ such that $\bm{m}(D_{\epsilon})>0$ where $D_\epsilon = \{ g\in G\csuchthat \mu (A_r\setminus gA_r ) > \epsilon \}$. Find $\delta >0$ such that $\mu (A_{r-\delta }\setminus A_r ) = \mu ( \{ x\in X\csuchthat r > \xi (x) \geq r-\delta \} ) < \epsilon /2$. Then for $g\in D_\epsilon$ we have $\mu (A_r \setminus gA_{r-\delta} ) > \epsilon /2$ and hence $\| \xi - \kappa _g(\xi ) \| _2^2 \geq \delta \epsilon /2$. Therefore, $0 = \int _{g\in D_{\epsilon}} \| \xi - \kappa _g(\xi ) \| _2 ^2 \, d\bm{m} \geq \bm{m}(D_\epsilon )\delta \epsilon /2 > 0$, a contradiction. For the reverse implication, approximate $\xi$ in $\| \cdot \| _2$-norm by $\xi$-measurable simple functions.
\end{proof}
\begin{definition}\label{def:WM}
Let $G\cc ^\sigma (X,\mu )$ be a p.m.p.\ action of $G$ and let $\kappa$ denote the associated Koopman representation on $L^2(X,\mu )$. Let $\mathscr{M}_0$ be a subsemigroup of $\mathscr{M}$. We say that $\sigma _{|\mathscr{M}_0}$ is {\bf ergodic} if every $\kappa _{|\mathscr{M}_0}$-invariant function in $L^2(X,\mu )$ is essentially constant. We say that $\sigma _{|\mathscr{M}_0}$ is {\bf weakly mixing} if $(\sigma \otimes \sigma )_{|\mathscr{M}_0}$ is ergodic.
\end{definition}
\begin{proposition}\label{prop:WMaction}
Let $G\cc ^\sigma (X,\mu )$ be a p.m.p.\ action of $G$ and let $\kappa$ denote the associated Koopman representation on $L^2(X,\mu )$. Let $\mathscr{M}_0$ be a subsemigroup of $\mathscr{M}$. Then the following are equivalent:
\begin{enumerate}
\item[i.] $\sigma _{|\mathscr{M}_0}$ is weakly mixing;
\item[ii.] $(\sigma \otimes \rho )_{|\mathscr{M}_0}$ is ergodic for every ergodic p.m.p.\ action $\rho$ of $G$;
\item[iii.] The restriction of $\kappa _{|\mathscr{M}_0}$ to $L^2(X,\mu ) \ominus \C 1_X$ is weakly mixing.
\end{enumerate}
\end{proposition}
\begin{proof}
This follows from Propositions \ref{prop:WM} and \ref{prop:ERG}.
\end{proof}
\subsection{Proof of Theorem \ref{thm:superrigid1}}
\begin{proof}[Proof of Theorem \ref{thm:superrigid1}]
For $g\in G$ let $C_H(g) = H\cap C_G(g)$. Define the set
\[
D = \{ g\in G \csuchthat \sigma _{|C_H(g)} \text{ has stable spectral gap} \} ,
\]
and let $G_0 = \langle H , D \rangle$. If $\bm{m}$ is a mean on $G$ which is invariant under conjugation by $H$ then, since $\sigma _{|H}$ has stable spectral gap, Lemma \ref{lem:ssg} implies that $\bm{m}(D) = 1$, so in particular $\bm{m}(G_0)=1$.
Let $w:G\times X \ra L$ be a cocycle with values in a group $L\in \mathscr{U}_{\mathrm{fin}}$. To show that $w$ untwists on $G_0$, we claim that it is enough to show that $w$ untwists on $H$. Indeed, assume that $w$ untwists on $H$. If $g\in D$ then $C_H(g)\leq gHg^{-1}\cap H$ and $\sigma _{|C_H(g)}$ has stable spectral gap, so in particular $\sigma _{|C_H(g)}$ is weakly mixing. We can therefore apply the last statement in Lemma 3.5 of \cite{Fu07} to conclude that $w$ untwists on all of $G_0$.
It remains to show that $w$ untwists on $H$. We may assume that $L$ is a closed subgroup of the unitary group $\mathscr{U}(N)$ of some finite von Neumann algebra $N$. Let $A= L^\infty (X,\mu )$ and view $w$ as a cocycle $w:G\ra \mathscr{U}(A\otimes N )$ for the action $\sigma \otimes \mathrm{id}_N$, i.e., satisfying $w_{gh}=w_g(\sigma _g \otimes \mathrm{id}_N)(w_h )$. We will use Popa's setup from Theorem 4.1 of \cite{Pop08}, with $A$ here taking the place of $P$. Namely, we let $M=(A\otimes N)\rtimes _{\sigma \otimes \mathrm{id}_N}G$
and we let $\wt{M} = (A\otimes A \otimes N)\rtimes _{\sigma \otimes \sigma \otimes \mathrm{id}_N}G$, and we view $M$ as a subalgebra of $\wt{M}$ so that the canonical unitaries $\{ u_g \} _{g\in G}\subset M$ implement $\sigma \otimes \mathrm{id}_N$ and $\sigma \otimes \sigma \otimes \mathrm{id}_N$ on $M$ and $\wt{M}$ respectively. We let $\tau$ denote the trance on $\wt{M}$.
Let $\{ \alpha _t \} _{t\in \R} \cup \{ \beta \} \subseteq \mathrm{Aut}(A\otimes A)$ denote the $s$-malleable deformation, and we extend $\beta$ and $\alpha _t$, $t\in \R$, to automorphisms of $\wt{M}$ by letting $\beta (x) = x = \alpha _t (x)$ if $x\in N\otimes \mathrm{L}G$. Let $\wt{u}_g = w_gu_g$ for $g\in G$, so that $g\mapsto \wt{u}_g$ is a homomorphism. Let $\wt{\pi}$ denote the representation of $G$ on $L^2(\wt{M}) =L^2(M)\otimes L^2(A)$ determined by $\wt{\pi} (g) ((xu_h )\otimes y ) = \mathrm{Ad}(\wt{u}_g)(xu_h)\otimes \sigma _g (y)$, for $x\in A\otimes N$, $y\in A$, $g,h\in G$.
\begin{claim}\label{claim:pv}
$\lim _{t\ra 0}\big( \sup _{\bm{m}\in C_{\mathscr{M}}(H)}\int _{g\in G} \| \alpha _t (\wt{u}_g)- \wt{u}_g \| _2 ^2 \, d\bm{m}(g) \big) = 0$.
\end{claim}
\begin{proof}[Proof of Claim \ref{claim:pv}]
Fix $\epsilon >0$. It suffices to show that there exists $t_\epsilon >0$, along with $S\subseteq H$ finite and $\delta >0$, such that if $p\in \Es{P}$ satisfies $\sup _{s\in S} \| sps^{-1} - p \| _1 < \delta$ then for all $t$ with $0\leq |t|\leq t_{\epsilon}$ we have $\int _G \| \alpha _t (\wt{u}_g)- \wt{u}_g \| _2 ^2 \, dp (g) < \epsilon$, since the claim will then follow using Lemma \ref{lem:Mazur}. Since $H\cc ^{\sigma }(X,\mu )$ has stable spectral gap, there exists a finite set $S\subseteq H$ and $\delta _0 >0$ such that if $\eta \in L^2(\wt{M})$ is a unit vector satisfying $\sup _{s\in S} \| \wt{\pi} (s)\eta - \eta \| _2 < \delta _0$, then $\| \eta - e(\eta ) \| _2 < \epsilon ^{1/2}/2$, where $e: L^2(\wt{M})\ra L^2(M)$ denotes the orthogonal projection. Since $S$ is finite, there exists $t_1>0$ such that for all $0\leq t_0\leq t_1$ we have $\sup _{s\in S} \| \alpha _{t_0} (\wt{u}_s)-\wt{u}_s \| _2 < \delta _0 /4$. Let $t_\epsilon = 2t_1$ and fix $t_0$ with $0\leq t_0\leq t_1$. Let $\delta = \delta _0 ^2/4$ and fix $p\in \Es{P}$ with $\sup _{s\in S} \| sps^{-1}-p \| _1 < \delta$.
Let $Q= \{ \wt{u}_g \} _{g\in G}''$. Then we may identify $\ell ^2(G)$ with $L^2(Q) \subset L^2(M)$ via $\delta _g \mapsto \wt{u}_g$. For each $q\in \Es{P}$, let $\eta _q = \sum _{g\in G} q(g)^{1/2}\wt{u}_g \in L^2(Q)$. We have $\tau (\alpha _{t_0}(\wt{u}_g)\wt{u}_h^*) = 0$ for all $g\neq h$, which implies that $\| \alpha _{t_0} (\wt{u}_g)\xi - \wt{u}_g\xi \| _2 = \| \alpha _{t_0} (\wt{u}_g) -\wt{u}_g \| _2 = \| \xi \alpha _{t_0}(\wt{u}_g) - \xi \wt{u}_g \| _2$ for every unit vector $\xi \in L^2(Q)$, and therefore $\sup _{s\in S} \| \alpha _{t_0} (\wt{u}_s) \eta _p - \eta _p \alpha _{t_0} (\wt{u}_s) \| _2 < \delta _0/2 + \sup _{s\in S}\| \eta _{sps^{-1}} - \eta _p \| _2 < \delta _0$. By replacing $t_0$ by $-t_0$ and applying $\alpha _{t_0}$ we obtain $\sup _{s\in S} \| \wt{\pi}(s) \alpha _{t_0}(\eta _p) - \alpha _{t_0}(\eta _p) \| _2< \delta _0$. Our choice of $\delta _0$ then implies that $\| \alpha _{t_0}(\eta _p)- e(\alpha _{t_0}(\eta _p )) \| _2 < \epsilon ^{1/2} / 2$, and hence by Popa's Transversality Lemma \cite[Lemma 2.1]{Pop08},
\begin{align*}
\int _{G}\| \alpha _{2t_0}(\wt{u}_g) - \wt{u}_g \| _2 ^2 \, dp &\leq 4\int _{G} \| \alpha _{t_0} (\wt{u}_g) - e(\alpha _{t_0}(\wt{u}_g)) \| _2 ^2 \, dp = 4 \| \alpha _{t_0}(\eta _p)- e(\alpha _{t_0}(\eta _p)) \| _2^2 < \epsilon .
\end{align*}
\qedhere[Claim \ref{claim:pv}]
\end{proof}
\begin{claim}\label{claim:2}
$\lim _{t\ra 0} \big( \sup _{h\in H}\| \alpha _t (\wt{u}_h)- \wt{u}_h \| _2^2 \big) =0$.
\end{claim}
\begin{proof}[Proof of Claim \ref{claim:2}]
Fix $\epsilon >0$. By Claim \ref{claim:pv} there exists $t_\epsilon >0$ such that $\int _G \| \alpha _t (\wt{u}_g)- \wt{u}_g \| _2 ^2 \, d\bm{m} <\epsilon /8$ for all $\bm{m}\in C_{\mathscr{M}}(H)$ and all $t$ with $0\leq |t|\leq t_\epsilon$. Fix any $h\in H$ along with $0\leq |t|\leq t_\epsilon$, and we will show that $\| \alpha _t (\wt{u}_h)-\wt{u}_h \| _2^2 < \epsilon$. For $\bm{m}\in C_{\mathscr{M}}(H)$ we have
\begin{align*}
\int _{G} \| \alpha _t &(\wt{u}_{hgh^{-1}})\wt{u}_h - \wt{u}_h\alpha _t (\wt{u}_g) \| _2^2 \, d\bm{m} \\
&\leq \Big[ \Big( \int _{G}\| \alpha _t (\wt{u}_{hgh^{-1}})- \wt{u}_{hgh^{-1}}\| _2^2 \, d\bm{m} \Big) ^{1/2} + \Big( \int _G \| \alpha _t (\wt{u}_g )- \wt{u}_g \| _2^2 \, d\bm{m} \Big) ^{1/2}\Big]^2 < \epsilon /2 .
\end{align*}
By replacing $t$ with $-t$ and applying $\alpha _t$ we obtain $\int _G \| \wt{u}_{hgh^{-1}}\alpha _t (\wt{u}_h)-\alpha _t(\wt{u}_h)\wt{u}_g \| _2^2 \, d\bm{m} <\epsilon /2$. Let $\eta _h = \alpha _t (\wt{u}_h)- e(\alpha _t(\wt{u}_h))$. Then by projecting onto $L^2(\wt{M})\ominus L^2(M)$ and using the last inequality we obtain
\begin{equation}\label{eqn:etah}
\int _G \| \wt{u}_{hgh^{-1}}\eta _h -\eta _h \wt{u}_g \| _2^2 \,d\bm{m} <\epsilon /2 , \ \ (\bm{m} \in C_{\mathscr{M}}(H) ) .
\end{equation}
{\bf Subclaim:} $\inf _{C_{\mathscr{M}}(H)} \sum _{i,j <n} \int _G |\langle \wt{u}_{hgh^{-1}}\eta _i , \eta _j \wt{u}_g \rangle | \, d\bm{m} = 0$ for any $\eta _0,\dots ,\eta _{n-1} \in L^2(\wt{M})\ominus L^2(M)$.
\begin{proof}[Proof of Subclaim] We may assume each $\eta _i$ is of the form $\eta _i = (a_i\otimes b_i \otimes c_i )u_{k_i}$, where $a_i \in A\otimes 1$, $b_i \in 1\otimes A$, $\tau (b_i ) = 0$, $c_i \in N$, $k_i\in G$, since such vectors span a dense subspace of $L^2(\wt{M})\ominus L^2(M)$. Since $w_l \in A \otimes 1 \otimes N$, we have
\begin{align*}
|\langle \wt{u}_{hgh^{-1}}\eta _i , \eta _j \wt{u}_g \rangle | &= |\langle w_{hgh^{-1}}u_{hgh^{-1}}(a_i\otimes b_i\otimes c_i)u_{k_i}, (a_j\otimes b_j\otimes c_j)u_{k_j}w_gu_g\rangle | \\
&\leq |\langle w_{hgh^{-1}}(\sigma _{hgh^{-1}}(a_i)\otimes c_i), (a_j\otimes c_j)(\sigma _{k_j}\otimes \mathrm{id}_N)(w_g)\rangle | \, | \langle \sigma _{hgh^{-1}}(b_i), b_j \rangle | \\
&\leq \| a_i \otimes c_i \| _2 \| a_j \otimes c_j \| _2 | \langle \sigma _{hgh^{-1}}(b_i), b_j \rangle | .
\end{align*}
Therefore, letting $C=\max _{i<n}\| a_i\otimes c_i \| _2$, for $\bm{m}\in C_{\mathscr{M}}(H)$ we have
\begin{equation}\label{eqn:C}
\sum _{i,j<n} \int _G |\langle \wt{u}_{hgh^{-1}}\eta _i , \eta _j \wt{u}_g \rangle | \, d\bm{m}(g) \leq C^2 \sum _{i,j<n}\int _G |\langle \sigma _{g}(b_i), b_j \rangle | \, d\bm{m} (g) .
\end{equation}
By Proposition \ref{prop:WM}, since $C_{\mathscr{M}}(H)\cc ^{\sigma}A$ is weakly mixing, the infimum over $\bm{m} \in C_{\mathscr{M}}(H)$ of the right hand side of \eqref{eqn:C} is $0$.\qedhere[Subclaim]
\end{proof}
By \eqref{eqn:etah} and the subclaim, we have
\[
\epsilon /2 \geq \sup _{C_{\mathscr{M}}(H)} \int _G \| \wt{u}_{hgh^{-1}}\eta _h -\eta _h \wt{u}_g \| _2^2 \, d\bm{m} \geq 2\| \eta _h \| _2 ^2 - 2\inf _{C_{\mathscr{M}}(H)} \int |\langle \wt{u}_{hgh^{-1}}\eta _h , \eta _h \wt{u}_g \rangle |\, d\bm{m} = 2 \| \eta _h \| _2 ^2 .
\]
Popa's Transversality Lemma now shows $\| \alpha _t (\wt{u}_h) - \wt{u}_h \| _2 ^2 \leq 4\| \eta _h \| _2 ^2 \leq \epsilon$. \qedhere[Claim \ref{claim:2}]
\end{proof}
By \cite{Pop07}, Claim \ref{claim:2} implies that $w$ untwists on $H$.
\end{proof}
\section{The AC-center, the inner radical, and linear groups}\label{sec:radicals}
\subsection{Proof of Theorem \ref{thm:I(G)}, parts i.\ through viii.}
\begin{proof}[Proof of Theorem \ref{thm:I(G)}, parts i.\ through viii.] We begin with the statements involving $\mathscr{AC}(G)$. It is clear that $\mathscr{AC}(G)$ and $\mathscr{I}(G)$ are characteristic subgroups of $G$. If $N_0$ and $N_1$ are normal subgroups of $G$ with both $G/C_G(N_0)$ and $G/C_G(N_1)$ amenable, then $G/C_G(N_0N_1) = G/(C_G(N_0)\cap C_G(N_1))$ is amenable. Therefore, $\mathscr{AC}(G)$ may be written as an increasing union $\mathscr{AC}(G)=\bigcup _{i\in \N} N_i$, with each $N_i$ normal in $G$ and $G/C_G(N_i)$ amenable. It follows that $G/C_G(\mathscr{AC}(G))$ is residually amenable since $C_G(\mathscr{AC}(G))= \bigcap _{i\in \N}C_G(N_i)$. Each of the groups $N_i$ is amenable since $N_i/Z(N_i)$ is isomorphic to a subgroup of the amenable group $G/C_G(N_i)$. This shows that $\mathscr{AC}(G)$ is amenable. Moreover, for each $i\in \N$, the action $N_i\rtimes G\cc N_i$ is amenable since it descends to an action of the amenable group $N_i\rtimes (G/C_G(N_i))$. If $\bm{m}_i$ is a $N_i\rtimes G$-invariant mean on $N_i$, then any accumulation point of $(\bm{m}_i)_{i\in \N}$ in the space of means on $G$ will be a mean witnessing that the action $\mathscr{AC}(G)\rtimes G \cc \mathscr{AC}(G)$ is amenable. It follows that $\mathscr{AC}(G) \leq \mathscr{I}(G)$.
To prove the remaining statements involving $\mathscr{I}(G)$ we will use the following lemma.
\begin{lemma}\label{lem:extendI(G)} Let $H$ and $K$ be normal subgroups of $G$.
\begin{enumerate}
\item Assume that $H\leq K$ and that the actions $H\rtimes G\cc H$ and $K/H \rtimes G/H \cc K/H$ are both amenable, with invariant means $\bm{m}_H$ and $\bm{m}_{K/H}$ respectively. Then the action $K\rtimes G \cc K$ is amenable with invariant mean $\bm{m}_K= \int _{kH\in K/H}k\bm{m}_H\, d\bm{m}_{K/H}$.
\item Assume that the actions $H\rtimes G\cc H$ and $K\rtimes G\cc K$ are both amenable, with invariant means $\bm{m}$ and $\bm{n}$ respectively. Then the action $HK\rtimes G\cc HK$ is amenable with invariant mean $\bm{m}\ast \bm{n}$.
\end{enumerate}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:extendI(G)}] Since $\bm{m}_H$ is invariant under left translation by $H$, for each $g\in G$ the mean $g\bm{m}_H$ only depends on the coset $gH\in G/H$. The mean $\bm{m}_K$ is therefore well-defined and it is straightforward to verify that it is $K\rtimes G$-invariant. This shows (1), and (2) can either be deduced from (1) or verified directly. \qedhere[Lemma \ref{lem:extendI(G)}]
\end{proof}
It follows from Lemma \ref{lem:extendI(G)}.(2) that $\mathscr{I}(G)$ may be written as an increasing union $\mathscr{I}(G)=\bigcup _{i\in \N}M_i$, where each $M_i$ is normal in $G$ and the action $M_i\rtimes G \cc M_i$ is amenable. If $\bm{m}_i$ is an invariant mean for the action $M_i\rtimes G\cc M_i$, then any accumulation point $\bm{m}$ of $(\bm{m}_i)_{i\in \N}$ will be an invariant mean for the action $\mathscr{I}(G)\rtimes G \cc \mathscr{I}(G)$. In particular, $\bm{m}$ witnesses that $\mathscr{I}(G)$ is amenable, and if $\mathscr{I}(G)$ is infinite then $\bm{m}$ also witnesses that $G$ is inner amenable. The proof of i.\ through v.\ is now complete.
vi. Let $\pi : G\ra G/N$ denote the projection map. Then the image under $\pi$ of an $\mathscr{I}(G)\rtimes G$-invariant mean on $\mathscr{I}(G)$ is an $\mathscr{I}(G)/N \rtimes G/N$-invariant mean on $\mathscr{I}(G)/N$. This shows that $\mathscr{I}(G)/N \leq \mathscr{I}(G/N )$. The reverse containment then follows by applying part (1) of Lemma \ref{lem:extendI(G)} to the groups $H=\mathscr{I}(G)$ and $K= \pi ^{-1}(\mathscr{I}(G/N ) )$.
vii. Part vi.\ implies that $\mathscr{I}(G/\mathscr{I}(G))=1$, and this in turn implies that $G/\mathscr{I}(G)$ is ICC since every finite conjugacy class in $G/\mathscr{I}(G)$ is contained in $\mathscr{AC}(G/\mathscr{I}(G))\leq\mathscr{I}(G/\mathscr{I}(G))=1$.
viii. This in fact holds more generally with $\mathscr{I}(G)$ replaced by any normal subgroup $N$ of $G$ for which $N\rtimes G\cc N$ is amenable. To see this, fix an invariant mean $\bm{n}$ for the action $N\rtimes G\cc N$. As in Lemma \ref{lem:extendI(G)}, we obtain a well-defined map
\[
\bm{m}\mapsto \bm{m}\ast\bm{n} = \int _{gN\in G/N}g\bm{n}\, d\bm{m}(gN)
\]
taking means on $G/N$ to means on $G$. This map is a section for the projection map on means, and since $\bm{n}$ is invariant under conjugation by $G$, this map takes conjugation invariant means on $G/N$ to conjugation invariant means on $G$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:I(G)}, parts ix.\ through xiv.}
The second half of Theorem \ref{thm:I(G)} will be deduced from the following spectacular theorem of S.G.\ Dani from \cite{Da85}, which appears to have been overlooked since its publication in 1985. In what follows, if $G\cc X$ is an action of a group $G$ then for $A\subseteq X$ let $\mathrm{stab}_G(A)$ denote the pointwise stabilizer of $A$ in $G$, and for $D\subseteq G$ let $\mathrm{fix}_X(D)$ denote the set of points in $X$ which are fixed by every element of $D$.
\begin{theorem}[Theorem 1.1 of \cite{Da85}]\label{thm:Dani}
Let $G\cc X$ be an amenable action of a group $G$ on a set $X$ and let $\bm{m}$ be a $G$-invariant mean on $X$. Suppose that the action satisfies the following two conditions:
\begin{enumerate}
\item[(1)] For every subset $A\subseteq X$ there exists a finite $A_0\subseteq A$ such that $\mathrm{stab}_G(A)=\mathrm{stab}_G(A_0)$.
\item[(2)] For every subset $D\subseteq G$ there exists a finite $D_0\subseteq D$ such that $\mathrm{fix}_X(D)=\mathrm{fix}_X(D_0)$.
\end{enumerate}
Then there exists a normal subgroup $N$ of $G$ such that $G/N$ is amenable and $\bm{m}(\mathrm{fix}_X(N)) =1$.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{thm:I(G)}, parts ix.\ through xiv.]
Assume that $G$ is linear. We first show that any conjugation invariant mean $\bm{m}$ on $G$ must concentrate on a normal subgroup $M$ of $G$ such that $G/C_G(M)$ is amenable. Consider the conjugation action $G\cc G$. For $A, D\subseteq G$ we have $\mathrm{stab}_G(A) = C_G(A)$, and $\mathrm{fix}_G(D) = C_G(D)$. Conditions (1) and (2) of Theorem \ref{thm:Dani} are therefore satisfied since $G$ satisfies the minimal condition on centralizers (see Remark \ref{rem:minC}). We conclude that there exists a normal subgroup $N$ of $G$ such that $G/N$ is amenable, and $\bm{m}(C_G(N))=1$. Take $M=C_G(N)$. Then $N\leq C_G(M)$, so $G/C_G(M)$ is amenable, as was to be shown. This also shows that xiii.\ holds.
ix.\ and x. By part iii., we may find an invariant mean $\bm{m}$ for the action $\mathscr{I}(G)\rtimes G\cc \mathscr{I}(G)$. Since $\bm{m}$ is conjugation invariant there exists a normal subgroup $M$ of $G$ with $G/C_G(M)$ amenable and $\bm{m}(M)=1$. Then $M\leq \mathscr{AC}(G)\leq \mathscr{I}(G)$ and, since $\bm{m}$ is invariant under left translation by $\mathscr{I}(G)$, we have equality $M=\mathscr{I}(G)$.
xi. Parts x.\ and ii.\ show that $\mathscr{I}(G)\leq C_G(C_G(\mathscr{I}(G)))\leq \mathscr{AC}(G)\leq \mathscr{I}(G)$.
xii. This follows from viii.\ and xiii.
xiv. From vi., we have $C_{G/N}(\mathscr{I}(G/N )) = C_{G/N}(\mathscr{I}(G)/N ) \geq C_G(\mathscr{I}(G))N /N$, hence $(G/N )/C_{G/N}(\mathscr{I}(G/N) )$ is amenable, and $\mathscr{I}(G/N)\leq \mathscr{AC}(G/N)$. Part ii.\ gives the reverse inclusion. It follows as in xi.\ above that $\mathscr{I}(G/N )$ coincides with its double centralizer. The group $(G/N)/\mathscr{I}(G/N) =(G/N)/(\mathscr{I}(G)/N) \cong G/\mathscr{I}(G)$ is not inner amenable by xii. Now let $\bm{m}_0$ be a conjugation invariant mean on $G/N$, and let $\bm{m}_1$ denote the projection of $\bm{m}_0$ to $G/\mathscr{I}(G)$. Then $\bm{m}_1$ is a conjugation invariant mean on $G/\mathscr{I}(G)$, so by viii.,\ $\bm{m}_1$ is the projection of some conjugation invariant mean $\bm{m}$ on $G$. By xiii.,\ $\bm{m}$ concentrates on $\mathscr{I}(G)$, hence $\bm{m}_1$ is the point mass at the identity, and therefore $\bm{m}_0$ concentrates on $\mathscr{I}(G)/N = \mathscr{I}(G/N)$.
\end{proof}
\subsection{Proof of Theorems \ref{thm:linear} and \ref{thm:Schmidt}}\label{sec:linear}
\begin{proof}[Proof of Theorem \ref{thm:linear}] The implication (1)$\Ra$(2) follows from Theorem \ref{thm:I(G)}.xiii., and (2)$\Ra$(1) is Theorem \ref{thm:I(G)}.v. Assume now that (2) holds and let $N=C_G(\mathscr{I}(G))\mathscr{I}(G)$. Then $G/N$ is amenable by Theorem \ref{thm:I(G)}.x., and $Z(N)=C_G(\mathscr{I}(G))\cap \mathscr{I}(G)$ by Theorem \ref{thm:I(G)}.xi., so (3) follows. If (3) holds then in the first alternative $Z(N)$ infinite and $Z(N)\leq \mathscr{AC}(G)\leq \mathscr{I}(G)$, and in the second alternative $M$ is infinite and $M\leq \mathscr{AC}(G)\leq \mathscr{I}(G)$, so (2) holds either way.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:Schmidt}]
Assume that $G$ is inner amenable and we will construct the desired action of $G$. This is straightforward if $G$ is stable, so we may assume that $G$ is inner amenable, but not stable. Then the group $N=C_G(\mathscr{I}(G))$ has infinite center $C$ (see Remark \ref{rem:final}), and by Theorem \ref{thm:I(G)} the group $K= G/N$ is amenable. Since $C$ is a countable abelian group, it possesses a free p.m.p.\ action $C\cc (Y,\nu )$ which is compact (for example, using a countable dense subset of $\widehat{C}$, inject $C$ as a subgroup of $\T ^\N$ and let $C$ act by translation on $\T ^\N$ equipped with Haar measure). Let $G\cc (X,\mu ) = (Y, \nu ) ^{G/C}$ be the coinduced action. This is a free weakly mixing action of $G$, and the restriction of this action to $C$ is an infinite diagonal product of compact actions of $C$, hence is itself a compact action. It follows that there exists a sequence $(c_n)_{n\in \N}$ in $C - 1$ which converges to the identity automorphism in the group $\mathrm{Aut}(X,\mu )$ equipped with the weak topology. The sequence $(c_n)_{n\in \N}$ is then asymptotically central in $[\mc{R}^N_X]$, and since $C$ acts freely, the sequence $(c_n)_{n\in \N}$ witnesses that the outer automorphism group of $[\mc{R}^N_X]$ is not Polish. Let $K\cc (Z,\eta )$ be a free ergodic action of $K$, and let $G\cc (X,\mu ) \otimes (Z ,\eta )$ be the diagonal product action where $G$ acts on $(Z,\eta )$ via the quotient map to $K$. This action of $G$ is free and ergodic, and as observed in Remark \ref{rem:overview} below, the construction in the proof of Theorem \ref{thm:extension} below yields a sequence $(T_n)_{n\in \N}$ witnessing that the outer automorphism group of $[\mc{R}^G_{X\times Z}]$ is not Polish.
For the converse, which holds even without the assumption that $G$ is linear, see \cite{JS87}.
\end{proof}
\section{Stability}\label{sec:stability}
\subsection{Kida's stability criterion}\label{sec:Kida}
In this section we employ the notation from \S\ref{sec:groupoid}. Let $(\mc{G},\mu )$ be a discrete p.m.p.\ groupoid and let $[\mc{G}]$ denote the {\bf full group} of $\mc{G}$, i.e., the collection of all local sections with domain equal to all of $\mc{G}^0$.
\begin{definition}\label{def:stabseq}
A sequence $(T_n)_{n\in \N}$ in $[\mc{G}]$ is said to be {\bf asymptotically central} if
\begin{enumerate}
\item[(i)] $\mu ( T_n ^0 A\triangle A) \ra 0$ for all measurable $A\subseteq \mc{G}^0$;
\item[(ii)] $\mu ( \{ x\in \mc{G}^0 \csuchthat (T_n\circ S)x = (S\circ T_n)x \} ) \ra 1$ for all $S\in [\mc{G}]$.
\end{enumerate}
A sequence $(T_n)_{n\in \N}$ in $[\mc{G}]$ is called a {\bf stability sequence} for $\mc{G}$ if it is asymptotically central and if furthermore there exists a sequence $(A_n)_{n\in \N}$ of measurable subsets of $\mc{G}^0$ such that
\begin{enumerate}
\item[(iii)] $(A_n)_{n\in \N}$ is asymptotically invariant for $\mc{G}$, i.e., $\mu (S^0A_n \triangle A_n ) \ra 0$ for all $S\in [\mc{G}]$;
\item[(iv)] $\mu (T_n^0A_n \triangle A_n ) \not\ra 0$.
\end{enumerate}
\end{definition}
\begin{remark}\label{rem:checkG}
Suppose that $\mc{G}=G\ltimes (X,\mu )$ is the translation groupoid associated to a p.m.p.\ action $G\cc (X,\mu )$ of a countable group $G$. We view each $T\in [\mc{G}]$ as a map from $X$ to $G$, so that $T^0 (x) = T(x)\cdot x$ for $x\in X$. We will make use of the observation \cite[\S 3.1]{Ki13a} that in this situation, a sequence $(T_n)_{n\in \N}$ in $[\mc{G}]$ is asymptotically central if and only if it satisfies (i) along with
\begin{enumerate}
\item[(ii${}'$)] $\mu ( \{ x\in X \csuchthat T_n(g\cdot x) = g T_n(x)g^{-1} \} ) \ra 1$ for all $g\in G$.
\end{enumerate}
Likewise, a sequence $(A_n)_{n\in \N}$ of measurable subsets of $X$ is asymptotically invariant for $\mc{G}$ if and only if $\mu (g\cdot A_n \triangle A_n ) \ra 0$ for all $g\in G$.
\end{remark}
The following theorem, due to Kida \cite{Ki13a}, provides an important criterion for demonstrating stability of a group.
\begin{theorem}[Theorem 1.4 of \cite{Ki13a}]\label{thm:criterion}
Let $G$ be a countable group and suppose that there exists a p.m.p.\ action $G\cc (X,\mu )$ of $G$ whose associated translation groupoid $G\ltimes (X,\mu )$ admits a stability sequence. Then $G$ is stable.
\end{theorem}
\subsection{Proof of Theorem \ref{thm:extension}}
\begin{proof}[Proof of Theorem \ref{thm:extension}]
Let $G\cc (X,\mu )$ be a p.m.p.\ action of $G$ such that $N\ltimes (X,\mu )$ admits a stability sequence. Let $K\cc (Z,\eta )$ be a free p.m.p.\ action of $K$ and let $G\cc (X,\mu ) \otimes (Z,\eta )$ be the diagonal product action, where $G$ acts on the second coordinate via the quotient map to $K$. In what follows we will often identify an element of $G$ with its image in $K$.
By Theorem \ref{thm:criterion}, it suffices to show that the translation groupoid $G\ltimes (X,\mu )\otimes (Z,\eta )$ admits a stability sequence. We will construct such a sequence $(T_n)_{n\in \N}$ which is moreover contained in $[N\ltimes (X,\mu )\otimes (Z,\eta )]$. Let $F_0\subseteq F_1\subseteq \cdots$ be an exhaustion of $G$ by finite subsets. By Theorem 3.1 of \cite{BT-D11}, since $K$ is amenable, for each $n\geq 0$ we may find a measurable function $\varphi _n : Z\ra K$ such that $\eta (C_n)>1-2^{-n}$, where $C_n = \{ z\in Z\csuchthat (\forall g\in F_n ) \ \varphi _n(g\cdot z) = g\varphi _n(z) \}$. Let $1_K\in Q_0\subseteq Q_1\subseteq \cdots$ be an exhaustion of $K$ by finite subsets such that for each $n\geq 0$ we have $\eta (D_n)>1-2^{-n}$, where $D_n = \{ z\in Z\csuchthat \varphi _n (z)\in Q_n \}$. Let $Z_n = \bigcap _{m\geq n}C_m\cap D_m$, so that $Z_0\subseteq Z_1\subseteq \cdots$, and $\eta (\bigcup _n Z_n) = 1$. After ignoring a null set we may assume that $\bigcup _n Z_n = Z$. Fix a section $\sigma : K \ra G$ for the map $G\ra K$ with $\sigma (1_K)=1_G$, and let $\rho : G\times K \ra N$ be the associated Schreier cocycle $\rho (g,k)= \sigma (gk)^{-1}g\sigma (k) \in N$.
By assumption, the groupoid $N\rtimes (X,\mu )$ admits a stability sequence $(S_i)_{i\in \N}$. After replacing $(S_i)_{i\in \N}$ by a subsequence if necessary we may assume that there exists an asymptotically invariant sequence $(B_i)_{i\in \N}$ for $N\rtimes (X,\mu )$ such that $\lim _i \mu (S_i^0B_i\triangle B_i ) >0$. Fix a sequence $\mc{B}_0\subseteq \mc{B}_1\subseteq \cdots$ of finite algebras of measurable subsets of $X$ whose union generates the measure algebra of $X$. By moving to a subsequence $(S_{i_n})_{n\in \N}$ and $(B_{i_n})_{n\in \N}$, which we will call $(S_n)_{n\in \N}$ and $(B_n)_{n\in \N}$ respectively, we may ensure that
\begin{enumerate}
\item[{\bf (C1)}] $\mu (S_n^0(\sigma (k)^{-1}\cdot A)\, \triangle \, \sigma (k)^{-1}\cdot A ) < 1/n$ for all $k\in Q_n$ and $A\in \mc{B}_n$.
\item[{\bf (C2)}] $\mu (W_n)>1-2^{-n}$, where
\[
W_n = \{ x\csuchthat S_n(\rho (g,k)\cdot (\sigma (k)^{-1}\cdot x))= \rho (g,k)S_n(\sigma (k)^{-1}\cdot x)\rho (g,k)^{-1} \text{ for all }g\in F_n, \ k \in Q_n \} .
\]
\item[{\bf (C3)}] $\mu (\rho (g,k)^{-1} B_n \triangle B_n ) < 1/n$ for all $g\in F_n$, $k\in Q_n$.
\end{enumerate}
Let $X_n = \bigcap _{m\geq n}W_m$, so that $X_0\subseteq X_1\subseteq \cdots$, and $\mu (\bigcup _n X_n) = 1$. After ignoring a null set we may assume that $\bigcup _n X_n = X$. For each $n\in \N$ and $(x,z)\in X\times Z$ define
\[
T_n(x,z) = \sigma (\varphi _n (z))S_n(\sigma (\varphi _n (z))^{-1}\cdot x ) \sigma (\varphi _n (z))^{-1} \in N .
\]
Then $T_n^0$ is an automorphism of $(X ,\mu )\otimes (Z , \eta )$, since for each fiber $(X_z,\mu _z ) := (X , \mu )\otimes ( \{ z \} , \updelta _z )$, the restriction of $T_n^0$ to $X_z$ is an automorphism $T_{n,z}^0 : (X_z,\mu _z)\ra (X_z,\mu _z )$. We now verify that $(T_n)_{n\in \N}$ satisfies properties (i)-(iv) of Definition \ref{def:stabseq} with respect to the groupoid $\mc{G}:=G\ltimes (X,\mu )\otimes (Z,\eta )$.
(i): It suffices to prove (i) for rectangles $D=A\times C$, with $A\in \bigcup _n \mc{B}_n$. For $z\in Z\setminus C$ we have $\mu _z(T_{n,z}^0(A\times C)_z\triangle (A\times C)_z ) =0$ for all $n$. For $z\in C$, if $n$ is large enough then $A\in \mc{B}_n$ and $z\in Z_n$, so $\varphi _n (z)\in Q_n$, hence
\begin{align*}
\mu _z (T_{n,z}^0(A\times C)_z\triangle (A\times C)_z) &= \mu (\sigma (\varphi _n (z))\cdot S^0_n(\sigma (\varphi _n (z))^{-1}\cdot A )\triangle A) \\
&= \mu (S^0_n(\sigma (\varphi _n (z))^{-1}\cdot A)\triangle \sigma (\varphi _n(z))^{-1}\cdot A ) < 1/n
\end{align*}
by {\bf (C1)}. Therefore, $(\mu \otimes \eta )(T_n^0(A\times C)\triangle (A\times C)) = \int_Z\mu _z(T_{n,z}^0(A\times C)_z \triangle (A\times C)_z ) \, d\eta \ra 0$ as $n\ra \infty$.
(ii${}'$): Fix $g\in G$. For all large enough $n\in \N$ we have $g \in F_n$, so if $(x,z)\in X_n\times Z_n$ then $\varphi _n(g\cdot z) = g\cdot \varphi _n(z)$ and $\varphi _n (z)\in Q_n$, hence by {\bf (C2)}, $\rho (g,\varphi _n (z))^{-1}S_n(\rho (g,\varphi _n (z))\cdot \sigma (\varphi _n (z))^{-1}\cdot x)\rho (g,\varphi _n (z)) = S_n(\sigma (\varphi _n (z))^{-1}\cdot x)$. Therefore, for all large enough $n$, if $(x,z)\in X_n\times Z_n$ then we have
\begin{align*}
T_n (g\cdot &(x, z)) =T_n(g\cdot x, g\cdot z ) \\
&= \sigma (\varphi _n (g\cdot z))S_n(\sigma (\varphi _n (g\cdot z))^{-1}\cdot (g\cdot x) ) \sigma (\varphi _n (g\cdot z))^{-1} \\
&= \sigma (g\cdot \varphi _n (z))S_n (\sigma (g \cdot \varphi _n (z))^{-1}\cdot (g\cdot x ))\sigma (g\cdot \varphi _n (z))^{-1} \\
&= g\sigma (\varphi _n (z))\rho (g,\varphi _n (z))^{-1}S_n(\rho (g,\varphi _n (z))\sigma (\varphi _n (z))^{-1}\cdot x )\rho (g,\varphi _n (z))\sigma (\varphi _n (z))^{-1}g^{-1} \\
&= g\sigma (\varphi _n (z))S_n(\sigma (\varphi _n (z))^{-1}\cdot x )\sigma (\varphi _n (z))^{-1}g^{-1} \\
&= gT_n (x,z)g^{-1},
\end{align*}
and since $X\times Z = \bigcup _n (X_n\times Z_n )$ the proof of (ii) is complete.
For each $n\in \N$ define the set $A_n = \{ (x,z)\in X\times Z\csuchthat x \in \sigma (\varphi _n (z)) \cdot B_n \}$. We will verify (iii) and (iv) using the sequence $(A_n)_{n\in \N}$.
(iii): Fix $g\in G$. For all large enough $n\in \N$ we have $g\in F_n$, so for $(x,z)\in X\times Z_n$ we have $(x,z)\in g^{-1}\cdot A_n \IFF x\in \sigma (\varphi _n (z))\cdot \rho (g,\varphi _n (z))^{-1}\cdot B_n$. Hence, by {\bf (C3)},
\[
(\mu \otimes \eta )(g^{-1}\cdot A_n \triangle A_n ) \leq \eta (Z\setminus Z_n) + \int _Z \mu (\rho (g,\varphi _n (z))^{-1}\cdot B_n \triangle B_n ) \, d\eta < \eta (Z\setminus Z_n) + 1/n \ra 0
\]
as $n\ra \infty$. This shows that $(A_n)_{n\in \N}$ is asymptotically invariant for $G\cc (X,\mu )\otimes (Z,\eta )$.
(iv): We have $T_n^0 A_n = \{ (x,z) \in X\times Z \csuchthat x\in \sigma (\varphi _n (z))\cdot S_n^0B_n ) \}$. It follows that $(\mu\otimes \eta )(T_n^0A_n\triangle A_n ) = \mu (S_n^0B_n\triangle B_n ) \not\ra 0$. This completes the proof.
\end{proof}
\begin{remark}\label{rem:overview}
Let $1\ra N \ra G \ra K \ra 1$ be a short exact sequence in which $K$ is amenable. Let $G\cc (X,\mu )$ be a p.m.p.\ action of $G$ and let $K\cc (Z,\eta )$ be a free p.m.p.\ action of $K$. Let $G\cc (X,\mu )\otimes (Z,\eta )$ be the diagonal product action where $G$ acts on $(Z,\eta )$ via the quotient map to $K$. The above proof constructs a map which takes asymptotically central sequences in $[N\ltimes (X,\mu )]$ to asymptotically central sequences in $[G\ltimes (X,\mu )\otimes (Z,\eta )]$, and which moreover takes stability sequences for $N\ltimes (X,\mu )$ to stability sequences for $G\ltimes (X,\mu )\otimes (Z,\eta )$. In addition, it follows from the construction that if $(S_n)_{n\in \N}$ is a sequence in $[N\ltimes (X,\mu ) ]$ witnessing that the outer automorphism group of $[\mc{R}^N_X]$ is not Polish, then the image of $(S_n)_{n\in \N}$ under this map will be a sequence witnessing that the outer automorphism group of $[\mc{R}^G_{X\times Z}]$ is not Polish (see \cite[Chapter I, \S 7]{Ke10}).
\end{remark}
\begin{comment}
\begin{remark}\label{rem:Schmidt}
In \cite{Sch86}, Schmidt raises the question of whether every inner amenable group $G$ possesses a free ergodic p.m.p.\ action $G\cc (X,\mu )$ which generates an orbit equivalence relation $\mc{R}^G_X$ for which the outer automorphism group of the full group $[\mc{R}^G_X]$ is not Polish, or equivalently, for which the full group $[\mc{R}^G_X]$ contains an asymptotically central sequence $(T_n)_{n\in \N}$ with $\liminf _n \mu ( \{ x \in X \csuchthat T_n x \neq x \} ) >0$. See also Problem 9.3 of \cite{Ke10}. We can now provide a positive answer to Schmidt's question when $G$ is linear, using Theorems \ref{thm:I(G)}, \ref{thm:linear}, and \ref{thm:linearstable}. This is straightforward if $G$ is stable, so we may assume that $G$ is inner amenable, but not stable. Then the group $N=C_G(\mathscr{I}(G))$ has infinite center $C$ (see Remark \ref{rem:final}), and by Theorem \ref{thm:I(G)} the group $K= G/N$ is amenable. Since $C$ is a countable abelian group, it possesses a free p.m.p.\ action $C\cc (Y,\nu )$ which is compact (for example, using a countable dense subset of $\widehat{C}$, inject $C$ as a subgroup of $\T ^\N$ and let $C$ act by translation on $\T ^\N$ equipped with Haar measure). Let $G\cc (X,\mu ) = (Y, \nu ) ^{G/C}$ be the coinduced action. This is a free weakly mixing action of $G$, and the restriction of this action to $C$ is an infinite diagonal product of compact actions of $C$, hence is itself a compact action. It follows that there exists a sequence $(c_n)_{n\in \N}$ in $C - 1$ which converges to the identity automorphism in the group $\mathrm{Aut}(X,\mu )$ equipped with the weak topology. The sequence $(c_n)_{n\in \N}$ is then asymptotically central in $[\mc{R}^N_X]$, and since $C$ acts freely, the sequence $(c_n)_{n\in \N}$ witnesses that the outer automorphism group of $[\mc{R}^N_X]$ is not Polish. Let $K\cc (Z,\eta )$ be a free ergodic action of $K$, and let $G\cc (X,\mu ) \otimes (Z ,\eta )$ be the diagonal product action where $G$ acts on $(Z,\eta )$ via the quotient map to $K$. This action of $G$ is free and ergodic, and as observed in Remark \ref{rem:overview} above, the construction in the proof of Theorem \ref{thm:extension} yields a sequence $(T_n)_{n\in \N}$ witnessing that the outer automorphism group of $[\mc{R}^G_{X\times Z}]$ is not Polish.
\end{remark}
\end{comment}
\begin{remark}\label{rem:finindex}
A variation of the proof of Theorem \ref{thm:extension} shows that if a group $G$ contains a finite index subgroup $H$ which is stable, then $G$ is stable. Take a p.m.p.\ action $H\cc (X_0 ,\mu _0 )$ such that $H\ltimes (X_0,\mu _0)$ admits a stability sequence $(S_n)_{n\in \N}$, and let $(B_n)_{n\in \N}$ be a sequence of asymptotically invariant sets for the action with $\lim _n \mu _0 (S^0_nB_n\triangle B_n ) \neq 0$. Let $G\cc (X,\mu )$ be the induced action, i.e., $X=X_0\times G/H$, $\mu$ is the product of $\mu _0$ with normalized counting measure, and $g\cdot (x_0 , kH ) = (\rho (g,kH)\cdot x_0 , gkH )$ for $g\in G$, $x_0 \in X_0$, $kH \in G/H$, where $\rho (g,kH)=\sigma (gkH)^{-1}g\sigma (kH)\in H$ and $\sigma :G/H \ra G$ is a section for the projection map $G\ra G/H$ with $\sigma (1H)=1$. Define $T_n \in [G\ltimes (X,\mu )]$ by $T_n (x _0 ,kH ) = \sigma (kH)S_n (x_0) \sigma (kH)^{-1} \in G$, and define $A_n = B_n\times G/H$. Then, using the sequence $(A_n)_{n\in\N}$, an argument similar to the proof of Theorem \ref{thm:extension} shows that $(T_n)_{n\in \N}$ is a stability sequence for $G\ltimes (X,\mu )$.
\end{remark}
\subsection{Proof of Theorem \ref{thm:extension2}} Throughout this subsection we work under the assumption that $1\ra N\ra G \ra K \ra 1$ is a short exact sequence in which $K$ is amenable.
\begin{lemma}\label{lem:LM}
Suppose that $N=LM$, where $L$ and $M$ are commuting normal subgroups of $N$ such that $M$ is amenable and $[N:L]=\infty$. Let $N/M \cc (X,\mu )$ and $N/L\cc (Y,\nu )$ be free p.m.p.\ actions of $N/M$ and $N/L$ respectively. Let $N$ act on $(X,\mu )$ and $(Y,\nu )$ via the quotient maps to $N/M$ and $N/L$ respectively, and let $N\cc (X,\mu )\otimes (Y,\nu )$ be the diagonal product action. Then the translation groupoid $N\ltimes (X,\mu )\otimes (Y,\nu )$ admits a stability sequence.
\end{lemma}
\begin{proof}
Let $C= L\cap M$ and let $M_0=M/C$. The action $M\cc (Y,\nu )$ descends to a free action $M_0 \cc (Y,\nu )$. Since the group $M_0$ is amenable, the equivalence relation $\mc{R}^{M_0}_Y$ is treeable. Therefore, by Theorem 1.1 of \cite{Ki14}, the p.m.p.\ groupoids $C\times M_0\ltimes (Y,\nu )$ and $M\ltimes (Y,\nu )$ are isomorphic. Here, $C\times M_0 \ltimes (Y,\nu )$ is the translation groupoid associated to the action $C\times M_0 \cc (Y,\nu )$, where $C$ acts trivially. Since $M_0$ is amenable and acts freely on $(Y,\nu )$, the groupoid $M_0\ltimes (Y,\nu )$ admits a stability sequence $(S_n)_{n\in\N}$ \cite{JS87}. For each $n\in \N$ define $S_n' \in [C\times M_0\ltimes (Y,\nu )]$ by $S_n'(y)= (1_C ,S_n(y))\in C\times M_0$. Then $(S_n')_{n\in \N}$ is a stability sequence for $C\times M_0\ltimes (Y,\nu )$. The image of this sequence under the above isomorphism is then a stability sequence $(T_n)_{n\in \N}$ for $M\ltimes (Y,\nu )$. Define $\wh{T}_n \in [M\ltimes (X,\mu )\otimes (Y,\nu )]$ by $\wh{T}_n(x,y) = T_n(y)$. Then $(\wh{T}_n)_{n\in \N}$ is a stability sequence for $N\ltimes (X,\mu )\otimes (Y,\nu )$.
\end{proof}
\begin{proof}[Proof of stability from hypothesis {\bf (H1)}]
Let $G/M \cc (X , \mu )$ and $G/L \cc (Y,\nu )$ be free p.m.p.\ actions of $G/M$ and $G/L$ respectively. Then $G$ acts on $(X,\mu )$ and $(Y,\nu )$ via the quotient maps to $G/M$ and $G/L$ respectively. Let $G\cc (X,\mu )\otimes (Y,\nu )$ be the diagonal product of these action. By Lemma \ref{lem:LM}, the groupoid $N\ltimes (X,\mu )\otimes (Y,\nu )$ admits a stability sequence, hence $G$ is stable by Theorem \ref{thm:extension}.
\end{proof}
\begin{proof}[Proof of stability from hypothesis {\bf (H6)}]
Let $N^*= N - \{ 1 \}$. Let $G$ act on $N^*$ by conjugation and consider the corresponding generalized Bernoulli action $G\cc (X,\mu )= ([0,1]^{N^*}, \lambda ^{N^*})$ given by $(g\cdot x )(h)= x(g^{-1}hg)$. Let $(c_n)_{n\in \N}$ and $(d_n)_{n\in \N}$ be sequences witnessing that $N$ is doubly asymptotically commutative. The proof of Proposition 9.8 of \cite{Ke10} shows that the sequence $(c_n)_{n\in \N}$, viewed as a sequence in $[N\ltimes (X,\mu )]$, is a stability sequence for $N\ltimes (X,\mu )$. Theorem \ref{thm:extension} then implies that $G$ is stable.
\end{proof}
We now show that {\bf (H3)} follows from each of the hypotheses {\bf (H2)}, {\bf (H4)}, and {\bf (H5)} (in fact, it can also be shown that {\bf (H3)} follows from {\bf (H6)}, but we will not need this). This is obvious for {\bf (H2)}. Assume now that {\bf (H4)} holds, so that $A\cap C_G(g)$ has finite index in $A$ for all $g\in N$. Then we can find a decreasing sequence $A=A_0\geq A_1 \geq \cdots$ of finite index subgroups of $A$ such that $N = \bigcup _m C_N(A_m)$. Then for all $m\in \N$ the pair $(C_N(A_m), A_m )$ does not have property (T), since $A_m$ is finite index in $A$ and $(N,A)$ does not have property (T). This shows that {\bf (H3)} holds. Finally, assume that {\bf (H5)} holds. After moving to a subsequence of $(c_n)_{n\in \N}$, we may assume that each of the subgroups $A_i = \langle \{ c_n\csuchthat n\geq i \} \rangle$, $i\in \N$, is abelian, so that $N = \bigcup _iC_N(A_i)$, where the union is increasing. Then each of the groups $(C_N(A_i), A_i )$ does not have property (T), since $A_i$ is infinite and $N$ has the Haagerup property. This verifies {\bf (H3)}. It remains to deduce stability of $G$ from {\bf (H3)}.
\begin{proof}[Proof of stability from hypothesis {\bf (H3)}]
We may assume that $N$ is not doubly asymptotically commutative, since otherwise we are done by the proof of stability from {\bf (H6)}. Let $(L_m)_{m\in \N}$ and $(D_m)_{m\in \N}$ be given by {\bf (H3)}. Let $F_0\subseteq F_1\subseteq \cdots$ be an increasing sequence of finite sets which exhaust $G$.
\begin{claim}\label{claim:sequence}
There exists an increasing sequence $H_0\leq H_1\leq \cdots$, of finitely generated subgroups of $N$ with $N=\bigcup _{m\in \N}H_m$, along with sequences $C_0\geq C_1\geq \cdots$, and $(N_m)_{m \in \N}$ of subgroups of $N$ such that, for all $m\in \N$,
\begin{itemize}
\item[1.] $C_m, H_m \leq N_m$, and $C_m = C_N(H_m) = Z(N_m)$,
\item[2.] The pair $(N_m, C_m )$ does not have relative property \emph{(T)},
\item[3.] $gH_mg^{-1} \leq H_{m+1}$ for all $g\in F_{m+1}$,
\item[4.] $C_m\geq g^{-1}C_{m+1}g$ for all $g\in F_{m+1}$.
\end{itemize}
\end{claim}
\begin{proof}[Proof of Claim \ref{claim:sequence}]
Since $N$ is not doubly asymptotically commutative there exists a finitely generated subgroup $H_0\leq N$ such that $C_N(H_0)$ is abelian. After moving to a subsequence of $(L_m)_{m\in \N}$ if necessary we may assume that $H_0\leq L_0$. We may then extend $H_0$ to a sequence $H_0\leq H_1\leq H_2\leq \cdots$, of finitely generated subgroups of $N$ with $N= \bigcup _{m\in\N} H_m$, and $H_m\leq L_m$ for all $m\in \N$. After moving to a further subsequence if necessary we may assume that property 3.\ is satisfied for all $m\in \N$. Let $C_m = C_N(H_m)$, so that $C_m$ is an abelian group containing $D_m$. Then the pair $(H_mC_m , D_m )$ does not have property (T), since $(H_mD_m,D_m )$ does not have property (T) and $(H_mC_m)/(H_mD_m)$ is amenable \cite{Jo05}. It follows that $(H_mC_m, C_m )$ does not have property (T). Let $N_m = H_mC_m$. Then $C_m\leq Z(N_m)\leq C_H(N_m)\leq C_H(H_m)= C_m$, so that both 1.\ and 2.\ are satisfied, and 4.\ follows from 1.\ and 3. \qedhere[Claim \ref{claim:sequence}]
\end{proof}
Fix an increasing sequence $S_0\subseteq S_1\subseteq \cdots$ of finite sets such that $S_m$ generates $H_m$ for all $m\geq 0$. Note that if $\pi _m$ is an irreducible unitary representation of $N_m$, then Schur's Lemma implies that $\pi _m (c)$ is a scalar multiple of the identity for all $c\in C_m$, since $C_m=Z(N_m)$. Property 4.\ of the claim then shows that $\pi _m (g^{-1}cg)$ is also a scalar for all $c\in C_{n}$, $g\in F_{n}$, $n>m$. The following is based on Lemma 4.1 of \cite{Ki13a}.
\begin{lemma}\label{lem:avg}
There exist sequences $(\pi _m )_{m\in \N}$, $(\xi _m )_{m\in \N}$, and $(c_m)_{m\in \N}$, where for each $m\in \N$, $\pi _m$ is an irreducible unitary representation of $N_{m}$, $\xi _m \in \Es{H}_{\pi _m}$ is a unit vector, and $c_m$ is an element of $C_{m}$, such that
\begin{enumerate}
\item[(i)] $\| \pi _m (s)\xi _m - \xi _m \| < 2^{-m}$ for all $s\in S_m$,
\item[(ii)] $| \pi _m (c_m) - 1 | > 1$,
\item[(iii)] $| \pi _j (g^{-1}c_mg) - 1 | < 2^{-m}$ for all $g\in F_m$ and $j<m$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $m\geq 0$ and assume inductively that we have already found $(\pi _j )_{j<m}$, $(\xi _j )_{j<m}$, and $(c_j)_{j<m}$. By compactness there exists a nonempty finite set $P\subseteq C_m$ such that for each $c\in C_m$ there exists some $d\in P$ with $\sup _{j<m}\sup _{g\in F_m} |\pi _j (g^{-1}cg)-\pi _j (g^{-1}dg) | < 2^{-m}$. Since $(N_m,C_m )$ does not have property (T), we may find an irreducible unitary representation $\pi _m$ of $N_m$ which has no nonzero $C_m$-invariant vector, satisfying $\sup _{d\in P} |\pi _m (d)-1| < 1/3$, along with a unit vector $\xi _m \in \Es{H}_{\pi _m}$, such that $\| \pi _m (s)\xi _m - \xi _m \| < 2^{-m}$ for all $s\in S_m$. Since $\pi _m (C_m)$ is nontrivial, there exists some $d_m \in C_m$ satisfying $|\pi _m (d_m) - 1 | > 4/3$. By our choice of $P$ there exists $d\in P$ such that $\sup _{j<m}\sup _{g\in F_m} | \pi _j (g^{-1}d_mg)-\pi _j(g^{-1}dg) | < 2^{-m}$. Let $c_m=d^{-1}d_m$. Then $\sup _{j<m}\sup _{g\in F_m}|\pi _j (g^{-1}c_mg) - 1 |< 2^{-m}$, and
\begin{align*}
|\pi _m (c_m) - 1 | &= |\pi _m (d_m) - \pi _m (d) | \geq |\pi _m (d_m) -1 | - | \pi _m (d) - 1| > 4/3 - 1/3 = 1 .\qedhere
\end{align*}
\end{proof}
For each $m\in \N$ let $\sigma _m: G/N_m \ra G$ be a section for the projection map $G\ra G/N_m$ with $\sigma _m (1N_m)=1$, and let $\rho _m : G\times G/N_m\ra N_m$ be the corresponding Schreier cocycle, $\rho _m (g,hN_m) = \sigma _m(ghN_m)^{-1}g\sigma _m(hN_m)$. Let $\wt{\pi}_m = \mathrm{Ind}_{N_m}^G (\pi _m )$ be the induced representation, i.e., $\wt{\pi}_m$ is the representation of $G$ on $\Es{H}_m = \Es{H}_{\pi _m}\otimes \ell ^2(G/N_m)$ given by $\wt{\pi}_m (g)(\xi \otimes \delta _{hN_m} ) = \pi _m(\rho _m(g,k))(\xi )\otimes \delta _{ghN_m}$. Let $\wt{\xi} _m= \xi _m \otimes \delta _{1N_m}$, so that $\wt{\xi} _m\in \Es{H}_m$ is a unit vector satisfying $\| \wt{\pi} _m (s)\wt{\xi} _m - \wt{\xi} _m \| = \| \pi _m (s)\xi _m - \xi _m \| < 2^{-m}$ for all $s\in S_m$ by property (i). By property (ii), for each $m\in \N$ we have $\| \wt{\pi}_m(c_m)\wt{\xi}_m - \wt{\xi}_m \| = \| (\pi _m (c_m) -1 )\xi _m \| = |\pi _m (c_m)-1 | >1$. Moreover, it follows from property (iii) that $\lim _{m\ra\infty} \| \wt{\pi }_j (c_m)\eta - \eta \| = 0$ for all $j\in \N$ and $\eta \in \Es{H}_j$. Let $G\cc (\Omega , \nu )$ denote the Gaussian action associated to the representation $\bigoplus _m \wt{\pi}_m$. As in the proof of Theorem 1.1.(i) of \cite{Ki13a}, we conclude that the sequence $(c_m)_{m\in \N}$, viewed as a sequence in the full group $[N\ltimes (\Omega , \nu )]$, is a stability sequence for $N\ltimes (\Omega ,\nu )$. We can now apply Theorem \ref{thm:extension} to conclude that $G$ is stable.
\end{proof}
The following proposition shows that, aside from the relative property (T) condition, the hypothesis {\bf (H4)} has a natural expression in terms of a conjugation invariant mean on $G$ which concentrates on $A$, as long as we assume that $A$ is finitely generated.
\begin{proposition}
Let $G$ be a countable group. Then the following are equivalent:
\begin{enumerate}
\item There exists an atomless conjugation invariant mean $\bm{m}$ on $G$ and an infinite finitely generated abelian group $A\leq G$ with $\bm{m}(A)>0$.
\item There exists an infinite finitely generated abelian group $A\leq G$ with $[G: \mathrm{comm}_G(A)]<\infty$ such that $\mathrm{comm}_G(A)/N$ is amenable, where $N$ is the kernel of the modular homomorphism from $\mathrm{comm}_G(A)$ into the abstract commensurator of $A$.
\end{enumerate}
Furthermore, the statement $(1')$, obtained from $(1)$ by replacing \emph{"$\bm{m}(A)>0$"} with \emph{"$\bm{m}(A)=1$"}, is equivalent to the statement $(2')$, obtained from $(2)$ by replacing \emph{"$[G:\mathrm{comm}_G(A)]<\infty$"} with \emph{"$G=\mathrm{comm}_G(A)$"}.
\end{proposition}
\begin{proof}
$(2)\Ra (1)$: Let $(A_n)_{n\in \N}$ enumerate the finite index subgroups of $A$, and for each $n\in \N$ choose a nonidentity element $a_n \in \bigcap _{i\leq n}A_i$. Let $\bm{m}_0$ be an accumulation point of $(\delta _{a_n})_{n\in \N}$ in the space of means on $G$. Then $\bm{m}_0$ concentrates on $A$, and is invariant under conjugation by $N$ since each $g\in N$ commutes with $a_n$ for cofinitely many $n\in \N$. Let $G_0 = \mathrm{comm}_G(A)$ and let $\bm{m}_{G_0/N}$ be a translation-invariant mean on $G_0/N$. Then the mean $\bm{m}_1 = \int _{gN\in G_0/N}g\bm{m}_0g^{-1}\, d\bm{m}_{G_0/N}$ is invariant under conjugation by $G_0$, and $\bm{m}_1$ concentrates on $A$ since each of the means $g\bm{m}_0g^{-1}$, $g\in G_0$, concentrates on $A$. Finally, the mean $\bm{m} = \frac{1}{[G:G_0]}\sum _{gG_0\in G/G_0} g\bm{m}_1g^{-1}$ is invariant under conjugation by $G$ and satisfies $\bm{m}(A)\geq \frac{1}{[G:G_0]}>0$. This also shows the implication $(2')\Ra (1')$.
$(1)\Ra (2)$: We may assume that $A$ has minimal rank among all $\bm{m}$-non-null finitely generated abelian subgroups of $G$. If $g\in G$ is such that $\bm{m}(gAg^{-1} \cap A ) >0$, then $gAg^{-1}\cap A$ has the same rank as $A$, so $gAg^{-1}\cap A$ has finite index in $A$. Let $G_0=\mathrm{comm}_G(A)$ and suppose toward a contradiction that $[G:G_0]=\infty$. Let $(g_n)_{n\in \N}$ be a sequence with $g_iG_0\neq g_jG_0$ for all $i\neq j$. Then, for all $i\neq j$, the group $g_j^{-1}g_iAg_i^{-1}g_j \cap A$ does not have finite index in $A$, hence $0 = \bm{m}(g_j^{-1}g_iAg_i^{-1}g_j \cap A )=\bm{m} (g_iAg_i^{-1}\cap g_jAg_j^{-1} )$. Therefore for all $n >0$ we have $1\geq \bm{m}(\bigcup _{i< n} g_i A g_i^{-1} ) =\sum _{i<n}\bm{m} (g_iAg_i^{-1})=n\cdot \bm{m}(A)$, a contradiction, since $\bm{m}(A)>0$. This shows that $[G:G_0]<\infty$.
Let $N = \{ g\in G_0 \csuchthat [A:A\cap C_G(g)]<\infty \}$ be the kernel of the modular homomorphism $G_0\ra \mathrm{comm}(A)$, and let $\varphi : G_0 \ra G_0/N$ be the projection to the group $G_0/N$. Suppose toward a contradiction that $\varphi (G_0)$ is nonamenable. For a subgroup $B\leq A$ let $N(B) = \{ g \in G_0\csuchthat [B:B\cap C_G(g)] <\infty \}$. Then $N(B)$ is a subgroup of $G_0$ with $N\leq N(B)$ for all $B\leq A$. Since $A$ is a finitely generated abelian group it satisfies the maximal condition on subgroups. Let $B_0\leq A$ be a maximal subgroup from the collection $\{ B\leq A\csuchthat \varphi (N(B)) \text{ is nonamenable}\}$; this collection is nonempty since it contains the trivial group by hypothesis. Since $N(A) = N$, the group $B_0$ has infinite index in $A$, and hence $\bm{m}(B_0)=0$. Observe that for any $a\in A - B_0$, if we let $B_1 = \langle a , B_0 \rangle$, then the group $\varphi (N(B_1))$ is amenable by maximality of $B_0$, and we have $C_{N(B_0)}(a) \leq N(B_1)$, so the group $\varphi (C_{N(B_0)}(a))$ is amenable. Let $Y$ be the saturation of $A - B_0$ under conjugation by $N(B_0)$. Then $\varphi (C_{N(B_0)}(y))$ is amenable for all $y\in Y$. It follows that for all $y\in Y$, the translation action $C_{N(B_0)}(y)\cc G_0/N$ is amenable. In addition, the conjugation action $N(B_0)\cc Y$ is amenable since $\bm{m}(Y)>0$. It follows (using Theorem \ref{thm:isoperim} for example) that the translation action $N(B_0)\cc G_0/N$ is amenable, and hence the group $\varphi (N(B_0))$ is amenable, a contradiction.
For $(1')\Ra (2')$, we take $A$ to have minimal rank among all $\bm{m}$-conull finitely generated abelian subgroups of $G$. Then for all $g\in G$ the group $gAg^{-1}\cap A$ is $\bm{m}$-conull, so $gAg^{-1}\cap A$ has finite index in $A$, and therefore $G = \mathrm{comm}_G(A)$. The rest proceeds as above.
\end{proof}
\subsection{Proof of Theorem \ref{thm:linearstable}}
\begin{proof}[Proof of Theorem \ref{thm:linearstable}]
(1)$\Ra$(2): Suppose first that $G$ is stable as witnessed by the free ergodic action $G\cc (X, \mu )$. Then $G\ltimes (X,\mu )$ admits a stability sequence $(T_n)_{n\in \N}$ \cite{JS87}. Let $\mc{U}$ be a non-principal ultrafilter on $\N$ and for $D\subseteq G$ let $\bm{m} (D) = \lim _{n\ra\mc{U} } \mu ( \{ x\in X\csuchthat T_n(x)\in D \} )$. Then $\bm{m}$ is a conjugation invariant mean on $G$, so $\bm{m}(\mathscr{I}(G))=1$ by Theorem \ref{thm:I(G)}.xiii. We may therefore assume without loss of generality that $(T_n)_{n\in \N}$ is contained in $[\mathscr{I}(G)\ltimes (X,\mu )]$. It follows that every subgroup of $G$ containing $\mathscr{I}(G)$ is stable. Let $N=C_G(\mathscr{I}(G))\mathscr{I}(G)$ and note that $\mathscr{I}(N) = \mathscr{I}(G)$ since $G/N$ is amenable. If $[N: C_G(\mathscr{I}(G))] = \infty$ then the image of $\mathscr{I}(G)$ in the amenable group $G/C_G(\mathscr{I}(G))$ is infinite, so the pair $(G/C_G(\mathscr{I}(G)) , N/C_G(\mathscr{I}(G)))$ does not have property (T), and hence $(G,\mathscr{I}(G))$ does not have property (T). We may therefore assume that $[N: C_G(\mathscr{I}(G))]<\infty$. This implies that the group $Z(N) = C_G(\mathscr{I}(G))\cap \mathscr{I}(G)$ has finite index in $\mathscr{I}(G)= \mathscr{I}(N)$. Suppose toward a contradiction that $(G,\mathscr{I}(G))$ has property (T). Since the group $G/N$ is amenable, the pair $(N, \mathscr{I}(G))$ has property (T), hence $(N,Z(N))$ has property (T) \cite{Jo05}. The group $N$ is stable since $\mathscr{I}(G)\leq N$, so Theorem 1.1.(2) of \cite{Ki13a} implies that $N/Z(N)$ is stable, and in particular $N/Z(N)$ is inner amenable \cite{JS87}. This is a contradiction since, by Theorem \ref{thm:I(G)}, every conjugation invariant mean on $N/Z(N)$ concentrates on the group $\mathscr{I}(N/Z(N)) = \mathscr{I}(N)/Z(N)$, which is finite. We conclude that the pair $(G , \mathscr{I}(G) )$ does not have property (T).
(2)$\Ra$(1): Assume that $(G,\mathscr{I}(G))$ does not have property (T). It follows that $\mathscr{I}(G)$ is infinite. Let $L= C_G(\mathscr{I}(G))$, let $M=\mathscr{I}(G)$, and let $N=LM$ so that $Z(N) = L\cap M$. The group $G/N$ is amenable by Theorem \ref{thm:I(G)}. If $[N:L]=\infty$ then hypothesis {\bf (H1)} holds, so $G$ is stable by Theorem \ref{thm:extension2}. If $[N:L]<\infty$ then $(N,Z(N))$ does not have property (T), since $(N,\mathscr{I}(G))$ does not have property (T) and $[\mathscr{I}(G):Z(N) ] =[N:L]$. This shows that hypothesis {\bf (H2)} holds, so $G$ is stable by Theorem \ref{thm:extension2}.
\end{proof}
\begin{remark}\label{rem:final}
It follows from Theorems \ref{thm:I(G)}, \ref{thm:linear}, and \ref{thm:linearstable}, that a linear group $G$ is inner amenable but not stable if and only if the group $\mathscr{I}(G)$ is infinite and is finite index over its center $C = C_G(\mathscr{I}(G))\cap \mathscr{I}(G)$, and the the pair $( C_G(\mathscr{I}(G)), C)$ has property (T).
\end{remark}
\subsection{Groups of piecewise projective homeomorphisms} Justin Moore has observed that an adaptation of the arguments of Brin-Squier \cite{BS85} and Monod \cite{Mo13} shows the following
\begin{lemma}\label{lem:Justin}
Let $G$ be a countable subgroup of $H(\R )$. Then the second derived subgroup $G''$ is either abelian or doubly asymptotically commutative.
\end{lemma}
\begin{proof} Assume first that $G$ is finitely generated. Then the set $U = \bm{\mathrm{P}}^1\setminus \mathrm{fix}(G)$ has finitely many connected components, each of which is an open interval. If $V\subseteq U$ is a union of a subset of these connected components then let $\varphi _{V}:G \ra H(\R )$ denote the homomorphism which sends $g\in G$ to the map $\varphi _V(g)$ which coincides with $g$ on $V$ and which is the identity elsewhere. In what follows, we fix an orientation of $\bm{\mathrm{P}}^1\setminus \{ \infty \}$.
\begin{claim}\label{claim:push}
For any compact subset $K\subseteq U$ there exists an element $g\in G$ such that $g(K)\cap K =\emptyset$.
\end{claim}
\begin{proof}[Proof of Claim \ref{claim:push}]
By induction on the number $n$ of connected components of $U$. If $n=1$ then it suffices to show that for any $p\in U$ we have $\sup _{g\in G}g(p) = \sup U$. Suppose otherwise and let $q= \sup _{g\in G}g(p) < \sup U$. Then $q\in U$, so we may find some $g\in G$ with $g(q)\neq q$, and after replacing $g$ by $g^{-1}$ if necessary we may assume that $g(q)>q$. If $(g_n)_{n\in \N}$ is any sequence in $G$ with $g_n(p)\ra q$ then $q\geq g(g_n(p))\ra g(q)>q$, a contradiction. Assume now that $U$ has $n+1$ connected components and fix one such component $V$. After making $K$ larger if necessary we may assume that $K\cap V$ is a closed interval. Apply the base of the induction to the group $\varphi _{V} (G)$ to obtain a group element $h\in G$ with $h(K\cap V)$ disjoint from $K\cap V$. Since $K\cap V$ is an interval, after replacing $h$ by $h^{-1}$ if necessary, this means that $\inf h(K\cap V) > \sup (K\cap V)$. Let $L = (K\setminus V) \cup h(K\setminus V)$. Then $L$ is a compact subset of $U\setminus V$, so we may apply the induction hypothesis to the group $\varphi _{U\setminus V}(G)$ to obtain a group element $f\in G$ satisfying $f(L)\cap L =\emptyset$. After replacing $f$ by $f^{-1}$ if necessary we may assume that $f(p)\geq p$, where $p= \inf h(K\cap V)$. Take $g=fh$. Then $\inf g(K\cap V) = f(p)\geq p >\sup (K\cap V)$, so $g(K\cap V)$ is disjoint from $K\cap V$. In addition, $g(K\setminus V)\cap (K\setminus V) =f(h(K\setminus V))\cap (K\setminus V) \subseteq f(L)\cap L = \emptyset$. Since $V$ is $G$-invariant, this shows that $g(K)\cap K=\emptyset$.
\qedhere[Claim \ref{claim:push}]
\end{proof}
Assume that $G''$ is nonabelian and fix two non-commuting elements $c_0,d_0\in G''$. As shown in Lemma 14 of \cite{Mo13}, the closure of the support of any element of $G''$ is a compact subset of $U$. Fix a finite subset $Q\subseteq G''$ and let $K$ be the union of the closures of the supports of all elements of $Q\cup \{ c_0,d_0 \}$. Apply the claim to find an element $g\in G$ with $g(K)\cap K=\emptyset$. Let $c=gc_0g^{-1}$ and let $d=gd_0g^{-1}$ so that $c,d\in G''$ and $cd\neq dc$. Then $c$ and $d$ both commute with each element of $Q$ since the support of $c$ and of $d$ are disjoint from the support of each element of $Q$. This shows that $G''$ is doubly asymptotically commutative.
When $G$ is not finitely generated we may write $G$ as an increasing union $G=\bigcup _n G_n$ with each $G_n$ finitely generated. Then $G''=\bigcup _n G_n''$, so if $G''$ is nonabelian then $G_n''$ is nonabelian for all large enough $n$. Now note that double asymptotic commutativity is preserved by directed unions.
\qedhere
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:HR}]
By \cite{Mo13}, $H(\R )$ is torsionfree, so any nontrivial amenable subgroup of $H(\R )$ is infinite, hence stable. If $G$ is a nonamenable subgroup of $H(\R )$ then $G''$ is nonabelian, hence doubly asymptotically commutative by Lemma \ref{lem:Justin}. Hypothesis {\bf (H6)} holds, using the short exact sequence $1\ra G''\ra G\ra G/G'' \ra 1$, hence Theorem \ref{thm:extension2} shows that $G$ is stable.
\end{proof}
\noindent{\bf Acknowledgements:} The author warmly thanks Justin Moore for explaining Lemma \ref{lem:Justin}, Adrian Ioana for helpful conversations and for allowing the inclusion of the joint Theorem \ref{thm:superrigid1}, and Yoshikata Kida for pointing out {\bf (H4)} from Theorem \ref{thm:extension2} and Corollary \ref{cor:gBS} and for a number of insightful comments on an earlier draft of this paper which helped improve the statement of Theorem \ref{thm:extension2}. The author was supported by NSF grant DMS 1303921.
|
\section{Introduction}
An unexpected discovery at RHIC/BNL in $D+Au$ reactions at $\sqrt{s}=200$~AGeV~\cite{Adare:2013piz} and
at LHC/CERN in $\sqrt{s}=5.02$~ATeV $p+Pb$ reactions~\cite{CMS:2012qk}
is the large magnitude of mid-rapidity
azimuthal anisotropy moments, $v_n{2}(k_T, \eta=0)$, that are remarkably
similar to those observed previously in non-central $Au+Au$
~\cite{Adams:2005dq
and in $Pb+Pb$
~\cite{Aamodt:2011by}
reactions.
See especially the preliminary ATLAS results in Fig. 24 of ref.~\cite{ATLASvn}
and the QM14 ATLAS talk by J. Jia in these proceedings~\cite{ATLASpAv1}.
In addition, the Beam Energy Scan (BES) at RHIC~\cite{Adamczyk:2013gw}
revealed a near $\sqrt{s}$ independence from 8~AGeV to 2.76~ATeV of
$v_n(p_T;s)$ in $A+A$ at fixed centrality that was also unexpected.
In high energy $A+A$, the systematics of $v_n\{2\ell\}(k_T,\eta=0)$ data
have been interpreted as
possible evidence for the near ``perfect fluidity'' of the
strongly-coupled Quark Gluon Plasmas (sQGP) produced in such
reactions~\cite{Romatschke:2007mq,
Heinz:2013th}. However, the recent observation of similar $v_n$ arising
from much smaller transverse size $p(D)+A$ systems and also the near beam energy
independence of the moments observed in the Beam Energy Scan (BES)~\cite{Adamczyk:2013gw}
from 7.7~AGeV to 2.76~ATeV in $A+A$ have posed a problem for the
perfect fluid interpretation because near inviscid hydrodynamics is
not expected to apply in space-time regions where the local
temperature field falls below the confinement temperature, $T(x,t) <T_c\sim
160 $~MeV. In that Hadron Resonance Gas (HRG) ``corona'' region, the
viscosity to entropy ratio is predicted to grow rapidly with
decreasing temperature~\cite{Danielewicz:1984ww} and the corona volume
fraction must increase and the volume of the
perfect fluid ``core'' with $T>T_c$ must decrease
when either the projectile atomic
number $A$ decreases or the center-of-mass (CM) energy
$\sqrt{s}$ decreases .
While hydrodynamic equations have been shown to be
{\em sufficient} to describe $p(D)+A$ data
with particular assumptions about initial and freeze-out boundary
conditions~\cite{Bozek:2011if}, its {\em necessity} as a
unique interpretation of the data is not guaranteed.
This point was underlined recently using a specific initial-state
saturation model~\cite{Dusling:2013oia} that was shown to be able to fit $p(D)+A$
correlation $v_{2n}$ even moments data without final-state interactions.
That saturation model has also been used in~\cite{Gale:2012rq} to specify
initial conditions for perfect fluid hydrodynamics in $A+A$.
However, in $p+A$ the transverse spatial structure
of initial conditions is not as well-controlled
because the gluon saturation scale scale, $Q_s$, is small and its fluctutations
over the transverse plane on sub-nucleon scales are more model dependent than in $A+A$.
The near independence of $v_n$ moments on beam energy observed in the BES~\cite{Adamczyk:2013gw}
at RHIC from 7.7 AGeV to 2.76~ATeV pose another serious challenges to
the uniqueness of the perfect fluid interpretations of the data
because previous hybrid fluid-HRG modeling~\cite{Teaney:2000cw} predicted a natural systematic reduction of the moments due to the increasing HRG corona fraction with decreasing beam energy.
The HRG corona fraction is expected to dilute flow signatures from the perfect fluid QGP core flow at lower energies.
The BES~\cite{Adamczyk:2013gw} data also a pose a challenge to color glass condensate (CGC)
gluon saturation models~\cite{Kharzeev:2004bw} used to specify initial conditions for hydrodynamic
flow predictions in $A+A$. This is because $Q_s^2$ is predicted to decrease with
$\log (s)$, and, thus, gluon saturation-dominated ``central rapidity region'' gluon fusion dynamics must switch
over into valence quark-diquark dominated ``fragmentation region'' inelastic
dynamics involving fragmentation of multiple quark-diquark beam jets.
In our GLVB paper~\cite{GLVB14} we explored the possibility that initial-state gluon bremsstrahlung,
sourced by Color Scintillating Arrays (CSA) of colored beam jet antennae, could partially
account for the above puzzling systematics of azimuthal harmonics. Non-abelian bremsstrahlung
is intrinsically azimuthally anisotropic. The pQCD-based GLVB model extends
the first order in opacity $\chi=1$ Gunion-Bertsch~\cite{Gunion:1981qs}
(GB) perturbative QCD bremsstrahlung
to all orders in opacity, $e^{-\chi}\sum_{n=1}^\infty\chi^n/n! \cdots$~\cite{Gyulassy:2000er},
Vitev-Gunion-Bertsch (VGB)~\cite{Vitev:2007ve} multiple interaction pQCD bremsstrahlung
for applications to $B+A$ nuclear collisions.
We show that VGB bremsstrahlung naturally leads on an
event-by-event basis to a hierarchy of
non trivial azimuthal asymmetry moments remarkably similar to that observed in
$p+A$ and peripheral $A+A$ at fixed $dN/d\eta$~\cite{Chatrchyan:2012wg}
\section{Results}
The non-abelian bremsstrahlung Gunion-Bertsch (GB)
formula~\cite{Gunion:1981qs,Vitev:2007ve} for the soft gluon radiation single inclusive
distribution for a triggered beam jet recoil momentum ${\bf q}=(q,\psi)$ is
\begin{eqnarray}
\frac{dN_g^1}{d\eta d^2 {\bf k}d^2{\bf q}} &= & \frac{C_R \alpha_s}{\pi^2}
\frac{\mu^2}{\pi(q^2+\mu^2)^2}
\frac{{\bf q}^2}{{\bf k}^2 ({\bf k}-{\bf q})^2}
\;\; . \label{GB1} \end{eqnarray}
Here, the
parton scattering elastic cross section is assumed to be
$d\sigma_{0}/d^2{\bf q}=\sigma_{0} \mu^2/\pi(q^2+\mu^2)^2$. The produced gluon has rapidity $\eta$ and
transverse momentum ${\bf k}=(k,\phi)$. Note especially that
the azimuthally asymmetric angular dependence has the simple form,
$dN_g= F_{kq}/(A_{kq} - \cos(\phi-\psi))$ of
the radiated $\phi$ relative to the reaction plane $\psi$ angles arising from
basic non-abelian interference effects. Note also the uniform rapidity-even, $\eta \approx \log (x E/k)$,
distribution of non-abelian bremsstrahlung.
In $p+A$ multiple target beam jets generally transform that uniform $\eta$ dependence into
a trapezoidal one, as discussed in~\cite{GLVB14}.
The GB azimuthal harmonics can then be analytically evaluated from
\begin{eqnarray}
\hspace{-0.2in}
v_n^{GB}(k,q,\psi)
&=& \int\frac{ d\phi}{2\pi} \cos(n \phi)\; \frac{(A_{kq}^2-1)^{1/2}}{ A_{kq} - \cos(\phi-\psi)}
\;\;=\;\; \cos[n \psi] \;(\; v_1^{GB}(k,q)\;)^n \; ,\label{gbsc}\\
v_1^{GB}(k,q) &=& (A_{kq}-\sqrt{A^2_{kq}-1})
\; \; , \label{vn1}
\end{eqnarray}
where (see \cite{GLVB14}) $A_{kq}= (k^2+q^2+\mu^2)/(2 k \, q) \ge 1$ implies that all harmonics
are peaked near $k\sim q$, vanish at $k=0$, and slowly decrease toward zero for $k\gg q$. In addition, the analytic single color antenna GB gluon harmonics
obey an approximate power law scaling with respect to the harmonic $n$ number:
\begin{eqnarray}
[v_n^{GB}(k,q,0)]^{1/n} &=& \;[v_m^{GB}(k,q,0)]^{1/m}
\; \; ,\label{vnscaling}\end{eqnarray}
that is similar to the scaling observed by ALICE, CMS and
ATLAS~\cite{Aamodt:2011b
} at LHC
and similar to perfect fluid harmonic scaling for the higher $n\ge 3$ moments
dominated by purely geometric fluctuations. We note also that unlike the low order CGC azimuthal harmonics,
the GB bremsstrahlung harmonics are non-vanishing and scale for all odd as well as even
moments $n$. We illustrate in Fig.~\ref{fig-4new}a the main features of azimuthal harmonics from
a single beam jet bremsstrahlung.
\begin{figure}[!tbh]
\centerline{\includegraphics[width=3.in]{GLVB14-fig4.pdf}
\hspace{0.1in} \includegraphics[width=3.5in]{GLVB14-fig5.pdf}
\caption{(Color online) (a) [left] Approximate $1/n$ power scaling
of $q$ averaged $ \langle v_n^{GB}(k,q,0)\rangle^{1/n}$ for a fixed reaction
plane $\psi=0$ from a single GB color antenna for all even and odd moments
is seen in the kinematic region
$k^2 < \langle q^2\rangle\equiv M\mu^2$. This scaling breaks down for for
$k > \sqrt{M}\mu$
because in the $\mu=0$ limit bremsstrahlung stricly vanish for $k>q$. (b) [right]
Schematic diagram corresponding
to coherent bremsstrahlung from the projectile beam jet at impact parameter ${\bf b}$.
The azimuthal distribution
is enhanced for transverse momenta ${\bf k}$ near the total accumulated
momentum transfer to the projectile
${\bf Q}_0\equiv \sum_{a} {\bf Q}_a$ from the $a=1,\cdots, M$
clusters of recoiling target beam jets. In addition, the bremsstrahlung is enhanced near
${\bf k}=-{\bf Q}_a$ from each incoherent recoiling target cluster
separated by $\Delta R_{ab} \sim 1/k$
in the transverse plane.
}
\label{fig-4new}
\end{figure}
In a given $p+A$ event a projectile
nucleon penetrates through a target nucleus $A$ at impact parameter
${\bf b}$, producing one projectile beam jet with rapidity $Y_P>0$.
There are $N\sim A^{1/3} <A$ target beam jets with $Y_T<0$ produced
with transverse coordinates ${\bf R}_i$ distributed within a distance $\sim 1/\mu$
from the impact parameter. The $N$ target dipoles naturally cluster
near the projectile impact parameter ${\bf{b}}$ as illustrated in
Fig.~(1b).
In a specific event, there are however only $1\le M\le N$
overlapping clusters that can radiate coherently toward
the negative rapidity $\eta<0$ hemisphere (see \cite{GLVB14}).
Incoherence of target clusters
bremsstrahlung is controlled by the transverse resolution scale
with $|{\bf R}_i-{\bf R}_j|> 1/k$.
Let $ {\bf Q}_a=\sum_{i\in I_a} {\bf q}_{i}$ denote the cumulative momentum transfer to the
projectile from target cluster $a$.
The total single inclusive bremsstrahlung distribution into a particular
mode $({\bf k}_1)$ has the form
\begin{equation}
dN^{M,N}= dN^{N}_{P}(\eta,{\bf k}_1 ; {\bf Q}_P)
+dN^{M,N}_{T}(\eta,{\bf k}_1;\{{\bf Q}_{a}\})
= \sum_{a=0}^{M}
\frac{ B_{k_1 Q_a}}{ ({\bf k}_1+{\bf Q}_a)^2+\mu_a^2 }
\;\; ,
\label{GBPTM}
\end{equation}
where we define
${\bf Q}_0 = -\sum_a {\bf Q}_a$ to
include the projectile beam jet contribution into the summation over target clusters. Note that for a semi exclusive event with all $M$ target recoil
momenta $Q_a$ and their azimuthal orientation $\psi_a$ determined, the bremsstrahlung
radiation is peaked near the $M+1$ cumulative momenta. However, averaging over all reaction planes forces all single particle $v_n{1}$ to vanish on the average.
Only 2 or higher particle correlation can reveal the intrinsic azimuthal anisotropy correlations above. Fortunately, in this CSA bremsstrahlung model
all $2\ell$ relative
azimuthal harmonics can also be evaluated analytically (see ~\cite{GLVB14}).
In the ``mean recoil'' approximation $Q\approx \bar{Q}$,
we find that a single GB antenna satisfies the generalized power scaling
law in case that subsets of the $2\ell$ gluons
have identical momenta. Suppose there are $1\le L\le 2\ell$
distinct momenta $K_r$ with $r=1,\cdots,L$ such
$m_r$ of the $2\ell$ gluons have momenta equal
to a particular value $K_r$ such that $\sum_{r=1}^{L} m_r= 2\ell$.
In this case
$
v_n^{M=1}\{2\ell\}(k_1,\cdots,k_{2\ell};\bar{Q} )
\approx \prod_{r=1}^L (v_n^{GB}(K_r,\bar{Q}))^{m_r}
= \prod_{r=1}^L
(v_1^{GB}(K_r,\bar{Q}))^{n m_r}$.
The approximate factorization and power scaling of azimuthal harmonics
from CSA coherent state non-abelian bremsstrahlung is similar
to ``perfect fluid hydrodynamic collective flow'' factorization and scaling,
but in this case no assumption about local equilibration or minimal viscosity
is necessary.
\section{Conclusions}
In this talk we summarized from Ref.~\cite{GLVB14}
some of the remarkable azimuthal correlation properties
of beam jet non-abelian bremsstrahlung even at the lowest order of perturbative QCD level
using the VGB generalization~\cite{Vitev:2007ve}
of GB~\cite{Gunion:1981qs} bremsstrahlung to all orders in opacity in $p+A$.
Of course, higher order and especially high gluon occupation number effects~\cite{Kharzeev:2004bw}
could modify the intricate initial-state chromo wave interference patterns.
However, the main lesson from this study is that in $p+A$ initial-state
wave interference phenomena may well dominate over any final-state
dynamics but appear as if ``perfect fluid'' or ``conformal holographic'' descriptions\cite{Basar:2013hea} were applicable on sub nucleon transverse scales.
Our analysis
shows that long range in $\eta$ multi-gluon, $1/n$ power law scaling,
azimuthal multipole
cumulant signatures are not unique to final-state perfect fluid
and dual AdS shock wave flows but can arise also
naturally from perturbative QCD features of initial state
bremsstrahlung from Color Scintillating Arrays of multiple beam jets.
A possible way to help discriminate between
initial-state interference harmonics and final-state flow harmonics
may be through the study of rapidity dependence of multi-gluon azimuthal harmonics
as discussed in~\cite{GLVB14}.
\section{Acknowledgements}
MG acknowledges support from the
US-DOE DE-FG02-93ER40764, DE-AC02-05CH1123. PL, TB, and MG
acknowledge support from Hungarian OTKA grants K104260, NK106119, and
NIH TET\_12\_CN-1-2012-0016. IV was supported in part by the US Department of Energy,
Office of Science, Office of Nuclear Physics.
|
\section{Introduction}
Relays are fundamental building blocks of wireless networks. One recently proposed channel model for relay networks is the \gls{mwrc}. Such a model applies to many communication architectures like the communication of several ground stations over a satellite, or wireless board-to-board communication in highly adaptive computing \cite{FettDATE13} where multiple chips exchange data with the help of another chip acting as relay. The \gls{mwrc} was first introduced in \cite{Gunduz2013}, where all users in the cluster send a message and are interested in decoding the messages of all other users in the cluster.
In \cite{Ong2010} the common-rate capacity of the \gls{awgn} \gls{mwrc} with full message exchange is given and it is shown that for three and more users this capacity is achieved by \gls{df} for \glspl{snr} below \unit[0]{dB} and compute-and-forward otherwise. In \cite{Chaaban2013} a constant gap approximation of the capacity region of the Gaussian 3-user \gls{mwrc} with full message exchange is given.
Besides spectral efficiency, another key performance metric in modern and future 5G wireless networks is \gls{ee}. From a mathematical standpoint, one well-established definition of the \gls{ee} of a communication system is the ratio between the system capacity or achievable rate and the total consumed power \cite{Miao2011,Isheden2012}. With this definition, the \gls{ee} is measured in bit/Joule. Previous results on \gls{ee} in relay systems mainly focus on regular \gls{af} or \gls{df} schemes and do not consider the \gls{mwrc}. In \cite{Zhou2011} the optimal placement of relays in cellular networks is investigated and is seen to provide power-saving gains. \cite{ZapTWC13} considers the bit/Joule definition of \gls{ee} and devises energy-efficient power control algorithms in interference networks. A cooperative approach is considered in \cite{ZapTSP13}, where a \gls{mimo} \gls{af} relay-assisted system is considered.
In this paper a 3-way relay channel is considered and both spectral and energy efficiency are analyzed and optimized.
In contrast to most other works on \glspl{mwrc}, we focus on a partial message exchange where each message is only destined for one receiver and, also, not every user sends a message to each other user. This makes it necessary to deal with interference at the receivers which complicates the analysis. However, it might also result in higher achievable rates due to less decoding constraints.
The contributions of the paper can be summarized as follows: 1) achievable sum rate expressions are derived for the \gls{af}, \gls{df}, and \gls{nnc} relaying protocols; 2) an outer bound for the capacity of the 3-way relay channel is derived and used for benchmarking purposes; 3) building on the derived achievable sum rate expressions, two algorithms for energy efficiency optimization are provided to jointly allocate the users' and the relays transmit powers.
We define the function $C(x) = \log_2(1+x)$ for $x\ge 0$.
\section{System Model}
We consider the symmetric 3-user \gls{siso} \gls{mwrc} with circular (i.e.,
partial) message exchange, \gls{awgn}, no direct user-to-user links, and full-duplex
transmission. The users are denoted as node 1 to 3 and the relay is node 0. We
define the set of all users as $\mathcal K = \{ 1, 2, 3 \}$ and the set of all
nodes as $\mathcal K_0 = \mathcal K \cup \{ 0 \}$.
The 3-user \gls{mwrc} consists of an uplink channel
$Y_0 = \sum_{k\in\mathcal K} X_i + Z_0$,
and downlink channels
$Y_k = X_0 + Z_k,\quad k\in\mathcal K$,
where $X_k$ and $Y_k$ are the complex valued channel input and output at node
$k\in\mathcal K_0$, respectively, and $Z_k$ is zero mean circularly symmetric
complex Gaussian noise with power $N_0$ at the relay and $N$ at all other
nodes.
All noise variables are mutually independent and the channel inputs are
\acrlong{iid} over time. All channel inputs have zero mean and an average power
constraint $\expval \abs{X_0}^2 \le P_0$
and $\expval \abs{X_k}^2 \le P$, for $k\in\mathcal K$.
We consider a circular message exchange as illustrated in Fig.~\ref{fig:msgex12} where
user $q(k)$ wants message $m_{k}$ with $q = [ 2, 3, 1 ]$. We also define $l(k)$ as the index of the interfering (i.e., unwanted) message at user $q(k)$ as $l = [3, 1, 2]$.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\begin{scope}[every node/.style=draw,circle]
\node (n1) {1};
\node (n2) at (0:3.5) {2};
\node (n3) at (60:3.5) {3};
\node (nR) at ($1/3*(n1)+1/3*(n2)+1/3*(n3)$) {0};
%
%
%
%
\end{scope}
\foreach \a/\da/{\mathbf d}} \newcommand{\Db}{{\mathbf D} in {1/dotted/dashed,2/dashdotdotted/dotted,3/dashed/dashdotdotted} {
%
%
\begin{scope}[thick]
\draw [\da,latex-,relative,out=350,in=190] (nR) to (n\a);
\draw [{\mathbf d}} \newcommand{\Db}{{\mathbf D},-latex,relative,out=10,in=170] (nR) to (n\a);
\end{scope}
}
\end{tikzpicture}
\end{center}
\vspace{-1em}
\caption{Illustration of the system model where node 0 is the relay and nodes 1 to 3 are the users. Messages travel along the different line styles.}
\label{fig:msgex12}
\end{figure}
A $(2^{nR_1}, 2^{nR_2}, 2^{nR_3}, n)$ code for the 3-user \gls{mwrc}
consists of
three message sets $\mathcal M_k = [1:2^{nR_k}]$, one for each user
$k\in\mathcal K$,
three encoders, where encoder $k\in\mathcal K$ assigns a symbol
$x_{ki}(m_k, y_k^{i-1})$ to each message $m_k \in\mathcal M_k$ and
received sequence $y_k^{i-1}$ for $i\in[1:n]$,
a relay encoder that assigns a symbol $x_{0i}(y_0^{i-1})$ to every
past received sequence $y_0^{i-1}$ for $i\in[1:n]$, and
three decoders, where decoder $k\in\mathcal K$ assigns an
estimate $\hat m_{q(k)} \in \mathcal M_{q(k)}$ or an error message $e$
to each pair $(m_k, y_k^n)$.
We assume that the message triple $(M_1, M_2, M_3)$ is uniformly distributed over
$\mathcal M_1 \times \mathcal M_2 \times \mathcal M_3$. The average probability
of error is defined as
$P_e^{(n)} = \Pr\left\{ \hat M_k \neq M_k \text{ for some } k\in\mathcal K \right\}$.
A rate triple $(R_1, R_2, R_3)$ is said to be achievable if there exists a
sequence of $(2^{nR_1}, 2^{nR_2}, 2^{nR_3}, n)$ codes such that
$\lim_{n\rightarrow\infty} P_e^{(n)} = 0$. The capacity region of the 3-user
\gls{mwrc} is the closure of the set of achievable rates. The sum rate is
defined as $R_\Sigma = \max\left\{ R_1 + R_2 + R_3 : \left( R_1, R_2, R_3
\right) \in \mathcal R \right\}$, where $\mathcal R$ is an achievable rate
region. Whenever $\mathcal R$ is the capacity region, we call $R_\Sigma$ the
sum capacity $C_\Sigma$.
\section{Bounds on the Sum Capacity}\label{Sec:AchievableRates}
We start our treatment of the symmetric 3-user \gls{mwrc} by deriving an upper
bound on the sum capacity and then continue with several inner bounds.
\subsection{Outer Bound}
This outer bound consists of the cut set bound in the
uplink and a downlink bound \cite{Yoo2009} that takes the side information
at the receivers into account.
\begin{lemma} \label{lem:outerbound}
The sum capacity of the symmetric 3-user \gls{mwrc} is upper bounded as
\begin{equation} \label{eq:outerbound}
C_\Sigma \le \min\left\{ \frac{3}{2} \Capa\left( \frac{P_0}{N} \right),\ 3 \Capa\left( \frac{P}{N_0} \right) \right\}.
\end{equation}
\end{lemma}
\begin{IEEEproof}
The proof is omitted due to space constraints.
\end{IEEEproof}
\subsection{Amplify-and-Forward}
We first consider \gls{af} relaying where the relay scales the observed signal by
a positive constant and broadcasts it back to the users. The transmitted
symbol at the relay is
$X_0 = \alpha Y_0$,
%
where $\alpha$ is a normalization factor chosen such that the transmit power
constraint at the relay is met, i.e.,
$\alpha = \sqrt{P'_0\,/ \left( \sum_{k\in\mathcal K} P'_k + N_0 \right)}$,
where $P'_k$, $k\in\mathcal K_0$, is the actual transmit power of node $k$
satisfying the average power constraints.
The receiver first removes its self-interference from the received signal and then decodes for its desired message while treating the remaining interference as noise.
We split the transmission into three equal length blocks and switch off user
$i$ in time slot $i$. This reduces interference and allows for higher transmission powers in the other
two time slots while still meeting the average power constraint.
\begin{lemma} \label{lem:af}
In the 3-user \gls{mwrc}, the sum rate
\begin{equation} \label{eq:af}
R_\Sigma^{AF} = \Capa\left( \frac{3 P P_0}{N_0 P_0 + 3 P N + N N_0} \right)
\end{equation}
is achievable with \gls{af} relaying and treating interference as noise at the receivers.
\end{lemma}
\begin{IEEEproof}
Omitted due to space limitations.
\end{IEEEproof}
\subsection{Decode-and-Forward}
In \gls{df} relaying, the relay completely decodes the messages of each user
and then broadcasts them back to all users. The achievable rate region is the
intersection of the capacity regions of the 3-user \acrlong{mac} and the \acrlong{bc}
with receiver side information and partial decoding at the receivers.
\begin{lemma} \label{lem:df}
In the 3-user \gls{mwrc}, the sum rate
\begin{equation} \label{eq:dfsr}
R_\Sigma^{DF} = \min\left\{ \frac{3}{2} \Capa\left( \frac{P_0}{N} \right),\ \Capa\left( \frac{3 P}{N_0} \right) \right\}
\end{equation}
is achievable with \gls{df} relaying.
\end{lemma}
\begin{IEEEproof}[Proof sketch]
The achievable rate region is given in \cite[Proposition 2]{Gunduz2013}.
Using the simplex algorithm and the fact that
$\Capa\left( \frac{2 P}{N_0} \right) < \Capa\left( \frac{P_0}{N} \right)$
implies
$\Capa\left( \frac{3 P}{N_0} \right) < \frac{3}{2} \Capa\left( \frac{P_0}{N} \right)$
we can prove \eqref{eq:dfsr}.
\end{IEEEproof}
\begin{remark}
The result from \cite{Gunduz2013} implements a full message exchange.
However, from the outer bound in \cite{Yoo2009} it can be seen that in the
symmetric case the relaxed decoding requirements due to the partial message
exchange considered here can not result in higher rates for \gls{df}.
\end{remark}
\begin{remark} \label{rem:caplowsnr}
For a completely symmetric scenario, \gls{df} is sum rate optimal in the low \gls{snr} regime.
To see this, let $P=P_0$ and
$N=N_0$ and define $S = \frac{P}{N}$. Then the bound in Lemma~\ref{lem:outerbound} is
$C_\Sigma \le \frac{3}{2} \Capa\left( S \right)$,
and
$R_\Sigma^{DF} = \min\left\{ \frac{3}{2} \Capa\left( S \right),\ \Capa\left( 3 S \right) \right\}$.
It is easily shown that for $S \le 3 + 2 \sqrt{3}$ the first term in the minimum
is dominant, i.e., $R_\Sigma^{DF} = \frac{3}{2}
\Capa\left( S \right)$. Since this is equal to
the outer bound, we have $C_\Sigma = \frac{3}{2} \Capa\left( S \right)$ for
\glspl{snr} up to $3 + 2 \sqrt{3} \approx \unit[8.1]{dB}$.
%
\end{remark}
\subsection{Noisy Network Coding}
\Gls{nnc} \cite{Lim2011} generalizes \acrlong{cf} to \acrlongpl{dmn}. For general
multi-message networks there are two different decoding methods to choose from:
\acrfull{snd} and \acrfull{ian}. Since, in general, none of the methods is
superior to the other, we evaluate both bounds. However, it turns out that
\gls{ian} is strictly worse than \gls{snd} and even than \gls{af}.
\begin{lemma} \label{lem:nnc-snd}
In the 3-user \gls{mwrc}, the sum rate
\begin{equation} \label{eq:nnc-snd}
R_\Sigma^{NNC-SND} = \frac{3}{2} \Capa\left( \frac{2 P P_0}{N_0 P_0 + 2 P N + N N_0} \right)
\end{equation}
is achievable with \gls{nnc} and simultaneous non-unique decoding.
\end{lemma}
\begin{IEEEproof}[Proof sketch]
Use \cite[Theorem 2]{Lim2011} and identify $\mathcal D_0 = \emptyset$ and
$\mathcal D_k = \{ q(k) \}$ for $k\in\mathcal K$. Assume $\hat Y_i = Y_i +
\hat Z_i$ with $\hat Z_i \sim \mathcal{CN}(0, Q_i)$ for $i\in\mathcal K_0$,
and $Q = \emptyset$, i.e., no time-sharing is used.
Then, the achievable rate region is
\begin{align*}
R_k &< \Capa\left( \frac{P}{N_0 + Q_0} \right),\\
\smashoperator{\sum_{i\in\mathcal K\setminus\{k\}}} R_i &< \min \left\{ \Capa\left( \frac{2 P}{N_0 + Q_0} \right),\ \Capa\left( \frac{P_0}{N} \right) - \Capa\left( \frac{N_0}{Q_0} \right) \right\},
\end{align*}
for each $k\in\mathcal K$.
With the simplex algorithm \cite{Luenberger2008} and after maximization over $Q_0$ we get \eqref{eq:nnc-snd}.
\end{IEEEproof}
\begin{remark} \label{rem:AFworseSND}
It can be shown that $R_\Sigma^{NNC-SND} \ge R_\Sigma^{AF}$.
\end{remark}
\begin{lemma} \label{lem:nnc-ian}
In the 3-user \gls{mwrc}, the sum rate
\begin{equation*}
R_\Sigma^{NNC-IAN} = 3 \Capa\left( \frac{P P_0}{2 P P_0 + N_0 P_0 + 3 P N + N N_0} \right)
\end{equation*}
is achievable with \gls{nnc} and treating interference as noise.
\end{lemma}
\begin{IEEEproof}
Similar to Lemma~\ref{lem:nnc-snd}.
\end{IEEEproof}
\begin{remark} \label{rem:IANworseAF}
%
It can be shown that $\strut R_\Sigma^{NNC-IAN} \le R_\Sigma^{AF}$. Thus, with
Remark~\ref{rem:AFworseSND}, $R_\Sigma^{NNC-IAN} \le R_\Sigma^{NNC-SND}$.
\end{remark}
\section{Energy Efficiency}
\label{sec:EE}
The \gls{ee} of the system is defined as the ratio between the achievable sum rate and the consumed power. The power consumed in the system is given by the sum of the transmit power of each user and of the relay plus the circuit power that is dissipated in each terminal to operate the devices. Moreover, the transmit power of each terminal should be scaled by a factor larger than $1$ to model the nonidealities of the power amplifier \cite{Isheden2012,circuit_power_consumption}. Namely, we can express the total power $P_{t}$ consumed in the network as $P_{t}=\phi P+\psi P_{0}+P_{c}$, with $P_{c}$ denoting the total circuit power consumed in all nodes, $\psi\geq 1$ being the inefficiency of the relay amplifier, and $\phi\geq 3$, accounting for the inefficiency of the amplifier of the three users.
Accordingly, the \gls{ee} can be defined as the ratio between the achievable sum rate and $P_{t}$.
Then, given the achievable sum rate expressions from Section~\ref{Sec:AchievableRates}, the \gls{ee} can be expressed in two different functional forms.
For the upper-bound and for the \gls{df} case we have
\[
{\rm EE}_{1}=\displaystyle\frac{\min\left\{a_{1}\Capa\left(\alpha_{1}\frac{P_{0}}{N}\right),\ a_{2}\Capa\left(\alpha_{2}\frac{P}{N_{0}}\right)\right\}}{\phi P+\psi P_{0}+P_{c}}\;,
\]
with $a_{1}$, $a_{2}$, $\alpha_{1}$, and $\alpha_{2}$ non-negative parameters.
For the \gls{af} and \gls{nnc} cases we have
\[
{\rm EE}_{2}=\displaystyle\frac{\alpha\Capa\left(\frac{PP_{0}}{aP+bP_{0}+c}\right)}{\phi P+\psi P_{0}+P_{c}}\;,
\]
with $\alpha$, $a$, $b$, and $c$ non-negative parameters.
In the following, \gls{ee} maximization will be carried out by means of fractional programming tools. In particular, we recall the following result from \cite{FracProgSS1983,NonlinearFracProg}. Consider the generic fractional problem
$%
\mathop{\max}_{\mbox{\boldmath $x$}\in\mathcal{S}} \frac{f(\mbox{\boldmath $x$})}{g(\mbox{\boldmath $x$})}
$ %
where ${\cal S}\in\mathbb{R}^n$, $f,g:\mathcal{S}\to\mathbb{R}$, with $f(\mbox{\boldmath $x$})\geq 0$ and $g(\mbox{\boldmath $x$})>0$. Define the function
$F(\lambda)=\mathop{\max}\limits_{\mbox{\boldmath $x$}\in\mathcal{S}}\left(f(\mbox{\boldmath $x$})-\lambda g(\mbox{\boldmath $x$})\right)$.
Then, maximizing $f(\mbox{\boldmath $x$})/g(\mbox{\boldmath $x$})$ is equivalent to finding the unique zero of $F(\lambda)$. This can be accomplished by means of Dinkelbach's algorithm \cite{NonlinearFracProg}, which only requires the solution of a sequence of convex problems, provided $f(\mbox{\boldmath $x$})$ and $g(\mbox{\boldmath $x$})$ are concave and convex, respectively, and that ${\cal S}$ is a convex set.
%
%
%
Moreover, it can be shown that the convergence rate of Dinkelbach's algorithm is superlinear \cite{NonlinearFracProg}.
\subsection{Maximization of ${\rm EE}_{1}$}
The maximization of ${\rm EE}_{1}$ is a non-concave and non-smooth problem. However, it can be reformulated as a smooth problem introducing the auxiliary variable $t$ as follows.
\begin{equation}\label{Prob:MaxEE1Smooth}
\left\{
\begin{array}{lll}
\displaystyle\max_{P,P_{0}}\quad \frac{t}{\phi P+\psi P_{0}+P_{c}}\\
{\rm s.t.}\;\;\; P\in[0;P^{max}]\;,\;P_{0}\in[0;P_{0}^{max}]\\
a_{1}\Capa\left(\alpha_{1}\frac{P_{0}}{N}\right)-t\geq 0\;,\quad a_{2}\Capa\left(\alpha_{2}\frac{P}{N_{0}}\right)-t \geq 0
\end{array}
\right.
\end{equation}
The numerator and denominator of the objective of \eqref{Prob:MaxEE1Smooth} are both linear and the constraints are convex. As a consequence, \eqref{Prob:MaxEE1Smooth} can be solved by means of Dinkelbach's algorithm with an affordable complexity.
\subsection{Maximization of ${\rm EE}_{2}$}
In this case, the optimization problem is formulated as
\begin{equation}\label{Prob:MaxEE2}
\left\{
\begin{array}{ll}
\displaystyle\max_{P,P_{0}}\displaystyle\frac{\alpha\Capa\left(\frac{PP_{0}}{aP+bP_{0}+c}\right)}{\phi P+\psi P_{0}+P_{c}}\\
{\rm s.t.}\;\;\; P\in[0;P^{max}]\;,\quad P_{0}\in[0;P_{0}^{max}]
\end{array}
\right.
\end{equation}
Problem \eqref{Prob:MaxEE2} is more challenging than Problem \eqref{Prob:MaxEE1Smooth} because the numerator of the objective function is not jointly concave in the optimization variables.
However, we observe that the numerator of the objective is separately concave in $P$ for fixed $P_{0}$ and vice versa. This suggests that a convenient way to tackle Problem \eqref{Prob:MaxEE2} is by means of the alternating maximization algorithm \cite{BertsekasNonLinear}, according to which we can alternatively optimize with respect to $P$ fixing the value of $P_{0}$, and with respect to $P_{0}$ for a fixed value of $P$.
The formal algorithm is reported next and labeled Algorithm 1.
\begin{algorithm}
\caption{Alternating maximization for Problem \eqref{Prob:MaxEE2}}
\label{Alg:AMEE2}
\begin{algorithmic}
\State
\texttt{Initialize} $P_{0}^{(0)}\in[0,P_{0}^{max}]$. \texttt{Set a tolerance} $\epsilon$. \texttt{Set} $n=0$;
\While{$\left|{\rm EE}_{2}^{(n)}-{\rm EE}_{2}^{(n-1)}\right| \leq \epsilon$}
\State \texttt{Given} $P_{0}^{(0)}$, \texttt{solve Problem} \eqref{Prob:MaxEE2} \texttt{with respect to} $P$ \texttt{to obtain the optimal} $P^{(n+1)}$;
\State \texttt{Given} $P^{(n+1)}$, \texttt{solve Problem} \eqref{Prob:MaxEE2} \texttt{with respect to} $P_{0}$ \texttt{to obtain the optimal} $P_{0}^{(n+1)}$;
\State $n=n+1$;
\EndWhile
\State \texttt{Output} $(P,P_{0})$.
\end{algorithmic}
\end{algorithm}
Each subproblem in Algorithm~\ref{Alg:AMEE2} can be globally solved by means of Dinkelbach's algorithm.
Moreover, the following proposition holds.
\begin{proposition}
Algorithm~\ref{Alg:AMEE2} converges to a stationary point of Problem \eqref{Prob:MaxEE2}.
\end{proposition}
\begin{IEEEproof}[Proofsketch]
After each iteration of Algorithm~\ref{Alg:AMEE2} the objective is not decreased. Hence, convergence follows since the objective is upper-bounded.
Convergence to a stationary point holds by virtue of \cite[Proposition 2.7.1]{BertsekasNonLinear}, which states that alternating maximization converges to a stationary point if: 1) the feasible set is the Cartesian product of closed and convex sets; 2) the objective is continuously differentiable on the feasible set; 3) the solution to each subproblem is unique. In our case, 1) and 2) are apparent. As for 3) it also holds because the objective function of each subproblem can be shown to be strictly pseudo-concave \cite{ZapTWC13}.
\end{IEEEproof}
\section{Discussion \& Numerical Results}
For a discussion and numerical evaluation of the presented transmission
schemes, we consider a completely symmetric scenario with $N = N_0$. We assume
$P = P_0$ for the spectral efficiency, and for the \gls{ee} evaluation, we
assume $P^{max} = P_0^{max}$, unit noise variance and no power loss at the
transmitter, i.e., $\psi = 1$ and $\phi = 3$. The shown performance has been
obtained using the algorithms proposed in Section~\ref{sec:EE}.
Fig.~\ref{fig:spectral_eff} shows the achievable sum rates from
Section~\ref{Sec:AchievableRates} as a function of the \gls{snr}. As noted in
Remarks~\ref{rem:AFworseSND} and \ref{rem:IANworseAF}, it can be observed that \gls{nnc} with \gls{snd}
achieves a higher sum rate than \gls{af} and \gls{af} achieves a higher sum
rate than \gls{nnc} with \gls{ian}. In the low \gls{snr} regime, the sum
capacity is achieved by \gls{df} (see Remark~\ref{rem:caplowsnr}). Starting at
approximately \unit[8]{dB}, \gls{df} stops being sum rate optimal but is still
better than all other considered transmission schemes. Starting from approximately
\unit[14.27]{dB} \gls{nnc} \gls{snd} is the best in terms of spectral efficiency.
Its gap to the outer bound is at most $\unit[1.5 \log_2\left( 1.5 \right)]{bit}
\approx \unit[0.877]{bit}$. In contrast, for all other considered schemes this
gap grows unbounded as $\SNR\rightarrow\infty$. Furthermore, the gap between
\gls{df} and \gls{af} approaches \unit[2]{bit} as $\SNR\rightarrow\infty$.
\Gls{nnc} \gls{ian} is clearly worse than every other employed scheme and
should not be considered in terms of spectral efficiency.
Fig.~\ref{fig:ee} shows the \gls{ee} as a function of the \gls{snr} for a
fixed circuit power $P_c = \unit[1]{W}$. First of all, it can be seen that the \gls{ee} saturates when $P_{max}$ exceeds a given value, which is lower than \unit[0]{dB} for all considered schemes. This is explained recalling that, unlike the achievable rate, the \gls{ee} is not increasing with the transmit powers, but instead admits an optimum transmit power level. If $P_{max}$ is larger than such power level, then it is not optimal to transmit at full power. This also explains why \gls{df} performs significantly better than all other schemes, including \gls{nnc} \gls{snd}. Indeed, due to the saturation of the \gls{ee}, the \gls{snr} range for which \gls{nnc} \gls{snd} yields a larger achievable sum rate than \gls{df} is not reached when \gls{ee} is optimized. Finally, as expected, \gls{nnc} \gls{snd} is better than \gls{af}, which is better than \gls{nnc} \gls{ian}.
However, as opposed to \gls{df}, \gls{af} does not require power-hungry analog
to digital conversion and digital signal processing at the relay which results
in significantly less power consumption. Furthermore, the decoders at the users
are also expected to consume less power due to the use of a (considerably
simpler) single user receiver. Thus, the higher achievable rates and the
resulting better \gls{ee} of \gls{df} over \gls{af} are obtained at the cost of a more complex
hardware and, hence, of a larger consumed circuit power. This suggests that the
comparison in Fig.~\ref{fig:ee} might be unfair and that the large gap between
\gls{df} and \gls{af} might in fact be smaller when the comparison is done on
equal grounds. Some insight on this issue is given in Fig.~\ref{fig:eeAFDF}, which shows the
\gls{ee} of \gls{df} as a function of its circuit power $P_c^{DF}$ and the
\gls{ee} of \gls{af} for a fixed circuit power $P_c^{AF} = \unit[1]{W}$. It can
be seen that, as expected, the gap to \gls{af} gets smaller with increasing
$P_c^{DF}$ and that \gls{af} might even outperform \gls{df} given a
significantly large $P_c^{DF}$.
\tikzset{/pgfplots/width = {\axisdefaultwidth}, /pgfplots/height= {0.85*\axisdefaultheight}}
\begin{figure}
\begin{tikzpicture}
\begin{axis} [
thick,
xlabel={SNR [dB]},
ylabel={Sum Rate [bit/s/Hz]},
ylabel near ticks,
grid=major,
xtick = {-20,-10,...,40},
xmin = -20,
xmax = 40,
minor x tick num = 1,
%
minor y tick num = 1,
yminorgrids = true,
mark repeat = 15,
legend pos=north west,
legend cell align=left,
cycle list name=color2,
]
\pgfplotstableread{sym_sumrates.dat}\tbl
\addplot+[mark phase = 0] table[y=bound] {\tbl};
\addlegendentry{Outer Bound};
\addplot+[mark phase = 8] table[y=nnc_snd] {\tbl};
\addlegendentry{NNC SND};
\addplot+[mark phase = 5] table[y=df] {\tbl};
\addlegendentry{DF};
\addplot+[mark phase = 11] table[y=af] {\tbl};
\addlegendentry{AF};
\addplot+[mark phase = 0] table[y=nnc_ifn] {\tbl};
\addlegendentry{NNC IAN};
\end{axis}
\end{tikzpicture}
\vspace{-2ex}
\caption{Spectral efficiency in the 3-user \gls{mwrc}; 1) Outer bound from
Lemma~\ref{lem:outerbound}, 2) \acrfull{nnc} with
\acrfull{snd} and 3) \acrfull{ian}, 4) \acrfull{af} and 5) \acrfull{df} plotted as a
function of the \gls{snr}.}
\label{fig:spectral_eff}
\end{figure}
\begin{figure}
\begin{tikzpicture}
\begin{axis} [
thick,
xlabel={$P^{max}$ [dB]},
ylabel={Energy Efficiency [bit/Hz/J]},
ylabel near ticks,
grid=major,
xtick = {-40,-30,...,20},
xmin = -30,
xmax = 10,
minor x tick num = 1,
%
minor y tick num = 1,
yminorgrids = true,
mark repeat = 150,
legend pos=north west,
legend cell align=left,
cycle list name=color2,
]
\pgfplotstableread{out_Pc=1_N=1.dat}\tbl
\addplot+[mark phase = 0] table[y=Outer_Bound] {\tbl};
\addlegendentry{Outer Bound};
\addplot+[mark phase = 15] table[y=NNC_SND] {\tbl};
\addlegendentry{NNC SND};
\addplot+[mark phase = 90] table[y=DF] {\tbl};
\addlegendentry{DF};
\addplot+[mark phase = 105] table[y=AF] {\tbl};
\addlegendentry{AF};
\addplot+[mark phase = 60] table[y=NNC_IFN] {\tbl};
\addlegendentry{NNC IAN};
\end{axis}
\end{tikzpicture}
\vspace{-2ex}
\caption{\Acrlong{ee} in the 3-user \gls{mwrc} of 1) \acrfull{nnc} with
\acrfull{snd} and 2) \acrfull{ian}, 3) \acrfull{af}, 4) \acrfull{df}, and 5) the outer
bound from Lemma~\ref{lem:outerbound} as a function of the \gls{snr} for fixed
circuit power $P_c = \unit[1]{W}$.}
\label{fig:ee}
\end{figure}
\begin{figure}
\begin{tikzpicture}
\begin{axis} [
thick,
xlabel={$P_c^{DF}$ [W]},
ylabel={Energy Efficiency [bit/Hz/J]},
ylabel near ticks,
grid=major,
xtick = {1, 10, 20, 30, 40, 50},
xmin = 1,
xmax = 50,
minor x tick num = 1,
%
mark repeat = 10,
legend pos=north east,
legend cell align=left,
cycle list name=color2,
cycle list shift=2,
]
\pgfplotstableread{DFstepPc.dat}\tbl
\pgfplotstablegetelem{1}{AF}\of\tbl
\let\AFval\pgfplotsretval
\addplot +[name path global=steppcDF, mark phase=5] table[y=DF] {\tbl};
\addlegendentry{DF};
\addplot +[name path global=steppcAF, mark phase=5] table[y=AF] {\tbl};
%
%
%
%
\addlegendentry{AF with $P_c^{AF} = 1$};
\path [inner sep=.1em, name intersections={of=steppcDF and steppcAF}]
(intersection-1) \pgfextra{\pgfgetlastxy{\macrox}{\macroy}\let\macrox\macrox}
node (int) {}
+(1em, 1em)
node (text) [anchor=south west] {$P_c^{DF} \approx \transformxdimension{\macrox}\pgfmathprintnumber[fixed,precision=2]{\pgfmathresult}$}
\pgfextra{\transformxdimension{\macrox}\global\let\Pcintersection\pgfmathresult};
\draw[latex-] (int.north east) -- (text.south west);
\end{axis}
\end{tikzpicture}
\vspace{-2ex}
\caption{\Acrlong{ee} in the 3-user \gls{mwrc} of \acrfull{df} as a function of
the circuit power $P_c^{DF}$ compared to \acrfull{af} with a fixed circuit power
$P_c^{AF} = \unit[1]{W}$ for an operating point of $P^{max} = \unit[10]{W}$. The intersection is at $P_c^{DF}
\approx \pgfmathprintnumber[fixed,precision=2]{\Pcintersection}\, W$.}
\label{fig:eeAFDF}
\end{figure}
\section{Conclusion}
In this paper, we studied both the spectral and the energy efficiency of the
3-user \gls{mwrc} with a partial message exchange. We provided analytic sum
rate expressions for the most common relaying schemes and discussed the solution of the
optimization problems arising in the calculation of the \gls{ee}.
\enlargethispage{-4.6cm}
We have seen that if we assume the same power consumption for all schemes,
\gls{df} performs best in terms of \gls{ee}. Moreover, the
energy-efficient performance of \gls{nnc} is not satisfactory due to the
fact that \gls{nnc} achieves a higher spectral efficiency only in the high
\gls{snr} regime, which is not the operating regime when \gls{ee} is
optimized. Furthermore, we have shown that \gls{af} might have better \gls{ee}
than the more complex \gls{df} if different circuit powers are assumed. This
assumption is reasonable since different hardware complexities imply different
circuit powers. Thus, to compare the \gls{ee} of the presented relaying schemes
in a fairer way, circuit power consumption models are necessary. This issue
will be addressed further in future work.
|
\section{Introduction}
Superfluidity plays a key role in many fields of physics, such as liquid
helium \cite{Leggett}, ultracold atomic gases \cite{Pitaevskii,Pethick},
quantum magnets \cite{Zapf}, and astrophysics \cite{Liberati}. In
particular, superfluid Bose-Fermi mixtures (SBFMs) in ultracold atomic gases
have attracted considerable interest recently. It is well known that for
liquid helium the superfluidity of bosonic $^{4}$He and fermionic $^{3}$He
can be achieved separately. However, the simultaneous superfluidity for the
two isotopes is remarkably prevented by the strong interactions between the
two species in spite of such a superfluid $^{4}$He-$^{3}$He mixture being a
long-sought object \cite{Rysti}. By contrast, a superfluid Bose-Fermi
mixture (SBFM) in ultracold atomic gases can be realized by using the
combination of the Feshbach resonance and radio-frequency techniques. Most
recently, a SBFM consisting of condensed $^{7}$Li bosons and ultracold $^{6}
Li fermions in two-spin states has been produced by Salomon's group \cit
{Ferrier-Barbut}. The experimental breakthrough provides new opportunities
to study the intriguing properties of SBFMs inaccessible in a Bose-Einstein
condensate (BEC) with arbitrary spin \cite{Ho1,YWu}, pure bosonic superfluid
mixtures, and pure fermionic superfluid mixtures. As a matter of fact, many
novel physical characteristics have been predicted theoretically in SBFMs,
including the saturation effect of nonlinear interaction in the
strong-coupling unitarity limit \cite{Adhikari1}, localized Bose-Fermi
bright soliton \cite{Adhikari2}, possibility of simulating dense quantum
chromodynamics matter \cite{Maeda}, phase transition overlapping with phase
separation \cite{Ramachandhran}, and the Faraday pattern generation \cit
{Abdullaev}.
All existing studies of the SBFMs refer to the nonrotating case. Considering
that one of the most striking hallmarks of a superfluid is its response to
rotation, in this paper we investigate the combined effects of rotation and
nonlinear interatomic interaction on the exact two-dimensional (2D)
topological structure and dynamics of a rotating SBFM, focusing on the
superfluid $^{7}$Li-$^{6}$Li mixture case as the prototype. Here we present
a phenomenological dissipation model \cite{Wen1} combining with the unitary
Schr\"{o}dinger equation for an interacting SBFM \cite{Adhikari1} to
describe the dynamics of a rotating SBFM. A ground-state phase diagram for a
nonrotating SBFM with specific parameters is given. We find that the
ground-state structure of a nonrotating SBFM is mainly determined by the
ratio of the boson-fermion (BF) scattering length to the boson-boson (BB)
scattering length. Furthermore, it is shown that the rotating SBFM with a
sufficiently large rotation frequency can display four steady structures: a
mixed phase and three typical layer separated phases. In particular, we show
that the generation of visible vortices in the fermionic superfluid displays
an evident hysteresis effect during the time evolution of a rotating SBFM.
Due to the Cooper pair entity of fermionic superfluid and the presence of BF
and fermion-fermion (FF) interactions, it is demonstrated that the
topological structure and the dynamics of a rotating SBFM are evidently
different from the usual cases of rotating two-component Bose-Einstein
condensates (BECs).
The paper is organized as follows. In Sec. II, we describe a
phenomenological model for a rotating SBFM. In Sec. III, we study the
equilibrium structure of the rotating SBFM. A phase diagram for a
nonrotating SBFM is given, and the phase structures of the rotating SBFMs
with various parameter values are analyzed. In Sec. IV, we discuss the
dynamics of vortex formation in a rotating SBFM. The conclusion is outlined
in the last section.
\section{Model}
We consider a superfluid Bose-Fermi mixture composed of superfluid bosons
and two-component superfluid fermions with equal populations of spin up and
down, corresponding to the most favorable condition for Cooper pairing. The
superfluid Bose-Fermi gas is confined in a harmonic trap rotating around the
$z$ axis with angular velocity $\Omega $. In the present work we use a
combined phenomenological model based on a phenomenological dissipation
model \cite{Wen1} and a Galilei-invariant nonlinear unitarity model \cit
{Adhikari1} to investigate the dynamics of the vortex formation and the
structure of the equilibrium state of a rotating SBFM. In this
phenomenological model, the order parameter $\Psi _{b}$ of a bosonic
superfluid (i.e., a Bose-Einstein condensate) and the order parameter $\Psi
_{p}$ of a fermionic superfluid \cite{Adhikari1} obey the coupled equations
in the rotating frame
\begin{eqnarray}
(i-\gamma )\hbar \frac{\partial \Psi _{b}}{\partial t} &=&\left[ -\frac
\hbar ^{2}\nabla ^{2}}{2m_{b}}+U_{b}+\mu _{b}(n_{b},a_{b})-\Omega L_{z
\right] \Psi _{b} \notag \\
&&+G_{bp}\left\vert \Psi _{p}\right\vert ^{2}\Psi _{b}, \label{SBE} \\
(i-\gamma )\hbar \frac{\partial \Psi _{p}}{\partial t} &=&\left[ -\frac
\hbar ^{2}\nabla ^{2}}{2m_{p}}+U_{p}+\mu _{p}(n_{p},a_{f})-\Omega L_{z
\right] \Psi _{p} \notag \\
&&+G_{bp}\left\vert \Psi _{b}\right\vert ^{2}\Psi _{p}, \label{SFE}
\end{eqnarray
where for simplicity we assume that the degree of dissipation $\gamma $ of
the fermionic superfluid is the same as that of the BEC. $m_{b}$ is the mass
of a bosonic atom, $m_{p}=2m_{f}$ is the mass of a Cooper pair with $m_{f}$
being the mass of a single fermion, and $L_{z}=i\hbar (y\partial
_{x}-x\partial _{y})$ denotes the $z$ component of the angular-momentum
operator. $U_{b}=m_{b}(\omega _{b}^{2}r^{2}+\omega _{zb}^{2}z^{2})/2$ is the
external trapping potential for the bosons, $U_{p}=m_{p}(\omega
_{f}^{2}r^{2}+\omega _{zf}^{2}z^{2})/2$ is the trapping potential for the
fermionic superfluid, and $\omega _{b}$($\omega _{f}$) and $\omega _{zb}$(
\omega _{zf}$) are the radial trap frequency and the axial trap frequency
for the bosons (fermions), respectively. $G_{bp}=4\pi \hbar
^{2}a_{bf}/m_{bf} $, where $a_{bf}$ is the BF $s$-wave scattering length,
and $m_{bf}=m_{b}m_{f}/(m_{b}+m_{f})$ is the BF reduced mass. The bulk
chemical potential of the BEC is given b
\begin{equation}
\mu _{b}(n_{b},a_{b})=\frac{\hbar ^{2}}{m_{b}}n_{b}^{2/3}f(n_{b}^{1/3}a_{b}),
\label{ChemPotentialBoson}
\end{equation
where
\begin{equation}
f(x)=\frac{4\pi \left( x+\alpha x^{5/2}\right) }{1+\lambda x^{3/2}+\beta
x^{5/2}}, \label{BosonFunction}
\end{equation
$n_{b}$ is the local density of bosons, and $a_{b}$ is the BB $s$-wave
scattering length. Here we choose $(\alpha -\lambda )=32/(3\sqrt{\pi })$,
\beta =4\pi \alpha /\eta $ with $\eta $ $=22.22$, $\alpha =32\xi /(3\sqrt
\pi })$, and $\lambda =32(\xi -1)/(3\sqrt{\pi })$ with $\xi =1.1$ \cit
{Adhikari1}. The bulk chemical potential of the Fermi superfluid is
expressed as \cite{Adhikari1,Recati,Adhikari3}
\begin{equation}
\mu _{p}(n_{p},a_{f})=\frac{2\hbar ^{2}}{m_{p}}(6\pi
^{2}n_{p})^{2/3}g(2^{1/3}n_{p}^{1/3}a_{f}), \label{ChemPotentialFermion}
\end{equation
where
\begin{equation}
g(x)=1+\frac{\delta x}{1-kx}, \label{FermionFunction}
\end{equation
$n_{p}=n_{f}/2$ is the local density of Cooper pairs with $n_{f}$ being the
total local density of fermions, $a_{f}$ denotes the attractive FF
scattering length, $\delta =20\pi /(3\pi ^{2})^{2/3}$, and $k=\delta
/(1-\zeta )$ with $\zeta =0.44$. The choices of the parameters in Eqs. (\re
{BosonFunction}) and (\ref{FermionFunction}) are consistent with the
unitarity and the Lee-Yang-Huang limits \cite{Adhikari1,Lee1,Lee2} as well
as the relevant results of the Monte Carlo calculations \cit
{Blume1,Carlson,Astrakharchik,Blume2,Chang}.
This phenomenological model is a variation of that in \cit
{Tsubota,Kasamatsu1} and a generalization of that of a rotating
single-component BEC \cite{Wen1}, and it has good predictive power \cit
{Wen2}. The largest merit of the phenomenological dissipation model is that
by virtue of this model one can obtain not only\ the steady states of a
rotating system but also reveal the dynamics of vortex formation. In
addition, the present model is valid from weak coupling to unitarity for
both bosons and fermions \cite{Adhikari1}.
As a matter of fact, the unitary Schr\"{o}dinger model for nonrotating
superfluid fermions was presented in Refs. \cit
{Adhikari1,Adhikari3,Salasnich}, where the bulk chemical potential was given
by Eqs. (\ref{ChemPotentialFermion}) and (\ref{FermionFunction}) (see also
Eqs. (10), (11) and (28) in Ref. \cite{Adhikari1}). The lowest-order term
g(x)=1$ in Eq. (\ref{FermionFunction}) results in the bulk chemical
potential of a Fermi superfluid in the absence of FF interaction ($a_{f}=0
). Furthermore, the next-order term $g(x)=1+\delta x$ can lead to a known
analytical result in the small-gas-parameter regime (medium value FF
scattering length $a_{f}$) as obtained in Ref. \cite{Lee1}. When
n_{p}^{1/3}a_{f}$ tends to negative infinity (the so-called unitary limit),
the function $g(x)$ obviously approaches an asymptotic value $(1-\delta /k)
. Thus the function (\ref{ChemPotentialFermion}) with Eq. (\re
{FermionFunction}) provides a smooth interpolation between the bulk chemical
potential of a Fermi superfluid in the weak-coupling limit and that in the
unitary limit for both the uniform case and the trapped case \cit
{Adhikari1,Adhikari3,Lee1}. It was shown that the results with the choice of
the fitting parameters $\delta =20\pi /(3\pi ^{2})^{2/3}$, $k=\delta
/(1-\zeta ),$ and $\zeta =0.44$ \cite{Adhikari1} agreed well with the
corresponding Monte Carlo data \cite{Blume2,Chang}. Therefore, we will use
the choice in the present work.
Next, we consider the two-dimensional problem by assuming the translation
invariance along the $z$ axis (i.e., $\omega _{zb}=\omega _{zf}=0$),
reducing the order parameters as $\Psi _{i}(\overrightarrow{r},t)=\psi
_{i}(x,y,t)/\sqrt{R_{z}}(i=b,p)$ with the typical size $R_{z}$ along the $z$
axis. After a straight-forward calculation, we obtain 2D coupled equations
\begin{eqnarray}
(i-\gamma )\hbar \frac{\partial \psi _{b}}{\partial t} &=&\left[ -\frac
\hbar ^{2}}{2m_{b}}\left( \nabla _{x}^{2}+\nabla _{y}^{2}\right) +F(\psi
_{b})-\Omega L_{z}\right] \psi _{b} \notag \\
&&+\left[ \frac{m_{b}\omega _{b}^{2}}{2}\left( x^{2}+y^{2}\right) +\frac
G_{bp}}{R_{z}}\left\vert \psi _{p}\right\vert ^{2}\right] \psi _{b},
\label{2DSBE} \\
(i-\gamma )\hbar \frac{\partial \psi _{p}}{\partial t} &=&\left[ -\frac
\hbar ^{2}}{2m_{p}}(\nabla _{x}^{2}+\nabla _{y}^{2})+J(\psi _{b})-\Omega
L_{z}\right] \psi _{p} \notag \\
&&+\left[ m_{f}\omega _{f}^{2}(x^{2}+y^{2})+\frac{G_{bp}}{R_{z}}\left\vert
\psi _{b}\right\vert ^{2}\right] \psi _{p}, \label{2DSFE}
\end{eqnarray
wher
\begin{equation}
F(\psi _{b})=\frac{4\pi \hbar ^{2}\left( a_{b}R_{z}^{-1}\left\vert \psi
_{b}\right\vert ^{2}+\alpha a_{b}^{5/2}R_{z}^{-3/2}\left\vert \psi
_{b}\right\vert ^{3}\right) }{m_{b}\left( 1+\lambda
a_{b}^{3/2}R_{z}^{-1/2}\left\vert \psi _{b}\right\vert +\beta
a_{b}^{5/2}R_{z}^{-5/6}\left\vert \psi _{b}\right\vert ^{5/3}\right) },
\label{Ffunction}
\end{equation
and
\begin{equation}
J(\psi _{p})=\frac{2\hbar ^{2}}{m_{p}}(6\pi ^{2})^{2/3}\left\vert \psi
_{p}\right\vert ^{4/3}\left( 1+\frac{2^{1/3}\delta a_{f}\left\vert \psi
_{p}\right\vert ^{2/3}}{1-2^{1/3}ka_{f}\left\vert \psi _{p}\right\vert ^{2/3
}\right) . \label{Jfunction}
\end{equation
The initial order parameters are normalized as $N_{i}=\iint \psi
_{i}(x,y,t=0)dxdy$ $(i=b,p)$ with $N_{b}$ being the initial number of bosons
and $N_{p}$ being the initial number of Cooper pairs ($N_{p}=N_{f}/2$ and
N_{f}$ is the total number of fermions). By introducing the notations
\omega _{bf}=(\omega _{b}+\omega _{f})/2$, $d_{0}=\sqrt{\hbar
/(2m_{bf}\omega _{bf})}$, $x_{0}=x/d_{0}$, $y_{0}=y/d_{0}$, $t_{0}=t\omega
_{bf}$ and $\Omega _{0}=\Omega /\omega _{bf}$, and replacing the wave
functions as $\psi _{i}\rightarrow \sqrt{N_{i}}\psi _{i}/d_{0}$, we obtained
the rescaled dimensionless 2D equations for the mixtur
\begin{eqnarray}
(i-\gamma )\frac{\partial \psi _{b}}{\partial t} &=&\left[ -\frac{m_{bf}}
m_{b}}(\nabla _{x}^{2}+\nabla _{y}^{2})+C_{1}\left\vert \psi _{p}\right\vert
^{2}-\Omega L_{z}\right] \psi _{b} \notag \\
&&+\frac{4\pi \hbar }{m_{b}\omega _{bf}}\frac{\left( A_{1}\left\vert \psi
_{b}\right\vert ^{2}+\alpha B_{1}\left\vert \psi _{b}\right\vert
^{4/3}\right) }{1+\lambda D_{1}\left\vert \psi _{b}\right\vert +\beta
E_{1}\left\vert \psi _{b}\right\vert ^{5/3}}\psi _{b} \notag \\
&&+\frac{m_{b}\omega _{b}^{2}}{4m_{bf}\omega _{bf}^{2}}\left(
x^{2}+y^{2}\right) \psi _{b}, \label{DimLess2DSBE} \\
(i-\gamma )\frac{\partial \psi _{p}}{\partial t} &=&\left[ -\frac{m_{bf}}
m_{p}}(\nabla _{x}^{2}+\nabla _{y}^{2})+C_{2}\left\vert \psi _{b}\right\vert
^{2}-\Omega L_{z}\right] \psi _{p} \notag \\
&&+A_{2}\left\vert \psi _{p}\right\vert ^{4/3}\left( 1+\frac{\delta
B_{2}\left\vert \psi _{p}\right\vert ^{2/3}}{1-kB_{2}\left\vert \psi
_{p}\right\vert ^{2/3}}\right) \psi _{p} \notag \\
&&+\frac{m_{f}\omega _{f}^{2}}{2m_{bf}\omega _{bf}^{2}}\left(
x^{2}+y^{2}\right) \psi _{p}, \label{DimLess2DSFE}
\end{eqnarray
where the number subscript $0$ is omitted for simplicity. The corresponding
coefficients are $A_{1}=a_{b}N_{b}d_{0}^{-2}R_{z}^{-1}$,
B_{1}=a_{b}^{5/2}N_{b}^{3/2}d_{0}^{-3}R_{z}^{-3/2}$, $C_{1}=8\pi
a_{bf}N_{p}R_{z}^{-1}$, $D_{1}=a_{b}^{3/2}N_{b}^{1/2}d_{0}^{-1}R_{z}^{-1/2}
, $E_{1}=a_{b}^{5/2}N_{b}^{5/6}d_{0}^{-5/3}R_{z}^{-5/6}$, $A_{2}=(2\hbar
/m_{p}\omega _{bf})(6\pi ^{2}N_{p}d_{0}^{-2})^{2/3}$,
B_{2}=(2N_{p}d_{0}^{-2})^{1/3}a_{f}$, and $C_{2}=8\pi a_{bf}N_{b}R_{z}^{-1}$.
In the following, we numerically solve the 2D coupled equations (\re
{DimLess2DSBE}) and (\ref{DimLess2DSFE}) which requires an enormous
computational effort. The initial ground-state order parameters $\psi
_{b}(x,y,t=0)$ and $\psi _{p}(x,y,t=0)$ of the system can be obtained by the
imaginary-time propagation method \cite{Zhang,Xu,Wen3} based on the
Peaceman-Rachford method \cite{Peaceman,Wen4}. Here we consider a superfluid
$^{7}$Li-$^{6}$Li mixture in harmonic traps satisfying $U_{b}(x,y)=m_{b
\omega _{b}^{2}\left( x^{2}+y^{2}\right) =U_{f}(x,y)=m_{f}\omega
_{f}^{2}\left( x^{2}+y^{2}\right) $, with $m_{b\text{ }}$being the mass of
^{7}$Li atom and $m_{f\text{ }}$ being the mass of the $^{6}$Li atom. The
parameters are chosen as $\omega _{b}=2\pi \times 100$ Hz, $R_{z}=10$ $\mu
m, and $\gamma =0.03$, corresponding to a temperature of about $0.1T_{c}$
\cite{Choi}. Recently, a mixture of $^{7}$Li superfluid and $^{6}$Li
superfluid has been realized by Salomon's group \cite{Ferrier-Barbut}. Thus
the above assumption is valid and feasible, and the relevant results can be
tested under the current experimental conditions.
\section{Steady structure of a rotating SBFM}
For convenience, we introduce two relative interaction strengths,
R_{1}=a_{bf}/a_{b}$ and $R_{2}=a_{f}/a_{b}$. The FF attractive scattering
length $a_{f}$ can be varied from zero to negative infinity, and the
absolute value of the BF scattering length $\left\vert a_{bf}\right\vert $
is assumed to be not too large (otherwise, it requires special attention
\cite{Adhikari1}).
\begin{figure}[tbp]
\centerline{\includegraphics*[width=7cm]{Figure1.eps}}
\caption{(Color online) Ground-state phase diagram of a nonrotating
superfluid $^{7}$Li-$^{6}$Li (Bose-Fermi) mixture, where $R_{1}=a_{bf}/a_{b}$
and $R_{2}=a_{f}/a_{b}$. The parameters are $N_{b}=1000$, $N_{p}=100$, and
a_{b}=50$ nm. Regions I--IV represent the system collapse regime, mixed
phase, layer-separated phase, and inlaid separated phase, respectively. }
\label{Figure1}
\end{figure}
Figure 1 shows the ground-state phase diagram of a static superfluid $^{7}
Li--$^{6}$Li mixture with fixed BB repulsive interaction. The corresponding
parameters are $a_{b}=50$ nm, $N_{b}=1000$, and $N_{p}=100$. There exist
four possible phases depending on the values of $R_{1}$ and $R_{2}$, and the
typical density profiles corresponding to phases II-IV are displayed in Fig.
2. In Fig. 1, region I denotes the system collapse regime. For a
sufficiently strong attractive BF interaction (here, $R_{1}<-0.5$, i.e.,
a_{bf}<-25$ nm), the system undergoes a simultaneous collapse of the density
profile of the BEC and that of the fermionic superfluid. Physically, the
critical value of BF scattering length is governed by the balance between
the kinetic energy of bosons and Cooper pairs and the mutual attractive BF
interaction. When the BF attraction becomes sufficiently strong, it can no
longer be stabilized by the kinetic energy. Therefore the mixture lowers its
energy via increasing the densities of bosons and Cooper pairs, and finally
the bosonic superfluid or the fermionic one or both the bosonic and
fermionic superfluids collapse simultaneously due to instability. Region II
represents a miscible phase in which the density profile of the BEC has a
much larger spatial extension than that of the fermionic superfluid due to
the BB repulsion and the FF attraction [see Figs. 2(a1) and 2(a2)]. Region
III is a layer-separated phase, where the BEC is completely expelled outside
the fermionic superfluid because of the strong BF repulsive interaction
[Figs. 2(b1) and 2(b2)], which is usually referred as demixing. The similar
mixing-demixing transition phenomenon has also been found in the studies of
degenerate boson-fermion mixtures \cite{Molmer,Roth,Capuzzi,Adhikari4,Wen5}
or two-component BECs \cite{Wen4,Ho2,Pu,Trippenbach,LWen}. Finally, region
IV marks an inlaid separated phase in which the fermionic superfluid is
divided into two segments and inlaid into the outskirts of the BEC [Figs.
2(c1) and 2(c2)].
\begin{figure}[tbp]
\centerline{\includegraphics*[width=7.8cm]{Figure2.eps}}
\caption{(Color online) Density profiles (a)--(c) correspond to phases
II--IV in Fig. 1, respectively. Here, $1$ labels the bosonic superfluid and
2$ denotes the fermionic one of the system. (a) $R_{1}=-0.4,R_{2}=-20,$ (b)
R_{1}=1,R_{2}=-20,$ and (c) $R_{1}=6,R_{2}=-20$. The other parameters are
the same as those in Fig. 1. The darker color area indicates the lower
density. $x$ and $y$ are in units of $d_{0}$. }
\label{Figure2}
\end{figure}
The ground-state structures of SBFMs are different from those of degenerate
boson-fermion mixtures \cite{Molmer,Roth,Capuzzi,Adhikari4,Wen5}. In the
latter case, there is no $s$-wave interaction between identical fermions in
the spin-polarized state due to the Pauli exclusion principle. Depending on
the BB and BF interactions, the density profile of a degenerate
boson-fermion mixture may display a core-shell-shaped separated phase, where
the Fermi gas forms a shell around or a core inside the BEC, or even both
\cite{Molmer,Roth,Capuzzi,Adhikari4,Wen5}. However, our simulation shows
that in a wide range of parameter values, the SBFM supports neither a
separated phase of fermions constituting a shell around the BEC nor a
staggered separated phase of fermions becoming both a shell around and a
core inside the BEC (see Figs. 1 and 2); even the total number of fermions
is much larger than that of bosons. In addition, for a fixed value of $R_{1}$
(i.e., a fixed BF interaction), the variation of the parameter $R_{2}$ (FF
interaction) dose not influence the phase structure of the SBFM (Fig. 1).
Furthermore, when the repulsive BF interaction of a SBFM is sufficiently
strong, the system exhibits an inlaid separated phase, where the fermionic
superfluid embeds into the periphery of the bosons by means of two fragment
superfluids [Figs. 2(c1) and 2(c2)]. It is known that the fundamental entity
of a fermionic superfluid is the Cooper pair \cit
{Ramachandhran,Adhikari5,Gubbels} and there exists attractive FF interaction
between the fermion with spin up and that with spin down. Thus the
center-of-mass momentum of each Cooper pair in the fermionic superfluid
remains zero, irrespective of the concrete value of $a_{f}$, which may
account for the above differences.
On the other hand, the equilibrium properties of the SBFMs are also
evidently different from those of two-component BECs \cit
{Wen4,Ho2,Pu,Trippenbach,LWen}. For the case of two-component BECs, by
varying the particle numbers of the two BECs or the ratios of intra- and
intercomponent interaction strengths, one can obtain many symmetric
separated phases \cite{Ho2,Pu,LWen}\ and even asymmetric separated phases
\cite{Wen4,Trippenbach}. Physically, there exist three kinds of actual $s
-wave interactions (including BB interaction, BF interaction, and FF
interaction) in a SBFM. In particular, when the bosonic and fermionic
scattering lengths $a_{b}$ and $\left\vert a_{f}\right\vert $ tend to
infinity (i.e., in the strong-coupling unitarity limit), both the bosonic
and fermionic interactions exhibit unitarity saturation effects due to the
constraints of quantum mechanics \cite{Adhikari1}. In this context, the
quasiparticle of a Cooper pair in the fermionic superfluid can just be
considered as a \textquotedblleft composite boson\textquotedblright\ rather
than a single bosonic atom, which largely results in the differences of the
ground-state structures.
In Fig. 3, we display the steady density profiles $\left\vert \psi
_{b}\right\vert ^{2}$ (row $1$) and $\left\vert \psi _{p}\right\vert ^{2}$
(row $2$), and the corresponding phase profiles of $\psi _{b}$ (row $3$) and
$\psi _{p}$ (row $4$) at $t=500$ for a harmonic trap rotating with $\Omega
=0.94$. Here the value of the phase varies continuously from $0$ to $2\pi $,
and the end point of the boundary between a $2\pi $ phase line and a $0$
phase line denotes a phase defect (i.e., a vortex with anticlockwise
rotation). The density profiles of three initial states corresponding to the
three columns\ (left, middle, and right) in Fig. 3 are given in Figs.
2(a)--2(c), respectively.
\begin{figure}[tbp]
\centerline{\includegraphics*[width=8cm]{Figure3.eps}}
\caption{(Color online) Steady density profiles (the top two rows) and phase
profiles (the bottom two rows) at $t=500$ after rotating the superfluid
Bose-Fermi mixture with $\Omega =0.94$, \ where $1$ and $3$ denote the
bosonic superfluid, while $2$ and $4$ represent the fermionic superfluid.
The initial states corresponding to three cases of $a,b,$ and $c$ (from left
to right) are illustrated in Figs. 2(a1,a2), Figs. 2(b1,b2), and Figs.
2(c1,c2), respectively. The other parameters are the same as those in Figs.
1 and 2. The value of phase varies continuously from $0$ to $2\protect\pi $.
The darker color area indicates lower density or phase. Here, $x$ and $y$
are in units of $d_{0}$. }
\label{Figure3}
\end{figure}
As shown in Figs. 3(a1) and 3(a3), when the rotating SBFM with an initial
state of $R_{1}=-0.4$ and $R_{2}=-20$ reaches an equilibrium state, one can
see that there are several visible vortices \cite{Wen1,Wen2} constituting a
triangular lattice in the periphery of the bosonic superfluid. In the
meantime, the trap center is occupied by a density peak of bosons due to the
attractive BF interaction. However, no visible vortex can be generated in
the density of the fermionic superfluid even for large rotation frequency of
$\Omega =0.94$ [see Fig. 3(a2)]. From Fig. 3(a4), there exist some phase
defects that are located on the outskirts of the superfluid fermions. Since
these phase defects are invisible in the \textit{in situ} density profile of
fermionic superfluid and contribute to neither the angular momentum nor the
energy of the SBFM, there are referred to as ghost vortices \cit
{Wen1,Tsubota,Kasamatsu1,Wen2,Wen4}. For the case of $R_{1}=1$ and
R_{2}=-20 $, there is a large circular density hole of bosons in the trap
center which looks like a giant vortex (a multiquantized vortex) and is
surrounded by an expected visible vortex lattice [Fig. 3(b1)]. In the phase
profile displayed in Fig. 3(b3), we find that there are three phase defects
in the trap center which are close to each other. The three singly quantized
phase defects show that the circular density hole is not a giant vortex.
Actually, they are known as singly quantized hidden vortices \cit
{Wen1,Wen2,Wen4,Mithun} because they carry significant angular momentum,
though they are invisible in the \textit{in situ} density profile of the
BEC. Only after including the hidden vortices can the Feynman rule be
satisfied \cite{Wen1,Fetter,Cooper}. As seen in Figs. 3(b2) and 3(b4), the
fermionic superfluid with ghost vortices distributing on the outskirts of
the cloud is completely pulled inside the trap center (i.e., the region of
the large density hole of the BEC) due to the competition between the BF
repulsion and the rotation repulsion, which indicates a fully separated
phase of the rotating SBFM.
With the further increase of the BF repulsive interaction, a nonrotating
SBFM will develop into an inlaid separated configuration, where the
superfluid fermions lie on the outer edges of the bosons via the form of two
fragments [Figs. 2(c1) and 2(c2)]. Counterintuitively, we find that the
steady structure of a rotating SBFM with $R_{1}=6$ and $R_{2}=-20$ is
similar to that with $R_{1}=1$ and $R_{2}=-20,$ as shown in Figs.
3(c1)--3(c4). Here the fermionic superfluid is separately repelled to the
trap center region rather than the outskirts of BEC, which is quite
different from the nonrotating case. In the presence of dissipation, the
steady visible vortex lattice in the BEC is mainly formed by the competition
between the rotating driving and the BB repulsive interaction. The BF
repulsion and the centrifugal force acting on the bosons tend to push the
fermionic superfluid toward the outside or the inside. Considering as well
the coherence of the fermionic superfluid with FF attractive interaction, it
evidently prefers to occupy the trap center, especially in the presence of a
central density hole for the bosons, because the corresponding potential
energy for the superfluid fermions is relatively small such that the system
energy of the SBFM reaches the minimum.
In Figs. 4(a1) and 4(a2), we show the density profiles of the BEC and the
fermionic superfluid in a static SBFM with $N_{b}=10^{3}$, $N_{p}=10^{5}$,
a_{b}=5$ nm, $R_{1}=6$, and $R_{2}=-20$. \ Compared with the layer-separated
phase in Figs. 2(a1) and 2(a2), here the torus density profile of the BEC
becomes thinner while the circular density distribution of the fermionic
superfluid gets larger. The steady structure at $t=500$ of the SBFM rotating
with $\Omega =0.92$ is displayed in Figs. 4(b1) and 4(b2), and the
corresponding phase profiles are given in Figs. 4(c1) and 4(c2). From Figs.
4(b1) and 4(c1), we can see that there are no visible vortices, but there
exist some hidden vortices in the BEC, where the trap center is occupied by
multiple close singly quantized hidden vortices. Here the central blurred
region in the phase profile mainly results from the inevitable numerical
errors or fluctuations in the numerical computations. In contrast to Fig. 3,
a triangular visible vortex lattice made of six visible vortices forms in
the rotating fermionic superfluid, as shown in Figs. 4(b2) and 4(c2).
Although there is attractive FF interaction between two fermions in a Cooper
pair, the fermionic superfluid can display a weak effective repulsive
interaction between the Cooper pairs under the appropriate parameters
according to Eqs. (\ref{SFE}), (\ref{ChemPotentialFermion}), (\re
{FermionFunction}), and (\ref{DimLess2DSFE}), which is consistent with the
relevant theoretical prediction \cite{Pieri}. This point may explain why
visible vortices can be generated in the rotating fermionic superfluid.
\begin{figure}[tbp]
\centerline{\includegraphics*[width=7.8cm]{Figure4.eps}}
\caption{(Color online) (a1), (a2) Ground-state density profiles of a static
superfluid $^{7}$Li-$^{6}$Li mixture with $N_{b}=10^{3}$ and $N_{p}=10^{5}$.
(b1), (b2) The steady density profiles at $t=500$ for the superfluid
Bose-Fermi system rotating with $\Omega =0.92$. (c1), (c2) The corresponding
phase profiles. The relevant parameters are $a_{b}=5$ nm, $R_{1}=6$, and
R_{2}=-20$. Here, $1$ denotes the bosonic superfluid, $2$ represents the
fermionic superfluid, and $x$ and $y$ are in units of $d_{0}$. The darker
color area indicates lower density or phase. }
\label{Figure4}
\end{figure}
\section{Dynamics of vortex formation in a rotating SBFM}
In order to reveal the vortex formation process in the rotating superfluid
mixture of bosons and Cooper pairs, we consider a bigger superfluid
Bose-Fermi system which requires more computation effort. In Figs. 5(a1) and
5(a2), we present the ground-state structure of a static SBFM, where the
parameters are $N_{b}=10^{4}$, $N_{p}=10^{5}$, $a_{b}=50$ nm, $R_{1}=6$, and
$R_{2}=-0.1$. The steady structure at $t=500$ of the SBFM rotating with
\Omega =0.96$ is shown in Figs. 5(b1) and 5(b2), and the corresponding phase
profiles are displayed in Figs. 5(c1) and 5(c2). The initial state of the
SBFM at $t=0$ has a shell-shaped separated structure with the fermionic
superfluid being surrounded by the bosonic superfluid, which is similar to
that in Figs. 2(b1) and 2(b2). In the presence of dissipation, a steady
visible vortex lattice forms eventually in the outer of the superfluid BEC
[Fig. 5(b1)], where the energy of the rotating SBFM reaches the minimum in
the rotating frame. This point is similar to the case of Fig. 3(b1). The
large density hole of the bosons corresponds to 22 singly quantized hidden
vortices indicated by the phase profile in Fig. 5(c1). The central blurred
region in the phase profile is mainly caused by the inevitable numerical
errors or fluctuations in the numerical computations. In contrast to the
case of $N_{b}=1000\ $and $N_{p}=100$ (Figs. 2 and 3), seven evident visible
vortices are generated in the fermionic superfluid [see Figs. 5(b2) and
5(c2)].
\begin{figure}[tbp]
\centerline{\includegraphics*[width=8.4cm]{Figure5.eps}}
\caption{(Color online) (a1), (a2) Ground-state density profiles of a static
superfluid $^{7}$Li-$^{6}$Li mixture with $N_{b}=10^{4}$ and $N_{p}=10^{5}$.
(b1), (b2) The steady density profiles at $t=500$ for the superfluid
Bose-Fermi system rotating with $\Omega =0.96$. (c1), (c2) The corresponding
phase profiles. The relevant parameters are $a_{b}=50$ nm, $R_{1}=6$, and
R_{2}=-0.1$. Here, $1$ denotes the bosonic superfluid, $2$ represents the
fermionic superfluid, and $x$ and $y$ are in units of $d_{0}$. The darker
color area indicates lower density or phase. }
\label{Figure5}
\end{figure}
The dynamical evolution of the rotating SBFM is illustrated in Fig. 6, where
the top two rows denote the time evolutions of the density profiles of the
BEC (row 1) and fermionic superfluid (row 2), while the bottom two rows
represent those of the phase profiles of the BEC (row 3) and fermionic
superfluid (row 4). The evolution times are $t=11.5$ (left), $t=20$
(middle), and $t=60$ (right), respectively. Initially, the superfluid
densities $\left\vert \psi _{b}(x,y,t=0)\right\vert ^{2}$ and $\left\vert
\psi _{p}(x,y,t=0)\right\vert ^{2}$ in a stationary isotropic harmonic trap
are shown in Figs. 5(a1) and 5(a2). With the development of time, the
boundary surface of the bosonic superfluid undergo complex turbulent
oscillation and many ghost vortices appear at the outskirts of the BEC,
which can be seen in Figs. 6(a1) and 6(a3). In contrast, the fermionic
superfluid remains basically unchanged, as shown in Figs. 6(a2) and 6(a4).
Essentially, these ghost vortices are generated by collective excitations
through the nonlinear atomic interactions and the Landau instability
associated with the negative excitation frequency \cite{Kasamatsu1,Wu}
because the rotating harmonic trap is isotropic and has rotation symmetry.
This characteristic is evident especially for the component of fermionic
superfluid [see Figs. 6(a2)--6(c2) and Figs. 6(a4)--6(c4)]. Thus here the
formation mechanism of topological defects is different from the case of a
rotating anisotropic harmonic potential, where ghost vortices mainly result
from the dynamical instability through the rapid modulation of trapping
anisotropy \cite{Tsubota,Kasamatsu1,Dalfovo,Madison}. The angular momentum
is transferred into the SBFM via the excitations of surface modes or the
generation of visible vortices and hidden vortices. With the further time
evolution, some ghost vortices penetrate into the BEC and become visible
vortices or hidden vortices due to the Landau instability [see Figs.
6(b1)--6(c1) and Figs. 6(b3)--6(c3)], where the visible vortices arrange
themselves irregularly. When the rotating SBFM reaches an equilibrium state,
the visible vortices form a triangular lattice such that the energy of the
system approaches the minimum in the rotating frame. Compared with the
rotating BEC, the vortex (including ghost vortex and visible vortex)
formation in the rotating fermionic superfluid exhibits an evident
hysteresis effect, as shown in Figs. 6 and 5.
\begin{figure}[tbp]
\centerline{\includegraphics*[width=8.4cm]{Figure6.eps}}
\caption{(Color online) Time evolutions of the density profiles $\left\vert
\protect\psi _{b}\right\vert ^{2}$ (row 1) and $\left\vert \protect\psi
_{p}\right\vert ^{2}$ (row 2) and the corresponding phase profiles (rows 3
and 4) after the superfluid Bose-Fermi system rotates with $\Omega =0.96$,
where the other parameters are the same as those in Fig. 5. The value of the
phase varies continuously from $0$ to $2\protect\pi $. The time is (a1-a4)
t=11.5$, (b1--b4) $t=20$, and (c1--c4) $t=60$. The dark color area indicates
the lower density or phase. Here $x$ and $y$ are in units of $d_{0}$, and $t$
is in units of $1/\protect\omega _{bf}$.}
\label{Figure6}
\end{figure}
The structure and dynamics of a rotating SBFM are remarkably different from
those of rotating two-component BECs. First, in the latter case, the
vortices in the two BECs are always generated simultaneously \cit
{Kasamatsu2,Woo,Mason}, while here the vortex formation in the superfluid
fermions is far later than that in the BEC due to the presence of Cooper
pair with attractive FF interaction. Second, in the presence of dissipation,
there are triangular visible vortex lattices and singly quantized hidden
vortices formed in a rotating SBFM. In our simulation, we did not observe
the generation of vortex sheet or square vortex lattice occurring in a
rotating two-component BEC \cite{Kasamatsu2,Mason,Kuopanportti,Ghazanfari}.
These structures as well as the giant vortex are suppressed by the
dissipation term in our simulation.
\section{Conclusion}
In summary, we have studied the structures for rotating and non-rotating
superfluid Bose-Fermi mixtures and the dynamics of vortex formation in a
rotating superfluid Bose-Fermi mixture. We show that the ratio of BF
interaction to BB interaction plays a key role in determining the
ground-state structure of a nonrotating SBFM. For fixed BB and BF $s$-wave
scattering lengths, the variation of FF (spin-up and -down) $s$-wave
scattering length does not influence the phase diagram for a nonrotating
SBFM. Furthermore, depending on the choice of parameters, we find that a
rotating SBFM with a sufficiently large angular velocity supports four
typical steady structures: a mixed phase and three layer-separated phases.
In addition, it is shown that different separated phases at the initial time
may result in almost the same steady structures. In particular, we find that
the generation of visible vortices in the fermionic superfluid exhibits an
evident hysteresis effect during the time evolution of a rotating SBFM. Our
results indicate that the topological structure and the dynamics of a
rotating SBFM are remarkably different from the cases of rotating
two-component BECs. We expect that our findings can be observed and tested
in future experiments. In the mean time, the present investigation provides
a way to further test the validity of the unitarity model \cite{Adhikari1}
for a SBFM.
\begin{acknowledgments}
We thank Biao Wu, Yongping Zhang, Li Mao and Yong Xu for helpful
discussions. L.W. acknowledges the research group of Professor Chuanwei
Zhang at The University of Texas at Dallas, where part of the computations
were carried out. This work was supported by the National Natural Science
Foundation of China (Grants No. 11475144, No. 11047033,and No. 11304270),
and the Ph.D. Foundation of Yanshan University (No. B846).
\end{acknowledgments}
|
\section{Introduction}
\label{sec:introduction}
Recently, Hersh and M{\'e}sz{\'a}ros introduced a new class of lattices, so-called SB-lattices~\cite{hersh14sb}. They showed that these lattices admit a certain edge-labeling, which implies that the order complex of every open interval of this lattice is homotopy equivalent to a sphere or a ball (hence the name). Equivalently, the M{\"o}bius function of such a lattice takes values only in $\{-1,0,1\}$. In the same paper they showed that every distributive lattice admits an SB-labeling, and they showed that the same is true for the weak order on a Coxeter group and for the Tamari lattice. We extend their results by showing that another class of lattices, so-called join-semidistributive lattice, belong to the class of SB-lattices as well.
Subsequently, we investigate a particular family of join-semidistributive lattices, namely the set of $\gamma$-sortable elements of a Coxeter groups equipped with Bruhat order. We denote this poset by $\mathcal{B}_{\gamma}$. The $\gamma$-sortable elements were defined by Reading and Speyer, see \cites{reading07sortable,reading11sortable}. To summarize, the first main result of this paper is the following.
\begin{theorem}\label{thm:main_sb}
Every join-distributive lattice is an SB-lattice. In particular, the Bruhat order on $\gamma$-sortable elements is an SB-lattice for every Coxeter group $W$ and every Coxeter element $\gamma\in W$.
\end{theorem}
Join-distributivity is a generalization of distributivity, and while working with the poset $\mathcal{B}_{\gamma}$ we observed that for some Coxeter groups and for some Coxeter elements the lattice $\mathcal{B}_{\gamma}$ is in fact distributive. This led us to the question whether we can characterize the (finite) Coxeter groups $W$ and the Coxeter elements $\gamma\in W$ for which $\mathcal{B}_{\gamma}$ is distributive. We approach this problem by looking for forbidden orientations of the Coxeter diagram of $W$, and we prove the following result.
\begin{theorem}\label{thm:main_distributive}
Let $W$ be a finite Coxeter group. There exists a Coxeter element $\gamma\in W$ such that $\mathcal{B}_{\gamma}$ is distributive if and only if $W$ is of type $A_{n},B_{n},H_{3}$ or $I_{2}(k)$.
\end{theorem}
The finite Coxeter groups appearing in Theorem~\ref{thm:main_distributive} are sometimes called the ``coincidental types'', since these groups enjoy a list of properties that distinguishes them from the other finite (complex) reflection groups, see \cite{fomin05generalized}*{Theorems~8.5~and~10.2}, \cite{miller15foulkes}*{Theorem~14}, \cite{reading08chains}, and \cite{williams13cataland}*{Remark~3.1.26}. Theorem~\ref{thm:main_distributive} adds another property to this list. We conclude this article with a conjectural characterization of the Coxeter elements $\gamma\in W$ for which $\mathcal{B}_{\gamma}$ is distributive.
\smallskip
This article is organized as follows: in Section~\ref{sec:preliminaries} we define SB-lattices, Coxeter groups and sortable elements. We subsequently prove Theorem~\ref{thm:main_sb} in Section~\ref{sec:join_distributive lattices}, and define the poset $\mathcal{B}_{\gamma}$. We conclude this paper by proving Theorem~\ref{thm:main_distributive} in Section~\ref{sec:distributivity}.
\section{Preliminaries}
\label{sec:preliminaries}
In this section, we recall the basic concepts needed in this article. For more background on SB-labelings, we refer to \cite{hersh14sb}. For any undefined notation and additional information on Coxeter groups and sortable elements, we refer to \cite{bjorner05combinatorics} and \cite{reading11sortable}, respectively.
\subsection{SB-Labelings}
\label{sec:sb_labelings}
Let $\mathcal{P}=(P,\leq)$ be a (possibly infinite) poset. An element $p\in P$ is \alert{covered by} another element $q\in P$ (denoted by $p\lessdot q$) if $p<q$ and there exists no $z\in P$ with $p<z<q$. Accordingly, $q$ \alert{covers} $p$ and the elements $p$ and $q$ form a \alert{cover relation} or an \alert{edge} in $\mathcal{P}$. The set $\mathcal{E}(\mathcal{P})=\bigl\{(p,q)\mid p\lessdot q\bigr\}$ is the \alert{Hasse diagram} of $\mathcal{P}$.
For $p\leq q$ we call a set of the form $[p,q]=\{z\in P\mid p\leq z\leq q\}$ a \alert{(closed) interval} of $\mathcal{P}$. A \alert{saturated chain} in an interval $[p,q]$ is a sequence $(p,z_{1},z_{2},\ldots,z_{k-1},q)$, where $z_{i}\in P$ for $i\in\{1,2,\ldots,k-1\}$ and $p\lessdot z_{1}\lessdot z_{2}\lessdot\cdots\lessdot z_{k-1}\lessdot q$.
A poset $\mathcal{P}$ is a \alert{lattice} if any two elements in $\mathcal{P}$ have a least upper bound (a \alert{join}) and a greatest lower bound (a \alert{meet}), denoted by $\vee$ and $\wedge$, respectively. A lattice is \alert{locally finite} if every interval is finite.
An \alert{edge-labeling} of $\mathcal{P}$ is a map $\lambda:\mathcal{E}(\mathcal{P})\to\Lambda$, for some set of labels $\Lambda$. An \alert{SB-labeling} of a lattice $\mathcal{P}$ is an edge-labeling $\lambda$ of $\mathcal{P}$ that satisfies the following properties for every $p,p_{1},p_{2}\in P$ with $p\lessdot p_{1},p_{2}$:
\begin{enumerate}[(i)]
\item $\lambda(p,p_{1})\neq\lambda(p,p_{2})$;
\item each saturated chain in the interval $[p,p_{1}\vee p_{2}]$ uses both labels $\lambda(p,p_{1})$ and $\lambda(p,p_{2})$ a positive number of times;\quad and
\item none of the saturated chains in the interval $[p,p_{1}\vee p_{2}]$ uses any other label besides $\lambda(p,p_{1})$ and $\lambda(p,p_{2})$.
\end{enumerate}
A locally finite lattice with a least element that admits an SB-labeling is called an \alert{SB-lattice}.
\begin{remark}
In fact, the original definition of an SB-labeling given in \cite{hersh14sb}*{Definition~3.2} was phrased a bit differently, but it was shown in \cite{hersh14sb}*{Theorem~3.5} that the above definition is equivalent to the original definition.
\end{remark}
SB-lattices enjoy the following nice property.
\begin{theorem}[\cite{hersh14sb}*{Theorem~3.8}]\label{thm:sb_mobius}
The M{\"o}bius function of an SB-lattice takes values only in $\{-1,0,1\}$.
\end{theorem}
\subsection{Coxeter Groups}
\label{sec:coxeter_groups}
A \alert{Coxeter group} is a group $W$ admitting a presentation
\begin{displaymath}
W = \bigl\langle s_{1},s_{2},\ldots,s_{n}\mid (s_{i}s_{j})^{m_{i,j}}=\varepsilon,\;\text{for}\;i,j\in\{1,2,\ldots,n\}\bigr\rangle,
\end{displaymath}
where $\varepsilon\in W$ is the identity and the numbers $m_{i,j}$ are either positive integers or the formal symbol $\infty$ for all $i,j\in\{1,2,\ldots,n\}$ such that $m_{i,j}\geq 2$ if $i\neq j$, and $m_{i,i}=1$. (We use the convention that $\infty$ is formally larger than every integer.) The elements in $S=\{s_{1},s_{2},\ldots,s_{n}\}$ are the \alert{Coxeter generators} of $W$, and $n$ is the \alert{rank} of $W$. A subgroup of $W$ that is generated by a subset $J\subseteq S$ is a Coxeter group in its own right and is called a \alert{standard parabolic subgroup} of $W$. A Coxeter group is called \alert{irreducible} if it is not isomorphic to a direct product of Coxeter groups of smaller rank. The finite irreducible Coxeter groups were completely classified by Coxeter in \cite{coxeter35complete}. This classification is best visualized using so-called Coxeter diagrams. The \alert{Coxeter diagram} of $W$, denoted by $\Gamma(W)$, is a labeled graph whose vertices are the Coxeter generators of $W$, and two generators $s_{i}$ and $s_{j}$ are connected by an edge if and only if $m_{i,j}\geq 3$. In addition, the edge between $s_{i}$ and $s_{j}$ is labeled by $m_{i,j}$ if $m_{i,j}\geq 4$. It is not hard to see that $W$ is irreducible if and only if its Coxeter diagram is connected. Figure~\ref{fig:coxeter_diagrams} shows the Coxeter diagrams of the finite irreducible Coxeter groups.
\begin{figure}
\centering
\begin{tabular}{c|l}
Name & Coxeter diagram\\
\hline
$A_{n},\quad n\geq 1$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,.2) node{};
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(3.67,0) node(sd){$\cdots$};
\draw(4.5,0) node(sn1){$s_{n-1}$};
\draw(6,0) node(sn){$s_{n}$};
\draw[thick](s1) -- (s2) -- (s3) -- (sd) -- (sn1) -- (sn);
\end{tikzpicture}}\\
$B_{n},\quad n\geq 2$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,.2) node{};
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(3.67,0) node(sd){$\cdots$};
\draw(4.5,0) node(sn1){$s_{n-1}$};
\draw(6,0) node(sn){$s_{n}$};
\draw(5.33,.15) node{\tiny $4$};
\draw[thick](s1) -- (s2) -- (s3) -- (sd) -- (sn1) -- (sn);
\end{tikzpicture}}\\
$D_{n},\quad n\geq 4$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(3.67,0) node(sd){$\cdots$};
\draw(4.5,0) node(sn2){$s_{n-2}$};
\draw(4.5,.75) node(sn1){$s_{n-1}$};
\draw(6,0) node(sn){$s_{n}$};
\draw[thick](s1) -- (s2) -- (s3) -- (sd) -- (sn2) -- (sn1);
\draw[thick](sn2) -- (sn);
\end{tikzpicture}}\\
$E_{6}$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(3,.75) node(s6){$s_{6}$};
\draw(4.5,0) node(s4){$s_{4}$};
\draw(6,0) node(s5){$s_{5}$};
\draw[thick](s1) -- (s2) -- (s3) -- (s4) -- (s5);
\draw[thick](s3) -- (s6);
\end{tikzpicture}}\\
$E_{7}$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(3,.75) node(s7){$s_{7}$};
\draw(4.5,0) node(s4){$s_{4}$};
\draw(6,0) node(s5){$s_{5}$};
\draw(7.5,0) node(s6){$s_{6}$};
\draw[thick](s1) -- (s2) -- (s3) -- (s4) -- (s5) -- (s6);
\draw[thick](s3) -- (s7);
\end{tikzpicture}}\\
$E_{8}$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(3,.75) node(s8){$s_{8}$};
\draw(4.5,0) node(s4){$s_{4}$};
\draw(6,0) node(s5){$s_{5}$};
\draw(7.5,0) node(s6){$s_{6}$};
\draw(9,0) node(s7){$s_{7}$};
\draw[thick](s1) -- (s2) -- (s3) -- (s4) -- (s5) -- (s6) -- (s7);
\draw[thick](s3) -- (s8);
\end{tikzpicture}}\\
$F_{4}$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,.2) node{};
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(4.5,0) node(s4){$s_{4}$};
\draw(2.25,.15) node{\tiny $4$};
\draw[thick](s1) -- (s2) -- (s3) -- (s4);
\end{tikzpicture}}\\
$H_{3}$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,.2) node{};
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(2.25,.15) node{\tiny $5$};
\draw[thick](s1) -- (s2) -- (s3);
\end{tikzpicture}}\\
$H_{4}$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,.2) node{};
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(3,0) node(s3){$s_{3}$};
\draw(4.5,0) node(s4){$s_{4}$};
\draw(3.75,.15) node{\tiny $5$};
\draw[thick](s1) -- (s2) -- (s3) -- (s4);
\end{tikzpicture}}\\
$I_{2}(k),\quad k\geq 5$ &
\raisebox{-.2cm}{\begin{tikzpicture}\small
\draw(4,.2) node{};
\draw(4,-.2) node{};
\draw(0,0) node(s1){$s_{1}$};
\draw(1.5,0) node(s2){$s_{2}$};
\draw(.75,.15) node{\tiny $k$};
\draw[thick](s1) -- (s2);
\end{tikzpicture}}\\
\end{tabular}
\caption{The Coxeter diagrams of the finite irreducible Coxeter groups.}
\label{fig:coxeter_diagrams}
\end{figure}
Since $S$ is a generating set of $W$, we can write every $w\in W$ as a product of Coxeter generators. The least number of generators needed to form $w$, is called the \alert{Coxeter length} of $w$, and will be written as $\ell_{S}(w)$. We say that a word $w=s_{i_{1}}s_{i_{2}}\cdots s_{i_{k}}$ is \alert{reduced} if $\ell_{S}(w)=k$.
\subsection{Sortable Elements}
\label{sec:sortable_elements}
Let $W$ be a Coxeter group of rank $n$. An element $\gamma\in W$ is called a \alert{Coxeter element} of $W$ if $\gamma=s_{\pi(1)}s_{\pi(2)}\cdots s_{\pi(n)}$ for some permutation $\pi$ of $\{1,2,\ldots,n\}$. Without loss of generality, we can restrict our attention to the Coxeter element $\gamma=s_{1}s_{2}\cdots s_{n}$. Consider the half-infinite word
\begin{displaymath}
\gamma^{\infty}=s_{1}s_{2}\cdots s_{n}\vert s_{1}s_{2}\cdots s_{n}\vert s_{1}\cdots.
\end{displaymath}
The vertical bars have no influence on the structure of the word, but shall serve for a better readability. Clearly, for every $w\in W$, every reduced word for $w$ can be written as a subword of $\gamma^{\infty}$. We call the lexicographically first subword of $\gamma^{\infty}$ that is a reduced word for $w$, the \alert{$\gamma$-sorting word} of $w$, and we denote it by $\gamma(w)$. We can write
\begin{displaymath}
\gamma(w) = s_{1}^{\delta_{1,1}}s_{2}^{\delta_{1,2}}\cdots s_{n}^{\delta_{1,n}}\vert s_{1}^{\delta_{2,1}}s_{2}^{\delta_{2,2}}\cdots s_{n}^{\delta_{2,n}}\vert\cdots\vert s_{1}^{\delta_{l,1}}s_{2}^{\delta_{l,2}}\cdots s_{n}^{\delta_{l,n}},
\end{displaymath}
for $l\in\mathbb{N}$ and $\delta_{i,j}\in\{0,1\}$ for $i\in\{1,2,\ldots,l\}$ and $j\in\{1,2,\ldots,n\}$. The \alert{$i$-th block} of $w$ is the set $b_{i}=\{s_{j}\mid\delta_{i,j}=1\}$. We say that $w$ is \alert{$\gamma$-sortable} if $b_{1}\supseteq b_{2}\supseteq\cdots\supseteq b_{l}$, and we write $C_{\gamma}$ for the set of $\gamma$-sortable elements of $W$. Further, define the set of filled positions of $w$ by
\begin{displaymath}
\alpha_{\gamma}(w)=\bigl\{(i-1)n+j\mid \delta_{i,j}=1\bigr\}.
\end{displaymath}
We notice that $\alpha_{\gamma}$ depends on the choice of reduced word for $\gamma$, while $C_{\gamma}$ does not.
\begin{remark}\label{rem:sortable_origin}
The concept of $\gamma$-sortability was introduced by Reading in \cite{reading07sortable} as a generalization of stack-sortability, and was used to define the family of Cambrian lattices associated with a Coxeter group \cites{reading06cambrian,reading11sortable}. The number of $\gamma$-sortable elements of a finite Coxeter group $W$ is the $W$-Catalan number, defined in \cite{bessis03dual}*{Section~5.2}, and Reading used the $\gamma$-sortable elements to provide a bridge between the noncrossing partitions of $W$ and the clusters of $W$, see \cite{reading07clusters}.
The concept of $\gamma$-sortability has been further extended by Armstrong in \cite{armstrong09sorting}, where he defined $\omega$-sortability for an arbitrary, not necessarily reduced word $\omega$ in the Coxeter generators of $W$, and if $\omega=\gamma^{\infty}$, then one obtains precisely the $\gamma$-sortable elements.
\end{remark}
\section{Join-Distributive Lattices}
\label{sec:join_distributive lattices}
In this section we prove the first part of Theorem~\ref{thm:main_sb}, namely that every join-distributive lattice admits an SB-labeling, see Theorem~\ref{thm:join_distributive_sb} below. Let us first recall the necessary definitions. A lattice $\mathcal{P}=(P,\leq)$ is \alert{meet-semidistributive} if for every three elements $p,q,r\in P$ with $p\wedge q=p\wedge r$ we have $p\wedge q=p\wedge(q\vee r)$. Moreover, $\mathcal{P}$ is \alert{upper semimodular} if for every two elements $p,q\in P$ with $p\wedge q\lessdot p,q$, we have $p,q\lessdot p\vee q$. Then, a lattice is \alert{join-distributive} if it is both meet-semidistributive and upper semimodular. See \cite{edelman80meet} for more information on join-distributive lattices.
Recall further that an \alert{antimatroid} is a pair $(M,\mathcal{F})$, where $M$ is a set and $\mathcal{F}\subseteq\wp(M)$ is a family of subsets of $M$ that satisfies the following properties:
\begin{enumerate}[(i)]
\item $\emptyset\in\mathcal{F}$; \quad and
\item if $X,Y\in\mathcal{F}$ with $Y\not\subseteq X$, then there exists some $x\in X\setminus Y$ such that $X\cup\{x\}\in\mathcal{F}$.
\end{enumerate}
The elements of $\mathcal{F}$ are called the \alert{feasible sets} of $(M,\mathcal{F})$. We have the following result.
\begin{theorem}[\cite{edelman80meet}*{Theorem~3.3}]\label{thm:join_distributive_antimatroid}
A lattice $\mathcal{P}$ is join-distributive if and only if there exists an antimatroid $(M,\mathcal{F})$ such that $\mathcal{P}\cong(\mathcal{F},\subseteq)$.
\end{theorem}
In view of this correspondence we can now conclude the following result.
\begin{theorem}\label{thm:join_distributive_sb}
Every join-distributive lattice admits an SB-labeling.
\end{theorem}
\begin{proof}
Let $\mathcal{P}$ be a join-distributive lattice. In view of Theorem~\ref{thm:join_distributive_antimatroid}, we can view $\mathcal{P}$ as a lattice of feasible sets of some antimatroid $(M,\mathcal{F})$, and thus every edge in $\mathcal{P}$ is determined by a pair $X,Y\in\mathcal{F}$ with $Y\setminus X=\{x\}$. This induces an edge-labeling of $\mathcal{P}$, which we will denote by $\lambda_{\mathcal{F}}$. Since $\mathcal{P}$ is upper-semimodular, it follows that for any $p,p_{1},p_{2}\in P$ with $p\lessdot p_{1},p_{2}$, where we write $\bar{p}=p_{1}\vee p_{2}$, we have $p_{1},p_{2}\lessdot\bar{p}$. Since $\mathcal{P}$ is meet-semidistributive, it follows that the interval $[p,\bar{p}]$ consists only of the four elements $p,p_{1},p_{2},\bar{p}$. Thus we have $\lambda_{\mathcal{F}}(p,p_{1})=\lambda_{\mathcal{F}}(p_{2},\bar{p})$ and $\lambda_{\mathcal{F}}(p,p_{2})=\lambda_{\mathcal{F}}(p_{1},\bar{p})$, as well as $\lambda_{\mathcal{F}}(p,p_{1})\neq\lambda_{\mathcal{F}}(p,p_{2})$. Hence $\lambda_{\mathcal{F}}$ is an SB-labeling of $\mathcal{P}$.
\end{proof}
\begin{corollary}\label{cor:join_semidistributive_mobius}
The M{\"o}bius function of a join-semidistributive lattice takes values only in $\{-1,0,1\}$.
\end{corollary}
\begin{proof}
This follows from Theorems~\ref{thm:sb_mobius} and \ref{thm:join_distributive_sb}.
\end{proof}
\begin{remark}
Join-distributivity can be seen as a generalization of distributivity, see \eqref{eq:meet_distributivity} and \eqref{eq:join_distributivity} below. In that sense, Theorem~\ref{thm:join_distributive_sb} generalizes \cite{hersh14sb}*{Theorem~5.1}, which states that every distributive lattice is an SB-lattice.
\end{remark}
All join-distributive lattices are by definition meet-semidistributive. Lattices that satisfy the meet-distributive law and the corresponding dual law are called \alert{semidistributive}. Obviously, every distributive lattice is also semidistributive, but semidistributive lattices need no longer be graded. It is known that the M{\"o}bius function of a semidistributive lattice takes values only in $\{-1,0,1\}$~\cite{farley13bijection}, and it would be interesting whether such lattices are always SB-lattices.
Another generalization of distributivity to ungraded lattices, so-called \alert{trimness}, was introduced by Thomas in \cite{thomas06analogue}. It is the statement of \cite{thomas06analogue}*{Theorem~7} that the M{\"o}bius function of a trim lattice also takes values only in $\{-1,0,1\}$, and again it would be interesting to know whether trim lattices are always SB-lattices.
An important example of lattices that belong to both previously mentioned classes of lattices are Reading's Cambrian semilattices, see \cite{reading11sortable}*{Theorem~8.1} and \cite{muehle15trimness}*{Theorem~1.1}. The Cambrian semilattices generalize the Tamari lattices to all Coxeter groups. Theorem~5.5 in \cite{hersh14sb} states that the Tamari lattices are SB-lattices. We could produce SB-labelings for some small Cambrian semilattices, but we could not find a uniform definition of such a labeling. We nevertheless pose the following conjecture.
\begin{conjecture}
Let $W$ be a Coxeter group, and let $\gamma\in W$ be a Coxeter element. The $\gamma$-Cambrian semilattice, \text{i.e.}\; the set $C_{\gamma}$ equipped with the weak order on $W$, is an SB-lattice.
\end{conjecture}
\begin{remark}
Recently, McConville has investigated a slightly weaker lattice property, namely \alert{crosscut-simpliciality} \cite{mcconville14crosscut}. It follows by definition that every SB-lattice is crosscut-simplicial. He showed in particular that meet-semidistributive lattices are crosscut-simplicial, see \cite{mcconville14crosscut}*{Theorem~1.2}.
\end{remark}
\subsection{The Bruhat Order on Sortable Elements}
\label{sec:bruhat}
In this section, we consider a special family of join-distributive lattices, namely the set of $\gamma$-sortable elements of a Coxeter group $W$ equipped with the Bruhat order, which we define next.
\begin{definition}
For $u,v\in W$ we write $u\leq_{B}v$ if and only if there exists a reduced word $v=a_{1}a_{2}\cdots a_{l}$ and indices $1\leq i_{1}<i_{2}<\cdots<i_{k}\leq l$ such that $u=a_{i_{1}}a_{i_{2}}\cdots a_{i_{k}}$. The partial order $\leq_{B}$ is called the \alert{Bruhat order} on $W$.
\end{definition}
Clearly the identity $\varepsilon$ is the least element with respect to $\leq_{B}$. Moreover, the poset $(W,\leq_{B})$ is graded by $\ell_{S}$, but it is in general not a lattice. Thus we restrict our attention to the subposet $\mathcal{B}_{\gamma}=(C_{\gamma},\leq_{B})$, and in what follows, we index poset-theoretic notions that refer to the Bruhat order on $\gamma$-sortable elements by ``B'', \text{i.e.}\; an interval in the poset $\mathcal{B}_{\gamma}$ will be denoted by $[u,v]_{B}$, and likewise for joins and meets. Recall that a lattice is \alert{finitary} if every principal order ideal is finite. In particular, finitary lattices are locally finite.
\begin{theorem}\label{thm:bruhat_lattice}
The poset $\mathcal{B}_{\gamma}$ is a finitary lattice for every Coxeter group $W$ and every Coxeter element $\gamma\in W$.
\end{theorem}
\begin{proof}
First of all, let $w\in C_{\gamma}$ with $\ell_{S}(w)=k$. The interval $[\varepsilon,w]_{B}$ is certainly finite, since $w$ has finite length. Moreover, using the terminology from above, it is easy to see that $w\leq_{B}w'$ if and only if $\alpha_{\gamma}(w)\subseteq\alpha_{\gamma}(w')$.
Let $u,u'\in C_{\gamma}$. The word $\bar{u}$ defined by $\alpha_{\gamma}(\bar{u})=\alpha_{\gamma}(u)\cup\alpha_{\gamma}(u')$ is again $\gamma$-sortable. In particular, $\bar{u}$ is the least upper bound for both $u$ and $u'$. Hence the interval $[\varepsilon,\bar{u}]_{B}$ is finite, and analogously to before we see that any two elements in this interval possess a join. Hence it is a classical lattice-theoretic result that $[\varepsilon,\bar{u}]_{B}$ is a lattice. It follows immediately that the meet of $u$ and $u'$ exists as well, and the proof is complete.
\end{proof}
Theorem~\ref{thm:bruhat_lattice} was already mentioned in \cite{armstrong09sorting}*{Section~6}. It should be remarked
that in general $\mathcal{B}_{\gamma}$ is an infinite lattice with no greatest element, which implies in particular that $\mathcal{B}_{\gamma}$ is no complete lattice. The following result is also implicit in \cite{armstrong09sorting}.
\begin{theorem}[\cite{armstrong09sorting}]\label{thm:bruhat_join_semidistributive}
The lattice $\mathcal{B}_{\gamma}$ is join-distributive for every Coxeter group $W$ and every Coxeter element $\gamma\in W$.
\end{theorem}
Hence the second part of Theorem~\ref{thm:main_sb} follows immediately from Theorem~\ref{thm:join_distributive_sb}.
\begin{corollary}\label{cor:bruhat_sb}
The lattice $\mathcal{B}_{\gamma}$ admits an SB-labeling for every Coxeter group $W$ and every Coxeter element $\gamma\in W$.
\end{corollary}
\begin{proof}
This follows from Theorems~\ref{thm:join_distributive_sb} and \ref{thm:bruhat_join_semidistributive}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main_sb}]
This follows from Theorem~\ref{thm:join_distributive_sb} and Corollary~\ref{cor:bruhat_sb}.
\end{proof}
\begin{corollary}\label{cor:bruhat_mobius}
The M{\"o}bius function of $\mathcal{B}_{\gamma}$ takes values only in $\{-1,0,1\}$ for every Coxeter group $W$ and every Coxeter element $\gamma\in W$.
\end{corollary}
The correspondence between join-distributive lattices and antimatroids allows us to associate an antimatroid with $\mathcal{B}_{\gamma}$. In general, this antimatroid is infinite, and thus the set of labels of the SB-labeling defined in the proof of Theorem~\ref{thm:join_distributive_sb} is potentially infinite. We notice that this labeling can be defined globally by
\begin{displaymath}
\lambda_{\gamma}:\mathcal{E}(\mathcal{B}_{\gamma})\to\mathbb{N},\quad (u,v)\mapsto\alpha_{\gamma}(v)\setminus\alpha_{\gamma}(u).
\end{displaymath}
An analogous labeling was used in \cite{kallipoliti13on} to prove topological properties of the Cambrian semilattices. Now consider the map
\begin{displaymath}
\eta:\mathbb{N}\to S,\quad i\mapsto \begin{cases}s_{n}, & \text{if}\;i\equiv 0\pmod{n},\\ s_{i\bmod{n}}, & \text{otherwise}.\end{cases}
\end{displaymath}
If we concatenate these two maps, then we obtain another SB-labeling of $\mathcal{B}_{\gamma}$, which in contrast to $\lambda_{\gamma}$ uses only a finite set of labels:
\begin{align}\label{eq:sb_bruhat}
b_{\gamma}:\mathcal{E}(\mathcal{B}_{\gamma})\to S,\quad (u,v)\mapsto\eta\bigl(\alpha_{\gamma}(v)\setminus\alpha_{\gamma}(u)\bigr).
\end{align}
\begin{example}
Figure~\ref{fig:bruhat_sortable} shows two Bruhat lattices labeled by the SB-labeling defined in \eqref{eq:sb_bruhat}. The lattice in Figure~\ref{fig:bruhat_sortable_a3} is associated with the Coxeter group $A_{3}$ and the Coxeter element given by the oriented Coxeter diagram
\raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,0) node(v3){$s_{3}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\end{tikzpicture}}. Figure~\ref{fig:bruhat_sortable_c2} shows the first seven ranks of the lattice associated with the affine Coxeter group $\tilde{C}_{2}$ subject to the Coxeter element given by the oriented Coxeter diagram
\raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(.5,.2) node{\tiny 4};
\draw(1,0) node(v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node(v3){$s_{3}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v3) -- (1.45,0);
\draw[thick](1.45,0) -- (v2);
\end{tikzpicture}}.
(See Section~\ref{sec:distributivity} for an explanation of the connection between Coxeter elements and orientations of the Coxeter diagram.)
\begin{figure}
\subfigure[$\mathcal{B}_{s_{1}s_{2}s_{3}}$ associated with $A_{3}$.]{\label{fig:bruhat_sortable_a3}
\begin{tikzpicture}\small
\def\x{1.6};
\def\y{1.2};
\draw(2*\x,1*\y) node(n1){$\varepsilon$};
\draw(1*\x,2*\y) node(n2){$s_{1}$};
\draw(2*\x,2*\y) node(n3){$s_{2}$};
\draw(3*\x,2*\y) node(n4){$s_{3}$};
\draw(1*\x,3*\y) node(n5){$s_{1}s_{2}$};
\draw(2*\x,3*\y) node(n6){$s_{1}s_{3}$};
\draw(3*\x,3*\y) node(n7){$s_{2}s_{3}$};
\draw(1*\x,4*\y) node(n8){$s_{1}s_{2}\vert s_{1}$};
\draw(2*\x,4*\y) node(n9){$s_{1}s_{2}s_{3}$};
\draw(3*\x,4*\y) node(n10){$s_{2}s_{3}\vert s_{2}$};
\draw(1.5*\x,5*\y) node(n11){$s_{1}s_{2}s_{3}\vert s_{1}$};
\draw(2.5*\x,5*\y) node(n12){$s_{1}s_{2}s_{3}\vert s_{2}$};
\draw(2*\x,6*\y) node(n13){$s_{1}s_{2}s_{3}\vert s_{1}s_{2}$};
\draw(2*\x,7*\y) node(n14){$s_{1}s_{2}s_{3}\vert s_{1}s_{2}\vert s_{1}$};
\draw(n1) -- (n2) node[fill=white] at(1.5*\x,1.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n1) -- (n3) node[fill=white] at(2*\x,1.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n1) -- (n4) node[fill=white] at(2.5*\x,1.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n2) -- (n5) node[fill=white] at(1*\x,2.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n2) -- (n6) node[fill=white] at(1.25*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n3) -- (n5) node[fill=white] at(1.75*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n3) -- (n7) node[fill=white] at(2.25*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n4) -- (n6) node[fill=white] at(2.75*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n4) -- (n7) node[fill=white] at(3*\x,2.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n5) -- (n8) node[fill=white] at(1*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n5) -- (n9) node[fill=white] at(1.5*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n6) -- (n9) node[fill=white] at(2*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n7) -- (n9) node[fill=white] at(2.5*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n7) -- (n10) node[fill=white] at(3*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n8) -- (n11) node[fill=white] at(1.25*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n9) -- (n11) node[fill=white] at(1.75*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n9) -- (n12) node[fill=white] at(2.25*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n10) -- (n12) node[fill=white] at(2.75*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n11) -- (n13) node[fill=white] at(1.75*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n12) -- (n13) node[fill=white] at(2.25*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n13) -- (n14) node[fill=white] at(2*\x,6.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\end{tikzpicture}}
\subfigure[The first seven ranks of $\mathcal{B}_{s_{1}s_{3}s_{2}}$ associated with $\tilde{C}_{2}$.]{\label{fig:bruhat_sortable_c2}
\begin{tikzpicture}\small
\def\x{1.6};
\def\y{1.2};
\draw(2*\x,1*\y) node(n1){$\varepsilon$};
\draw(1*\x,2*\y) node(n2){$s_{1}$};
\draw(2*\x,2*\y) node(n3){$s_{2}$};
\draw(3*\x,2*\y) node(n4){$s_{3}$};
\draw(1*\x,3*\y) node(n5){$s_{1}s_{2}$};
\draw(2*\x,3*\y) node(n6){$s_{1}s_{3}$};
\draw(3*\x,3*\y) node(n7){$s_{3}s_{2}$};
\draw(1*\x,4*\y) node(n8){$s_{1}s_{2}\vert s_{1}$};
\draw(2*\x,4*\y) node(n9){$s_{1}s_{3}s_{2}$};
\draw(3*\x,4*\y) node(n10){$s_{3}s_{2}\vert s_{3}$};
\draw(.5*\x,5*\y) node(n11){$s_{1}s_{2}\vert s_{1}s_{2}$};
\draw(1.5*\x,5*\y) node(n12){$s_{1}s_{3}s_{2}\vert s_{1}$};
\draw(2.5*\x,5*\y) node(n13){$s_{1}s_{3}s_{2}\vert s_{3}$};
\draw(3.5*\x,5*\y) node(n14){$s_{3}s_{2}\vert s_{3}s_{2}$};
\draw(1*\x,6*\y) node(n15){$s_{1}s_{3}s_{2}\vert s_{1}s_{2}$};
\draw(2*\x,6*\y) node(n16){$s_{1}s_{3}s_{2}\vert s_{1}s_{3}$};
\draw(3*\x,6*\y) node(n17){$s_{1}s_{3}s_{2}\vert s_{3}s_{2}$};
\draw(.5*\x,7*\y) node(n18){$s_{1}s_{3}s_{2}\vert s_{1}s_{2}\vert s_{1}$};
\draw(2*\x,7*\y) node(n19){$s_{1}s_{3}s_{2}\vert s_{1}s_{3}s_{2}$};
\draw(1*\x,8*\y) node(n20){$s_{1}s_{3}s_{2}\vert s_{1}s_{3}s_{2}\vert s_{1}$};
\draw(3*\x,8*\y) node(n21){$s_{1}s_{3}s_{2}\vert s_{1}s_{3}s_{2}\vert s_{3}$};
\draw(n1) -- (n2) node[fill=white] at(1.5*\x,1.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n1) -- (n3) node[fill=white] at(2*\x,1.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n1) -- (n4) node[fill=white] at(2.5*\x,1.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n2) -- (n5) node[fill=white] at(1*\x,2.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n2) -- (n6) node[fill=white] at(1.25*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n3) -- (n5) node[fill=white] at(1.75*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n3) -- (n7) node[fill=white] at(2.25*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n4) -- (n6) node[fill=white] at(2.75*\x,2.25*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n4) -- (n7) node[fill=white] at(3*\x,2.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n5) -- (n8) node[fill=white] at(1*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n5) -- (n9) node[fill=white] at(1.5*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n6) -- (n9) node[fill=white] at(2*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n7) -- (n9) node[fill=white] at(2.5*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n7) -- (n10) node[fill=white] at(3*\x,3.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n8) -- (n11) node[fill=white] at(.75*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n8) -- (n12) node[fill=white] at(1.25*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n9) -- (n12) node[fill=white] at(1.75*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n9) -- (n13) node[fill=white] at(2.25*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n10) -- (n13) node[fill=white] at(2.75*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n10) -- (n14) node[fill=white] at(3.25*\x,4.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n11) -- (n15) node[fill=white] at(.75*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n12) -- (n15) node[fill=white] at(1.25*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n12) -- (n16) node[fill=white] at(1.75*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n13) -- (n16) node[fill=white] at(2.25*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n13) -- (n17) node[fill=white] at(2.75*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n14) -- (n17) node[fill=white] at(3.25*\x,5.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n15) -- (n18) node[fill=white] at(.75*\x,6.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n15) -- (n19) node[fill=white] at(1.5*\x,6.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n16) -- (n19) node[fill=white] at(2*\x,6.5*\y) {\tiny\color{gray!80!black}$s_{2}$};
\draw(n17) -- (n19) node[fill=white] at(2.5*\x,6.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n18) -- (n20) node[fill=white] at(.75*\x,7.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\draw(n19) -- (n20) node[fill=white] at(1.5*\x,7.5*\y) {\tiny\color{gray!80!black}$s_{1}$};
\draw(n19) -- (n21) node[fill=white] at(2.5*\x,7.5*\y) {\tiny\color{gray!80!black}$s_{3}$};
\end{tikzpicture}}
\caption{Two Bruhat lattices of sortable elements associated with the Coxeter groups $A_{3}$ and $\tilde{C}_{2}$, respectively. Their edges are labeled by the SB-labeling defined in \eqref{eq:sb_bruhat}.}
\label{fig:bruhat_sortable}
\end{figure}
\end{example}
\section{Distributivity of the Bruhat Order on Sortable Elements}
\label{sec:distributivity}
Recall that a lattice $\mathcal{P}=(P,\leq)$ is \alert{distributive} if it satisfies one of the two following, equivalent, properties for all $p,q,r\in P$:
\begin{align}
p & \wedge(q\vee r) = (p\wedge q) \vee (p\wedge r)\label{eq:meet_distributivity}\\
p & \vee(q\wedge r) = (p\vee q) \wedge (p\vee r)\label{eq:join_distributivity}
\end{align}
Armstrong remarked in \cite{armstrong09sorting} that for a certain Coxeter element of the Coxeter group $A_{n}$ the lattice $\mathcal{B}_{\gamma}$ coincides with the lattice of order ideals of the root poset of $A_{n}$. (For any undefined terminology, we refer once more to \cite{bjorner05combinatorics}.) Hence this particular lattice is distributive. However, Armstrong remarked that this ``phenomenon, unfortunately, does not persist for all types''. In this section we partially answer the question for which finite Coxeter groups and which Coxeter elements the lattice $\mathcal{B}_{\gamma}$ is distributive.
\begin{lemma}\label{lem:subgroups_ideals}
Let $W$ be a Coxeter group, and $\gamma\in W$ a Coxeter element. If $W'$ is a standard parabolic subgroup of $W$ and $\gamma'\in W'$ denotes the restriction of $\gamma$ to $W'$, then $\mathcal{B}_{\gamma'}$ is an order ideal of $\mathcal{B}_{\gamma}$.
\end{lemma}
\begin{proof}
Since $W'$ is a subgroup of $W$, every element $w'\in W'$ lies in $W$ as well. Since $W'$ is a standard parabolic subgroup of $W$, we conclude that there exists some $J\subseteq S$ such that $W'$ is generated by $J$. Since $\gamma'$ is the restriction of $\gamma$ to $W'$, we conclude that $\gamma'$ is the subword of $\gamma$ that is obtained by deleting the letters not in $J$. Hence if $w'\in W'$ is $\gamma'$-sortable, then it is also $\gamma$-sortable. It follows immediately if $w\in W$ and $w'\in W'$ satisfy $w\leq_{B}w'$, then we have $w\in W'$. (Otherwise, the $\gamma$-sorting word of $w$ contains a letter not in $J$, which then implies $\alpha_{\gamma}(w)\not\subseteq\alpha_{\gamma}(w')$. This, however, contradicts $w\leq_{B}w'$.)
\end{proof}
\begin{remark}\label{rem:subgroups_intervals}
In particular, if $W$ is finite, then each standard parabolic subgroup of $W$ induces an interval of $\mathcal{B}_{\gamma}$.
\end{remark}
Now recall that each Coxeter element $\gamma\in W$ induces an orientation $\Gamma_{\gamma}(W)$ of the Coxeter diagram of $W$ as follows: an edge between $s_{i}$ and $s_{j}$ is oriented
\raisebox{-.15cm}{\begin{tikzpicture}\small
\draw(0,0) node(si){$s_{i}$};
\draw(1,0) node(sj){$s_{j}$};
\draw(.5,.15) node{\tiny $a$};
\draw[->,thick](si) -- (.55,0);
\draw[thick](.55,0) -- (sj);
\end{tikzpicture}}
if and only if $s_{i}$ precedes $s_{j}$ in every reduced word for $\gamma$, see \cite{shi97enumeration}*{Section~1.3}. The next result shows which orientations of a Coxeter diagram induce non-distributive intervals of $\mathcal{B}_{\gamma}$.
\begin{proposition}\label{prop:forbidden_subgraphs}
Let $W$ be a Coxeter group, let $\gamma\in W$ be a Coxeter element, and let $\Gamma_{\gamma}(W)$ be the Coxeter diagram of $W$ with the orientation induced by $\gamma$. If $\Gamma_{\gamma}(W)$ contains one of the following induced subgraphs, then $\mathcal{B}_{\gamma}$ is not distributive:
\begin{enumerate}[(i)]
\item \raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(.75,.2) node{\tiny a};
\draw(1.5,0) node(v2){$s_{i_{2}}$};
\draw(2.25,.2) node{\tiny b};
\draw(3,0) node(v3){$s_{i_{3}}$};
\draw[thick,->](v2) -- (.7,0);
\draw[thick](.7,0) -- (v1);
\draw[thick,->](v2) -- (2.3,0);
\draw[thick](2.3,0) -- (v3);
\end{tikzpicture}} for $i_{1},i_{2},i_{3}\in\{1,2,\ldots,n\}$, and $a,b\geq 3$,
\item \raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(1.5,0) node(v2){$s_{i_{2}}$};
\draw(2.25,.2) node{\tiny a};
\draw(3,0) node(v3){$s_{i_{3}}$};
\draw[thick,->](v2) -- (.7,0);
\draw[thick](.7,0) -- (v1);
\draw[thick,->](v3) -- (2.2,0);
\draw[thick](2.2,0) -- (v2);
\end{tikzpicture}} for $i_{1},i_{2},i_{3}\in\{1,2,\ldots,n\}$, and $a\geq 4$,
\item \raisebox{-.7cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(1.5,0) node(v2){$s_{i_{2}}$};
\draw(3,.5) node(v3){$s_{i_{4}}$};
\draw(3,-.5) node(v4){$s_{i_{3}}$};
\draw[thick,->](v1) -- (.8,0);
\draw[thick](.8,0) -- (v2);
\draw[thick,->](v3) -- (2.2,.23);
\draw[thick](2.2,.23) -- (v2);
\draw[thick,->](v2) -- (2.3,-.26);
\draw[thick](2.3,-.26) -- (v4);
\end{tikzpicture}} for $i_{1},i_{2},i_{3},i_{4}\in\{1,2,\ldots,n\}$,
\item \raisebox{-.7cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(1.5,0) node(v2){$s_{i_{2}}$};
\draw(3,.5) node(v3){$s_{i_{4}}$};
\draw(3,-.5) node(v4){$s_{i_{3}}$};
\draw[thick,->](v1) -- (.8,0);
\draw[thick](.8,0) -- (v2);
\draw[thick,->](v3) -- (2.2,.23);
\draw[thick](2.2,.23) -- (v2);
\draw[thick,->](v4) -- (2.2,-.23);
\draw[thick](2.2,-.23) -- (v2);
\end{tikzpicture}} for $i_{1},i_{2},i_{3},i_{4}\in\{1,2,\ldots,n\}$,
\item \raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(1,0) node(v2){$s_{i_{2}}$};
\draw(1.5,.2) node{\tiny a};
\draw(2,0) node(v3){$s_{i_{3}}$};
\draw(3,0) node(v4){$s_{i_{4}}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}} for $i_{1},i_{2},i_{3},i_{4}\in\{1,2,\ldots,n\}$, and $a\geq 4$,
\item \raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(1,0) node(v2){$s_{i_{2}}$};
\draw(2,0) node(v3){$s_{i_{3}}$};
\draw(2.5,.2) node{\tiny a};
\draw(3,0) node(v4){$s_{i_{4}}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\draw[thick,->](v3) -- (2.55,0);
\draw[thick](2.55,0) -- (v4);
\end{tikzpicture}} for $i_{1},i_{2},i_{3},i_{4}\in\{1,2,\ldots,n\}$, and $a\geq 5$, \quad or
\item \raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{i_{1}}$};
\draw(1,0) node(v2){$s_{i_{2}}$};
\draw(2,0) node(v3){$s_{i_{3}}$};
\draw(2.5,.2) node{\tiny a};
\draw(3,0) node(v4){$s_{i_{4}}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}} for $i_{1},i_{2},i_{3},i_{4}\in\{1,2,\ldots,n\}$, and $a\geq 5$.
\end{enumerate}
\end{proposition}
\begin{proof}
Suppose that $\Gamma_{\gamma}(W)$ contains an induced subgraph of form (i). Then, in particular, we have the $\gamma$-sortable elements $x=s_{i_{2}}s_{i_{1}}\vert s_{i_{2}},y=s_{i_{2}}s_{i_{1}}s_{i_{3}}$, and $z=s_{i_{2}}s_{i_{3}}\vert s_{i_{2}}$. We have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{2}}s_{i_{1}}s_{i_{3}}\vert s_{i_{2}}\\
& = s_{i_{2}}s_{i_{1}}\vert s_{i_{2}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{2}}s_{i_{1}} \vee_{B} s_{i_{2}}\\
& = s_{i_{2}}s_{i_{1}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
If $\Gamma_{\gamma}(W)$ contains an induced subgraph of form (ii), then consider the elements $x=s_{i_{2}}s_{i_{1}}\vert s_{i_{2}},y=s_{i_{3}}s_{i_{2}}s_{i_{1}}$, and $z=s_{i_{3}}s_{i_{2}}\vert s_{i_{3}}s_{i_{2}}$. We have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{3}}s_{i_{2}}s_{i_{1}}\vert s_{i_{3}}s_{i_{2}}\\
& = s_{i_{2}}s_{i_{1}}\vert s_{i_{2}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{2}}s_{i_{1}}\vee_{B}s_{i_{2}}\\
& = s_{i_{2}}s_{i_{1}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
If $\Gamma_{\gamma}(W)$ contains an induced subgraph of the form (iii), then consider the elements $x=s_{i_{4}}s_{i_{2}}s_{i_{3}}\vert s_{i_{2}},y=s_{i_{1}}s_{i_{4}}s_{i_{2}}s_{i_{3}}$, and $z=s_{i_{1}}s_{i_{4}}s_{i_{2}}\vert s_{i_{1}}s_{i_{4}}s_{i_{2}}$. We have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{1}}s_{i_{4}}s_{i_{2}}s_{i_{3}}\vert s_{i_{1}}s_{i_{4}}s_{i_{2}}\\
& = s_{i_{4}}s_{i_{2}}s_{i_{3}}\vert s_{i_{2}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{4}}s_{i_{2}}s_{i_{3}}\vee_{B}s_{i_{4}}s_{i_{2}}\\
& = s_{i_{4}}s_{i_{2}}s_{i_{3}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
If $\Gamma_{\gamma}(W)$ contains an induced subgraph of the form (iv), then consider the elements $x=s_{i_{1}}s_{i_{3}}s_{i_{2}}\vert s_{i_{1}}s_{i_{3}}s_{i_{2}}, y=s_{i_{1}}s_{i_{4}}s_{i_{2}}\vert s_{i_{1}}s_{i_{4}}s_{i_{2}}$, and $z=s_{i_{3}}s_{i_{4}}s_{i_{2}}\vert s_{i_{3}}s_{i_{4}}s_{i_{2}}$. We have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{1}}s_{i_{3}}s_{i_{4}}s_{i_{2}}\vert s_{i_{1}}s_{i_{3}}s_{i_{4}}s_{i_{2}}\\
& = s_{i_{1}}s_{i_{3}}s_{i_{2}}\vert s_{i_{1}}s_{i_{3}}s_{i_{2}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{1}}s_{i_{2}}\vert s_{i_{1}}\vee_{B}s_{i_{3}}s_{i_{2}}\vert s_{i_{3}}\\
& = s_{i_{1}}s_{i_{3}}s_{i_{2}}\vert s_{i_{1}}s_{i_{3}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
If $\Gamma_{\gamma}(W)$ contains an induced subgraph of the form (v), then consider the elements $x=s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}, y=s_{i_{4}}s_{i_{3}}\vert s_{i_{4}}$, and $z=s_{i_{1}}s_{i_{2}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}$. We have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}\\
& = s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{4}}s_{i_{3}}\vert s_{i_{4}}\vee_{B}s_{i_{2}}s_{i_{3}}\vert s_{i_{2}}s_{i_{3}}\\
& = s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
If $\Gamma_{\gamma}(W)$ contains an induced subgraph of the form (vi), then consider the elements $x=s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}, y=s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}\vert s_{i_{2}}$, and $z=s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{3}}$. We have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{1}}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}\\
& = s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}\vert s_{i_{2}}\\
& \kern1cm \vee_{B}s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\\
& = s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}}s_{i_{3}}s_{i_{4}}\vert s_{i_{2}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
If $\Gamma_{\gamma}(W)$ contains an induced subgraph of the form (vii), then consider the elements $x=s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}},y=s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}$, and $z=s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{4}}s_{i_{3}}$. We
have
\begin{align*}
x\wedge_{B}(y\vee_{B}z) & = x\wedge_{B}s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{1}}s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\\
& = s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}},\qquad\text{and}\\
(x\wedge_{B}y)\vee_{B}(x\wedge_{B}z) & = s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}\\
& \kern1cm \vee_{B} s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{4}}\\
& = s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}}s_{i_{3}}\vert s_{i_{2}}s_{i_{4}},
\end{align*}
which contradicts \eqref{eq:meet_distributivity}.
\end{proof}
We obtain the following corollary immediately.
\begin{corollary}\label{cor:type_d_e}
If $W=D_{n}$, for $n\geq 4$, $W=E_{n}$, for $n\in\{6,7,8\}$, $W=F_{4}$, or $W=H_{4}$, and $\gamma\in W$ is a Coxeter element, then $\mathcal{B}_{\gamma}$ is not distributive.
\end{corollary}
\begin{proof}
First consider $W=D_{4}$. The eight orientations of $\Gamma(D_{4})$ are shown below.
\begin{center}\begin{tabular}{cccc}
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,.5) node[gray](v3){$s_{3}$};
\draw(2,-.5) node[gray](v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->,gray](v2) -- (1.55,.27);
\draw[thick,gray](1.55,.27) -- (v3);
\draw[thick,->,gray](v2) -- (1.55,-.27);
\draw[thick,gray](1.55,-.27) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,.5) node[gray](v3){$s_{3}$};
\draw(2,-.5) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v2) -- (1.55,.27);
\draw[thick,gray](1.55,.27) -- (v3);
\draw[thick,->](v2) -- (1.55,-.27);
\draw[thick](1.55,-.27) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,.5) node(v3){$s_{3}$};
\draw(2,-.5) node[gray](v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->](v3) -- (1.45,.23);
\draw[thick](1.45,.23) -- (v2);
\draw[thick,->,gray](v2) -- (1.55,-.27);
\draw[thick,gray](1.55,-.27) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,.5) node[gray](v3){$s_{3}$};
\draw(2,-.5) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v2) -- (1.55,.27);
\draw[thick,gray](1.55,.27) -- (v3);
\draw[thick,->](v4) -- (1.45,-.23);
\draw[thick](1.45,-.23) -- (v2);
\end{tikzpicture}
\\
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,.5) node(v3){$s_{3}$};
\draw(2,-.5) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v3) -- (1.45,.23);
\draw[thick](1.45,.23) -- (v2);
\draw[thick,->](v2) -- (1.55,-.27);
\draw[thick](1.55,-.27) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,.5) node(v3){$s_{3}$};
\draw(2,-.5) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,.27);
\draw[thick](1.55,.27) -- (v3);
\draw[thick,->](v4) -- (1.45,-.23);
\draw[thick](1.45,-.23) -- (v2);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,.5) node(v3){$s_{3}$};
\draw(2,-.5) node(v4){$s_{4}$};
\draw[thick,->](v2) -- (.45,0);
\draw[thick](.45,0) -- (v1);
\draw[thick,->](v3) -- (1.45,.23);
\draw[thick](1.45,.23) -- (v2);
\draw[thick,->](v4) -- (1.45,-.23);
\draw[thick](1.45,-.23) -- (v2);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,.5) node(v3){$s_{3}$};
\draw(2,-.5) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v3) -- (1.45,.23);
\draw[thick](1.45,.23) -- (v2);
\draw[thick,->](v4) -- (1.45,-.23);
\draw[thick](1.45,-.23) -- (v2);
\end{tikzpicture}\\
\end{tabular}\end{center}
The first four orientations in the first row correspond to case (i) in Proposition~\ref{prop:forbidden_subgraphs}, the first three orientations in the second row correspond to case (iii) in Proposition~\ref{prop:forbidden_subgraphs}, and the fourth orientation in the second row corresponds to case (iv) in Proposition~\ref{prop:forbidden_subgraphs}. Hence $\mathcal{B}_{\gamma}$ cannot be distributive for Coxeter elements inducing these orientations.
If $W=D_{n}$, for $n>4$, or $W=E_{n}$, for $n\in\{6,7,8\}$, then we conclude from Figure~\ref{fig:coxeter_diagrams} that $W$ has a standard parabolic subgroup isomorphic to $D_{4}$. In view of Lemma~\ref{lem:subgroups_ideals} and Remark~\ref{rem:subgroups_intervals}, we conclude that $\mathcal{B}_{\gamma}$ contains a non-distributive interval, and hence cannot be distributive itself.
Now let $W=F_{4}$. The eight orientations of $\Gamma(F_{4})$ are shown below.
\begin{center}\begin{tabular}{ccc}
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->,gray](v3) -- (2.55,0);
\draw[thick,gray](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v2) -- (1.55,0);
\draw[thick,gray](1.55,0) -- (v3);
\draw[thick,->](v3) -- (2.55,0);
\draw[thick](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v2) -- (1.55,0);
\draw[thick,gray](1.55,0) -- (v3);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}
\\
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v2) -- (.45,0);
\draw[thick](.45,0) -- (v1);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->,gray](v3) -- (2.55,0);
\draw[thick,gray](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->,gray](v2) -- (1.55,0);
\draw[thick,gray](1.55,0) -- (v3);
\draw[thick,->,gray](v3) -- (2.55,0);
\draw[thick,gray](2.55,0) -- (v4);
\end{tikzpicture}
\\
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node(v3){$s_{3}$};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 4};
\draw(2,0) node(v3){$s_{3}$};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v3) -- (1.45,0);
\draw[thick](1.45,0) -- (v2);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}
& \\
\end{tabular}\end{center}
The first four orientations correspond to case (i) in Proposition~\ref{prop:forbidden_subgraphs}, the last two orientations in the second row correspond to case (ii) in Proposition~\ref{prop:forbidden_subgraphs}, and the two orientations in the third row correspond to case (v) in Proposition~\ref{prop:forbidden_subgraphs}. Hence $\mathcal{B}_{\gamma}$ cannot be distributive for Coxeter elements inducing these orientations.
Now let $W=H_{4}$. The eight orientations of $\Gamma(H_{4})$ are shown below.
\begin{center}\begin{tabular}{ccc}
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->,gray](v3) -- (2.55,0);
\draw[thick,gray](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v2) -- (1.55,0);
\draw[thick,gray](1.55,0) -- (v3);
\draw[thick,->](v3) -- (2.55,0);
\draw[thick](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node[gray](v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->,gray](v2) -- (.45,0);
\draw[thick,gray](.45,0) -- (v1);
\draw[thick,->,gray](v2) -- (1.55,0);
\draw[thick,gray](1.55,0) -- (v3);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}
\\
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v2) -- (.45,0);
\draw[thick](.45,0) -- (v1);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->,gray](v3) -- (2.55,0);
\draw[thick,gray](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v2) -- (.45,0);
\draw[thick](.45,0) -- (v1);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->,gray](v4) -- (2.45,0);
\draw[thick,gray](2.45,0) -- (v3);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node[gray](v2){$s_{2}$};
\draw(2,0) node[gray](v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node[gray](v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->,gray](v3) -- (1.45,0);
\draw[thick,gray](1.45,0) -- (v2);
\draw[thick,->,gray](v4) -- (2.45,0);
\draw[thick,gray](2.45,0) -- (v3);
\end{tikzpicture}
\\
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,0) node(v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\draw[thick,->](v3) -- (2.55,0);
\draw[thick](2.55,0) -- (v4);
\end{tikzpicture}
&
\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,0) node(v3){$s_{3}$};
\draw(2.5,.2) node{\tiny 5};
\draw(3,0) node(v4){$s_{4}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\draw[thick,->](v4) -- (2.45,0);
\draw[thick](2.45,0) -- (v3);
\end{tikzpicture}
& \\
\end{tabular}\end{center}
The first four orientations correspond to case (i) in Proposition~\ref{prop:forbidden_subgraphs}, and the last two orientations in the second row correspond to case (ii) in Proposition~\ref{prop:forbidden_subgraphs}. The first orientation in the third row corresponds to case (vi) in Proposition~\ref{prop:forbidden_subgraphs}, and the second orientation in the third row corresponds to case (vii) in Proposition~\ref{prop:forbidden_subgraphs}.
\end{proof}
\begin{proposition}\label{prop:distributive_coincidental}
If $W=A_{n}$ for $n\geq 1$, $W=B_{n}$ for $n\geq 2$, $W=H_{3}$ or $W=I_{2}(k)$ for $k\geq 5$, then there exists a Coxeter element $\gamma\in W$ such that $\mathcal{B}_{\gamma}$ is distributive.
\end{proposition}
\begin{proof}
Let $W=A_{n}$ and let $\gamma$ be the Coxeter element that induces the linear orientation
\raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,0) node(v3){$\cdots$};
\draw(3,0) node(v4){$s_{n}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.5,0);
\draw[thick](1.5,0) -- (v3);
\draw[thick,->](v3) -- (2.6,0);
\draw[thick](2.6,0) -- (v4);
\end{tikzpicture}}.
It follows from the bijection in \cite{bandlow01area} that $\mathcal{B}_{\gamma}$ is isomorphic to the lattice of classical Dyck paths under dominance order. This lattice is known to be distributive, see for instance \cite{ferrari05lattices}*{Corollary~2.2}. The same bijection implies also that $\mathcal{B}_{\gamma}$ is isomorphic to the lattice of order ideals of the root poset of $A_{n}$.
Let $W=B_{n}$ and let $\gamma$ be the Coxeter element that induces the linear orientation
\raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,0) node(v3){$\cdots$};
\draw(3.2,0) node(v4){$s_{n-1}$};
\draw(3.8,.2) node{\tiny 4};
\draw(4.2,0) node(v5){$s_{n}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.5,0);
\draw[thick](1.5,0) -- (v3);
\draw[thick,->](v3) -- (2.6,0);
\draw[thick](2.6,0) -- (v4);
\draw[thick,->](v4) -- (3.85,0);
\draw[thick](3.85,0) -- (v5);
\end{tikzpicture}}.
It follows from the bijection in \cite{stump13more}*{Section~3} that $\mathcal{B}_{\gamma}$ is isomorphic to the lattice of type-$B$ Dyck paths under dominance order. This lattice is known to be distributive, see \cite{muehle13heyting}*{Theorem~2.9}. The same bijection implies also that $\mathcal{B}_{\gamma}$ is isomorphic to the lattice of order ideals of the root poset of $B_{n}$.
Let $W=I_{2}(k)$ for $k\geq 5$, and denote by $s_{1}$ and $s_{2}$ the Coxeter generators of $W$. We have $\mathcal{B}_{s_{1}s_{2}}\cong\mathcal{B}_{s_{2}s_{1}}$, and this lattice is trivially distributive. It is also isomorphic to the lattice of order ideals of the ``root poset'' of $I_{2}(k)$ defined by Armstrong in \cite{armstrong09generalized}*{Figure~5.15}.
Let $W=H_{3}$ and let $\gamma$ be the Coxeter element that induces the linear orientation
\raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(1.5,.2) node{\tiny 5};
\draw(2,0) node(v4){$s_{3}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.55,0);
\draw[thick](1.55,0) -- (v3);
\end{tikzpicture}}.
We can easily check by computer that the resulting lattice $\mathcal{B}_{\gamma}$ is distributive. However, in this case, $\mathcal{B}_{\gamma}$ is \emph{not} isomorphic to the lattice of order ideals of the ``root poset'' of $H_{3}$ defined by Armstrong in \cite{armstrong09generalized}*{Figure~5.15}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main_distributive}]
This follows immediately from Proposition~\ref{prop:distributive_coincidental}.
\end{proof}
We conclude this section with the following conjecture.
\begin{conjecture}\label{conj:list_exhaustive}
For finite Coxeter groups, the list in Proposition~\ref{prop:forbidden_subgraphs} is exhaustive, \text{i.e.}\; if $W$ is a finite Coxeter group, $\gamma\in W$ is a Coxeter element and the orientation $\Gamma_{\gamma}(W)$ of the Coxeter diagram of $W$ induced by $\gamma$ does not contain one of the induced subgraphs listed in Proposition~\ref{prop:forbidden_subgraphs}, then $\mathcal{B}_{\gamma}$ is distributive.
\end{conjecture}
\begin{remark}
The claim of Conjecture~\ref{conj:list_exhaustive} for $W=H_{3}$ can be verified by computer. For $W=B_{3}$, the only orientation other than the one in Proposition~\ref{prop:distributive_coincidental} that is conjectured to yield a distributive lattice $\mathcal{B}_{\gamma}$ is
\raisebox{-.16cm}{\begin{tikzpicture}\small
\draw(0,0) node(v1){$s_{1}$};
\draw(1,0) node(v2){$s_{2}$};
\draw(2,0) node(v3){$\cdots$};
\draw(3.2,0) node(v4){$s_{n-1}$};
\draw(3.8,.2) node{\tiny 4};
\draw(4.2,0) node(v5){$s_{n}$};
\draw[thick,->](v1) -- (.55,0);
\draw[thick](.55,0) -- (v2);
\draw[thick,->](v2) -- (1.5,0);
\draw[thick](1.5,0) -- (v3);
\draw[thick,->](v3) -- (2.6,0);
\draw[thick](2.6,0) -- (v4);
\draw[thick,->](v5) -- (3.75,0);
\draw[thick](3.75,0) -- (v4);
\end{tikzpicture}}. For $W=A_{n}$ there are several more options.
\end{remark}
|
\section{Introduction}
\label{sec:intr}
Algorithmic quantum simulation~\cite{Fey82,Llo96,BACS07a,BN09},
which is digital quantum simulation with pre-specified bounded-error output~\cite{Sand13},
is important for simulating many-body dynamics~\cite{WBHS11,RWS12},
quantum-state generation and
dissipative quantum-state engineering~\cite{VWC09,MDPZ12},
quantum thermodynamics~\cite{TD00,PW09},
nonequilibrium quantum phase transitions~\cite{SAA+10,SMN+13},
testing element distinctness~\cite{Chi09},
and solving linear equations~\cite{HHL09}
and differential equations~\cite{Ber14}.
Experimental quantum simulation~\cite{BN09}
has been demonstrated in quantum computing implementations such as
ion traps~\cite{LHN+11,BSK+12,BMS+11},
atoms in optical lattice~\cite{LSA+07,WMBL11},
and superconducting circuits~\cite{GSP14}.
Whereas unitary evolution generated by a self-adjoint Hamiltonian
has so far been the major research focus,
algorithmic quantum simulation of quantum channels
(i.e., completely-positive trace-preserving mappings)~\cite{Sti55,Cho75,Kra83}
and open-system dynamics~\cite{NC00} is a nascent and exciting research area
both theoretically~\cite{KBG+11,WBOS13} and experimentally~\cite{SMN+13,FPK+12}.
Quantum channel simulators can play vital roles in quantum simulation for the study of
quantum non-Markovian effects~\cite{CGM+12,SPA+12},
dissipative quantum many-body dynamics~\cite{SMN+13},
and also for modelling quantum noise for the test of various protocols~\cite{TB05,MPG+13}.
Our aim is to develop a classical algorithm
that designs quantum circuits,
which simulate accurately an arbitrary qudit
(corresponding to states described by positive semidefinite trace-class operators~$\rho$
acting on the~$d$-dimensional Hilbert space $\mathscr{H}_d\cong\mathbb{C}^d$)
channel~$\mathcal{E}$
on fault-tolerant quantum computers
within a specified error tolerance~$\epsilon$
with respect to the diamond norm~\cite{AKN97}.
This quantum circuit needs to be executed efficiently with respect to~$\epsilon$,
i.e., with quantum space and time resources scaling as polylog$\frac{1}{\epsilon}$.
We treat classical space resources, i.e.~dits ($d$-dimensional digits, $d=2$ for bits), as free.
Previously two different cases have been considered:
Markovian qudit channel simulation
but with the Liouvillian rather than the channel as input~\cite{KBG+11}
and for a general single-qubit channel
as input with an efficient simulation circuit as output~\cite{WBOS13}
employing qubit extreme channel theory~\cite{Cho75,RSW02}.
Efficient and accurate algorithmic quantum simulation,
such that the output is delivered with minimal resources and within the pre-specified error tolerance,
is vital for constructing quantum simulators in the near term that answer computational
problems.
Generalizing from simulating qubit channels to qudit channels is not straightforward
because decomposing an arbitrary qudit channel into a convex combination of generalized extreme
channels is an open problem in quantum information~\cite{Rus07}.
We circumvent this obstacle by decomposing approximately, rather than exactly,
into a convex sum of generalized extreme channels,
and we construct a classical optimization algorithm~\cite{BV04}
that devises circuits for simulating generalized extreme channels so that the entire
qudit channel can be simulated by random concatenations of generalized extreme channel simulators.
The circuits devised by our algorithm show how to realize
algorithmic quantum-channel simulation.
We approach the problem of constructing a qudit channel circuit simulator
by constructing an algorithm whose inputs are the description of~$\mathcal E$,
the tolerance~$\epsilon$,
and the dimension~$d$ of the qudit Hilbert space
and delivering an output comprising a description of a simulation circuit
and the actual error~$\tilde{\epsilon}$.
In contrast to the case for unitary channels,
which can be constructed as a concatenation of other unitary channels,
the non-unitary channel is not such a simple sequence of channels~\cite{WC08}
thereby resulting in complicated approach to quantum channel simulation.
A direct procedure for a quantum simulation of a quantum channel is to
employ Stinespring dilation~\cite{Sti55},
which replaces the qudit channel $\mathcal{E}$
by a unitary channel $\mathcal{U}$,
with up to~$d^3$ dimensions,
followed by a partial trace over the environment
to recover the description of the channel $\mathcal{E}$.
The circuit for simulating the channel by
a dilated unitary channel acting on a Hilbert space of dimension~$d^3$
generically requires $O(d^6)$
single-qubit and two-qubit gates~\cite{BBC+95,MV06},
obtained from a small universal instruction set using a Solovay-Kitaev gate decomposition
approach~\cite{DN06,KMM13b,Kli13}.
Therefore, the time cost is $O(d^6)$ and the space cost is three qudits.
In the interest of bringing algorithmic quantum-channel simulation to its lowest possible
cost for experimental expediency,
we employ the procedure of approximately decomposing the channel
into a convex combination of generalized extreme channels.
The simulation circuit has a time cost of $O(d^2\log \frac{d^2}{\epsilon} )$,
which is the same as the time cost for simulating a unitary qudit channel,
hence a lower bound~\cite{BBC+95,BOB05}.
Furthermore this procedure requires a spatial resource of just two qudits
plus random dits,
hence reduces the quantum space cost by a third.
This work comprises three parts.
In Sec.~\ref{sec:extdecomp} we present our method for
the construction of extreme channels.
In Sec.~\ref{sec:circ} we construct quantum circuits for generalized extreme channels.
In Sec.~\ref{sec:algo} we discuss the quantum channel simulation algorithm.
We conclude briefly and provide supporting information in Appendix.
\section{Extreme quantum channels}
\label{sec:extdecomp}
A quantum channel $\mathcal{E}\in\mathscr{S}_d$,
the set of all channels for qudits of dimension~$d$,
can be represented as
\begin{equation}
\label{eq:channel}
\mathcal{E}(\rho)
=\sum_{i=0}^m K_i\rho K_i^\dagger
\end{equation}
for all states~$\rho$
with a set of linearly independent Kraus operators~\cite{Kra83}
\begin{equation}
\{K_i:\mathscr{H}_d\to\mathscr{H}_d\}
\end{equation}
such that
\begin{equation}
\sum_{i=0}^m K_i^\dagger K_i=\mathds{1},
\end{equation}
and Kraus rank~\cite{Kra83}
\begin{equation}
m+1\leq d^2.
\end{equation}
A channel is extreme if and only if it cannot be written as a convex sum of other channels.
Equivalently a channel is extreme if and only if $m$ is bounded above by~$d-1$
and $\{K_i^\dagger K_j\}$ is a linearly independent set~\cite{Cho75}.
A channel is called a generalized extreme channel if its Kraus rank is at most~$d$,
and a generalized extreme channel which is not extreme is called a quasi-extreme channel~\cite{Rus07}.
Clearly the set of generalized extreme channels contains both extreme channels and quasi-extreme channels.
Next we propose a Kraus operator-sum representation
for an arbitrary rank-$d$ extreme or quasi-extreme channel.
First, we construct the sum representation
using the Heisenberg-Weyl basis
\begin{equation}
\left\{X_i Z_j;i,j\in\mathbb{Z}_d\right\}
\end{equation}
for
\begin{equation}
X_i =\sum_{\ell=0}^{d-1}|\ell\rangle\langle \ell+i|,\;
Z_j =\sum_{\ell=0}^{d-1}\operatorname{e}^{\operatorname{i}2\pi \ell j/d}|\ell\rangle\langle \ell|.
\end{equation}
\begin{proposition}
\label{prop:ext}
A rank-$d$ extreme channel $\mathcal{E}^\text{e}\in\mathscr{S}_d$
can be represented by
\begin{equation}
\label{eq:extKraus}
\mathcal{E}^\text{e}(\rho)=\sum_{i=0}^{d-1}K_i\rho K_i^\dagger
\end{equation}
for any Kraus operators satisfying
\begin{equation}
\label{eq:F}
K_i:=W F_i V,\;
F_i:=X_iE_i,\;
E_i:=\sum_{j=0}^{d-1}a_{ij}Z_j,\,
i\in\mathbb{Z}_d,
\end{equation}
for any unitary operators
\begin{equation}
V, W \in SU(d),
\end{equation}
provided that
\begin{equation}
\{a_{ij}\in \mathbb{C}\}
\end{equation}
is chosen such that the set $\{F_i^\dagger F_j\}$
is linearly independent and
\begin{equation}
\sum_{i=0}^{d-1} F_i^\dagger F_i=\mathds{1}
\end{equation}
is satisfied.
\end{proposition}
\begin{proof}
Per definition, Eq.~(\ref{eq:extKraus}) holds for any rank-$d$ extreme channel
with~$\{K_i^\dagger K_j\}$ being linearly independent.
Thus, the proof focuses on showing that the ansatz~(\ref{eq:F})
yields arbitrary linearly independent operators $\{K_i^\dagger K_j\}$.
Linear independence of $\{K_i^\dagger K_j\}$
requires that
\begin{equation}
\label{eq:Xi}
\Xi:=\sum_{i,j=0}^{d-1}\gamma_{ij}K_i^\dagger K_j=0
\iff\gamma_{ij}\equiv 0 \; \forall i,j.
\end{equation}
From Eq.~(\ref{eq:F}),
\begin{equation}
\Xi=V^\dagger \left(\sum_{ij}\gamma_{ij}F_i^\dagger F_j\right)V.
\end{equation}
This is a unitary conjugation of the sum in parentheses
so we ignore $V$ in the proof.
Therefore, we need to require linear independence of $\{F_i^\dagger F_j\}$. For
\begin{equation}
\label{eq:b}
b_{i\mu\nu}:=\sum_{k,l=0}^{d-1} a^*_{ik}a_{i+\mu,l} \operatorname{e}^{\operatorname{i}2\pi[\mu l+\nu(l-k)]/d},\;
\mu\in\mathbb{Z}_d,
\end{equation}
we observe that
\begin{equation}
F_i^\dagger F_{i+\mu}=\sum_{\nu=0}^{d-1} b_{i\mu\nu} |\nu\rangle\langle \nu+\mu|;
\end{equation}
hence
\begin{equation}
\text{tr}[(F_i^\dagger F_{i+\mu})^\dagger F_j^\dagger F_{j+\mu'} ]=0
\end{equation}
for $\mu\neq \mu'$.
Now we partition
\begin{equation}
\label{eq:partition}
\left\{F_i^\dagger F_{i+\mu};i,\mu\in\mathbb{Z}_d\right\}
\to\left\{\{F_i^\dagger F_{i+\mu};i\in\mathbb{Z}_d\};\mu\in\mathbb{Z}_d\right\}.
\end{equation}
For $\{F_i^\dagger F_{i+\mu}\}$ to be a linearly independent set,
each subset must be linearly independent.
For each subset,
\begin{equation}
\Xi_\mu:=\sum_{i=0}^{d-1} \gamma_{i,i+\mu} F_i^\dagger F_{i+\mu}
\end{equation}
so
\begin{equation}
\Xi =\sum_{\mu=0}^{d-1} \Xi_\mu.
\end{equation}
Then
$\Xi\equiv 0$
implies $\Xi_\mu \equiv 0 \; \forall \mu$.
Now we establish linear independence of $\{F_i^\dagger F_j\}$
by constraining each subset~(\ref{eq:partition}).
First we map each matrix $F_i^\dagger F_{i+\mu}$
to a vector
\begin{equation}
\bm{b}_{i\mu}:=(b_{i\mu\nu}).
\end{equation}
Then linear independence of
\begin{equation}
\{F_i^\dagger F_{i+\mu}\}
\end{equation}
is ensured by the condition that the determinant of each matrix
\begin{equation}
B_\mu:=(\bm{b}_{i\mu})
\end{equation}
is nonzero;
i.e.,
\begin{equation}
\text{det} B_\mu\neq 0 \; \forall \mu
\end{equation}
(except for a zero-measure subset of $\{a_{ij}\}$).
Then $\Xi_\mu\equiv 0$ implies
\begin{equation}
\gamma_{i,i+\mu}\equiv 0 \; \forall i,\mu,
\end{equation}
which establishes linear independence of $\{F_i^\dagger F_j\}$
hence also $\{K_i^\dagger K_j\}$.
As $\{F_i^\dagger F_j\}$ spans $\mathcal{B}(\mathscr{H}_d)$,
the space of bounded linear operators on $\mathscr{H}_d$,
hence is a basis.
Composition with~$V$ and also $W$ ensures that an arbitrary basis
$\{K_i^\dagger K_j\}$ can be realized for an extreme channel.
Consequently,
the proof showing extremality of the channel~(\ref{eq:extKraus}) is complete.
\end{proof}
\begin{corollary}
The set of Kraus operators~$F_i\left(\{a_{ij}\in\mathbb{C}\}\right)$
has at most $d^2-d$ independent real parameters.
\label{cor:parameters}
\end{corollary}
\begin{proof}
We prove the statement using the property of the Choi-Jamio{\l}kowski state~\cite{Cho75,Jam72}
\begin{equation}
\label{eq:C}
\mathcal{C} := \mathcal{E} \otimes\mathds{1} (|\eta\rangle\langle\eta|),\;
|\eta\rangle=\sum_{i=0}^{d-1}|i\rangle|i\rangle \in \mathscr{H}_d \otimes \mathscr{H}_d,
\end{equation}
for a channel $\mathcal{E}$ and~$\{|i\rangle\}$ the computational basis of $\mathscr{H}_d$.
With Prop.~\ref{prop:ext}
and defining $\tilde{a}_{i,l+i}:=\sum_{j=0}^{d-1}a_{ij}\operatorname{e}^{\operatorname{i}2\pi(l+i) j/d}$,
we find that
the Choi-Jamio{\l}kowski state $\mathcal{C}^\text{e}$ corresponding to $\{F_i\}$ is
\begin{equation}\label{eq:extchoi}
\mathcal{C}^\text{e}=\sum_{i=0}^{d-1} \sum_{k,l=0}^{d-1}\tilde{a}^*_{i,k+i} \tilde{a}_{i,l+i}|l,l+i\rangle\langle k,k+i|,
\end{equation}
which is a~$d$-sparse, rank-$d$ positive semidefinite matrix
with at most $d^2$ real parameters.
Constrained by normalization,
$\{F_i\}$ has at most $d^2-d$ independent parameters.
\end{proof}
When the set $\{K_i^\dagger K_j\}$ is not linearly independent
our construction~(\ref{eq:F}) yields non-extreme yet quasi-extreme channels,
which are in the closure of the set of extreme channels~\cite{Rus07}.
As mentioned in the proof of Prop.~\ref{prop:ext},
the set of extreme channels dominates the set of all generalized extreme channels.
Also, both extreme and quasi-extreme channels with rank smaller than~$d$ can be realized if
some of the Kraus operators are zero matrices.
\begin{corollary}
\label{coro:quasiext}
A rank-$d$ generalized extreme channel $\mathcal{E}^\emph{g}\in\mathscr{S}_d$
can be represented by
\begin{equation}
\mathcal{E}^\emph{g}(\rho)=\sum_{i=0}^{d-1}K_i\rho K_i^\dagger
\end{equation}
with Kraus operators~(\ref{eq:F})
for any unitary operators $V, W \in SU(d)$
and
\begin{equation}
\sum_{i=0}^{d-1} F_i^\dagger F_i=\mathds{1}.
\end{equation}
\end{corollary}
\begin{proof}
From the construction~(\ref{eq:F}), it holds $\text{tr} F_i^\dagger F_j =0$ for $i\neq j$,
which means the set $\{F_i\}$ (also $\{K_i\}$) is linearly independent.
The unitary operators $V$ and $W$ can take this set to an arbitrary linearly independent set
with the same cardinality.
This proves that any rank-$d$ channel can be written in the proposed form.
\end{proof}
\section{Quantum circuits for generalized extreme channels}
\label{sec:circ}
Now that we have the sum representation of the (quasi)extreme channel~$\mathcal{E}^\text{e}$
with respect to $\{F_i\}$~(\ref{eq:F}),
we construct a quantum circuit for simulating state evolution through the channel
by employing Stinespring dilation~\cite{Sti55}.
The Kraus operators $\{F_i\}$
can be realized by a channel
\begin{equation}
\label{eq:Uchannel}
\mathcal{U}(\bullet):= U\bullet U^\dagger,\;
U\in SU(d^2)
\end{equation}
with~$U$ acting on the system~(s) qudit and
an ancillary~$d$-dimensional ancilla~(a) qudit
such that $F_i=_\text{a}\!\!\!\langle i| U|0\rangle_\text{a}$.
Such Kraus operators~$\{F_i\}$ trivially satisfy
linear independence and
\begin{equation}
\sum_iF_i^\dagger F_i=\mathds{1}.
\end{equation}
In principle the quantum circuit for a generalized extreme channel could be constructed in three stages:
solve the Kraus decomposition in Prop.~\ref{prop:ext},
then use the Kraus operators~$\{F_i\}$ to construct the unitary operator
based on Stinespring dilation,
and finally decompose it into a quantum circuit comprising gates from a finite
universal instruction set.
However, this method is stymied by the intractability of the nonlinear algebra\"{i}c equations that arise
from the Kraus decomposition so this approach is not viable.
Instead we adopt a different tack,
which is to find the quantum circuit by optimization.
In this approach, we construct the circuit
for a generalized extreme channel as a sequence of instruction-set gates
and optimize over the set of circuits
such that
the diamond-norm distance~\cite{AKN97}
(rather than the induced Schatten one-norm~\cite{KBG+11,WBOS13})
\begin{equation}
\|\mathcal{E}-\mathcal{\tilde{E}}\|_{1\rightarrow 1}
:=\max_\rho\|\mathcal{E}(\rho)-\mathcal{\tilde{E}}(\rho)\|_1
\end{equation}
between the input channel~$\mathcal E$ and
the approximate channel~$\tilde{\mathcal E}$ satisfies
\begin{align}
\label{eq:diamond}
\|\mathcal{E}-\mathcal{\tilde{E}}\|_\diamond
:=\|\mathcal{E}\otimes\mathds{1}-\mathcal{\tilde{E}}
\otimes\mathds{1}\|_{1\rightarrow 1}\leq \epsilon.
\end{align}
The diamond-norm distance is preferred
as it gives worst-case gate error,
and has the operational meaning that the probability of distinguishing between
the two channels from their outputs is
\begin{equation}
\frac{1+\epsilon/2}{2}.
\end{equation}
Next we present the single- and two-qudit gate set for this circuit construction.
Three types of single-qudit gates are specified by
\begin{equation}
X_{jk} :=|j\rangle\langle k|+|k\rangle\langle j|,
\end{equation}
by the Givens rotation, which is a two-level unitary gate~\cite{NC00}
\begin{equation}
G_{jk}(\theta) :=\cos\theta(|j\rangle\langle j|+|k\rangle\langle k|)
+\sin\theta(|k\rangle\langle j|-|j\rangle\langle k|),
\end{equation}
and by the gate~$X_i$ from the Heisenberg-Weyl basis ($i,j,k\in \mathbb{Z}_d$).
Our gate notation implies an identity operator acting on the rest of the space.
We augment these gates by their two-qudit controlled counterparts
\begin{equation}
CX_{jk}:=|j\rangle_\text{s}\langle j|\otimes X_{jk}
\end{equation}
and
\begin{equation}
CG_{jk}(\theta):=|j\rangle_\text{s}\langle j|\otimes G_{jk}(\theta)
\end{equation}
with the system as control,
and
\begin{equation}
CX_i:= X_i\otimes|i\rangle_\text{a}\langle i|
\end{equation}
with the ancilla as control.
We introduce a qudit multiplexer,
which generalizes the qubit case~\cite{SBM06},
as a sequence of two controlled Givens rotations
\begin{equation}
M_{jk}(\alpha,\beta):=CG_{jk}(\alpha)CG_{kj}(-\beta)
\end{equation}
depicted in Fig.~\ref{fig:mult},
with the proof of the circuit equivalence
following straightforwardly
for the qubit case~\cite{BBC+95}.
\begin{figure}
\begin{indented}
\item[]
\includegraphics[width=12.5cm]{Multiplexer.eps}
\end{indented}
\caption{
Circuit diagram for a multiplexer~$M_{jk}(\alpha,\beta)$.
Each wire represents an evolving qudit in the~$d$-ary representation
with~\textcircled{j} and~\textcircled{k}~$d$-ary control operation,
and
$\gamma_1\equiv\frac{1}{2}\left(\beta-\alpha+\frac{\pi}{2}\right)$
and $\gamma_2\equiv\frac{1}{2}\left(\beta+\alpha-\frac{\pi}{2}\right)$.
}
\label{fig:mult}
\end{figure}
\begin{proposition}
\label{prop:cir}
Given any $\mathcal{V}(\bullet):=V\bullet V^\dagger$
and $\mathcal{W}(\bullet):=W\bullet W^\dagger$ with~$V,W\in SU(d)$,
any channel $\mathcal{W}\left(\text{tr}_\text{a}\mathcal{U}\right)\mathcal{V}$
is extreme provided that
\begin{equation}\label{eq:extU}
U:= \prod_{i=d-1}^1 CX_i \prod_{j=d-1}^1\prod_{k=j-1}^0 M_{jk}(\alpha_{jk},\beta_{jk}),
\end{equation}
for all but a zero-measure subset of the rotation-angle sets~$\{\alpha_{jk}\}$ and~$\{\beta_{jk}\}$
with at most $(d^2-d)/2$ elements per set.
\end{proposition}
\begin{proof}
We prove the theorem by showing that the partial trace of~$U$ (\ref{eq:extU})
yields Kraus operators $F_i=_\text{a}\!\!\!\langle i| U|0\rangle_\text{a}$
that satisfy the normalization and linear independence conditions
of Prop.~\ref{prop:ext}.
To this end we define~$U'$ as a product of controlled-Givens rotations such that
\begin{equation}
U=\left(\prod_{i=d-1}^1 CX_i\right)U'.
\end{equation}
We define
\begin{equation}
\{u_{i\ell}\in \mathbb{R};i,\ell\in\mathbb{Z}_d\}
\end{equation}
such that
\begin{equation}
U' |0\rangle_\text{a}|\ell\rangle_\text{s}
= \sum_{i=0}^{d-1}u_{i\ell}|i\rangle_\text{a}|\ell\rangle_\text{s}.
\end{equation}
The unitary operator $U'$ corresponds to a channel
with diagonal Kraus operators $\{E_i\}$
such that
\begin{equation}
E_i|\ell\rangle_\text{s}=u_{i\ell}|\ell\rangle_\text{s}
\end{equation}
as
\begin{align}
\label{eq:U'}
U' |0\rangle_\text{a} |\ell\rangle_\text{s}
= \sum_{i=0}^{d-1} |i\rangle_\text{a} \langle i| U' |0\rangle_\text{a} |\ell\rangle_\text{s}
= \sum_{i=0}^{d-1} |i\rangle_\text{a} E_i |\ell\rangle_\text{s}.
\end{align}
We can identify $E_i$ in Eq.~(\ref{eq:U'})
with~$E_i$ in Eq.~(\ref{eq:F})
by setting
\begin{equation}
u_{i\ell}\equiv\tilde{a}_{i\ell}:=\sum_{j=0}^{d-1}a_{ij}\operatorname{e}^{\operatorname{i}2\pi\ell j/d}.
\end{equation}
Reincorporating the gates $CX_i$ yields
\begin{align}
\label{eq:}
U |0\rangle_\text{a} |\ell\rangle_\text{s}
= \sum_{i=0}^{d-1} |i\rangle_\text{a} X_i E_i |\ell\rangle_\text{s}
= \sum_{i=0}^{d-1} |i\rangle_\text{a} F_i |\ell\rangle_\text{s}.
\end{align}
A projection $|i\rangle_\text{a}\langle i|$ on the ancilla corresponds to the action of $F_i$ on the system.
The angles $\alpha_{jk}$ and $\beta_{jk}$ can be chosen (e.g., randomly) to satisfy
the linear independence of $\{F_i^\dagger F_j\}$.
This means the circuit $U$ realizes the Kraus operators $\{F_i\}$ for an extreme channel.
As there are ${d \choose 2}$ multiplexers,
the total number of independent parameters
is consistent with Corollary~\ref{cor:parameters}.
\end{proof}
As for constructing Kraus operators,
when the set $\{F_i^\dagger F_j\}$ is not linearly independent,
the circuit~(\ref{eq:extU}) realizes quasi-extreme channels.
As a result, the circuit in Prop.~\ref{prop:cir}
successfully yields simulations of arbitrary generalized extreme channels.
\section{Quantum channel simulation algorithm}
\label{sec:algo}
In this section, we first describe the classical optimization algorithm to design the quantum circuit,
and then we analyze the resultant quantum circuit in terms of space and time cost.
\subsection{Classical optimization algorithm to design the quantum circuit}
The first step in developing the quantum-circuit design procedure is to show the existence of
decomposition of a channel into a convex sum of generalized extreme channels.
Following from Ruskai's Conjectures~2, 3, 4 and 5~\cite{Rus07},
any channel $\mathcal{E}\in\mathscr{S}_d$ can be expressed as
\begin{equation}\label{eq:chadecom}
\mathcal{E}=\sum_{\imath=1}^{d} p_\imath \mathcal{E}^\text{g}_\imath, \quad
\sum_{\imath=1}^{d} p_\imath=1, \quad 0 \leq p_\imath \leq 1,
\end{equation}
for generalized extreme channels $\{\mathcal{E}^\text{g}_\imath\}\subset\mathscr{S}_d$ of Kraus rank $\leq d$.
The upper bound for the convex sum is not guaranteed to be~$d$,
but this upper bound is implied by Ruskai's conjectures,
which we adopt in our algorithm.
Even assuming that the upper bound holds,
an analytical formula for such a decomposition is unknown.
Now we describe the algorithm for the simulation of a general qudit channel.
The algorithm accepts the dimension~$d$ of the Hilbert space,
the description of a channel $\mathcal{E}$
and an error tolerance $\epsilon$ as input.
The output is a quantum circuit
(with an output of zero reserved for the case that the algorithm aborts before finding a satisfactory circuit)
and a bound~$\tilde{\epsilon}$ on the resultant circuit with
respect to the actual channel~$\mathcal E$ being simulated.
Our algorithmic procedure is as follows.
Based on Ruskai's conjecture,
we assume that any given channel can be decomposed into
a~$d$-fold sum of generalized extreme channels~(\ref{eq:chadecom}),
and we know from Prop.~\ref{prop:cir}
a description of the circuit for any generalized extreme channel.
If the assumed decomposition does not hold, the algorithm can fail and an output of zero for the circuit description ensues.
Thus, Eq.~(\ref{eq:chadecom}) and Prop.~\ref{prop:cir} together
inform us that a quantum circuit for the qudit channel can be realized by
choosing generalized extreme channel circuits randomly with each $\imath^\text{th}$ circuit chosen
with probability~$p_\imath$.
Our algorithm initially chooses a set of~$d$ generalized extreme channels randomly
and tests whether the resultant guessed channel~$\tilde{\mathcal{E}}$ is within distance~$\epsilon$
of the correct channel~$\mathcal E$.
Typically the guessed channel fails to be within the error tolerance
so we employ an optimization algorithm to pick a new~$\tilde{\mathcal{E}}$
and try again.
This procedure is repeated until a satisfactory circuit is found or
aborted if the optimization routine fails to find a good circuit within a
pre-specified number of trials.
We now determine the number of parameters in~$\mathcal E$ for optimization.
From Eq.~(\ref{eq:chadecom})
we see that there are~$d-1$ parameters of~$\{p_\imath\}$.
The unitary matrices~$V$ and~$W$ in Prop.~\ref{prop:ext} could be constructed as
products
\begin{equation}
V=\prod_i V_i,\;
W=\prod_j W_j
\end{equation}
with as many unitary operators~$\varkappa$
in the two products as needed to provide enough parameters for the optimization.
As there are~$d$ generalized extreme channels and $d^2-1$ free parameters in $SU(d)$,
we have
\begin{equation}
\varkappa d(d^2-1)
\end{equation}
free parameters associated with~$V$ and~$W$.
We add this number of parameters to the number of parameters for~$d$ generalized extreme channels,
namely $d(d^2-d)$ with~$d^2-d$ the number of free angles $\{\alpha_{jk},\beta_{jk}\}$,
and then add these to the number of probabilities~$\{p_\imath\}$.
The total number of parameters for the approximate channel should satisfy the inequality
\begin{equation}
\label{eq:count}
\varkappa d(d^2-1)+d(d^2-d)+(d-1)\geq d^4-d^2
\end{equation}
with the right-hand side corresponding to the number of parameters that specify the qudit channel.
For the most efficient simulation,
we minimize~$\varkappa$ so
\begin{equation}
\label{eq:paracount}
\varkappa=\left\lceil\frac{(d-1)(d^2+d+1)}{d(d+1)}\right\rceil.
\end{equation}
As an example, a qutrit channel has 72 parameters,
but our optimization is over 92 parameters.
Our analysis reduces to the qubit case~\cite{WBOS13}.
In that case~$d=2$ so the channel~$\mathcal E$ has 12 parameters
whereas the optimization of~$\mathcal E$ is over 17 parameters.
We note that the optimization precludes an efficient circuit-design algorithm even
in the qubit case, contrary to the earlier claim~\cite{WBOS13}.
The final step for the algorithm is to construct the objective function
for the optimization problem.
Mathematically we represent the correct $\mathcal{E}$ channel
by the Choi-Jamio{\l}kowski state~$\mathcal{C}$,
and the approximate circuit is represented by the state
\begin{equation}
\label{eq:C'}
\mathcal{C}'=\sum_\imath p_\imath \mathcal{C}^\text{g}_\imath.
\end{equation}
Channel decomposition in the Choi-Jamio{\l}kowski state representation is elaborated
in~\ref{sec:choirep}.
Our goal is to find the best possible~$\mathcal{C}'$ by optimization.
The objective function for optimization is given by the trace distance~$D_t(\mathcal{C}, \mathcal{C}')$,
which bounds the $\diamond$-norm distance~(\ref{eq:diamond})
between two channels~$\mathcal{E}$
and~$\mathcal{E}'$ according to~\cite{Wat13}
\begin{equation}\label{eq:cbound}
D_t(\mathcal{C}, \mathcal{C}')\leq \frac{\epsilon}{2}
\implies \|\mathcal{E}-\mathcal{E}'\|_{\diamond}\leq\epsilon.
\end{equation}
The trace distance is a convex function over the set of quantum states~\cite{NC00}.
Each $\mathcal{C}_\imath^\text{g}$ can be parameterized by a set of rotation angles
$\{\theta_{\imath\jmath}: \jmath=1,\dots, \varkappa (d^2-1)\}$ for the prior and posterior unitary operators,
and a set of rotation angles $\{\varphi_{\imath\jmath}: \jmath=1,\dots, d^2-d\}$,
which denote the sets $\{\alpha_{jk}\}$ and $\{\beta_{jk}\}$ altogether
from Eq.~(\ref{eq:extU}).
The range of the objective function is
\begin{equation}\label{eq:range}
0\leq D_t(\mathcal{C}, \mathcal{C}')\leq d.
\end{equation}
The optimization is to find~$\mathcal{C}'$ such that $D_t$ is minimized according to
\begin{equation}
\min_{\{\{p_\imath\},\{\theta_{\imath\jmath}\},\{\varphi_{\imath\jmath}\}\}}
D_t(\mathcal{C}, \mathcal{C}').
\end{equation}
The minimization over probabilities~$\{p_\imath\}$~(\ref{eq:C'})
is subject to the constraint
$\sum_\imath p_\imath -1=0$.
Our algorithm employs a simple nonlinear programming method~\cite{YS06}
on channels generated by partial trace of Haar-random-generated unitary operators on the dilated
space~\cite{Toth08}.
We simulate on MATLAB$^\circledR$ using MultiStart and GlobalSearch algorithms;
simulated annealing was less effective.
We have demonstrated numerically that our optimization algorithm
is successful for systems of up to four dimensions.
Our simulations yield errors of order~$10^{-2}\sim10^{-4}$ for qubit channels,
$10^{-2}$ for qutrit channels,
and $10^{-1}$ for two-qubit channels.
The errors for the case $d=4$ from the numerical simulation is rather large yet acceptable for demonstrating
the efficacy of our algorithm.
For high-accuracy simulation,
significantly greater computational resources are required.
As the system dimension~$d$ increases,
we expect at least a quadratic increase in run-time of the simulation with respect to~$d$
due to the built-in method employed by GlobalSearch or MultiStart program.
Moreover, given that resources are finite, e.g.\ run-time,
numerical optimization is not even guaranteed to succeed due to becoming stuck at
certain points in the parameter space.
Such problems are quite generic for optimization problems.
In order to illustrate how the simulation works,
a concrete example for simulating one randomly chosen qutrit channel is presented in~\ref{sec:qutrit},
and the pseudo-code of the algorithm is presented in~\ref{sec:pseudocode}.
\subsection{Space and time cost of quantum simulation circuit}
Here we consider the time and space cost for the quantum circuit
to simulate a generalized extreme qudit channel~$\mathcal{E}^\text{g}$
on a quantum computer based on qudits and single- and two-qudit unitary gates.
The generalized extreme qudit channel is dilated to a unitary operator~$U$ on two qudits,
which contains a sequence of multiplexers and a sequence of $CX_i$ gates acting between the system
and the ancillary qudits,
and also a prior qudit rotation and a posterior qudit rotation acting on the system.
An arbitrary single qudit rotation can be decomposed into a product of at most $d(d-1)/2$,
which is $O(d^2)$, two-level unitary gates ($\S$4.5.1, \cite{NC00}).
A controlled-Givens rotation $CG_{jk}(\theta)$ can be realized by
two Givens rotations and a $CX_{jk}$ gate, similar to the qubit case~\cite{BBC+95}.
The sequence of $CX_i$ gates can be realized by classically controlled $X_i$ gates since the ancillary system is traced out,
with an $X_i$ gate acting on the system conditioned on a $|i\rangle\langle i|$ projection on the ancilla.
As a result, the generalized extreme channel circuit can be
realized by a product of $O(d^2)$ $CX_{jk}$ gates and
continuously-parameterized Givens rotations.
To assess the cost of the quantum circuit,
we employ the Solovay-Kitaev-Dawson-Nielsen algorithm for qudits~\cite{DN06}.
From the error tolerance $\epsilon$, which is an algorithmic input for circuit design,
any Givens rotation can be approximated by an
\begin{equation}
O\left(\log\frac{d^{2}}{\epsilon}\right)
\end{equation}
sequence of universal qudit gates~\cite{DN06}.
As a result, the number of elementary gates,
hence computational time cost, of the generalized extreme channel circuit is
\begin{equation}
\label{eq:cost}
O\left(d^2\log\frac{d^{2}}{\epsilon}\right),
\end{equation}
and the space cost is two qudits.
The circuit corresponding to~$\tilde{U}$
yields an approximation~$\mathcal{\tilde{E}}^\text{g}$
to the desired generalized extreme channel~$\mathcal{E}^\text{g}$.
From~\cite{WBOS13}
\begin{equation}\label{eq:wbos13}
\|\mathcal{E}^\text{g}-\mathcal{\tilde{E}}^\text{g}\|_{1\rightarrow 1}\leq2 \|U-\tilde{U}\|
=2\|U\otimes\mathds{1}-\tilde{U}\otimes\mathds{1}\|,
\end{equation}
we obtain
\begin{equation}
\label{eq:skbound}
\|U-\tilde{U}\|\leq\frac{\epsilon}{2}\implies
\|\mathcal{E}^\text{g}-\mathcal{\tilde{E}}^\text{g}\|_{\diamond}\leq\epsilon.
\end{equation}
From strong convexity and the chain property of trace distance,
relations (\ref{eq:cbound}) and~(\ref{eq:skbound}) above
together ensure the desired simulation accuracy~(\ref{eq:diamond}).
Finally simulating an arbitrary channel is implemented by
probabilistically implementing different generalized extreme channels
according to the distribution $\{p_\imath\}$~(\ref{eq:chadecom}).
The space and time costs of a single-shot implementation of the channel
are one dit and two qudits for space and
the classical time cost for generating the random dits plus
\begin{equation}
O\left(d^2\log\frac{d^{2}}{\epsilon}\right)
\end{equation}
quantum gates.
In other words,
the quantum computational cost for simulating a random qudit channel is the same
as for simulating the generalized extreme channel,
and the additional cost is only classical:
dits plus running a random-number generator.
This cost can be explained by recognizing that the qudit channel simulator is
simply a randomized generalized extreme channel simulator.
On the other hand, estimating qudit observables accurately could require many shots,
with the number of shots depending on the particular observable.
\section{Conclusions}
\label{sec:conc}
In this work, we have presented a classical circuit-design algorithm for constructing
quantum simulation circuits for accurately simulating arbitrary qudit channels.
Our algorithm employs channel decomposition into a convex sum of generalized extreme channels,
leading to quantum circuits only for simulating generalized extreme channels,
which consume less computational resources.
In particular, we propose an ansatz for any extreme and quasi-extreme channels,
which has a concise mathematical structure.
We also show that the classical circuit-design algorithm can be
formalized as an optimization problem,
and we have performed numerical proof-of-principle simulations for low dimensional systems.
Our quantum channel simulation scheme transcends the standard quantum circuit model,
in that our circuit exploits resources other than quantum gates and measurements.
Using classical resources, e.g., bits, we can reduce the demand for quantum resources.
In particular, we show that it is possible to achieve the circuit lower bound~$O(d^2)$
for simulating non-unitary processes, which is the same lower bound for simulating
unitary operators.
Due to such a significant reduction of circuit cost,
our method is especially suitable for experimental implementation in the near future.
Our algorithm is designed for implementation in a fault-tolerant quantum computer
in order to achieve optimal efficiency,
but our simulation scheme can be implemented sooner with current technologies that are not fault-tolerant,
such as superconducting circuits, trapped ions or photons~\cite{BN09}.
Such systems admit more than two levels hence support qudits.
A proper error analysis for any physical system is required to assess its feasibility in the absence
of our fault-tolerance assumption.
This work generalizes the previous result of single-qubit-channel quantum simulation~\cite{WBOS13}
based on the extreme channel theory developed in Ref.~\cite{RSW02}.
Although a closed form of channel decomposition remains elusive,
our method provides an alternative approach to tackle this open problem
of channel decomposition into a convex sum of generalized extreme channels for the qudit case~\cite{Rus07}.
Although our algorithm relies on a conjectured upper bound to the number of
generalized extreme channels required to decompose the given channel~\cite{Rus07},
our numerical simulations succeed in delivering circuit designs that we verify work up to qudits of dimension 4.
If the conjectured upper bound for channel decomposition does not hold,
our algorithm can be modified by increasing the upper bound as an input to the algorithm,
but our algorithm is not guaranteed to be tractable under failure of this conjecture.
However, our algorithm is valuable as it works well for low dimensions,
and, if it fails at higher dimensions, superior algorithms could be tried.
\section{Acknowledgments}
The authors acknowledges AITF, NSERC and USARO for financial support.
This project was supported in part by the National Science Foundation under Grant No.\ NSF PHY11-25915.
We thank D. W. Berry, R. Colbeck,
M. C. de Oliveira, I. Dhand, R. Iten, R. Sweke, and E.\ Zahedinejad for
both valuable discussions and critical comments.
Also we thank J. Eisert for pointing out that the classical optimization is not convex.
|
\section{Introduction}
There is a growing interest among particle, condensed matter and atomic
physicists on the behavior of the shear viscosity $\eta$ and the degree to
which the ratio $\eta/s$, where $s$ is the entropy density, is close to
the lower bound $\eta/s\geq\hbar/(4\pi k_{B})$
conjectured by Kovtun, Son, and Starinets (KSS) \cite{Kovtun}.
The KSS conjecture has led to renewed interest as to
which fluids in nature are ``perfect'' fluids, i.e., those that come as
close as possible to minimizing the conjectured bound.
For atomic Fermi gases there have been experimental \cite{ThomasI,ThomasII}
and theoretical \cite{GuoII} studies of this ratio which suggest a
close approach to the KSS bound. These have been in
the specific unitary regime; no such studies are yet available
for the bosonic superfluids.
Theories of $\eta/s$ in graphene \cite{Fritz}, along with high $T_{C}$
superconductor experiments \cite{Rameau},
have also made claims that these exhibit an
$\eta/s$ near the KSS bound.
It is of interest then to perform concrete calculations of this ratio,
particularly in the presence of the many body physics which gives
rise to superfluidity. Thus, in this paper we address
the quasi-particle
contribution to $\eta/s$
for bosonic and fermionic superfluids. We arrive at a fairly generic
behavior for this ratio of the form
$\eta/s \propto T /\gamma(T)$, where $\gamma(T)$ is the
temperature
dependent inverse quasi-particle lifetime.
While classical arguments have led to a prediction
of this form, we show that it holds as well for the extreme non-classical
regimes in systems with widely different quasi-particle dispersion relations.
A scaling of $\eta/s$ of this form has also been argued to apply near
quantum critical points, or for nodal $d$-wave superfluids, and in these
contexts, it has been argued that the behavior is
suggestive of near perfect fluidity \cite{Rameau}.
More generally,
here we present a comparison of the dissipative transport properties of bosonic
and fermionic superfluids mostly at low temperatures, where bosonic superfluid theories exist
and are controlled. For the fermionic superfluids there is no restriction on temperature, and one can, furthermore,
probe the behavior of an interesting normal phase (with a pseudogap) in the regime of
BCS-BEC. In addition to the shear viscosity $\eta$, we focus
on the quasi-particle contributions to transport (which
are the exclusive contributions)
in the thermal conductivity $\kappa$, the $\omega \neq 0$
mass conductivity $\sigma$, and the off-diagonal
thermo-electric coefficients.
These are to be distinguished from condensate contributions,
which dominate the $\omega \equiv 0$ mass conductivity.
For the latter, a proper theory of transport has to deal with a number of
subtle features involving gauge invariance and the important
constraint in bosonic transport in which the (two particle) density
excitation spectrum or sound modes are intimately coupled to the single
particle excitations.
Our starting point for bosonic superfluids
is due to Wong and Gould \cite{Wong},
and Talbot and Griffin \cite{Griffin}, and known as the
``one-loop approximation''.
We note that prior to the recent focus on trapped atomic gas superfluids there
were puzzles concerning the behavior of the shear viscosity
alone, which originated in contrasting observations of the fermionic and
bosonic counterparts of liquid helium.
For example,
$^{3}\mathrm{He}$ and $^4\mathrm{He}$ exhibit a
remarkable difference in their shear viscosity for low
temperatures. At temperatures below the critical temperature
the shear viscosity of fermionic $^{3}\mathrm{He}$ has been
measured \cite{Roobol} to be a decreasing function
of decreasing temperature, whereas the shear
viscosity of bosonic $^4\mathrm{He}$ \cite{Woods} is an increasing function of
decreasing temperature.
In this paper we suggest that these differences are understood
as reflecting the different dispersions of the quasiparticles. Indeed,
a central theme of this work is that the shear viscosity itself provides
a sensitive measure of
the nature of the quasi-particle excitations.
Our calculations show that the one-loop Bogoliubov theory
for bosons and the BCS theory for fermions are
formally strikingly similar.
Nevertheless, primarily as a result
of differences in the quasi-particle excitation spectrum, as
well as the statistics, there
are important differences in superfluid
transport. Quite generally, in bosonic systems, because the dispersion relation is gapless,
the transport coefficients increase more
rapidly as a function of temperature when compared to the gapped fermionic systems.
This results in low temperature transport in the bosonic case being more
accessible experimentally.
\section{Theory}
Previous studies of superfluid transport have
relied heavily on kinetic theory and a Boltzmann equation
ansatz \cite{Khalatnikov, GriffinBookI}. A less widely applied approach has
been the use of linear response theory and Kubo formulae,
which we will use here.
The advantage of the Kubo formulae approach is that
by relating directly to Green's function diagrammatics one has better control over the
processes included in transport and the appropriate
constraints. This enables a more systematic imposition of perturbation expansions,
which is especially crucial when considering bosonic superfluids. Also important are the constraints which must
be imposed on separating the contributions associated with longitudinal and transverse correlations,
since it is only in the former that the condensate will directly enter.
Finally, a subtle but important issue here arises in the shear viscosity, for example, where the
Kubo formula shows that there are multiple response functions which enter in addition
to the simplest stress tensor-stress tensor correlator \cite{Bradlyn}. It is not as apparent how to include
these in a Boltzmann based approach.
Since it is likely
that dissipation in the ultracold gases
is linked to the details of the experimental set up, we will introduce dissipation
via a phenomenological parametrization within the Kubo approach.
The philosophy behind our phenomenological approach
to dissipation is similar to that articulated by Kadanoff and Martin,
who emphasized the importance
of the Kubo based correlation functions
and their symmetries \cite{KadanoffMartinII}.
In related work on superconductors \cite{KadanoffMartinI},
they argued for the suitability of introducing a
parametrization of the lifetimes associated with transport.
In building any phenomenology it is important to emphasize that
inter-particle collisions can not be the sole source of dissipation
in mass transport, as in the particle conductivity. This
particular transport coefficient
reflects the fact that the total momentum would
be conserved (in the presence of Galilean invariance)
without other sources of momentum relaxation.
By contrast, the strength of Boltzmann theory is that if all
the details of the processes giving rise to dissipation are well
established, one can incorporate dissipation via
specific collision integrals.
Within a Boltzmann based theory of bosonic superfluid transport,
there is a fairly extensive review
by Griffin \cite{GriffinBookI} in which the shear viscosity and the thermal
conductivity are addressed. At a qualitative
level our Kubo calculations are consistent with this earlier work,
but we also include additional transport coefficients. A Kubo formulation
of the shear viscosity of a trapped Bose condensed gas
was studied in \cite{Shahzamanian}
within the second order Beliaev approximation, but
this analysis
did not incorporate the contribution from anomalous Green's
functions.
This paper revisits this earlier work (albeit without a trap)
and with the important inclusion
of the anomalous Green's functions, which are a crucial component for
a consistent treatment of superfluid transport.
\subsection{Physical Analysis of the Quasi-particle Regime}
In order to further understand the role of the phenomenological
inverse lifetime, we address $\eta/s$ using
a simple classical argument \cite{KovtunII}.
The entropy density $s$ of a weakly interacting system is proportional
to the quasi-particle number density $n$:
\begin{align}s\sim k_B n.\nonumber\end{align}
The shear viscosity is proportional to the product of the average energy
per particle $\epsilon$, and the mean free time between collisions $\tau \equiv 1/\gamma$:
\begin{align}\eta \sim n \epsilon/\gamma.\nonumber\end{align}
Then, assuming $\epsilon\sim k_B T$, the ratio $\eta/s$ is
\begin{align}\label{eq:EtaS}\eta/ s\sim T/\gamma.\end{align}
In order for the quasi-particle picture to be valid, the particles
must be long lived: $\hbar \gamma <<k_BT$ so that the ratio $\eta/s$ is far above the KSS bound:
$$\eta/s >>\hbar/4\pi k_B.$$
Importantly, we
will show in this paper that, even in the non-classical regime, for both
bosonic and fermionic superfluids, an equation of the form given in Eq. (\ref{eq:EtaS}) results.
In this way the quasi-particle regime should be understood as a regime
where the system is far from being a perfect fluid.
\subsection{Overview of Our Transport Results}
We begin by summarizing our results, which serve to emphasize
the formal similarity of the bosonic one-loop transport theory
with the fermionic BCS transport theory.
We define the general transport coefficients $(L_{ij})$ via particle $(\mathbf{J}_{p})$
and heat $(\mathbf{J}_{Q})$ current densities as follows
\begin{align}
\label{eq:Jp}\mathbf{J}_{p} &= -L_{11} \nabla \mu- L_{12} \nabla T, \\
\label{eq:JQ}\mathbf{J}_{Q} &= -L_{21} \nabla \mu- L_{22} \nabla T,
\end{align}
where $\nabla\mu$ and $\nabla T$ represent imposed gradients
of the chemical potential (analogous to the electric field for a
charged system) and the temperature. (We work in units where $\hbar=k_{B}=e=1$.)
Here the particle or mass conductivity $\sigma\equiv L_{11}$ and the
thermal conductivity $\kappa\equiv L_{22}$. The off-diagonal transport coefficients
appear, for example, in the quasi-particle thermopower.
For a superfluid it should be stressed that the various
correlation functions that enter into the $L_{ij}$
may be distinct for longitudinal and transverse properties.
This distinction is most important for the mass conductivity,
as the longitudinal contribution reflects the condensate
(and diverges at zero momentum and frequency) while the
transverse contribution reflects the quasi-particles.
The shear viscosity is also represented in terms of this
transverse response.
Following the approach of Kadanoff and Martin \cite{KadanoffMartinI},
lifetime effects are phenomenologically incorporated by introducing
the parameter $\gamma(T)$. In this context $\gamma^{-1}$ was
introduced as a lifetime required to restore local equilibrium to a system
perturbed from the equilibrium state. It may therefore be regarded as an additional experimental parameter
for the particular system of interest. In the context of superfluids $\gamma^{-1}$
can be associated with quasi-particle lifetime processes, which in certain cases
are known \cite{Khalatnikov}.
Using the correlation functions which will appear in Eqs. (\ref{eq:X11}$-$\ref{eq:X22})
below, we find that, for bosons, the
transport coefficients (in 3d) are
\begin{align}\label{eq:NB}\eta^{B}&=\int_{0}^{\infty}dk\
\frac{k^6}{30\pi^2 m^2}\left(\frac{\xi_{\mathbf{k}}}{E_{\mathbf{k}}}\right)^2
\left(-\frac{\partial n(E_{\mathbf{k}})}{\partial E_{\mathbf{k}}}\right)\frac{1}{\gamma},
\\\label{eq:LB}\mathrm{Re}L^{B}_{ij}&=T^{1-j}\int_{0}^{\infty}dk\
\frac{k^{4}}{6\pi^2m^2}\xi_{\mathbf{k}}^{i+j-2}
\left(-\frac{\partial n(E_{\mathbf{k}})}{\partial E_{\mathbf{k}}}\right)\frac{1}{\gamma}.\end{align}
Note we have evaluated $\eta^{B}, \mathrm{Re}L^{B}_{ij}$ in the limit $\omega\rightarrow0$.
We introduce $n_{0}$ as the condensate density and $g$
as the interaction strength. The Hugenholtz-Pines theorem determines
the chemical potential, in the Bogoliubov
approximation, as $\mu^{B}=n_{0}g$.
The free particle dispersion relation is $\epsilon_{\mathbf{k}}=\tfrac{k^2}{2m}$
and we define $\xi_{\mathbf{k}}=\epsilon_{\mathbf{k}}+\mu^{B}$.
The Bogoliubov quasi-particle dispersion relation is then
$E_{\mathbf{k}}^2=\xi_{\mathbf{k}}^2-(\mu^{B})^2$.
We define $n(x)=[e^{x/T}-1]^{-1}$ as the Bose-Einstein distribution function.
The same calculations performed above for bosons can be performed
for strict BCS fermions. The only differences that arise are
a sign factor due to the different statistics, a degeneracy factor of two
due to spin and a redefinition of the dispersion relation.
For fermions the transport coefficients are
\begin{align}\label{eq:NF}\eta^{F}&=\int_{0}^{\infty}dk\ \frac{k^6}{15\pi^2 m^2}
\left(\frac{\xi_{\mathbf{k}}}{E_{\mathbf{k}}}\right)^2\left(-\frac{\partial f(E_{\mathbf{k}})}{\partial E_{\mathbf{k}}}\right)\frac{1}{\gamma},
\\\label{eq:LF}\mathrm{Re}L^{F}_{ij}&=T^{1-j}\int_{0}^{\infty}dk\
\frac{k^4}{3\pi^2m^2}\xi_{\mathbf{k}}^{i+j-2}\left(-\frac{\partial f(E_{\mathbf{k}})}{\partial E_{\mathbf{k}}}\right)\frac{1}{\gamma}.\end{align}
where $\xi_{\mathbf{k}}=\epsilon_{\mathbf{k}}-\mu^{F}$, $E_{\mathbf{k}}^2=\xi_{\mathbf{k}}^2+\Delta^2,$ and $f(x)=[e^{x/T}+1]^{-1}$.
(We have again evaluated $\eta^{F}, \mathrm{Re}L^{F}_{ij}$ in the limit $\omega\rightarrow0$.)
This expression for the shear viscosity has been obtained previously in \cite{GuoII}.
Similarly the mass and thermal conductivities $L_{11}, L_{22}$
are consistent with results obtained from BCS theory \cite{KadanoffMartinI}.
Our emphasis here is that a comparison between Eqs. (\ref{eq:NB}$-$\ref{eq:LB})
and Eqs. (\ref{eq:NF}$-$\ref{eq:LF}) shows the striking similarities
between the transport coefficients in bosonic and fermionic superfluids.
A key difference between the transport coefficients
arises from the soft quasi-particle excitations for
bosons as opposed to the gapped excitations for fermions.
\subsection{Details of the Derivation}
We proceed now to derive
Eqs. (\ref{eq:NB}$-$\ref{eq:LF}).
In linear response theory
the response of a system perturbed slightly from
thermal equilibrium is expressed in terms of correlation
functions of the unperturbed system \cite{KadanoffMartinII}.
Equations (\ref{eq:Jp}$-$\ref{eq:JQ}) lead to four possible
correlation functions involving combinations of particle and heat or energy
currents.
These four
correlation functions are
\begin{align}\label{eq:X}\tensor{\chi}_{ij}(x_{1}-x_{2},\tau_{1}-\tau_{2})=-\langle
T_{\tau}j_{i}(x_{1},\tau_{1})j_{j}(x_{2},\tau_{2})\rangle,\end{align} where
$i,j\in\{1,2\}$.
The particle and heat currents which appear above
are defined as \cite{KadanoffMartinI}
\begin{align}\label{eq:EC}j_{1}&=-\frac{i}{2m}\left(\nabla_{1}-\nabla_{1}'\right)\left.
\psi^{+}(1')\psi(1)\right|_{1'=1^{+}},
\\\label{eq:HC}j_{2}&=-\frac{i}{2m}\left(\partial_{t_{2}}\nabla_{2}'+\partial_{t_{2}}'\nabla_{2}\right)\left.
\psi^{+}(2')\psi(2)\right|_{2'=2^{+}}.\end{align}
For a generic superfluid correlation function $\tensor{\chi}_{ij}$, it is convenient to decompose into
longitudinal and transverse
components which are given by
$\chi_{ij}^{L}=\tfrac{\mathbf{q}\cdot
\tensor{\chi}_{ij}\cdot\mathbf{q}}{q^2},
\chi_{ij}^{T}=\tfrac{1}{2}(\sum_{\alpha}{\chi}^{\alpha\alpha}_{ij}
-\chi_{ij}^{L}).$
We define the Fourier transform by
$\tensor{\chi}_{ij}(x_{1}-x_{2},\tau_{1}-\tau_{2})
=\tfrac{1}{\beta}\sum_{i\omega_{m}}\int\tfrac{d^{3}q}{(2\pi)^3}\tensor{\chi}_{ij}
(\mathbf{q},i\omega_{m})e^{i\mathbf{q}\cdot(\mathbf{x}_{1}-\mathbf{x}_{2})}e^{-i\omega_{m}(\tau_{1}-
\tau_{2})}$.
Then the Kubo formulas for the transport coefficients, except those associated with $\chi_{11}$, are
\begin{align}\mathrm{Re}L_{ij}=-T^{1-j}\mathrm{lim}_{\mathbf{q}\rightarrow0}\frac{\mathrm{Im}\chi^{L}_
{ij}(\mathbf{q},\omega)}{\omega},\quad i,j\neq1.\end{align}
Using this definition, one can compute
transport coefficients $\mathrm{Re}L^{B}_{ij}, i,j\neq1$
for the bosonic case and $\mathrm{Re}L^{F}_{ij}, i,j\neq1$
for the fermionic case.
The quasi-particle contribution to the mass conductivity and the shear
viscosity (for which there is no condensate
component) depend only on the transverse component of
$\tensor{\chi}_{11}$ and are given by
\cite{KadanoffMartinII}
\begin{align}\mathrm{Re}\sigma(\omega\neq0)&=-\mathrm{lim}_{\mathbf{q}\rightarrow0}
\frac{\mathrm{Im}\chi_{11}^{T}(\mathbf{q},\omega)}{\omega},
\\\label{eq:N}\eta&=-m^2\mathrm{lim}_{\omega\rightarrow0}\mathrm{lim}_{\mathbf{q}\rightarrow0}
\frac{\omega}{q^2}\mathrm{Im}\chi_{11}^{T}(\mathbf{q},\omega).\end{align}
By limiting consideration in $\sigma$ to $\omega \neq 0$, we focus on the
quasi-particle transport. The total mass conductivity (which includes the condensate) is
$\mathrm{Re}\sigma(\omega)=\mathrm{Re}\sigma(\omega\neq0)+\frac{\pi
n_{s}}{m}\delta(\omega),$ where $\tfrac{n_{s}}{m}$ is the superfluid
density. The mass conductivity of the condensate
is infinite but all condensate thermoelectric coefficients vanish.
More specifically the condensate enters directly into only
$L_{11}$. Finally, we note that the
Onsager relation between the associated transport coefficients is
$L_{12}=L_{21}/T$.
\begin{figure*}
\subfloat{\includegraphics[width=2.25in,clip]{Sigma.eps}}\hspace{5mm}
\subfloat{\includegraphics[width=2.25in,clip]{Kappa.eps}}
\caption{(Color online) The normalized (a) mass and (b) thermal
conductivity coefficients for
bosonic (dashed blue line) and fermionic (solid red line) superfluids as functions of
$T/T_{C}$. The left and right axes are associated with the bosonic
and fermionic coefficients, respectively. The fermionic transport
coefficients are normalized by the normal state expressions
$\sigma_{N}=\sigma(\Delta=0), \kappa_{N}=\kappa(\Delta=0)$.
In the bosonic case we define $T/T_{C}$=$T/\mu^{B}$.
In the Bogoliubov approximation $n_{0}$ is the particle number at $T=0$.
Thus our calculations are confined to $T/T_{C}<<1$ and
so we use the normalization $\sigma_{0}=\tfrac{(2m\mu^3)^{1/2}}{6\pi^2\gamma},
\kappa_{0}=\tfrac{(2m\mu^5)^{1/2}}{6\pi^2\gamma}$. Here the
mass conductivity corresponds to the $\omega \neq 0$ contribution.}
\label{fig:Transport}
\end{figure*}
\subsection{Bosonic One Loop Approximation and Correlation Functions}
In order to evaluate the various $\tensor{\chi}_{ij}$
we introduce the appropriate Green's functions.
These functions are well established for the case of fermionic BCS superfluids.
For the bosonic case, the
one-loop approximation is based on the
Bogoliubov Green's functions and thus involves the Bogoliubov quasi-particle dispersion relation.
The Green's functions in the Bogoliubov approximation are
\begin{align}
G(K)&=\frac{u_{\mathbf{k}}^2}{i\omega_{n}-E_{\mathbf{k}}}-\frac{v_{\mathbf{k}}^2}{i\omega_{n}+E_{\mathbf{k}}},\\
F(K)&=-u_{\mathbf{k}}v_{\mathbf{k}}\left(\frac{1}{i\omega_{n}-E_{\mathbf{k}}}-\frac{1}{i\omega_{n}+E_{\mathbf{k}}}\right),
\end{align}
where $u_{\mathbf{k}}^2=\tfrac{1}{2}(1+\xi_{\mathbf{k}}/E_{\mathbf{k}}),v_{\mathbf{k}}^2=u_{\mathbf{k}}^2-1.$
Because bosonic superfluid theories involve a controlled perturbation in the
interaction strength, they lead to a clear hierarchy of diagrams and
we can restrict attention in the dilute fluid limit to those involving one or
at most two Green's functions. The latter constitute the ``loops" of the
transport approximation.
For transverse response functions,
the only diagrams that contribute are those that cannot be
divided into two parts by removing one line representing a single-particle propagator.
Such diagrams are called proper. \\\noindent The condensate contributions to a generic
correlation function (dependent on a single-particle Green's function)
are not proper, and therefore do not contribute to the transverse response functions.
It follows that, for a one-loop theory, the transverse component of a
generic correlation function is completely determined by diagrams containing only two single-particle Green's functions.
For longitudinal correlation functions, other than $L_{11}$, there are no condensate contributions and
again the leading order contribution involves two single-particle Green's functions.
In the superfluid phase there are two such Green's functions (the anomalous and normal Green's functions.)
At this bosonic one-loop level we
relate these correlation functions to
the imaginary time single particle Green's functions in position
space, given by
$G(x,\tau)$ (normal), $F(x,\tau)$ (anomalous), defined by:
$\langle T_{\tau}\psi(x_{1})
\psi^{+}(x_{2})\rangle=-G(x_{1}-x_{2})+n_{0}$ and $\langle
T_{\tau}\psi(x_{1})\psi(x_{2})\rangle=-F(x_{1}-x_{2})+n_{0}$.
For convenience, we make the following definitions:
the four vector summation $\sum_{K}\equiv-\tfrac{1}{\beta}\sum_{i\omega_{n}}
\int\frac{d^{3}k}{(2\pi)^3}$, the vertex factors
$\mathbf{v}_{1}=\left(\tfrac{\mathbf{k}+\tfrac{1}{2}\mathbf{q}}{m}\right),
\mathbf{v}_{2}=\frac{\mathbf{q}}{2m},$ and
$\mathbf{v}_{3}=\left((i\omega_{n}+i\omega_{m})\frac{\mathbf{k}}{2m}+
i\omega_{n}\frac{\mathbf{k+q}}{2m}\right)$. The dissipative parameter $\gamma$ previously introduced also serves to analytically
continue the Matsubara frequencies $i\omega_{m}$ to
real frequencies $\omega$ via: $i\omega_{m}=\omega+i\gamma$.
With these definitions, the
four momentum space correlation functions can be computed.
The particle current-particle current correlation function is
given by:
\begin{align}\label{eq:X11}
&\tensor{\chi}_{11}(\mathbf{q},i\omega_{m})=n_{0}\mathbf{v}_{2}\mathbf{v}_{2}
[G(Q)+G(-Q)-F(Q)-F(-Q)]\nonumber\\&+\sum_{K}\mathbf{v}_{1}\mathbf{v}_{1}\left[G(K)G(K+Q)-F(K)F(K+Q)\right].
\end{align}
The particle current-heat current correlation function is:
\begin{align}\label{eq:X12}&\tensor{\chi}_{12}(\mathbf{q},i\omega_{m})
=\sum_{K}\mathbf{v}_{1}\mathbf{v}_{3}\nonumber\\&\times\left[G(K)G(K+Q)+F(K)F(K+Q)
\right].\end{align}
The heat current-particle current correlation function is:
\begin{align}&\tensor{\chi}_{21}(\mathbf{q},i\omega_{m})=\sum_{K}\mathbf{v}_{3}\mathbf{v}_{1}\nonumber\\&\times\left[G(K)G(K+Q)-F(K)F(K+Q)\right].
\end{align}
The heat current-heat current correlation function is:
\begin{align}\label{eq:X22}&\tensor{\chi}_{22}(\mathbf{q},i\omega_{m})=\sum_{K}\mathbf{v}_{3}\mathbf{v}_{3}\nonumber\\&\times\left[G(K)G(K+Q)+F(K)F(K+Q)
\right].\end{align}
Our expressions in Eqs. (\ref{eq:X11}$-$\ref{eq:X22}) contain
all possible contributions to the irreducible transverse response functions \cite{GriffinBookII}.
Note that, the correlation functions
$\tensor{\chi}_{12}$ and
$\tensor{\chi}_{21}$ differ in the relative sign of
the contribution from the anomalous Green's functions.
It follows that, in order to satisfy the Onsager
relation, the anomalous Green's functions must give no contribution to
the transport coefficients $L_{12}$ and $L_{21}$. This is confirmed
explicitly by direct calculation.
As can be seen, the particle current-particle
current correlation function ($\tensor\chi_{11}$) which appears in Eq.
(\ref{eq:X11}), unlike all the other $\tensor\chi_{ij}$, contains a term proportional to the
condensate density $n_{0}$. This term is purely longitudinal and of no
interest here. In order to ensure
charge conservation (via the longitudinal or $f$-sum rule)
in the superfluid phase, the condensate requires a consistent treatment,
analogous to collective mode effects in fermionic superfluids.
Finally, from the definitions of the transport coefficients,
combined with the correlation functions in
Eqs. (\ref{eq:X11}$-$\ref{eq:X22}) and the Bogoliubov
Green's functions, the resulting transport coefficients
$\eta^{B}$ and $\mathrm{Re}L^{B}_{ij}$ are given by
the expressions in Eqs. (\ref{eq:NB}$-$\ref{eq:LB}).
Figure (\ref{fig:Transport}) shows the comparison between the
normalized low temperature bosonic and
fermionic transport coefficients, corresponding to mass and
thermal conductivity. It is clear from the
figures that the quasi-particle transport coefficients
at these low $T$ differ by several orders of magnitude. This is
due to the differences in the quasi-particle excitation
spectrum. From an experimental perspective, it appears
rather prohibitive to measure very low temperature transport properties of
Fermi systems. By contrast it appears Bose systems lend themselves to
these low $T$ studies.
\subsection{Low Temperature Analysis}
In general, the bosonic transport
coefficients exhibit power law behavior, whereas the
fermionic transport coefficients exhibit an
exponentially suppressed response. Explicitly, in the low temperature limits
($T<<\mu^{B},T_{C}$) we find that for bosons
\begin{align}
\mathrm{Re}L^{B}_{ij}&\rightarrow\frac{2\pi^2}{45\gamma}m^{1/2}(\mu^{B})^{i+j-9/2}T^{5-j},
\end{align}whereas for fermions
\begin{align}
\mathrm{Re}L^{F}_{11}&\rightarrow\frac{2g(E_{F})p_{F}^2}{3m^2\gamma}\left(\frac{2\pi\Delta_{0}}{T}\right)^{1/2}e^{-\Delta_{0}/T},\\
\mathrm{Re}L^{F}_{22}&\rightarrow\frac{2g(E_{F})p_{F}^2}{3m^2\gamma}\left(\frac{2\pi\Delta_{0}^3}{T}\right)^{1/2}e^{-\Delta_{0}/T},
\end{align}
where $p_{F}$ is the Fermi-momentum, $\Delta_{0}=\Delta(T\rightarrow0)$, and $g(E_{F})$ is the density of states at $E_F$.
In BCS theory, assuming the chemical potential is of order
$\mu^{F}\sim E_{F}$ and with exact particle-hole symmetry, $\mathrm{Re}L^{F}_{12}\rightarrow0$.
\begin{figure*}
\subfloat{\includegraphics[width=2.25in,clip]{Eta.eps}}\hspace{5mm}
\subfloat{\includegraphics[width=2.25in,clip]{EtaS_Ratio.eps}}
\caption{(Color online) (a) The normalized shear viscosity of
bosonic (dashed blue line) and fermionic (solid red line) superfluids as functions of
$T/T_{C}$. The left and right axes are associated with the bosonic
and fermionic coefficients, respectively. The fermionic shear viscosity is normalized by the normal state expression $\eta_{N}=\eta(\Delta=0)$. In the bosonic case we define $T/T_{C}$=$T/\mu^{B}$. In the Bogoliubov approximation $n_{0}$ is the particle number at $T=0$. Thus our calculations are confined to $T/T_{C}<<1$ and so we use the normalization $\eta_{0}=\tfrac{(2m^3\mu^5)^{1/2}}{15\pi^2\gamma}$.
(b) The low temperature limit of the shear viscosity to entropy density ratio. Both limits are linear in $T/\gamma(T)$, and so have the same functional form. The only difference is the associated axes.}
\label{fig:Eta}
\end{figure*}
\section{Calculation of $\eta$ and $\eta/s$}
Kovtun, Son and Starinets (KSS) \cite{Kovtun} have made an interesting
conjecture concerning the shear viscosity. They conjecture that
any relativistic quantum field theory at finite temperature and zero
chemical potential has a ratio of shear viscosity to entropy density
satisfying the bound $\eta/s\geq\hbar/(4\pi k_{B})$. Despite the
construction of certain systems that violate the KSS bound
\cite{Cherman}, the KSS conjecture has lead to renewed interest in
what the perfect fluids in nature are, i.e., those that come as
close as possible to minimizing the conjectured bound. It has been
shown by KSS that fluids that saturate this bound are those with a
dual gravity description.
An interesting feature of the KSS bound is that it is independent of
the speed of light $c$. Therefore, a non-relativistic quantum system
is a possible candidate for the perfect fluid. Here we
investigate the magnitude of
$\eta/s$ arising from quasi-particle
transport in the bosonic one-loop and fermionic BCS superfluids.
A variant of the KSS conjecture extends the applicability of the conjectured bound of $\eta/s$ to
the case of non-zero chemical potential \cite{Son}. If we allow
$\mu^{B}, \mu^{F}\neq0$, then the low temperature entropy limits for bosons and fermions are
\begin{align}
\label{eq:SB}s^{B}&\rightarrow\frac{2\pi^2}{45}\left(\frac{m}{\mu^{B}}\right)^{3/2}T^3,\\
\label{eq:SF}s^{F}&\rightarrow 2g(E_{F})\left(\frac{2\pi\Delta_{0}^3}{T}\right)^{1/2}e^{-\Delta_{0}/T}.
\end{align}
Similarly the low temperature shear viscosity limits for bosons and fermions are
\begin{align}
\label{eq:EtaB}\eta^{B}&\rightarrow\frac{2\pi^2}{225\gamma}\left(\frac{m}{\mu^{B}}\right)^{3/2}T^4,\\
\label{eq:EtaF}\eta^{F}&\rightarrow\frac{2g(E_{F})p_{F}^4}{15m^2\gamma}\left(\frac{2\pi T}{\Delta_{0}}\right)^{1/2}e^{-\Delta_{0}/T}.
\end{align}
Depending on the temperature dependence of the quasi-particle lifetimes
($\gamma^{-1}$) , the bosonic shear viscosity can exhibit
an upturn for low temperatures. However, due to the exponentially
suppressed term, the fermionic shear viscosity is not expected to exhibit an
upturn, regardless of the parameter $\gamma(T)$.
Using the low temperature limits of $s$ and $\eta$ in Eqs. (\ref{eq:SB}$-$\ref{eq:EtaF}),
we obtain the ratio $\eta/s$ for bosons and fermions:
\begin{align}\label{eq:ESB}\eta^{B}/s^{B}&\rightarrow\frac{1}{5}\frac{T}{\gamma},\\
\label{eq:ESF}\eta^{F}/s^{F}&\rightarrow\frac{4}{15}\left(\frac{E_{F}}{\Delta_{0}}\right)^2
\frac{T}{\gamma},\end{align}
It should be noted that
once the entropy density is included, both bosons and fermions exhibit
the same $T/\gamma(T)$ dependence in their $\eta/s$ ratios.
This derivation confirms the arguments
given earlier, namely that systems with a quasi-particle description
have a large ratio of $\eta/s$, with the generic form $\eta/s\sim T/\gamma(T)$.
An example of the temperature dependence used
for $\gamma$ in $^{4}$He \cite{Khalatnikov} would predict an upturn in
$\eta/s$ at low $T$ for the bosonic superfluid case.
While both bosonic and fermionic cases considered have a similar
functional form for the ratio $\eta/s$, the low temperature limits
of the entropy and shear viscosity of bosons and fermions are markedly different:
Figure (\ref{fig:Eta})
presents a plot of the normalized shear viscosity and the ratio $\eta/s$ for
bosonic and fermionic BCS superfluids.
While $\eta$ is highly suppressed for the fermionic case (as compared
with a bosonic superfluid), in the $\eta/s$ ratio the fermionic contribution
is highly enhanced. This is due to the fact that for fermions there are
two different energy scales present, $E_{F}$ and $\Delta_{0}$, while for bosons $\mu$ is the
only energy scale. Equation (\ref{eq:ESF}) shows that there is a factor of
the ratio of these two energy scales $(E_F/\Delta_0)^2$ which appears.
This is of course a very large number in the strict BCS limit, which
reflects the fact that the entropy in BCS theory is much smaller than
the shear viscosity.
\subsection{BCS-BEC Crossover}
\begin{figure*}
\includegraphics[width=5.5in,clip]{Eta_BCS_BEC.eps}
\caption{(Color online) The normalized shear viscosity in BCS-BEC crossover theory.
The shear viscosity is normalized by the normal state expression $\eta_{N}=\eta(T=T^{*})$.
The parameters for each plot are (a) $T_{C}=0.12T_{F}, T^{*}=0.13T_{F}, 1/k_{F}a=-1$,
(b) $T_{C}=0.26T_{F}, T^{*}=0.50T_{F}, 1/k_{F}a=0$, and (c) $T_{C}=0.21T_{F}, T^{*}=1.28T_{F}, 1/k_{F}a=1$.}
\label{fig:Eta_BCS_BEC}
\end{figure*}
Our primary analysis has been the transport properties of both bosonic and
fermionic superfluids in the temperature regime far below the critical temperature.
While theories of bosonic superfluidity are restricted to low temperatures,
fermionic superfluids can be studied up to $T_C$.
Equally interesting is the behavior in the normal phase ($T>T_{C}$)
of superfluids
in the presence of stronger attractive interactions, as associated with
the crossover from BCS to BEC.
In this normal or ``pseudogap" phase, pairing persists above $T_C$
and is expected to lead to suppressed shear viscosity.
With the discovery of the trapped atomic gases one has access to
fermionic superfluids with variable attraction, as parameterized
by the dimensionless scattering length $1/k_{F}a$, where
$1/k_{F}a=0$ is the so-called unitary regime.
One particular BCS-BEC scenario \cite{GuoII} by members of our
group (which is explicitly sum-rule consistent)
has addressed $\eta^{F}(T)$ theoretically over the entire
range of temperatures using as a framework the
BCS wave function for the ground state (with arbitrary attraction
and self consistent chemical potential). Here one finds
that $\eta^{F}=$
\begin{align}
&\int_{0}^{\infty}dk\ \frac{k^6}{15\pi^2 m^2}
\left(\frac{\xi_{\mathbf{k}}}{E_{\mathbf{k}}}\right)^2\left(1-\frac{\Delta^2_{pg}}{E_{\mathbf{k}}^2}\right)
\left(-\frac{\partial f(E_{\mathbf{k}})}{\partial E_{\mathbf{k}}}\right)\frac{1}{\gamma},
\end{align}
In this scenario there are two gap functions
$\Delta_{sc}$ and $\Delta_{pg}$, where the first represents the order
parameter, and the second the contribution to the excitation gap
associated with non-condensed pairs. The total gap is given by
$\Delta^2=\Delta^{2}_{sc}+\Delta^{2}_{pg}$, so that the
excitation gap $E_{\mathbf{k}}$ takes the usual BCS form in terms of
the full gap $\Delta$.
This previous work \cite{GuoII} and the above
equation shows that the effect of the non-condensed pairs, associated with
the pseudogap $\Delta_{pg}$, is to reduce the shear viscosity. This is due to the fact that
when pairs are present there are fewer fermions to contribute to the shear viscosity.
In Fig. (\ref{fig:Eta_BCS_BEC}) we show plots of the normalized shear viscosity
for this BCS-BEC crossover scenario.
The plots from left to right correspond to passing from the BCS side
of unitarity, to unitarity, to the BEC side.
The exponential suppression of transport in fermionic superfluids is
then reflected in the behavior of $\eta$: as the gap $\Delta$ increases
in size from BCS to BEC the shear viscosity is accordingly
suppressed at the lowest $T$. It should be emphasized that the BEC limit
still reflects the pairing of fermions and will not coincide
with Bogoliubov descriptions of bosons, as the latter involves
boson-boson interactions. These are not incorporated into the generalized
BCS wave function.
Of particular interest are the observations
\cite{ThomasI,ThomasII}
that the unitary superfluids have a shear viscosity which
is close to the KSS bound. This makes them very different
from the strict BCS superfluids with extremely large
$\eta/s$, which we studied earlier in this paper. There are hints from
Eq. (\ref{eq:ESF})
(which shows that $\eta/s$ for strict BCS fermions contains
a prefactor
$(T_{F}/T_{C})^2$) that as unitarity is approached and
$T_{F}/T_{C}$ becomes order one, that $\eta/s$ is significantly
reduced relative to the strict BCS case.
Indeed, the analogous prefactor $\sim(T_{F}/T_{C})^2$ is now several orders of magnitude smaller
than its counterpart for BCS theory.
As addressed in \cite{GuoII} and seen explicitly
in Fig. (\ref{fig:Eta_BCS_BEC}), the presence of the
non-condensed pairs via $\Delta_{pg}$ does not affect the
exponential suppression of $\eta$ at low temperatures. However,
at the same time the entropy acquires an additional bosonic
contribution
($s=s^{F}+s^{B}$) \cite{Stajic}, where $s^{B}$ dominates at low $T$ and is a power of $T$.
Thus, the ratio $\eta/s$
will not be a linear function of $T/\gamma$, as was found for strict
BCS superfluids; rather, these near-unitary superfluids will exhibit
near perfect fluidity.
We summarize this discussion by emphasizing that our theory of
BCS-BEC crossover is a
theory of fermions and the
BEC limit does not include the direct effects of
inter-boson interactions (which give rise to the sound
mode excitations of Bogoliubov theory). These sound-mode
effects do, of course, appear in the collective modes as
the Nambu Goldstone bosons. Such collective modes must
be included in some transport coefficients and must
not be included in others. More precisely the Nambu Goldstone
modes in fermionic superfluids
of this type couple to the longitudinal response. They
do not couple to the transverse response, of which the
shear viscosity is one example. Our previous work
on $\eta$ for the BCS-BEC system used a theoretical approach
which analytically satisfied the transverse sum rule \cite{GuoII}.
A failure to satisfy the sum rules
is one of the best internal checks on whether or not,
and precisely where,
collective modes must be included in a given response function.
\section{Conclusions}
We have compared
the $\omega \rightarrow 0$ mass conductivity, the shear viscosity, and the thermal conductivity
in bosonic and fermionic superfluids based on a Kubo formula approach within
the one-loop Bogoliubov and closely related
BCS approximation. At this level
of approximation, our work
demonstrates the formal (albeit non-quantitative)
similarity between the transport
behavior of both superfluid types.
The transverse response functions
do not contain condensate contributions. Similarly for
the longitudinal thermoelectric coefficients (aside from the
$\omega \equiv 0$ mass conductivity) no condensate
contributions appear.
Thus, it is appropriate to characterize these coefficients
entirely in terms of their quasi-particle contributions,
as we have done here.
Of central interest here is the fact that even though the
shear viscosity for Bogoliubov and BCS superfluids
have dramatically
different temperature dependence,
their ratios in terms of the entropy density have
precisely the same linear $T/\gamma(T)$ dependence (where $\gamma(T)$ is the inverse
quasi-particle lifetime) with
very different prefactors.
When considering the extension of BCS theory to BCS-BEC
crossover near unitarity, we find a very different temperature
dependence. Here because there are both bosonic and fermionic
degrees of freedom, there is no simple $T/\gamma(T)$ scaling.
Indeed, due to the suppression of the shear viscosity,
it appears that unitary Fermi gases are a candidate
for nearly perfect fluids.
We stress that
pure bosonic or Bogoliubov theories of superfluidity have a structure
not exhibited in the BCS-BEC crossover; this difference
arises due to the soft dispersion relation
present in the long wavelength limit of Bogoliubov theory.
Similarly, inter-boson interactions are not
directly present at the level of a BCS-based theory of unitarity. Here the dominant
many body physics is an attraction between fermions, as distinct
from boson-boson interactions. Recent experiments \cite{ThomasIII} seem to confirm
this exponential suppression in the low temperature
shear viscosity as unitarity is approached.
We end by noting that essentially all reasonable
models for the temperature dependence of the transport
lifetime will give an upturn in $\eta/s$ at low $T$, but not,
for the case of fermions, in $\eta$ itself. This
appears consistent with the observed differences between
helium-3 and helium-4 superfluids \cite{GuoI}.
This work is supported by NSF-MRSEC Grant
0820054. We thank Adam Ran\c con for many helpful
conversations.
\vskip5mm
|
\section{Related Work}\label{related}
In this section, we give an overview of the text-based authentication
mechanisms and the proposed techniques to gain resilience against
observation attacks.
\textbf{Traditional password.} Recall-based user-chosen textual password
is the most widely used authentication scheme on the web. However, it is
vulnerable to online guessing attacks~\cite{captcha,guessing08} that
exploit password reuse and predictable patterns, since users use the
same password on an average of six different accounts~\cite{pwhabit} and
follow predictable strategies (e.g., using dictionary words or names)
while creating passwords~\cite{pwpattern1,pwpattern2}.
To motivate users to create stronger passwords, different password
restriction policies~\cite{pwpattern1,pwmeter13} have been deployed,
such as increasing minimum length of passwords, using combination of
different types of characters, asking users to change passwords at
regular intervals, and using password strength meters. However, in
separate studies,
Proctor et al.~\cite{proctor_sec} and Shay et al.~\cite{pwpattern1} reported that
password restriction policies do not necessarily lead to more secure
passwords. Rather, in some cases, they adversely affect the
memorability.
\textbf{Mnemonic Password.} Kuo et al.~\cite{mnemonic} studied the
guessing-resistance of user-selected mnemonic phrase based passwords, in
which the user chooses a memorable phrase and uses a character (often
the first letter) to represent each word in the
phrase. Results~\cite{mnemonic} show that user-selected mnemonic
passwords are slightly more resistant to brute-force attacks than
traditional text-based passwords, . Mnemonic passwords are found more
predictable when users choose common phrases to create their
passwords. A properly chosen dictionary may increase the success rate in guessing mnemonic passwords~\cite{mnemonic}.
\textbf{System-assigned password.} Randomly assigned textual passwords
provide better resilience against online guessing attacks in comparison
to user-chosen text-based passwords~\cite{yan_sec}. Wright et
al.~\cite{text_recog} compared the usability of three different
system-assigned textual password schemes: i) Recognition-based words,
ii) Recall-based words, and iii) Recall-based random letters. In the
recognition-based scheme, the user has to recognize a set of words, each
word being displayed on screen amongst a set of distracter words. In the
second scheme, users have to remember a list of four whole words, which
serves as one password. In the last scheme, users are assigned random
passwords of six lower-case letters. The password space for all the
conditions are kept same ($28$ bits). Results for memorability show that
word recall performs the worst and no significant difference is found
between recognition and letter recall. Furthermore, the time required to
login is significantly different for all pairings, and notably worst in
the recognition condition.
\textbf{System-assigned passphrase.} Shay et al.~\cite{passphrase}
investigated the usability of system-assigned passphrases
(space-delimited sets of natural language words), which did not show
significant improvement over the system-assigned textual passwords of
similar entropy. The study~\cite{passphrase} indicates that a majority of participants in each condition wrote down their passwords/passphrases, and around half of the participants who did
not write them down failed to recall their authentication secrets after
two days.
\textbf{PTP.} Forget et al.~\cite{persuation,forget_thesis} proposed the
Persuasive Text Passwords (PTP) scheme, in which, after a password is
created by the user, PTP improves its security by placing
randomly-chosen characters at random positions into the password. Users
may shuffle to re-position the random characters until they find a
suitable combination to memorize. PTP is resilient against attacks
exploiting password reuse and predictable patterns. However, knowing that a system uses PTP and knowing how PTP
works would allow attackers to refine their cracking strategies. The
memorability for PTP is found to be just $25$\% when two random
characters are inserted at random positions~\cite{forget_thesis}.
\textbf{Cognitive question.} The primary investigation of Furnell
et al.~\cite{qa04} reported that $70$\% users preferred to have
cognitive questions for authentication. However, most of the questions
used in their study (e.g., ``What is the name of your favorite
relation?", ``What is your favorite shape?") offer very limited entropy
and thus can be easily guessed by attackers. Also, their approach is
vulnerable to shoulder surfing and keystroke loggers. Later
studies~\cite{no_sec,qa_just09,qa_fb} performed comprehensive analysis
on cognitive questions to figure out the usability and security of
different types of questions. We have carefully considered their findings to
design Q-A. A detailed discussion follows in the next section.
\textbf{Resilience to observation attacks.} Textual passwords are vulnerable to {\em external observation attacks}, such as shoulder surfing~\cite{shoulder06}. De Luca et
al.~\cite{shoulder_chi13} examined the impact of fake cursors in
providing resilience against shoulder surfing when passwords are entered
through an on-screen keyboard. However, results show that the users act
predictably to identify their active cursor (e.g., moving the mouse
cursor to the border of the interaction area or moving the mouse in
small circles), which may make it easy for the shoulder surfer to find
the real cursor and subsequently the authentication secret from
on-screen keyboard entries.
Gaining the user's credentials through malware (e.g., keystroke loggers,
mouse loggers) is called an {\em internal observation
attack}~\cite{design_space}. Text-based passwords are prone to
keystroke loggers~\cite{keylogger}. A few tricks have been proposed to
hide passwords from keystroke loggers, such as typing fake characters in
multiple text-boxes at the time of entering the
password~\cite{keylogger}. These tricks, however, require users to be
conscious about keystroke loggers and take proactive measures, which is
difficult~\cite{pwhabit,pwpattern2}.
\textbf{Conclusion and open problems.} As we see from the above discussion, textual password schemes that
provide better memorability are vulnerable to guessing attacks (e.g.,
traditional and mnemonic passwords). On the other hand, text-based passwords with higher resilience against guessing attacks suffer from memorability problems (e.g., system assigned passwords and passphrases,
PTP). Moreover, textual passwords are vulnerable to observation attacks.
{\em Thus, despite a large body of research, it remains a critical challenge in password research to build an authentication system that offers both high memorability and guessing resilience, with robustness against observation attacks.}
\section{Conclusion and Future Work}\label{conc}
Our study on Q-A finds that cognitive questions possess good possibilities in primary authentication in terms of both usability and security. We note that in addition to providing sufficient entropy to prevent online guessing attacks, Q-A is resilient to both internal and external observation attacks. Memorability in Q-A was found to be $100$\% in our study. Users' feedbacks reveal that they are highly satisfied with the memorability and security of this scheme, which would be most appropriate for online accounts with high security requirements (e.g., online banking, e-commerce~\cite{pw_diary}).
Now that we have observed promising memorability for single Q-A password in a lab setting, it seems reasonable to plan for a field study with larger and more diverse populations in order to examine if cognitive questions can provide a usable solution to the memorability and interference problems with multiple passwords.
\section{Discussion}\label{disc}
The login success rate for Q-A was found to be $100$\% in both Recall-1
and Recall-2. Here we discuss the anticipated factors that might have
played an important role to gain such high a performance during
authentication.
\begin{itemize}
\item Q-A asks for already known information that are related to the
long term memory of a user. So, the user does not have to memorize any
artificial information as her authentication secret.
\item Q-A provides users with the flexibility to choose any six
questions from a set of twenty, which they find most applicable to them.
\item The basic restrictions (see~\S\ref{sec_adv}) should have motivated
users to enter the correct answer instead of a weak placeholder one
(`Aab'). Real answers to cognitive questions have been shown to be
more memorable than placeholder answers~\cite{no_sec}.
\item Being specific with the question helps to accurately recall the
answer in Q-A. For example: instead of asking ``Where did your father
and mother meet?'', we would ask ``In what city or town did your
father and mother meet?'
\item Q-A makes users focus on fetching a letter from the given position
in her answer, which reduces the chance of typing mistakes that occur
when they have to enter the whole answer.
\item Q-A is case-insensitive, while confusion between capital and
lowercase letters has been shown to be a prominent reason for making
mistakes when answering cognitive questions~\cite{qa04}.
\end{itemize}
\subsection{Limitations and Ecological Validity}
Our participants were young and university educated, which represents a large number of frequent Web users, but may not generalize to the entire population. As the study was performed in a lab setting, we were only able to gather data from $22$ participants.
Although field studies may provide superior ecological validity, lab studies have been preferred to examine brain-powered memorability of passwords~\cite{ecology13}. Moreover, lab studies have the advantage of taking place in a controlled setting, which helps to establish performance bounds and figure out whether field tests are worthwhile in future research.
Our results for login time are conservative since they reflect initial use. As noted in existing research~\cite{design_space}, login time likely decreases with frequent use of a scheme. Chiasson et al.~\cite{passpoint3} find that users get more attentive while entering passwords in a lab study, which can contribute to less login error and higher login time than the casual login attempts in real-life scenario. Further long-term studies in a real life setting could provide additional insights on the training effect.
\section{Introduction}
Compared with the tremendous advancement of digital devices over the
past few decades, passwords are a stubborn legacy technology that hangs
on for want of something better.
Recall-based user-chosen textual password is the most widely used authentication scheme on the Web, but it is fraught with security problems because of password reuse~\cite{pwhabit} and predictable patterns~\cite{pwpattern1,pwpattern2} that make this scheme vulnerable to online guessing attacks~\cite{captcha,guessing08}. Password restriction policies are found to have limited impact on ensuring security and in some cases adversely affect password memorability~\cite
furnell07,pwpattern1}. System-assigned random passwords provide higher security, but suffer from memorability problems~\cite{yan_sec}. Multiple variants~\cite{passphrase,text_recog,persuation,forget_thesis} have been proposed to improve memorability in this respect, yet none of these schemes has shown significant improvement for the examined usability
metrics.
Users have cognitive limitations that define our potential for
interaction with computers. Existing password systems fail to fully
address these limitations, nor do they leverage humans' cognitive
strengths.
\subsection{Motivation}
The most prominent dilemma studied in the research on user
authentication is the delicate balance between memorability and
preventing successful guessing. Existing schemes rely on information
specifically memorized for the purpose of authentication, such as the
string of characters that make up a textual password. This means that
the information is less memorable for users than information that users
already know because it is meaningful for them. This in turn means that
users tend to choose passwords that are easy to guess.
Beyond guessing resistance, however, other security concerns have been
shown to be at least as important. Florencio et al. point
out~\cite{strongpw} that observation attacks, such as \textit{shoulder
surfing} (external observation attack) and \textit{keystroke loggers}
(internal observation attack~\cite{design_space}), appear to be the most
prevalent attacks, in which strong passwords are just as susceptible to
being stolen by an attacker as weak ones. They find that none of the
password ``best practices'' offers any real protection against these
attacks. This observation is also supported by the study of Florencio
and Herley~\cite{lockout}, who examined the password policies of $75$
different websites and found that their password policies address only
guessing attacks and, in few cases, passwords reuse across sites.
The susceptibility to shoulder surfing and keystroke loggers is
generally high when users log in from public
computers~\cite{keylogger}. A two-week-long field study~\cite{pw_diary}
on real life Web usage found that half of the participants used public
computers to access online accounts.
Thus, one of the most challenging tasks in user authentication today is to develop a text-based authentication scheme that satisfies the following requirements: i) Provides protection against observation attacks, ii) Offers good memorability, and iii) Provides sufficient entropy to prevent online brute-force attacks.
\subsection{Contributions}
In this paper, we explore the potential of {\em cognitive questions} in addressing these challenges by incorporating the scientific
understanding of long-term memory to advance the security and usability
properties that can be provided in an authentication system.
In particular, we propose {\em Q-A}, a new authentication scheme in which a user registers by selecting and answering six cognitive
questions (e.g., ``Who was your favorite high school teacher?'') from a
set of twenty questions. During login, the user is shown one question at
a time and asked to enter the letter at a random position in her
answer. This process is repeated for all six questions, and correctly
entering all six letters is required to authenticate. The contributions
of Q-A are given below (see~\S\ref{design} for a detailed discussion).
\textbf{Usability.} A survey study~\cite{survey} on $25$ different
password schemes found that memorability can be supported by leveraging
pre-existing user-specific knowledge, rather than requiring users to
memorize new information that is artificially constructed or
random. This finding inspires our application of cognitive questions in
Q-A, which gains usability advantages over other text-based password
schemes~\cite{mnemonic,persuation,text_recog,forget_thesis} in the
following way:
\begin{itemize}
\item Authentication secrets are easier to remember in Q-A, since it
asks users for already known information, while the other schemes
query for specifically memorized information.
\item Psychology studies~\cite{grecog, kintsch, tulving73} reveal that
it is difficult to remember information spontaneously without memory
cues, where cognitive questions work as cues to retrieve the
corresponding answers from a long-term
memory~\cite{qa04,forget_thesis}.
\item Our scheme allows users to enter case-insensitive letters for
authentication, while traditional textual passwords may contain
numbers, upper-case letters, and special characters, which may require
additional time and effort to enter, especially from portable devices
(e.g., cellular phones).
\end{itemize}
In our study, users had a $100$\% memorability rate after one week, which
was significantly higher than that for randomly assigned six-character
passwords consisting of only lowercase letters.
They reported high levels of satisfaction with the usability and
security of Q-A, despite a long average time to log in.
\textbf{Security.} To the best of our knowledge, Q-A is the first
text-based authentication scheme that addresses both internal and
external observation attacks. The security features of our system are as
follows:
\begin{itemize}
\item Typically, cognitive questions are prone to \textit{guessing by acquaintances attacks} that exploit the knowledge about personal information of a user~\cite{qa_just09}. The analysis of Just and Aspinall~\cite{qa_just09} shows that \textit{three} questions are sufficient to guarantee reasonable security for authentication based on cognitive questions. Our study on Q-A satisfied these security requirements, where a set of questions were carefully selected considering both usability and security metrics~\cite{no_sec,qa_just09}, and each user was asked to answer \textit{six} questions from this set to measure the usability of Q-A for a system with high security requirements.
\item Q-A offers the \textit{variant response} feature, where users' entries vary across different login sessions to gain resilience against observation attacks.
\item Although each security question offers relatively low entropy compared with a moderately strong password~\cite{qa_just09}, selecting random letters from the answers to six different questions makes for a string of characters that is partially random and therefore hard to guess~\cite{yan_sec}. The theoretical password space ($28$ bits) offered by Q-A is sufficient to prevent online brute-force attacks~\cite{lockout}.
\end{itemize}
With Q-A, we explore the idea of using information that
is meaningful to users and already held in long-term memory for a
primary authentication mechanism. By developing and evaluating the Q-A
design, we show that such a mechanism can be built with reasonable
usability and security properties. We hope that this inspires
researchers to further investigate ways to leverage such information for
better schemes.
\section{Acknowledgement}
\bibliographystyle{IEEEtran}
\section{Results}
In this section, we discuss the results of the user study described
in~\S\ref{study}. We label the login performance of participants in
session 1 and session 2 as \textit{Recall-1} and \textit{Recall-2},
respectively. Here, we tested the following hypotheses:
\subsection{Hypothesis 1}
\vspace{0.03cm}
\noindent $H1_a$: \textit{The login success rate for Q-A and control passwords would not significantly differ in Recall-1.}
\vspace{0.03cm}
\noindent $H1_b$: \textit{The login success rate for Q-A would be significantly higher than that of control passwords in Recall-2.}
\vspace{0.05cm}
In Q-A, users don't have to memorize any new authentication secrets, as
their password instead comes from cognitive questions related to their
real life. However, in the control condition, the user is required to
memorize a random string of characters as her password. In Recall-1,
users were asked to enter their control passwords within a short period
of learning it. So, we hypothesized that the login success rate for Q-A
and control passwords would not significantly differ in Recall-1, but
that Q-A would perform significantly better than control passwords in
Recall-2, in terms of login success rate.
We observed a $100$\% login success rate for Q-A in both Recall-1 and
Recall-2. In the control condition, login success rate was $91$\% in
Recall-1 and $77$\% in Recall-2. Whether or not a participant
successfully authenticated is a binary measure, so we use McNemar's
tests when we analyze the login success rate for our within-subjects experiment.
Our analysis shows that login success rate did not differ significantly between Q-A and control
conditions in Recall-1, $\mathcal{X}^{2}(1, N = 22) = 0.5$, $p =
0.24$. In Recall-2, however, the login success rate for Q-A was
significantly higher than the control passwords, $\mathcal{X}^{2}(1, N =
22) = 3.2$, $p<.05$. $H1_a$ and $H1_b$ are supported by these results.
\begin{figure}[t]
\centering
\includegraphics[width=85mm]{graphs/bar2}
\vskip -15 pt
\caption{Login success rate}
\label{fig:success}
\end{figure}
\subsection{Hypothesis 2}
\vspace{0.05cm}
\noindent $H2$: \textit{There would be a significant difference in login time between Q-A and control passwords in both Recall-1 and Recall-2.}
\vspace{0.05cm}
Since Q-A requires entering the letter of a random position for as many
as six answers, we hypothesized that for both Recall-1 and Recall-2, the
login time for the control passwords would be significantly less than
that of Q-A. Table~\ref{tab:time} summarizes the results for login time,
which shows that the mean login time in Q-A was $53.9$ seconds in Recall-1 (median: $51$ seconds) and $56.9$ seconds in Recall-2 (median: $53$ seconds). The mean login time ($38.1$ seconds in Recall-1, $43.7$ seconds in Recall-2) in the control condition seems to be affected by the outliers, where the median login time was $11.5$ seconds in Recall-1 and $13$ seconds in Recall-2.
We did not get matched pair of subjects while comparing the time for successful logins in our study conditions, since a number of participants who succeeded to log in using Q-A, failed in control condition. So, instead of Wilcoxon signed-rank test, we used a Wilcoxon-Mann-Whitney test to evaluate the difference in login time between Q-A and control passwords. Wilcoxon tests are similar to t-tests, but make no assumptions about the distributions of the compared samples, which is appropriate to the count data in these conditions. In Recall-1, we found a significant difference in login time ($W=72$, $p<0.01$). The difference was significant in Recall-2 as well ($W=350$, $p<0.05$). All these findings provide enough evidence to support $H2$.
\begin{table}[b]
\renewcommand{\arraystretch}{1.3}
\caption{Descriptive statistics for registration and login time (in seconds)}
\vspace{0.0cm}
\centering
\begin{tabular}{ccccc}
\textbf{Scheme}&Study&Mean&Median&SD\\
\hline
\multirow{3}{*}{\textbf{Q-A}}&Registration & $116.6$ & $103$ & $43$ \\
&Recall-1 & $53.9$ & $51$ & $17.8$\\
&Recall-2 & $56.9$ & $53$ & $20.5$\\
\rule{0pt}{4ex}\multirow{3}{*}{\textbf{Control}}&Registration & $48.1$ & $45.5$ & $27.1$ \\
&Recall-1 & $38.1$ & $11.5$ & $70.8$\\
&Recall-2 & $43.7$ & $13$ & $55$\\
\end{tabular}
\label{tab:time}
\end{table}
\subsection{Hypothesis 3}
\vspace{0.05cm}
\noindent $H3$: \textit{During login in Q-A, the mean time required to enter a letter of the given position would vary significantly for different positions in the answer in both Recall-1 and Recall-2.}
\vspace{0.05cm}
The cognitive effort required to fetch a letter from the third position
of an answer might be higher than entering a letter in the first
position. So, we hypothesized that the mean time required to enter a
letter would vary significantly for different positions in the answer.
We used a Kruskal-Wallis test for Hypothesis 3, which is similar to
ANOVA, but does not make any assumption about the distributions of the
compared samples, and is thus appropriate in this context. The
Kruskal-Wallis test demonstrated that for Recall-1, the mean time
required to enter a letter varied significantly for different positions
in the answers, $\mathcal{X}^{2}(4)= 15.32$, $p<0.01$. However, for
Recall-2, the differences were not statistically significant,
$\mathcal{X}^{2}(4)= 4.79$, $p=0.31$. So, $H3$ is partially supported by
our results.
\subsection{Notable Findings}
\vspace{0.03cm}
\subsubsection{Number of attempts}
A successful authentication in Q-A requires entering the correct letter
for six questions, but we didn't restrict the number of attempts a user
could make before entering her authentication secret correctly. It
allowed us tracking the number of mistakes a user committed before she
was able to log in successfully. Q-A showed quite promising results in
this respect.
In Recall-1, $82$\% participants succeeded on the first attempt to enter
the correct letter for all six questions. The other participants
made a mistake for only one question and entered the correct letter in
the second attempt. In Recall-2, $73$\% participants succeeded on the
first attempt for all six questions. Of the other six participants
($27$\%) who made a mistake in a question on the first attempt, four
succeeded in the second attempt and the other two required no more than
four attempts before entering the correct letter. One participant made
mistakes in two questions; no other participant made any mistakes in
more than one question in Recall-2.
For the control passwords in Recall-1, $82$\% of participants succeeded
to log in on the first attempt. One participant required $10$ attempts to
log in successfully, as he did not notice that \textit{Caps Lock} was
on. Q-A overcomes this limitation, since the users' responses are not
case sensitive. In Recall-2, $68$\% of participants entered their
control passwords correctly in the first attempt. The participants who
were successfully authenticated after multiple attempts in Recall-2
required four attempts on average with a maximum of eight attempts.
\subsubsection{Registration}
In Q-A, participants were able to complete registration in less than two minutes on average (see Table~\ref{tab:time}). The mean time to register in the control condition was $48.1$ seconds. A Wilcoxon signed-rank test (appropriate for matched pair of subjects) reveals that there existed a significant difference in registration time between Q-A and control conditions ($W=450.5$, $p<0.01$).
\begin{table}[t]
\renewcommand{\arraystretch}{1.3}
\caption{Questionnaire responses for the usability of Q-A. Scores are
out of 10. * indicates that the scale was reversed.} \centering
\begin{tabular}{lcc}
\multicolumn{1}{c}{Questions}&Mode&Median\\
\hline
Logging in using Q-A password was easy&$10$&$8.5$\\
Q-A passwords are easy to remember&$9$&$9$\\
With practice, I could quickly & \multirow{2}{*}{$10$}&\multirow{2}{*}{$9$}\\
enter my Q-A password &&\\
\rule{0pt}{4ex}*I found Q-A too time-consuming (i.e., I &\multirow{2}{*}{$4$}&\multirow{2}{*}{$4.5$}\\
found Q-A to not take too much time) & & \\
\rule{0pt}{4ex}*I prefer system-assigned textual &\multirow{3}{*}{$8$}&\multirow{3}{*}{$7$} \\
passwords to Q-A (i.e., I prefer Q-A to & & \\
system-assigned textual passwords) & &\\
\rule{0pt}{4ex}I could easily use Q-A every day&$10$&$8$\\
I could easily use Q-A every week&$10$&$9$\\
\end{tabular}
\label{tab:usability}
\end{table}
\begin{table}[b]
\renewcommand{\arraystretch}{1.3}
\caption{Questionnaire responses for the applicability of Q-A in
different online accounts. Scores are out of 10.}
\centering
\begin{tabular}{r@{\hskip 0.8cm}cc
Online accounts&Mode&Median\\
\hline
Bank&$9$&$9$\\
Webmail&$9$&$6$\\
Social Networking&$5$&$5$\\
University Portal&$6$&$6.5$\\
E-commerce&$10$&$8$\\
\end{tabular}
\label{tab:applicable}
\end{table}
\textbf{Correlations.} We were interested to see if the registration
time for a participant would have any correlation with her login time
and the required number of attempts for a successful authentication in
Q-A. Our results show that the registration time and login time in Q-A
were not strongly correlated in either Recall-1 ($r=0.25$) or Recall-2
($r=0.26$). Also, there existed no strong correlation between the
registration time and the required number of attempts in Q-A for a
successful login in either Recall-1 ($r=-.05$) or in Recall-2 ($r=.20$).
\textbf{Number of mistakes.} During registration in Q-A, answers are
shown as asterisks or dots to reduce the risk of shoulder surfing, and
the users have to re-enter an answer to confirm. Our results show that
users did not make any mistake when confirming the answers. We have
deployed some basic restrictions in our system to guard against poor
answers (e.g., minimum three characters in an answer, no repeat answers
between questions, at least two different letters in an answer). In this
respect, six participants attempted to enter an identical answer for
multiple questions, and when they saw the error message they entered
distinct answers for each question.
\subsection{User Opinion and Perception}
In order to gain an understanding of users' perceptions on the usability and applicability of Q-A, we asked them to answer two sets of Likert scale questions at the end of the second session. We used
ten-point Likert scales, where anchors were included on the bi-polar ends of the scale ($1$ indicating \textit{strong disagreement} and $10$ equalling \textit{strong agreement} with the given statement). We
reversed some of the questions to avoid bias; thus the scores marked with (*) were reversed before calculating the mode and median. A higher score always indicates a more positive result for Q-A. Since Likert scale data are ordinal, it is most appropriate to calculate mode and median for Likert-scale responses~\cite{mean}.
\textbf{Usability.} The perceptions of users on the usability of Q-A are illustrated in Table~\ref{tab:usability}. Users showed a high degree of satisfaction on the usability (e.g., memorability, ease of login) of our scheme, and they preferred Q-A over system-assigned textual passwords. The participants responded positively (i.e., mode and median were higher than neutral) about the ease of using Q-A either weekly or daily. Although they expressed concerns regarding the authentication time in Q-A, they reported that with practice they could log in quickly with this scheme.
\begin{table}[!b]
\renewcommand{\arraystretch}{1.3}
\caption{Questionnaire responses to the requirement of recording
passwords in real life}
\centering
\begin{tabular}{rccccc
\multicolumn{1}{c}{\bf Scheme\hspace{0.3cm}}&Never&Rarely&Sometimes&Often&Always\\
\hline
{\bf Q-A}\hspace{0.3cm} &$55$\% & $36$\% & $9$\% & $0$\% & $0$\%\\
{\bf Control}\hspace{0.3cm} &$0$\% & $14$\% & $46$\% & $36$\% & $4$\%\\
\end{tabular}
\label{tab:record}
\end{table}
\textbf{Applicability.} Table~\ref{tab:applicable} shows users' perceptions on the applicability of Q-A in different types of online accounts. Our findings suggest that most of the participants would strongly prefer to use Q-A for online accounts with high security requirements, such as banking and e-commerce accounts.
The participants were given an open-ended question at the end of second session to express their opinion about Q-A. In general, the feedbacks were positive and encouraging. They expressed high degree of satisfaction on the security features of Q-A. For example, one participant reported, ``I would use it for banks and other websites, where I use confidential info.'' Several of them were interested to know if Q-A would be deployed commercially.
\textbf{Password recording.} All the participants reported that they did not write down their control nor Q-A passwords for this study. We also asked them if they would require writing down their Q-A or control passwords if they would use them in real life. The results (see Table~\ref{tab:record}) are quite promising in this respect: $55$\% of participants reported that they would \textit{never} need to record their Q-A password, while $36$\% of participants would \textit{rarely} need to. Only two ($9$\%) of the participants mentioned that they would \textit{sometimes} require to write down their Q-A password.
For control passwords, $46$\% of participants would \textit{sometimes} require and $36$\% of participants would \textit{often} need to record their password.
\section{The Q-A Design and Scientific Motivations}\label{design}
\begin{figure}[t]
\centering
\includegraphics[width=70mm, height=25mm]{graphs/ques4}
\vskip -15 pt
\caption{A screenshot of a demo of the Q-A.}
\label{fig:ques}
\end{figure}
In this section, we first explain the basic design of Q-A. We then describe our approach to address the usability and security concerns with cognitive questions, followed by a detailed discussion on the usability and security features offered by Q-A.
Q-A is based on cognitive questions (e.g., ``What is the name of your favorite childhood teacher?''), which inherently leverage existing long-term memories, information that is known to users based on their
life experience. Q-A invokes the answers to cognitive questions in a novel way. At the time of registration, the user is shown a set of $20$ carefully selected questions, of which she must select six
questions to answer. These answers, in total, constitute her authentication secret. We choose the $20$ questions carefully to ensure that each one asks for alphabetical answers with a high amount of entropy, such as the name of a person or a location.
During login, a user is presented with all six selected questions, one at a time. She does not have to enter the whole answers. Instead, for each question, the user is asked to enter a single letter from a
particular position in her answer. Every time a user logs in, this position is randomly chosen by the system individually for each question. So, in a login session, a user may be asked to enter the letter in the \textit{second position} for the first question, in the \textit{fourth position} for the second question, and so on. These positions will vary in the subsequent login sessions. For example, if a user's answer is ``Anderson", and at a given login session she is asked to enter the second letter of this answer, she has to enter `n'. In this way, a user has to correctly enter a letter for all six questions for a successful login.
\subsection{Usability and Security of Cognitive Questions}
In this section, we describe our approach in Q-A to address the usability and security concerns with cognitive questions.
\subsubsection{Usability}
Just and Aspinall defined three metrics to measure the usability of a
cognitive question: i) Applicability, ii) Memorability, and iii)
Repeatability~\cite{qa_just09}. This provides a useful guide to
selecting appropriate questions for Q-A. We carefully considered these
metrics while selecting questions for our study.
\textbf{Applicability.} Not every user can effectively answer every
cognitive question. Users in our system choose any six questions from a
set of twenty questions that they find most applicable to them. Offering
a greater number of questions would further increase the applicability
of the system.
\textbf{Memorability.} Research shows that a user can easily recall the
answers of cognitive questions that are related to her long-term
memory~\cite{no_sec,qa_just09}. Selecting such questions ensures that
the user does not need to devote much cognitive effort in learning the
Q-A passwords, as they are simply answers that she already knows.
A study by Schechter et al. reveals that the participants who used weak
placeholder answers during the study failed to recall them
later~\cite{no_sec}. So, with a few basic restrictions
in place to guard against poor answers (see~\S\ref{sec_adv}), it should
typically be easier for a user to remember the real answer to a question
than a weak placeholder answer (`Aab').
The study of Furnell et al.~\cite{qa04} shows that confusion between
capital and lowercase letters is a prominent reason for making mistakes
when answering a cognitive question. We address this by ignoring case in
our scheme.
\textbf{Repeatability.} Just and Aspinall indicate that repeatability
can be improved by providing users with the fixed format of answers to
the questions that ask about dates or locations (e.g., different formats
for date: `Feb-05, 1992', `02-05-1992',
`02/05/1992')~\cite{qa_just09}.
To provide higher entropy, we recommend to avoid questions that ask for
numerical answers, and this includes questions about specific dates. For
location-related questions, instead of imposing any specific format on
users, we prefer to be more specific with the questions to ensure
repeatability. For example: instead of asking ``Where did your father
and mother meet?'', we would ask ``In what city or town did your father
and mother meet?'' We recommend providing a fixed format for the answer
to a question in which being specific with the question does not resolve
the repeatability issues.
\subsubsection{Security}
The most important security concerns with cognitive questions include:
i) User-created questions, ii) Guessing by acquaintances attacks, and
iii) Common answers. We address these security issues in the following
way.
\textbf{User-created questions.} If users are allowed to freely create their own questions, many users will not choose sufficiently secure questions~\cite{qa_just09}. On the other hand, it may have usability concerns for many users if the questions are strictly assigned by the system. Q-A balances these trade-offs by asking users to select any six questions from a larger set.
\textbf{Guessing by acquaintances.} Typically, cognitive questions are prone to targeted guessing attacks, in which attackers exploit the knowledge about personal information of a user~\cite{qa_just09,no_sec}. One's mother's maiden name and Social Security Number (SSN) in the US are particularly well-known examples of such questions. It is possible to select suitable questions, but still the amount of entropy in a single answer is typically lower than for a password~\cite{no_sec,qa_just09,qa_fb}.
Just and Aspinall~\cite{qa_just09} show that {\em three} questions are sufficient to guarantee reasonable security for cognitive-question-based authentication. In our study, we asked users to answer {\em six} questions to measure the usability of Q-A for a system with high security
requirements.
\textbf{Common answers.} The answers to some cognitive questions are generally common among users. For example, \textit{blue} or \textit{pink} may be common answers to the question ``What's your favorite color?". The prior studies~\cite{no_sec,qa_just09} have found that carefully selected questions can make common answers less of an issue for most users. The questions for Q-A were selected carefully based on the prior usability and security analysis on cognitive questions~\cite{no_sec,qa_just09,qa04,qa_fb}.
\subsection{Usability Features and Memory Retrieval in Q-A}\label{use_adv}
In this section, we state the usability advantages of Q-A from the perspective of cognitive psychology. Q-A offers two important usability advantages over other text-based authentication systems~\cite{mnemonic,persuation,text_recog,forget_thesis}:
\begin{itemize}
\item \textbf{Known Information:} Q-A queries for already known
information, while the other schemes query for specifically memorized
information.
\item \textbf{Memory Cues:} Q-A provides users with the questions that
work as cues to retrieve the corresponding answers from user's memory,
so as to log in successfully.
\end{itemize}
\textbf{Known Information.} From the viewpoint of cognitive
psychology~\cite{episodic72,encoding}, Q-A is closely tied with the
concept of \textit{episodic memory}~\cite{episodic72,encoding}, which
refers to the autobiographical event that the user can accurately recall,
since she was part of it. Episodic memory incorporates the time and place of a personally meaningful event with the associated feeling and contextual information. For example,
recalling the celebration of new millennium's eve involves figuratively
traveling back in time to precisely remember the place and the people
associated with that event. Thus, cognitive-question-based
authentication systems like Q-A aid password memorability, since a user
does not need to memorize the answers specifically for authentication to
an online account.
\textbf{Memory Cues.} Psychology research has shown that it is difficult
to remember information spontaneously without memory cues~\cite{grecog,
kintsch, tulving73} . This suggests that authentication schemes should
provide users with cues to aid memory retrieval. Ellis and Kvavilashvili
state that memory cues support \textit{prospective
memory}~\cite{memory09}, which is the ability to generate, retain, and
later recall information in the appropriate
context. \textit{Encoding specificity theory}~\cite{encoding} postulates that the most effective cues are those that are present at the time of remembering. If semantic information about a
word is processed at the time of learning, then that information can
successfully be used to cue memory. Thus, the word `Millennium Eve' can
only be used to cue memory of the word `New York' if the subject encodes
the semantic information linking the two objects at the time of
encoding.
\textit{Generate-recognize theory}~\cite{grecog} and
\textit{Associative-strength theory}~\cite{ellis89} also focus on the
effectiveness of cues in aiding memorability~\cite{kintsch,
tulving73}. Generate-recognize theory~\cite{grecog}
speculates that retrieval is a two-step process, where in the generate
phase, a list of candidate words is formed by searching the long-term
memory. Then in the recognize phase, the list of words is evaluated to
see if they can be recognized as the sought out memory. According to
this theory, a cue can help not only in generating a relevant candidate
list, but also in recognizing the appropriate word from that
list. Associative-strength theory~\cite{ellis89} states that a cue
becomes effective if it has previously occurred with the remembered
event in the past. The theory assumes that memory is structured as a
network that connects all items in memory, and items in memory with
stronger ties between them make better cues.
\subsection{Security Features in Q-A}\label{sec_adv}
In this section, we describe the security features offered by Q-A.
\subsubsection{Guessing resistance} The theoretical password space ($28$ bits) offered by Q-A is sufficient to prevent online brute-force attacks~\cite{lockout}.
\textbf{Theoretical password space.} The questions in our scheme ask users for alphabetical answers (e.g., ``what was the name of your first teacher?"), where the answers are case-insensitive. During login, the user is asked to enter the character at a given position in her answer. In this respect, the size of the domain for an alphabetical entry ($26$) is larger than a numerical entry ($10$). Thus, alphabetical answers provide higher resilience against online brute-force attacks~\cite{guessing07,guessing08}, in which an attacker tries all elements within a search space to obtain the password. The user has to enter a letter for each of six questions. The space for this condition will therefore be $log_{2} {(26)}^6\approx 28$ bits. Florencio and Herley's study on security policies~\cite{lockout} suggests that $20$ bits of theoretical password space suffices for an online environment with lockout rules.
\textbf{Effective password space.} Both in Q-A and in the study of password policies, only theoretical entropy is considered, though letters from random positions in the answers to cognitive questions may also have more effective entropy than user-selected passwords prone to dictionary attacks. In Q-A, the effective password space may vary for different questions, which requires to consider variation in answers in addition to general letter frequency. For example, the letter frequencies of names of people and locations will vary significantly by country and region. It would be an interesting area for future work to analyze the effective entropy of Q-A considering the letter frequency found in answers to different questions and the variation in answers in different languages, ethnicities, and countries.
\textbf{Guard against poor answers.} We have deployed some basic restrictions in our system to guard against poor answers (e.g. minimum three characters in an answer, no repeat answers between questions, at
least two different letters in an answer), which might motivate users to give their real answers that, with the right questions, should have a large space of both possible and probable answers.
\subsubsection{Observation attacks}
Having authentication information vary across login sessions is known as {\em variant response}~\cite{survey}. During login, the variant response feature makes our system more resilient against observation attacks like shoulder surfing, as compared to other text-based password schemes where the same set of characters is entered as password in every login session.
\textbf{Shoulder surfing.} When the user enters her credentials, either at registration or login, the answers are shown as asterisks or dots (as with regular password entry) to minimize the risk of shoulder surfing. During login, a user enters the letter at a given position in her answer, and to learn that letter and its position in the answer, the shoulder surfer needs to observe both the monitor and keyboard at a time, which has been found to be difficult in practice~\cite{shoulder06}. The requirements to observe the letters for all six questions further increase the hurdles for an attacker.
Even if a shoulder surfer can learn the letter and its position, he is likely to be asked to enter a letter of different position when he tries to log in as the user. Only a good guess of the entire answer for all six questions gives the attacker a reasonable chance of logging in.
The shoulder surfer may attempt to gain the user's credentials when she enters the entire answer to a question at the time of registration. Although answers are shown as asterisks or dots (as with regular password entry) to reduce the risk of shoulder surfing, we recommend that the users register in a secure environment (e.g., avoiding public terminals) to ensure maximum security.
\textbf{Keystroke and mouse loggers.} Gaining a user's authentication credentials through malware, such as keystroke loggers and mouse loggers, is called an {internal observation attack}~\cite{design_space}. A system provides resilience against keystroke/mouse loggers when the keyboard/mouse entries for authentication vary across subsequent login sessions~\cite{survey}. Thus, the variant response feature in Q-A offers higher resilience against keystroke loggers as compared to a password system where the same set of letters is entered (using keyboard) during each login session. Our system is clearly resilient to mouse loggers, as we don't use mouse input.
The security of Q-A can be further improved by asking users to answer more than six questions during registration and then drawing six at random for each authentication attempt. In this case, when an attacker tries to log in as a user, he may get different questions than the ones he gains through observation attacks. We will explore the usability of this approach in future work.
\subsubsection{Social engineering attacks}
Social engineering refers to the psychological manipulation of people so that they divulge confidential
information~\cite{survey}. Phishing~\cite{survey} is a common form of social engineering attack.
\textbf{Phishing.} In a phishing attack, users are redirected to fraudulent websites to enter their credentials. In this case, the phishing victim will very likely get different questions from the ones she normally uses to log in, since a phisher would not typically have access to the user's login sessions. This means that not only will the user enter information that is useless to the attacker, she may realize that something is wrong and end the session.
\textbf{Fake calls.} Users may be tricked to reveal credentials by any means, e.g., phone calls from a fake help desk or credit company~\cite{survey}. Q-A does not provide direct resilience to social engineering when the users disclose their authentication secrets to the attacker. However, a site with high security requirements can use unique questions to avoid question and answer reuse and overlap between sites so that attackers cannot gain access to other accounts by acquiring password for a single account. We propose to explore this issue further in future work.
\section{Study Design}\label{study}
Two prior text-based authentication systems~\cite{passphrase,text_recog} aim to provide a usable solution to the security issues with passwords. Both papers compared their scheme with system-assigned random textual passwords (\textit{control passwords}), where the results did not show significant improvement over control passwords in terms of memorability. We compare Q-A with control passwords and keep the theoretical password space the same ($28$ bits) for both conditions to examine whether Q-A offers better memorability in this respect.
In this study, we used a within-subjects design consisting of two experimental conditions (e.g., Q-A and control passwords). Using a within-subjects design controls for individual differences, and permitted the use of statistically stronger hypothesis tests. The experiments performed as part of this research received approval from our university's Institutional Review Board (IRB) for human subject research.
\subsection{Participants, Apparatus and Environment}
For this experiment, we recruited $22$ university students ($6$ women, $16$ men) from diverse backgrounds, including majors from Biology, Engineering, Interdisciplinary Study, etc. The mean age of the participants was $27$. They make regular use of the Internet and websites that require authentication. Each participant was compensated with a \$10 gift card for participating in this study.
To administer this experiment, we created two realistic and distinct websites outfitted with the password scheme according to our two conditions: Q-A and control passwords. For both conditions, we used banking sites, but designed the interfaces with a different look and feel. Upon successful login at each bank site, the participants were forwarded to a dummy account overview page, to give them the feeling of
using online banking services.
We carefully reviewed the prior usability and security studies on cognitive questions~\cite{no_sec,qa_just09,qa04,qa_fb} to select the $20$ questions for this study. The lab studies were conducted with one participant at a time to allow the researchers to observe the user's interaction with the system.
\subsection{Procedure}
We conducted the experiment in two sessions, each lasting around $30$ minutes. The second session took place one week after the first one to test memorization of the password. Note that the one-week delay reflects a common interval used in authentication studies (e.g.,~\cite{face_age,text_recog,bdas}). A field study~\cite{pw_diary} on real-life web usage found that one week is larger than the maximum average interval for a user between her subsequent logins to any of her important accounts~\cite{pw_diary}. Thus, we used this interval to examine the usability of our scheme.
\subsubsection{Session 1}
Before starting the experiment, we provided a consent form for the participants
to read and sign, if they agreed. To compensate for the novelty effect, we asked the participants to perform one practice trial for authentication with Q-A. We did not collect data for this
practice trial. The subsequent steps of the experiment are as follows:
\textbf{Sign-up.} During sign-up (we alternatively use the term {\em registration}) with Q-A, participants were shown a set of twenty questions, and they selected any six questions to answer, which constituted their password. Participants were also asked to sign-up with another site, where they were assigned a random six character password (control password).
To control for order effects, we employed counter balancing, so that all the participants would not see the schemes in the same order. Half of the participants used Q-A first, and the other half used the control password first.
\textbf{Distraction.} After sign-up, the participants were asked to count down in threes from a randomly chosen four-digit number for $45$ seconds. This type of distraction flushes the textual working memory~\cite{short_mem} and simulates a longer passage of time by focusing their attention on a separate, cognitively-difficult task. Participants were then given questionnaire that gathered demographic information.
\textbf{Login (Recall-1).} The participants were asked to log into each of those same sites, to demonstrate that the passwords had been memorized. Participants who were unable to reproduce their passwords during login, were shown the passwords.
\subsubsection{Session 2}
The second appointment took place one week after the first one. The participants were asked to log into each of the two sites (\textit{Recall-2}). After they had finished, an anonymous paper-based survey was conducted to get their opinion on the overall experiences of using the authentication schemes. Participants were then compensated and thanked for their time.
|
\section*{Supplementary information}
In the main paper we use a single set of simulation parameters, namely $S_{\mathrm{nem}}=1$, $A=0.08$, $C=0.5$, $L=0.005$, $K=0.01$, $\Gamma=0.1$, $M=0.1$, $\eta=1/6$, $b=0.1$, $\zeta=0.3$, $\rho=40$ and $\chi=0.005$. Here we consider the effect of changing some of these parameters. The first variation we consider is a five-fold increase in $A$ and $K$ to $A=0.4$ and $K=0.05$. This has the effect of increasing the interfacial tension between the two fluid phases while leaving the interfacial width approximately unchanged. The second variation is a ten-fold decrease in the mobility parameters, such that $\Gamma=0.01$ and $M=0.01$. This slows down the rate of approach to equilibrium, and hence weakens the role of thermodynamics compared to flow. And third, we consider removing the substrate friction by setting $b=0$.\\
In figure~\ref{fig:SIdrop}, we compare, for the different parameters, the evolution of an initially circular drop, as shown in Fig.~1 of the main paper. Although there is variation in the exact drop shape and the rate at which it evolves, all four parameter sets show the same qualitative stages of behaviour: elongation, active anchoring, and the eventual undulatory instability.\\
In figure~\ref{fig:SIband}, we compare the evolution of an initially straight stripe, as investigated in Fig.~3 of the main paper. Again, the four situations differ in the details; e.g. in the high surface tension case the amplitude of the interface undulations is smaller and the point defects do not have an isotropic core, while the frictionless case evolves more quickly. But notwithstanding these differences, the same qualitative stages are observed. The stripe undergoes a waving instability, in which the concave parts of the interface form sharp cusps while the convex parts gain only a small curvature. Finally, singular point defects, of charge +1/2, are emitted from the interface into the bulk nematic.
\begin{figure}
\includegraphics[width=\textwidth]{SIdrop.png}
\caption{Evolution in time of an initially circular drop of extensile active fluid for the various parameter sets.}
\label{fig:SIdrop}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{SIband.png}
\caption{Evolution in time of a band of extensile active fluid for the various parameter sets.}
\label{fig:SIband}
\end{figure}
\end{document}
|
\section{Introduction}
Let $K$ be a field with a discrete valuation $\mathfrak v$ and ring of integers $\mathcal O_K$ and $C$ an irreducible, smooth, algebraic curve of genus $g\geq 1$ defined over $K$ and function field $K (C)$. The discriminant $\mathfrak D_{C/K}$ is an important invariant of the function field of the curve and therefore of the curve. Since the discriminant is a polynomial given in terms of the coefficients of the curve, then it is an ideal in the ring of integers $\mathcal O_K$ of $K$. The valuation of this ideal is a positive integer.
A classical question is to find an equation of the curve such that this valuation is minimal, in other words the discriminant is minimal.
When $g=1$, so that $C$ is an elliptic curve, there is an extensive theory of the minimal discriminant ideal $\mathfrak D_{C/K}$. Tate \cite{ta-75} devised an algorithm how to determine the Weierstrass equation of an elliptic curve with minimal discriminant as part of his larger project of determining Neron models for elliptic curves. The main focus of this paper is to extend their work to superelliptic curves, full details and proofs are intended in \cite{rachel}.
The paper is organized as follows. In section 2 we give the basic definitions for genus $g\geq 2$ superelliptic curves isomorphism classes of which correspond to projectively equivalent classes of degree $d$ binary forms. For a binary form $f(X, Z)$ and a matrix
$M = \begin{bmatrix} a, b \\ c, d \end{bmatrix}$, such that $M \in GL_2 (k)$, we have that $f^M:= f(aX + bZ, cX + dZ)$ has discriminant
$\Delta (f^M) = ( \det M )^{d(d-1)} \cdot \Delta (f)$.
This property of the discriminant is crucial in our algorithm which is explained in Section~5.
In section 3 we define the discriminant of a genus $g \geq 2$ superelliptic curve $\mathcal X_g$ defined over an algebraically number field $K$. We follow the classical theory and define the discriminant for local fields and then generalize it to global fields.
In Section 4, we summarize briefly Tate's algorithm and a modified version of it by Laska \cite{la-82}. Since the case of the elliptic curves is the simplest case this hopefully gives the reader an idea of how things work out in higher genus.
In Section 5, we generalize the algorithm to all superelliptic curves. This algorithm computes a Weierstrass equation with minimal discriminant for all superelliptic curves. Details and proofs are intended to be described in \cite{rachel}.
\section{Preliminaries}
Let $\mathcal X_g$ be a superelliptic curve of genus $g\geq 2$ with affine equation
\begin{equation}\label{w-eq-super}
y^n = f(x, 1) = a_d x^d + \cdots a_1 x + a_0
\end{equation}
defined over and algebraic number field $K$. Obviously the set of roots of $f(x)$ does not determine uniquely the isomorphism class of $\mathcal X_g$ since every coordinate change in $x$ would change the set of these roots. Such isomorphism classes are classified by the invariants of binary forms.
For any algebraically closed field $k$ let $k[X,Z]$ be the polynomial ring in two variables and let $V_d$ denote the $(d+1)$-dimensional subspace of $k[X,Z]$ consisting of homogeneous polynomials
\begin{equation} \label{eq1}
f(X,Z) = a_0X^d + a_1X^{d-1}Z + ... + a_dZ^d
\end{equation}
of degree $d$. Elements in $V_d$ are called \textit{binary forms} of degree $d$. $GL_2(k)$ act as a group of automorphisms on $ k[X, Z] $ as follows:
\begin{equation}
M =
\begin{pmatrix} a &b \\ c & d
\end{pmatrix}
\in GL_2(k), \textit{ then }
\quad M \begin{pmatrix} X\\ Z \end{pmatrix} =
\begin{pmatrix} aX+bZ\\ cX+dZ \end{pmatrix}
\end{equation}
Denote by $f^M$ the binary form $f^M (X, Z) := f(aX+bZ, cX+dZ)$.
It is well known that $SL_2(k)$ leaves a bilinear form (unique up to scalar multiples) on $V_d$ invariant.
Consider $a_0$, $a_1$, ... , $a_d$ as parameters (coordinate functions on $V_d$). Then the coordinate ring of $V_d$ can be identified with $ k[a_0 , ... , a_d] $. For $I \in k[a_0, ... , a_d]$ and $M \in GL_2(k)$, define $I^M \in k[a_0, \dots , a_d]$ as follows
\begin{equation} \label{eq_I}
{I^M}(f):= I( f^M)
\end{equation}
for all $f \in V_d$. Then $I^{MN} = (I^{M})^{N}$ and Eq.~(\ref{eq_I}) defines an action of $GL_2(k)$ on $k[a_0, \dots , a_d]$.
A homogeneous polynomial $I\in k[a_0, \dots , a_d, X, Z]$ is called a \textbf{covariant} of index $s$ if $I^M(f)=\delta^s I(f)$,
where $\delta =\det(M)$. The homogeneous degree in $a_0, \dots , a_d$ is called the \textbf{ degree} of $I$, and the homogeneous degree in $X, Z$ is called the \textbf{ order} of $I$. A covariant of order zero is called \textbf{invariant}. An invariant is a $SL_2(k)$-invariant on $V_d$.
Let $f(X, Z)$ and $g(X, Z)$ be binary forms of degree $n$ and $m$ respectively with coefficients in $k$. We denote the \textbf{r-transvection} of two binary forms $f$ and $g$ by $(f,g)^r$.
It is a homogeneous polynomial in $k[X, Z]$ and therefore a covariant of order $m+n-2r$ and degree 2.
A very important invariant is the discriminant of the binary form. In the classical way, the discriminant is defined as
$ \Delta = \prod_{i \neq j } (\alpha_i - \alpha_j)^2$,
where $\alpha_1, \dots \alpha_d$ are the roots of $f(x, 1)$. It is a well-known result that it can be expressed in terms of the transvectians. For example, for binary sectics we have $\Delta = J_{10}$ and for binary octavics $\Delta (f) = J_{14}$; see \cite{rachel} for details.
\begin{lem}
i) The discriminant of a degree $d$ binary form $f(X, Z)\in k[X, Z]$ is and $SL_2 (k)$-invariant of degree $2d-2 $.
ii) For any $M \in GL_2 (k)$ and any degree $d$ binary form $f$ we have that
\[ \Delta (f^M) = \left( \det M \right)^{d (d-1) } \, \Delta (f) \]
\end{lem}
\section{Discriminant of a curve}
The concept of a minimal discriminant for elliptic curves was defined by Tate and others in the 1970-s; see \cite{ta-75}. Such definitions and results we generalized by Lockhart in \cite{lockhart} for hyperelliptic curves. In this section we briefly generalize the concept of the minimal discriminant to all superelliptic curves.
Let $K$ be a local field, complete with respect to a valuation $\mathfrak v$. Let $\mathcal O_K$ be the ring of integers of $K$, in other words
$\mathcal O_K = \{ x\in K \, | \, \mathfrak v (x) \geq 0\}$.
We denote by $\mathcal O_K^\ast$ the group of units of $\mathcal O_K$ and by $\m$ the maximal ideal of $\mathcal O_K$. Let $\pi$ be a generator for $\m$ and $k=\mathcal O_K / \m$ the residue field. We assume that $k$ is perfect and denote its algebraic closure by $\bar k$.
Let $\mathcal X_g$ be a superelliptic curve of genus $g \geq 2$ defined over $K$ and $P$ a $K$-rational point on $\mathcal X_g$. By a suitable change of coordinates we can assume that all coefficients of $\mathcal X_g$ are in $\mathcal O_K$. Then, the discriminant $\Delta \in \mathcal O_K$. In this case we say that the equation of $\mathcal X_g$ is \textbf{integral}.
An equation for $\mathcal X_g$ is said to be a \textbf{minimal equation} if it is integral and $\mathfrak v (\Delta)$ is minimal among all integral equations of $\mathcal X_g$. The ideal $I=\m^{\mathfrak v (\Delta)}$ is called the \textbf{minimal discriminant} of $\mathcal X_g$.
\newcommand\fa{\mathfrak a}
Let us assume now that $K$ is an algebraic number field with field of integers $\mathcal O_K$. Let $M_K$ be the set of all inequivalent absolute values on $K$ and $M_K^0$ the set of all non-archimedean absolute values in $M_K$.
We denote by $K_\mathfrak v$ te completion of $K$ for each $\mathfrak v \in M_K^0$ and by $\mathcal O_\mathfrak v$ the valuation ring in $K_\mathfrak v$. Let $\mathfrak p_v$ be the prime ideal in $\mathcal O_K$ and $\m_v$ the corresponding maximal ideal in $K_\mathfrak v$. Let $(\mathcal X, P)$ be a superelliptic curve of genus $g\geq 2$ over $K$.
If $\mathfrak v \in M_K^0$ we say that $\mathcal X$ is \textbf{integral at $\mathfrak v$} if $\mathcal X$ is integral when viewed as a curve over $K_\mathfrak v$. We say that $\mathcal X$ is \textbf{minimal at $\mathfrak v$} when it is minimal over $K_\mathfrak v$.
An equation of $\mathcal X$ over $K$ is called \textbf{integral} (resp. \textbf{minimal}) over $K$ if it is integral (resp. minimal) over $K_\mathfrak v$, for each $\mathfrak v \in M_K^0$.
Next we will define the minimal discriminant over $K$ to be the product of all the local minimal discriminants. For each $\mathfrak v \in M_K^0$ we denote by $\Delta_\mathfrak v$ the minimal discriminant for $(\mathcal X, P)$ over $K_\mathfrak v$. The \textbf{minimal discriminant} of $(\mathcal X, P)$ over $K$ is the ideal
\[ \Delta_{\mathcal X / K} = \prod_{\mathfrak v \in M_K^0} \m_\mathfrak v^{\mathfrak v (\Delta_\mathfrak v) } \]
We denote by $\fa_\mathcal X$ the ideal $ \fa_\mathcal X = \prod_{\mathfrak v \in M_K^0} \mathfrak p_\mathfrak v^{\mathfrak v (\Delta_\mathfrak v) }$.
In \cite{rachel} we prove that
\begin{thm}
Let $(\mathcal X_g, P)$ be a superelliptic curve over $\mathbb Q$. Then its global minimal discriminant $\Delta\in \mathbb Z$ is unique (up to multiplication by a unit). There exists a minimal Weierstrass equation corresponding to this $\Delta$.
\end{thm}
Next we briefly describe how this minimal Weierstrass equation is determined for superelliptic curves. Full details and further analysis of discriminants of superelliptic curves is intended in \cite{rachel}.
\section{Elliptic curves and Tate's algorithm}
Let $E$ be an elliptic curve defined over a number field $K$ with equation
\begin{equation}\label{w-eq-ell} y^2 + a_1 xy + a_3 y = x^3 + a_2 x^2 + a_4 x +a_6. \end{equation}
For simplicity we assume that $E$ is defined over $\mathbb Q$, the algorithm works exactly the same for any algebraic number field $K$.
We would like to find an equation
\begin{equation}\label{w-eq-2} y^2 + a_1^\prime xy + a_3^\prime y = x^3 + a_2^\prime x^2 + a_4^\prime x +a_6^\prime. \end{equation}
such that the discriminant $\Delta^\prime$ of the curve in Eq.~\eqref{w-eq-2} is minimal. Since we want the new equation to have integer coefficients then the only transformations we can have are
\[ x = u^2 x^\prime + r, \qquad y = u^3 y^\prime + u^2 s x^\prime + t\]
for $u, r, s, t \in \mathbb Z$ and $u \neq 0$. The coefficients of the two equations are related as follows:
\[
\begin{split}
& ua_1^\prime = a_1 + 2s, \\
& u^3 a_3^\prime = a_3 + r a_1 + 2 t, \\
& u^2 a_2^\prime = a_2 - sa_1 + 3 r - s^2, \\
\end{split}
\qquad
\begin{split}
& u^4 a_4^\prime = a_4 - s a_3 + 2 r a_2 - (t+rs)a_1 + 3 r^2 - 2st \\
& u^6 a_6^\prime = a_6 + r a_4 + r^2 a_2 + r^3- ta_3 - rta_1 - t^2 \\
& u^{12} \Delta^\prime = \Delta \\
\end{split}
\]
The version of the algorithm below is due to M. Laska; see \cite{la-82}. \\
\noindent \textsc{Step 1:} Compute the following
\[
\begin{split}
c_4 & = (a_1^2+4a_2)^2 - 24(a_1a_3 +2a_4), \\
c_6 & = - (a_1^2+4a_2)^3 + 36(a_1^2 + 4 a_2)(a_1a_3 + 2 a_4) - 216 (a_3^2 +4a_6)
\end{split}
\]
\noindent \textsc{Step 2:} Determine the set $S$ of integers $u \in \mathbb Z $ such that there exist $x_u$, $y_u \in \mathbb Z$ such that
$ u^4 = x_u c_4$ and $ u^6 y_u = c_6$. Notice that $S$ is a finite set. \\
\noindent \textsc{Step 3:} Choose the largest $u \in S$, say $u_0$ and factor it as $u_0 = 2^{e_2} \, 3^{e_3} \, v$, where $v$ is relatively prime to 6. \\
\noindent \textsc{Step 4:} Choose
\[
a_1^\prime, a_3^\prime \in \left\{ \sum_{i=1}^n \alpha_i w_i \, | \, \alpha_i = 0 \, \textit{ or } 1 \, \right\} \, \textit{ and } \,
a_2^\prime \in \left\{ \sum_{i=1}^n \alpha_i w_i \, | \, \alpha_i = -1, 0 \, \textit{ or } 1 \, \right\}
\]
subject to the following conditions:
\[ (a_1^\prime)^4 \equiv x_u \mod 8, \quad (a_2^\prime)^3 \equiv - (a_1^\prime)^6 - y_u \mod 3. \]
\noindent \textsc{Step 5:} Solve the following equations for $a_4^\prime$ and $a_6^\prime$
\[
\begin{split}
x_u & = ({a_1^\prime}^2+4{a_2^\prime})^2 - 24({a_1^\prime}{a_3^\prime} +2{a_4^\prime}), \\
y_u & = - ({a_1^\prime}^2+4{a_2^\prime})^3 + 36({a_1^\prime}^2 + 4 {a_2^\prime})({a_1^\prime}{a_3^\prime} + 2 {a_4^\prime}) - 216 ({a_3^\prime}^2 +4{a_6^\prime})
\end{split}
\]
\noindent \textsc{Step 6:} Solve the equations for $s, r, t$ successively
\[ u a_1^\prime = a_1 + 2s, \quad u^2 a_2^\prime = a_2 - s a_1 + 3 r - s^2, \quad u^3 a_3^\prime = a_3 + r a_1 + 2t \]
For these values of $a_1^\prime, \dots , a_6^\prime$ the Eq.~\eqref{w-eq-2} is the desired result.
For a complete version of the algorithm see \cite{la-82}.
\section{Superelliptic curves with minimal discriminant}
Let $\mathcal X_g$ be a genus $g\geq 2$ superelliptic curve with equation as in \eqref{w-eq-super}.
The discriminant of $\mathcal X_g$ is the discriminant of the binary form $f(x, z)$, hence an invariant of homogenous degree ${\delta }= 2d -2$ and $\Delta_f \in \mathcal O_K$.
Let $M \in GL_2 (K)$ such that $\det M = \lambda$. Then from remarks in section 2 we have that $ \Delta (f^M) = \lambda^{d(d-1)} \, \Delta(f)$.
We perform the coordinate change $x \to \frac 1 {u^n} x$ on $f(x)$. Then the new discriminant is $\Delta^\prime = \frac 1 {u^{n \cdot d(d-1)} } \, \cdot \Delta $.
\begin{lem} A superelliptic curve $\mathcal X_g$ with integral equation
\[ y^n = a_d x^d + \cdots a_1 x + a_0 \]
is in minimal form if $\mathfrak v (\Delta) < n d(d-1)$.
\end{lem}
Hence, if we choose $u\in \mathbb Z$ such that $u^{ n d(d-1) }$ divides $\Delta$, then $\Delta^\prime$ becomes smaller. Indeed, we would like to choose the largest such $u$. In the process we have to make sure that for the $u$'s that we pick we do get an equation of a superelliptic curve isomorphic to $\mathcal X_g$.
Hence, we factor $\Delta$ as a product of primes, say $\Delta = p_1^{\alpha_1} \cdots p_r^{\alpha_r}$, and take $u$ to be the product of those powers of primes with exponents $\alpha_i \geq n d (d-1)$. For primes $p=2, 3$ we have to be more careful since in our exposition above we have assumed that the characteristic of the field is $\neq 2, 3$.
In \cite{rachel} we give a description of all the steps of the algorithm which is also implemented for genus 3 hyperelliptic curves and for triagonal curves $y^3=f(x)$ up to $\deg f \leq 8$.
The main result of \cite{rachel} is the following:
\begin{thm}
Let $\mathcal X$ be a genus $g \geq 2$ superelliptic curve defined over an algebraic number field $K$ and $P$ a $K$-rational point on $\mathcal X$. For the pair $(\mathcal X, P)$ the
global minimal discriminant $\Delta_{\mathcal X, P} \in \mathcal O_K$ is unique (up to multiplication by a unit). Moreover, there exists a minimal Weierstrass equation corresponding to this discriminant $\Delta_{\mathcal X, P}$.
\end{thm}
The theorem also provides the blueprint for the algorithm which is much more involved then the case of elliptic curves described in \cite{ta-75} and \cite{la-82} and the case of hyperelliptic curves described in \cite{liu}.
\bibliographystyle{amsplain}
\begin{bibdiv}
\begin{biblist}
\bibselect{bibl}
\end{biblist}
\end{bibdiv}
\end{document}
Let $C/K$ is a hyperelliptic curve of genus $g \geq 2$ and $\sigma$ denote its hyperelliptic involution. By definition, the field of invariants
$K(C)^{\<\sigma\>} $ is the function field of a smooth projective line $\mathbb P^1 (K)$.
A choice of a generator $x$ of $K(C)^{\< \sigma \>}$ over $K$ defines a natural smooth model $\mathbb P^1_x (\mathcal O_K)$ of $\mathbb P^1 (K)$, with $\mathbb P^1_x = Spec(\mathcal O_K[x]) \cup Spec(\mathcal O_K[1/x])$. A \textit{Weierstrass model} for $C/K$ is the integral closure of $\mathbb P^1_x$ in $K(C)$. Let $B$ denote this Weierstrass model. Then there exists a basis $\{1,y \}$ for $B$ over $\mathcal O_K[x]$ such that
\[ E : \qquad y^2+ Q(x) y= P(x), \]
with $\deg (Q(x)) \leq g+1$, and $\deg(P(x))\leq 2g+2$. To such an integral equation $(E)$ is associated a discriminant $\Delta (E)$. The integer
\[ \mathfrak v (\Delta_W):= \mathfrak v (\Delta(E)),\]
is independent of $x$ modulo the action of $PGL_2 (\mathcal O_K)$ on $K(x)$.
A \textit{minimal Weierstrass model} is a Weierstrass model $W /\mathcal O_K $ such that $ \mathfrak v (\Delta_W)$ is minimal among all Weierstrass models of $C/K$. Minimal Weierstrass models always exist, but may not be unique. The model $W/\mathcal O_K$ is smooth if and only if $\mathfrak v (\Delta_W)=0$, and $W$ is regular if $\mathfrak v (\Delta_W)=1$; see \cite{liu} for details.
Liu in \cite{liu} defined for each closed point $p_0 \in W$ a non-negative integer $\lambda (p_0)$ which is a measure of the singularity of the point $p_0$. He uses this notion of multiplicity to prove an effective criterion for a Weierstrass model to have minimal discriminant. He then compares a minimal Weierstrass model with the regular minimal model of $X/\mathcal O_K$ of $C/K$. Let $\tilde W \to W$ denote the minimal desingularization of $W$. The author gives a criterion for $\tilde W$ to be equal to $X$. He shows in particular that if $C(K)$ is not empty, then $\tilde W =X$. When $C(K)$ is empty, it may happen that $X$ is obtained from $\tilde W$ by contracting the strict transform in $\tilde W$ of the special fiber of $W$. As a corollary, Liu shows that, up to isomorphism, a given $C/K$ has only finitely many minimal Weierstrass models.
He has implemented over $\mathbb Z[\frac 1 2]$ an algorithm to find the Weierstrass equations with minimal discriminant for curves of genus 2.
In this paper we study the minimal discriminant for superelliptic curves of arbitrary genus $g \geq 1$. We generalize the work Tate \cite{ta-75}, Laska \cite{la-82}, Lockhart \cite{lockhart} and Liu \cite{liu} to superelliptic curves. Defining the discriminant of superelliptic curves it is a rather straightforward task from the definition of the discriminant of elliptic and hyperelliptic curves; see \cite{se-2} among others for basic definitions of discriminants over function fields.
Furthermore, we provide an algorithm to determine the Weierstrass equation of superelliptic curves with minimal discriminant and implement this algorithm for hyperelliptic curves of genus 3.
|
\section{Introduction}
\subsection{}
Let
\begin{equation*}
\label{eqn:TriangularArray}
\begin{matrix}
b_1^{(1)} & {} & {} & {} \\
b_1^{(2)} & b_2^{(2)} & {} & {} \\
b_1^{(3)} & b_2^{(3)} & {b_3^{(3)}} & {} \\
\vdots & \vdots & \vdots & \ddots
\end{matrix}
\end{equation*}
\noindent
be a triangular array of integers,
the elements of which are strictly
decreasing along rows. The array
\eqref{eqn:TriangularArray} gives
rise to a sequence $\Omega^{(N)}$ of planar
domains via the following construction.
Fix a coordinate system in the plane whose axes meet at a
$120^{\degree}$ angle.
We specify $\Omega^{(N)}$ by specifying its boundary,
which consists of two piecewise linear components.
One component of
$\partial\Omega^{(N)}$ --- the lower boundary --- is simply the
horizontal axis in the plane. The other component --- the
upper boundary --- is built in three steps. First,
construct the line parallel to the lower boundary passing
through the point $(0,N)$. Second, affix $N$ outward-facing
unit triangles to this line such that the midpoints of their
bases have horizontal coordinates
$b_1^{(N)} > \dots > b_{N}^{(N)}$.
Finally, erase the bases of these triangles. We will refer to
$\Omega^{(N)}$ as the \emph{sawtooth domain} of rank $N$ with
boundary conditions $(b_1^{(N)},\dots,b_{N}^{(N)})$.
\begin{figure}
\includegraphics{tiling-fade-nolabels-eps-converted-to.pdf}
\caption{\label{fig:Tiling}A lozenge tiling of
a sawtooth domain of rank $6$.}
\end{figure}
\subsection{}
A \emph{lozenge} is a unit rhombus in the plane whose
sides are parallel to one of the coordinate axes, or to the line
bisecting the obtuse angle between them. Lozenges
are thus divided into three classes: left-leaning, right-leaning,
and vertical. Given a lozenge tiling of $\Omega^{(N)}$,
as in Figure \ref{fig:Tiling}, the horizontal
line through $(0,k)$ ``threads'' exactly
$k$ vertical tiles, or ``beads'',
and the beads on adjacent
threads interlace, as in Figure
\ref{fig:Beads}.
Let $T^{(N)}$ be a
uniformly random lozenge tiling of $\Omega^{(N)}$,
and let $b_{k1}^{(N)} > \dots > b_{kk}^{(N)}$
be the horizontal coordinates of the
centroids of the beads on the
$k$th thread through $T^{(N)}$.
The main result of this note is a limit theorem
for the $k$-dimensional
random vector $(b_{k1}^{(N)},\dots,b_{kk}^{(N)})$,
in the regime where $N \rightarrow \infty$
with $k$ fixed.
\subsection{}
Suppose there exists a positive integer $M$ such
that, for each $N \geq 1$,
\begin{equation*}
\{b_1^{(N)} > \dots > b_N^{(N)}\} \subseteq
\{MN > \dots > -MN\}.
\end{equation*}
\noindent
Let $\nu^{(N)}$ be the probability measure which
places mass $1/N$ at each of the points $b_i^{(N)}/N$.
Suppose that $\nu^{(N)}$ converges weakly to $\nu$, the
probability measure on $[-M,M]$ with moment
sequence $\psi_1,\psi_2,\psi_3,\dots$.
\begin{thm}
\label{thm:Main}
For each $N \geq 1$ and $1\leq k\leq N$, set
\begin{equation*}
\tilde{b}_{kl}^{(N)} =
\frac{\frac{b_{kl}^{(N)}}{\sqrt{N}}-
(\psi_1-\frac{1}{2})\sqrt{N}}
{\psi_2-\psi_1^2-\frac{1}{12}}, \quad
1 \leq l \leq k.
\end{equation*}
\noindent
For any fixed $k$, the random vector
$(\tilde{b}_{k1}^{(N)},\dots,\tilde{b}_{kk}^{(N)})$
converges weakly to the ordered list of
eigenvalues of a $k \times k$
GUE random matrix as $N \rightarrow \infty$.
\end{thm}
Note that $\psi_1$ and $\psi_2-\psi_1^2$ are,
respectively, the mean and variance of $\nu$,
while the numbers $1/2$ and $1/12$ are
the mean and variance of the uniform
probability measure on $[0,1]$.
\subsection{}
Given that the law of large numbers for $T^{(N)}$
manifests as the convergence
of the height function of the normalized tiling
$N^{-1}T^{(N)}$ to a deterministic
limit, the so-called \emph{limit shape} \cite{DM,KO,Petrov},
the $N^{-1/2}$ scaling
in Theorem
\ref{thm:Main} is natural.
Indeed, as discussed in
\cite{OR}, the arctic curve separating the
frozen and liquid regions of $T^{(N)}$ which
emerge as $N \rightarrow \infty$
resembles a parabola near the point where
it is tangent to the lower boundary of
$\Omega^{(N)}$. For boundary conditions producing
an arctic curve which actually is a parabola,
see \cite{NY1,NY2}.
\begin{figure}
\includegraphics{tiling-beads-eps-converted-to.pdf}
\caption{\label{fig:Beads}Interlacing beads and their coordinates.}
\end{figure}
\subsection{}
The connection between the joint distribution
of vertical tiles near the frozen boundary and
GUE eigenvalues was first studied by
by Okounkov and Reshetikhin \cite{OR}.
For a special class of boundary conditions,
Theorem \ref{thm:Main} was proved
by Johansson and Nordenstam \cite{JN}.
In a slightly different (but equivalent) form,
Theorem \ref{thm:Main} was obtained in full
generality by Gorin and Panova
\cite{GP} as a consequence of their general approach to
Schur function asymptotics.
In this note, we present a different approach
to Theorem \ref{thm:Main} in which the
usual tools of integrable probability (e.g. determinantal processes,
steepest descent analysis) play no role. Instead,
our argument is based on the combinatorial
interpretation of the Harish-Chandra/Itzykson-Zuber
integral discovered in \cite{GGN3}.
\subsection{}
Work on this paper began while the author was
a Professeur Invit\'e at
Universit\' e Paris Diderot in
the Spring of 2014. I am grateful to
G. Chapuy and S. Corteel for
the invitation to visit.
While writing this article, I benefited
from stimulating correspondence
with V. Gorin and G. Panova.
I am indebted to
M. Lacroix for producing the figures which
accompany this note.
\section{Proof of Theorem \ref{thm:Main}}
\subsection{}
Let us replace the
$k$-dimensional random vector
$(b_{k1}^{(N)},\dots,b_{kk}^{(N)})$
with the random Hermitian matrix
\begin{equation*}
B_k^{(N)} = U_k
\begin{bmatrix}
b_{k1}^{(N)} & {} & {} \\
{} & \ddots & {} \\
{} & {} & b_{kk}^{(N)}
\end{bmatrix}
U_k^{-1},
\end{equation*}
\noindent
where $U_k$ is a random matrix drawn from
normalized Haar measure on the
unitary group $U(k)$. By the \emph{Laplace transform} of
$B_k^{(N)}$, we mean the function on $k \times k$
complex semisimple matrices $A$ defined by
\begin{equation*}
\label{eqn:Laplace}
A \mapsto \mathbf{E}[e^{\operatorname{Tr} AB_k^{(N)}}],
\end{equation*}
\noindent
where $\mathbf{E}$ denotes expectation. In the
case $k=1$, this function coincides with the classical
two-sided Laplace transform encoding the distribution
of the horizontal coordinate of the bottom bead.
\subsection{}
The Laplace transform of $B_k^{(N)}$
depends only on the
eigenvalues of $A$, and
thus may be considered as a function
of $k$ complex variables. This function
is analytic, because the distribution
of $B_k^{(N)}$ in $H(k)$, the space of
$k \times k$ Hermitian matrices,
is compactly supported.
Explicitly,
\begin{equation*}
\begin{split}
L_k^{(N)}(a_1,\dots,a_k) &=
\sum_{\{b_1 > \dots > b_k\} \subset \mathbb{Z}}
\mathbf{P}(b_{k1}^{(N)}=b_1,\dots,b_{kk}^{(N)}=b_k) \\
&\times\int\limits_{U(k)} e^{\operatorname{Tr} \operatorname{diag}(a_1,\dots,a_k)U
\operatorname{diag}(b_1,\dots,b_k)U^{-1}} \mathrm{d}U,
\end{split}
\end{equation*}
\noindent
where the sum is over all $k$-point
particle configurations on the integer
lattice and $\mathbf{P}$ is the uniform
probability measure on lozenge
tilings of $\Omega^{(N)}$. The integral
over $U(k)$ is just
the Laplace transform of the uniform
probability measure
on the set of $k \times k$ Hermitian
matrices with eigenvalues
$b_1 > \dots > b_k$. That is,
$L_k^{(N)}$ is
the Laplace transform of a mixture of
\emph{orbital measures}.
If $k=N$, the bead locations are deterministic,
and we are dealing with the Laplace transform of a
pure orbital measure. The following proposition
reduces our workload to the analysis of the
Laplace transforms of pure orbital measures.
\begin{prop}
\label{prop:Key}
For any integers
$1 \leq k \leq N$,
\begin{equation*}
L_k^{(N)}(a_1,\dots,a_k) =
\left( \prod_{i=1}^k \frac{a_i}{e^{a_i}-1}
\right)^{N-k} L_N^{(N)}(a_1,\dots,a_k,0,\dots,0).
\end{equation*}
\end{prop}
\begin{proof}
The proof is a combination of
three standard facts from
the representation theory of the complex
general linear group $GL(N)$.
First, the isomorphism classes
of irreducible rational representations
of $GL(N)$ are indexed by $N$-point particle
configurations on $\mathbb{Z}$.
This is a classical result,
see e.g. \cite{Weyl}.
Second,
given a particle configuration
$\{b_1 > \dots > b_N\} \subset \mathbb{Z}$,
the corresponding normalized irreducible character
\begin{equation*}
\frac{\chi^{(b_1,\dots,b_N)}(e^{a_1},\dots,e^{a_N})}
{\chi^{(b_1,\dots,b_N)}(1,\dots,1)}
\end{equation*}
\noindent
equals the twisted Laplace transform
\begin{equation*}
\prod_{1\leq i<j \leq N} \frac{a_i-a_j}{e^{a_i}-e^{a_j}}
\int\limits_{U(N)}
e^{\operatorname{Tr} \operatorname{diag}(a_1,\dots,a_N)U\operatorname{diag}(b_1,\dots,b_N)U^{-1}}
\mathrm{d}U
\end{equation*}
\noindent
of the uniform measure on Hermitian
matrices with spectrum $\{b_1 > \dots > b_N\}$.
This identity is independently due to Harish-Chandra
\cite{HC}, and Itzykson and Zuber \cite{IZ}
--- it is the $U(N)$ case of the Kirillov
character formula \cite{Kirillov}.
The third and final ingredient is
the branching rule for irreducible
characters of $GL(N)$ under restriction
to $GL(N-1)$:
\begin{equation*}
\chi^{(b_1,\dots,b_N)}(e^{a_1},\dots,e^{a_{N-1}},1)
=\sum_{\{c_1 > \dots > c_{N-1}\}\subset \mathbb{Z}}
\chi^{(c_1,\dots,c_{N-1})}(e^{a_1},\dots,e^{a_{N-1}}),
\end{equation*}
\noindent
where the sum is over all configurations of
$N-1$ particles on $\mathbb{Z}$ which interlace
with the configuration $\{b_1 > \dots > b_N\}$.
A proof of the branching rule may be found in
\cite[Chapter 8]{GW}. Iterating the branching rule
$N-k$ times and applying the Harish-Chandra formula
yields the stated formula for $L_k^{(N)}$ in
terms of $L_N^{(N)}$.
\end{proof}
\subsection{}
Consider the analytic function
$\mathbb{C} \times \mathbb{C}^N \times \mathbb{C}^N
\rightarrow \mathbb{C}$
defined by
\begin{equation*}
\label{eqn:HCIZ}
(z;a_1,\dots,a_N;b_1,\dots,b_N) \mapsto
\int\limits_{U(N)} e^{z\operatorname{Tr}
\operatorname{diag}(a_1,\dots,a_N)U\operatorname{diag}(b_1,\dots,b_N)U^{-1}}
\mathrm{d}U.
\end{equation*}
\noindent
This is the famous Harish-Chandra/Itzykson-Zuber
integral. The parameter $z$ may
be called the \emph{coupling constant}, as a
reference to its origin in the spectral analysis
of coupled random semisimple
matrices with $AB$-interaction
\cite{AvM,IZ}.
The HCIZ integral enjoys
a natural $S(N) \times S(N)$ symmetry: it
is invariant under permutation of the $a$'s
amongst themselves, and the $b$'s amongst
themselves.
Combining this symmetry
with the fact that the Newton power-sums
form a linear basis of the algebra of
symmetric polynomials, we may present the Maclaurin
series of the logarithm of the HCIZ
integral in the form
\begin{equation*}
\begin{split}
&\log \int\limits_{U(N)} e^{z\operatorname{Tr}
\operatorname{diag}(a_1,\dots,a_N)U\operatorname{diag}(b_1,\dots,b_N)U^{-1}}
\mathrm{d}U \\
=&\sum_{d=1}^{\infty} \frac{z^d}{d!}
\sum_{\alpha,\beta \vdash d} C_N(\alpha,\beta)
p_{\alpha}(a_1,\dots,a_N)p_{\beta}(b_1,\dots,b_N),
\end{split}
\end{equation*}
\noindent
where the internal sum is over all pairs
of Young diagrams with $d$ cells.
\subsection{}
The coefficients $C_N(\alpha,\beta)$
have the following combinatorial
interpretation. Consider the
Cayley graph of the symmetric group $S(d)$ as generated
by the conjugacy class of transpositions.
Equip this graph with the Biane-Stanley
edge labelling \cite{Biane,Stanley}, wherein each
edge corresponding to
the transposition $(s\ t)$ is tagged
with $t$, the larger of the
two numbers interchanged. The
$d=4$ case is
shown in Figure \ref{fig:Cayley}, where $2$-edges
are drawn in blue, $3$-edges in yellow, and $4$-edges
in red.
A walk on the Cayley graph is said to be
\emph{monotone} if the labels of the edges it traverses
form a weakly increasing sequence.
A walk is \emph{transitive} if its steps
and endpoints together generate a transitive
subgroup of $S(d)$. Given
two partitions $\alpha,\beta \vdash d$,
and a nonnegative integer $r$, let $\vec{H}^r(\alpha,\beta)$
be the number of $r$-step monotone,
transitive walks on $S(d)$ which
begin at a permutation of cycle type $\alpha$ and
end at a permutation of cycle type $\beta$.
\begin{thm}[\cite{GGN3}]
\label{thm:LeadingDerivatives}
For any $1 \leq d \leq N$, and
any $\alpha,\beta \vdash d$,
we have
\begin{equation*}
C_N(\alpha,\beta) = \frac{1}{N^d}
\sum_{r=0}^{\infty} (-1)^r
\frac{\vec{H}^r(\alpha,\beta)}{N^r}.
\end{equation*}
\end{thm}
\begin{figure}
\includegraphics{CayleyS4-Palettec-eps-converted-to.pdf}
\caption{\label{fig:Cayley} $S(4)$ with the Biane-Stanley
edge-labelling.}
\end{figure}
\subsection{}
The number $H^r(\alpha,\beta)$, which counts walks
as above, but without the monotonicity constraint, is
a \emph{double Hurwitz number}. The double Hurwitz
numbers are important quantities in classical
and modern enumerative geometry, see
\cite{GJV,Okounkov}.
Reversing a classical construction due to
Hurwitz \cite{Hurwitz}, we have that
\begin{equation*}
\frac{1}{d!} H^r(\alpha,\beta) =
\sum_{(X,f)} \frac{1}{|\operatorname{Aut}(X,f)|},
\end{equation*}
\noindent
where the sum runs over all isomorphism
classes of pairs $(X,f)$ in which $X$ is
a compact, connected Riemann surface and $f:X \rightarrow
\mathbf{P}^1$ is a degree $d$ mapping
to the Riemann sphere with profile
$\alpha$ over $\infty$, profile $\beta$
over $0$, and simple ramification over
the $r$th roots of unity.
By the Riemann-Hurwitz formula, such a branched
covering exists if and only if
\begin{equation*}
\label{eqn:RiemannHurwitz}
g = \frac{r+2-\ell(\alpha)-\ell(\beta)}{2}
\end{equation*}
\noindent
is a non-negative integer, in which
case $g$ is the genus of $X$. Here
$\ell(\alpha)$ is the number of parts
in the partition $\alpha \vdash d$, and
likewise for $\ell(\beta)$.
We write
$H^r(\alpha,\beta) = H_g(\alpha,\beta)$,
with the understanding that $r$
and $g$ determine one another via
Riemann-Hurwitz.
\subsection{}
Following the terminology of \cite{GGN3},
we refer to the numbers
$\vec{H}^r(\alpha,\beta)=\vec{H}_g(\alpha,\beta)$ as the
\emph{monotone} double Hurwitz numbers.
The expansion in Theorem
\ref{thm:LeadingDerivatives} may
equivalently be written
\begin{equation}
\label{eqn:LeadingDerivatives}
C_N(\alpha,\beta) =
(-1)^{\ell(\alpha) + \ell(\beta)}
N^{2-d-\ell(\alpha)-\ell(\beta)}
\sum_{g \geq 0} \frac{\vec{H}_g(\alpha,\beta)}{N^{2g}}.
\end{equation}
\noindent
This expansion renders the asymptotics
of the HCIZ integral transparent in
virtually any scaling regime. In particular,
one obtains the following limits.
\begin{prop}
\label{prop:FreeCumulant}
Under the hypotheses of Theorem
\ref{thm:Main},
for any fixed $d \in \mathbb{N}$
and $a_1,\dots,a_k \in \mathbb{C}$, we
have
\begin{equation*}
\begin{split}
&\lim_{N \rightarrow \infty}\frac{1}{N}
\sum_{\alpha,\beta \vdash d} C_N(\alpha,\beta)
p_{\alpha}(a_1,\dots,a_k)
p_{\beta}(b_1^{(N)},\dots,b_N^{(N)}) \\
=&\ p_d(a_1,\dots,a_k)
\sum_{\beta \vdash d} (-1)^{1+\ell(\beta)}
\vec{H}_0(d,\beta)\psi_{\beta},
\end{split}
\end{equation*}
\noindent
where $\psi_\beta=\prod_i \psi_{\beta_i}$.
\end{prop}
\begin{proof}
According to \eqref{eqn:LeadingDerivatives},
we have
\begin{equation*}
\begin{split}
&\frac{1}{N}
\sum_{\alpha,\beta \vdash d} C_N(\alpha,\beta)
p_{\alpha}(a_1,\dots,a_k)
p_{\beta}(b_1^{(N)},\dots,b_N^{(N)}) \\
=& \sum_{\alpha \vdash d} (-1)^{\ell(\alpha)}
\frac{p_{\alpha}(a_1,\dots,a_k)}{N^{\ell(\alpha)-1}}
\sum_{\beta \vdash d} (-1)^{\ell(\beta)}
\frac{
p_{\beta}(\frac{b_1^{(N)}}{N},\dots,\frac{b_N^{(N)}}{N})}
{N^{\ell(\beta)}}
\sum_{g=0}^{\infty} \frac{\vec{H}_g(\alpha,\beta)}{N^{2g}}
\end{split}
\end{equation*}
\noindent
for any $N \geq d$.
From the definition of $\vec{H}_g(\alpha,\beta)$,
we have the upper bound
\begin{equation*}
\vec{H}_g(\alpha,\beta) \leq (d!)^{2g +
\ell(\alpha)+\ell(\beta)} \leq
(d!)^{2g+2d}.
\end{equation*}
\noindent
Thus
\begin{equation*}
\sum_{g=0}^{\infty} \frac{\vec{H}_g(\alpha,\beta)}{N^{2g}}
= \vec{H}_0(\alpha,\beta) +O\left(\frac{1}{N^2}\right)
\end{equation*}
\noindent
as $N \rightarrow \infty$, uniformly in $\alpha,\beta$.
The weak convergence of $\nu^{(N)}$ to $\nu$,
the measure on $[-M,M]$ with moments
$\{\psi_m : m \in \mathbb{N}\}$,
is equivalent to the limits
\begin{equation*}
\lim_{N \rightarrow \infty}
\frac{p_m(\frac{b_1^{(N)}}{N},\dots,\frac{b_N^{(N)}}{N})}{N}
=\psi_m, \quad m \in \mathbb{N}.
\end{equation*}
\end{proof}
\subsection{}
The numbers $\vec{H}_g(d,\beta)$ are \emph{one-part}
monotone double Hurwitz numbers; their classical
counterparts $H_g(d,\beta)$ are analyzed in \cite{GJV}.
The sum
\begin{equation*}
K_d = \sum_{\beta \vdash d} (-1)^{1+\ell(\beta)}
\vec{H}_0(d,\beta)\psi_{\beta}
\end{equation*}
\noindent
which emerges in Proposition \ref{prop:FreeCumulant}
is an element of $\mathbb{Z}[\psi_1,\dots,\psi_d]$,
homogeneous of degree $d$ with respect to the grading
$\deg(\psi_m)=m$. In fact, $K_d$
is, up to a simple factor, the $d$th \emph{free cumulant}
$\kappa_d$ of the measure $\nu$:
\begin{equation}
\label{eqn:FreeCumulant}
K_d = (d-1)! \kappa_d.
\end{equation}
We recall that the free cumulants of a probability
measure are obtained by replacing the lattice of all
partitions with the lattice of noncrossing partitions
in the moment-cumulant formula, see e.g. \cite{NS}.
The identity \eqref{eqn:FreeCumulant} may be
established in a purely combinatorial way, by
viewing the noncrossing partition lattice
$NC(d)$ as the set of geodesic paths
$(1)\dots (d) \rightarrow (1\ \dots\ d)$
on the Cayley graph of $S(d)$ and using the
Kreweras antiautomorphism. For our purposes,
we only require explicit knowledge of $K_1$
and $K_2$, which can be computed directly
from the definition of the monotone
double Hurwitz numbers:
\begin{equation*}
\begin{split}
&\vec{H}_0(1,1)=1 \implies K_1=\psi_1 \\
&\vec{H}_0(2,2) = \vec{H}_0(2,11) = 1 \implies
K_2=\psi_2-\psi_1^2.
\end{split}
\end{equation*}
\noindent
We thus leave the proof of
\eqref{eqn:FreeCumulant}
to the interested reader.
\subsection{}
The absolute summability of the series
\begin{equation*}
\sum_{d=1}^{\infty} \frac{z^d}{d!} K_d
\end{equation*}
\noindent
follows from \cite[Theorem 3.4]{GGN3}.
Arguing as in \cite[Theorem 4.1]{GGN3},
Proposition \ref{prop:FreeCumulant}
may be promoted to the following
scaling limit of the HCIZ integral,
which is closely related to the
results of \cite{CS,GM}.
\begin{prop}
\label{prop:FPlimit}
Let $k \in \mathbb{N}$ be fixed.
Under the assumptions of Theorem
\ref{thm:Main}, there exists $\varepsilon > 0$
such that
\begin{equation*}
\frac{1}{N}\log
\int\limits_{U(N)} e^{z\operatorname{Tr}\operatorname{diag}(a_1,\dots,a_k,0,\dots,0)
U\operatorname{diag}(b_1^{(N)},\dots,b_N^{(N)})U^{-1}}
\mathrm{d}U
\rightarrow
\sum_{d=1}^\infty \frac{z^d}{d!} p_d(a_1,\dots,a_k)K_d,
\end{equation*}
\noindent
uniformly on compact subsets of
$\{(z;a_1,\dots,a_k) \in \mathbb{C} \times \mathbb{C}^k
: |za_i| < \varepsilon \}$.
\end{prop}
Tuning the coupling constant to $z=N^{-1/2}$,
Proposition \ref{prop:FPlimit} yields the following corollary.
\begin{cor}
\label{cor:TilingLimit}
Let $k \in \mathbb{N}$ be fixed.
Under the assumptions of Theorem
\ref{thm:Main}, we have
the $N \rightarrow \infty$
asymptotic expansion
\begin{equation*}
\log
\int\limits_{U(N)} e^{\frac{1}{\sqrt{N}}
\operatorname{Tr}\operatorname{diag}(a_1,\dots,a_k,0,\dots,0)
U\operatorname{diag}(b_1^{(N)},\dots,b_N^{(N)})U^{-1}}
\mathrm{d}U \sim
\sum_{d=1}^\infty
\frac{K_d}{d!} p_d(a_1,\dots,a_k)N^{1-\frac{d}{2}},
\end{equation*}
\noindent
uniformly on compact subsets
of $\mathbb{C}^k$.
\end{cor}
\subsection{}
Combining Corollary \ref{cor:TilingLimit}
with the fact that
\begin{equation*}
\log \frac{a}{e^a-1}=
-\log \frac{e^a-1}{a} = -\frac{1}{2} \frac{a^1}{1!}
- \frac{1}{12} \frac{a^2}{2!} + \dots
\end{equation*}
\noindent
is negative one times the generating
function for the classical cumulants
$c_1,c_2,\dots$ of uniform measure on
$[0,1]$, Proposition \ref{prop:Key}
yields the asymptotic expansion
\begin{equation*}
\log L_k^{(N)}(\frac{a_1}{\sqrt{N}},\dots,
\frac{a_k}{\sqrt{N}}) \sim \sum_{d=1}^\infty
\frac{K_d-c_d}{d!} p_d(a_1,\dots,a_k)N^{1-\frac{d}{2}},
\end{equation*}
\noindent
uniformly on compact subsets of $\mathbb{C}^k$.
In particular,
\begin{equation*}
\log L_k^{(N)}(\frac{a_1}{\sqrt{N}},\dots,
\frac{a_k}{\sqrt{N}})
=\sqrt{N}(\psi_1-\frac{1}{2})p_1(a_1,\dots,a_k)
+\frac{1}{2}(\psi_2-\psi_1^2-\frac{1}{12})p_2(a_1,\dots,a_k)
+ O\left( \frac{1}{\sqrt{N}} \right)
\end{equation*}
\noindent
as $N \rightarrow \infty$.
Since a $k \times k$ standard GUE random
matrix $X_k$ is characterized by the
log-Laplace transform
\begin{equation*}
\log \mathbf{E}[e^{\operatorname{Tr} AX_k}]=
\frac{1}{2}\operatorname{Tr} A^2,
\end{equation*}
\noindent
and since $H(k)$ is a finite-dimensional
Euclidean space with the inner product
$\langle A, B\rangle = \operatorname{Tr} AB$, Theorem
\ref{thm:Main} follows from the above
quadratic approximation and the L\'evy
continuity theorem.
\bibliographystyle{amsplain}
|
\section{Introduction} \label{sec intro}
Let $M$ be an oriented compact Riemannian manifold of dimension $n=4h-1$, $h \ge 1$, and consider the Atiyah-Patodi-Singer operator $\mathcal{D}$ (\textit{APS-operator} for short) defined on the space of smooth even forms $\Omega^{ev}(M) = \bigoplus_{p=0}^{2h-1} \Omega^{2p}(M)$ by
\begin{equation} \label{D}
\mathcal{D} \phi = (-1)^{h+p-1} (*d-d*) \phi
\end{equation}
with $\phi\in \Omega^{2p}(M)$, where $\Omega^{2p}(M)$ denotes the degree $2p$ forms.
This operator is closely related to the signature operator. In fact,
$\mathcal{D}$ is the tangential boundary operator of the signature operator $\mathcal{S}$ acting on a $4h$-dimensional manifold $\tilde M$ having $M$ as its boundary.
By compactness of $M$, $\mathcal{\mathcal{\mathcal{D}}}$ has a discrete spectrum, $\mathrm{Spec}_\mathcal{D}(M)$, of real eigenvalues $\lambda$ with finite multiplicity $d_\lambda$ which accumulate only at infinity. The \emph{eta series}
\begin{equation} \label{Eta(s)}
\eta(s) = \sum_{0 \not = \lambda \in \mathrm{Spec}_\mathcal{D}(M)} sign(\lambda) \, |\lambda|^{-s}, \qquad \mathrm{Re}(s)> n,
\end{equation}
defines a holomorphic function having a meromorphic continuation to $\C$, also denoted by $\eta(s)$, having (possibly) simple poles in the set $\{n-k : k\in \N_0\}$. Remarkably, $\eta(s)$ is holomorphic at $s=0$ and the value $\eta = \eta(0)$
is called the \emph{eta invariant} of $\mathcal{D}$.
Both $\mathcal{D}$ and $\eta(s)$ were first introduced and studied by Atiyah, Patodi and Singer in a sequel of 3 classical papers \cite{APS1}, where they also proved the regularity of $\eta(s)$ at the origin in the case of odd dimension. The finiteness of $\eta$ in any dimension is due to P.B.\@ Gilkey (\cite{Gi}). Actually, the results in \cite{APS1} and \cite{Gi} are valid for arbitrary elliptic differential operators.
\sk
Any orientable compact flat manifold (in what follows \textit{cfm} for short) is isometric to
$M_\G = \G \backslash \R^n$, with $\G$ an orientable Bieberbach group, i.e.\@ a discrete, cocompact, torsion-free subgroup of
$\mathrm{I}^+(\R^n) = \mathrm{SO}(n) \ltimes \R^n$, the orientation preserving isometriy group of $\R^n$.
Thus,
$\G = \langle \gamma= BL_b, L_\Lambda \rangle$, with $L_\Lambda = \{L_\lambda : \lambda \in \Lambda \}$,
where $B \in \mathrm{SO}(n)$, $L_b$ denotes translation by $b\in \R^n$, $\g^k \ne \mathrm{Id}$ for every $k\in \N$ and $\Lambda$ is a $B$-stable lattice in $\R^n$. We will usually identify the point group with the holonomy group, that is $F\simeq \Lambda \backslash \G$.
Since $BL_\lambda B^{-1}=L_{B\lambda} \in L_\Lambda$ for any $B\in \mathrm{SO}(n), \lambda \in \Lambda$, conjugation by $F$ in
$\Lambda \simeq \Z^n$ defines the integral holonomy representation $\rho$ of $\G$ (which does not determine $M_\Gamma$ uniquely, in general). A \textit{$G$-manifold} is a cfm with holonomy group $F\simeq G$. In this paper we will be concerned with the case $F\simeq \Z_{2^r}$, for $r\in \N$.
\msk
In \cite{MP.JGA11}, we give a general expression for $\eta(s)$ and $\eta$ for $\mathcal{D}$ acting on an arbitrary orientable cfm
$M_\G$ (see Theorems 3.3, 3.5 and 4.2). Also, simpler expressions can be found in its sequel \cite{MP.AGAG12} in the case of cyclic holonomy group (see Proposition \ref{AGAG}). There, it is shown that the computation of $\eta(s)$ can be reduced to the case when $F$ is cyclic (\cite{MP.AGAG12}, Proposition~5.1).
This, enabled us to compute the $\eta$-invariant of some nice families of cfm's, such as $F$-manifolds with
$F\simeq \Z_p, \Z_p\times \Z_q, \Z_2^k$ and $\Z_p^k$, with $p,q$ odd primes and $k\ge 2$, or even with non-abelian holonomy group $F$, where $F$ is of order 8 or $F$ is metacyclic (dihedral and of odd order). Expressions for $\eta(s)$ and $\eta$ on cfm's for the spin Dirac operator $D$ were studied in \cite{MP.TAMS06} (the general case and $\Z_2^k$-manifolds), in \cite{Po.UMA05} ($\Z_4$-manifolds) and in \cite{MP.PAMQ09} and \cite{GMP} ($\Z_p$-manifolds, $p$ odd prime).
\msk
The main goal of this paper is to compute $\eta(s)$ and $\eta$ for $\mathcal{D}$ on orientable $\Z_{2^r}$-manifolds
of odd dimension $n=2m+1$, $m$ odd.
In this way, the present paper can be considered as a natural continuation of \cite{MP.JGA11} and \cite{MP.AGAG12}.
We remark that these manifolds are not classified (as it is the case for $\Z_p$ manifolds, $p$ prime) since there is no such classification available for integral representations of $\Z_{2^r}$, $r \ge 2$.
\sk
A brief outline of the paper is as follows. In Section \ref{sec rotation}, we study the rotation angles of a matrix $B\in \mathrm{GL}(n)$ of order $2^r$. Some trigonometric identities involving sines at special angles of the form $\tf{k\ell \pi}{2^r}$, with $k,\ell$ odd integers, are given in Propositions \ref{prop. sines2} and \ref{prop. sumsines} (see also Propositions \ref{lema. cots} and \ref{prop. sumcots} for related identities involving cotangents).
These identities play a key role in the computation of $\eta(s)$ and $\eta$, allowing great simplifications in the steps of the proofs, which lead to the final expressions in the statements.
\sk
Section \ref{sec etas} is devoted to the computation of $\eta(s)$ for arbitrary $\Z_{2^r}$-manifolds in all dimensions $n=2m+1$, $m$ odd. Let $M_\G$ be a $\Z_{2^r}$-manifold with $\gamma=BL_b$ the generator of $\Lambda\backslash \G$.
In Theorems \ref{teo.etasz2r} we show that, in the non-trivial cases,
\begin{equation} \label{etasgen}
\eta(s) = \tf{\eta}{(2^{r+1-\nu}\pi \lambda_{_B})^s} (\zeta(s,\tf 14)-\zeta(s, \tf 34)) = \tf{2\eta}{(2^{r-1-\nu}\pi \lambda_{_B})^s} L(s,\chi_4),
\end{equation}
where $\nu \in \N\cap [0,r-2]$ and $\lambda_{_B}\in \R$ are certain constants depending on the metric,
$\zeta(s,a) = \sum_{n=0}^\infty (n+a)^{-s}$ is the Hurwitz zeta function, with $a \in (0,1]$, $\mathrm{Re}(s)>1$, and $L(s,\chi_4)$ is the $L$-function associated to the primitive Dirichlet character mod 4. By Proposition \ref{etaux1}, the $\eta$-invariant has the simple expression
$$\eta = \pm \, 2^{f(B)-2} \,,$$
where $\pm$ is a sign depending subtly on $\g$, and $f(B)$ is the number of irreducible factors of the characteristic polynomial of $B$ depending on $r$ and $n$ (see \eqref{fb}).
\sk
Next, in Section \ref{sec.family}, we introduce an infinite family $\mathcal{F}$ of orientable $\Z_{2^r}$-manifolds of dimensions
$n=4h-1$, $h\ge 1$, each having integral holonomy representation.
In Theorem~\ref{teoH1} we compute the first (co)homology groups over $\Z$ and $\Z_2$ of $M\in \mathcal{F}$. This allows us to get the $\eta$-invariant of $M$ (and hence $\eta(s)$, by \eqref{generaletas} and \eqref{generaleta0}) in topological terms; namely
$$\eta(M) = -\tf 12 |\mathrm{Tor}(H_1(M,\Z))|.$$
We also posed some queries, see Questions \ref{q1}, \ref{q2} and \ref{q3}.
In Section \ref{sec eta*}, we show that for any $\Z_{2^r}$-manifold $M$, there is some $M_{k(r)} \in \mathcal{F}$ such that
$\eta(M) = \eta(M_{k(r)})$. This allows us to prove that the set of possible values of $\eta(M)$, with $M$ ranging over all $\Z_{2^r}$-manifolds, $r\ge 1$, is
$$\{0\}\cup \{\pm 2^{k} : k\in \N_0\}.$$
Then, we show that there are infinite families of $\Z_{2^r}$-manifolds with $\eta=2^k$, for each $k$ (with growing dimensions).
Moreover, there is a number $n_{r,k}$ such that for every $n\ge n_{r,k}$ there is a $\Z_{2^r}$-manifold of dimension $n$ with $\eta=2^k$.
\sk
Finally, in the last section, we compare expression \eqref{cyclic eta0} for the $\eta$-invariant of any $\Z_{2^r}$-manifold, with Donnelly's expression \eqref{doneli}, proved for a special class of $\Z_{2^r}$-manifolds (see Proposition \ref{etaz2rdoneli}).
\section{Rotation angles for order $2^r$ matrices} \label{sec rotation}
As we shall later see, the results in this section are crucial for the determination of $\eta(s)$ in Theorem \ref{teo.etasz2r}.
We will first study the rotation angles of a matrix in $\mathrm{GL}(n,\Z)$ of order $2^r$, and then compute some trigonometric identities related to them.
\subsection{Maximal torus and rotation angles}
For any $d\in \N$, let $\mathcal{U}_d =\{\omega \in \C : \omega^d=1\}$ be the group of complex $d$-th roots of unity and denoteby
$\mathcal{U}_d^*=\{\omega \in\mathcal{U}_d : \mathrm{ord}(\omega) =d\}$ the subgroup of primitive roots in $\mathcal{U}_d$.
Then, for any $N\in \N$, we clearly have $\mathcal{U}_N = \bigcup_{d \mid N} \mathcal{U}_d^*$.
The cyclotomic polynomial of order $d$ is defined by $\Phi_d(x)= \prod_{\omega \in \mathcal{U}_d^*} (x-\omega) \in \C[x]$.
It is known that $\Phi_d(x) \in \Z[x]$ is irreducible over $\Q$ of degree $\varphi(d)$, where $\varphi$ is the Euler totient function. We have the relation
\begin{equation}\label{cyclotomic decomp}
x^n-1 = \prod_{d\mid n} \Phi_d(x).
\end{equation}
\begin{lema}\label{lema charpol}
Let $B \in \mathrm{GL}(n,\Z)$ of order $N \in \N$. Then, the characteristic polynomial $p_{_B}(x)$ of $B$ has the prime factorization
\begin{equation}\label{eq. charpol}
p_{_B}(x) = \prod_{d\mid N} \Phi_d(x)^{c_d}
\end{equation}
in $\Z[x]$, with $c_d\ge0$ for any $d\mid N$ and $c_N\ge 1$. Also, $\deg p_{_B} = \sum_{d\mid N} c_d \, \varphi(d)$.
\end{lema}
\begin{proof}
We know that $p_{_B}(x)$ is a monic polynomial in $\Z[x]$, all of whose roots are in $\mathcal{U}_N$.
Since $\Z[x]$ is a unique factorization domain, $p_{_B}(x)=\prod_j p_j(x)$, with $p_j(x) \in \Z[x]$ monic irreducible for each $j$.
We will show that the only monic irreducible polynomial in $\Z[x]$, with roots in $\mathcal{U}_h$ is $\Phi_h(x)$, $h\mid N$. Let $\alpha$ be a root of $p_{_B}(x)$ and let $\sigma \in \mathrm{Gal}(\Q(\alpha)/\Q)$
with $\sigma(\alpha)=\alpha^k$ for $(k,h)=1$. We have $p_{_B}(\alpha^k) = p_{_B}(\sigma(\alpha)) = \sigma(p_{_B}(\alpha)) = 0$,
for any $(k,h)=1$. Thus, $\Phi_h(x) \mid p_{_B}(x)$ in $\Q[x]$, hence in $\Z[x]$, and, by irreducibility, it must be one of the $p_j(x)$'s. Thus, we get \eqref{eq. charpol} with $c_d\ge 0$ for each $d\mid N$.
Clearly, we must have $c_N\ge 1$ for $B$ to have order $N$.
The assertion on the degree is obvious.
\end{proof}
For $B\in \mathrm{GL}(n,\Z)$ we define $f(B)$ to be the \emph{number of irreducible factors} in the prime decomposition of the characteristic polynomial $p_{_B}(x)$ of $B$ in $\Z[x]$. Thus, by \eqref{eq. charpol},
\begin{equation}\label{fb}
1\le f(B)= \sum_{d \, \mid \, o(B)} c_d
\end{equation}
where $o(B)$ is the order of $B$. If $B= (\begin{smallmatrix} B' & \\ & 1 \end{smallmatrix})$, then $f(B') = f(B)-1$ is the number of irreducible factors in the prime decomposition of $p_{_{B'}}(x)= p_{_B}(x)/(x-1)$.
\sk
Let $n=2m+1$.
Since $\mathrm{SO}(n)$ is a compact connected Lie group, it contains a maximal torus
$T_n = \{x(t_1,\ldots,t_m) : t_1,\ldots,t_m\in \R \}$, where
\begin{equation}\label{xs}
x(t_1,\ldots,t_m) := \mathrm{diag}(x(t_1),\ldots,x(t_m),1)
\end{equation}
with $x(t) = (\begin{smallmatrix} \cos t & -\sin t \\ \sin t & \cos t \end{smallmatrix})$, $t\in \R$.
Also, $T_{n-1} = \{x(t_1,\ldots,t_m) : t_1,\ldots,t_m\in \R \}$, where now
$x(t_1,\ldots,t_m) = \mathrm{diag}(x(t_1),\ldots,x(t_m))$.
\sk
Let $\G\subset I^+(\R^n)$ be a Bieberbach group and for any $BL_b\in \G$ put $n_B := \dim (\R^n)^B$. Note that $n_{\pm B}$ is the multiplicity of the $(\pm 1)$-eigenvalues of $B$.
Since $B\in \mathrm{SO}(n)$, $B$ is conjugate to some element $x_B\in T_n$. Thus, there is $C\in \mathrm{GL}(n)$ such that $CBC^{-1}=x_B$. Since $x_B$ fixes $e_n$, $B$ fixes $u=C^{-1}e_n$ and hence $n_{x_B}=n_B\ge 1$
(this is known by other methods, see for instance \cite{MR}).
Also, note that $p_{_B}(x)=p_{x_B}(x)$ and therefore $f(B)=f(x_B)$.
\begin{prop}\label{lema angles}
Suppose $\G$ is an orientable Bieberbach group of dimension $n=2m+1$.
Let $BL_b \in \G$ with $B$ of order $N$. Then, $B$ is conjugate in $T_{2m}$ to $x_B=x(t_1,t_2\ldots,t_m)$ or to
$x_B'=x(-t_1,t_2,\ldots,t_m)$ with
\begin{equation}\label{xB ord N}
x_B = x(\underbrace{ j_{s,1} \tf{2\pi}{d_s}, \ldots, j_{s,f_s} \tf{2\pi}{d_s} }_{c_{d_s}\ge 1},\ldots, \underbrace{ j_{1,1} \tf{2\pi}{d_1}, \ldots, j_{1,f_1} \tf{2\pi}{d_1} }_{c_{d_2}}, \underbrace{0}_{c_1-1}),
\end{equation}
where
$1 = d_1 < d_2 < \cdots < d_s = N$ are all the divisors of $N$ and each $c_{d_j}$ is the exponent of
$\Phi_{d_j}(x)$ in the prime factorization of $p_{x_B}(x)$ in \eqref{eq. charpol}, such that
$$(j_{i,k}, d_i)=1, \qquad 1\le j_{i,k} \le \lfloor \tf{d_i}{2} \rfloor, \qquad 1 \le k \le f_i = \tf{\varphi(d_i)}{2},$$
for each $2\le i \le s$.
\end{prop}
\begin{proof}
Since $B' \in \mathrm{SO}(n-1)$, $B'$ is conjugate to some element $x_B$ in $T_{n-1}$.
There are two conjugacy classes in $\mathrm{SO}(n-1)$, and hence $B$ is conjugate to $x_B=x(t_1,t_2,\ldots,t_m)$ or to $x_B'=x(-t_1,t_2,\ldots,t_m)$ with $0\le t_i < 2\pi$ for $i=1,\ldots,m$ (see the comments in between (3.8) -- (3.10) in \cite{MP.JGA11}).
Since $B$ is of order $N>1$, it is clear that the rotation angles must be of the form $\tf{2j\pi}{N}$, for certain
$1\le j \le [\tf{N-1}2]$. Also, $o(B)=N$ implies $c_N \ge 1$.
Thus, if $\omega \in \mathcal{U}_N^*$, all the $\varphi(N)$ primitive $N$-th roots of unity $\{\omega^j : (j,N)=1\}$ are roots of
$p_{_B}(x)$. Since complex roots appear in conjugate pairs, it suffices to consider the angles in $[0,\pi]$. Hence, the rotation angles corresponding to $\mathcal{U}_N^*$ are the $\tf{\varphi(N)}2$ angles $\tf{2j\pi}{N}$ with $(j,N)=1$, $1\le j\le [\tf N2]$, each with multiplicity $c_N$. Similar arguments apply for every $\zeta \in \mathcal{U}_d^*$ with $d\mid N$ and $c_d \ge 1$.
\end{proof}
As a direct consequence of the previous results, we can be more precise in the case $N=2^r$.
\begin{coro}\label{coroz2r}
Let $M_\G$ be a $\Z_{2^r}$-manifold of dimension $n$, with $F = \langle B \rangle$. Then,
\begin{equation}\label{eq. charpol2}
p_{_B}(x) = (x-1)^{c_1} \prod_{k=1}^r (x^{2^{k-1}}+1)^{c_{2^k}}
\end{equation}
with $n=\deg p_{_B} = c_1+ \sum_{k=1}^r 2^{k-1} c_{2^k}$ and $c_1, c_{2^r}\ge 1$.
Also,
\begin{equation}\label{xB2r}
x_B = x(\underbrace{ \tf{\pi}{2^{r-1}}, \tf{3\pi}{2^{r-1}}, \ldots, \tf{(2^{r-1}-1)\pi}{2^{r-1}} }_{c_{2^r}}, \ldots,
\underbrace{ \tf{\pi}{4}, \tf{3\pi}{4} }_{c_8}, \underbrace{ \tf{\pi}{2} }_{c_4}, \underbrace{ \pi}_{c_2},
\underbrace{ 0 }_{c_1-1}).
\end{equation}
\end{coro}
\begin{proof}
First we note that we can assume that $B\in \mathrm{O}(n,\Z)$. In fact, $B \in \mathrm{O}(n)$ and since $\Lambda\simeq \Z^n$ we have $\rho(B) \in \mathrm{GL}(\Lambda) \simeq \mathrm{GL}(\Z^n) \simeq \mathrm{GL}(n,\Z)$ and hence $\det(\rho(B))=\det(B)$ and $p_B(x) = p_{\rho(B)}(x)$.
By \eqref{cyclotomic decomp} and $x^{2^k}-1=(x^{2^{k-1}}-1)(x^{2^{k-1}}+1)$ we have that
$\Phi_{2^k}(x) = x^{2^{k-1}}+1$, $k\ge 1$. Since the divisors of $2^r$ are $1,2,2^2,\ldots, 2^r$, by Lemma \ref{lema charpol} with $N=2^r$ we have $p_{_B}(x) = (x-1)^{c_1} (x+1)^{c_2} (x^2+1)^{c_4} (x^4+1)^{c_8} \cdots (x^{2^{r-1}}+1)^{c_{2^r}}$
and hence \eqref{eq. charpol2} follows, with $c_{2^r}\ge 1$.
Relative to the angles, all the $j_{i,k}$'s in Proposition \ref{lema angles} are odd, and hence $B$ is conjugate in $T_{2m}$ to
$x_B = x(t_1,t_2,\ldots,t_m)$ or to $x_B' = x(-t_1,t_2,\ldots,t_m)$ with $x_B$ as in \eqref{xB2r}.
Clearly, we have $n= 2^{r-1} c_{2^r} + \cdots + 4c_8 + 2c_4 + c_2 +(c_1-1) + 1 = \deg p_{_B}$. Since $n_B$ is the multiplicity of 1 as eigenvalue of $B$, and $n_B\ge 1$, we have $c_1\ge 1$.
\end{proof}
\subsection{Trigonometric identities}
The rotation angles of order $2^r$ matrices give rise to interesting non-trivial trigonometric identities involving products and alternating sums of sines and cotangents (see Propositions \ref{lema. cots} and \ref{prop. sumcots}). The following notations will be useful in the sequel. For any positive integer $N$ define $I_N=\{i \in \N: 1 \le i\le N\}$ and
\begin{equation}\label{In}
I_N^* = \{j\in I_N : j \,\mathrm{odd}\}.
\end{equation}
We will need the following technical result.
\begin{lema} \label{lema sns}
Let $k,N\in \N$. If $k$ is odd and $N$ is even then
\begin{equation} \label{sn*}
S_N^*(k) := \sum_{j\in I_{N-1}^*} [\tf{jk}{N}] = \tf{(k-1)N}{4} \in \Z,
\end{equation}
where $[\,\cdot\,]$ is the floor function. In particular, $S_N^*(k) \in 2\Z$ for $N=2^r$, $r\ge 2$.
\end{lema}
\begin{proof}
Let $S_N(k) := \sum_{j=1}^{N-1} [\tf{jk}{N}]$ and note that $S_N(k) = \sum_{j=1}^{N/2} [\tf{jk}{N}] + \sum_{j=1}^{N/2} \big[ \tf{(N-j)k}{N} \big] - [\tf N2 \tf kN]$. Thus, we have $\big[ \tf{(N-j)k}{N} \big] = [k-\tf{jk}{N}] = k+[-\tf{jk}{N}] = k-1-[\tf{jk}{N}]$, since $[x+n]=[x]+n$ for $n\in \Z$ and $[-x]=-1-[x]$ for
$x\not \in \Z$. Hence,
$$S_N(k) = \sum_{1\le j \le N/2} [\tf{jk}{N}]+ \Big( \tf{(k-1)N}2 - \sum_{1\le j \le N/2} [\tf{jk}{N}] \Big) + [\tf k2] =
\tf{(k-1)N}2 + \tf{k-1}2 = \tf{(k-1)(N-1)}2.$$
Now,
$$S_N(k) = S_N^*(k) + S_{\f N2}^*(k) + \cdots + S_4^*(k) + S_2^*(k) = S_N^*(k) + S_{\f N2}(k),$$
by induction, and hence,
$S_N^*(k) = S_N(k) - S_{\f N2}(k) = \tf{(k-1)}{2} \tf N2$, and we are done.
\end{proof}
\begin{prop} \label{prop. sines2}
If $r,k$ are positive integers with $r\ge 2$ and $k$ odd then
\begin{equation} \label{sines2}
\prod_{j\in I_{2^r}^*} \sin \big( \tf{jk\pi}{2^r} \big) = \tf{1}{2^{2^{r-1}-1}}.
\end{equation}
\end{prop}
\begin{proof}
Let $N=2^{r}$ and assume $k$ odd. We will first prove that the product in \eqref{sines2} does not depend on $k$.
For any fixed odd $k$, there are unique integers $q_j, r_j$ such that $jk=q_j N + r_j$ with $0\le r_j \le N-1$. If $j$ is odd, then $r_j$ must be odd, and hence $r_j\ge 1$. Thus,
\begin{equation}\label{signsin}
\sin(\tf{jk\pi}{N}) = \sin(q_j\pi + \tf{r_j \pi}{N}) = (-1)^{q_j} \sin(\tf{r_j \pi}{N}).
\end{equation}
By modularity, for any $k$ odd we have $\{ 1, 3,\ldots, N-1 \} = \{ k, 3k, \ldots, k(N-1)\}$ mod $N$ and hence
$\{ 1, 3,\ldots, N-1 \} = \{r_1,r_3,\ldots,r_{N-1}\}$ mod $N$. Therefore, by \eqref{signsin}
\begin{equation}\label{prodsins}
\prod_{j\in I_{2^r}^*} \sin(\tf{jk\pi}{N}) =
(-1)^{S_N^*(k)} \prod_{j\in I_{2^r}^*} \sin(\tf{r_j\pi}{N}) =
\prod_{j\in I_{2^r}^*} \sin(\tf{j\pi}{N}),
\end{equation}
where we have used that $q_j=[\tf{jk}{N}]$ and $S_N^*(k)$ as in \eqref{sn*} is even by Lemma \ref{lema sns}, since $r\ge 2$.
\sk It only remains to compute the last product in \eqref{prodsins}.
We will fisrt prove
\begin{equation} \label{sines}
\prod_{j\in I_{2^{r-1}}^*} \sin \big( \tf{j \pi }{2^r} \big) = \tf{\sqrt 2}{2^{2^{r-2}}}.
\end{equation}
Assume that $r\ge 2$, the case $r=1$ being trivial. Let us define
$\Pi_{2^r} := \prod_{j=1}^{2^{r-1}} \sin(\tf{j\pi}{2^r})$ and $\Pi_{2^r}^* := \prod_{j\in I_{2^{r-1}}^*} \sin(\tf{j\pi}{2^r})$.
We recall the identity $\prod_{j=1}^{[\f d2]} \sin(\tf{j\pi}d) = \sqrt{d} / 2^{\f {d-1}2}$,
with $d\in \N$ (see \cite{MP.PAMQ09}, Lemma~3.1, for a proof). By taking $d=2^r$ we have that
$\Pi_{2^r} = 2^{\f r2 - \f{2^r-1}2}$. By induction, we have
$$\Pi_{2^r} = \Pi_{2^r}^* \Pi_{2^{r-1}}^* \cdots \Pi_{2^2}^* \Pi_{2^1}^* = \Pi_{2^r}^* \Pi_{2^{r-1}}$$
for $r\ge 2$, and thus
$$\Pi_{2^r}^* = \f{\Pi_{2^r}}{\Pi_{2^{r-1}}} = \f{2^{\f r2 - \f{2^r-1}2}}{2^{\f{r-1}2 - \f{2^{r-1}-1}2}} = \f{\sqrt 2}{2^{2^{r-2}}},$$
as desired.
Now, by symmetry, $\sin(\tf{(N-j)\pi}{N}) = \sin(\tf{j\pi}{N})$, for $0<j<\f N2$. By using this and \eqref{sines} we obtain
$$\prod_{j\in I_{N}^*} \sin(\tf{j\pi}{N}) = \prod_{j\in I_{2^{r-1}}^*} \big( \sin(\tf{j\pi}{2^{r}}) \big)^2 =
(\Pi_{2^{r}}^*)^2 = \big( \tf{\sqrt 2}{2^{2^{r-2}}} \big)^2 = \tf{1}{2^{2^{r-1}-1}},$$
and thus the proposition follows.
\end{proof}
For $r\in \N$ and $t,\omega \in \Z$ define the sums
\begin{equation}\label{S,r,t}
\mathcal{S}_{r,t}(\omega) := \sum_{k\in I_{2^r}^*} \, (-1)^{[\tf k2]} \; \sin(\tf{k \omega \pi}{2^t})
\end{equation}
\begin{prop}\label{prop. sumsines}
Let $r, \nu \in \N$ and $t,\ell \in \Z$ with $\ell$ odd. If $t\le r$ then
\begin{equation}\label{sumsines}
\mathcal{S}_{r,t}(2^\nu \ell) = \left\{ \begin{array}{cl}
(-1)^{[\tf{\ell}{2}]} \, 2^{r-1} & \qquad \text{if } t=\nu +1, \msk \\
0 & \qquad \text{if } t \ne \nu+1. \end{array} \right.
\end{equation}
\end{prop}
\begin{proof}
For $t\le 0$ the result is trivial, thus assume $t\ge 1$. Suppose first that $\nu=0$.
Consider $t=1$. Since for any $k$ odd, $\sin(\tf{k \pi}{2}) = (-1)^{[\tf k2]}$ and
$\sin(\tf{k \ell \pi}{2}) = (-1)^{[\tf{k\ell}{2}]} = (-1)^{[\tf k2]} (-1)^{[\tf{\ell}{2}]}$, with $\ell$ odd, then we have
$$\mathcal{S}_{r,1}(\ell) = \sum_{k\in I_{2^r}^*} (-1)^{[\tf k2]} \, \sin(\tf{k \ell \pi}{2}) =
\sum_{k\in I_{2^r}^*} (-1)^{[\tf{\ell}{2}]} = (-1)^{[\tf{\ell}{2}]} \, 2^{r-1}.$$
Now let $t>1$.
If $t=r$, we have the angles $\tf{\pi}{2^r}, \tf{3\pi}{2^r},\ldots,\tf{(2^r-1) \pi}{2^r}$ in $[0,\pi]$, and we compare the contributions of $\tf{k\ell \pi}{2^r}$ and $\tf{(2^r-k) \ell \pi }{2^r}$. Since for any positive odd integer $k$ we have
$\sin(\tf{(2^r-k)\ell\pi}{2^r}) = (-1)^{\ell+1} \sin(\tf{k\ell \pi}{2^r})$
and $(-1)^{[\f{2^r-k}{2}]} = (-1)^{[2^{r-1} - \f k2]} = (-1)^{[-\tf{k}{2}]} = (-1)^{[\tf k2]+1}$,
we get
$$(-1)^{[\f{2^r-k}{2}]} \sin(\tf{(2^r-k)\pi}{2^r}) = (-1)^\ell (-1)^{[\f{k}{2}]} \sin(\tf{k \ell \pi}{2^r})$$
and thus, since $\ell$ is odd, the contributions of the angles $\tf{k\pi}{2^r}$ and $\tf{(2^r-k)\pi}{2^r}$ in \eqref{sumsines} cancel each other out.
If $t<r$, there are more angles to consider. However, by modularity, it is enough to consider the angles
$$\tf{\pi}{2^t}, \tf{3\pi}{2^t},\ldots,\tf{(2^t-1) \pi}{2^t},\tf{(2^t+1)\pi }{2^t},\tf{(2^t+3)\pi }{2^t},\ldots,\tf{(2^{t+1}-1) \pi }{2^t}$$
in the interval $(0,2\pi)$. In this case, we compare the angle $\tf{k\ell\pi}{2^t}$ with $\tf{(2^t+k)\ell\pi}{2^t}$, for any $k=1,3,5,\ldots,2^t-1$. Since $\sin(\theta + \pi) = -\sin(\theta)$ and $(-1)^{[\f{2^t+k}{2}]} = (-1)^{[\f k2]}$, we see again that the contributions of $\tf{k\ell\pi}{2^t}$ and $\tf{(2^t+k)\ell\pi}{2^t}$ cancel out.
In this way, $\mathcal{S}_{r,t}(\ell)=0$ for $t>1$.
Now, consider the case $\nu \ge 1$. Clearly, $\mathcal{S}_{r,t}(2^\nu \ell)=0$ for $\nu\ge t$. For $1\le \nu \le t-1$, note that
$\mathcal{S}_{r,t}(2\ell)=\mathcal{S}_{r,t-1}(\ell)$. Hence, by induction,
$$\mathcal{S}_{r,t}(2^\nu \ell)=\mathcal{S}_{r,t-\nu}(\ell)=\delta_{t,\nu+1} \, (-1)^{[\tf{\ell}{2}]} \, 2^{r-1},$$
where $\delta$ is the Kronecker function, and thus \eqref{sumsines} holds.
\end{proof}
\section{Eta function and $\eta$-invariant} \label{sec etas}
Let $\G$ be an $n$-dimensional Bieberbach group with translation lattice $\Lambda$ and point group $F$ and
let $\g=BL_b\in \G$. If $x_B = x(t_1,\ldots,t_m) \in T_{n-1}$
is conjugate to $B$, denote the angles $t_1,\ldots,t_m$ of $B$ by $t_1(x_B), \ldots,t_m(x_B)$ and put
$$(\Lambda \backslash \G)' =\{ BL_b\in \Lambda \backslash \G : B\in F_1'\}$$
where $F_1' = \{B \in F : n_B=1 \text{ and } t_i(x_B) \not \in \pi\Z, \: 1\le i \le m \}$.
If $B\in F_1$, choose $v_B \in \Lambda^*$ such that
\begin{equation}\label{vB}
(\Lambda^*)^B = \Z v_B \qquad \text{and put} \qquad \lambda_{_B}=\|v_B\|.
\end{equation}
Then, if $o(B)$ denotes the order of $B$, we have (see \cite{MP.JGA11}, Lemma 4.1)
\begin{equation}\label{elegama}
\ell_\gamma := o(B) \langle v_B, b \rangle \in \Z.
\end{equation}
\sk
We now recall the expressions of $\eta(s)$ for $\Z_N$-manifolds, that will be our starting point.
\begin{prop}[\cite{MP.JGA11}, Prop.\@ 2.2] \label{AGAG}
Let $M_\Gamma$ be an orientable $\Z_N$-manifold of dimension $n = 2m+1$, with $m$ odd, and $\Gamma = \langle \gamma , L_\Lambda \rangle$ where $\gamma = BL_b$.
If $r=1$ or $n_B > 1$ then $\eta(s)=0$. On the contrary, if $r\ge 2$ and $n_B =1$ then
\begin{equation} \label{etas cyclic}
\eta(s) = -\sigma_{v_B} \, \tf{2^{m+1}}{N(2 \pi \lambda_{_B})^s} \sum_{\substack{k=1 \sk \\ \g^k \in (\Ld \backslash \G)' }}^{N-1} \, \tf{1}{o(B^k)^s} \, \Big( \prod_{j=1}^m \sin (k t_j) \Big) \, \sum_{j=1}^{o(B^k)-1} \sin \big( \tfrac{2\pi j k \ell_\gamma}{N} \big) \, \zeta(s, \tfrac{j}{o(B^k)})
\end{equation}
where $v_B$ and $\ell_\g$ are as in \eqref{vB} and \eqref{elegama} respectively, and
\begin{equation} \label{cyclic eta0}
\eta = -\sigma_{v_B}\, \tf{2^{m}}{N} \sum_{k=1}^{N-1} \sum_{\g^k \in (\Ld \backslash \G)'}
\big( \prod_{j=1}^m \sin (k t_j) \big) \cot(\tf{\pi k \ell_\g}{N}),
\end{equation}
where $\sigma_{v_B} \in \{\pm 1\}$ is a sign depending on the conjugacy class of $x_B$ in $T_{n-1}$.
\end{prop}
\noindent
\textit{Note.} See \cite{MP.JGA11}, (3.10)--(3.11), for details on the sign $\sigma_{v_B}$.
\subsection{Eta function}
We now give the eta function of $\mathcal{D}$ for an arbitrary $\Z_{2^r}$-manifold $M_\G$.
We recall here that the class of ${\Z_{2^r}}$-manifolds are not classified, since there is no such classification for the integral representations of $\Z_{2^r}$, for $r\ge 3$.
However, all the information needed to compute $\eta(s)$ is (roughly) contained in the angles of the rotational part of the elements of $\G$.
\sk
We will express $\eta(s)$ in terms of a Dirichlet $L$-function. We recall that for $\chi : \Z \rightarrow \C^*$ a Dirichlet character modulo $N$ one has the Dirichlet series $L(s,\chi) = \sum_{n=1}^\infty \f{\chi(n)}{n^s}$, absolutely convergent for $\mathrm{Re} \, (s) >1$. It has an analytic continuation to the whole $\C$ given by
\begin{equation}\label{L&zetas}
L(s,\chi) = \tf{1}{N^s}\sum_{j=1}^N \chi(j) \, \zeta(s, \tf jN) \,.
\end{equation}
We will be concerned with $\chi_4$, the \textit{primitive} character mod $4$, defined by
$\chi_4(1) = 1$, $\chi_4(3) = -1$ and $\chi_4(0) = \chi_4(2) = 0$, that is
$\chi_4(n) = \sin(\tf{\pi n}{2})$ for $n\in \Z$.
\msk
The promised result is as follows.
\begin{teo} \label{teo.etasz2r}
Let $M_\G$ be any orientable $\Z_{2^r}$-manifold of dimension $n \equiv 3$ mod 4, with $\G= \langle \g= BL_b, L_\Lambda \rangle$. For $\ell_\g$ as in \eqref{elegama} put $\ell_\g = 2^\nu \ell$ with $\ell$ odd and $\nu \ge 0$. If $n_B>1$ or $r=1$, then $\eta(s)=0$. If $r\ge 2$ and $n_B=1$ then Spec$_\mathcal{D}(M_\G)$ is asymmetric and we have
\begin{equation}\label{generaletas}
\eta(s) = \tf{\sigma_\g \, 2^{f(B)-1}}{(2^{r+1-\nu} \pi \lambda_{_B})^{s}} \{ \zeta(s,\tf 14) - \zeta(s,\tf 34) \}
= \tf{\sigma_\g\, 2^{f(B)-1}}{(2^{r-1-\nu} \pi \lambda_{_B})^{s}} \, L(s,\chi_4)
\end{equation}
with $0\le \nu \le r-2$ and $\sigma_\g = \sigma_{v_B} \, (-1)^{[\f{\ell}{2}]+1} \in \{\pm 1\}$, where $\chi_4$ is the primitive Dirichlet character mod 4 and $f(B)$, $\lambda_B$ are as in \eqref{fb}, \eqref{vB} respectively.
\end{teo}
\begin{proof}
We will apply Proposition \ref{AGAG}. It is known that $\Z_2$-manifolds have symmetric spectrum and hence trivial $\eta(s)$ (\cite{MP.JGA11}, Proposition 3.11, or \S 6.1.1 in \cite{MP.AGAG12}). Also, $\eta(s)=0$ if $n_B>1$. Thus, from now on, we assume that
$r\ge 2$ and $n_B=1$. Let $n=2m+1$, $m$ odd.
We now study the ingredients in \eqref{etas cyclic}. The angles are given by \eqref{xB2r}.
Also, $n_B=1$ implies $c_1=1$ and $\g^k \in (\Lambda \backslash \G)'$ if and only if $k$ is odd and $c_2=0$. Thus,
\begin{equation}\label{xB2r0}
x_B = x(\underbrace{ \tf{\pi}{2^{r-1}}, \tf{3\pi}{2^{r-1}}, \ldots, \tf{(2^{r-1}-1)\pi}{2^{r-1}} }_{c_{2^r}}, \ldots, \underbrace{ \tf{\pi}{8}, \tf{3\pi}{8}, \tf{5\pi}{8}, \tf{7\pi}{8} }_{c_{16}}, \underbrace{ \tf{\pi}{4}, \tf{3\pi}{4} }_{c_8}, \underbrace{ \tf{\pi}{2} }_{c_4}).
\end{equation}
Note that $o(B^k)=2^r$ for every $k$ odd. Relative to the dimension, we have that
\begin{equation}\label{mB}
m = 2^{r-2} c_{2^r} + \cdots + 4 c_{16} + 2 c_8 + c_4.
\end{equation}
By Corollary \ref{coroz2r}, we know that $c_{2^r} \ge 1$. The condition $m$ odd implies that $c_4$ is also odd, and, in particular, $c_4 \ge 1$.
By putting all this information in \eqref{etas cyclic} we get
\begin{equation} \label{etaux1}
\eta(s) = - \tf{\sigma_{v_B} \, 2^{m+1-r}}{(2^{r+1}\pi \|v_B\|)^s} \, \sum_{k\in I_{2^r}^*}
\, P_k(x_B) \, \sum_{j=1}^{2^r-1} \sin \big( \tf{\pi jk\ell_\g}{2^{r-1}} \big) \, \zeta(s,\tf{j}{2^r}) \,,
\end{equation}
where we have used \eqref{In} and the notation $P_k(x_B) = \prod_{j=1}^m \sin(kt_j)$.
By \eqref{xB2r}, we have
$$ P_k(x_B) = \Bigg( \prod_{\substack{j=1 \sk \\ j \text{ odd}}}^{2^{r-1}}
\sin (\tf{jk\pi}{2^{r-1}}) \Bigg)^{c_{2^r}}
\Bigg( \prod_{\substack{j=1 \sk \\ j \text{ odd}}}^{2^{r-2}}
\sin (\tf{jk\pi}{2^{r-2}}) \Bigg)^{c_{2^{r-1}}} \cdots \Bigg( \prod_{\substack{j=1 \sk \\ j \text{ odd}}}^{2^{2}}
\sin (\tf{jk\pi}{2^{2}}) \Bigg)^{c_8} \big( \sin (\tf{k\pi}{2})\big)^{c_4}.$$
For $r=2$, we have
$P_k(x_B) = (\sin (\tf{k\pi}{2}))^{c_4} = (-1)^{[\tf k2]}$, since $c_4$ is odd.
For $r\ge 3$, by Proposition \ref{prop. sines2}, we get (this is a first key step)
$$P_k(x_B) = (-1)^{[\tf k2]} \, \prod_{i=2}^{r-1} \Bigg( \prod_{j\in I_{2^i}^*}
\sin (\tf{jk\pi}{2^{i}}) \Bigg)^{c_{2^{i+1}}} =
(-1)^{[\tf k2]} \, \prod_{i=2}^{r-1} \big( \tf{1}{2^{2^{i-1}-1}} \big)^{c_{2^{i+1}}} = (-1)^{[\tf k2]} \, 2^{-S_r},$$
where we have put $S_r := \sum\limits_{i=2}^{r-1} (2^{i-1}-1) \, c_{2^{i+1}}$.
By \eqref{mB} and \eqref{fb}, since $c_1=1$,
$c_2=0$, we have
$$S_r = \sum_{i=2}^{r-1} 2^{i-1} c_{2^{i+1}} - \sum_{i=2}^{r-1} c_{2^{i+1}} = (m-c_4) - (f(B')-c_4) = m-f(B)+1.$$
Therefore, we get
$P_k(x_B) = (-1)^{[\tf k2]} \, 2^{-m + f(B) -1}$ and putting this in \eqref{etaux1} we obtain
\begin{equation}
\label{etaux2}
\eta(s) = - \tf{\sigma_{v_B} \, 2^{f(B)-r}}{(2^{r+1}\pi \|v_B\|)^s} \, \sum_{k\in I_{2^r}^*}
(-1)^{[\tf k2]} \underbrace{\sum_{j=1}^{2^r-1} \sin \big( \tf{\pi jk\ell_\g}{2^{r-1}} \big) \, \zeta(s,\tf{j}{2^r})}_{:= \xi_{k,\ell_\g,r}(s)}.
\end{equation}
Now, since $\sin(\tf{(2^r-j)k\pi}{2^{r-1}}) = -\sin(\tf{jk\pi}{2^{r-1}})$ for every $j,k, \in \N$, we can write
\begin{eqnarray*}
\xi_{k,\ell_\g,r}(s) &=&
\sum_{j=1}^{2^{r-1}-1} \sin \big( \tf{\pi jk\ell_\g}{2^{r-1}} \big) \, \big\{ \zeta(s,\tf{j}{2^r}) - \zeta(s,1-\tf{j}{2^r}) \} \\
& = & \sum_{t=2}^r \sum_{j\in I_{2^{t-1}}^*}
\sin \big( \tf{\pi jk\ell_\g}{2^{t-1}} \big) \, \big\{ \zeta(s,\tf{j}{2^t}) - \zeta(s,1-\tf{j}{2^t}) \},
\end{eqnarray*}
where we have put together the contribution of the angles of the $N$-th roots of unity of the same order $2^t$, $1\le t \le r-1$.
In this way, the sum in \eqref{etaux2} becomes
\begin{eqnarray*}
\sum_{k\in I_{2^r}^*}
(-1)^{[\tf k2]} \, \xi_{k,\ell_\g,r}(s) &=& \sum_{t=2}^r \sum_{j\in I_{2^{t-1}}^*}
\Bigg( \sum_{k\in I_{2^r}^*} (-1)^{[\tf k2]} \, \sin \big( \tf{\pi jk\ell_\g}{2^{t-1}} \big) \Bigg) \,
\big\{ \zeta(s,\tf{j}{2^t}) - \zeta(s,1-\tf{j}{2^t}) \big \}.
\end{eqnarray*}
By Proposition \ref{prop. sumsines}, the sum between parentheses equals
$$\mathcal{S}_{r,t-1}(2^\nu j\ell) = \delta_{t-1,\nu+1}(-1)^{[\f{j\ell}{2}]} 2^{r-1}$$
for $t\le r$ (where $\ell_\gamma=2^\nu \ell$ with $\ell$ odd); i.e.\@ it does not vanish only for $t=\nu+2$ and hence $0\le \nu \le r-2$
(this is a second key step).
Thus, since $(-1)^{[\tf{j\ell}{2}]} = (-1)^{[\tf{j}{2}]} (-1)^{[\tf{\ell}{2}]}$, we get
$$\sum_{k\in I_{2^r}^*} (-1)^{[\tf k2]} \, \xi_{k,\ell_\g,r}(s) = (-1)^{[\tf{\ell}2]} \, 2^{r-1} \, \sum_{j\in I_{2^{\nu+1}}^*} (-1)^{[\tf j2]} \, \big\{ \zeta(s,\tf{j}{2^{\nu+2}}) - \zeta(s,1-\tf{j}{2^{\nu+2}}) \big\}.$$
Putting this information in \eqref{etaux2}, we get expression
\begin{equation} \label{preetas}
\eta(s) = \sigma_\g \, \f{ 2^{f(B)-1}}{(2^{r+1}\pi \lambda_{_B})^{s}} \, \sum_{j \in I_{ 2^{\nu+1}}^*} (-1)^{[\tf j2]}
\big( \zeta(s, \tf{j}{2^{\nu+2}}) - \zeta(s,1-\tf{j}{2^{\nu+2}}) \big)
\end{equation}
where $\sigma_\gamma =-\sigma_{v_B} (-1)^{\f{\ell}{2}}$.
\msk
Now, denote by $F_\nu(s)$ the function given by the sum in \eqref{preetas}, i.e.\@
$$F_\nu(s) = \sum_{j \in I_{ 2^{\nu+1}}^*} (-1)^{[\tf j2]}
\big( \zeta(s, \tf{j}{2^{\nu+2}}) - \zeta(s,1-\tf{j}{2^{\nu+2}}) \big).$$
Note that $L(s,\chi_4)=\tf{1}{4^s} \big( \zeta(s,\tf14) - \zeta(s,\tf34) \big)$ by \eqref{L&zetas}.
Thus, $F_0(s)= 4^s L(s,\chi_4)$ and hence \eqref{generaletas} holds for $\nu=0$ (i.e.\@ $t=2$).
We now consider the remaining cases, that is $\nu>0$ (i.e.\@ $t\ge 3$). Suppose $\chi$ is a Dirichlet character mod $k$. Then, there is some primitive Dirichlet character $\psi$ mod $d$, with $d\mid k$ (the conductor), such that $\chi=\psi \cdot \chi_{1,k}$, where $\chi_{1,k}$ is the principal Dirichlet character mod $k$, and
\begin{equation}\label{L-ind}
L(s,\chi) = L(s,\psi) \prod_{p\mid k} (1 - \tf{\chi(p)}{p^s}).
\end{equation}
Now, for $t \ge 3$, let $\chi_{2^t}$ be the real Dirichlet character mod $k=2^t$ induced by the primitive charater $\chi_4$, that is
$\chi_{2^t}=\chi_4 \cdot \chi_{1,2^t}$. Thus, $\chi_{2^t}(2j)=0$ and $\chi_{2^t}(2j+1)=(-1)^{[\f j2]}$.
On one hand, since $k=2^t$, by \eqref{L-ind} we have
\begin{equation}\label{L1}
L(s,\chi_{2^t}) = L(s,\chi_4) (1 - \tf{\chi_4(2)}{2^s}) = L(s,\chi_4).
\end{equation}
On the other hand, since $\chi_{2^t}(2^t-j) = \chi_{2^t}(-1)\chi_{2^t}(j)=-\chi_{2^t}(j)$, by \eqref{L&zetas} we have
\begin{equation} \label{L2}
\begin{split}
L(s,\chi_{2^t})
= \tf{1}{(2^t)^s} \sum_{j\in I_{2^{t-1}}^* } (-1)^{[\f j2]} \, \big( \zeta(s,\tf{j}{2^t}) - \zeta(s,1-\tf{j}{2^t}) \big)
= \tf{1}{(2^t)^s} F_{t-2}(s).
\end{split}
\end{equation}
From \eqref{L1} and \eqref{L2}, we have
$$F_\nu(s) = 2^{\nu s} L(s,\chi_{2^{\nu+2}}) = 2^{\nu s} L(s,\chi_4),$$
with $t=\nu+2$, and therefore \eqref{generaletas} holds for any $0\le \nu \le r-2$.
\sk
Finally, note that by \eqref{Eta(s)} we have
$\eta(s) = \sum_{\lambda>0} (d_\lambda^+ - d_{\lambda}^-) \lambda^{-s}$,
where $d_{\ld}^{\pm}$ is the multiplicity of $\pm \ld$. Therefore, symmetry in the spectrum implies $\eta(s) = 0$ for every $s$, a contradiction. Thus the spectrum is asymmetric and the result follows.
\end{proof}
Notice that, by \eqref{etas cyclic}, the dependence of $\eta(s)$ on the metric is given by the numbers $\nu \in \N_0$ (with $\ell_\g =
2^\nu \ell$, $\ell$ odd) and $\lambda_B=\|v_B\|\in \R$ where $(\Ld^*)^B = \Z v_B$ and $\g=BL_b$ is the generator of
$\Ld\backslash \G$.
\begin{rem}
Expression \eqref{generaletas} for $\eta(s)$ in Theorem \ref{teo.etasz2r} is very simple, compare with the general expression \eqref{etas cyclic}. By \eqref{etaux2}, a priori, all the functions $\zeta(s,\tf{1}{2^t}), \zeta(s,\tf{3}{2^t}), \ldots, \zeta(s,\tf{2^t-1}{2^t})$ should appear in $\eta(s)$, for every $2\le t \le r$. However, as the proof of the theorem shows, great cancellations take place and Hurwitz zeta functions for only one $t$, namely
$\zeta(s,\tf{j}{2^{\nu+2}})$ for $j=1,3,5, \ldots, 2^{\nu+2}-1$, contribute to $\eta(s)$.
\end{rem}
\begin{rem}\label{reti}
Let $M_\G$ be an orientable $\Z_{2^r}$-manifold, $r\ge 2$, of dimension $n=2m+1$, $m$ odd, with $\G= \langle BL_b, L_\Lambda \rangle$ and $n_B=1$. Suppose $B=\mathrm{diag}(B',1)$ with $B'\in \mathrm{SO}(n-1)$ and
\begin{equation}\label{reticulin}
\Lambda = \Ld' \bigoplus^\perp \Z e_n
\end{equation}
with $\Ld'$ a $B'$-stable lattice in $\R^{n-1}$. In this case, $\ell_\g=\ell$ is odd, that is $\nu=0$, for if not $\G$ would have torsion elements other than the identity. This is the case, for instance, for the tetracosm $M_1$ (see \eqref{tetracosmo}), the $\Z_{2^r}$-manifolds in family $\mathcal{F}$ (see Section \ref{sec.family}) and for the exceptional $\Z_p$-manifolds, $p$ odd prime (see \cite{GMP}, Proposition 2.2). This is not however the case in general, as Example \ref{nonstandard} below shows.
\end{rem}
\begin{rem} \label{rem 35}
By (the proof of) Theorem \ref{teo.etasz2r} and its previous results, the eta function of an $n$-dimensional $\Z_{2^r}$-manifold ($n=4h-1$) is non-trivial if and only if $r\ge 2$ and $c_1=1$, $c_2=0$, $c_{2^r}, c_4\ge 1$ with $c_4$ odd.
Hence, by \eqref{eq. charpol},
$$p_{_B}(x) = \Phi_1(x) \Phi_4 (x)^{c_4} \Phi_{2^r}(x)^{c_{2^r}} g(x)$$
where $g(x) =\prod_{i=3}^{r-1} \Phi_{2^i}(x)^{c_{2^i}}$ with $c_{2^i}\ge 0$, $i=3,\ldots,r-1$.
\end{rem}
\begin{defi} \label{def nr}
For any fixed $r\ge 2$, let $n_r$ be the \emph{minimal dimension} for a $\Z_{2^r}$-manifold having non-trivial eta function.
\end{defi}
Thus, $n_2=3$ for $\Z_4$-manifolds and for $r\ge 3$ we have
\begin{equation}\label{nr}
n_r = \min\limits_{B\in G_{n,r}^*} \{ \deg p_{_B} \} = 2^{r-1}+3, \qquad r\ge 3,
\end{equation}
where $G_{n,r}^* = \{B\in \mathrm{GL}(n) : o(B) = 2^r \text{ and $B$ has no $(-1)$-eigenvalues} \}$, corresponding to the decomposition $p_{_B}(x) = \Phi_1(x) \Phi_4 (x) \Phi_{2^r}(x) = (x-1)(x^2+1)(x^{2^{r-1}}+1)$ in the above remark,
that is $g(x)=1$ (i.e.\@ $c_8=c_{16}=\cdots = c_{2^{r-1}} =0$).
In the next section we will show that $\Z_{2^r}$-manifolds having non-trivial eta function exist for every $r\ge 2$ and in every dimension $n\ge n_r+4k$, $k\ge0$.
\subsection{Eta invariant and examples}
As a direct consequence of Theorem \ref{teo.etasz2r} we obtain the $\eta$-invariant of any $n$-dimensional $\Z_{2^r}$-manifold. It is, up to a sign, a positive power of 2 which does not depend on $n$ nor on $r$, but on the nature of the integral holonomy representation only.
\begin{prop} \label{prop eta0}
Let $M_\G$ be a $\Z_{2^r}$-manifold of dimension $n\equiv 3$ mod 4 with $F= \langle B \rangle$ and put $\ell_\g=2^\nu \ell$ with $0\le \nu \le r-2$ and $\ell$ odd. If $n_B>1$ or $r=1$, then $\eta=0$. If $n_B=1$ and $r\ge 2$, then
\begin{equation}\label{generaleta0}
\eta(M_\G) = \sigma_\g \, 2^{f(B)-2} \in \Z,
\end{equation}
where $\sigma_\g=-\sigma_{v_B} \, (-1)^{[\f{\ell}{2}]} \in \{\pm 1\}$
and $f(B)$ is defined in \eqref{fb}.
\end{prop}
\begin{proof}
By \eqref{generaletas}, we can get the $\eta$-invariant simply by evaluation of $\eta(s)$ at $s=0$, using that $\zeta(0,a) = \tf12 -a$.
Thus,
$$\eta(M_\G) = \eta(0) = \sigma_\g \, 2^{f(B)-1} \{(\tf 12 - \tf 14) - (\tf12-\tf 34)\} = \sigma_\g \, 2^{f(B)-2}.$$
We have $c_{2^r}, c_4 \ge 1$, by the comments after \eqref{mB}, and hence $f(B)\ge 2$. Thus, $\eta(M_\G) \in \Z$, and the result follows.
\end{proof}
\begin{rem}
(i) Expression \eqref{generaleta0} is very simple (compare with the general formula \eqref{cyclic eta0}).
\sk (ii) The $\eta$-invariant is, up to sign, determined by $f(B)$. To compute this number, one only needs to know the integral holonomy representation. In given examples, one can use some mathematical software to compute the characteristic polynomial $p_{_B}(x)$. To determine the sign, one needs to know $\g$ and $\Ld$ explicitly.
\sk (iii) Clearly, $\eta(s)$ determines the $\eta$-invariant. The converse is not true in general, because of the dependence on $\nu$ and $\lambda_B$ in \eqref{generaleta0}. However, $\eta(s)=0$ if and only if $\eta=0$.
\sk (iv) By \eqref{generaleta0}, the \textit{reduced eta invariant} $\bar \eta = \tf12 (\eta+d_0)$ mod $\Z$ of any $\Z_{2^r}$-manifold is either 0 or $\tf12$. Also, $d_0 = \dim \ker \mathcal{D}$ can be computed by using Theorem 3.5 and Proposition~3.7 in \cite{MP.JGA11}.
\end{rem}
\msk
In dimension 3, there are $\Z_{2^r}$-manifolds for $r=1,2$ only. Up to diffeomorphism, there are three $\Z_2$-manifolds and there is only one $\Z_4$-manifold, given by
\begin{equation}\label{tetracosmo}
M_1 = \G \backslash \R^3, \quad \G = \langle \g = BL_{\tf{e_3}{4}}, L_{\Z^3}\rangle, \quad
B = \mathrm{diag}(J_1,1), \quad J_1=\left( \begin{smallmatrix} & -1 \\ 1 & \end{smallmatrix} \right).
\end{equation}
This manifold, known as the \emph{tetracosm} after \cite{RC}, gives a nice example in spectral geometry, being one of the `spectral twins', i.e.\@ the only two isospectral-on-functions and non-isometric compact 3-manifolds (\cite{DR}).
\begin{coro} \label{coro eta0}
Let $M_\G$ be a $\Z_{2^r}$-manifold of dimension $n\equiv 3$ mod 4. If $n \ge 7$ then $\eta \in 2\Z$. If $n=3$ then
$\eta = 0$ for $\Z_2$-manifolds and $\eta = \pm 1$ for $\Z_4$-manifolds.
\end{coro}
\begin{proof}
We know that $f(B)\ge 2$ and, by \eqref{generaleta0}, $\eta(M_\G) \in 2\Z$ if and only if $f(B)\ge 3$. By Proposition \ref{coroz2r} we have $c_{2^r}, c_1\ge 1$. By Remark \ref{rem 35}, $\eta\ne 0$ if and only if $r\ge 2$ and $c_1=0$, $c_2=0$, $c_4,c_{2^r}\ge 1$.
Thus, in dimension 7, if $\eta \ne 0$, we have $c_1, c_4, c_{2^r}\ge 1$ with $r>2$, and hence $\eta\in 2\Z$.
On the other hand, $f(B)=2$ if and only if $r=2$ and $c_4=1$ (i.e.\@ $r=2$, $m=1$) and thus $M_\G$ is a 3-dimensional $\Z_4$-manifold.
Thus, $M_\G$ is diffeomorphic to the tetracosm $M_1$ defined in \eqref{tetracosmo}. It is known that $\eta(M_1) = -1$ (see \cite[\S 3.1]{MP.AGAG12}).
Since the $\eta$-invariant is preserved by diffeomorphisms up to sign, and this sign changes with a change of orientation (see \cite{APS1}), the result readily follows.
\end{proof}
Next, we will illustrate with two examples the method for computing $\eta$-invariants given by Proposition \ref{prop eta0}.
\begin{ejem}[\emph{tetracosm}]
\label{tetracosm}
We now compute the $\eta(s)$ and $\eta$-invariant of $M_1$ (see \eqref{tetracosmo}).
\sk (i) Note that $p_{_B}(x) = (x-1)(x^2+1)$, hence $f(B)=2$ and
$\eta(M_1) = -\sigma_B(-1)^{[\ell_\g/2]}$. Also, $v_B=e_3$ and $x_B=x(\tf{\pi}{2})$, then $\sigma_B=1$ and $\ell_\g=1$.
Thus $\eta(M_1)=-1$ and
$$\eta(s) = \tf{-2}{(8\pi)^s} \, \{\zeta(s,\tf14)-\zeta(s,\tf34)\}.$$
\sk (ii) Suppose now that we take $\g^3 = B^3L_{\f34 e_3}$ as the generator of $\G$. Since
$B^3 = \mathrm{diag}(-J_1,1)$, we have $x_{B^3} = x(-\tf{\pi}{2})$, hence $\sigma_{B^3} = -\sigma_B=-1$, and $\ell_{\g^3}=3$. Thus,
we also get $\eta(M_1)= -1$. On the other hand, let $M_1' = \G' \backslash \R^3$, with $\G' = \langle \g' = B^3 L_{\f14 e_3}, L_{\Z^3}\rangle$.
We have $\sigma_{B^3}=-1$ and $\ell_\g = 1$ and thus $\eta(M_1')= 1$. Taking $(\g')^3=(B^3)^3L_{\f34 e_3}=BL_{\f34 e_3}$ as the generator of $\G'$, we also get $\eta(M_1')=1$.
\sk (iii) Any diffeomorphism between cfm's is given by conjugation of the corresponding Bieberbach groups by an element in the affine group. Suppose that $C(BL_b)C^{-1} = B'L_b$ with $C \in \mathrm{GL}(3)$. Then, $CBC^{-1}L_{Cb}= B^3L_b$, i.e.\@ $CBC^{-1}=B^3$ and $Cb=b$. Then, one can take $C = \mathrm{diag}(1,-1,1)$ and thus
$\varphi_{_C} (M_1) = C \G C^{-1} \backslash \R^3 = \G' \backslash \R^3 = M_1'$.
Since $\det C = -1$, $\varphi_{_C}$ is an orientation reversing diffeomorphism between $M_1$ and $M_1'$, and $M_1' = M_1^-$, the tetracosm with the opposite orientation. We saw in (i) that $\eta(M_1)=-\eta(M_1^-)$.
\end{ejem}
In Section \ref{sec.family} we will define a family of $\Z_{2^r}$-manifolds (with $\eta \ne 0$) having integral holonomy representations of a special kind (see \eqref{Cr}, \eqref{Bj}), that we will refer to as \textit{regular $\Z_{2^r}$-representations}.
It is difficult, in general, to construct non-regular $\Z_{2^r}$-representations. One way to do that, is to look up at $\Z_{2^r}$-manifolds in the classification of low dimensional cfm's ($\dim \le 7$) in \textsc{carat} \cite{carat}, and assemble some different representations together (taking some care with the lattices). However, the resulting associated $\Z_{2^r}$-manifold will have $\eta=0$, in general.
In the next example we construct a $\Z_8$-manifold having non-regular integral representation with $\eta \ne 0$.
\begin{ejem}[\textit{$\Z_8$-manifold, integral holonomy representation, $\eta \ne 0$}]
\label{nonstandard}
Consider the matrix $\tilde B = \mathrm{diag}(K,J_1) \in \mathrm{SL}(5,\Z)\times \mathrm{SO}(2,\Z) \subset \mathrm{GL}(7,\Z) $, where
$$K = {\tiny \left( \begin{array}{rrrrr}
1 & 0 & 0 & 1 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & -1& 0 \\
-1 & 0 & 0 & 0 &-1 \\
-1 &-1 & 0 & -1& 0
\end{array} \right)} \qquad \text{and} \qquad J_1 = {\scriptsize \left( \begin{array}{rr} 0&-1\\ 1&0 \end{array} \right)}.$$
It is immediate to check that $K$ has order 8, hence with eigenvalues $\pm e^{\f{\pi i}{4}}, \pm e^{\f{3\pi i}{4}}, 1$ and that $J_1$ has order 4 with eigenvalues $\pm i$. Take the lattice $\Ld = \Ld_5 \oplus \Ld_2 \subset \R^7$, where $\Ld_5 = \Z f_1 \oplus \cdots \oplus \Z f_5$ is $K$-stable and $\Ld_2 = \Z e_6 \oplus \Z e_7$. Also, one checks that $n_{\tilde B} = n_K=1$ with
$\Lambda^{\tilde B} = \Lambda^K = \Z (f_1-f_5)$.
We claim that
$$\tilde \G = \langle \tilde \g = \tilde B L_{\f12 f_1}, L_{\Ld}\rangle$$
is a discrete cocompact torsion-free subgroup of $\mathrm{Aff}(\R^7)$.
In fact, the lattice $\Ld_5$ exists, since $K$ appears as a subrepresentation of the 6-dimensional Bieberbach group with point group
$\Z_8$ given by the $\Z$-class labelled 468.1.2 in \cite{carat}.
Also, by looking at the Biberbach group with holonomy group $\Z_2 \times \Z_8$ given by the $\Z$-class labelled 4407.1.3,
one deduces that $\tf12 f_1$ can be used as translation vector for $K$ (and $\tilde B$), hence giving rise to a torsion-free group.
Since $\tilde B$ is conjugate in $\mathrm{GL}(7)$ to
$B = \mathrm{diag}(x(\tf{\pi}{4}),x(\tf{3\pi}{4}),x(\tf{\pi}{2}),1) \in T_7 \subset \mathrm{SO}(7)$ (or to $B' = \mathrm{diag}(x(\tf{-\pi}{4}),x(\tf{3\pi}{4}),x(\tf{\pi}{2}),1)$, hence $\sigma_{v_B} = 1$ or $-1$, respectively), there exists $C\in \mathrm{GL}(7)$ such that $C\tilde B C^{-1}= B$. In this way, we have that
$$ \G = C\tilde \G C^{-1} = \langle \g = \tilde B L_{\f{1}{2}C f_1}, L_{C\Ld} \rangle \subset \mathrm{I}^+(\R^7) \,,$$
and hence $\tilde \G$ is a 7-dimensional Biberbach group. It is thus clear that $M = \tilde \G \backslash \R^7$ is an orientable $\Z_8$-manifold having integral holonomy representation given by the matrix $B$.
Now, one has that $f(B)=3$, since
$$p_B(x) = p_K(x) p_{J_1}(x) = \big((x-1)(x^4+1)\big) (x^2+1) = \Phi_1(x)\Phi_4(x)\Phi_8(x),$$
and $\eta(M) = \pm 2$, by \eqref{generaleta0}. The sign $\sigma_\g=\pm 1$ can be determined provided one knows $C$ and $\Ld_5$ explicitly.
\end{ejem}
\begin{rem}
In \cite{Sz2}, by using results in \cite{Do}, the $\eta$-invariant of 7-dimensional cfm's $M$ having cyclic holonomy group with a special holonomy representation are computed.
The expression for $\eta(M)$ involves sums of products of cotangents at special angles.
For such $M$, Theorem 1 claims that $\eta(M)\in \Z$. However, the integrality of $\eta$ comes out after computations with `Mathematica'.
There are 126 non-diffeomorphic $\Z_{2^r}$-manifolds in this family (roughly $\tf 13$ of the total), involving only $r=1,2,3$.
Proposition~\ref{prop eta0} assures that indeed $\eta \in \Z$ for these $\Z_{2^r}$-manifolds, and also allows one to compute the $\eta$-invariant in the cases not covered by the mentioned theorem (see the table before Ex.\@ 2 in \cite{Sz2}).
\end{rem}
\section{A distinguished family of $\Z_{2^r}$-manifolds} \label{sec.family}
\subsection{The family $\mathcal{F}$}
For any $r \in \N$, we will construct an infinite family of orientable $\Z_{2^r}$-manifolds in dimensions $n=4h-1$, $h\ge 1$,
each having irreducible integral holonomy representation with $F \subset \mathrm{SO}(n,\Z)$.
For $r\in \N$, let $I_r$ be the $r \times r$ identity matrix and put
\begin{equation}\label{Cr}
C_r = \left( \begin{smallmatrix} & -1 \\ I_{2^r-1} & \end{smallmatrix} \right)
\qquad \text{and} \qquad
J_r = \left( \begin{smallmatrix} & J_{r-1} \\ I_{2^r} & \end{smallmatrix} \right)
\end{equation}
where $J_0 = (-1)$. Thus, for instance, $C_1 = \left( \begin{smallmatrix} &-1 \\ I_1& \end{smallmatrix} \right) = J_1$, $J_2=\left( \begin{smallmatrix} &J_1 \\ I_2& \end{smallmatrix} \right)$ and {\small $J_3=\left( \begin{smallmatrix} &&J_1 \\ &I_2& \\ I_4&& \end{smallmatrix} \right)$}. It is easy to check that $C_r, J_r \in \mathrm{SO}(2^r,\Z)$.
For instance, for the $r\le 3$ we have
{\small
\begin{center}
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{c|c|c|c|c}
$r$ & $C_r$ & $J_r$ & order & size \\ \hline
1 & $\left( \begin{smallmatrix} & -1 \\ 1 & \end{smallmatrix} \right)$ & $\left( \begin{smallmatrix} & -1 \\ 1 & \end{smallmatrix} \right)$ & 4 & 2 \\
2 & ${\scriptsize \left( \begin{smallmatrix} &&& -1 \\ 1&&& \\ &1&& \\ &&1& \end{smallmatrix} \right)}$ & ${\scriptsize \left( \begin{smallmatrix} &&& -1 \\ &&1& \\ 1&&& \\ &1&& \end{smallmatrix} \right) }$
& 8 & 4 \\
3 & $\tiny{ \left( \begin{smallmatrix} &&&&&&& -1 \\ 1&&&&&&& \\ &1&&&&&& \\ &&1&&&&& \\ &&&1&&&& \\ &&&&1&&& \\ &&&&&1&& \\ &&&&&&1& \end{smallmatrix} \right) } $ & $\tiny{ \left( \begin{smallmatrix} &&&&&&& -1 \\ &&&&&&1& \\ &&&&1&&& \\ &&&&&1&& \\ 1&&&&&&& \\ &1&&&&&& \\ &&1&&&&& \\ &&&1&&&& \end{smallmatrix} \right)}$
& 16 & 8
\end{tabular}
\end{center}}
Since $C_r$ is the companion matrix of the $2^{r+1}$-th cyclotomic polynomial $\Phi_{2^{r+1}}(x) = x^{2^{r}}+1$, it has order $o(C_r)=2^{r+1}$ and its eigenvalues are the primitive $2^{r+1}$-th roots of unity. Similarly, one can check that $J_r$ has order $2^{r+1}$ and is conjugate to $C_r$ in GL$(2^r,\R)$, hence with the same eigenvalues as $C_r$.
Thus, the rotation angles for both $C_r$ and $J_r$ are
$$\tf{\pi}{2^r}, \tf{3\pi}{2^r}, \tf{5\pi}{2^r}, \ldots, \tf{(2^r-1)\pi}{2^r} \,.$$
Since $C_r$ and $J_r$ do not have $\pm 1$-eigenvalues, we have $n_{\pm C_r} = n_{\pm J_r} = 0$, for every $r\ge 2$.
\sk
For any integer $r \ge 2$ and $j_{r-1}, \ldots, j_1 \in \N_0$, with $j_{r-1}>0$, we put $j(r)=(j_{r-1},\ldots,j_2,j_1)$ and define the matrices
\begin{gather}
\begin{aligned} \label{Bj}
B_{{j}(r)} & =
\mathrm{diag} \big( \underbrace{C_{r-1},\ldots, C_{r-1}}_{j_{r-1}}, \ldots, \underbrace{C_1,\ldots, C_1}_{j_1}, 1 \big), \sk \\
B_{{j}(r)}' & = \mathrm{diag} \big( \underbrace{J_{r-1},\ldots, J_{r-1}}_{ j_{r-1}}, \ldots, \underbrace{J_1,\ldots, J_1}_{j_1}, 1 \big).
\end{aligned}
\end{gather}
It is clear that $B_{j(r)}$ and $B_{j(r)}'$ have order $2^{r}$ and belong to $\mathrm{SO}(n,\Z)$, where
\begin{equation}\label{dimension}
n = 2^{r-1} j_{r-1} + \cdots + 4j_2 + 2j_1 + 1 = 2(2^{r-2} j_{r-1} + \cdots + 2j_2 + j_1) + 1 \,.
\end{equation}
Thus, $n=2m+1$ with $m = 2^{r-2} j_{r-1} + \cdots + 2j_2 + j_1$, and $m$ is odd if and only if $j_1$ is odd.
More generally, for each $r, n$ and $j_{r-1}, \ldots,j_1$ satisfying \eqref{dimension} we can take pairs $k_i, k_i' \in \N_0$ such that $k_i+k_i'=j_i$ for $i=1,\ldots, r-1$ and define
\begin{equation}\label{Bkk'}
B_{k(r)} = \mathrm{diag} \big( \underbrace{C_{r-1},\ldots, C_{r-1}}_{k_{r-1}}, \underbrace{J_{r-1},\ldots, J_{r-1}}_{k'_{r-1}}, \ldots,
\underbrace{C_1,\ldots, C_1}_{k_1}, \underbrace{J_1,\ldots, J_1}_{k_1'}, 1 \big)\,.
\end{equation}
Since $C_1=J_1$, for simplicity we will take $k_1=j_1$ and $k_1'=0$.
In this way, $B_{k(r)} = B_{j(r)}$ if $k(r) = (j_{r-1},0,\ldots,j_2,0,j_1)$ and $B_{k(r)} = B_{j(r)}'$ if $k(r) = (0,j_{r-1},\ldots,0,j_2,j_1)$.
Also, $n_{B}=1$ and $x_B$ is as given in \eqref{xB2r0} with $c_{2^i+1}=j_i$, $i=1,\ldots,r-1$, that is
\begin{equation}\label{xjr}
x_B = x_{j_{r-1},\ldots,j_1} = x(\underbrace{ \tf{\pi}{2^{r-1}}, \tf{3\pi}{2^{r-1}}, \ldots, \tf{(2^{r-1}-1)\pi}{2^{r-1}} }_{j_{r-1}},
\ldots, \underbrace{ \tf{\pi}{8}, \tf{3\pi}{8}, \tf{5\pi}{8}, \tf{7\pi}{8} }_{j_3}, \underbrace{ \tf{\pi}{4}, \tf{3\pi}{4} }_{j_2}, \underbrace{ \tf{\pi}{2} }_{j_1}).
\end{equation}
Define the Bieberbach groups
\begin{equation}\label{Z2r group}
\Gamma_{k(r)} := \langle \g= B_{k(r)} L_{b_r}, L_{\Ld} \rangle, \qquad b_r = \tf{1}{2^r} e_n
\end{equation}
where $j_1$ is odd and $\Ld = \Z e_1 \oplus \cdots \oplus \Z e_n$ is the canonical lattice in $\R^n$. Since $B_{k(r)} \in \mathrm{SO}(n)$ and $F = \langle B_{k(r)} \rangle \simeq \Z_{2^r}$, we have the associated orientable $\Z_{2^r}$-manifold
\begin{equation}\label{Z2r man}
M_{k(r)} := \Gamma_{k(r)} \backslash \R^n
\end{equation}
of dimension $n=2m+1 \equiv 3 \mod 4$.
\begin{defi} \label{defi f}
For a fixed $r$, let $\mathcal{F}_{r}$ denote the set of all $\Z_{2^r}$-manifolds as in \eqref{Z2r man} and let
\begin{equation}\label{familyr}
\mathcal{F} = \bigcup_{r =1}^\infty \mathcal{F}_{r}.
\end{equation}
Also, put $\mathcal{F}(n) = \{M \in \mathcal{F} : \dim M = n \}$ and $\mathcal{F}_r(n)
= \mathcal{F}_r\cap \mathcal{F}(n)$.
\end{defi}
Then, we have that $\mathcal{F}(n)$ consists of $\Z_{2^r}$-manifolds with $1 \le r \le t= \lceil \log_2 n \rceil $. In other words, if $2^{t-1} < n \le 2^{t}-1$,
$$\mathcal{F}(n) = \mathcal{F}_1(n) \cup \mathcal{F}_2(n) \cup \cdots \cup \mathcal{F}_t(n).$$
\sk
Moreover, for any fixed $r$, the number of $\Z_{2^r}$-manifolds in dimension $n$ of the form $M_{j(r)}$ equals the number of partitions
of $n$ into the first $r$-powers of two, i.e., $2^t$ with $0\le t \le r-1 $. This number is known as the \textsl{binary partition function} and is denoted by $b(n)$. Since $j_1 \ge 1$, we have that $\#\mathcal{F}_r(n) \ge b(n-3)$. Mahler showed that the logarithm of $b(n)$ grows like $(\log n)^2/2\log 2$ as $n$ grows to infinity (\cite{Ma}, see also \cite{Pe}).
Thus, asymptotically, we have
$$\#\mathcal{F}_r(n) \sim (n-3)^{\sqrt{\log_2(n-3)}}.$$
On the other hand note that, for a given $r$, the minimal dimension for $M\in \mathcal{F}_r$ is $n=2^{r-1}+3$ for $r\ge 3$ (since $j_{r-1},j_1\ge1$ and $j_i=0$ for $2\le i\le r-2$) and $n=3$ for $r=2$ (since $j_{r-1}=j_1$ in this case). Thus,
$\min \{\dim M : M\in \mathcal{F}_r\} = n_r$.
\begin{ejem} \label{tetra}
We now describe the manifolds in $\mathcal{F}$ in the lowest dimensions. There is only one 3-manifold in $\mathcal{F}$,
the tetracosm $M_1$ in \eqref{tetracosmo}.
In dimension 7, there are 3 manifolds in $\mathcal{F}$; two $\Z_8$-manifolds $M_{1,0,1}, M_{0,1,1}$, determined by
$\mathrm{diag}(J_2,J_1,1)$, $\mathrm{diag}(C_2,J_1,1)$, and the $\Z_4$-manifold $M_3$ given by
$\mathrm{diag}(J_1,J_1,J_1,1)$. In dimension 11, we have 8 manifolds in $\mathcal{F}$. Two $\Z_{16}$-manifolds $M_{1,0,0,0,1}, M_{0,1,0,0,1}$, given by the matrices $\mathrm{diag}(C_3,J_1,1)$, $\mathrm{diag}(J_3,J_1,1)$; five $\Z_8$-manifolds
$M_{2,0,1}$, $M_{1,1,1}$, $M_{0,2,1}$, $M_{1,0,3}$, $M_{0,1,3}$ given respectively by $\mathrm{diag}(C_2,C_2,J_1,1)$,
$\mathrm{diag}(C_2,J_2,J_1,1)$, $\mathrm{diag}(J_2,J_2,J_1,1)$, $\mathrm{diag}(C_2,J_1,J_1,J_1,1)$, $\mathrm{diag}(J_2,J_1,J_1,J_1,1)$; and the $\Z_4$-manifold $M_5$ determined by $\mathrm{diag}(J_1,J_1,J_1,J_1,J_1,1)$.
\end{ejem}
\subsection{Homology and $\eta$-invariants}
We will now compute the first integral homology and cohomology groups for all the manifolds in $\mathcal{F}$, showing that $H_1(M,\Z)$ has only 2-torsion.
\begin{teo} \label{teoH1}
Let $M=M_{k(r)} \in \mathcal{F}$. Then
\begin{equation}\label{H1}
H_1(M,\Z) \simeq \Z \oplus \Z_2^{j_{r-1} + \cdots + j_1}
\end{equation}
with $j_i=k_i+k_i'$ for $1\le i \le r-1$, and
\begin{equation}\label{H^1}
H^1(M,\Z) \simeq \Z, \qquad H^1(M,\Z_2) \simeq \Z_2^{j_{r-1}+\cdots+j_1+1}.
\end{equation}
\end{teo}
\begin{proof}
We will first compute $H_1(M_\G,\Z)= \G / [\G,\G]$, where $\G=\G_{k(r)}=\langle \g, L_{e_1}, \ldots, L_{e_n}\rangle$. There are 3 kind of commutators:
$[L_\ld,L_{\ld'}]=I$, $[\g,L_\ld]$ and $[\g,\g']$.
Since $F$ is cyclic, every element in $\G$ is of the form $\g^i L_\ld$ and thus
$[\g^i L_\ld,\g^j L_\ld'] = [\g^i, \g^j][L_\ld,L_{\ld'}] = I$ for every $i,j \in \Z$, $\ld,\ld'\in \Ld$. Also,
$[\g,L_\ld] = BL_b L_\ld L_{-b} B^{-1} L_{-\ld} = L_{B\ld-\ld}$.
Therefore,
$$[\G,\G] = \langle [\g,L_\ld] : \ld \in \Ld \rangle = L_{(B-I)\Ld},$$
We now study the action of each of the blocks $C_i$ and $J_i$ on $\Ld_i \subseteq \Ld$, with $\Ld_i \simeq \Z^{2^i}$. For every
$1\le i \le r-1$, let $e_1, \ldots, e_{2^i}$ be any $\Z$-basis of $\Ld_i$. For $C_i$, we have
$$(C_i - I_{2^i})e_j = e_{j+1} - e_j \quad 1\le j\le 2^i-1 \qquad \text{and} \qquad (C_i - I_{2^i})e_{2^i} =-e_1 - e_{2^i}.$$
By putting $f_j = e_{j+1} -e_j$ for $1\le j\le 2^i-1$ and $f_{2^i} = -e_1 - e_{2^i}$, we have that
$$(f_1 + \cdots + f_{j-1}) - (f_j + \cdots + f_{2^i}) = 2e_j, \qquad 1\le j\le 2^i,$$
and thus $2e_j \in L_{(C_i-I_{2^i})\Ld_i}$ for every $1\le j \le 2^i$. Furthermore, since
$L_{e_{j}}-L_{e_{j-1}} \in [\G,\G]$ for $1\le j\le 2^i$, we have that
$L_{e_j} \sim L_{e_k}$, $1\le j,k\le 2^i$, in the quotient $\G/[\G,\G]$. Therefore,
$$\langle L_{e_1},\ldots,L_{e_{2^i}}\rangle / \langle L_{f_1},\ldots,L_{f_{2^i}} \rangle \simeq \Z_2,$$
and thus, every block $C_i$ gives rise to one $\Z_2$ in the quotient $\G / L_{(B-I)\Ld}$.
On the other hand, for the block $J_i$, it will be convenient to look up at the $e_j$'s with $j$ in the intervals
$1 \le j \le 2^{i-1}$,\quad $2^{i-1}+1 \le j \le 2^{i-1}+2^{i-2}$,\quad $2^{i-1}+2^{i-2}+1 \le j \le 2^{i-1}+2^{i-2}+2^{i-3}$, and so on.
We then have
$$(J_i - I_{2^i})e_{j(l,k)} = e_{2^{i-l}+k} - e_{j(l,k)}, \qquad 1\le k\le 2^i, \: 1\le l\le r,$$
where
$$j(l,k) = 2^{i-1}+2^{i-2}+\cdots +2^{i-(l-1)} + k.$$
As before, by putting $g_j=e_{2^{i-l}+k} - e_{j(l,k)}$ and looking at the sums of the form $\sum_{j} \pm g_j$ one sees that
$2e_j \in L_{(J_i-I_{2^i})\Ld_i}$ for every $1\le j \le 2^i$.
Also, we again have that $L_{e_{j}} \sim L_{e_{k}} \in [\G,\G]$ for $1\le j,k \le 2^i$. Thus, every block $J_i$ induces a $\Z_2$ in the quotient $\G / L_{(B-I)\Ld}$.
Moreover, since $\g^{2^r}=L_{e_n}$, it is clear that $\langle \g, L_{e_n}\rangle$ generates an infinite cyclic group in $\G / L_{(B-I)\Ld}$. As a result of all these observations we get expression \eqref{H1}.
\sk
Now, it is known that $H^n(M,\Z) \simeq F_n \oplus T_{n-1}$, where $F_n$ and $T_n$ are the free and the torsion part of $H_n(M,\Z)$, respectively. Since $H_0(M,\Z) \simeq \Z$, by \eqref{H1} we get $H^1(M,\Z) = \Z$, as desired. Finally, by the universal coefficient theorem, we have
\begin{eqnarray*}
H^1(M,\Z_2) & \simeq & \mathrm{Hom}(H_1(M,\Z),\Z_2) \oplus \mathrm{Ext}(H_0(M,\Z),\Z_2) \\
& \simeq & \mathrm{Hom}(\Z\oplus \Z^{j_{r-1}+\cdots+j_2+j_1},\Z_2) \,,
\end{eqnarray*}
where we have used that $\mathrm{Ext}(H_0(M,\Z),\Z_2) \simeq \mathrm{Ext}(\Z,\Z_2) \simeq 0$, for $\Z$ is projective. Finally, since $\mathrm{Hom}(\oplus G_i,G) = \bigoplus \mathrm{Hom}(G_i,G)$ and $\mathrm{Hom}(\Z,\Z_2) = \mathrm{Hom}(\Z_2,\Z_2) =
\Z_2$ we obtain that $H^1(M,\Z_2)\simeq \Z_2^{j_{r-1}+\cdots+j_2+j_1+1}$, and the proof is now complete.
\end{proof}
As a result, we can count the number of spin structures of the manifolds in $\mathcal{F}$.
\begin{coro}\label{spins}
Every $M=M_{k(r)} \in \mathcal{F}$ is spin and has $2^{j_{r-1}+\cdots+j_1+1}$ spin structures.
\end{coro}
\begin{proof}
By applying the methods used in \cite{MP.MZ04} or \cite{Po.UMA05} (see Theorem 2.1 or Proposition 2.2, respectively, and their previous comments), one can prove that $M$ is a spin manifold. The number of spin structures of $M$ is then given by
$\#H^1(M,\Z_2) = 2^{j_{r-1}+\cdots+j_1+1}$ (see \cite{LM}).
\end{proof}
\begin{quest} \label{q1}
Let $a(r)=(a_{r-1},a_{r-1}',\ldots,a_2,a_2',j_1)$ and $b(r)=(b_{r-1},b_{r-1}',\ldots,b_2,b_2',j_1)$ be two $(2r-1)$-tuples satisfying
$a_i+a_i' = b_i+b_i' = j_i$ for $2\le i \le r-1$, but with $a(r)\ne b(r)$. Then, $M_{a(r)}$ and $M_{b(r)}$ have, in general, different integral representations. Since they have the same eigenvalues, $C_r$ and $J_r$ are conjugate in $\R$. Are they still conjugate in $\Z$? In other words, are $M_{a(r)}$ and $M_{b(r)}$ equivalent as compact flat manifolds? Theorem \ref{teoH1} gives no answer to this question.
\end{quest}
\begin{rem}
In \cite{Po.UMA05}, (2.1)--(2.3), we have defined a family $\mathcal{F}^n = \{M_{j,k,l}\}$ of $n$-dimensional $\Z_4$-manifolds, $n=2m+1$, and we obtained that $H_1(M_{j,k,l},\Z) \simeq \Z^l \oplus \Z_2^{j+k}$ (see Lemma~2.1). If $m$ is odd and $\Lambda =\Z^n$, the manifolds $M_{j,0,1} \in \mathcal{F}^n$ are exactly the $\Z_4$-manifolds of the form $M_{j(r)}=M_j \in \mathcal{F}_2$ in this paper. This is in agreement with \eqref{H1}.
Corollary \ref{spins} says, for example, that the tetracosm $M_1$ has $2^{1+1}=4$ spin structures. This is also in coincidence with Proposition 2.2 and Corollary 2.3 in \cite{Po.UMA05}.
\end{rem}
\begin{rem}
It is possible to define a bigger family $\tilde{\mathcal{F}}$, using matrices of order $\le 2$. That is, we can add $2 \times 2$ blocks
$J = (\begin{smallmatrix} & 1 \\ 1 & \end{smallmatrix})$ and $\pm I = (\begin{smallmatrix} \pm 1& \\ & \pm 1 \end{smallmatrix})$.
Thus, consider the matrix
$$B_{\kappa(r)} := \mathrm{diag} \big( B'_{k(r)}, \underbrace{J,\ldots, J}_{k_0}, \underbrace{-I,\ldots, -I}_{k_0'},
\underbrace{I,\ldots,I}_{i_0}, 1 \big)$$
with $k_0,k_0'$ even and where $B'_{k(r)}$ denotes the matrix $B_{k(r)}$ in \eqref{Bkk'} with the 1 in the last position removed. Similarly as in \eqref{Z2r group}--\eqref{Z2r man} we define the group $\Gamma_{\kappa(r)}$ and the corresponding $\Z_{2^r}$-manifold $M_{\kappa(r)}$ of dimension
$n=2^{r-1}j_{r-1} + \cdots + 4j_2 + 2(j_1 + j_0 + i_0) +1$, where $j_0 = k_0+k_0'$. Hence, $j_1 + i_0$ must be odd for $n$ to be congruent to 3 mod 4. For $M_{\kappa(r)} \in \tilde{\mathcal{F}}$, proceeding similarly as in the proof of Theorem \ref{teoH1}, one can obtain
$$H_1(M_{\kappa(r)},\Z) = \Z_2^{j_{r-1} + \cdots + j_1 + k_0'} \oplus \Z^{k_0 + i_0+1}.$$
Similar results can be obtained for $H^1(M_{\kappa(r)},\Z)$ and $H^1(M_{\kappa(r)},\Z_2)$. However, we have $\eta(M_{\kappa(r)})=0$ unless $M_{\kappa(r)} \in \mathcal{F}$, i.e.\@ $j_0=i_0=0$ and $M_{\kappa(r)} = M_{k(r)}$.
\end{rem}
As a consequence of our previous results we have that, for manifolds in $\mathcal{F}$, the $\eta$-invariant has a strong topological meaning. In fact, it is related with the order of the torsion group $T$ (which equals the 2-torsion) of the first integral homology group and with the number of spin structures.
\begin{prop} \label{prop.etasz2r}
If $M=M_{k(r)} \in \mathcal{F}$ then and
\begin{equation} \label{etaH1}
\eta (M) = -2^{j_{r-1}+\cdots +j_1-1} = - \tf12 |T| = \tf14 \# \mathrm{Spin}(M) \ne 0,
\end{equation}
where $T$ is the torsion subgroup of $H_1(M,\Z)$ and $\mathrm{Spin}(M)=\{\text{spin structures on $M$}\}$.
\end{prop}
\begin{proof}
Suppose $M=M_{k(r)}$ of dimension $n=2m+1$, with $m$ odd, and $k_i+k_i'=j_i$ for $1\le i \le r$.
Then, $m=2^{r-2}j_{r-1} + \cdots +2j_2+j_1$ with $j_i = c_{2^{i+1}}$ for $1\le i \le r-1$.
Since $v_B=e_n$ we have $\ell_\g=1$, $\nu=0$ and $\ell_{_B}=1$, and thus the expression for $\eta(s)$ follows from \eqref{generaletas}.
Since $\sigma_B=1$ and $f(B) = j_{r-1}+\cdots +j_1+1$, by \eqref{generaleta0}, we have $\eta=-2^{j_{r-1}+\cdots + j_1-1}$. The result readily follows from \eqref{H1} in Theorem \ref{teoH1}.
\end{proof}
\begin{quest}\label{q2}
For the $\Z_8$-manifold $M$ of Example \ref{nonstandard}, Theorem \ref{teoH1} and Proposition \ref{prop.etasz2r} do not apply. Proceeding similarly as in (the proof of) Theorem \ref{teoH1}, we can check that $H_1(M,\Z) = \Z \oplus \Z_2^{2}$ and,
since $\eta(M)=\pm 2$, we still have $\eta(M)=\pm \tf12 |T|$ as in \eqref{etaH1}.
Does this phenomenon hold in general or is there a $\Z_{2^r}$-manifold $M\notin \mathcal{F}$ with non-trivial $\eta$-invariant
such that $\eta(M) \ne -\tf12 |T|$?
\end{quest}
\begin{quest}\label{q3}
If $M$ is an arbitrary $\Z_{2^r}$-manifold, is $\eta(M)$ completely determined by $H_1(M,\Z)$ as it is the case for manifolds in $\mathcal{F}$? The best we can say is the statement of Lemma \ref{lemin} below.
\end{quest}
\begin{rem}\label{eta ell}
By definition, every $\Z_{2^r}$-manifold $M \in \mathcal{F}$ has a generator $\g = B L_b$ with $b=\tf{1}{2^r} e_n$ and hence $\ell_\g=1$. For a given $M$, we can define $M_{\ell}$ with $b$ replaced by $b_\ell=\tf{\ell}{2^r} e_n$, $\ell$ odd (these manifolds are diffeomorphic but non-isometric to each other). Clearly, we have $\ell_\g=\ell$ and
$\eta (M_\ell) = (-1)^{[\ell/2]} \, \eta (M)$, for every $\ell$ odd.
\end{rem}
Note that, for fixed $r$, the $\eta$-invariant determines the eta function $\eta(s)$.
In fact, by Theorem \ref{teo.etasz2r} and Proposition \ref{prop eta0}, we have
\begin{equation} \label{etas F}
\eta (s,M) = -\tf{2\eta(M)}{(2^{r-1-\nu}\pi \lambda_B)^s} L(s,\chi_4).
\end{equation}
Since manifolds in $\mathcal{F}$ has $\nu=0$ ($\ell_\g=1$) and $\lambda_B=1$, it is clear that if
$M, M' \in \mathcal{F}_r$ then
$$\eta(M)=\eta(M') \quad \Rightarrow \quad \eta(s,M)=\eta(s,M').$$
This may not be the case in general because of the numbers $r$, $\nu$ and $\lambda_B$;
since it could well happen that $2^{r-1-\nu}\lambda_B = 2^{r'-1-\nu'}\lambda_B'$.
\section{The image of $\eta^*$} \label{sec eta*}
We will consider the following subfamilies of the set $\mathcal{M}$ of all $\Z_{2^r}$-manifolds:
the subset $\mathcal{M}_r$ of $\Z_{2^r}$-manifolds with fixed $r$ (arbitrary dimension), the subset $\mathcal{M}(n)$ of
$n$-dimensional $\Z_{2^r}$-manifolds ($r$ arbitrary) and the subset $\mathcal{M}_r(n) = \mathcal{M}_r \cap \mathcal{M}(n)$, with both $r$ and $n$ fixed. Clearly, $\mathcal{F} \subset \mathcal{M}$, $\mathcal{F}_r \subset \mathcal{M}_r$, $\mathcal{F}(n) \subset \mathcal{M}(n)$ and
$\mathcal{F}_r(n) \subset \mathcal{M}_r(n)$.
We will now study the image of the map
\begin{equation} \label{eta*}
\eta^* \, : \, \mathcal{M} \rightarrow \Z, \qquad M \mapsto \eta(M)
\end{equation}
and its restrictions $\eta_r^*$ to $\mathcal{M}_r$ and $\eta_{(n)}^*$ to $\mathcal{M}(n)$.
By Proposition \ref{prop eta0}, it is clear that
$$\mathrm{Im} \, \eta^* \subseteq \{0,\pm 1\} \cup \{\pm 2^k : k\in \N\}.$$
The extreme cases, i.e.\@ $\Z_2$-manifolds and 3-dimensional $\Z_{2^r}$-manifolds, are trivial ones, since
$\eta=0$ if $r=1$ and $\eta = 0, \pm 1$ if $n=3$; i.e.\@ $\mathrm{Im}\, \eta_1^* = \{0\}$ and $\mathrm{Im}\, \eta_{(3)}^* = \{0,\pm 1\}$.
Since we are considering dimensions $n\equiv 3$ mod 4, for $M\in \mathcal{F}$, $n$ is as given in \eqref{dimension} with $j_1$ odd. Out of all possible partitions of $n$ into powers of 2, the 2-adic expansion of $n$ is a proper partition with the minimum number of parts. Let $\tau(n)$ be this number, i.e.\@
\begin{equation} \label{tau}
\tau(n) = \tau \qquad \text{if} \qquad n=2^{a_\tau} + \cdots + 2^{a_2} + 2^{a_1}, \quad 0\le a_1 < a_2 < \cdots < a_\tau.
\end{equation}
Equivalently, if $n= b_m 2^m + \cdots + b_3 2^3 + b_2 2^2 +b_1 2+b_0$ for some $m$, with $b_i\in\{0,1\}$, $1\le i \le m$,
then $\tau(n)$ equal the number of non zero $b_i$'s.
In our case, we have $\tau(3)=2$ and $\tau(n)\ge 3$ for $n\ge 7$.
Furthermore, $\tau(n_r)=3$ for any $r\ge 2$ and $\tau(n)\ge 4$ for $n\ne n_r$, where we recall that, by \eqref{nr} and
\eqref{dimension},
\begin{equation} \label{nr2}
n_r = \min_{M\in \mathcal{F}_r} \{ \dim M \} = \min_{M\in \mathcal{M}_r} \{\dim M: \eta(M)\ne 0\} = 2^{r-1}+3.
\end{equation}
Let us see that given a general $\Z_{2^r}$-manifold $M$ with $\eta\ne 0$, the value $\eta(M)$ can be obtained by the eta invariant of some manifold in $\mathcal{F}$.
\begin{lema} \label{lemin}
If $M\in \mathcal{M}$ then $\eta(M)=0$ or $\eta(M) = \eta(M_{k(r)})$ for some $M_{k(r)} \in \mathcal{F}$.
\end{lema}
\begin{proof}
Let $M$ be any orientable $\Z_{2^r}$-manifold of dimension $n$, with point group
$F=\langle B\rangle$.
By Proposition \ref{prop eta0}, $\eta(M)=0$ if $n_B>1$ or $r=1$, while $\eta(M) = \pm 2^{f(B)-2}$ otherwise.
Assume that $n_B=1$ and $r\ge 2$.
Since $B$ is conjugate to $x_B=x_{j_{r-1},\ldots,j_1}$ as in \eqref{xjr} (or to $x_B'$, see Proposition \ref{lema angles}), $B$ is also conjugate in $\mathrm{GL}(n)$ to the matrix $C_B = \mathrm{diag}(B_{j(r)},1)$ with $k(r)=(j_{r-1},\ldots,j_1)$ as in \eqref{Bj}. Thus, $\eta(M)$ is determined (up to sign) by the number of matrices $C_i$ or $J_i$, $1\le i \le r-1$, in $C_B$. Finally, if $\eta(M) = -\eta(M_{k(r)})$, then considering $M_{k(r)}^-$ (the manifold $M_{k(r)}$ with the opposite orientation) we get $\eta(M) = \eta(M_{k(r)}^-)$.
\end{proof}
\begin{teo}\label{im eta}
Let $n=2m+1$, with $m$ odd, be fixed. Then, $\mathrm{Im}\, \eta_{(n)}^* = \{0\}$ for $r=1$, $\mathrm{Im}\, \eta_{(n)}^* = \{0,\pm 2^{m-1} \}$ for $r=2$ and
\begin{equation}\label{Imn*}
\mathrm{Im}\, \eta_{(n)}^* = \{0,\pm 2^{\tau(n)-2}, \pm 2^{\tau(n)-1}, \ldots, \pm 2^{m-1}\}, \qquad r\ge 3.
\end{equation}
In particular,
$\mathrm{Im}\, \eta_{(n_r)}^* = \{ 0,\pm 2, \pm 2^2,\ldots, \pm 2^{2^{r-2}} = \pm 2^{\f{n_r-3}{2}} \}$.
Therefore, $\mathrm{Im}\, \eta_1^* = \{ 0\}$, $\mathrm{Im}\, \eta_2^* = \{0,\pm 4^k\}_{k\in \N_0}$
and $\mathrm{Im}\, \eta_r^* = \{0,\pm 2^k\}_{k\in \N}$ for $r\ge 3$.
\end{teo}
\begin{proof}
By Lemma \ref{lemin}, we can assume that $M \in \mathcal{F}_r(n)$ and, without loss of generality, that $M=M_ {j(r)}$ as in \eqref{Bj}--\eqref{Z2r man}. By Proposition~\ref{prop eta0}, $\operatorname{Im} \eta_{(3)}^* =\{0,\pm 1\}$, with the values $\pm 1$ attained by the tetracosm. From now on, suppose that $n\ge 7$ and $\eta\ne 0$. Hence, by \eqref{generaleta0}, $n_B=1$ and $r\ge 2$.
Fix the dimension
$$n= 2(2^{r-2} j_{r-1} + \cdots + 2j_2 + j_1) +1 \qquad j_{r-1}>0, \quad \text{$j_1$ odd.}$$
If $r=2$, then $n=2 j_1+1=2m+1$ and thus $f(B) = m+1$. Hence, $\eta(M) = \pm 2^{m-1}$ by \eqref{generaleta0}. The positive value $2^{m-1}$ is attained by the manifold $M_m \in \mathcal{F}$ determined by $B_m = \mathrm{diag}(C_1,\ldots,C_1,1)$, with $C_1$ repeated $m$-times.
\sk
Let $r\ge 3$. If $n=n_r=2^{r-1}+3$, all the values $2, 4, 8, \ldots, 2^{m-1}$ are attained by the $\eta$-invariants
on the following manifolds. $M^{(1)}=M_{1,0,\ldots,0,1} \in \mathcal{F}_r$ determined by the matrix $\mathrm{diag}(C_{r-1},C_1,1)$, has 3 blocks and hence $\eta(M^{(1)})=2$. By replacing the block $C_{r-1}$ by 2 blocks $C_{r-2}$ we get
$M^{(2)} = M_{0,2,0,\ldots,0,1}\in \mathcal{F}_{r-1}$ determined by $\mathrm{diag}(C_{r-2}, C_{r-2}, C_1, 1)$ with 4 blocks, and hence
$\eta(M^{(2)})=2^{4-2}=4$. Now, by replacing one block $C_{r-2}$ from the previous matrix by 2 blocks $C_{r-3}$ we get $M^{(3)} = M_{0,1,2,0,\ldots,0,1} \in \mathcal{F}_{r-1}$ determined by $\mathrm{diag}(C_{r-2}, C_{r-3}, C_{r-3}, C_1, 1)$ with 5 blocks, and hence $\eta(M^{(2)})=2^{5-3}=8$. It is clear that, by repeating this \textsl{`splitting block'} procedure, i.e.\@ by replacing some block $C_{j_i}$ of $M^{(i)}$ by 2 blocks $C_{j_i-1}$ (this changes the $\eta$-invariant keeping the dimension unaltered), we get a finite sequence of $\Z_{2^t}$-manifolds
$$M^{(1)}, M^{(2)}, \ldots, M^{(m-1)}$$
(with different $t$'s, $3\le t\le r$), determined by diagonal block matrices with $3,4,\ldots, m+1$ blocks respectively, and hence, with corresponding $\eta$-invariants
$2^1,2^2,\ldots, 2^{m-1}$.
Since $n=2m+1$, we have
$m=2^{r-2}+1$ and hence $m-1=2^{r-2}=\tf{n_3-3}{2}$.
Here, $M^{(m-1)}$ is the $\Z_4$-manifold determined by the matrix $\mathrm{diag}(C_1, \ldots, C_1,1)$, with $C_1$ repeated $m$ times.
The splitting of blocks are not unique, but there is at least one. Also, by dimension issues, it is clear that $2^{m-1}$ is the maximum posible value for $\eta$.
In case $n \ne n_r$, we proceed similarly as before. We begin with $M^{(1)}=M_{j_{r-1},\ldots,j_1}$, where $j_{r-1}=j_1=1$ and
$j_{i_1},\ldots,j_{i_{\tau-3}} \ge 1$ with $\tau=\tau(n)$ --the other $j_k$'s being $0$--, such that
$$n=2^{r-1} + (\sum\limits_{k=1}^{\tau-3} 2^{i_k}j_{i_k}) + 3 = n_r + \sum\limits_{k=1}^{\tau-3} 2^{i_k}j_{i_k}.$$
By replacing some block $C_{j_i}$ by 2 blocks $C_{j_i-1}$ and iterating this process, we get a sequence
$M^{(1)}, M^{(2)},\ldots, M^{(m-\tau+2)}$ of $\Z_{2^t}$-manifolds (with different $t$'s, $t\le r$), respectively determined by diagonal block matrices with $\tau,\tau+1,\ldots, m+1$ blocks, and hence with corresponding $\eta$-invariants $2^{\tau-2},2^{\tau-3},\ldots, 2^{m-1}$.
\sk
To get trivial $\eta$-invariants for any $n$ and $r$, just take the previous manifolds and replace one $J_1$ by
$\pm I =(\begin{smallmatrix} \pm 1 &0 \\ 0& \pm 1 \end{smallmatrix})$ (these manifolds will be not in $\mathcal{F}$). To get the negative values, since $\eta(M^-) = -\eta(M)$, we just need to change the orientation of every $M$ previously used.
The remaining assertions in the statement follow directly from the previous ones, and the result is thus proved.
\end{proof}
\begin{ejem}\label{ejemplin}
Here, we illustrate the results in Theorem \ref{im eta}. Using \eqref{Imn*}, in Table~\ref{tabla2}, we give the values of $\eta(M)$ for $M$ a $\Z_{2^r}$-manifold in the lowest dimensions. For each fixed value of $n$, we give the 2-adic expansion, $\tau(n)$, the highest possible $r$ for a $\Z_{2^r}$-manifold in this dimension,and all the allowed values of $\eta$ (only non-negative values for simplicity).
{\small
\renewcommand{\arraystretch}{1}
\begin{table}[h]
\caption{$\eta$-invariants for $\Z_{2^r}$-manifolds, $3\le n \le 63$.}
\label{tabla2}
\begin{center}
\begin{tabular}{l|c|c|l}
dimension $n$ & $\tau(n)$ & $\max r$ & $\eta$-invariant \\
\Xhline{2\arrayrulewidth}
$3=2+1=n_2$ & 2 & 2 & $0, 1$ \\ \hline
$7=4+2+1=n_3$ & 3 & 3 & $0, 2,2^2$ \\ \hline
$11=8+2+1=n_4$ & 3 & 4 & $0, 2, 2^2, 2^3, 2^4$ \\
$15=8+4+2+1$ & 4 & 4 & $0, 2^2, 2^3, 2^4, 2^5, 2^6$ \\\hline
$19=16+2+1=n_5$ & 3 & 5 & $0, 2, 2^2,\ldots, 2^8$ \\
$23=16+4+2+1$ & 4 & 5 & $0, 2^2, 2^3,\ldots, 2^{10}$ \\
$27=16+8+2+1$ & 4 & 5 & $0, 2^2, 2^3,\ldots, 2^{12}$ \\
$31=16+8+4+2+1$ & 5 & 5 & $0, 2^3, 2^4,\ldots, 2^{14}$ \\ \hline
$35=32+2+1=n_6$ & 3 & 6 & $0, 2, 2^2,\ldots, 2^{16}$ \\
$39=32+4+2+1$ & 4 & 6 & $0, 2^2, 2^3, \ldots, 2^{18}$ \\
$43=32+8+2+1$ & 4 & 6 & $0, 2^2, 2^3, \ldots, 2^{20}$ \\
$47=32+8+4+2+1$ & 5 & 6 & $0, 2^3, 2^4, \ldots, 2^{22}$ \\
$51=32+16+2+1$ & 4 & 6 & $0, 2^2, 2^3, \ldots, 2^{24}$ \\
$55=32+16+4+2+1$ & 5 & 6 & $0, 2^3, 2^4, \ldots, 2^{26}$ \\
$59=32+16+8+2+1$ & 5 & 6 & $0, 2^3, 2^4, \ldots, 2^{28}$ \\
$63=32+16+8+4+2+1$ & 6 & 6 & $0, 2^4, 2^5, \ldots, 2^{30}$
\end{tabular}
\end{center}
\end{table}}
\end{ejem}
We will show that there exists an infinite number of infinite families of $\Z_{2^r}$-manifolds in $\mathcal{F}$ (with growing dimensions) having constant
$\eta$-invariant, one for every possible positive power of two.
Let $n_{r,k}$ denote the minimal dimension for a $\Z_{2^r}$-manifold with $\eta=2^k$, i.e.
\begin{equation} \label{nrk}
n_{r,k} = \min_{M\in \mathcal{M}_r} \{\dim M \,:\, \eta(M)=2^k\}.
\end{equation}
By Theorem \ref{im eta}, this number is well defined for $r\ge 3$, $k\ge1$ and for $r=2$, $k=1$ or $k$ even. Note that $n_{r,1}=n_r$ and for $r\ge 2, k\ge 1$,
by \eqref{generaleta0}, we have
\begin{align*}
n_{r,1} & \le n_{r,2} \le n_{r,3} \le \cdots \le n_{r,k} \le n_{r,k+1} \le \cdots \\
n_{2,k} & \le n_{3,k} \le n_{4,k} \le \cdots \le n_{t,k} \le n_{t+1,k} \le \cdots
\end{align*}
We now compute these dimensions.
\begin{lema}\label{lem nrk}
We have $n_{2,2k} = 3+2k$ for $k\ge 0$ and for any $r\ge 3$, $k\ge 1$
$$n_{r,k}= n_r+[\tf k2]= 2^{r-1}+[\tf k2]+3.$$
\end{lema}
\begin{proof}
Let $M\in \mathcal{M}_r$ with $r\ge 2$. By Lemma \ref{lemin} we can assume that $M=M_{j(r)}\in \mathcal{F}$.
If $r=2$, it is clear that $n_{2,2k}=2k+3$, attained by the manifold $M_{2k+1}$ determined by the matrix
$B_{2k+1} = \mathrm{diag}(J_1, \ldots, J_1, 1)$, with $J_1$ repeated $2k+1$ times.
Now consider the case $r\ge 3$. It is clear that $n_{r,1}=n_r=2^{r-1}+3$, attained by the manifold $M_{1,0,\ldots,0,1}$ determined by the matrix $B_{1,0,\ldots, 0,1}=\mathrm{diag}(J_{r-1},J_1,1)$. To get $\eta=2^2$ we need 4 blocks, so the minimum dimension where this can be achieved is $7+4=11$ given by the manifold $M_{1,0,\ldots,0,1,1}$ determined by the matrix $B_{1,0,\ldots,0,1,1}=\mathrm{diag}(J_{r-1},J_2,J_1,1)$. Similarly, for $\eta=2^3$, the minimum dimension is also $11$, attained by the manifold $M_{1,0,\ldots,0,3}$ determined by the matrix $B_{1,0,\ldots,0,3}=\mathrm{diag}(J_{r-1},J_1,J_1,J_1,1)$ (just split one block $J_2$ by two $J_1$'s). In general, the minimal dimension needed to get $\eta=2^{2k}$ or $2^{2k+1}$ is the same. In fact,
$n_{r,2k}= 2^{r-1} + 4+2(2k-2)+3$ and $n_{r,2k+1}= 2^{r-1} + 2((2k+1)-1)+3$, that is
$$n_{r,2k}=n_{r,2k+1}=2^{r-1}+4k+3=n_r+4k,$$
from which the expression in the statement directly follows.
These dimensions are respectively attained by the manifolds $M_{2,2k-1}$ and $M_{1,2k+1}$, determined by the matrices
$B_{1,0,\ldots,0,1,2k-1}$ $=\mathrm{diag}(J_{r-1},J_2,J_1,\ldots,J_1,1)$, $J_1$ repeated $2k-1$ times, and $B_{1,0,\ldots,0,2k+1}=$
$\mathrm{diag}(J_{r-1},J_1,\ldots,J_1,1)$, $J_1$ repeated $2k+1$ times.
\end{proof}
\begin{ejem} (i) Let us construct $\Z_{2^7}$-manifolds $M$, $M'$ with $\eta(M)=2^{11}$, $\eta(M')=2^{10}$. Since $[\tf{11}2]=5$, the minimal dimension for $M$ is given by $n_{7,11}= n_7+4\cdot 5= 2^6+3+20= 87$. We need 13 `blocks', so take $M=M_{1,0,0,0,0,11}$ given by $B=\mathrm{diag}(E_6,J_1,\ldots,J_1,1)$ with $J_1$ repeated $11$ times, where $E_6=C_6$ or $J_6$.
As before, $n_{7,10}=87$. Now, we need 12 blocks, thus we take $M'=M_{1,1,0,0,0,10}$ given by $B=\mathrm{diag}(E_6,E_2, J_1,\ldots,J_1,1)$,
$J_1$ repeated 10 times, with $E_i=C_i$ or $J_i$ for $i=2,6$.
\sk (ii)
In Table \ref{tabla3} below, we give all the $\Z_{2^r}$-manifolds in family $\mathcal{F}$ having $\eta=2^3$ in dimensions $n=4k+3 \le 35$. This manifolds are defined by diagonal block matrices $B_{j(r)}$ with the blocks in the set $\{J_4, J_3, J_2, J_1, J_0\}$, as defined in \eqref{Bj}.
\begin{table}[h]
\caption{all $\Z_{2^r}$-manifolds in $\mathcal{F}$ with $\eta=8$ in $\dim n \le 35$.}
\label{tabla3}
\begin{center}
\begin{tabular}{c|c|ccccc|c|c}
$\dim$ & partition of $n$ & $J_4$ & $J_3$ & $J_2$ & $J_1$ & $J_0$ & $r$ & $F$ \\
\hline
11 & 4+2+2+2+1 & 0 & 0 & 1 & 3 & 1 & 3 & $\Z_8$ \\
15 & 4+4+4+2+1 & 0 & 0 & 3 & 1 & 1 & 3 & $\Z_8$ \\
15 & 8+2+2+2+1 & 0 & 1 & 0 & 3 & 1 & 4 & $\Z_{16}$\\
19 & 8+4+4+2+1 & 0 & 1 & 2 & 1 & 1 & 4 & $\Z_{16}$\\
23 & 8+8+4+2+1 & 0 & 2 & 1 & 1 & 1 & 4 & $\Z_{16}$\\
23 & 16+2+2+2+1& 1 & 0 & 0 & 3 & 1 & 5 & $\Z_{32}$\\
27 & 8+8+8+2+1 & 0 & 3 & 0 & 1 & 1 & 4 & $\Z_{16}$\\
27 & 16+4+4+2+1& 1 & 0 & 2 & 1 & 1 & 5 & $\Z_{32}$\\
31 & 16+8+4+2+1& 1 & 1 & 1 & 1 & 1 & 5 & $\Z_{32}$
\end{tabular}
\end{center}
\end{table}
\end{ejem}
We now show that there are infinite families of $\Z_{2^r}$-manifolds with prescribed constant $\eta$-invariant and growing dimensions.
\begin{coro} \label{coro =eta}
For every positive integer $k$ there is a family $\mathcal{G}_k=\{M_i\}_{i=1}^\infty \subset \mathcal{F}$
such that $\eta(M_i)= 2^k$ for every $i$ and
$\dim M_i \nearrow \infty$. In particular, for every $r\ge 2$, one can take each $M_r \in \mathcal{G}_k$ having holonomy group of order $2^r$ and minimal dimension $n_{r,k}$.
\end{coro}
\begin{proof}
It is sufficient to prove the second claim in the statement. Let $k\ge 1$, $r\ge 2$.
By \eqref{generaleta0}, to get $\eta=2^k$ , we need to take a manifold in $\mathcal{F}$ with exactly $k+2$ blocks.
Consider the matrix
$$B_r = \mathrm{diag}(J_{r-1},\underbrace{J_2,\ldots,J_2}_{(k-1)-\mathrm{times}},J_1,1).$$
The induced manifold $M_r$ in $\mathcal{F}$ has holonomy group of order $2^r$, $\eta=2^k$ and dimension
$$\dim M_r = 2^{r-1}+4(k-1)+1 =2^{r-1}+4k-3\ge n_{r,k}.$$
To obtain minimal dimensions, just proceed as in the proof of Lemma \ref{lem nrk}, taking $B_r = \mathrm{diag}(J_{r-1},J_2,J_1,\ldots,J_1,1)$ or $B_r = \mathrm{diag}(J_{r-1},J_1,\ldots,J_1,1)$, depending whether $k$ is even or odd.
\end{proof}
\section{Comparison with Donnelly's expression for $\eta$} \label{sec doneli}
In \cite{MP.JGA11}, we gave an expression for the $\eta$-invariant of any cfm (i.e., arbitrary holonomy representations, any lattice) following a direct approach; i.e., we first computed the multiplicities of the eigenvalues, then we found $\eta(s)$ and finally we
obtained $\eta$ by evaluation at $s=0$ (see Theorems 3.3, 3.5 and 4.2 in \cite{MP.JGA11}).
On the other hand, Donnelly has previously obtained an expression for the $\eta$-invariant for more general manifolds $M$ (compact oriented Riemannian manifolds with a Lie group acting by isometries on it) in an indirect way (\cite{Do}, Theorem 3.4), by first computing the signature of a manifold $\tilde M$ with $M=\partial \tilde M$, and then using the Atiyah-Patodi-Singer index theorem for manifolds with boundary in \cite{APS1}.
For a very special kind of cfm's, namely those having holonomy group $F \subset \textrm{SO}(n,\Z)$ and canonical lattice $\Lambda=\Z^n$, the expression for $\eta$ drastically simplifies (\cite{Do}, Proposition 4.12).
Since both methods are quite different, i.e.\@ representation theoretical vs.\@ topological ones, it is the author feeling that it is interesting to compare both expressions for the $\eta$-invariant, when possible (i.e.\@ in the special case considered in \cite{Do}).
\sk
Donnelly considered cfm's having holonomy group $F\subset \mathrm{SO}(n,\Z)$ where each $B\in F$ is of the form $B=\mathrm{diag}(B',1)$, $B' \in \mathrm{SO}(n-1,\Z)$. This is a rather restricted family, since if $B\in \mathrm{SO}(n,\Z)$ then necessarily
\begin{equation}\label{bsonz}
B e_i = \pm e_{j(B,i)} \qquad 1 \le i \le n-1, \qquad Be_n=e_n,
\end{equation}
where $\{e_i\}$ is the canonical basis of $\R^n$. Such $B$ induces a permutation matrix $P_B$ given by $P_B e_i = e_{j(B,i)}$. Then, we have the decomposition
$P_B = P_{B,1} \cdots P_{B,c}$
into disjoint cycles. In other words, $c=c(B)$
is the number of orbits of the action of $B$ on the basis vectors. Clearly, $c(B)=c(B')+1\ge 2$. Notice that the matrices $B_{k(r)}$ in \eqref{Bkk'} satisfy condition \eqref{bsonz} and that
$$c(B_{k(r)}) = j_{r-1} + \cdots + j_1 + 1 = f(B_{k(r)}).$$
\sk
We now recast Donnelly's expression for the $\eta$-invariant in our present notations (with $\alpha=1$ the trivial representation).
\begin{prop}[\cite{Do}, Proposition 4.12] \label{teo.donnelly}
Let $M_\Gamma$ be a compact flat manifold of dimension $n=4h-1$ with translation lattice $\Lambda = \Z^n$
such that every $\g=BL_b \in \G$ is of the form $B=\operatorname{diag}(B',1)$ with $B'\in \mathrm{SO}(n-1,\Z)$ and $b=ae_n$, with $a\in \R$. Then,
\begin{equation}\label{doneli}
\eta^{\mathrm{[Do]}}(M_\G) = \tf{(-1)^h}{|F|} \sum_{BL_b\in (\Lambda \backslash \G)'} 2^{c(B')} \,
\Big( \prod_{j=1}^{m} \cot \big( \tf{t_j(x_B)}{2} \big) \Big) \, \cot( \pi \langle e_n, b \rangle ).
\end{equation}
\end{prop}
Note that \eqref{doneli} is formally very similar to \eqref{cyclic eta0}, which is valid for arbitrary cfm's.
Both expressions involve sums of trigonometric products at special angles. The main difference seems to be the factors $2^{c(B)}$ for $BL_b\in \G$.
\sk
Since $\Z_{2^r}$-manifolds in $\mathcal{F}$ satisfy the conditions in Proposition \ref{teo.donnelly}, they are specially suited for comparison, since Donnelly's formula apply in this case.
We will first need some trigonometric identities.
\begin{prop} \label{lema. cots}
For any $r\in \N$, $r\ge 2$ and $k$ odd, the following identity holds
\begin{equation}
\label{prodcots}
\prod_{j\in I_{2^{r-1}}^*} \cot(\tf{jk\pi}{2^r}) = 1.
\end{equation}
\end{prop}
\begin{proof}
By using $\sin 2\theta = 2 \sin \theta \cos \theta$ and \eqref{sines2}, for any $r\ge 2$, we get
\begin{equation}\label{cos}
\prod_{j\in I_{2^{r}}^*}
\cos(\tf{j\pi}{2^r}) =
\f{ \prod_{j\in I_{2^{r}}^*}
\sin(\tf{2j\pi}{2^r})} {2^{2^{r-2}} \prod_{j\in I_{2^{r}}^*}
\sin(\tf{j\pi}{2^r})} = \f{\tf{1}{2^{2^{r-2}-1}}}{2^{2^{r-2}} \tf{\sqrt 2}{2^{2^{r-2}}}} = \f{\sqrt 2}{2^{2^{r-2}}}.
\end{equation}
Now, the identities \eqref{signsin} and \eqref{prodsins} also hold for $\cos(\f{jk\pi}{N})$ changing every sine by the corresponding cosine. Therefore, the product in \eqref{prodcots} does not depend on $k$. Thus,
$\prod_{j\in I_{2^r}^*} \cot(\tf{jk \pi}{N}) = \prod_{j\in I_{2^r}^*} \cot(\tf{j \pi}{N})=1$,
by \eqref{sines} and \eqref{cos}.
\end{proof}
\begin{prop} \label{prop. sumcots}
Let $r, \ell \in \N$ with $\ell$ odd. If $N$ is either $2^r$ or $2^{r-1}$, with $r\ge 2$, then
\begin{equation}
\label{sumcots}
\sum_{j\in I_{N}^*} (-1)^{[\tf k2]} \, \cot(\tf{k \ell \pi}{2^r}) = (-1)^{[\tf{\ell}{2}]} \, \tf N2 \,.
\end{equation}
\end{prop}
\begin{proof}
We will denote by $\Sigma_{2^r}$ and $\Sigma_{2^r}'$ the sums in \eqref{sumcots} corresponding to $N=2^r$ and $2^{r-1}$, respectively.
Note that, $\cot \big( \tf{(2^r-k)\ell \pi}{2^r} \big) = -\cot \big( \f{k \ell \pi}{2^r} \big)$, and hence
$$\Sigma_{2^r} = \cot(\tf{\ell \pi}{2^r}) - \cot(\tf{3 \ell \pi}{2^r}) + \cdots +
\cot(\tf{(2^r-3) \ell \pi}{2^r}) - \cot(\tf{(2^r -1)\ell\pi}{2^r}) = 2 \, \Sigma_{2^r}' \,.$$
Thus, it is enough to prove that $\Sigma_{2^r}' = (-1)^{[\tf{\ell}{2}]}2^{r-2}$, which we will do by induction on $r$. It is immediate to check that $\cot(\tf{\ell\pi}{4}) = (-1)^{[\f{\ell}2]}$ for any odd $\ell$ and hence, $\Sigma_{2^2}'=(-1)^{[\f{\ell}2]} 2$, and the first step in the induction holds. For the general step, we have
$$\Sigma_{2^r}' = \sum_{k\in I_{2^{r-1}}^*}
(-1)^{[\tf k2]} \cot(\tf{k \ell \pi}{2^r}) = \sum_{k\in I_{2^{r-2}}^*}
(-1)^{[\tf k2]} \; \Big(\! \underbrace{\cot(\tf{k \ell \pi}{2^r}) - \cot\big( \tf{(2^{r-1}- k) \ell \pi}{2^r} \big)}_{C_{k,r}} \Big) \,.$$
Since $\cos \big( \tf{(2^{r-1}-k)\ell\pi}{2^r} \big) = (-1)^{[\f{\ell}2]} \sin(\tf{k\ell\pi}{2^r})$ and $\sin \big( \tf{(2^{r-1}-k)\ell\pi}{2^r} \big) = (-1)^{[\f{\ell}2]} \cos(\tf{k\ell\pi}{2^r})$, by using
$\cos 2\theta = \cos^2 \theta - \sin^2 \theta$ and $\sin 2\theta = 2\sin \theta \cos \theta$, we have
$$C_{k,r} = \f{\cos(\tf{k\ell\pi}{2^r})}{\sin(\tf{k\ell\pi}{2^r})} - \f{\sin(\tf{k\ell\pi}{2^r})}{\cos(\tf{k\ell\pi}{2^r})}
= \f{\cos^2(\tf{k\ell\pi}{2^r}) - \sin^2(\tf{k\ell\pi}{2^r})}{\sin(\tf{k\ell\pi}{2^r}) \cos(\tf{k\ell\pi}{2^r})}
= \f{2 \cos(\tf{k\ell\pi}{2^{r-1}})}{\sin(\tf{k\ell\pi}{2^{r-1}})} = 2 \cot(\tf{k\ell\pi}{2^{r-1}}).$$
In this way, by the inductive hypothesis we get
$$\Sigma_{2^r}' = 2 \sum_{k\in I_{2^{r-2}}^*}(-1)^{[\tf k2]} \cot(\tf{k \ell \pi}{2^{r-1}})
= 2\, \Sigma_{2^{r-1}}' = (-1)^{[\f{\ell}2]} 2^{r-2}.$$
and the result thus follows.
\end{proof}
For $\Z_{2^r}$-manifolds in $\mathcal{F}$, the expression \eqref{doneli} can be explicitly computed.
\begin{prop} \label{etaz2rdoneli}
If $M=M_{k(r)} \in \mathcal{F}$ then $\eta^{\mathrm{[Do]}}(M) = (-1)^h \, 2^{j_{r-1} + \cdots + j_1-1}$.
\end{prop}
\begin{proof}
Starting from \eqref{doneli}, and using that $\sigma_B=1$, $\ell_\g=1$, $F = \langle B \rangle$ is cyclic of order $N=2^r$, with $B=B_{k(r)}$, and that $B^k\in F_1'$ if and only if $k$ is odd, we get
\begin{equation}\label{etadon1}
\eta^{\mathrm{[Do]}}(M_\G) = (-1)^h \ 2^{c(B')-r} \, \sum_{k\in I_{2^r}^*} \,
\Big( \prod_{j=1}^{m} \cot \big( \tf{k t_j}{2} \big) \Big) \, \cot( \tf{\pi k}{2^r})
\end{equation}
since $c(B^k)=c(B)$ for any $k$ odd.
Now, by \eqref{xjr}, we have
\begin{eqnarray*}
\prod_{j=1}^{m} \cot \big( \tf{k t_j}{2} \big)
& = &
\Bigg( \prod_{k\in I_{2^r}^*}
\cot (\tf{jk\pi}{2^{r+1}}) \Bigg)^{j_{r-1}} \cdots \Bigg( \prod_{k\in I_{8}^*}
\cot (\tf{jk\pi}{16}) \Bigg)^{j_3} \Bigg( \prod_{k\in I_{4}^*}
\cot (\tf{jk\pi}{8}) \Bigg)^{j_{2}} \cot(\tf{\pi k}{4})^{j_1}
\\ & = &
(-1)^{[\tf k2]j_1} \, \prod_{i=2}^{r-1} \Bigg( \prod_{k\in I_{2^i}^*}
\cot (\tf{jk\pi}{2^{i+1}}) \Bigg)^{j_{i}} = (-1)^{[\tf k2]}
\end{eqnarray*}
where in the last equality we have used that $j_1$ is odd and Proposition \ref{lema. cots}.
Hence, by \eqref{etadon1},
$$\eta^{\mathrm{[Do]}}(M_\G) = (-1)^h \, 2^{c(B')-r} \, \sum_{k\in I_{2^r}^*} (-1)^{[\tf k2]} \, \cot( \tf{\pi k}{2^r})$$
and thus, by Proposition \ref{prop. sumcots},
$\eta^{\mathrm{[Do]}}(M_\G) = (-1)^h \, 2^{c(B')-r} \, 2^{r-1} =(-1)^h \ 2^{j_{r-1}+\cdots + j_1 -1}$,
as asserted.
\end{proof}
\begin{rem} \label{etadomp}
(i) In relation to (and continuing with) Remark \ref{eta ell}, notice that taking $\ell_\g=\ell$ odd in \eqref{etadon1}, and
using Proposition \ref{prop. sumcots}, we have $\eta^{[\mathrm{Do}]} (M_\ell) = (-1)^{[\ell/2]} \, \eta^{[\mathrm{Do}]} (M)$.
\sk (ii) By Corollary \ref{prop.etasz2r} and Proposition \ref{etaz2rdoneli} we have
$\eta^{\mathrm{[Do]}} = (-1)^{h+1} \, \eta^{\mathrm{[MP]}}$. In this way, we see that there is a difference in sign between both results, precisely when $h$ is odd, i.e.\@ in dimensions $n \equiv 3$ mod 8.
\end{rem}
Donnelly's expression \eqref{doneli} works for any cfm whose integral holonomy representation is restricted to $\mathrm{SO}(n,\Z)$.
In light of the previous results, one may ask if Proposition \ref{teo.donnelly} can be generalized to hold with more generality.
For $\Z_{2^r}$-manifolds, by Propositions \ref{prop.etasz2r} and \ref{etaz2rdoneli} and Lemma \ref{lemin} we see
that Proposition \ref{teo.donnelly} holds for arbitrary $\Z_{2^r}$-manifolds; that is, for any lattice and any integral holonomy representation. However, the condition on the holonomy representation cannot be removed for $F \not \simeq \Z_{2^r}$,
as the next example shows.
\begin{ejem}
Up to diffeomorphism, there is only one $\Z_3$-manifold in dim 3, the \emph{triscosm} $M_3 = \Gamma_{3} \backslash \R^3$ where $\G_3 = \langle BL_{\f{e_3}{3}}, L_{\Ld} \rangle$, $B = \textrm{diag}(B',1)$, $B' = ( \begin{smallmatrix} 0&-1 \\ 1&-1 \end{smallmatrix})$, and
$\Ld = \Ld_{_\mathcal{H}} \oplus \Z e_3$, with $\Ld_{_\mathcal{H}} = \Z e_1 \oplus \Z f_2$, $f_2= -\f12 e_1 + \f{\sqrt 3}2 e_2$, the plane hexagonal lattice. It is known that $\eta(M_3)=-\tf 23$ (\cite{LR}, also \cite{MP.AGAG12}, \cite{Sz2}). However, applying \eqref{doneli}, since $x_{B^k}=x(\tf{2k\pi}3)$ and $c(B^k)=2$, $k=1,2$, one gets
$$\eta^{[\mathrm{Do}]}(M_3) = -\tf 23 \sum_{k=1,2} \cot(\tf{k\pi}{3})^2 = -\tf 23 (\tf 13+\tf 13) = -\tf 49.$$
Here, $B' \notin \textrm{SO}(2)$ and, after conjugation, we get $\big( \begin{smallmatrix} \cos(2\pi/3)&-\sin(2\pi/3) \\
\sin(2\pi/3)& \cos(2\pi/3) \end{smallmatrix} \big) \in \textrm{SO}(2) \smallsetminus \mathrm{SO}(2,\Z)$.
\end{ejem}
The previous example is a particular case of $\Z_p$-manifolds, $p$ odd prime (classified by Charlap \cite{Ch}). See \cite{GMP}, \cite{MP.TAMS06}, \cite{MP.PAMQ09}, \cite{MP.JGA11} and \cite{MP.AGAG12} for details on the construction/classification of $\Z_p$-manifolds, topological properties and computations of $\eta(s)$ and $\eta$ for both the spin Dirac and the APS operators.
\begin{rem} \label{zpmanifolds}
Donnelly's expession \eqref{doneli} cannot be applied to $\Z_p$-manifolds, $p$ odd prime, because $\Z_p$-manifolds having non-trivial $\eta$-invariant have integral holonomy representations taking values in $\mathrm{GL}(n,\Z) \smallsetminus \mathrm{SO}(n)$
(see \S2 in \cite{MP.PAMQ09}, \S4 in \cite{MP.AGAG12}). As before, one can show (with a little bit more effort) that by applying expression \eqref{doneli} anyhow we get a different result for $\eta$ that when using formula \eqref{cyclic eta0}.
\end{rem}
\subsection*{Acknowledgments}
I am very grateful to Professor Roberto Miatello for useful conversations which led this paper to the final form.
In particular, I am in debt with him for pointing out to me the subfamily of manifolds in $\mathcal{F}$ determined by the matrices $B'$ in \eqref{Bj}.
|
\section{Introduction}
The iron-based high-$T_c$ superconductors (IBSs) can be classified into two main groups: iron pnictides and iron chalcogenides. \cite{review-AdvPhys,review-NatPhys,review-RMP}
Following the IBSs, it was found that most of their nickel-based counterparts with the same crystal structure are also superconducting. \cite{NBS1,NBS2,NBS3,NBS4}
The nickel-based superconductors (NBSs) exhibit some distinct features: (i) the superconducting transition temperature $T_c$ is pretty low, usually lower than 5 K; \cite{RonningF} (ii) the Fermi surfaces are more complicated and three-dimensional; \cite{Terashima,Subedi,ZhouB} (iii) there is no evidence for the existence of an antiferromagnetic order neighboring the superconducting state.
It is important to know whether the pairing mechanism of NBSs is different from IBSs. Clarifying the superconducting gap symmetry and structure will provide important clues. For the IBSs, it has been shown that the gap structure is quite elusive, varying substantially from member to member and as a function of doping. \cite{RPP} While most IBSs have multiple nodeless gaps (likely $s_\pm$-wave), some of them show nodal superconductivity. \cite{RPP} In the case of NBSs, however, the superconducting gap appears more ``conventional". Both specific heat and thermal conductivity measurements suggest fully gapped $s$-wave superconductivity in BaNi$_2$As$_2$. \cite{Kurita} Isovalent phosphorus doping does not change its gap structure, \cite{BaNi2AsP} in contrast to that observed in nodal superconductor BaFe$_2$(As$_{1-x}$P$_x$)$_2$. \cite{Hashimoto} Low-temperature magnetothermal conductivity $\kappa(T,H)$ measurements also rule out the presence of nodes in the superconducting gap of SrNi$_2$P$_2$. \cite{NKurita}
While fully gapped $s$-wave superconductivity seems to be a universal feature of nickel pnictides, it is not so clear for nickel chalcogenides, in which heavy-electron behavior was observed. \cite{Neilson,R.Neilson,FangMH} In the three nickel chalcogenides, KNi$_2$Se$_2$ ($T_c$ $\simeq$ 0.80 K), KNi$_2$S$_2$ ($T_c$ $\simeq$ 0.46 K), and TlNi$_2$Se$_2$ ($T_c$ $\simeq$ 3.7 K), the electronic specific-heat coefficient $\gamma$ is 44, 68, and 40 mJ mol$^{-1}$ K$^{-2}$, respectively. \cite{Neilson,R.Neilson,FangMH} The estimated effective electron mass $m^{\ast}$ can be as high as 24$m_e$ in KNi$_2$S$_2$. \cite{R.Neilson} The $\gamma(H)$ of TlNi$_2$Se$_2$ exhibits a square root field dependence, which is usually seen in nodal superconductors.\cite{FangMH} Therefore, it is of great interests to investigate the superconducting gap structure of nickel chalcogenides. For KNi$_2$Se$_2$ and KNi$_2$S$_2$, only polycrystalline samples were synthesized so far. Fortunately, sizable high-quality single crystals of TlNi$_2$Se$_2$ have been successfully grown. \cite{FangMH}
Low-temperature heat transport is an established bulk technique to study the superconducting gap structure.\cite{Louis}
In this paper, we present the thermal conductivity measurements of TlNi$_2$Se$_2$ single crystals down to 50 mK ($\sim$ $T_c$/70). The relatively tiny residual linear term in zero field and its slow field dependence in low fields suggest nodeless superconducting gap. In the whole field range, the $\kappa_0(H)/T$ curve shows a concave to convex evolution (``$S$'' shape), which was previously also observed in BaNi$_2$As$_2$ and SrNi$_2$P$_2$. Multi-gap superconductivity is introduced to explain this common behavior of NBSs.
\section{Experiment}
Single crystals of TlNi$_2$Se$_2$ were grown using the self-flux method. \cite{FangMH} The dc magnetization was measured by a superconducting quantum interference device (MPMS, Quantum Design). Two samples, labeled as A and B, were used in the transport measurements. The two samples were cleaved to rectangular shape with dimensions of $\sim$2.0 $\times$ 0.5 mm$^2$ in the \emph{ab} plane and $\sim$40 $\mu$m along the \emph{c} axis. Contacts were made directly on the sample surfaces with silver paint, which were used for both resistivity and thermal conductivity measurements. The contacts are metallic with typical resistance of 10 m$\Omega$ at 2 K. In-plane thermal conductivity was measured in a dilution refrigerator, using a standard four-wire steady-state method with two RuO$_2$ chip thermometers, calibrated \emph{in situ} against a reference RuO$_2$ thermometer. Magnetic fields were applied along the \emph{c} axis and perpendicular to the heat current. To ensure a homogeneous field distribution in the samples, all fields were applied at temperatures above $T_c$ for transport measurements.
\section{Results and discussion}
\begin{figure}
\includegraphics[width=0.4\textwidth]{Fig1.eps}
\caption{(Color online) (a) Low-temperature dc magnetization of TlNi$_2$Se$_2$ single crystal measured with zero-field-cooled (ZFC) and field-cooled (FC) processes. (b) Low-temperature in-plane resistivity of TlNi$_2$Se$_2$ single crystal (sample A). The solid line is a fit of the data between 4 and 8.5 K to $\rho = \rho_0 + AT^2$. Inset shows the $\rho$($T$) curves of sample A and B up to room temperature. After normalizing the value of sample B at 290 K to that of sample A, the two curves are nearly identical.}
\end{figure}
Figure 1(a) shows the low-temperature dc magnetization of TlNi$_2$Se$_2$ single crystal. The onset of the superconducting transition is at 3.7 K. The sharp drop of diamagnetic signal and its quick saturation (below 3.2 K) indicate the sample is of high quality.
Figure 1(b) plots the in-plane resistivity of TlNi$_2$Se$_2$ samples. To reduce the uncertainty associated with geometric factor, we normalize the resistivity of sample B to sample A at $T$ = 290 K. The two resistivity curves are nearly identical after normalization, as seen in the inset. Later we will use the normalized geometric factor for sample B. The resistivity decreases monotonically with lowering the temperature, followed by a sharp superconducting transition. The $T_c$ defined by $\rho = 0$ is 3.7 K, which is consistent with the onset of diamagnetic transition. Fermi-liquid behavior $\rho \sim T^2$ is observed at low temperature. The fit of $\rho(T)$ data between 4 and 8.5 K to $\rho = \rho_0 + AT^2$ gives the residual resistivity $\rho_0$ = 0.60 $\mu\Omega$ cm for sample A and $\rho_0$ = 0.61 $\mu\Omega$ cm for sample B. The residual resistivity ratio (RRR) is about 120, which is much higher than the nickel pnictides BaNi$_2$As$_2$ and SrNi$_2$P$_2$. \cite{Kurita,NKurita}
\begin{figure}
\includegraphics[width=0.447\textwidth]{Fig2.eps}
\caption{(Color online) Low-temperature thermal conductivity of TlNi$_2$Se$_2$ single crystals.
(a) Sample A and sample B in zero field.
(b) Thermal conductivity of sample A under magnetic fields up to $H$ = 1 T. Data below $H_{c2}$ = 0.8 T are fitted to $\kappa/T$ = $\kappa_0/T$ + $bT^2$, as represented by the solid lines.
The dashed line is the normal-state Wiedemann-Franz law expectation $L_0/\rho_0$, with $L_0$ = 2.45 $\times$ 10$^{-8}$ W $\Omega$ K$^{-2}$ and $\rho_0$ = 0.60 $\mu\Omega$ cm.}
\end{figure}
Figure 2(a) shows the low-temperature thermal conductivity of TlNi$_2$Se$_2$ sample A and B at zero field, plotted as $\kappa/T$ versus $T^2$.
The measured thermal conductivity $\kappa$ can be expressed as $\kappa$ = $\kappa_e$ + $\kappa_{ph}$, the sum of electron contribution $\kappa_e$ and
phonon contribution $\kappa_{ph}$. Due to their distinct temperature dependence at low temperatures, the two contributions can be well separated by fitting the data to
\begin{eqnarray}
\kappa = aT + bT^\alpha,
\end{eqnarray}
where $aT$ is the residual linear term of electrons and $bT^{\alpha}$ is the phonon contribution in the boundary scattering limit. Usually $2 < \alpha \leq 3$, which depends on the effect of specular reflection of phonons at the sample surfaces. \cite{Mike,LiSY} For both samples A and B, the fitting parameter $\alpha$ in zero field is very close to 3, therefore we fix it to 3. Note that for BaNi$_2$As$_2$ and SrNi$_2$P$_2$ single crystals, the parameter $\alpha$ is also 3. \cite{Kurita,NKurita} It seems that the effect of specular reflection of phonons at the surfaces is very weak for NBS single crystals.
In Fig. 2(a), the fittings give $\kappa_0/T \equiv a$ = 0.18 $\pm$ 0.02 mW K$^{-2}$ cm$^{-1}$ and 0.05 $\pm$ 0.01 mW K$^{-2}$ cm$^{-1}$ for sample A and B, respectively. Comparing with our experimental error 0.005 mW K$^{-2}$ cm$^{-1}$,\cite{LiSY} these $\kappa_0/T$ values are not negligible. However, they are actually very tiny, if we compare them with the normal-state Wiedemann-Franz law expectation $\kappa_{N0}/T$ = $L_0/\rho_0 \approx$ 40 mW K$^{-2}$ cm$^{-1}$. The ratio ($\kappa_0/T$)/($\kappa_{N0}/T$) of TlNi$_2$Se$_2$ is only 0.44$\%$ (sample A) and 0.12$\%$ (sample B). For nodal superconductors, a substantial $\kappa_0/T$ in zero field contributed by the nodal quasiparticles has been found.\cite{Proust,Suzuki,Hill} For example, $\kappa_0/T$ of the overdoped ($T_c$ = 15 K) $d$-wave cuprate superconductor Tl$_2$Ba$_2$CuO$_{6+\delta}$ (Tl-2201) is 1.41 mW K$^{-2}$ cm$^{-1}$, $\sim$36\% $\kappa_{N0}/T$. \cite{Proust} For the $p$-wave superconductor Sr$_2$RuO$_4$ ($T_c$ = 1.5 K), $\kappa_0/T$ = 17 mW K$^{-2}$ cm$^{-1}$ was reported, more than 9\% $\kappa_{N0}/T$. \cite{Suzuki} The multi-gap nodal heavy-fermion superconductor PrOs$_4$Sb$_{12}$ has $\kappa_0/T$ = 0.46 mW K$^{-2}$ cm$^{-1}$, $\sim$7$\%$ $\kappa_{N0}/T$. \cite{Hill} In this context, the tiny percentage of ($\kappa_0/T$)/($\kappa_{N0}/T$) observed in TlNi$_2$Se$_2$ suggests that the very small $\kappa_0/T$ may not come from nodal quasiparticals.
In fact, the finite value of $\kappa_0/T$ in zero field can be theoretically estimated for a quasi-two-dimensional \emph{d}-wave superconductor: \cite{th1,th2}
\begin{eqnarray}
\frac{\kappa_0}{T}\simeq\frac{\hbar}{2\pi} \frac{\gamma_Nv_F^2}{\triangle_0},
\end{eqnarray}
where $\gamma_N$ is the electronic specific heat coefficient in the normal state, $v_F$ is the Fermi velocity, and $\triangle_0$ stands for the maximum of the superconducting gap.
$\upsilon_F$ = 5.48 $\times$ 10$^4$ m s$^{-1}$, $\gamma_N$ = 40 mJ mol$^{-1}$ K$^{-2}$ and $\triangle_0 = 2.01 k_BT_c$ can be obtained from a former work. \cite{FangMH} In case that TlNi$_2$Se$_2$ is a quasi-two-dimensional \emph{d}-wave superconductor, we estimate $\kappa_0/T$ $\simeq$ 3.22 mW K$^{-2}$ cm$^{-1}$, which should be $\sim$8$\%$ $\kappa_{N0}/T$ of our samples. This value is much higher than what we observed in both sample A and B, therefore, the superconducting gap of TlNi$_2$Se$_2$ is not consistent with the $d$-wave scenario. The very small $\kappa_0/T$ in zero field may result from tiny non-superconducting impure phase in the samples.
\begin{figure}
\includegraphics[width=0.44\textwidth]{Fig3.eps}
\caption{(Color online) (a) Normalized $\kappa_0/T$ of TlNi$_2$Se$_2$ as a function of $H/H_{c2}$. For comparison, similar data are shown for the single band $s$-wave superconductor Nb, \cite{Lowell} the multiband $s$-wave superconductor NbSe$_2$, \cite{Boaknin} the $d$-wave superconductor Tl-2201, \cite{Proust} two nickel-pnictide superconductors BaNi$_2$As$_2$ and SrNi$_2$P$_2$. \cite{Kurita,NKurita}
(b) Ratio of thermal conductivity to heat capacity ($\kappa_0/T$)/($C_{res}/T$) in TlNi$_2$Se$_2$. The end of the rapid increase regime at $H^{\ast}$ indicates the complete suppression of the smaller gap(s).}
\end{figure}
The field dependence of $\kappa_0/T$ can provide further information of the superconducting gap structure.\cite{Louis} The thermal conductivity under magnetic fields for sample A is shown in Fig. 2(b). Similar results are obtained for sample B but not shown here, since the data is noisier at $H$ $>$ 0.2 T. Upon applying magnetic fields, vortices are gradually introduced into the sample. The unpaired electrons inside the vortices contribute to $\kappa_0(H)/T$.
Although the curves of 0.8 and 1 T is not smooth, one can still see that $\kappa_0/T$ roughly meets the Wiedemann-Franz law expectation $L_0/\rho_0$ = 40.8 mW K$^{-2}$ cm$^{-1}$. We determine the bulk upper critical field $H_{c2}$ = 0.8 T, which agrees with the value estimated from resistivity measurements. \cite{FangMH} The data in different fields below $H_{c2}$ are also fitted to $\kappa/T = \kappa_0/T + bT^2$, as represented by the solid lines in Fig. 2(b).
Normalized $\kappa_0(H)/T$ of TlNi$_2$Se$_2$ as a function of $H/H_{c2}$ is presented in Fig. 3(a), together with the single band $s$-wave superconductor Nb, \cite{Lowell} the multiband $s$-wave superconductor NbSe$_2$, \cite{Boaknin} the $d$-wave curpate superconductor Tl-2201, \cite{Proust} and two nickel-pnictide superconductors BaNi$_2$As$_2$ and SrNi$_2$P$_2$. \cite{Kurita,NKurita}
For single band $s$-wave superconductor Nb, the $\kappa_0(H)/T$ changes little even up to 40\% $H_{c2}$. \cite{Lowell}
While in nodal superconductor Tl-2201, small field can yield a quick growth in the quasiparticle density of states (DOS) due to Volovik effect, \cite{Volovik} and the low field $\kappa_0(H)/T$ is roughly $\sqrt{H}$ dependant. \cite{Proust}
In the case of NbSe$_2$, the distinct $\kappa_0(H)/T$ behavior was well explained by multiple superconducting gaps with different magnitudes. \cite{Boaknin}
Of the above three archetypal examples, the field dependence of $\kappa_0(H)/T$ for TlNi$_2$Se$_2$ resembles NbSe$_2$ most. Since NBSs also have several bands across the Fermi level, \cite{Terashima,ZhouB,Subedi,RonningF} it is natural to explain the $\kappa_0(H)/T$ behavior of TlNi$_2$Se$_2$ with multiple gaps. As was done in the case of NbSe$_2$, \cite{Boaknin} one can estimate the ratio of smaller gap to larger gap $\triangle_s/\triangle_l$ by plotting the ratio of thermal conductivity to heat capacity ($\kappa_0/T$)/($C_{res}/T$) as a function of $H/H_{c2}$.
In the vortex state, residual specific heat is associated with the unpaired electron DOS, while $\kappa_0/T$ manifests the tunneling ability of those unpaired electrons. Thus ($\kappa_0/T$)/($C_{res}/T$) represents the degree of delocalization of quasiparticles in the vortex state.\cite{Boaknin}
In a simple two-gap model, both Fermi sheets contribute to the rise of ($\kappa_0/T$)/($C_{res}/T$) below a characteristic field $H^{\ast}$. In the regime of $H^{\ast}$ $<$ $H$ $<$ $H_{c2}$, the smaller gap is completely suppressed and only the Fermi sheet with larger gap contributes to the rise of ($\kappa_0/T$)/($C_{res}/T$).
In Fig. 3(b), we plot the ratio ($\kappa_0/T$)/($C_{res}/T$) as a function of $H/H_{c2}$. The residual specific heat $C_{res}(H)/T$ is adopted from Ref. 19.
Two distinct regimes can be resolved: after a rapid initial increase, ($\kappa_0/T$)/($C_{res}/T$) reaches a weak $H$-dependent regime. The end of the rapid increase at $H^{\ast}$ $\simeq$ 0.36 $H_{c2}$ indicates the complete suppression of the smaller gap. Considering that the upper critical field is related to the superconducting gap by $H_{c2}$ $\propto$ $\triangle^2$/$v_F^2$, the characteristic field $H^{\ast}$ allows us to estimate the gap ratio $\triangle_s$/$\triangle_l$ $\simeq$ 0.6. In Ref. 19, the specific heat data can be best fitted by two-gap BCS model with $\triangle_s$/$\triangle_l$ $\simeq$ 0.42, \cite{FangMH} which is qualitatively consistent with our thermal conductivity analysis.
We then compare the $\kappa_0(H)/T$ behavior of TlNi$_2$Se$_2$ to those of BaNi$_2$As$_2$ and SrNi$_2$P$_2$. \cite{Kurita,NKurita} From Fig. 3(a), the three $\kappa_0(H)/T$ curves show a common ``$S$'' shape (concave in low fields and convex in high fields). Previously, this ``$S$''-shape curve of $\kappa_0(H)/T$ was interpreted as the consequence of $H_{c2}$ distribution in the BaNi$_2$As$_2$ and SrNi$_2$P$_2$ crystals due to the sample quality. \cite{Kurita,NKurita}
While for our TlNi$_2$Se$_2$ single crystals, the sharp diamagnetic transition shown in Fig. 1(a) suggests that there should be no $H_{c2}$ distribution, despite the possible existence of tiny impure phase. The RRR of TlNi$_2$Se$_2$ is also much higher than those of BaNi$_2$As$_2$ and SrNi$_2$P$_2$, pointing to cleaner sample. For TlNi$_2$Se$_2$, the Ginzberg-Landau coherence length $\xi$ = 20.3 nm is calculated from the equation
\begin{eqnarray}
\xi = [\frac{\Upphi_0}{2\pi H_{c2}(0)}]^{\frac{1}{2}},
\end{eqnarray}
where $\Upphi_0$ = 2.07 $\times$ 10$^{-7}$Oe cm$^2$ is the flux quantum.
According to the relationship
\begin{eqnarray}
\frac{\kappa}{T} = \frac{1}{3}\gamma\upsilon_Fl_e,
\end{eqnarray}
the electron mean free path $l_e$ = 677 nm is estimated. The ratio $l_e$/$\xi$ = 33.3 ($\gg$ 1) places our TlNi$_2$Se$_2$ single crystal in the clean limit. Therefore the ``$S$''-shape $\kappa_0(H)/T$ curve of TlNi$_2$Se$_2$ should not be explained by bad sample quality. Since all three compounds have the same crystal structure and similar electronic structure, \cite{RonningF} their common ``$S$''-shape field dependence of $\kappa_0(H)/T$ may have the same origin --- the multiple nodeless gaps.
We note that the normalized $\kappa_0(H)/T$ has been numerically simulated for two band $s$-wave state with unequal gap sizes, which successfully explained the experimental data of Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ with different Co-doping by systematically varying the ratio $\triangle_s$/$\triangle_l$. \cite{Bang} However, this kind of calculation can not reproduce the pronounced ``$S$''-shape $\kappa_0(H)/T$ curve in Fig. 3(a), since the simulation is based on an assumption that each band possesses equal weighting of quasiparticle DOS. \cite{Bang} To get the ``$S$''-shape $\kappa_0(H)/T$ curve, one may need to assume that those bands with smaller gap possesses more quasiparticle DOS than those bands with larger gap. Further numerical simulations are needed to reproduce this common feature observed in NBSs.
\section{Conclusion}
In summary, we have measured the thermal conductivity of TlNi$_2$Se$_2$ single crystals down to 50 mK.
The relatively tiny $\kappa_0/T$ and weak field dependance of $\kappa_0(H)/T$ in low fields suggest nodeless superconducting gap.
The $\kappa_0(H)/T$ curve shows an ``$S$''-shape, which was previously also observed in BaNi$_2$As$_2$ and SrNi$_2$P$_2$. This common feature of nickel-based superconductors is explained by multiple nodeless superconducting gaps. A characteristic field $H^{\ast} \simeq$ 0.36 $H_{c2}$ was identified from apparent slope change in ($\kappa_0/T$)/($C_{res}/T$), which gives the ratio $\triangle_s$/$\triangle_l \simeq$ 0.6 in TlNi$_2$Se$_2$.
\\
\\
\begin{center}
{\bf ACKNOWLEDGEMENTS}
\end{center}
This work is supported by the Natural Science Foundation of China (Grants No. 91021016, No. 91221303, No. 11374261, and No. 11204059), the Ministry of Science and Technology of China (National Basic Research Program No. 2012CB821402, No. 2012CB821404, and No. 2011CBA00103), and the Program for Professor of Special Appointment (Eastern Scholar) at Shanghai Institutions of Higher Learning.
|
\section{Introduction}
There is a qualitative transition in the sophistication of the mathematical tools required to describe mechanical systems of a few free bodies to those dominated by
the behavior of continuous fields.
Second-order ordinary differential equations are indispensable in the study of point-system mechanics, particularly in a first year course, and many curricula are
organized around first understanding the properties of these equations (particularly in the case in which the equation of motion has the form
$\ddot{x} = const$).
In a first course in electromagnetism, the fundamental quantities to understand are the distributed electromagnetic fields, $\vec{E},\vec{B}$, with
dynamics given by Maxwell's Equations using the language of partial differential equations with appropriate boundary conditions. Even very good companion works can give the impression to undergraduates that the ideas of ``divergence, gradient, and curl'' have their primary (perhaps only) role in describing electricity and magnetism. \cite{dgcandallthat}
Thus, beginning students get the impression that these mathematical tools are somehow wedded to the separate subjects of the physics canon, rather than being independent, widely applicable tools.
My aim here is to present a mechanics problem which is usually dealt with in the algebra-based mechanics curriculum and recast it using the mathematical language
students expect in electricity and magnetism. I have two aims. First, I will connect Archimedes' principle (usually developed as a ``just-so" statement of how a ``buoyant force" works) to real forces acting on real bodies.
The ideas of buoyancy, density, and volume are known areas of difficulty for science teaching candidates, K-12 students, and even quite advanced students majoring
in physics. \cite{loverude1,loverude2}
As we shall see, the buoyant force is merely a consequence of the superposition of many forces acting
on the outer edge of an object, and thus is no more mysterious than any other force in mechanics.
Secondly, I wish to give an undergraduate a context in which vector analysis has a definite application in the most familiar, concrete physics a student knows, namely mechanics.
Thus, the integration of these mathematical tools and the ``net forces" ideas in this nexus of problems and placing that integrated
understanding firmly within the grasp
of an intermediate-level student of physics is the goal.
It is not my intention to merely display yet another derivation of the principle, although I will certainly do that.
Rather, it is my intention to integrate this basic idea into our curriculum in a fundamental way.
If this work is used to merely guide a new set of derivations in a lecture hall, its purpose will be unfulfilled.
There is a growing consensus across the
academy that the value of integrating methods and content, and creating a unified, flexible intellectual framework supporting lifelong learning is
the eventual goal of a baccalaureate course of study --- even in professional and STEM fields, perhaps {\it particularly so}.
Too often a university education can be experienced as a list of topics or courses logically disconnected from each other.
The work of integrating the education into a seamless whole can be left to a student, and may or may not occur. \cite{aacu_integrate}
Physics is particularly blessed in this regard, in that our discipline celebrates the unification of disparate phenomena in a way that few disciplines have a history of doing.
There are a number of excellent introductory texts that take this integrated point of view as opposed to a historical development. The ``right"
ideas are developed at the beginning of study, and the special cases amenable to analysis (exact or approximate) are labeled as such.\cite{mandi}
In the physics program at CSU Long Beach where I am a faculty member, we have for a number of years been intentional with this very issue, stemming from our activity as a PhysTEC supported site
in 2010.\cite{phystec}
One element of this program was to develop a Learning Assistant Program on the Colorado model, along with a course in physics pedagogy \cite{phys390}
we use to train physics majors to teach in our lower-level laboratories as Learning Assistants. \cite{la_model}
A part of that course, and the key to explaining to physics majors why this low-level material is so hard to learn (and teach) is discussing
the excellent article by Reddish on mental models in physics learning. \cite{reddish}
One of the most striking features of this treatment is the realization that mental structures of physics in a student's head are incomplete, disorganized, and even for very well-trained students (and faculty) there exists ``holes'' in even very basic concepts.
In taking this course, students are required to be present in a science tutoring center, and as is often the case, have duties to help students with courses they themselves had never taken. Algebra-based mechanics and K-6 level physical sciences are never taken by physics students, and we assume that advanced students (with
advanced mathematical and physics tools) should be able to handle the more elementary ideas in those ``precursor'' courses.
It is a relatively simple problem that stumped my stable of learning assistants that prompted the development below --- a problem using Archimedes' principle.
It is my experience that anything involving fluids is inordinately tricky to explain.
The basic functioning of a centrifuge can be understood at an engineering level and qualitatively well enough for the devices to work correctly in the hands of research-caliber scientists and engineers.
Yet, the explanation in terms of hydrostatic equilibrium in a non-inertial frame is subtle with tendrils reaching up into very sophisticated mechanics.\cite{taylor,goldstein}
Of the many treatments of the principle (starting with a good exposition of Archimedes' words themselves \cite{whatsaid} and even quite interesting
derivations centering on energy ideas \cite{apfrome}) not many cement the principle into the basic foundation of mechanics
many students understand and can already use quite well.
Even in a static case, in which Archimedes' principle and buoyancy should be sufficient to solve many problems, my group of junior level physics students ``knew'' Archimedes' principle, but were unable to apply it.
And, embarrassingly enough, likewise for myself.
Only when I had put the problem into the language of vector analysis did Archimedes' principle actually make sense to me (in the sense that I had connected it to my own ideas of mechanics).
The point of this paper is to elucidate the conditions that lead me to ``reconstruct'' a theorem of vector calculus Archimedes himself must have discovered: \cite{joke}
\begin{equation}
\varoiint d \vec{A} p = \iiint dv \vec{\nabla} p
\label{at}
\end{equation}
which some will recognize as the three-dimensional version of the ``mean-value theorem,'' but which I will simply regard as ``Archimedes' Theorem" ---
a fitting companion to Gauss' law.
I will first describe the elementary problem that revealed to us all that we had a buoyancy gap in our understanding.
Then I will describe hydrostatic equilibrium and the new vector identity, and in the discussion describe some applications and ``clicker'' style questions.
\section{Elementary Problem}
The problem that send my physics major Learning Assistants off into a cycle of recursive algebra is a relatively straightforward problem encountered in the algebra-based mechanics sequence. Here is a variation:
{\it A hot air balloon of volume $V$ is in equilibrium surrounded by air of density $\rho_a$. At $t=0$ ballast of mass $m_o$ is released from the balloon, and it is observed to accelerate upwards with an acceleration $a$. What is the final mass of the balloon (in terms of $m_o, \rho_a, g, V$)?}
These were students, by and large, who had just taken our junior-level course in mechanics, or had completed a course in introductory modern physics and differential equations, so there was no issue with their preparation.
Yet, they reported that they spent an inordinate amount of time circling around the problem without coming to a resolution, much to the consternation of the the first-year students who were hoping to get some help on the problem in our tutoring center.
What my students knew, and what they had connected to the stock of problems they had personally solved, were two different things. It was only the presence of our Master Teacher in Residence, Ms. Kathryn Beck of Bolsa Grande High School that saved our collective bacon by remarking that the buoyant force is just $\rho_a V g$ both before the ballast had been dropped and after.\cite{tir} Given that this force is the same before and after leads immediately to the observation that
\begin{equation}
M = m_o g / a
\end{equation}
In short, my students had ``heard'' of Archimedes' principle (as I had myself), but in never having had to use it in a physical context (this sort of problem has fallen out of our ``calculus''-based mechanics curriculum).
They were quite confident in their use of free body diagrams, but the nature of the buoyant force was not connected from the beginning to the end of the problem.
And, once I had the benefit of Ms. Beck's observation (she was clearly very, very familiar with this sort of problem, having taught it many years in many guises and having
charged generations of students to construct and predict the displacement of cardboard boats) I was left with a conundrum. There is in reality
no {\bf buoyant} force, there is only the effect of the surrounding air pressure on the balloon. As I had drilled into my own mind in the course of working through my own Ph.D. thesis \cite{pickett}, only
gradients of the pressure field give forces.
The presence of a pressure gradient is the {\bf only} way the air can affect the balloon, not through an unsystematic
``watch me pull a rabbit out of a hat'' buoyant force.
\section{Pressure}
And, here is how I connect the buoyant force and the presence of a non-uniform pressure field.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.5]{schematic.eps}
\caption{A small cube of the ambient air at density $\rho$ is subjected to an inward force on its upper face, an upward force on its lower face, and its weight.
An abstract ``net force" or ``free body" diagram is at right.}
\label{schematic}
\end{figure}
The air around the balloon is in hydrostatic equilibrium. Thus, a small cube of air with edge lengths $dx, dy, dz$ is in equilibrium when:
\begin{equation}
(\rho_a dx dy dz) g = P(x,y,z) dx dy - P(x,y,z+dz) dx dy
\end{equation}
Here, the pressure on the top face of the box produces a downward force, and the pressure on the bottom face produces an upward force.
Rearranging a bit gives a differential equation for the pressure:
\begin{equation}
\rho_a g = - \partial_z P
\label{pgrad}
\end{equation}
with the familiar solution:
\begin{equation}
P = P_o - \rho_a g z,
\label{hydro}
\end{equation}
where $P_o$ is the pressure at $z=0$. This treatment is elementary enough, certainly not worth writing a manuscript to share.
So, now we know the pressure field around the balloon. What is the microscopic mechanism of transferring force to the balloon from this pressre field? Again, mirroring the constructions in Ref.~\cite{dgcandallthat} I consider an infinitesimal element of the surface area (with normal directed outward) of the ballon, $d\vec{A}$.
As in Figure~\ref{at_schematic}, the air puts a force on this area element, $d\vec{F} = -P d\vec{A}$.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.5]{at_fig2.eps}
\caption{An arbitrary shaped object is impinged by pressure forces from the
surrounding medium. Given the pressure field, $P(\vec{r})$, our task is to
add up all of the forces resulting from each infinitesimal directed area element,
$d\vec{A}$. Also shown schematically is a ``box discretization'' of the sort
used to motivate theorems in vector calculus.\cite{dgcndallthat}}
\label{at_schematic}
\end{figure}
Note that there is a pressure field inside the balloon as well, but if we are concerned with the balloon as the system, all such internal forces cancel in pairs --- a distinct advantage of the Matter and Interactions curriculum\cite{mandi} is the clear demarcation between system and surroundings in every phsyical situation.
Thus, the total force from the air is a sum over all of these small air forces:
\begin{equation}
\vec{F} = \varoiint - d\vec{A} P
\label{fair}
\end{equation}
This looks suspiciously like the ``flux'' expression in Gauss' Law, except $\vec{F}_a$ is a vector
quantity and not a scalar formed by the flux of a vector field through a closed surface.
Here is the clever bit, which generates a new (to me) theorem of vector calculus. Let us dot bot sides of Eq.~\ref{fair} with $\hat{z}$:
\begin{equation}
F_z = \varoiint - d \vec{A} \cdot (P \hat{z}) =
\iiint dv -\vec{\nabla} ( P \hat{z} ).
\end{equation}
The second equality follows form the divergence theorem applied to the vector field $P \hat{z}$.
Thus, we have
\begin{equation}
F_z = - \iiint dv \partial_z P = - (-g) \iiint dv \rho_a = g M_a,
\end{equation}
where the pressure gradient in Eq.~\ref{pgrad} is used in the second equality.
Eureka.
This gives just the $z$ component of the total force, and in this case, as the pressure has no gradients in either the $x$ nor $y$ directions, this is all the analysis can tell us, although if $P$ depends on each of $x,y, z$ nontrivially, we have derived the the full version of Archimedes' theorem,
Eq.~\ref{at}.
\section{Discussion}
The development of Archimedes' Principle as a consequence of hydrostatic equilibrium, Eq.~\ref{hydro} and a generalization of the divergence theorem,
Eq.~\ref{at} is clearly within the ability of an advanced undergraduate to use, and, in my opinion, relaxes the appearance of the Archimedes' Principle from an {\it ad
hoc} statement and grounds it solidly upon the development of general equilibrium, forces, and superposition of forces.
It is altogether more clear to an advanced student {\it why the principle is true} which allows it to be used in different contexts.
The design of numerous hand-made ``bob'' accelerometers is a case in point. Here, an object less dense than water is submerged completely, yet tethered to the
bottom of a clear, plastic bottle.
The bob reacts quickly to an applied acceleration ... accelerate forward, and the bob immediately leans forward, and the magnitude of the
acceleration can be measured by the angle ot the tilt.
I have seen many students fumbling with {\bf why} this should occur. There is some loose talk of ``water rushing backward, the bob getting out of the way,'' and
yet this is clearly an unsystematic treatment.
Suppose the bottle is accelerated along the $x$-axis.
A pressure field is immediately set up in the fluid (on the timescale it would take a shockwave to traverse the bottle) with a large pressure at the rear
of the bottle, and a smaller pressure at the front.
This component of the gradient of the pressure field has to match the hydrostatic equilibrium condition, Eq.~\ref{hydro}, with an acceleration $\vec{a}$:
\begin{equation}
P = P_o + (\vec{g}-\vec{a}) \cdot \vec{r}.
\end{equation}
Then, an application of Eq.~\ref{at} yields
\begin{equation}
\vec{F}_{fluid} = (\vec{a}-\vec{g}) M_{fluid}.
\end{equation}
Thus, for a quiescent bottle, there is a net pressure force in the direction opposite ``$\vec{g}$'', but given the equivalence principle, we should be able to replace
$\vec{g}$ with an accelerating noninertial frame with acceleration equal to $-\vec{g}$ with no observable effects.
And, if we moved the whole discussion off into the Internationa Space Station, we would have $\vec{g}=0$, and to get the bob to point away form the bottom of the bottle, we would have to accelerate the bottle in the $\hat{z}$ direction with the magnitude of the terrestrial $g$ to get the same behavior.
A suitable probe for the reasonableness of this whole discussion is the following ``clicker"-style question.
{\it A helium-filled balloon is suspended from a thread one meter from the floor of an enclosed elevator. At the instant you cut the thread, the cable supporting the
elevator is severed. While you are hurtling toward certain severe injury, you observe the balloon
\begin{itemize}
\item{A)} Rise toward the ceiling of the elevator.
\item{B)} Remain one meter from the floor of the elevator.
\item{C)} Fall toward the floor of the elevator.
\end{itemize}
}
The correct answer is B), of course, for a very physical reason that is opaque in the usual reasoning of the principle. In the freely falling elevator, the pressure field is uniform, and hence the net force from the air on the helium balloon is zero. The balloon thus has only its weight force acting on it, and it freely falls along with the elevator.
An unsophisticated reading of the problem will assume that the buoyant force is {\bf always} acting, so the balloon will not only rise upward, it will rise upward faster and faster until it hits the roof of the elevator. A suitable demonstration (in a very special case) is found in dropping a water-filled container with a neutrally buoyant object floating in its midsection.
If the buoyant force always acts, the ball will remain in equilibrium while the container falls,
and he top of the container will fall toward the ball.
Yet, in the freely falling system, the water, ball, and container all should share the same acceleration.
And further, we have a solid explanation in terms of real forces in an inertial frame for the functioning of a centrifuge. For, as the centrifuge spins, there is an
acceleration field of the form $ - \omega^2 \vec{r}_{\perp}$.
Thus, the equivalent of the hydrostatic condition is (in cylindrical coordinates)
\begin{equation}
\partial_r P = -\omega^2 r \rho_l
\end{equation}
with the analog of Eq.~\ref{at} is then
\begin{equation}
F_{fluid} = omega^2 \rho_l \iiint dv = \omega^2 \rho_l M_l R
\end{equation}
where $R$ radial position of the geometric center of the displaced volume. This is exactly the position of the center of mass of the fluid that would exists had the
object not been submerged in the fluid.
The dependence on $R$ on $F_{fluid}$ provides an enhancement of the density-discriminating effect, crucial for the operation of a gas-centrifuge.
But again, this is solely a function of adding up the pressure on each surface of an object, and determining the net force.
From there, we can connect to a student's early training in free body diagrams (and the Momentum Principle, in the language of Ref.~\cite{mandi}.
What is going on here is extremely clear from a computational point of view. As in Figure~\ref{at_schematic}, we attempt to calculate the effect of the pressure forces
on the finite elements in the entire body under consideration. As the pressures are all equal in the $x$ and $y$ directions, it is just the $z$ direction
that I will concentrate upon here.
Each element of volume has two faces, and the pressures at neighboring surfaces give canceling forces. It is just the pressure at the top, and the pressure
at the bottom of the overall column that is important, and this gradient is just adding up the increments in the weights of the fluid elements.
What we have here is little more than
\begin{equation}
\Delta P = dx dy (P_{top} - P_{n-1} + P_{n-1} - P_{n-2} + \cdots P_1 - P_{bot})
\end{equation}
with
\begin{equation}
(P_{i} - P_{i-1}) dx dy = (dx dy dz) \rho g.
\end{equation}
A student asked to {\bf compute the effect of the pressure gradient} will find Eq~\ref{at} as a trivial restatement of a massive internal cancellation
of terms.
Exactly this sort of exercise is well within the capability of students completing a computation-enriched curriculum.\cite{mandi,taylor,goldstein}
\section{Conclusion}
Here, I have related an episode revealing a ``hole'' in the preparation of undergraduate students in physics at CSU Long Beach
(and perhaps in many more institutions beside), and proposed a solution.
The solution brings the mathematics of vector analysis back into the physics curriculum between the hiatus in electrodynamics courses
(just when it would be very useful {\it e.g.} in an introduction to quantum mechanics).
And, the argument provides a clear insight into how deeply Archimedes himself understood the theorems of vector calculus.\cite{joke}
|
\section{Introduction}
We are interested in efficient algorithms for computing a small percentage
of eigenpairs of a large Hermitian matrix $A$ that is
either sparse or structured (i.e., the matrix--vector product $Ax$
can be computed efficiently.) Often these eigenpairs correspond to the algebraically smallest eigenvalues.
This type of problem arises, for example,
in the context of Kohn-Sham density functional theory (DFT) based
electronic structure calculations of large molecules or
solids.
When the dimension of the matrix $n$ is above $10^6$, for example,
even half a percent of $n$ amounts to more than 5,000 eigenvalues.
When that many eigenpairs are needed, many of the existing algorithms
such as the Lanczos algorithm~\cite{Saad-book3}, the block Davidson algorithm~\cite{Davidson:75, Saad-book3},
which is widely used in the electronic structure calculation community,
and the Locally Optimal Block Preconditioned Conjugate Gradient (LOBPCG)
algorithm~\cite{Knyazev:01} are often not adequate or efficient.
This is true even when vast computational resources are available.
One of the main obstacles to achieving high performance
in the existing algorithms is the cost associated with solving
a projected eigenvalue problem whose dimension is at least as
large as the number of eigenvalues to be computed. The solution
of this projected eigenvalue problem is part of the so-called
Rayleigh--Ritz (RR) procedure used to extract approximations to
the eigenpairs from a carefully constructed subspace.
When the number of desired
eigenpairs is large, the cost for performing this step,
which scales cubically with respect to the dimension of the projected
matrix, cannot be ignored. To speed up the computation, one may use
the ScaLAPACK library~\cite{scalapack} to perform the dense eigenvalue
calculation in parallel on a distributed memory parallel computer.
However, for this type of calculation, it is generally difficult
to achieve good scalability beyond a few hundred processors.
Attempts have been made in recent work to address the issue of
high RR cost in large-scale eigenvalue computations.
One approach is based on the idea of spectrum slicing~\cite{Schofield.Chelikowsky.Saad:12} in which the spectrum of $A$ is divided into multiple small
intervals, and eigenvalues belonging to different intervals are computed
simultaneously. This algorithm is potentially scalable and does not
suffer from high RR cost. However, dividing the spectrum in an optimal
way is nontrivial. Furthermore, computing interior clustered eigenvalues
can be difficult. Another approach is based on solving the eigenvalue
problem as a penalized trace minimization~\cite{Wen.Yang.Liu.Zhang:13TR}.
By moving the orthonormality constraint to the objective function as
a penalty term, this scheme can use unconstrained optimization
techniques without performing frequent RR calculations.
However, the efficiency of the method depends on an optimal choice of
the penalty parameter, which may not be easy to obtain. The significance of reducing the
intensity of RR calculations was pointed out in earlier works as well, e.g., by Stewart and Jennings~\cite{Stewart.Jennings:81}.
In this paper, we present an algorithm that reduces the number of
the RR calculations.
Our approach is similar to the Davidson-Liu and LOBPCG methods in the sense that
a preconditioned
short-term recurrence is used to update the approximation to the
desired
invariant subspace.
A key difference in the proposed scheme is
that the coefficients of the short-term recurrence are obtained by
solving $k/q$ independent $3q \times 3q$ eigenvalue problems instead of one large
$2k \times 2k$ or $3k \times 3k$ eigenvalue problem, where $k$ is the number of desired eigenpairs and $q$ is a chosen block size independent of $k$.
Instead of large RR computations at every iteration, periodic basis orthogonalization is performed in the new algorithm.
The computational kernels used in this orthogonalization step typically run more
efficiently than dense diagonalization on high performance
parallel computers.
The idea of replacing the solution of a large projected eigenproblem by a sequence of smaller problems
has been considered by Knyazev in the context of the LOBPCG II algorithm~\cite{Knyazev:01}, as a means to reduce the dimension of the LOBPCG trial
subspace from $3k$ to $k$. While the approach significantly reduces the RR cost compared to the original version of LOBPCG, its application within LOBPCG II
does not eliminate the solution of a possibly large dense eigenproblem at every iteration. Specifically, instead of solving a $3k$-by-$3k$ eigenproblem as in the
original LOBPCG method, LOBPCG II solves a $k$-by-$k$ eigenproblem, which is still costly for large $k$.
Our approach, which we refer to as the Projected Preconditioned Conjugate Gradient (PPCG) algorithm,
can be easily implemented by making a relatively small modification
to the existing schemes implemented in many scientific software
packages, such as planewave DFT based electronic structure calculation
software packages. We will show by numerical examples that PPCG
indeed outperforms the current state-of-the-art algorithms implemented in the
widely used Quantum Espresso (QE) planewave density functional electronic structure software package~\cite{QE-2009}
In this work we only consider the the standard eigenvalue problem $A x = \lambda x$.
The generalization of the new algorithm to the case of the generalized eigenvalue problem $A x = \lambda B x$
is straightforward and can be performed without factoring the Hermitian positive definite matrix $B$.
The paper is organized as follows. In section~\ref{sec:tracemin}, we discuss a few
optimization based approaches for large-scale eigenvalue computations.
We present the basic version of the PPCG algorithm in section~\ref{sec:basic}.
The connection between our approach and other available algorithms is
discussed in section~\ref{sec:connect}.
A number of practical aspects for implementing the new method are addressed in section~\ref{sec:practical}.
Section~\ref{sec:example} contains numerical results. Conclusions are given in section~\ref{sec:concl}.
\section{Trace Minimization} \label{sec:tracemin}
The invariant subspace associated with the $k$ algebraically smallest
eigenvalues of $A$ can be computed by solving the following constrained
optimization problem
\begin{equation}
\min_{X^*X=I} \frac{1}{2}\mbox{trace}(X^* A X),
\label{eq:tracemin}
\end{equation}
where $X\in \mathbb{C}^{n \times k}$.
There exist several approaches for large-scale eigenvalue computations
that are based directly on formulation~\eqref{eq:tracemin}.
These approaches treat the eigenvalue problem
as the minimization problem, which allows applying relevant optimization techniques for computing the targeted
eigenpairs.
A number of popular algorithms for computing invariant subspaces
are based on gradient type methods
for minimizing~\eqref{eq:tracemin}.
In particular, projecting the gradient of the objective function in~\eqref{eq:tracemin}
along the tangent of the orthonormality constraint $X^*X=I$
yields the residual
\begin{equation}
R = (I-XX^\ast)AX = AX - X(X^*AX),
\label{eq:resid}
\end{equation}
which can be chosen as the search direction in an optimization
algorithm designed to solve ~\eqref{eq:tracemin}.
A preconditioner $T$ can be introduced to yield a modified
search direction $T R$.
In the simplest version of the Davidson-Liu algorithm, a new approximation
$\bar{X}$ is constructed by taking it to be a linear combination
of $X$ and $TR$, i.e., we write
\[
\bar{X} = X C_1 + TR C_2,
\]
where $C_1, C_2 \in \mathbb{C}^{k \times k}$ are chosen to minimize
the trace of $A$ within the subspace spanned by columns of $X$ and $TR$.
The optimal $C_1$ and $C_2$ can be obtained by computing the
lowest $k$ eigenpairs of the projected $2k \times 2k$ eigenvalue problem
\begin{equation}\label{eq:projev}
(S^* A S)C = (S^* S) C\Omega, \ \ C^* (S^*S) C = I,
\end{equation}
where $S = [ X, \ TR ]$, $C \in \mathbb{C}^{2k\times k}$, and
$\Omega \in \mathbb{R}^{k \times k}$ is a diagonal matrix that
contains the $k$ algebraically smallest eigenvalues. We can then
take $C_1$ to be the first $k$ rows of $C$ and $C_2$ to contain
the remaining rows of $C$.
The main steps of the simplest version of the Davidson-Liu algorithm
are outlined in Algorithm~\ref{alg:davidson}.
\begin{algorithm}[htbp]
\begin{small}
\begin{center}
\begin{minipage}{5in}
\begin{tabular}{p{0.5in}p{4.5in}}
{\bf Input}: & \begin{minipage}[t]{4.0in}
The matrix $A$, a preconditioner $T$ and
the starting guess of the invariant subspace
$X^{(0)} \in \mathbb{C}^{n \times k}$ associated
with the $k$ smallest eigenvalues of $A$, $X^{(0)*}X^{(0)} = I$;
\end{minipage} \\
{\bf Output}: & \begin{minipage}[t]{4.0in}
An approximate invariant subspace $\mathbb{C}^{n \times k}$
associated with $k$ smallest eigenvalues of $A$;
\end{minipage}
\end{tabular}
\begin{algorithmic}[1]
\STATE $X \leftarrow X^{(0)}$;
\WHILE {convergence not reached}
\STATE $R \leftarrow T(AX - X(X^* A X))$;
\STATE $S \gets [X, \ R]$;
\STATE Find eigenvectors $C$ associated with the $k$ smallest eigenvalues $\Omega$ of~\eqref{eq:projev};
\STATE $X \leftarrow S C$;
\ENDWHILE
\end{algorithmic}
\end{minipage}
\end{center}
\end{small}
\caption{Simplest version of Davidson-Liu algorithm}
\label{alg:davidson}
\end{algorithm}
Algorithm~\ref{alg:davidson} can be modified to accumulatively include multiple $TR$
blocks computed from different iterations in $S$, which gives the conventional Davidson method~\cite{Davidson:75}.
Alternatively, the algorithm's iterations can be altered to include the so-called ``conjugate'' direction $P$
as part of the search space $S$,
i.e., one can let
$S \leftarrow [X, TR, P]$ and solve a $3k \times 3k$
projected eigenvalue problem~\eqref{eq:projev}.
The block $P$ can be constructed as a linear combination of
$TR$ and $P$ computed at the previous iteration.
Such a modification leads to the locally optimal block preconditioned
conjugate gradient (LOBPCG) algorithm originally proposed in~\cite{Knyazev:01}.
When a good preconditioner $T$ is available, as is the case
for planewave based electronic structure calculations,
both the simplest version of the Davidson-Liu algorithm and
the LOBPCG algorithm can converge rapidly. The number of iterations
required by LOBPCG to reach convergence is often smaller than that taken
by the Davidson-Liu algorithm, but each Davidson iteration is
slightly cheaper because it solves a $2k \times 2k$ instead
of a $3k \times 3k$ projected eigenvalue problem.
When $k$ is relatively small, such extra cost per iteration is
negligible. However, when $k$ is relatively large (e.g.,
on the order of thousands or more) the cost of solving the
projected eigenvalue problem, which we refer to as the
RR cost, can no longer be ignored.
Although the RR eigenvalue problem can be solved in
parallel using the ScaLAPACK library, the performance
of this part of the calculation generally does not
scale well beyond a few hundred cores. Although some
progress has recently been made on speeding up symmetric dense
eigenvalue calculation on distributed memory parallel
computers \cite{ELPA, ELEMENTAL}, the performance of the latest algorithms still
lags behind that of level three BLAS and other computational
building blocks of electronic structure codes.
Another optimization based approach for eigenvalue computations
was proposed by Sameh and Wisniewski~\cite{sw1982,st2000}. Their TRACEMIN algorithm is different from the
gradient type schemes applied to~\eqref{eq:tracemin}.
It relies on the trace minimization procedure, which solves
a sequence of correction problems of the form
\begin{equation}
\min_{X^\ast \Delta = 0} \mbox{trace}(X-\Delta)^\ast A (X-\Delta).
\label{eq:trmin_corr}
\end{equation}
The solution of~\eqref{eq:trmin_corr} is obtained
by iteratively solving the projected linear system
\[
(M A M) \Delta = MAX, \ \ X^{\ast} \Delta = 0,
\]
where $M = I-XX^{\ast}$. A RR procedure is then performed
within the subspace spanned by the columns of $X-\Delta$ in each step to
produce a new approximation to the solution of~\eqref{eq:tracemin}.
\section{The Projected Preconditioned Conjugate Gradient algorithm}
\label{sec:basic}
In this section, we present a preconditioned conjugate gradient type of
scheme to find a solution of the minimization problem~\eqref{eq:tracemin}.
The proposed approach is motivated by the gradient projection techniques for constrained optimization (e.g.,~\cite{Haftka.Gurdal:92, Nocedal.Wright:99, Polyak_book_eng:87}).
Given a function $f(x)$ whose minimum is sought over a set $Q$ defined by constraints,
the general framework of gradient projection methods is to iteratively perform a sequence of updates
$\bar x \gets x + \gamma s$, where the updated approximation $\bar x$ is allowed to leave the set $Q$ that represents
feasible regions.
The new approximation, however, is then projected back to the feasible set
$Q$, i.e., the new iterate $x$ is defined
as $x \gets M_Q \bar x$, where $M_Q$ is an appropriately
defined projector onto $Q$.
The search direction $s$ can be defined in a number of ways.
For example,
it can be chosen as the gradient $\nabla f(x)$ of $f$ evaluated at the current approximation
$x$~\cite{Levitin.Polyak:66, Goldstein:64}. In equality
constrained optimization, the search direction is often taken to be the
projection of the gradient onto the tangent of constraints, i.e.,
$s = M \nabla f(x)$, where $M$ is a corresponding projection
defined in terms of the normal of the equality constraint
that implicitly defines the region $Q$ in which $x$ must lie~\cite{Rosen1:60, Rosen2:61}.
In particular, for the equality constraint $X^*X = I$ in~\eqref{eq:tracemin},
$M = I-XX^{\ast}$.
We consider an extension of the gradient projection approach to
trace minimization~\eqref{eq:tracemin} in which the approximation
to the
minimizer is updated as follows:
\begin{equation}\label{eq:ppcg}
\bar X \gets X C_X + W C_W + P C_P, \quad X \gets M_Q \bar X, \quad \end{equation}
where the search direction $W = (I - X X^*) T R$ is given by the preconditioned residual $T R = T(AX - X (X^* A X))$
projected onto the tangent of the orthonormality constraint and
$P = (I - X X^*) (W^{\prime} C_W^{\prime} - P^{\prime} C_P^{\prime})$
represents a conjugate direction in the same tangent space. The ``prime notation''
refers to the corresponding quantities from the previous step.
Extracting the best approximation
from the subspace spanned by the columns of $X$, $W$ and $P$ (as traditionally done) would require
a RR calculation that is costly when the number of desired
eigenpairs is large. To reduce such cost, we relax the optimality requirement on the search
parameters $C_X$, $C_W$, and $C_P$ in~\eqref{eq:ppcg}, and allow them to introduce non-orthogonality
in the updated columns of $\bar X$. This places $\bar X$ outside of the feasible region given by the orthogonality
constraint, which is remedied by the subsequent application of a projector $M_Q$.
Specifically,
let us restrict $C_X$, $C_W$, and $C_P$ to be diagonal matrices.
An advantage of such a restriction is that the iteration parameters associated
with each column of the updated $X$ can be determined independently.
More generally, it is possible to allow $C_X$,
$C_W$ and $C_P$ to be block diagonal matrices
with small diagonal blocks (e.g., $5\times 5$ or $10 \times 10$ blocks). One can then expect that, if properly chosen, the extra degrees of
freedom introduced by these diagonal blocks can reduce the iteration count.
The general block diagonal formulation will be discussed in section~\ref{subsec:block}.
Let $C_X = \text{diag} \{ \alpha_1, \ldots,\alpha_k \}$, $C_W = \text{diag}\{ \beta_1, \ldots,\beta_k \}$,
and $C_P = \text{diag}\{ \gamma_1, \ldots,\gamma_k \}$.
Then iteration~\eqref{eq:ppcg} gives a sequence of $k$ single-vector
updates
\begin{equation}\label{eq:upd_col}
\bar x_j \gets \alpha_j x_j + \beta_j w_j + \gamma_j p_j, \quad j = 1, \ldots,k;
\end{equation}
where $\bar x_j$, $x_j$, $w_j$, and $p_j$ denote the $j$th columns of $\bar X$, $X$, $W$, and $P$, respectively.
Let us choose $\alpha_j$, $\beta_j$, and $\gamma_j$ in such a way that
each corresponding updated column $\bar x_j$ yields the minimizer of $x^* A x$, subject to
the normalization constraint $\|x\| = 1$, over the corresponding subspace
spanned by $x_j$, $w_j$, and $p_j$.
Clearly, the computations of the parameter triplets are independent of each other, and can be
performed by solving $k$ separate $3$-by-$3$ eigenvalue problems.
As a result of the decoupled steps~\eqref{eq:upd_col}, the columns $\bar x_j$ are generally not orthogonal to each other.
Moreover,
they can all converge to the same eigenvector associated with the smallest
eigenvalue of~$A$ without any safeguard in the algorithm. To overcome this issue, we project the updated block $\bar X$
back onto the orthonormality constraint $X^* X = I$ by
performing a QR factorization of $X$ and setting the new approximation $X$ to the obtained orthogonal factor.
This step corresponds to the action of applying the projector $M_Q$ in~\eqref{eq:ppcg}, which we
further denote by $X \leftarrow \texttt{orth}(\bar X)$.
\begin{algorithm}[htbp]
\begin{small}
\begin{center}
\begin{minipage}{5in}
\begin{tabular}{p{0.5in}p{4.5in}}
{\bf Input}: & \begin{minipage}[t]{4.0in}
The matrix $A$, a preconditioner $T$, and a starting guess of the invariant subspace $X^{(0)} \in \mathbb{C}^{n \times k}$
associated with the $k$ smallest eigenvalues of $A$;
\end{minipage} \\
{\bf Output}: & \begin{minipage}[t]{4.0in}
An approximate invariant subspace $X \in \mathbb{C}^{n \times k}$
associated with the $k$ smallest eigenvalues of $A$;
\end{minipage}
\end{tabular}
\begin{algorithmic}[1]
\STATE $X \gets \texttt{orth} (X^{(0)})$; $P \gets \left[ \ \right]$;
\WHILE {convergence not reached}
\STATE $W \gets T (AX - X(X^* A X))$;
\STATE $W \gets (I - XX^*)W$;
\STATE $P \gets (I - XX^*)P$;
\FOR {$j = 1, \ldots, k$}
\STATE $S \gets [x_j , w_j , p_j]$;
\STATE Find the smallest eigenpair ($\theta_{\min}, c_{\min}$) of $S^*AS c = \theta S^* Sc$, where $c^* S^*S c = 1$;
\STATE $\alpha_j \gets c_{\min}(1)$, $\beta_j \gets c_{\min}(2)$; and $\gamma_j \gets c_{\min}(3)$ ($\gamma_j = 0$
at the initial step);
\STATE $p_j \gets \beta_j w_j + \gamma_j p_j$;
\STATE $x_{j} \gets \alpha_j x_j + p_j$.
\ENDFOR
\STATE $X \gets \texttt{orth}(X)$;
\STATE If needed, perform the Rayleigh-Ritz procedure within $\mbox{span}(X)$;
\ENDWHILE
\end{algorithmic}
\end{minipage}
\end{center}
\end{small}
\caption{The projected preconditioned conjugate gradient (PPCG) algorithm}
\label{alg:ppcg0}
\end{algorithm}
The proposed approach is outlined in
Algorithm~\ref{alg:ppcg0} that we refer to as the Projected
Preconditioned Conjugate Gradient (PPCG) algorithm.
Note that in practical implementations we require the method to perform
the RR procedure every once in a while (step 14).
Such periodic RR computations allow ``rotating'' the columns of $X$ closer to the targeted eigenvectors.
They also provide opportunities for us to identify converged eigenvectors and
deflate them through a locking mechanism
discussed in
section~\ref{sec:practical}.
In our experiments, we typically perform an RR calculation every 5-10 iterations,
which is significantly less frequent compared to the existing eigensolvers that
perform the RR procedure at each step.
In principle, the RR procedure in step 14 of Algorithm~\ref{alg:ppcg0} can be omitted.
In this case, the columns of the iterates $X$ generally do not converge to
eigenvectors, i.e.,
$X$ only represents some orthonormal basis of the approximate invariant subspace.
However, in a number of our test problems,
the absence of step 14 led to the convergence deterioration. Therefore, performing a periodic RR step can be helpful to
ensure the eigensolver's robustness. We will return to this discussion in section~\ref{sec:practical}.
An important element of Algorithm~\ref{alg:ppcg0} is the orthonormalization of the block $X$ in step~13. It is clear that
the $\mbox{orth}(X)$ procedure is well-defined if and only if it is applied to a full-rank matrix, i.e., if and only if the
single-vector sweep in Steps 6-12 yields a block $X$ of linearly independent vectors.
The following theorem shows that the linear independence among
columns of $X$ is guaranteed if all the parameters $\alpha_j$ are nonzero.
\begin{theorem}\label{thm:1}
Let vectors $\bar x_j$ be computed according to~\eqref{eq:upd_col}, where $x_j$, $w_j$, and $p_j$ denote the $j$th columns of $X$, $W$, and $P$, respectively; and let $X^*X = I$.
Then the matrix $\bar X = [\bar x_1, \ldots, \bar x_k]$ is full-rank if $\alpha_j \neq 0$ for all $j$.
\end{theorem}
\begin{proof}
Let us write~\eqref{eq:upd_col} in the matrix form as
\begin{equation}\label{eq:expandx}
\bar X = X C_X + K,
\end{equation}
where $K = W C_W + P C_P$; with $C_X$, $C_W$, $C_P$ being the diagonal matrices of iteration coefficients
$\alpha_j$, $\beta_j$, and $\gamma_j$, respectively.
We assume, on the contrary, that columns of $\bar{X}$ are linearly dependent.
Then there exists a vector $y\neq 0$ such that $\bar{X}y=0$. It follows from~\eqref{eq:expandx}
and the conditions $X^* X=I$ and $X^{*}K=0$ that $C_Xy=0$. Since $C_X$
is diagonal, then at least one of its diagonal elements
must be zero.
This contradicts the assumption that $\alpha_j \neq 0$ for all $j$.
\end{proof}
Theorem~\ref{thm:1} provides us with a simple indicator of rank deficiency in
$\bar{X}$. In the case when $\bar{X}$ becomes rank deficient, there is a simple
way to fix the problem. We can simply backtrack and exclude the $P$ block
in \eqref{eq:ppcg} and take a steepest descent-like step by recomputing
the coefficients in $C_X$ and $C_W$.
The next theorem shows that, in this case, if the preconditioner $T$ is Hermitian positive
definite (HPD), the recomputed $\alpha_j$'s are guaranteed to be nonzero and
therefore the updated $X$ is full rank, unless some columns of $W$ become
zero, which indicate the convergence of some eigenvectors that should be
deflated at an earlier stage.
\begin{theorem}\label{thm:2}
Let $P \equiv 0$ in \eqref{eq:ppcg}
so that the update in $X$ are computed as
\begin{equation}\label{eq:upd_col_trunc}
\bar x_j \gets \alpha_j x_j + \beta_j w_j, \quad j = 1, \ldots,k.
\end{equation}
If the preconditioner $T$ is~HPD and the $j$th column $r_j$ of the residual
$AX - X(X^*AX)$ is nonzero, then $\bar X = [\bar x_1, \ldots, \bar x_k]$
must have a full rank.
\end{theorem}
\begin{proof} It follows from Theorem 1 that if $\bar{X}$ is rank deficient,
then there is at least one $j$ such that $\alpha_j = 0$. Since
$\alpha_j$ is the first component of an eigenvector $c = (\alpha_j, \beta_j)^*$ of the 2-by-2 projected eigenproblem
\begin{equation}\label{eq:proj2}
\left(
\begin{array}{cc}
x_j^*A x_j & x_j^* A w_j \\
x_j^*A w_j & w_j^* A w_j
\end{array}
\right) c = \theta
\left(
\begin{array}{cc}
x_j^* x_j & 0 \\
0 & w_j^* w_j
\end{array}
\right) c,
\end{equation}
the coefficient $\beta_j$ must be nonzero, and the matrix on the left hand
side of \eqref{eq:proj2} must be diagonal, i.e., $x_j^* A w_j = 0$.
However, since
\[
x_j^*Aw_j = (Xe_j)^*A (I-XX^*) T \left[(I-XX^*)AX\right]e_j = r_j^* T r_j,
\]
$r_j$ must be zero since $T$ is assumed to be HPD. This contradicts the
assumption that $r_j$ is nonzero. Hence, it follows that $\alpha_j \neq 0$,
and $\bar{X}$ must be full~rank.
\end{proof}
We note, however, that the situation where $\alpha_j$ is zero is very unlikely in practice
and, in particular, has never occurred in our numerical tests.
\begin{comment}
The following theorem suggests that the linear independence among
columns of $X$ is guaranteed under some mild assumptions that
are easily satisfied in practice.
\begin{theorem}\label{thm:1}
Let vectors $\bar x_j$ be computed according to~\eqref{eq:upd_col}, where $x_j$, $w_j$, and $p_j$ denote the $j$th columns of $X$, $W$, and $P$, respectively; and let $X^*X = I$. Then the matrix $\bar X = [\bar x_1, \ldots, \bar x_k]$ is
full-rank if $r_j \neq 0$ for all $j$ and the preconditioner $T$ is nonsingular.
\end{theorem}
\begin{proof}
Let us write~\eqref{eq:upd_col} in the matrix form as
\begin{equation}
\bar X = X C_X + K, \label{eq:expandx}
\end{equation}
where $K = W C_W + P C_P$; with $C_X$, $C_W$, $C_P$ being the diagonal matrices of iteration coefficients $\alpha_j$, $\beta_j$, and $\gamma_j$ respectively.
{\color{red} If columns of $\bar{X}$ are linearly dependent, }
{\color{blue} Let us assume, on the contrary, that columns of $\bar{X}$ are linearly dependent.}
Then there exists a vector
$y\neq 0$ such that $\bar{X}y=0$. It follows from \eqref{eq:expandx}
and the conditions $X^* X=I$ and $X^{*}K=0$ that $C_Xy=0$. Since $C_X$
is diagonal, then at least one of its diagonal elements, say $\alpha_j$,
must be zero.
{\color{red}
Since $\alpha_j$ is the first component of an eigenvector of the $3\times 3$
projected matrix
\begin{equation}
\left(
\begin{array}{ccc}
x_j^* Ax_j & x_j^* A w_j & x_j^* A p_j \\
w_j^* Ax_j & w_j^* A w_j & w_j^* A p_j \\
p_j^* Ax_j & p_j^* A w_j & p_j^* A p_j \\
\end{array}
\right),
\label{eq:blocka}
\end{equation}
it is not difficult to show that \eqref{eq:blocka} is block diagonal and
$x_j^*A w_j = x_j^* A p_j = 0$.
}
{\color{blue} Not true. Since $\alpha_j$ is the first component of an eigenvector of the $3\times 3$
projected pencil $(K, M)$, where
\begin{equation}
K = \left(
\begin{array}{ccc}
x_j^* Ax_j & x_j^* A w_j & x_j^* A p_j \\
w_j^* Ax_j & w_j^* A w_j & w_j^* A p_j \\
p_j^* Ax_j & p_j^* A w_j & p_j^* A p_j \\
\end{array}
\right),
M = \left(
\begin{array}{ccc}
x_j^* x_j & 0 & 0 \\
0 & w_j^* w_j & w_j^* p_j \\
0 & p_j^* w_j & p_j^* p_j \\
\end{array}
\right),
\label{eq:blocka}
\end{equation}
it is easy to show that either $x_j^*A w_j = x_j^* A p_j = 0$ or $\beta_j (x_j^*A w_j) + \gamma_j (x_j^* A p_j) = 0$.
We can put $\beta_j (x_j^*A w_j) + \gamma_j (x_j^* A p_j) \neq 0$ to the theorem's statement as an assumption. Then the only
possibility would be $x_j^*A w_j = x_j^* A p_j = 0$ and the proof follows.
}
However, since
\[
x_j^*Aw_j = (Xe_j)^* {\color{red}(I-XX^*)}AT\left[(I-XX^*)AX\right]e_j = r_j^* T r_j,
\]
{\color{blue} This is a nice short derivation. There is a small typo. Should be :
\[
0 = x_j^*Aw_j = (Xe_j)^*A (I-XX^*) T \left[(I-XX^*)AX\right]e_j = r_j^* T r_j,
\]
}
and $T$ is nonsingular, {\color{red} $r_j$ must be zero which
contradicts our assumption}. {\color{blue} requiring that $T$ is nonsingular is not enough. $T r_j$ can be orthogonal to $r_j$ for a non-zero $r_j$, although this is unlikely in practice.
Should request that $T$ is HPD, which ensures that $r_j^* T r_j > 0$}.
Therefore, $\bar{X}$ must be full
rank.
\end{proof}
Note that when $\|r_j\|$ is zero or extremely small, it is an
indication that $X$ contains a converged eigenvector or an invariant
subspace. In this case, a deflation technique should be used to
extract and lock the converged invariant subspace. Otherwise,
the above theorem suggests that the new approximation $\bar{X}$ is
of full rank, and the $\mbox{orth}(X)$ operation in Step 13 is
well defined.
\end{comment}
\begin{comment}
We consider a Projected Gradient (PG) method, introduced by Levitin and Polyak~\cite{Levitin.Polyak:66},
and Goldstein~\cite{Goldstein:64},
for a problem of minimizing a function $f(x)$ over a set $Q$ that is typically given by a
number of constraints on $x$.
Assuming that $f$ is differentiable, in its general form, the PG iteration can be written as
\begin{equation}\label{eq:pg}
x^{(i + 1)} = P_Q ( x^{(i)} + \gamma^{(i)} \nabla f (x^{(i)}) ),
\end{equation}
where $\nabla f (x)$ denotes the gradient of $f(x)$ and $P_Q$ is a projection onto $Q$ defined by
\begin{equation}\label{eq:projQ}
P_Q (x) = \text{arg}\min_{y \in Q} \|x - y\|.
\end{equation}
The PG scheme represents an extension of the classical gradient methods (see, e.g., [?, ?]) to the case of
constrained optimization. The idea behind~\eqref{eq:pg} is, first, to ignore the constraints on $x$ and perform a
standard gradient minimization step for the unconstrained problem. This gives a point outside of $Q$,
which is then projected back onto the set using $P_Q$ in~\eqref{eq:projQ}.
In fact,~\eqref{eq:pg} defines a family of optimization methods, where each particular instance is determined by the
choice of the iteration parameter $\gamma^{(i)}$ and the structure of the set $Q$. In particular, under certain assumptions
on $\gamma^{(i)}$ and $Q$, the method~\eqref{eq:pg} is proven to converge to a local minimizer of $f(x)$;
see, e.g.,~\cite{Levitin.Polyak:66, Goldstein:64,McCormick.Tapia, Bersekas, Dunn, Calamai.More}.
We are interested in adapting the framework of the PG approach to the computation of the invariant subspace associated
with $k$ smallest eigenvalues of a Hermitian matrix $A$. For this purpose, we consider the following trace minimization problem:
\[
\min_{X \in Q} \frac{1}{2} \text{trace}( (X^*X)^{-1} X^* A X ), \quad Q = \left\{X \in \mathbb{C}^{n \times k} : \ X^* X = I \right\},
\]
where the objective function is a block generalization of the Rayleigh Quotient $(x,Ax)/(x,x)$ and $Q$ is a set of all $n$-by-$k$
matrices with orthonormal columns. A minimizer of this problem represents an orthonormal basis of the targeted invariant subspace.
Let $X^{(i)}$ be a current approximation of the invariant subspace, such that $X^{(i)*} X^{(i)} = I$. Then the gradient
of the objective function at $X^{(i)}$ is the residual $R^{(i)} = AX^{(i)} - X^{(i)}(X^{(i)*}AX^{(i)})$.
Following the PG approach, let us omit the constraints, and perform an optimization step based
on the current approximation and the gradient. The result is then projected back onto $Q$:
\begin{equation}\label{eq:pg_eig}
X^{(i + 1)} = P_Q (X^{(i)} C_X^{(i)} + R^{(i)} C_R^{(i)}).
\end{equation}
Here, $C_X^{(i)} = \text{diag} \{ \alpha^{(i)}_1, \ldots,\alpha^{(i)}_k \}$ and
$C_R^{(i)} = \text{diag}\{ \beta^{(i)}_1, \ldots,\beta^{(i)}_k \}$ are the diagonal matrices of iteration coefficients.
In contrast to the original PG formulation~\eqref{eq:pg}, we consider a generalization where more than one (2$k$) iteration parameter are
defined at each step. As we will demonstrate later, the increase in the number of properly chosen iteration parameters naturally speeds up the convergence of
the scheme.
Let $x_j^{(i)}$ and $r_j^{(i)}$ denote the $j$th columns of $X^{(i)}$ and $R^{(i)}$, respectively. Since $C_X^{(i)}$ and $C_R^{(i)}$ are assumed to be diagonal,
scheme~\eqref{eq:pg_eig} updates all columns separately. In particular, we choose iteration parameters $\alpha^{(i)}_j$ and $\beta^{(i)}_j$, such that
the updated columns $z_j = \alpha^{(i)}_j x_j^{(i)} + \beta^{(i)}_j r_j^{(i)}$ satisfy
\begin{equation}
\min_{\|\ z_j \| = 1} z_j^* A z_j, \quad j = 1, 2,...,k.
\label{eq:surmodel}
\end{equation}
Clearly, minimizations~\eqref{eq:surmodel} are performed independently of each other. Each of them represents a separate
minimizations of the Rayleigh Quotient over the subspaces spanned by a pair of the orthogonal vectors $x^{(i)}$ and $w^{(i)}$,
which is accomplished by solving a two-dimensional eigenvalue problem.
Given a block $Z$ of the updated columns $z_j$, the new approximate invariant subspace $X^{(i+1)} = P_Q(Z)$ is defined as a matrix
with orthonormal columns that is the closest, in a certain norm, to $Z$. We address this step in more detail in the next section,
where we further generalize the PG method~\eqref{eq:pg_eig}--\eqref{eq:surmodel} to a preconditioned conjugate gradient type iteration.
\section{The Projected Preconditioned Conjugate Gradient Algorithm}
Given a preconditioner $T$, let us consider the following extension of the PG method~\eqref{eq:pg_eig}--\eqref{eq:surmodel}:
\begin{equation}\label{eq:ppcg}
X^{(i + 1)} = P_Q (X^{(i)} C_X^{(i)} + W^{(i)} C_W^{(i)} + P^{(i)} C_P^{(i)}),
\end{equation}
where $W^{(i)} = (I - X^{(i)*}X^{(i)}) T R^{(i)}$ and $P^{(i)} = (I - X^{(i)*}X^{(i)}) X^{(i-1)}$. Similar to~\eqref{eq:pg_eig},
we assume that the iteration matrices are diagonal, i.e., $C_X^{(i)} = \text{diag} \{ \alpha^{(i)}_1, \ldots,\alpha^{(i)}_k \}$,
$C_W^{(i)} = \text{diag}\{ \beta^{(i)}_1, \ldots,\beta^{(i)}_k \}$, and $C_P^{(i)} = \text{diag}\{ \gamma^{(i)}_1, \ldots,\gamma^{(i)}_k \}$.
Scheme~\eqref{eq:ppcg} represents a projected variant of the three-term recurrent form of the preconditioned conjugate gradient method [Knyazev, Polyak-book].
Here we require that the preconditioned gradient is projected onto the orthogonal complement of $X^{(i)}$, so that the blocks $X^{(i)}$ and $W^{(i)}$ are orthogonal,
by analogy with $X^{(i)}$ and $R^{(i)}$ in the unpreconditioned case of the PG method~\eqref{eq:pg}. This orthogonalization turns out to play a crucial role in enabling
regular convergence of the method. It will also allow us to justify feasibility of certain steps of the resulting algorithm discussed below.
Same considerations apply to the projection in the definition of the ``conjugate direction'' $P^{(i)}$.
Similar to~\eqref{eq:pg_eig}, let us define the iteration parameters $\alpha^{(i)}$, $\beta^{(i)}$, and $\gamma^{(i)}$ in~\eqref{eq:ppcg} in such a way that
the updated columns
\begin{equation}\label{eq:upd_col}
z_j = \alpha^{(i)}_j x_j^{(i)} + \beta^{(i)}_j w_j^{(i)} + \beta^{(i)}_j p_j^{(i)}, \quad j = 1, \ldots,k;
\end{equation}
satisfy~\eqref{eq:surmodel}.
In this case, the Rayleigh Quotient is minimized separately, $k$ times, on the subspaces spanned by the columns $x^{(i)}_j$, $w^{(i)}_j$,
and $p^{(i)}_j$ of the matrices $X^{(i)}$, $W^{(i)}$, and $P^{(i)}$, respectively.
As a result, one has to solve $k$ $3 \times 3$ eigenvalue problems. This is in contrast, e.g., to the $3k \times 3k$ eigenvalue problems that arise in the LOBPCG algorithm,
or the $2k \times 2k$ subspace diagonlizations in the Davidson-Liu approach.
Clearly, the updated columns of $z_j$ in~\eqref{eq:upd_col} are generally not orthogonal to each other.
Moreover, there is a potential danger that they can be linear dependent, because, according to~\eqref{eq:surmodel}, each $z_j$
tends to converge to the eigenvector associated with the smallest eigenvalue of $A$.
The following proposition shows that this does not happen if the parameters $\alpha^{(i)}_j$ in~\eqref{eq:upd_col} are nonzero.
\begin{proposition}
If $\alpha_j^{(i)} \neq 0$, then all the columns $z_j$ in~\eqref{eq:upd_col} are linearly independent.
\end{proposition}
\textbf{Proof.}
Let $Z = X^{(i)} C_X^{(i)} + (I - X^{(i)*}X^{(i)}) W^{(i)} + (I - X^{(i)*}X^{(i)}) P^{(i)}$ be the
matrix of the updated columns $z_j$ in~\eqref{eq:upd_col}.
Then, due to the presence of the projectors $I - X^{(i)*}X^{(i)}$ in the definitions of $W^{(i)}$ and $P^{(i)}$,
$Z^* X^{(i)} = C_X^{(i)}$, which is a diagonal matrix with nonzero diagonal elements $\alpha^{(i)}_j$.
Thus the columns of $Z$ and $X^{(i)}$ represent two bi-orthogonal sets of vectors.
This implies that the columns of $Z$ are linearly independent.
To see the above, let us assume that the columns of $Z$ are linearly dependent.
From the bi-orthogonality of $Z$ and $X^{(i)}$, we have $z_j^* x_l^{(i)} = 0$ for any $j \neq l$ and $z_j^* x^{(i)}_j = \alpha_j^{(i)} \neq 0$ for all $j$.
Since, by our assumption, columns of $Z$ are linearly dependent,
there exists a column $z_s$ that can be expressed as a linear combination of other columns of $Z$, i.e., $z_s = \sum_{ j \neq s } c_j z_j$.
But then $z_s^* x_s^{(i)} = \sum_{ j \neq s } c_j z_j^* x^{(i)}_s = 0$, which contradicts our assumption that
all $\alpha_j^{(i)} = z_j^* x_j^{(i)} \neq 0$. Therefore, the columns of $Z$ must be linearly independent. \textbf{End Proof}
Let us now address the question of defining the projector $P_Q$ in~\eqref{eq:projQ}.
Given a matrix $Z$, its closest, with respect to a unitary norm, matrix with orthonormal columns is
an orthonormal factor $U$ in the polar decomposition\footnote{Let $Z$ be an $n$-by-$k$ matrix, then $Z = U H$, where $U$ is an $n$-by-$k$
matrix with orthonormal columns and $H \equiv (A^* A)^{1/2}$ is positive semidefinite, is called a polar decomposition of $Z$; see, e.g.,~\cite{Horn}.
} of $Z$~\cite{?}.
In particular, if $Z$ is a block of the intermediate updates~\eqref{eq:upd_col} and if columns of $Z$ are linearly independent,
then the new approximation should be defined as $X^{(i+1)} \equiv P_Q(Z) = U$, where $U = Z (Z^* Z)^{-1/2}$.
Clearly, $Z$ and the optimal approximation $U$ represent exactly the same subspace. Therefore, the projector $P_Q$ can be viewed as a procedure
for selecting an appropriate basis of $Z$. However, the construction of the polar factor $U$ requires diagonalizing the $k$-by-$k$ matrix $Z^* Z$, which is
undesirable in the context where $k$ is large, and is the step that we want to avoid in the final algorithm. Therefore, we propose to replace the polar
decomposition of $Z$ by the QR factorization~\cite{?}, and set $X^{(i+1)}$ to be the corresponding orthonormal factor. In other words, $X^{(i+1)}$ is obtained simply by
orthonormalizing the updated columns $z_j$ in~\eqref{eq:upd_col}.
The QR decomposition represents a significantly faster computation compared to the diagonalization procedure, and can scale to a larger number of processors.
As we demonstrate in our numerical experiments, sacrificing the optimality in the definition~\eqref{eq:projQ} does not affect the convergence.
Apparently, this can be attributed to the fact that $Z$ turns out to be nearly orthogonal after each sweep of the column updates~\eqref{eq:upd_col}.
In fact, we have considered pushing the projection requirement even further and tried to entirely remove the orthogonalization at a number of steps.
We discuss this issue in more detail in Section~\ref{sec:practical}.
\end{comment}
\begin{comment}
The algorithm we propose below attempts to reduce the
number of RR eigenvalue calculations. Our basic idea
is to construct a surrograte model by relaxing the
$k(k+1)/2$ orthonormality constraints $X^TX = I$
in \eqref{tracemin}, and replacing them
with $k$ normalization constraints
$\|Xe_j\| = 1$, $j=1,2,...,k$, where $e_j$ is the $j$th column
of the identity matrix.
Such relaxation allows us performs $k$ separate
Rayleigh quotient minimizations using a
single-vector version of the LOBPCG updating
scheme to reduce the trace of $X^{\ast} AX$.
Because each columns of $X$ is updated
independently, we only need to solve $k$ $3 \times 3$ eigenvalue
problems instead of a $3k \times 3k$ eigenvalue problem.
Clearly, the updated columns of $X$ are generally not
orthogonal to each other when the orthonormality constraint $X^TX = I$
is relaxed. Moreover, without other additional
constraints, they may all converge to the eigenvector associated
with the smallest eigenvalue of $A$.
We use three strategies to overcome this potential problem
and ensure the new iterative procedure converge to
the solution of \eqref{eq:tracemin}.
\begin{enumerate}
\item When minimizing the surrogate model
\begin{equation}
\min_{\|x_j^{(i)}\| = 1} \langle x_j^{(i)}, A x_j^{(i)} \rangle, \ \ \mbox{for} \ \
j = 1, 2,...,k,
\label{eq:surmodel}
\end{equation}
at the $i$th iteration, we choose the search direction
to be tangent to the orthonormality constraint
$\langle X^{(i)}, X^{(i)} \rangle = I$.
If each column of $X$ is updated by a preconditioned
conjugate gradient type of scheme:
\[
x_j^{(i+1)} \leftarrow \alpha x_j^{(i)} + \beta K r_j^{(i)} + \gamma p_j^{(i-1)},
\]
where $r_j^{(i)} = Ax_j^{(i)} - \theta_j x_j^{(i)}$, $K$ is a preconditioner,
$\theta_j = \langle x_j^{(i)}, A x_j^{(i)} \rangle$,
and $p_j^{(i-1)}$ is search direction computed in the previous step,
we perform the following orthogonalization step
\[
r_j^{(i)} \leftarrow r_j^{(i)} - X^{(i)} \langle X^{(i)}, r_j^{(i)} \rangle \ \ \mbox{and} \ \
p_j^{(i-1)} \leftarrow p_j^{(i-1)} - X^{(i)} \langle X^{(i)}, p_j^{(i-1)} \rangle
\]
to make sure
$\langle X^{(i)}, r_j^{(i)} \rangle = 0$ and $\langle X^{(i)}, p_j^{(i-1)} \rangle = 0$
hold for all $j= 1, 2,...,k$.
\item After we update each column of $X^{(i)}$ by reducing
$\langle x_j^{(i)}, A x_j^{(i)} \rangle$ under the normalization
constraint $\| x_j^{(i)} \| = 1$, we project the updated $X^{(i)}$
back onto the orthonormality constraint $X^TX = I$ by
performing a QR factorization denoted by
\[
X^{(i+1)} \leftarrow \mbox{QR}(X^{(i)}).
\]
\item We periodically (every 10-20 iterations) perform the
Rayleigh-Ritz procedure with the subspace spanned by
columns of $X^{(i)}$ to reposition the starting point
for each Rayleigh-Quotient minimization.
When each column of $X$ is sufficiently close to an eigenvector
associated with a distinct eigenvalue of $A$, minimizing
$k$ Rayleigh quotients separately is likely to be just as effective
as minimizing the trace of $X^TAX$.
\end{enumerate}
\end{comment}
\section{Relation to other algorithms}\label{sec:connect}
The PPCG algorithm is similar to a variant of the LOBPCG
algorithm called LOBPCG II presented in~\cite{Knyazev:01},
in the sense that it
updates
the subspace via $k$ independent $3 \times 3$ RR computations rather than one large $3k \times 3k$ one, as in LOBPCG.
As discussed in the next section, the PPCG algorithm generalizes this to $3q \times 3q$ RR computations, where $q$ is a chosen block size, to increase convergence rate and better exploit available computational kernels.
The main difference, however, is that in LOBPCG II,
the RR procedure is performed within the subspace
spanned by the columns of $X$ in each iteration,
whereas in PPCG, $X$ is merely orthogonalized,
and RR computations are invoked only periodically
(e.g., every 5 or 10 iterations). Another difference is
related to the construction of the blocks $W$ and $P$. In contrast
to the proposed PPCG algorithm, LOBPCG II does not carry out the orthogonalizations
of the preconditioned residuals and conjugate directions against the
approximate invariant subspace $X$.
Furthermore, to allow the replacement of RR with periodic orthogonalization, the definition of the PPCG residuals has been generalized to the case in which $X$ is not necessarily formed by a basis of Ritz vectors and
the matrix $X^*AX$ is not necessarily diagonal.
As a result, the separate minimizations in
PPCG and LOBPCG II are performed with respect to different
subspaces, and the methods
are not equivalent even if the PPCG block size is $1$.
The TRACEMIN algorithm~\cite{sw1982} also becomes similar to the PPCG algorithm
if the RR procedure is performed periodically. Instead of minimizing
several Rayleigh quotients, TRACEMIN solves several
linear equations using a standard preconditioned conjugate gradient (PCG)
algorithm. Typically, more than one PCG iteration is needed to
obtain an approximate solution to each equation.
The PPCG method can also be viewed as a compromise between
a full block minimization method such as the LOBPCG method,
which converges rapidly but has a higher RR cost per iteration,
and a single vector method combined with an appropriate
deflation scheme (also known as a band-by-band method), which
has a negligible RR cost but slower overall convergence rate because
one eigenpair is computed at a time. Also, band-by-band methods
cannot effectively exploit the concurrency available in multiplying
$A$ with a block of vectors, and hence are often slower in practice
on high performance parallel computers.
\section{Practical aspects of the PPCG algorithm}\label{sec:practical}
In this section, we address several practical aspects of the PPCG algorithm
that are crucial for achieving high performance. We first consider the
generalization of the single-vector updates in~\eqref{eq:upd_col}
to block updates.
\subsection{Block formulation}\label{subsec:block}
As indicated earlier, we can allow $C_X$, $C_W$ and
$C_P$ to be block diagonal.
In this case,
$C_X = \mbox{diag}\{ C_{X_1}, \ldots, C_{X_s}\}$,
$C_W = \mbox{diag}\{ C_{W_1}, \ldots, C_{W_s}\}$,
and $C_P = \mbox{diag}\{ C_{P_1}, \ldots, C_{P_s}\}$,
and $X$, $W$ and $P$ can be partitioned conformally as
$X = [X_1, X_2, \ldots , X_s]$,
$W = [W_1, W_2, \ldots , W_s]$, and
$P = [P_1, P_2, \ldots , P_s]$,
where the $j$th subblocks of $X$, $W$, and $P$ contain $k_j$ columns, and
$\sum_{j=1}^s k_j = k $.
The single-column sweep~\eqref{eq:upd_col} is then replaced by block updates
\begin{equation}\label{eq:upd_block}
\bar X_j \gets X_j C_{X_j} + W_j C_{W_j} + P_j C_{P_j}, \quad j = 1,\ldots,s.
\end{equation}
After columns of $\bar X = [\bar X_1, \bar X_2, \ldots , \bar X_s]$
are orthonormalized, we obtain a new approximation which is used as a
starting point for the next PPCG iteration. For each $j$, the block
coefficients $C_{X_j}$, $C_{W_j}$, and $C_{P_j}$
in~\eqref{eq:upd_block} are chosen to minimize the trace~\eqref{eq:tracemin}
within span$\{X_j, W_j, P_j\}$. This is equivalent to
computing the $k_j$ smallest eigenvalues and corresponding eigenvectors
of $3 k_j$-by-$3 k_j$ eigenvalue problems~\eqref{eq:projev} with
$S = [X_j, W_j, P_j]$. Thus, the block formulation of the PPCG algorithm performs
$s$ iterations of the ``for'' loop in lines 6-12 of Algorithm~\ref{alg:ppcg0}.
Note that, in the extreme case where $s = 1$, i.e., the splitting corresponds to the whole block, the PPCG algorithm becomes equivalent to
LOBPCG~\cite{Knyazev:01}. In this case the matrices $C_X$, $C_W$,
and $C_P$ are full and generally dense, and the updated
solution is optimal within the subspace spanned by all columns of $X$, $W$, and $P$.
As we will demonstrate in section~\ref{sec:example}, making
$C_X$, $C_W$ and $C_P$ block diagonal
generally leads to a reduction in the number of outer iterations required
to reach convergence. However, as the block size increases, the
cost associated with solving $s$ $3k_j \times 3k_j$ eigenvalue
problems also increases. The optimal choice of $k_j$ will be problem- and computational platform
dependent. Heuristics must be developed to set $k_j$ to an appropriate
value in an efficient implementation.
In our current implementation, we set $k_j$ to a constant $sbsize$ with
the exception that the last block of $X$, $W$, and $P$ may contain a
slightly different number of columns.
In principle, one can choose different $k_j$ values for each subblock. For example, this could be helpful if additional information about the distribution of $A$'s spectrum is available.
In this case, a proper uneven splitting could potentially allow for a better resolution of eigenvalue clusters.
\subsection{Convergence criteria}
We now discuss appropriate convergence criteria for terminating
the PPCG algorithm. If, instead of individual eigenpairs, we are only interested in the invariant
subspace associated with the smallest eigenvalues of $A$,
we may use the following relative subspace
residual norm
\begin{equation}
\frac{\| AX - X(X^\ast AX) \|_F} {\| X^\ast A X\|_F},
\label{eq:resnrm}
\end{equation}
as a metric to determine when to terminate the PPCG iteration, where $\|\cdot\|_F$ is the Frobenius
norm.
No additional multiplication of $A$ with $X$ is required
in the residual calculation. Checking the subspace
residual does not require a RR calculation.
Since the above measure monitors the quality of the whole approximate invariant subspace,
it does not allow one
to see whether the
subspace
contains good approximations to some of the eigenvectors
that can be locked and deflated from subsequent
PPCG iterations. This is a reason periodic
RR can be helpful. We will discuss
deflation in section~\ref{sec:deflation}.
In some cases, especially in the early PPCG iterations in which
significant changes in $X$ can be observed, it may not be
necessary to check the subspace residual. Since the objective of
the algorithm is to minimize the trace of $X^*AX$, it is reasonable
to use the relative change in the trace, which can be computed
quickly, as a measure for terminating the PPCG iteration.
To be specific, if $X$ and $X^{\prime}$ are approximations
to the desired invariant subspace obtained at the current and previous
iterations, we can use
\[
\frac{\left | \mathrm{trace}\left( X^* A X \right) -
\mathrm{trace}\left(X^{\prime *}A X^{\prime} \right) \right |}
{\mathrm{trace}\left(X^* A X\right)}
< \tau,
\]
as a criterion for terminating the PPCG iteration, where
$\tau$ is an appropriately chosen tolerance.
This criterion is often used in an iterative eigenvalue
calculation called within each self-consistent
field iteration for solving the Kohn--Sham nonlinear
eigenvalue problem~\cite{Hohenberg.Kohn:64, Kohn.Sham:65}.
\subsection{Buffer vectors}
When the $k$th eigenvalue is not well separated from the $(k+1)$st
eigenvalue of $A$, the convergence of that eigenvalue may be
slow in subspace-projection based solution methods, see, e.g., \cite{Knyazev:01}.
As noted in~\cite{Kn.Ar.La.Ov:07}, one way to overcome this, is to expand the block $X$
with $l$ additional columns $Y_l$ (a standard approach in electronic structure calculations) which we call \textit{buffer vectors}. In this case,
we set $X \gets [X, Y_l]$ and
apply Algorithm~\ref{alg:ppcg0} to the extended block with $k' = k+l$ columns. The main difference is that
one has to monitor the convergence only to the invariant subspace that is associated with the $k$ wanted
eigenvalues, i.e., only the initial $k$ columns of the expanded $X$ should be used to evaluate the convergence
metrics discussed in the previous section.
It is clear that introducing buffer vectors increases the cost of the algorithm, per iteration.
For example, the cost of matrix--block multiplications with $A$ becomes higher, more work is required to perform dense
linear algebra (BLAS3) operations, the number of iterations of the inner ``for'' loop in lines 6-12 increases to $k' = k+l$ steps, etc.
However, the number of buffer vectors $l$ is normally chosen to be small relative to $k$, e.g., $1-5\%$ of the number of targeted eigenpairs,
and the increase in the computational work per iteration is relatively small, while the decrease in iterations required to reach convergence can be substantial.
\subsection{Orthogonal projection of the search direction}
The projector $I - XX^\ast$ in the definition of search directions $W$ and $P$ turns out to be crucial for achieving rapid
convergence of the PPCG algorithm. This is in contrast to the LOBPCG algorithm,
where applying $I - XX^\ast$ is not as important,
at least not in exact arithmetic, because a new approximation
to the desired invariant subspace is constructed from the subspace
spanned by the columns of $X$, $W$ and $P$. The use of $I-XX^\ast$
in the construction of $W$ and $P$ does not change that subspace.
However, in PPCG, application of $I - XX^\ast$ affects the low-dimensional
subspaces spanned by individual columns of $X$, $W$, and $P$.
\begin{figure}[htbp]
\begin{center}%
\includegraphics[width=6cm]{FIG/proj_effects_sih4.eps}
\includegraphics[width=6cm]{FIG/proj_effects_Li318.eps}
\end{center}
\caption{Effects of the projection steps 4 and 5 of Algorithm~\ref{alg:ppcg0}
on convergence. The PPCG variants are applied to compute 10 (left) and 2,000 (right) lowest eigenpairs of the
converged Kohn-Sham Hamiltonian of the silane molecule (left) and the Li318 lithium-ion electrolyte (right).
}\label{fig:proj}
\end{figure}
As demonstrated in Figure~\ref{fig:proj}, the PPCG algorithm can be very sensitive to the orthogonal projector used in steps 4 and 5 of Algorithm~\ref{alg:ppcg0}. We observe that removing either of the two projection steps can lead to
a severe deterioration of convergence.
In Figure~\ref{fig:proj} (left) the PPCG algorithm is used to compute the 10
lowest eigenpairs of a Kohn-Sham Hamiltonian of the SiH4 (silane) molecule.
In Figure~\ref{fig:proj} (right), a similar computation is performed
for the Li318 (lithium-ion electrolyte) system, where $2,000$ eigenpairs are sought.
Interestingly, we observed that the effects of applying $I - XX^\ast$ are more pronounced in the cases where $A$ has multiple
eigenvalues. Furthermore, in some experiments, we noticed that skipping the projector only in $P$ may not alter the convergence; see Figure~\ref{fig:proj} (right).
Nevertheless, we recommend keeping $(I - XX^*)$ for computing both $W$ and $P$
to achieve robust convergence.
\subsection{Orthogonalization of the approximate invariant subspace}\label{subsec:qr}
There are a number of ways to obtain an orthonormal basis of $X$ after
its columns have been updated by the ``for'' loop in lines 6-12 of
Algorithm~\ref{alg:ppcg0}. If the columns of $X$ are far from
being linearly dependent, an orthonormalization procedure based on using
the Cholesky factorization of $X^{\ast} X = R^{\ast}R$, where $R$
is a unit upper triangular matrix, is generally efficient. In this case,
$X$ is orthonormalized by
\[
X \leftarrow X R^{-1},
\]
i.e., step 13 of Algorithm~\ref{alg:ppcg0} is given by the
QR decomposition based on the Cholesky decomposition (the Cholesky QR factorization).
If the columns of $X$ are almost orthonormal, as expected when
$X$ is near the solution to the trace minimization problem, we may
compute the orthonormal basis as
$
X \gets X (X^* X)^{-1/2}
$,
where $(X^* X)^{-1/2} = (I + Y)^{-1/2}$ can be effectively approximated by several
terms in the Taylor expansion of $f(x) = \sqrt{1 + y}$. This gives the
following orthogonalization procedure:
\[
X \leftarrow X (I - Y/2 + 3Y^2/8 - 5Y^3/16 + \cdots), \quad Y = X^* X - I.
\]
Since $Y$ is likely to be small in norm, we may only need three or four terms
in the above expansion to obtain a nearly orthonormal $X$.
An attractive feature of this
updating scheme is that it uses only dense matrix--matrix multiplications
which can be performed efficiently on modern high performance
computers. Note that the updated matrix represents an orthonormal factor in the polar
decomposition~\cite{Horn.Johnson:90} of $X$, which gives a matrix with orthonormal columns that is closest to $X$~\cite{Fan.Hoffman:55}.
\begin{figure}[htbp]
\begin{center}%
\includegraphics[width=6cm]{FIG/orth_skip.eps}
\end{center}
\caption{Effects of removing the orthonormalization step 13 of
Algorithm~\ref{alg:ppcg0}, when computing the $4$ lowest
eigenpairs of a Kohn-Sham Hamiltonian associated with the silane molecule.
The problem size is $2,103$.}\label{fig:orth_skip}
\end{figure}
We have observed that columns of $X$ often remain nearly orthonormal
after they are updated according to \eqref{eq:upd_block}.
Motivated by this observation, we experimented with performing
orthonormalization periodically (step 13 of Algorithm~\ref{alg:ppcg0})
in the PPCG iteration. Figure~\ref{fig:orth_skip} demonstrates the effects
of this strategy on the convergence of the algorithm.
The plotted convergence curves correspond to PPCG runs in which
orthonormalization is performed every $t$ steps, where $t = 1, 5, 8, 10, 15$.
In our implementation, we use the Cholesky QR factorization to
orthonormalize the columns of $X$.
As can be seen in Figure~\ref{fig:orth_skip}, skipping the QR factorization
of $X$ for a small number of PPCG iterations (up to 5 steps in this example)
does not affect the convergence of the algorithm. In this case, the loss of
orthogonality, which can be measured by $\|X^* X - I\|_F$, is
at most $O(10^{-1})$. Note that when the columns of
$X$ are not orthonormal, the projectors in steps 4 and 5 of
Algorithm~\ref{alg:ppcg0} become approximate projectors.
Nevertheless, the convergence of PPCG is not substantially affected as
long as $t$ is not too large.
However, if we reduce the frequency of the QR factorizations significantly,
the number of PPCG iterations required to reach convergence starts to increase.
For this example, when we perform the QR factorization every $15$ iterations,
PPCG fails to converge (within an error tolerance of $10^{-2}$) within 40
iterations. In this case, the loss of orthogonality in $X$ reaches
$O(1)$, which severely affects the convergence of the method.
Thus, in order to gain extra computational savings, we can devise an
optional heuristic based on the measured loss of orthogonality.
We can decide to skip step 13 of Algorithm~\ref{alg:ppcg0} if the
loss of orthogonality is relatively small.
\subsection{Periodic RR computation}
For many problems, the orthogonal projection of $W$ and $P$ against
columns of $X$ and subsequent orthogonalization are not enough to ensure that $X$ converges rapidly
to a basis of the desired invariant subspace.
We found that a practical remedy for avoiding possible convergence
degradation or failure is to perform the RR procedure periodically,
which has the effect of systematically repositioning the $j$th column
of $X$ towards the eigenvector associated with the $j$th eigenvalue of $A$.
In this case, since
each column of $X$ is forced to be sufficiently close to an eigenvector, minimizing
$k$ Rayleigh quotients separately
becomes
just as effective
as minimizing the trace of $X^*AX$ under the orthonormality constraint.
Therefore, in practical implementations, we perform RR periodically,
even though PPCG has been observed to converge without performing this step for some problems.
This is done in step 14 of Algorithm~\ref{alg:ppcg0}.
Another reason periodic RR may be advantageous is
that it provides an opportunity to lock converged eigenvectors
and reduce the number of sparse matrix vector multiplications
required to find the remaining unconverged eigenvectors.
Clearly, introducing the periodic RR calculation
increases the cost of some PPCG iterations and can potentially
make the algorithm less scalable due to the lack of scalability
of the dense eigensolver. However, the extra cost can be
offset by accelerated convergence of the algorithm and reduced number of sparse
matrix vector multiplications once some approximate eigenvectors
have converged.
In our PPCG implementation,
we control the frequency of the RR calls by a parameter
\textit{rr\_period}. Our numerical experiments suggest that a
good value of \textit{rr\_period} is between 5 and 10.
Because good approximations to desired eigenvectors do not
emerge in the first few PPCG iteration, \textit{rr\_period}
can be set to a relatively large value and then
decreased in later iterations when many converged eigenvector
approximations can be found and locked.
\subsection{Locking converged eigenvectors} \label{sec:deflation}
Even before the norm of the subspace residual $R \equiv AX - X(X^{\ast} A X)$
becomes small, some of the Ritz vectors associated with the subspace
spanned by columns of $X$ can become accurate. However, these Ritz
vectors generally do not reveal themselves through the norm of each
column of $R$ because each column of $X$ may consist of a linear
combination of converged and unconverged eigenvector approximations.
The converged eigenvector approximations
can only be revealed through the RR procedure.
As mentioned earlier, by performing an RR calculation, we rotate
the columns of $X$ to the Ritz vectors which are
used as
starting points for independent Rayleigh quotient minimization carried out
in the inner loop of the next PPCG iteration.
Once the converged Ritz vectors are detected, we lock these
vectors by keeping them in the leading columns of $X$.
These locked vectors are not updated in inner loop of the PPCG
algorithm until the next RR procedure is performed. We do not
need to keep the corresponding columns in the $W$ and $P$ matrices.
However, the remaining columns in $W$ and $P$ must be
orthogonalized against all columns of $X$. This ``soft locking'' strategy is along the lines of that described in~\cite{Kn.Ar.La.Ov:07}.
A detailed description of the PPCG algorithm, incorporating the above practical aspects, is given in Algorithm~\ref{alg:ppcg} of Appendix A.
\section{Numerical examples}\label{sec:example}
In this section, we give a few examples of how the PPCG algorithm
can be used to accelerate both the SCF iteration and the band structure
analysis in Kohn-Sham DFT calculations. We implement PPCG within
the widely used Quantum Espresso (QE) planewave pseudopotential density functional electronic structure code~\cite{QE-2009}, and compare the performance
of PPCG with that of the state-of-the-art Davidson solver implemented in QE. The QE code
also contains an implementation of a band-by-band conjugate
gradient solver. Its performance generally lags behind that of
the Davidson solver, however, especially when a large number of eigenpairs
is needed. Therefore, we compare to the Davidson solver here.
In QE, the Davidson algorithm can construct a subspace $Y$ of dimension
up to $4k$ before it is restarted. However, when the number of desired
eigenpairs $k$ is large, solving a $4k \times 4k$ projected eigenvalue
problem is very costly. Therefore, in our tests, we limit the subspace dimension
of $Y$ to $2k$.
The problems that we use to test the new algorithm are listed in
Table~\ref{tab:probs}. A sufficiently large supercell is used
in each case so that we perform all calculations at the $\Gamma$-point
only. The kinetic energy cutoff (ecut), which determines the
the number of planewave coefficients ($n_G$), as well as the
number of atoms ($n_a$) for each system are shown.
Norm conserving pseudopotentials are used to construct the
Kohn-Sham Hamiltonian.
Therefore, all eigenvalue problems we solve
here are standard eigenvalue problems, although our algorithm can be
easily modified to solve generalized eigenvalue problems.
The local density approximation (LDA)~\cite{Kohn.Sham:65} is used for
exchange and correlation. The particular choice of
pseudopotential and exchange-correlation functional is not
important here, however, since we focus only on the performance of the
eigensolver.
The distribution of eigenvalues for each problem
is plotted in Figure~\ref{fig:dos}. We can see that there are
several clusters of eigenvalues for Li318 and bulk Si.
They are insulating and semiconducting, respectively.
No visible spectral gap can be observed for Graphene512. It is known
to be a metallic system. The test cases thus encompass the full range of electronic structures from insulating to metallic.
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Problem & $n_{a}$ & ecut (Ryd) & $n_G$ \\ \hline
Li318 & 318 & 25 & 206,691 \\ \hline
Graphene & 512 & 25 & 538,034 \\ \hline
bulk Si & 1000 & 35 & 1,896,173 \\ \hline
\end{tabular}
\caption{Test problems}
\label{tab:probs}
\end{center}
\end{table}
\begin{figure}[htbp]
\centering
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/DOS_Li318.eps}
\caption{Li318}\end{subfigure}\hfill
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/DOS_graphene512.eps}
\caption{graphene512}\end{subfigure}\hfill
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/DOS_sicluster.eps}
\caption{bulk Si}\end{subfigure}
\caption{Eigenvalue distributions for test problems.}
\label{fig:dos}
\end{figure}
All tests were performed on Edison,
a Cray XC30 supercomputer maintained at the National Energy Research
Scientific Computer Center (NERSC) at Lawrence Berkeley National Laboratory.
Each node on Edison has two twelve-core Intel 2.4GHz
``Ivy Bridge" processor sockets. It is equipped with
64 gigabyte (GB) DDR3 1600MHz shared memory. However,
memory access bandwidth and latency are nonuniform across all cores.
Each core has its own 64 kilobytes
(KB) L1 and 256 KB L2 caches. A 32MB L3 cache is shared among
12 cores. Edison nodes are connected by a Cray Aries network with
Dragonfly topology and 23.7 TB/s global bandwidth.
We follow the same parallelization strategy implemented in the QE
package to perform the multiplication of the Hamiltonian and
a wavefunction and distributed dense matrix--matrix multiplications.
The most expensive part of the Hamiltonian and wavefunction
multiplication is the three-dimensional
FFTs. Each FFT is parallelized over $nz$ cores, where $nz$ is the
number of FFT grid points in the third dimension. Multiple FFTs
can be carried out simultaneously on different cores when the
``-ntg" option is used.
We do not use the multi-threaded feature of QE.
The planewave coefficients are partitioned and distributed by
rows. Therefore, the dense matrix--matrix multiplications are
carried out by calling the DGEMM subroutine in BLAS3 on
each processor and performing a global sum if necessarily.
The ScaLAPACK library is used to solve the dense projected
eigenvalue problem and to perform the Cholesky factorization
required in the Cholesky QR factorization. Because
ScaLAPACK requires a 2D square processor grid for these
computations, a separate communication group that typically
consists of fewer computational cores is created to
complete this part of the computation.
Table~\ref{tab:sqgrid} gives the default
square processor configurations generated by QE when a certain
number of cores are used to solve the Kohn-Sham problem.
Although we did not try different configurations exhaustively,
we found the default setting to be close to optimal, i.e.,
adding more cores to perform ScaLAPACK calculations generally
does not lead to any improvement in timing because of the
limited amount of parallelism in dense eigenvalue and
Cholesky factorization computations and the communication
overhead.
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c|c|}
\hline
ncpus & Processor grid\tabularnewline
\hline
\hline
200 & 10x10\tabularnewline
\hline
400 & 14x14\tabularnewline
\hline
800 & 20x20\tabularnewline
\hline
1,600 & 28x28\tabularnewline
\hline
2,400 & 34x34\tabularnewline
\hline
3,000 & 38x38\tabularnewline
\hline
\end{tabular}
\caption{Default ScaLAPACK processor grid configurations used by QE
for different total core counts.}
\label{tab:sqgrid}
\end{center}
\end{table}
}
\subsection{Band energy calculation}
We first show how PPCG performs relative to the Davidson algorithm
when they are used to solve a single linear eigenvalue problem
defined by converged electron density and its corresponding
Kohn-Sham Hamiltonian. Table~\ref{tab:bandperf} shows
the total wall clock time required by both the block Davidson
algorithm and the PPCG algorithm for computing the $k$ lowest eigenpairs
of a converged Kohn-Sham Hamiltonian. This is often known as the
band structure calculation, although we only compute band energies
and corresponding wavefunctions at the $\Gamma$-point of the Brillouin zone.
In all these calculations, we perform the RR procedure
every 5 iterations. Depending on the problem, the subblock
size \textit{sbsize} is chosen to be 5 or 50. In our experience, such
choice of \textit{sbsize} leads to satisfactory convergence behavior of the
PPCG algorithm (we address this question in more detail below).
For all tests, the number of buffer vectors \textit{nbuf} is set to 50.
{\small
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Problem & ncpus & $k$ & sbsize & Time PPCG & Time Davidson\tabularnewline
\hline
\hline
Li318 & 480 & 2,062 & 5 & 49 (43) & 84 (27) \tabularnewline
\hline
Graphene512 & 576 & 2,254 & 50 & 97 (39) & 144 (36) \tabularnewline
\hline
bulk Si & 2,400 & 2,550 & 50 & 189 (78) & 329 (77) \tabularnewline
\hline
\end{tabular}
\caption{Comparison of the total wall clock time (in seconds) used by PPCG and Davidson
to compute the lowest $k$ eigenpairs. Numbers in parentheses correspond to iteration counts.}
\label{tab:bandperf}
\end{center}
\end{table}
}
For both Li318 and Graphene512, we terminate the Davidson iteration
when the relative subspace residual norm defined in~\eqref{eq:resnrm}
is less than $tol = 10^{-2}$. Since residual norms are only calculated
when the Davidson iteration is restarted, the actual residual norm
associated with the approximate solution produced by the Davidson algorithm
may be much less than $tol$ upon termination. We use that relative residual
norm as the stopping criterion for the PPCG algorithm.
For bulk Si, we set $tol$ to $10^{-3}$.
The results shown in Table~\ref{tab:bandperf} indicate that PPCG performs
much better on the test problems than Davidson's method. We observe almost
a factor of two speedup in terms of wall clock time.
Note that the number of iterations required by PPCG is noticeably higher
than that of Davidson's method in some cases (e.g., Li318). However, most
PPCG iterations are much less expensive than Davidson iterations
because they are free of RR calculations. The reduced number of RR
calculations leads to better overall performance. As expected,
as the \textit{sbsize} value increases, the difference in total number of
outer iterations between PPCG and Davidson becomes smaller.
For example, for bulk Si and Graphene512, where \textit{sbsize} is relatively large ($50$), the number of iterations taken by PPCG and Davidson are
almost the same.
\begin{figure}[htbp]
\centering
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/Li318_nbnd_2062.eps}
\caption{Li318 ($k = 2062$)}\end{subfigure}\hfill
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/graphene512_nbnd_2254.eps}
\caption{Graphene512 ($k = 2254$)}\end{subfigure}\hfill
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/sicluster_nbnd_2550.eps}
\caption{bulk Si ($k = 2550$)}\end{subfigure}
\caption{Convergence of the Davidson and PPCG algorithms for band structure calculation.}
\label{fig:Liband}
\end{figure}
Figure~\ref{fig:Liband} shows how the Frobenius norms of the subspace
residuals change with respect to the elapsed time for
all three test problems reported in Table~\ref{tab:bandperf}.
Note that at some point, both PPCG and the Davidson method start to
converge more rapidly. The change in convergence rate is the result of
locking the converged eigenpairs, which significantly reduces
computational cost in performing $AX$.
In Tables~\ref{tab:Li318prof} and~\ref{tab:Siprof}, we provide a more
detailed timing breakdown for both PPCG and Davidson algorithms when
they are used to solve Li318 and bulk Si, respectively. We can clearly see that the wall clock time
consumed by the Davidson run is dominated by RR calculations.
The RR cost is significantly lower in PPCG. However, such
a reduction in RR cost is slightly offset by the additional cost
of performing Cholesky QR, which we enable in each PPCG iteration.
Its cost represents roughly 15\% of the total.
As has been discussed in Section~\ref{subsec:qr}, the number of these factorizations can, however, be further reduced, which
will lead to even more efficient PPCG.
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Computation & PPCG & Davidson \tabularnewline
\hline
\hline
GEMM & 16 & 11 \tabularnewline
\hline
$AX$ & 10 & 6 \tabularnewline
\hline
RR & 13 & 66 \tabularnewline
\hline
CholQR & 8 & 0 \tabularnewline
\hline
\end{tabular}
\caption{Timing profiles (in seconds) for PPCG and Davidson
when they are used to compute the 2,062 lowest eigenpairs of
the Li318 problem on 480 cores.}
\label{tab:Li318prof}
\end{center}
\end{table}
}
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Computation & PPCG & Davidson \tabularnewline
\hline
\hline
GEMM & 27 & 41 \tabularnewline
\hline
$AX$ & 94 & 96 \tabularnewline
\hline
RR & 40 & 191 \tabularnewline
\hline
CholQR & 19 & 0 \tabularnewline
\hline
\end{tabular}
\caption{Timing profiles (in seconds) for PPCG and Davidson
when they are used to compute the 2,550 lowest eigenpairs of
the bulk Si problem on 2,400 cores.}
\label{tab:Siprof}
\end{center}
\end{table}
}
We also note from Table~\ref{tab:Li318prof} that PPCG may spend more time in
performing dense matrix--matrix multiplications (GEMM) required to orthogonalize $W$ and $P$ against the current approximation to the desired invariant subspace
than the Davidson algorithm.
We believe the relatively high cost of GEMM operations is due to
the 1D decomposition of the planewave coefficient matrix used in QE, which
is less than optimal for machines with many processors.
The performance of GEMM depends on the size of the matrices being
multiplied on each processor. In the case of Li318, the dimension
of the local distributed $X$ is $720 \times 2062$, which does not lead
to optimal single-processor GEMM performance when $X^*X$ or
similar matrix--matrix multiplications are computed.
For bulk Si, the dimension of the local distributed
$X$ is $2024 \times 2550$, which is nearly a perfect square. The
optimized BLAS on Edison is highly efficient for matrices of this
size.
It also appears that PPCG can spend more time in performing $AX$ than
the Davidson method. The higher $AX$ cost in PPCG
can be attributed to its delayed locking of converged eigenpairs.
Because PPCG performs RR periodically, locking must also be performed
periodically even though many eigenpairs may have converged before
the next RR procedure is called. Although we use a constant
RR frequency value \textit{rr\_period}, it is possible to choose
it dynamically. In the first few PPCG iterations in which
the number of converged eigenvectors is expected to be low, we should
not perform the RR procedure too frequently, in order to reduce the RR cost.
However, when a large number of eigenvectors start to converge, it may be
beneficial to perform the RR procedure more frequently to lock the
converged eigenvectors as soon as they appear, and hence reduce the number of
sparse matrix multiplications. In our tests, we observe that
setting \textit{rr\_period} to a value between 5 and 10 typically yields
satisfactory performance.
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm]{FIG/RRSTEP_effects_iter.eps}
\includegraphics[width=6cm]{FIG/RRSTEP_effects.eps}
\caption{Effect of RR frequency on the convergence of
PPCG for the Li318 problem.}
\label{fig:rreffects}
\end{figure}
In Figure~\ref{fig:rreffects}, we demonstrate this finding by considering the effects of the RR frequency on the convergence of PPCG in terms of
iteration count (left) and time (right) for the Li318 system. We can see that calling
the RR procedure periodically is crucial for reaching convergence.
Completely removing the RR calculation generally leads to a (near) stagnation
of the algorithm. At the same time, it can be seen from
Figure~\ref{fig:rreffects} (left), that performing the RR procedure too
frequently does not necessarily accelerate PPCG convergence. For this
particular example, performing RR calculations every 15 iterations in fact results
in essentially the same convergence rate as that observed in
another PPCG run in which the RR computation is performed at every step.
The effect of the RR frequency becomes more clear when we examine
the change of residual norm with respect to the wall clock time.
As shown in Figure~\ref{fig:rreffects} (right),
the best \textit{rr\_period} value for Li318 is 5, i.e., invoking the
RR procedure every five steps achieves a good balance between
timely locking and reduction of RR computations. As mentioned above, in principle, one can vary the \textit{rr\_period} values during the solution process.
As a heuristic, \textit{rr\_period} can be set to a relatively large
number in the first several iterations, and then be gradually reduced
to provide more opportunities for locking converged eigenvectors.
We tried such a strategy for Li318 but did not observe significant
improvement. Therefore, we
leave \textit{rr\_period} at $5$ in all runs.
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm]{FIG/SBSIZE_effects_iter.eps}
\includegraphics[width=6cm]{FIG/SBSIZE_effects.eps}
\caption{Effect of block size on the convergence of
PPCG for the Li318 problem.}
\label{fig:blocksize}
\end{figure}
Figure~\ref{fig:blocksize} shows the effect of the block size \textit{sbsize} on PPCG convergence for the Li318 problem.
One can see that increasing the \textit{sbsize} value results in a smaller
number of iterations required to achieve the desired tolerance
(Figure~\ref{fig:blocksize}, left). This behavior is expected, because in the limiting case where
\textit{sbsize}$ = k$, PPCG becomes the LOBPCG method, which
optimizes in the full $3k \times 3k$ subspace in each iteration.
Figure~\ref{fig:blocksize} (right) demonstrates the effect of
\textit{sbsize} on the solution time. Remarkably, a larger
block size, which leads to a reduced iteration count, does not
necessarily result in better overall performance, even though
it tends to reduce outer iterations. This is in
part due to the sequential implementation of the {\tt for} loop in
step 11 in the current implementation. Since each block minimization
in the inner loop can take a non-negligible amount of time,
the inner {\tt for} loop can take a significant amount
of time, even though the loop count is reduced.
On the other hand, setting \textit{sbsize}$=1$ is not desirable either,
because of that tends to slow convergence and increase the outer PPCG iteration
count. Furthermore, the inner minimization cannot effectively
take advantage of BLAS3 operations in this case. For Li318, we observe that the
best \textit{sbsize} value is $5$.
In Figure~\ref{fig:pscale}, we plot how the wall clock times of PPCG and
Davidson change with respect to the number of cores when
applied to the Li318 and bulk Si problems.
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm]{FIG/Li318_nbnd_2062_scale.eps}
\includegraphics[width=6cm]{FIG/sicluster_nbnd_2550_scale.eps}
\caption{Scaling of the Davidson and PPCG algorithms
when used to compute $2,062$ and $2,550$ bands of
the converged Kohn-Sham Hamiltonian of the Li318 (left) and bulk Si (right) systems, respectively.}
\label{fig:pscale}
\end{figure}
We observe that both algorithms exhibit nearly
perfect parallel scalability when a relatively small number
of cores are used in the computation. However, as the
number of cores increases, the performance of both
algorithms stagnates.
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}
\cline{2-7}
\multicolumn{1}{c|}{} & \multicolumn{6}{c|}{ncpus} \tabularnewline
\cline{1-7}
\multicolumn{1}{|c|}{Computation} & 48 & 96 & 192 & 384 & 480 & 768 \tabularnewline
\hline
\hline
GEMM & 77 & 44 & 26 & 18 & 16 & 16 \tabularnewline
\hline
$AX$ & 33 & 28 & 17 & 13 & 10 & 11 \tabularnewline
\hline
RR & 40 & 24 & 16 & 13 & 13 & 13 \tabularnewline
\hline
CholQR & 28 & 16 & 9 & 8 & 8 & 9 \tabularnewline
\hline
Total & 186 & 116 & 71 & 53 & 49 & 50 \tabularnewline
\hline
\end{tabular}
\caption{Scaling of different computational components of PPCG for Li318.}
\label{tab:li318_ppcg}
\end{center}
\end{table}
}
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}
\cline{2-7}
\multicolumn{1}{c|}{} & \multicolumn{6}{c|}{ncpus} \tabularnewline
\cline{1-7}
\multicolumn{1}{|c|}{Computation} & 48 & 96 & 192 & 384 & 480 & 768 \tabularnewline
\hline
\hline
GEMM & 54 & 29 & 17 & 12 & 11 & 11\tabularnewline
\hline
$AX$ & 22 & 19 & 13 & 8 & 6 & 7\tabularnewline
\hline
RR & 370 & 170 & 105 & 77 & 66 & 66\tabularnewline
\hline
Total & 449 & 219 & 136 & 98 & 84 & 85\tabularnewline
\hline
\end{tabular}
\caption{Scaling of different computational components of the Davidson algorithm for Li318.}
\label{tab:li318_dav}
\end{center}
\end{table}
}
A closer look at the timing profiles consisting of wall clock
time used by different computational kernels, as shown in
Tables~\ref{tab:li318_ppcg}--\ref{tab:sicluster_dav},
reveals that the lack of scalability at high core count is
caused by the poor parallel scaling of both $AX$ and $GEMM$ calculations
at such core counts. A similar picture is observed for other cases
we have tested.
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}
\cline{2-7}
\multicolumn{1}{c|}{} & \multicolumn{6}{c|}{ncpus} \tabularnewline
\cline{1-7}
\multicolumn{1}{|c|}{Computation} & 200 & 400 & 800 & 1,600 & 2,400 & 2,800 \tabularnewline
\hline
\hline
GEMM & 202 & 104 & 57 & 35 & 27 & 26 \tabularnewline
\hline
$AX$ & 247 & 165 & 129 & 106 & 94 & 92 \tabularnewline
\hline
RR & 142 & 77 & 48 & 40 & 40 & 41 \tabularnewline
\hline
CholQR & 66 & 91 & 21 & 18 & 19 & 21 \tabularnewline
\hline
Total & 685 & 399 & 266 & 209 & 189 & 188 \tabularnewline
\hline
\end{tabular}
\caption{Scaling of different computational components of PPCG for bulk Si.}
\label{tab:sicluster_ppcg}
\end{center}
\end{table}
}
{\small
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}
\cline{2-7}
\multicolumn{1}{c|}{} & \multicolumn{6}{c|}{ncpus} \tabularnewline
\cline{1-7}
\multicolumn{1}{|c|}{Computation} & 200 & 400 & 800 & 1,600 & 2,400 & 2,800 \tabularnewline
\hline
\hline
GEMM & 248 & 138 & 76 & 47 & 41 & 38 \tabularnewline
\hline
$AX$ & 253 & 169 & 133 & 111 & 96 & 96 \tabularnewline
\hline
RR & 474 & 303 & 214 & 189 & 191 & 189 \tabularnewline
\hline
Total & 986 & 615 & 425 & 348 & 329 & 323 \tabularnewline
\hline
\end{tabular}
\caption{Scaling of different computational components of Davidson's algorithm for bulk Si.}
\label{tab:sicluster_dav}
\end{center}
\end{table}
}
The less than satisfactory scalability of GEMM is likely due
to the 1D partition of the planewave coefficients in QE.
We are aware of a recent change in the QE design to allow
planewave coefficients to be distributed on a 2D
processor grid such as the one used in ABINIT~\cite{ABINIT1, ABINIT2, Bottin.Leroux.Knyazev.Zerah:08} and
Qbox~\cite{QBOX}. However, the new version of the code is
still in the experimental stage at the time of this writing.
Hence we have not tried it. Once the new version of QE becomes
available, we believe the benefit of using PPCG to
compute the desired eigenvectors
will become even more substantial.
The poor scalability of $AX$ is due to the overhead related
to the all-to-all communication required in 3D FFTs. When a
small number of cores are used, this overhead is relatively
insignificant due to the relatively large ratio of
computational work and communication volume. However, when a
large number of cores are used, the amount of computation performed
on each core is relatively low compared to the volume of communication.
We believe that one way to reduce such overhead is to perform each
FFT on fewer than $n_z$ cores, where $n_z$ is the number of FFT grid
point in the third dimension. However, this would require
a substantial modification of the QE software.
\subsection{SCF calculation}
We ran both the block Davidson (Algorithm~\ref{alg:davidson}) and the
new PPCG algorithm to compute the solutions to the Kohn-Sham equations
for the three systems list in Table~\ref{tab:probs}. To account
for partial occupancy at finite temperature, we set the number
of bands to be computed to $k = 886$ for Li318; $k = 1,229$
for Graphene512; and $k = 2,000$ for bulk Si.
In general, it is not necessary to solve the linear
eigenvalue problem to high accuracy in the first few
SCF cycles because the Hamiltonian itself has not converged.
As the electron density and Hamiltonian
converge to the ground state solution, we should gradually
demand higher accuracy in the solution to the linear eigenvalue problem.
However, because the approximate invariant subspace obtained
in the previous SCF iteration can be used as a good starting
guess for the eigenvalue problem produced in the current
SCF iteration, the number of Davidson iterations required to
reach high accuracy does not necessarily increase.
The QE implementation of Davidson's algorithm uses
a heuristic to dynamically adjust the convergence tolerance
of the approximate eigenvalues as the electron density and Hamiltonian
converge to the ground-state solution. In most cases, the
average number of Davidson iterations taken in each SCF cycle is
around 2. We have not implemented the same heuristic for
setting a dynamic convergence tolerance partly because we do not
always have approximate eigenvalues. To be comparable to
the Davidson solver, we simply set the maximum number of
iterations allowed in PPCG to 2. In all our test cases, two
PPCG iterations were taken in each SCF cycle.
{\small
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Problem & ncpus & sbsize & PPCG & Davidson\tabularnewline
\hline
\hline
Li318 & 480 & 5 & 35 (40) & 85 (49)\tabularnewline
\hline
Graphene512 & 576 & 10 & 103 (54) & 202 (57)\tabularnewline
\hline
bulk Si & 2,000 & 5 & 218 (14) & 322 (14)\tabularnewline
\hline
\end{tabular}
\caption{Comparison of total wall clock time (in seconds) used by PPCG and Davidson algorithms to compute solutions of the Kohn-Sham equations.
Numbers in parentheses correspond to SCF iteration counts.}
\label{tab:bestperf_scf}
\end{center}
\end{table}
}
In Table~\ref{tab:bestperf_scf}, we report the overall time
used in both the Davidson and PPCG versions of the
SCF iteration for all three test problems. The SCF
convergence tolerance, which is used to terminate the SCF
iteration when the estimated total energy error predicted by
the so-called Harris--Foulkes energy functional~\cite{Har85, FouHay89}
is sufficiently small,
is set to $10^{-6}$ for Li318 and bulk Si.
It is set to $10^{-4}$ for Graphene512 because the SCF iteration
converges more slowly for this problem, hence takes much longer to run.
Note that the SCF convergence tolerance is defined internally by QE, which
we did not modify. Thus both the Davidson and PPCG versions of the SCF
iteration are subject to the same SCF convergence criterion.
We also use the same (``plain'') mixing and finite temperature smearing in
both the Davidson and PPCG runs.
Note that, following the discussion in section~\ref{subsec:qr},
we omit the Cholesky QR step in PPCG for the reported runs. Since only
two eigensolver iterations are performed per SCF iteration, this did not affect convergence and
resulted in a speedup of the overall computation. The \textit{sbsize} parameter has been set to $5$ for the Li318
and bulk Si systems, and to 10 for Graphene512.
\begin{figure}[htbp]
\centering
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/Li318_SCF_cv.eps}
\caption{Li318}\end{subfigure}\hfill
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/graphene_SCF_cv.eps}
\caption{Graphene512}\end{subfigure}\hfill
\begin{subfigure}{.33\textwidth}\includegraphics[width=\textwidth]{FIG/sicluster_SCF_cv.eps}
\caption{bulk Si}\end{subfigure}
\caption{Convergence of the SCF iteration with Davidson and PPCG algorithms.}
\label{fig:scf_cv}
\end{figure}
The convergence curves corresponding to the SCF runs of Table~\ref{tab:bestperf_scf} are shown in Figure~\ref{fig:scf_cv}.
We can clearly see that the PPCG based SCF iteration can be
nearly twice as fast as the Davidson based iteration. The figure also demonstrates the effects of skipping the Cholesky QR step, which further reduces the
PPCG run time.
It appears that, for the Li318 and Graphene512 examples,
a slightly fewer number of SCF iterations is needed to reach convergence
when PPCG is used to solve the linear eigenvalue problem in
each step.
In Figure~\ref{fig:pscale_scf}, we compare scalability of the SCF iteration based on the PPCG (without Cholesky QR) and Davidson algorithms.
We report the results for Li318 (left) and bulk Si (right). Similar to the case of the band structure calculations, both schemes scale
approximately up to the same number of cores for Li318, whereas scalability of PPCG is slightly better for the bulk Si example.
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm]{FIG/Li318_SCF_scale.eps}
\includegraphics[width=6cm]{FIG/sicluster_SCF_scale.eps}
\caption{Scaling of SCF iterations with the Davidson and PPCG algorithms
for Li318 (left) and bulk Si (right) systems.}
\label{fig:pscale_scf}
\end{figure}
\section{Conclusions}\label{sec:concl}
We presented a projected preconditioned conjugate gradient (PPCG) algorithm
for computing an invariant subspace associated with the smallest eigenvalues
of a large Hermitian matrix.
The key feature of the new algorithm is that it performs
fewer Rayleigh-Ritz computations, which are often the bottleneck in
iterative eigensolvers when the number of required eigenpairs is relatively
large (e.g., over thousands).
We discussed a number of practical issues that must be addressed in order to
implement the algorithm efficiently.
We implemented the PPCG algorithm within the widely used Quantum Espresso
(QE) planewave pseudopotential electronic structure software package.
We demonstrated that PPCG is nearly two times faster than the existing
state-of-the-art Davidson algorithm implemented in QE for a number of test problems.
We believe further performance gains can be achieved in PPCG relative
to other algorithms if the multiplication
of $A$ with a block of vectors $X$ and the dense matrix multiplications
such as $X^{\ast}X$ are implemented in a scalable fashion.
\paragraph{Acknowledgments.}
The authors thank Dr.~Erik~Draeger at the Lawrence Livermore National Laboratory for insightful comments and discussions.
|
\section{introduction}
J. Wiegold \cite{W76} showed that non-abelian free groups are not invariably generated. Kantor, Lubotzky and Shalev \cite{KLS} conjectured (in the form of a question) that every non-elementary Gromov hyperbolic group is non-invariably-generated. We show that this is indeed the case for the much wider class of convergence groups. We prove that every non-elementary torsion-free convergence group admits a proper (infinite rank) free subgroup that meets all the conjugacy classes. For convergence groups with torsion, we obtain a similar result, where the free subgroup is replaced by (an infinite) free product of cyclic groups, and we sometimes have to mod out by a finite normal subgroup.
\subsection{Invariable generation}
Recall that a subset $S$ of a group $G$ {\it invariably generates} $G$ if $G= \langle s^{g(s)} | s \in S\rangle$ for every choice of $g(s) \in G,s \in S$. One says that a group $G$ is {\it invariably generated}, or shortly IG, if such $S$ exists, or equivalently if $S=G$ invariably generates $G$.
Equivalently, $G$ is IG iff no proper subgroup meets every conjugacy class, or in other words, iff every transitive permutation representation on a non-singleton set admits a fixed-point-free element.
A well known simple counting argument shows that every finite group is IG. Obviously, abelian groups are IG. More generally J. Wiegold \cite{W76} showed that the class of IG groups is closed under extensions, hence contains all virtually solvable groups. Clearly this class is also closed to quotients.
On the other hand the class IG is not closed under direct union; for instance the group of finitely supported permutations of ${\mathbb{N}}$ is clearly not IG, since every element admits a conjugate whose support is contained in $2{\mathbb{N}}$. Moreover, Wiegold \cite{W77} gave an example of an IG group whose commutator is not IG, proving in particular that the class IG is not subgroup closed.
Let us list some classical examples of non-IG groups:
\begin{itemize}
\item Groups with a single non-trivial conjugacy class are non-IG (cf. \cite{HNN,Osin}).
\item $G=\text{SO}(3)$. To see that consider the standard rotation action of $G$ on the sphere $S^2$, and recall that every element fixes some point (and its antipode).
More generally, every nonabelian connected compact group is non-IG, since every abelian subgroup is contained in a maximal one and all maximal abelian subgroups are conjugated.
\item $G=\text{GL}_n({\mathbb{C}}),~n\ge 2$. Indeed every matrix can be conjugated to the Borel subgroup $B$ of upper triangular matrices. More generally, by \cite[Proposition 2.4]{KLS} a non-virtually-solvable linear algebraic group over an algebraically closed field is never IG.
\end{itemize}
\medskip
In \cite{W76}, Wiegold proved that the free group on two (or more) letters $F_{\{ a,b,\ldots\}}$ is non-IG by producing a list $L$ of conjugacy class representatives which are jointly independent. To recall his construction, let $\{ w_n\}$ be conjugacy class representatives which start and end with a non-zero power of $b$, then take $L=\{w_n^{a^n}:n\in{\mathbb{N}}\}$.
W. M. Kantor, A. Lubotzky and A. Shalev asked in \cite{KLS} (among many other questions) whether every non-elementary Gromov hyperbolic group is non-IG. In this note we treat this problem for the much wider class of convergence groups.
\subsection{Convergence groups} \label{sec:convergence}
Convergence groups were introduced by H. Furstenberg in \cite{Fur} under the name ``Dynkin groups".
The term {\it convergence groups}
was given by Gehring and Martin in \cite{GM:Convergence} (who studied properties of
Kleinian groups through their action on the boundary, thus restricted to
convergence group actions on spheres).
These are groups admitting an action on a compact metrizable space $K$ such that the corresponding action on triples $\{(x_1,x_2,x_3)\in K^3: i\ne j\Rightarrow x_i\ne x_j\}$ is proper (see also the equivalent Definition \ref{defn:convergence} below).
This class includes
Gromov hyperbolic groups
\cite{Tukia:convergence, Tukia:Uniform_convergence_groups, Bowditch:characterisation_hyperbolic}, relatively
hyperbolic groups \cite{Yaman:Relatively_Hyp_Convergence}, as well as any group acting properly discontinuously on a complete locally compact Gromov hyperbolic space, and in particular every non-elementary discrete subgroup of a rank one simple Lie group. Obviously, this class is subgroup closed.
\medskip
The main purpose of this paper is to prove:
\begin{thm}\label{thm}
Let $\Gamma$ be a non-elementary convergence group admitting a faithful minimal convergence action. Then $\Gamma$ has an independent set $I$ consisting of one representative of every non-trivial conjugacy class, which generates a proper subgroup $\langle I\rangle\lneqq \Gamma$.
\end{thm}
By definition, a set $I$ in a group is {\it independent} if the subgroup $\langle I\rangle$ is the free product of the cyclic groups $\langle\gamma\rangle,~\gamma\in I$.
The kernel $\ker(\curvearrowright)$ of a convergence action $\Gamma\curvearrowright X$ is always finite, and in many natural examples it is trivial. Theorem \ref{thm} could be stated without the faithfulness assumption, in which case $I$ would consist of representatives of the conjugacy classes not belonging to the finite group $\ker(\curvearrowright)$.
\begin{cor}
Non-elementary convergence group are non-IG.
\end{cor}
\section{Definitions and useful properties of convergence groups}
We refer to \cite{Bowditch:Convergence} for supplementary background on convergence groups.
\begin{definition}
\label{defn:collapsing} An infinite set $\Phi$ of homeomorphisms
of a compact topological space $X$ is called {\it collapsing} with respect to a
pair of (not necessarily distinct) points $(a,r)$ if for every
pair of compact sets $K \subset X \setminus \{r\}$ and $L \subset
X \setminus \{a\}$, the set $\{\phi \in \Phi : \phi\cdot K \cap L \ne
\emptyset \}$ is finite. We shall then call $a$ the attracting
point and $r$ the repelling point of $\Phi$. A set $\Phi$ is {\it
collapsing} if it is collapsing with respect to some pair of
points.
\end{definition}
\begin{definition}
\label{defn:convergence} An action of a group $\Gamma$ on a compact Hausdorff space $X$ is
said to have the convergence property if every infinite subset $\Phi < \Gamma$ contains an
infinite subset $\Phi' \subset \Phi$ which is collapsing. A {\it convergence group} is a group
that admits a convergence action on some compact metrisable space consisting of more than two points.
\end{definition}
It follows easily from the definition that the kernel of every convergence action is finite. In particular a convergence group with no
finite normal subgroups admits a faithful convergence action.
It also follows from the definition that every element of infinite order in $\Gamma$ fixes
either one or two points of $X$. As a Corollary we obtain the ``usual'' classification of elements
into three mutually exclusive categories: elements of finite order are called {\it elliptic},
elements of infinite order fixing exactly one point are called {\it parabolic} and elements of
infinite order fixing two points of $X$ are called {\it loxodromic}.
We may denote the fixed points of a non-torsion element $g$ by $g^+$ and $g^-$, and think of them as the future limit and past limit, so that the following is satisfied:
\begin{itemize}
\item $\forall x\ne g^-$ we have $g^n\cdot x\to g^+$,
\item $g^+=g^-$ iff $g$ is parabolic.
\end{itemize}
Moreover, given any pair of neighbourhoods $g^-\in R,g^+\in A$ there is $n_0$ such that for all $n\ge n_0$, $g^n\cdot (X\setminus R)\subset A$.
For $x\ne g^-$ and a set $U\subset X$ containing $g^+$ in its interior, we denote by $N(g,U,x)\in{\mathbb{N}}$ the minimal integer such that
$$
n\ge N(g,U,x)\Rightarrow g^n\cdot x\in U.
$$
A convergence group is called elementary if
it is finite or if it stabilizes a nonempty subset of $X$ with at most $2$ elements.
The {\it limit set} $L(\Gamma)\subset X$ is characterised as the set of all limit points $\lim \gamma_n\cdot x$, where $x\in X$ is constant and $(\gamma_n)$ is a sequence of distinct elements in $\Gamma$, or as the closure of the set $\{ g^+:g\in\Gamma~\text{non-torsion}\}$, assuming $\Gamma$ is non-torsion, which is usually the case (see Lemma \ref{lem:lox}). The assumption that $\Gamma$ is non-elementary is equivalent to $\text{Card}(L(\Gamma))>2$ in which case $L(\Gamma)$ is the unique minimal $\Gamma$-invariant compact set and it is perfect, i.e. has no isolated points. The minimality of $\Gamma\curvearrowright L(\Gamma)$ means that every orbit is dense. Replacing $X$ by $L(\Gamma)$ we may always suppose that the action of $\Gamma$ on $X$ is minimal.
The following two lemmas are well known and elementary:
\begin{lem}\label{lem:lox}
A non-elementary convergence
group contains a loxodromic element. Moreover the set of pairs $\{ (g^-,g^+)\in X^2:g~\text{is loxodromic}\}$ is dense in $L(\Gamma)^2$.
\end{lem}
\begin{proof}[Sketched proof]
If $\Phi\subset \Gamma$ is a collapsing sequence with attracting and repelling points $a,r\in L(\Gamma)$, we may, up to multiplying $\Phi$ by some $\gamma_0\in\Gamma$, suppose that $a\ne r$. Let $O_a,O_r$ be disjoint non-complementary open neighbourhoods of $a,r$ respectively, and pick $\phi\in\Phi$ with $\phi\cdot (X\setminus O_r)\subset O_a$. This forces $\phi$ to be loxodromic, proving the existence statement.
To prove the density statement, let $\phi$ be an arbitrary loxodromic element and let $U,V\subset L(\Gamma)$ be disjoint non-complementary open sets. Since the action $\Gamma\curvearrowright L(\Gamma)$ is minimal, there are $\alpha,\beta\in\Gamma$ such that $\alpha\cdot \phi^+\in U$ and $\beta\cdot \phi^-\in V$. Pick $n\in{\mathbb{N}}$ sufficiently large so that $\phi^n\cdot (X\setminus \beta^{-1}V)\subset \alpha^{-1}\cdot U$. Then we have
$\alpha \phi^n \beta^{-1} \cdot (X\setminus V)\subset U$. This implies that $g=\alpha \phi^n \beta^{-1}$ is loxodromic with $(g^-,g^+)\in V\times U$.
\end{proof}
\begin{lem}\label{lem:placement}
Let $\Gamma\curvearrowright X$ be a minimal non-elementary convergence action. Given a proper closed subset $\Sigma\subset X$ and an open subset $O\subset X$ there is an element $\delta\in\Gamma$ for which $\delta\cdot\Sigma\subset O$.
\end{lem}
\begin{proof}
By Lemma \ref{lem:lox} there is a loxodromic element $g\in \Gamma$ with $g^+\in O$ and $g^-\notin \Sigma$. Thus for a sufficiently large $n$ we have $g^n\cdot \Sigma\subset O$.
\end{proof}
We will make use of the fact that:
\begin{prop}\label{lem:countable}
A convergence group is countable.
\end{prop}
\begin{proof}
Let $\Gamma\curvearrowright X$ be a convergence action on a metrisable compact space $X$. Pick three distinct points $a,b,c\in X$ and a countable dense set $D\subset X$. Suppose by way of contradiction that $\Gamma$ is uncountable. Since for every $g\in\Gamma$, $g\cdot D$ is dense, and $X$ is second countable, a pigeon hall type argument implies that for some $x,y,z\in D$, the set
$$
\{ g\in\Gamma: g\cdot x\in U_a, g\cdot y\in U_b,g\cdot z\in U_c\}
$$
is uncountable and in particular non-empty for any triple of open sets $a\in U_a,b\in U_b,c\in U_c$. Thus we can produce a sequence of distinct elements $g_n\in \Gamma$ such that $g_n\cdot (x,y,z)\to (a,b,c)$ in $X^3$, in contrast to the convergence property.
\end{proof}
The following proposition is essential for our purpose:
\begin{prop}[Main Proposition]\label{prop:main}
Let $\Gamma\curvearrowright X$ be a minimal non-elementary convergence action. Then for every non-torsion element $\gamma\in\Gamma$ there are two sets $\Omega^+,\Omega^-$ such that $\gamma^n\cdot (X\setminus \Omega^-)\subset \Omega^+,~\forall n\in{\mathbb{N}}$ and the complement $X\setminus (\Omega^-\cup \Omega^+)$ has a non-empty interior.
\end{prop}
\begin{proof}
Let $A$ be a neighbourhood of $\gamma^+$ which is not dense in $X$.
Set
$$
A_n:=\{ x\in X:N(\gamma,\overline{A},x)\le n\}.
$$
Then
\begin{itemize}
\item the $A_n$ are closed,
\item $A_n\subset A_{n+1}$,
\item $\cup_n A_n\supset X\setminus\{ \gamma^-\}$, and
\item $A_{n+1}=\gamma^{-1}\cdot A_n$.
\end{itemize}
It follows from the Baire category theorem that there is a minimal $n\in{\mathbb{N}}$ such that $A_n\setminus A_1$ has a non-empty interior.
Then $A_n\setminus A_{n-1}$ has a non-empty interior. However, as the sets $A_i\setminus A_{i-1}$ are all homeomorphic, we deduce that $n=2$.
Thus we may take
$$
\Omega^+=A_1~\text{and}~\Omega^-=X\setminus A_2.
$$
\end{proof}
Let us say that a group action $\Gamma\curvearrowright X$ on a topological space is {\it generically free} if the fixed point set of any non-trivial element (an element not belonging to the kernel) is nowhere-dense. As Shahar Mozes pointed out to me, it is easy to construct an example of a non-elementary discrete subgroup of $\text{Aut}(T)$, where $T$ is a $4$-regular tree with a non-trivial elliptic element that fixes an open set in the boundary $\partial(T)$. The next proposition shows that this cannot happen for minimal convergence actions:
\begin{prop}
A non-elementary minimal convergence action is generically free.
\end{prop}
\begin{proof}
We will argue by way of contradiction.
Suppose that $\Gamma$ admits a minimal non-elementary convergence action $\Gamma\curvearrowright X$ which is not generically free, and let $\gamma\in\Gamma$ be an element whose fixed point set $A=\text{Fix}(\gamma)$ is a proper closed subset with a non-empty interior $A^\circ$.
In view of Lemma \ref{lem:placement} there is an element $\delta\in\Gamma$ such that $\delta^{-1}\cdot A\varsubsetneq A^\circ$. This implies that $\delta^{n-1}\cdot A\varsubsetneq \delta^{n}\cdot A$, for every $n\in {\mathbb{N}}$. Let $\gamma_n=\gamma_{n-1}^\delta=\gamma^{\delta^n}$, then $\text{Fix}(\gamma_n)=\delta^n\cdot A$ form a strictly increasing sequence of closed sets, and in particular the set $\{\gamma_n:n\in{\mathbb{N}}\}$ is infinite.
On the other hand, since $X$ is perfect, the non-empty open set $A^\circ$ contains $3$ distinct points, and as the action is convergence, its pointwise stabiliser is finite. A contradiction.
\end{proof}
In view of Baire's category theorem we deduce:
\begin{cor}\label{cor:wandering}
Let $\Gamma\curvearrowright X$ be a non-elementary minimal convergence action, then there is a point $x\in X$ with a trivial stabiliser.
\end{cor}
For an element $\gamma\in\Gamma$, let us say that a subset $\Sigma\subset X$ is $\gamma$-wandering if all the $\langle\gamma\rangle$-translations of $\Sigma$
are pairwise disjoint or, in other words, if $\Sigma\cap\gamma^n\cdot\Sigma=\emptyset$ whenever $n\notin \text{ord}(\gamma){\mathbb{Z}}$.
In particular, a singleton set is $\gamma$-wandering iff the corresponding $\langle\gamma\rangle$-orbit is faithful. Obviously a subset of a $\gamma$-wandering set is also $\gamma$-wandering.
\begin{cor}\label{cor:open-wandering}
Let $\Gamma\curvearrowright X$ be a faithful minimal convergence action. Then for every $\gamma\in\Gamma$ there is a $\gamma$-wandering open set $O=O(\gamma)$ in $X$.
\end{cor}
\begin{proof}
Suppose first that $\gamma$ has infinite order and let $\Omega^+,\Omega^-\subset X$ be the corresponding sets given in Proposition \ref{prop:main}. By Proposition \ref{prop:main}, $X\setminus (\Omega^+\cup\Omega^-)$ is $\gamma$-wandering and contains an open set.
Next consider the finite order case. By Corollary \ref{cor:wandering}, there is a $\gamma$-wandering point $x\in X$. By continuity of the action, if $O$ is a sufficiently small neighbourhood of $x$ then $O$ is $\gamma$-wandering.
\end{proof}
\section{The proof of the Theorem \ref{thm}}
Let us reformulate Theorem \ref{thm}:
\begin{thm}\label{thm:torsion}\label{thm:main}
Let $\Gamma$ be a group admitting a faithful minimal convergence action $\Gamma\curvearrowright X$ with $\text{card}(X)> 2$. Then there is a set of representatives $I$ for the non-trivial conjugacy classes of $\Gamma$, such that $H=\langle I\rangle$ is the free product of the cyclic groups $\{\langle \gamma_i\rangle\}_{\gamma_i\in I}$ and is strictly contained in $\Gamma$.
\end{thm}
\begin{rem}
In fact, for any function
$$
F:\{ \text{non-trivial conjugacy classes of}~\Gamma\}\to {\mathbb{N}}\cup\{\infty\},
$$
one can produce an independent set $I_F$ which intersects each non-trivial conjugacy class $C$ exactly $F(C)$ times.
\end{rem}
\begin{lem}\label{lem:wandering}
Let $\Gamma\curvearrowright X$ be a minimal faithful convergence action. Given an element $\gamma\in\Gamma$ and a proper closed subset $\Sigma\subset X$, there is an element $\delta=\delta(\gamma,\Sigma)$ for which $\Sigma$ is $\gamma^\delta$-wandering.
\end{lem}
\begin{proof}
Let $O$ be a $\gamma$-wandering open set given by Corollary \ref{cor:open-wandering}.
Let $\delta\in\Gamma$ be an element satisfying $\delta\cdot \Sigma\subset O$ (see Lemma \ref{lem:placement}). Then $\Sigma$ is $\delta^{-1}\gamma\delta$-wandering.
\end{proof}
The following
proposition is most suitable in our situation:
\begin{prop}\label{prop:ping-pong}
Suppose that $\alpha_n,~n=1,2,\ldots$ are elements of $\Gamma$ and $\Omega_n$ are pairwise disjoint sets, such that $X\setminus \Omega_n$ is $\alpha_n$-wandering for $n=1,2,\ldots$. Then
$$
\Delta:=\langle\alpha_n,~n=1,2,\ldots\rangle=*_n\langle\alpha_n\rangle
$$
is the free product of the cyclic groups $\langle\alpha_n\rangle,~n=1,2,\ldots$.
Moreover the limit set $L(\Delta)$ is contained in $\overline{\cup_n\Omega_n}$.
\end{prop}
\begin{proof}
The proof of the first statement is by a standard ``ping-pong argument". To keep the argument clean let us suppose that $X\setminus\cup_n\Omega_n$ is non-empty and contains the point $x$. For $i\in{\mathbb{N}}$ and $j\in {\mathbb{Z}}$ which is not a multiple of $\text{ord}(\alpha_i)$, let $X_i^j=\alpha_i^j\cdot(X\setminus\Omega_i)$ and observe that by the assumptions of the proposition the sets $X_i^j$ are pairwise disjoint. We claim that for $\alpha\in\Delta$ we can recover the unique reduced expression of $\alpha$ as a word in the $\alpha_i$'s by a recursive algorithm based on the position of $\alpha\cdot x$. Indeed, if $\alpha=\alpha_{i_1}^{j_1}\alpha_{i_2}^{j_2}\cdots\alpha_{i_k}^{j_k}$ then $\alpha\cdot x\in X_{i_1}^{j_1}$, and $\alpha_{i_1}^{-j_1}\alpha\cdot x\in X_{i_2}^{j_2}$, etc.
In order to show the second statement recall that $L(\Delta)$ is the set of limits $\lim_n\gamma_n\cdot x$ where $(\gamma_n)$ runs over all sequences in $\Delta$ for which the limit exists. Since for every $\gamma\in \Delta$, $\gamma\cdot x\in \cup_{i,j} X_i^j\subset \cup_i\Omega_i$ the statement follows.
\end{proof}
\begin{proof}[The proof of Theorem \ref{thm:torsion}]
Let $\Gamma\curvearrowright X$ be a faithful minimal convergence group action. Let $\{C_n\}_{n\in{\mathbb{N}}}$ be the non-trivial conjugacy classes of $\Gamma$, and let $\Omega_n,~n\in{\mathbb{N}}$ be arbitrary pairwise disjoint open subsets of $X$ whose union is not dense in $X$. In view of Lemma \ref{lem:wandering} we can pick a representative $\gamma_n\in C_n$ for each $n$, such that $X\setminus\Omega_n$ is $\gamma_n$-wandering. By Proposition \ref{prop:ping-pong} the set $\{\gamma_n\}$ is independent, i.e. $\Delta:=\langle\gamma_n:n\in{\mathbb{N}}\rangle$ is the free product $*_n\langle \gamma_n\rangle$. Finally, since $\cup\Omega_n$ is not dense, the limit set of $\Delta$ is a proper subset of $X$ and therefore $\Delta$ is a proper subgroup of $\Gamma$.
\end{proof}
{\it Acknowledgement:} I would like to thank Uri Bader for a valuable remark that enabled removing an unnecessary assumption that appeared in the first manuscript of this paper.
|
\section{Almost constant scalar curvature metric and K-semistability}
\subsection{Definition of an almost cscK metric}
Let $X$ be a polarized manifold by an ample line bundle $L$. In \cite{D4}, S.K. Donaldson introduced the notion of a test configuration for the couple $(X,L)$. Essentially, it consists in a $\mathbb{C}^*$-equivariant flat projective family $\pi:(\mathcal{X},\mathcal{L})\rightarrow \mathbb{C}$ with generic fibre $\pi^{-1}(t)$ isomorphic to $(X,L)$ for $t\neq 0$ while the central fibre, that may be singular, has a $\mathbb{C}^*$ action. Using this action, one can define a numerical invariant $F_1(\mathcal{X},\mathcal{L})$, which generalizes the Futaki invariant. Then $(X,L)$ is said to be K-stable if $F_1(\mathcal{X},\mathcal{L})>0$ for any non trivial test-configurations with general fibre $(X,L)$. We refer to \cite{RT, Thnotes} for details about this definition and to \cite{Stoppa2} for the notion of non trivial test-configuration. The Yau-Tian-Donaldson conjecture asserts that the existence of a cscK in the class $c_1(L)$ is equivalent to the K-stability of $(X,L)$.\\
It is natural to ask what is happening at the limit case for K-semistability. \\
Let us denote $\hat{s}_L$ the average of the scalar curvature in the class $c_1(L)$, which is a topological invariant. For $\omega$ a K\"ahler metric in $c_1(L)$, we denote $\mathrm{scal}(\omega)$ its scalar curvature. From the main result of \cite{D5}, we know that if there exists a sequence of metric $\omega_\epsilon$ such that $$\Vert\mathrm{scal}(\omega_\epsilon)-\hat{s}_L\Vert_{L^2}\rightarrow 0,$$ then the manifold is K-semistable. The converse is an open question as far as we know.
\begin{definition}
Given $(X,L)$ a projective manifold, we say that there exists an almost cscK metric in the class $c_1(L)$ in $\mathrm{C}^r$ topology $(r\in \mathbb{N})$ if there is a family of K\"ahler metrics $\omega_\epsilon\in c_1(L)$ such that $$\Vert\mathrm{scal}(\omega_\epsilon)-\hat{s}_L\Vert_{\mathrm{C}^r}\rightarrow 0$$ when $\epsilon\rightarrow 0$.
\end{definition}
In the case of the anticanonical class, this definition appeared first in \cite{Bando} where it is related to the existence of a lower bound for the Mabuchi K-energy. Obviously, from Donaldson's result, a manifold $(X,L)$ endowed with an almost cscK metric is K-semistable.
\subsection{Construction of almost cscK metric}\label{Sect1}
Let $E$ be an irreducible Mumford semistable vector bundle of rank 2 over a polarized manifold $(B,L_B)$, given by a non-split exact sequence of line bundles
$$0\rightarrow L_1 \rightarrow E \rightarrow L_2 \rightarrow 0,$$
with $c_1(L_1)=c_1(L_2)$. Let us assume that $h_1,h_2$ are projectively flat metrics on $L_1,L_2$ satisfying $F_{h_1}=\deg_L(L_1)\omega=F_{h_2}$ with $\omega$ a cscK metric in the class $c_1(L_B)$.
Consider the holomorphic structure on $E$ that has the following form
$$\bar{\partial}_E=\begin{pmatrix}
\bar{\partial}_{L_1} & \alpha \\
0 & \bar{\partial}_{L_2}
\end{pmatrix}
$$
where $\alpha$ is a smooth section of $\Omega^{0,1}(Hom(L_1,L_2))$, see \cite[Chapter I, Section 6]{Kob}. Then
one has for the curvature of $E$, and denoting $\mu(E)$ the slope of $E$,
\begin{align*}
\Vert F_{E}- \mu(E)\mathrm{Id}_E\omega\Vert_{\mathrm{C}^r} \leq&\Vert F_{h_1}- \deg(L_1)\omega \Vert_{\mathrm{C}^r} + \Vert F_{h_2}- \deg(L_2)\omega\Vert_{\mathrm{C}^r}\\
& + 2\Vert \alpha\Vert_{\mathrm{C}^r}^2 + 2\Vert \bar\partial^* \alpha\Vert_{\mathrm{C}^r}^2
\end{align*}
We can do a gauge change of the form $g=\begin{pmatrix} \xi & 0 \\ 0 & \xi^{-1} \end{pmatrix}$ and we obtain $g(\bar{\partial}_E)=\begin{pmatrix}
\bar{\partial}_{L_1} & \xi^{-2}\alpha \\
0 & \bar{\partial}_{L_2}
\end{pmatrix}.$
For any any $\epsilon>0$ and for any $r>0$, we can find the gauge transformation $\xi$ such that $$2\xi^{-2}(\Vert \alpha\Vert_{\mathrm{C}^r}^2 + \Vert \bar\partial^* \alpha\Vert_{\mathrm{C}^r}^2)< \epsilon.$$ This provides a structure $h_E$ (depending on the parameters $\epsilon,r$) such that $$ \Vert F_{E,h_E}-\mu(E)\mathrm{Id}_E\omega\Vert_{\mathrm{C}^r}< \epsilon.$$ Note that fixing the holomorphic structure with variation of the metric, or fixing the metric with variation of the holomorphic structure is geometrically equivalent in this setup. Therefore, we have obtained an approximate Hermitian-Einstein structure in the sense of \cite[Chapter IV]{Kob}, from which we deduce an almost cscK metric on $\mathbb{P}(E)$ using the next lemma.\\
From now we assume that $E$ is ample, that is $\mathcal{O}_{\mathbb{P}(E)}(1)$ is a positive line bundle (without loss of generality we can tensorize $E$ by a sufficiently ample line bundle $L_C$, use the identification $\mathbb{P}(E\otimes L_C)\simeq \mathbb{P}(E)$ and the induced approximate Hermitian structure).
\begin{lemma}\label{tech}
From $h_E$ hermitian metric on the bundle $E$, one can define a metric $\hat{h}_E$ on $\mathcal{O}_{\mathbb{P}(E)}(1)$ which curvature is denoted $\hat{\omega}_E$ and is a K\"ahler form. Then at $v\in \mathbb{P}(E)$, with $\pi(v)=x\in C$, one has pointwise
$$\hat{\omega}_E = \pi^* \left(\frac{\sqrt{-1}}{\Vert v\Vert^2_{h_E}}\langle F_{E,h_E}(v),v\rangle_{h_E}\right)+{\omega_{FS}}_{\vert \mathbb{P}(E)_x}$$
and ${\omega_{FS}}_{\vert \mathbb{P}(E)_x}$ is the Fubini-Study metric at $\mathbb{P}(E)_x$.
\end{lemma}
We refer to \cite[Chapter V, §15.C]{De2} for a proof. A direct consequence of the previous lemma and the existence of an approximate Hermitian-Einstein structure with respect to a cscK metric is the following proposition.
\begin{proposition}\label{prop0}
Let $E\rightarrow B$ be an ample Mumford semistable rank 2 vector bundle induced by a non-split exact sequence of projectively flat line bundles as above over a cscK polarized manifold $(B,L)$. Then for any $r\in \mathbb{N}$, there is an almost cscK metric on the ruled surface $\pi\colon \mathbb{P}(E)\to B$ in $\mathrm{C}^r$ topology.
\end{proposition}
If the base manifold is a curve of genus $g>1$, line bundles are automatically projectively flat, the exact sequence does not split for $L_1$ not isomorphic to $L_2$ since $h^1(C,L_2\otimes L_1^*)=g-1>0$ and there exists a cscK metric on the base manifold. Thus we obtain the next result.
\begin{corollary}\label{cor1}
Consider $E$ a rank 2 vector bundle on a curve $C$ of genus $\geq 2$. Assume that $E$ is Mumford semistable. Then for any $r\in \mathbb{N}$, there is an almost cscK metric on the ruled surface $\pi\colon \mathbb{P}(E)\to C$ in $\mathrm{C}^r$ topology.
\end{corollary}
Note that if $E$ is not irreducible, then it is actually a direct sum of line bundles and thus we know the existence of a genuine cscK metric on the projectivisation, see for instance \cite{AT1}.
\subsection{Computation of the Donaldson-Futaki invariant}
\begin{proposition}\label{prop}
Consider $E$ an ample irreducible Mumford semistable vector bundle which is not stable over a curve of genus $g>1$. Then $(\mathbb{P}(E),\mathcal{O}_{\mathbb{P}(E)}(1))$ is not K-polystable and not asymptotically Chow polystable.
\end{proposition}
\begin{proof}
This is a consequence of \cite[Theorem 5.13]{RT}, where it is done a computation of the Donaldson-Futaki invariant for the test configuration induced by a deformation to the normal cone of $\mathbb{P}(F)$ where $F$ is any subbundle of $E$. This computation shows that the Donaldson-Futaki invariant for such a test-configuration is a multiple of the differences of slopes $\mu(E)-\mu(F)$. Remark that with \cite[Proposition 4.1, Theorem 4.5]{DV1}, it is also proved that $\mathcal{O}_{\mathbb{P}(E)}(1)$ is not asymptotically Chow polystable.
\end{proof}
\begin{proposition}\label{prop3}
Assume that $E\rightarrow B$ an ample Mumford semistable rank 2 vector bundle induced by a non-split exact sequence of projectively flat line bundles over a cscK polarized manifold $(B,L)$, as constructed in Section \ref{Sect1}. Then $(\mathbb{P}(E),\mathcal{O}_{\mathbb{P}(E)}(1))$ is not K-polystable and not asymptotically Chow polystable.
\end{proposition}
\begin{proof}
Let us denote $b=\dim_\mathbb{C} B$. We compute the Donaldson-Futaki invariant $F_1$ for the test configuration induced by a deformation to the normal cone of $\mathbb{P}(L_1)$. Note that $\mu(L_1)=\mu(E)$. As explained in \cite{RT} (see also \cite{KR1}), $F_1=a_1b_0-a_0b_1$
where one has defined
\begin{align*}
p(r)&=h^0(\mathbb{P}(E),\mathcal{O}_{\mathbb{P}(E)}(r))=a_0r^{b+1}+a_1r^b+...\\
w(r)&=\sum_{i=0}^r ih^0(B,L_1^i\otimes L_2^{r-i})=b_0r^{b+2}+b_1r^{b+1}+...
\end{align*}
Now, under our assumptions, using the fact that $c_1(E)=2c_1(L_1)$, the polynomials $p$ and $w$ are proportionals. This shows that $F_1$ vanishes and a similar reasoning can be done to get that the Chow weight associated to this test configuration also vanishes.
\end{proof}
\begin{corollary}\label{cor}
There exist examples $(X,L)$ of polarized manifolds such that $L$ is K-semistable and not K-polystable. In particular, there are examples of non convergent sequence of almost cscK metrics. For any irreducible Mumford semistable bundle $E$ (not Mumford stable) of rank 2 over a curve of genus $g\geq 2$, there are integral classes on the ruled surface $X=\mathbb{P}(E)$ that are K-semistable and not K-polystable.
\end{corollary}
\begin{proof}
We fix a bundle as in the statement and apply Corollary \ref{cor1} to produce a sequence of an almost cscK metric. The existence of such metric implies in turn that $(\mathbb{P}(E),\mathcal{O}_{\mathbb{P}(E)}(1)))$ is K-semistable from \cite{D5}. On another hand, the automorphism group $Aut(\mathbb{P}(E))$ is actually trivial, see \cite{S}. Therefore if the sequence of almost cscK metric was convergent, it would converge towards a cscK metric and $(X,L)$ would be K-stable by \cite{D1,Stoppa1}. This would contradict Proposition \ref{prop}.
\end{proof}
Corollary \ref{cor} gives a positive answer to a conjecture of J. Stoppa \cite{Sto3}. Our construction can be easily modified to produce examples of K-semistable not K-stable manifolds in any dimension using vector bundles of higher rank over a curve (using an induction argument on their Harder-Narasimhan filtration, see \cite{Ke00}) or using Proposition \ref{prop3}. Note that Fano examples of K-semistable but not K-polystable threefolds have been found by G. Tian by considering small deformations of the Mukai-Umemura threefold.
\section{Almost balanced metric and Asymptotic Chow semistability}
\subsection{Definition of almost balanced metric}
Let us recall that one can define for $X$ a submanifold of $\mathbb{P}^N$ the center of mass of $X$
as
$$\mu(X)=\int_X \frac{zz^*}{\vert z\vert^2} d\mu_{FS} - \frac{\mathrm{Vol}(X)}{N+1}\mathrm{Id}\in \sqrt{-1}Lie(SU(N+1))$$
considering $\mathbb{P}^N$ as a co-adjoint orbit in the Lie algebra of $SU(N+1)$. The Chow weight of $X$ with respect to $A$, hermitian matrix,
is $$FCh(A,X)=\mathrm{tr}(\mu(X)\cdot A)=\int_X \frac{z^* A z}{\vert z \vert^2}d\mu_{FS} - \frac{\mathrm{Vol}(X)}{N+1}\mathrm{tr}(A)$$
and the definition can be extended to any algebraic cycles. It is a classical fact, based on Kempf-Ness theory, that $FCh(A,e^{tA}\cdot X)$ is an increasing function of $t$ and we refer \cite[Proposition 5]{D5} for details. In particular for $\overline{X}$ the limiting Chow cycle of $e^{tA}\cdot X$ as $t\rightarrow -\infty$, we get $$FCh(A,X)\geq FCh(A,\overline{X}).$$ This provides the inequality
\begin{equation}\Vert \mu(V)\Vert_{2} \,\Vert A\Vert_2 \geq -FCh(A,\overline{X}) \label{ineq3}
\end{equation}
where one has defined the norm $\Vert T\Vert_2^2= \sum \vert \lambda_i\vert^2$ for $\lambda_i$ eigenvalues of the hermitian matrix $T$, taking into account their multiplicities. This is the finite dimensional analogue of the main theorem of \cite{D5} that we used previously.
Let us now introduce a notion of almost-balanced metrics.
\begin{definition}\label{albal}
Given $(X,L)$ a projective manifold, we say that there exists a sequence of almost balanced metrics if for all $k>>0$ and all $\epsilon>0$,
there exists a hermitian metric $h_{k,\epsilon}$ on $L^k $ such that the Bergman function satisfies $$\Big\Vert\rho(h_{k,\epsilon}) - \frac{N_k+1}{\mathrm{Vol}_L(X)}\Big\Vert_{\mathrm{C}^0}\leq \epsilon.$$
\end{definition}
Note that $\rho(h_{k,\epsilon})=\sum_{i=1}^{N_k+1} \vert s_i\vert^2_{h_{k,\epsilon}}\in C^{\infty}(X,\mathbb{R}_+)$ for $N_k+1=h^0(L^k)$ and $\{s_i\}_{i=1,..,N_k+1}$ an orthonormal basis of $H^0(L^k)$ with respect to the $L^2$-inner product induced by $h_{k,\epsilon}$. The existence of an almost balanced metric for $(X,L)$ implies, using \eqref{ineq3} and Hilbert-Mumford criterion, that $(X,L)$ is asymptotically Chow stable since the Chow weight of the limiting Chow cycle along a test-configuration cannot be strictly negative (this appears also clearly in \cite[Equation (16)]{D5} where the lower order terms are the higher Chow weights associated to the one-parameter subgroup action).
\subsection{Construction of almost balanced metric}
Let us consider $\hat{h}_\epsilon$ a hermitian metric on $\mathcal{O}_{\mathbb{P}(E)}(1)$ with the same notations as in the previous section.
Then the Bergman function for $(\mathcal{O}_{\mathbb{P}(E)}(k),\hat{h}_\epsilon^k)$ has an asymptotic expansion
\begin{equation}\rho(\hat{h}_\epsilon^k)=k^{r}+ k^{r-1}\frac{\mathrm{scal}(\omega_{\epsilon})}{2}+k^{r-2}a_2+...+k^{r-q}a_q\label{asympt}\end{equation}
where $r$ is the rank of the bundle $E\rightarrow C$ and $\omega_{\epsilon}=c_1(\hat{h}_\epsilon)$.
The writing of \eqref{asympt} means the following inequality holds in $\mathrm{C}^0$-topology (it will be sufficient to work in that topology in the sequel)
\begin{equation}\Big\Vert \rho(\hat{h}_\epsilon^k)- \left( k^{r}+ k^{r-1}\frac{\mathrm{scal}(\omega_{\epsilon})}{2}+k^{r-2}a_2+...+k^{r-q}a_q\right) \Big\Vert_{\mathrm{C}^0}\leq C_q(\hat{h}_\epsilon)k^{n-q-1}.\label{asympt2}\end{equation}
The terms $a_i$ involve at most the $(2i-2)$-th first covariant derivatives of the curvature $\omega_{\epsilon}$.
\begin{lemma}\label{fixed}
If the metric $\omega_\epsilon$ is bounded from below and bounded in $\mathrm{C}^{2q}$ norm by a constant $\delta$ with some reference metric, then
the constant $C_q(\hat{h}_\epsilon)$ in Equation \eqref{asympt2} depends actually only on $q$ and the constant $\delta$.
\end{lemma}
This is well known, see for instance \cite[Proposition 6]{D1}. \\
We are now coming back to the setup of Section \ref{Sect1} and shall construct a sequence of almost-balanced metric. Since we work in the smooth category, it is not difficult to adapt the reasoning in order to obtain an approximate Hermitian-Einstein structure $h_{\epsilon}^\infty$ in the following sense
\begin{equation}\label{uniform2}\Vert F_{E,h_{\epsilon}^\infty}-\mu(E)\mathrm{Id}_E\,\omega\Vert_{\mathrm{C}^\infty}< \epsilon.\end{equation}
Furthermore we can assume that $h_E$ is real analytic. If $h_E$ is not real analytic we may use a slight generalization of Tian's result \cite{Ti1} of approximation of a positive hermitian metric by a sequence of Bergman type metrics in smooth topology (for a discussion on the smooth convergence see \cite{Ru}). Actually, we can pull-back the canonical metric on the universal bundle $U_{(2)} $ over the Grassmannian $Gr(2,H^0(B,E\otimes L^s))$ for $s>>1$, which provides a sequence of real analytic metrics. This sequence is convergent towards the metric $h_E$ in smooth topology thanks to the asymptotic result for the Bergman kernel of $E\otimes L^s$ that can be found in \cite{W2}.
\\
On $X=\mathbb{P}(E)$, the curvature $\omega^\infty_\epsilon$ of the associated real analytic metric $\hat{h}_\epsilon^\infty$ on $\mathcal{O}_{\mathbb{P}(E)}(1)$ is bounded in $\mathrm{C}^\infty$ norm and is positive. This can be seen by expressing the curvature of $\hat{h}_\epsilon^\infty$ in terms of the curvature terms of $h_\epsilon^\infty$, see Lemma \ref{tech}. We can apply Lemma \ref{fixed} and for $0<\epsilon<\epsilon_0$ we get a uniform expansion of the Bergman function of $\hat{h}_\epsilon^\infty$ with constant depending only on the maximum order of the expansion and $\epsilon_0$, and hence we denote the constant in Equation \eqref{asympt2} by $C_q(\epsilon_0)$.
\begin{lemma}\label{uniform3}
In the above setting, there is a constant $C_\infty>0$ depending only on $\epsilon_0$ such that $C_q(\epsilon_0)\leq C_\infty^q$. In other words, the growth of the error constant in \eqref{asympt2} when taking higher order expansion is at most exponential.
\end{lemma}
\begin{proof}
This is a consequence of the techniques used in the proof of \cite[Theorem 1.3]{L-L} (see also Theorem 1.2 of the associated announcement paper). We shall use the notations of the quoted paper. The Bergman function is given by the sum of an orthonormal basis of sections that can be taken as the union of a peaked section and vanishing sections at $x$ that have been orthonormalized. In order to orthonomalize these sections, it is necessary to inverse a matrix formed of the inner products $\langle S,S\rangle_{L^2}$ and $\langle S,T\rangle_{L^2}$ as it is done in \cite[Section 5]{L-L}.
Since we deal with analytic metrics, we can do the Taylor expansion of the involved metrics (on the line bundle and the volume form) and we get an expansion of the $L^2$ norm $\langle S,S\rangle_{L^2}$ of the peak section $S$ of the form $\langle S,S\rangle_{L^2}=\frac{1}{k^n}(1+\sum_{i\geq 1} \beta_i k^{-i})$. By convergence of this Taylor expansion, one has for a certain uniform constant $C>0$ that depends on the metric, $\vert \beta_i\vert \leq C^i$ for $i\geq 1$. In \cite[Theorem 4.1 (3)]{L-L}, it is proved a uniform bound on the $L^2$-inner product $\langle S,T\rangle_{L^2}$ between a holomorphic section $S$ peaked at the point $x$ and sections $T$ that vanish at order $p'>0$ at $x$. Together these two uniform controls provide the expected growth on the error term $C_q$ of the expansion of the Bergman function. In particular we have shown that in \eqref{asympt2}, one has the control $\vert a_i \vert <\frac{c_0}{\gamma^i}$ for a uniform constant $\gamma>0$.
\end{proof}
Furthermore, the terms $a_i$ enjoy the property of being polynomial expressions in the curvature and its covariant derivatives. For any integer $q$, we can find a metric $\hat{h}_{\epsilon,q}^\infty$ such that the term
$$\left( k^{r}+ k^{r-1}\frac{\mathrm{scal}(\omega_{\epsilon})}{2}+k^{r-2}a_2+...+k^{r-q}a_q\right),$$
which satisfies a similar property, is also constant up to an error term of the form $\epsilon/2$ in $\mathrm{C}^0$ norm while we are are still under the assumptions of Lemma \ref{fixed}. This is a consequence of the uniform control in $\mathrm{C}^\infty$ topology of the curvature of $\omega^\infty_\epsilon$ using \eqref{uniform2}. \\
Furthermore, using Lemma \ref{uniform3} and taking $k>C_\infty$, we can impose $$k^{r-q-1}C_{q}(\epsilon_0)\leq k^{r-1}\left(\frac{C'_\infty}{k}\right)^q<\epsilon/2$$ in Equation \eqref{asympt2} by choosing $q$ large enough. Hence, for $k>>0$, we have obtained a metric $\hat{h}_{k,\epsilon}=(\hat{h}_{\epsilon,q}^\infty)^k $ which is almost balanced in the sense of Definition \ref{albal}.\\
Eventually, with the examples considered in Section \ref{Sect1} and Proposition \ref{prop}, we have proved the following result.
\begin{theorem}\label{thm2}
There exist examples $(X,L)$ of smooth polarized manifolds such that $L$ is asymptotically Chow semistable and not K-polystable.
For instance, for any irreducible Mumford semistable bundle $E$ (not Mumford stable) of rank 2 over a curve of genus $g\geq 2$, any K\"ahler integral class on the ruled surface $X=\mathbb{P}(E)$ is asymptotically Chow semistable, not asymptotically Chow polystable and not K-polystable.
\end{theorem}
\medskip
Note that Chow semistability implies K-semistability, see \cite{Thnotes}. Therefore Theorem \ref{thm2} implies straightforward Corollary \ref{cor}.
We chose to present both results in order to stress how one could expect to find examples of K-semistable but non Chow semistable manifolds. Actually, one could try to find a sequence of almost cscK metrics in $\mathrm{C}^0$ norm such that some high order covariant derivatives of the Riemann curvature tensor are unbounded.
\section{Parabolic structures and conic K\"ahler metrics \label{cscK conic}}
\subsection{Construction of stable parabolic structure}
Given a vector bundle $E$ over a curve, we provide an elementary construction that allows to construct a stable parabolic structure on $E$.
\subsubsection{From an unstable to a semistable parabolic structure\label{sub1}}
Suppose that $E$ has rank 2 and is not semistable.
We have an exact sequence
$$
0\longrightarrow F_1\longrightarrow E\longrightarrow
F_2\longrightarrow 0
$$
such that\\
(i) $F_i$ are line bundles,\\
(ii) $\deg(F_1)>\mu(E)=\frac{\deg(F_1)+\deg(F_2)}{2}>\deg(F_2)$.\\
We set
\begin{equation}\label{A}
A:=\deg(F_1)-\mu(E)=\mu(E)-\deg(F_2)>0.
\end{equation}
We take a large integer $N$ such that
$2A/N<1$ and fix some distinct points $p_i\in X$ for $i=1,\ldots,N$.
We take subspaces
$V_i$ of $E_{|p_i}$ $(i=1,\ldots,N)$ of dimension 1 such that
$V_i\not\subset F_{1|p_i}$.\\
We define the parabolic filtration $\mathcal{F}_{j}(p_i)$ at $p_i$ by $$\mathcal{F}_{0}=E_{|p_i}\supset \mathcal{F}_{1}=V_i \supset \mathcal{F}_{2}=\{0\}$$
with weight $0$ for $\mathcal{F}_0$ and $2A/N$ for $\mathcal{F}_1$. The, the degree of the parabolic bundle $E_{\ast}$ is
$$
\mathrm{par}\deg(E_{\ast})=\deg(E)+\sum_{i=1}^N 2A/N=\deg(E)+2A$$
Hence, $\mathrm{par}\mu(E_{\ast})=\mu(E)+A$.
The degree of the induced parabolic bundle $F_{1\ast}$ is
$$
\mathrm{par}\deg(F_{1\ast})=\deg(F_1)=\mu(E)+A=\mathrm{par}\mu(E_{\ast})$$
If $F\subset E$ is a subsheaf of rank one such that $F\not\subset F_1$,
then there exists a non-trivial morphism $F\longrightarrow F_2$.
Hence, we have $\deg(F)\leq \deg(F_2)$ and thus, using (\ref{A}),
$$
\mathrm{par}\mu(F_{\ast})=\mathrm{par}\deg(F_{\ast})
\leq
\deg(F)+2A
\leq
\mu(E)+A
=\mathrm{par}\mu(E_{\ast})
$$
Hence, we have proved that $E_{\ast}$ is parabolic semistable.
\subsubsection{From a semistable to a stable parabolic structure}
Let $E_{\ast}$ be semistable (not stable) parabolic bundle of rank $2$ over
$(X,D)$, where $D$ is a finite subset of $X$. We wish to modify the parabolic structure
so that the new parabolic bundle is stable. One of the following holds:
\begin{enumerate}
\item $E_{\ast}\simeq F_{0\ast}\otimes \mathbb{C}^2$,
\item $E_{\ast}\simeq F_{1\ast}\oplus F_{2\ast}$,
\item there exists a non-split exact sequence
$$0\longrightarrow F_{1\ast}\longrightarrow
E_{\ast}\longrightarrow F_{2\ast}\longrightarrow 0,$$
where $F_{i\ast}$ are parabolic bundles of rank one.
\end{enumerate}
We choose $p'_1,p'_2,p'_3\in X\setminus D$
and subspaces $V_i\subset E_{|p'_i}$ of dimension 1
with the following property: if $F_{\ast}\subset E_{\ast}$
with $\mathrm{par}\deg(F_{\ast})=\mathrm{par}\mu(E_{\ast})$, then $F_{|p'_i}=V_i$ may happen
for at most one $i$.\\
Let us choose a real number $1>\epsilon>0$.
We consider the parabolic filtration $\mathcal{F}'_{j}(p'_i)$ at $p'_i$ by $$\mathcal{F}'_{0}=E_{|p'_i}\supset \mathcal{F}'_{1}=V_i \supset \mathcal{F}'_{2}=\{0\}$$
with weight $0$ for $\mathcal{F}'_0$ and weight $\epsilon$ for $\mathcal{F}'_1$.
We also consider the same parabolic structure
at the points of $D$.
We obtain a parabolic bundle $E^{(\epsilon)}_{\ast}$. We have
$$
\mathrm{par}\mu(E^{(\epsilon)}_{\ast})
=\mathrm{par}\mu(E_{\ast})+3\epsilon/2.
$$
For $F_{\ast}\subset E_{\ast}$
such that $\mathrm{par}\deg(F_{\ast})=\mathrm{par}\mu(F_{\ast})=\mathrm{par}\mu(E_{\ast})$,
we have
$$
\mathrm{par}\deg(F^{(\epsilon)}_{\ast})
\leq
\mathrm{par}\deg(F_{\ast})+\epsilon
<\mathrm{par}\mu(E^{(\epsilon)}_{\ast})
$$
If $\epsilon$ is sufficiently small,
then
$\mathrm{par}\mu(F^{(\epsilon)}_{\ast})
<\mathrm{par}\mu(E^{(\epsilon)}_{\ast})$
for any $F_{\ast}$ such that
$\mathrm{par}\mu(F_{\ast})<\mathrm{par}\mu(E_{\ast})$.
Finally, with the new induced parabolic structure, $E_{\ast}^{(\epsilon)}$ is parabolic stable over the rank 2 bundle $E$. We explain now how to deal with higher rank bundles.
\subsubsection{Higher rank and corollaries}
By induction on the rank of the bundle, we have a generalization of our reasoning on any vector bundle over a curve. Let $r\geq 2$ be the rank of $E$. Since the degrees of the subbundles of $E$ are bounded from above, let choose $F_2$ the maximal destabilizing subbundle of $E$. Then $F_1=E/F_2$ is a vector bundle and we have the exact sequence $$ 0 \rightarrow F_1 \rightarrow E \rightarrow F_2 \rightarrow 0.$$
Let $F$ a subbundle of $E$. If $F\subset F_1$, then by induction, we can find a stable parabolic stable structure on $F_1$ that we denote $(F_{1})_{\ast_1}$ and thus $\mathrm{par}\mu(F_{\ast_{_1}}) < \mathrm{par}\mu((F_{1})_{\ast_1})$. Let us assume that $A=\mathrm{par}\mu((F_1)_{\ast_1})- \mathrm{par}\mu(E_{\ast_1})$ is non negative. It remains to show that one can refine the considered parabolic structure $\ast_1$ so that $\mathrm{par}\mu(F_{\ast_{_2}}) < \mathrm{par}\mu((F_{1})_{\ast_2}) \leq \mathrm{par}\mu(E_{*_2})$ with respect to a new structure $\ast_2$. This is done as in the previous subsection by choosing an adapted filtration so that
\begin{align*}
\mathrm{par}\deg(E_{\ast_2}) &=\mathrm{par}\deg(E_{\ast_1}) + rA,\\
\mathrm{par}\deg((F_1)_{\ast_2}) &=\mathrm{par}\deg((F_1)_{\ast_1}), \\
\mathrm{par}\deg((F_2)_{\ast_2}) &\leq \mathrm{par}\deg((F_2)_{\ast_1})+rA.
\end{align*}
If $F\not\subset F_1$, then there is a non trivial morphism $F\longrightarrow F_2$ and eventually $\mathrm{par}\mu(F_{\ast_2})\leq \mathrm{par}\mu(E_{*_2})$. We get the parabolic semistability of $(E_{\ast_2})$. We apply the same reasoning as in the previous subsection to derive the following result.
\begin{theorem}\label{process}
Given $E$ a holomorphic vector bundle over a curve, one can find sufficiently many points $p_i$ and sufficiently small weights $\beta>0$ such that the associated parabolic structure $E_\ast$ is Mumford parabolic stable. If $E$ is Mumford semistable and has rank 2, it is sufficient to consider 3 points.
\end{theorem}
Theorem \ref{process} restricted to rational weights and the main result of \cite{Ro1} provide a new proof of an old result of C. Lebrun and M. Singer \cite[Theorem 3.11]{LS}. Note that the assumption on the genus is made to kill the non trivial holomorphic vector fields.
\begin{corollary}
Consider a ruled surface $S$ over a curve $C$ of genus $g\geq 2$. The blow up of $S$ at sufficiently many points admits a cscK metric.
\end{corollary}
Of course, a different proof can be given using the work of C. Arezzo- F. Pacard. The ruled surface is birationnally equivalent to the product $C\times \mathbb{P}^1$, which means that a blow-up of $S$ is also a blow-up of $C\times \mathbb{P}^1$, on which there exists a cscK metric. Then it is sufficient to apply the main result of \cite{AP1}. Our construction has the advantage to be more constructive.
\subsection{Construction of constant scalar curvature K\"ahler metric with conic singularities}
Let's start with $E$ a semistable bundle over a curve $C$. From the previous section (Theorem \ref{process}), we have obtained a stable parabolic structure $E_*$ along the points $\mathcal{P}=\{p_1,...,p_{m_\mathcal{P}}\}$, with $m_\mathcal{P}\geq 3$ and weight less than $\beta_0>0$. We shall see that there is an Hermitian-Yang-Mills metric with respect to a K\"ahler-Einstein metric with conic singularity along the associated points to the parabolic structure $E_*$.
Firstly, let us recall the notion of K\"ahler metric with conical singularity, focusing on the complex dimensional one case.
\begin{definition}
Let $C$ be a complex curve. Let $\mathcal{P}=\{p_j\}_{j=1,..,m_\mathcal{P}}\subset C$ be a finite set of points. Let $\beta=(\beta_1,..,\beta_{m_\mathcal{P}})$ with $0<\beta_i\leq 1$ be the cone angles. Given a point $p_i\in \mathcal{P}$, label a local chart $(V_{p_i},z_1)$ centered at $p_i$ as local cone chart. A K\"ahler metric with conical singularity and cone angle $2\pi \beta_i$ along $p_i$ (in short a conical K\"ahler metric) is a closed positive (1,1) current and a smooth K\"ahler metric on $C\setminus \mathcal{P}$ such that in a local cone chart $V_{p_i}$ its K\"ahler form is quasi-isometric to the cone flat metric
\begin{equation}\label{local}\omega_{cone}=\frac{\sqrt{-1}}{2}\beta_i^2 \vert z_1\vert^{2(\beta_i-1)} dz_1\wedge d\bar z_1.\end{equation}
\end{definition}
Over a complex curve, the notions of K\"ahler class and a pointwise conformal class are equivalent. We can apply the work of M. Troyanov \cite{Tro} and therefore fixing at any $m_\mathcal{P}$ points, with at least $m_\mathcal{P}\geq 3$, a curve $C$ admits a conical K\"ahler-Einstein metric $\omega_\beta$ along these fixed points for any angle between 0 and 1. Let us consider a model cone metric of the form $$\tilde{\omega}_\beta=\frac{\sqrt{-1}}{c_\beta}\sum_{i=1}^{m_\mathcal{P}} \partial\bar\partial \vert \sigma_i\vert^{2\beta}+\omega$$
where $\omega$ is a smooth K\"ahler form on $C$, $\beta<\beta_0$, the sections $\sigma_i$ vanish exactly at $p_i$ and $c_\beta>0$ is large enough so that $\tilde{\omega}_\beta$ is positive over the curve. A direct computation shows that actually $\tilde{\omega}_\beta$ satisfies the previous definition. Now, the behavior of conical K\"ahler-Einstein metrics are pretty well understood and it has been derived some Laplacian estimates for it. From \cite[Theorem 1]{Bre1} (see also \cite[Theorem A]{CGP}, \cite[Section 5.2]{GP} and \cite[Theorems 1 and 2]{Yao}), we know that the conical K\"ahler-Einstein metric is uniformly equivalent to the model cone metric for small enough angle, i.e there exists a certain constant $\delta>0$ such that
\begin{equation}
\frac{1}{\delta}\tilde{\omega}_\beta < \omega_\beta < \delta\tilde{\omega}_\beta.
\label{unif}
\end{equation}
We shall explain now how to apply Simpson's theory \cite{Si1} in our context. Firstly, it is well known that a conical K\"ahler metric has finite volume. Moreover there is an exhaustion function $\varphi$ of $C$ such that $\Delta_{\omega_\beta}\varphi$ is bounded, and Sobolev inequality holds with respect to $\omega_\beta$. Both facts are checked in \cite[Proposition 4.1]{Li1} where the arguments use only the local expression of the metric \eqref{local} and thus are still valid for $\omega_\beta$.
Moreover, in dimension one, analytic stability and parabolic stability coincide. In \cite[Section 3]{Li1}, it is constructed a metric $K_0$ on $E_*$ such that its curvature satisfies
$\vert \Lambda_{\tilde{\omega}_\beta} F_{K_0}\vert_{K_0}$ is bounded on $C$ for small enough angle $\beta>0$. Therefore $\vert \Lambda_{{\omega}_\beta} F_{K_0}\vert_{K_0}$ is still bounded using \eqref{unif}. By stability of the parabolic structure $E_*$, Simpson's theorem \cite[Theorem 1]{Si1} imply the existence of a Hermitian-Yang-Mills metric $H_E$ on $E$ satisfying the Hermitian-Einstein equation \begin{equation}F_{H_E}=\frac{\mathrm{tr} F_{H_E}}{r}\mathrm{Id}_E\,\omega_\beta={\textrm{par}\mu(E)}\mathrm{Id}_E\, \omega_\beta\label{HEeqn}
\end{equation}
over $C\setminus \mathcal{P}$, with $r$ the rank of $E$.
Let us now consider the ruled manifold $X=\mathbb{P}E$ and $\pi:X\rightarrow C$ the projection on the base manifold. We denote $D\subset \mathbb{P}(E)$ the induced divisor from the preimage of $\mathcal{P}$.
The metric $H_E$ on $E$ induce a metric $\hat{H}_E$ on $\mathcal{O}_{X}(1)$. Actually for any point $p\in C$, $a,b\in E_{\vert p}$, and $\gamma\in E_{\vert p}^*$, one can define locally the metric $\hat{H}_E$ by
$\hat{H}_E(\hat u, \hat v)= \frac{\gamma(u)\overline{\gamma(v)}}{\Vert \gamma\Vert^2_{H_E} }$. \\
The curvature of $\hat{H}_E$ on $X$ is denoted $\hat{\omega}_E$. Since its restriction to the fibre is the Fubini-Study, it is non degenerate on $\mathbb{P}E_{\vert p}$ for $p\in C\setminus \mathcal{P}$. From Equation
\eqref{HEeqn}, it satisfies $$\hat{\omega}_E={\textrm{par}\mu(E)} \pi^*\omega_\beta,$$ see \cite[Section 1]{Fuj}. Consequently, for any $m$ large enough and from the properties of $\omega_\beta$, the metric $\hat{\omega}_E+m\pi^* \omega_\beta$ is a closed positive current, a K\"ahler metric with conical singularities along $D$ and has constant scalar curvature on $X \setminus D$.
\begin{theorem}\label{thm2b}
Given $E$ a vector bundle over a curve $C$, there exists a smooth divisor $D\subset \mathbb{P}(E)$ and a constant scalar curvature K\"ahler on $\mathbb{P}(E)$ with conical singularity along $D$. If $E$ is semistable and has rank 2, $D$ can be given by the preimage of 3 points of $C$.
\end{theorem}
We address now some remarks. By changing $m$, this theorem provides in particular examples of cscK cone metric in the class of $[\hat{\omega}_{E}+m\pi^*\omega_\beta]$ which are not K\"ahler-Einstein. In \cite{RoSi}, it is explained in details how to obtain scalar-flat K\"ahler metrics with orbifold singularities when one restricts to the particular case of a ruled surface and the parabolic structure has rational weights, using Mehta-Seshadri theorem. Also, remark that from the point of view of extremal K\"ahler metrics, some conical metrics are constructed in \cite{Lih1} using the formalism developed by Apostolov et al in \cite{ACGT2}, while G. Sz\'ekelyhidi studied the geometric splitting of a ruled surface given by a non stable manifold under the Calabi flow in \cite{Sz1}. From the point of view of stability, we expect that Theorem \ref{thm2b} provides examples of log-K-stable manifolds in the sense of \cite{D8}.
\bigskip
\noindent {\bf Acknowledgments.} We are very grateful to Vestislav Apostolov, Takuro Mochizuki, Yann Rollin, Xiaowei Wang and Kai Zheng for useful conversations.
This research was partially supported by the French Agence Nationale de la Recherche - ANR project EMARKS.
{\small
|
\section{Introduction}\label{sec-intro}
The Jaynes-Cummings model (JCM) which is extensively used to perform a full quantum mechanical description of the atom-field interaction \cite{JCM1,JCM2}, gives a pattern to solve the interaction between the single-mode quantized electromagnetic field and the two-level atom in the rotating wave approximation (RWA).
Relative facility of this model, in addition to appearing some non-trivial phenomena and remarkable properties such as Rabi oscillations of population inversion (which indicate the energy exchanges between field and matter), the collapses and revivals of these oscillations \cite{shore} etc gets the model in the core of this field of researches. Moreover, experimental observation of the collapse and subsequent revival, Rabi oscillations and a single-mode two-photon maser has been reported \cite{walther}.
To expand and modify the JCM, many generalizations have been proposed in recent decades. Intensity-dependent (nonlinear) JCM was suggested in \cite{suk1,suk2} where the dependence of atom-field coupling on the light intensity is described, and then has been used by others \cite{selective1,selective2,selective3}.
Dynamical behavior of the JCM beyond the RWA has been studied in \cite{naderi} in which the effects of the counter-rotating terms on various dynamical properties such as the atomic population inversion, photon counting statistics, quantum phase properties of the cavity field etc have been discussed. Quantum properties of a $\Lambda$-type three-level atom interacting with a single-mode field in a Kerr medium with intensity-dependent coupling and in the presence of the detuning parameters have been studied by us \cite{us}.
The ability of the nonlinear JCM in generating a class of $SU(1,1)$ coherent states of the Gilmore-Perelomov type and also $SU(2)$ group was shown by one of us \cite{miry1}. In addition, as a results of a system in which a two-level atom interacts alternatively with a dispersive quantized cavity field and a resonant classical field, it is recently proposed a theoretical scheme from which the nonlinear elliptical states can be generated \cite{miry2}. In particular and in direct relation to the present work, more recently nonlinear interaction between a three-level atom (in a $\Lambda$ configuration) and a two-mode cavity field in the presence of a cross-Kerr medium and its deformed counterpart \cite{newhonarasa}, intensity-dependent atom-field coupling and the detuning parameters has been discussed by us \cite{usJOSA}. Briefly, we studied the effects of these latter parameters on a few physical properties of the obtained state vector of the entire quantum system. Specifically, using the von Neumann approach, we firstly obtained the time evolution of the field entropy by which the amount of entanglement between subsystems has been determined. In addition, the position-momentum entropic uncertainty relation is evaluated, from which the entropy squeezing in position/momentum component is investigated. \\
Now, due to the fact that, nonclassical states have received considerable attention in various fields of research, such as quantum optics, quantum cryptography and quantum communication \cite{application1,application2,application3}, meanwhile in the continuation of our recent work in \cite{usJOSA} and for completing it,
the main goal of the present paper is to investigate individually and simultaneously the effects of intensity-dependent coupling, deformed Kerr medium and detuning parameters on some physical properties, consist of {\it number-phase entropic squeezing}, {\it sub-Poissonian statistics}, {\it Cauchy-Schwartz inequality} (CSI) and two different types of {\it squeezing} of the state vector of the atom-field system under our consideration. As will be observed, the considered parameters in our model allow one to tune the depth and domain of the mentioned nonclassicality features (the aspects with no classical analogue) with respect to time. It is also noteworthy to declare that, as is well-known, there is a logical link between these nonclassicality features and Glauber-Sudarshan $P$-distribution function \cite{GSPfunction1,GSPfunction2,GSPfunction3}. By this we mean that the occurrence of nonclassicality via each of the above indicators leads to the non-positivity of the corresponding $P$-function. But, since deriving this function is usually a hard task (if not impossible), this paper is allocated to study the above nonclassicality indicators which are really independent of each other and of the previously considered properties in \cite{usJOSA}, too.
To make our motivations of this presentation more clear, we give a few words on the notability of the considered physical (nonclassical) criteria. Starting with the number-phase entropic uncertainty relation, it should be mentioned that, studying the quantum phase properties leads to the investigating the variation of the phase relation between the photons of the field. Moreover, it has been shown that for the single-mode JCM the evolution of the phase variance, as well as the phase distribution can carry certain information on the collapse-revival phenomenon of the corresponding atomic inversion \cite{dung}. Altogether, to achieve this purpose, it ought to be mentioned that, for discussing the dynamical behaviour of the phase properties, we will suitably extend the Pegg-Barnett formalism \cite{pb} to be utilizable for our considered model. Regarding CSI, it may be noted that, the violation of CSI may be generally led to the violation of Bell inequalities \cite{walls}. Also, it is shown that, the violation of CSI can be observed in a two-photon interference experiment with a light source in a pair coherent state \cite{ghosh}. Moreover, concerning the importance of the squeezing phenomenon, it is worthwhile to illustrate that two-mode squeezed state can be regarded as the output state of an ideal two-photon device \cite{caves}. In addition, it is reported that, in particular there exists a close link between sum squeezing phenomenon and sum-frequency generation (combination of two photons with different frequencies $\omega_{a}$ and $\omega_{b}$, and generation of a photon with frequency $\omega_{c}=\omega_{a}+\omega_{b}$, by the presence of a nonlinear medium and through the second-order susceptibility $\chi^{(2)}$ \cite{boyd}) that has been explored by Hillery \cite{hillery}. All above explanations which attracted a great deal of attention in recent researches motivate one to explore the considered nonclassicality features for the complicated atom-field interaction system which has been analytically solved by us \cite{usJOSA}. \\
The paper is organized as follows. In the next section, a brief review of obtaining the state vector of the atom-field system will be given, which has been extensively explained in \cite{usJOSA}. In section \ref{phyproperties}, after investigating the number-phase entropic uncertainty relation by considering the two-mode Pegg-Barnett formalism, from which the entropy squeezing can be investigated, some of the nonclassical criteria of the obtained states such as Mandel $Q$ parameter, CSI and two different types of squeezing parameter are studied. Finally, section \ref{conclusion} contains a summary and concluding remarks.
\section{A brief review on the model}
Let us give a brief review on the model which has been extensively discussed by us in \cite{usJOSA}, in which a three-level atom (in a $\Lambda$ configuration) interacts with a two-mode quantized electromagnetic field which oscillates with frequencies $\Omega_{1}$ and $\Omega_{2}$ in an optical cavity surrounded by a Kerr medium with intensity-dependent coupling. According to the generalized JCM, the Hamiltonian for this system in the RWA can be written as $\hat{H} =\hat{H}_{0}+\hat{H}_{1}$ where ($\hbar=c=1$)
\begin{eqnarray}\label{hamiltonih0}
\hat{H}_{0}&=& \sum_{j=1}^{3} \omega_{j}\hat{\sigma}_{jj}+\sum_{j=1}^{2} \Omega_{j} \hat{a}^{\dag}_{j} \hat{a}_{j},
\end{eqnarray}
and
\begin{eqnarray}\label{hamiltonih1}
\hat{H}_{1} &=& \chi \hat{\R}_{1}^{\dag} \hat{\R}_{1} \hat{\R}_{2}^{\dag} \hat{\R}_{2}+\lambda_{1}(\hat{\A}_{1} \hat{\sigma}_{12}+\hat{\sigma}_{21}\hat{\A}_{1}^{\dag}) \nonumber \\
&+& \lambda_{2}(\hat{\A}_{2}\hat{\sigma}_{13}+\hat{\sigma}_{31}\hat{\A}_{2}^{\dag}),
\end{eqnarray}
where $\hat{\sigma}_{ij}$ is the atomic lowering and raising operator between $|i\rangle$ and $|j\rangle$ defined by $\hat{\sigma}_{ij}=|i\rangle \langle j|,(i,j=1,2,3),\hat{a}_{j}$ ($\hat{a}_{j}^{\dag}$) is the bosonic annihilation (creation) operator of the field mode $j$, $\chi $ denotes the third-order susceptibility of cross-Kerr medium and the constants $\lambda_{1},\lambda_{2}$ determine the strength of the atom-field couplings. In above relation, $\hat{\R}_{j}=\hat{a}_{j}\hat{g}_{j}(\hat{n}_{j})$ and $\hat{\A}_{j}=\hat{a}_{j}\hat{f}_{j}(\hat{n}_{j})$ with $\hat{\R}_{j}^{\dag}$ and $\hat{\A}_{j}^{\dag}$ as their respective conjugate hermitians, where $\hat{n}_{j}=\hat{a}_{j}^{\dag}\hat{a}_{j}$. Note that, $g_{j}(\hat{n}_{j})$ and $f_{j}(\hat{n}_{j})$ are two generally different operator-valued functions which depend on the light intensity and specify the deformation of the Kerr medium and intensity-dependent atom-field coupling, respectively. \\
The wave function $|\psi(t)\rangle$ corresponding to the whole system may be proposed in the form
\begin{eqnarray}\label{say}
|\psi(t)\rangle&=&\sum_{n_{1}=0}^{+\infty}\sum_{n_{2}=0}^{+\infty}q_{n_{1}}q_{n_{2}}\Big[ A(n_{1},n_{2},t) e^{-i\gamma_{1} t}|1,n_{1},n_{2}\rangle \nonumber \\
\hspace{-2cm}&+& B(n_{1}+1,n_{2},t)e^{-i\gamma_{2} t}|2,n_{1}+1,n_{2}\rangle \nonumber \\
\hspace{-2cm}&+& C(n_{1},n_{2}+1,t)e^{-i\gamma_{3} t}|3,n_{1},n_{2}+1\rangle \Big],
\end{eqnarray}
where $q_{n_{1}}$ and $q_{n_{2}}$ describe the amplitudes of the initial states of the field associated to each mode and $A$, $B$ and $C$ are the time-dependent atomic probability amplitudes which have to be evaluated. Also we have set
\begin{eqnarray}
\gamma_{1}&\doteq&\omega_{1}+n_{1}\Omega_{1}+n_{2}\Omega_{2}, \nonumber \\
\gamma_{2}&\doteq&\omega_{2}+(n_{1}+1)\Omega_{1}+n_{2}\Omega_{2}, \nonumber \\
\gamma_{3}&\doteq&\omega_{3}+n_{1}\Omega_{1}+(n_{2}+1)\Omega_{2}.
\end{eqnarray}
By applying the probability amplitudes method, and after some lengthy but straightforward manipulations we obtained the probability amplitudes $A$, $B$ and $C$ (which determine the explicit form of the wave function of whole system) in the following form \cite{usJOSA}:
\begin{eqnarray}\label{abc}
A(n_{1},n_{2},t)&=&-e^{-i\Delta_{2}t}\sum_{j=1}^{3}(\mu_{j}+V_{B})b_{j}e^{i\mu_{j}t}, \nonumber \\
B(n_{1}+1,n_{2},t)&=&\sum_{j=1}^{3}\kappa_{1}\;b_{j} e^{i\mu_{j} t}, \nonumber \\
C(n_{1},n_{2}+1,t)&=&\frac{e^{i(\Delta_{3}-\Delta_{2})t}}{\kappa_{2}}\sum_{j=1}^{3}\Big[(\mu_{j}+V_{B}) \nonumber \\
&\times& (\mu_{j}+V_{A} - \Delta_{2}) - \kappa_{1}^{2}\Big]b_{j}e^{i\mu_{j}t},
\end{eqnarray}
where
\begin{eqnarray}\label{vkardan}
\mu_{j}&=&-\frac{1}{3}x_{1}+\frac{2}{3}\sqrt{x_{1}^{2}-3x_{2}}\cos\left[ \theta+\frac{2}{3}(j-1)\pi \right], \;j=1,2,3, \nonumber \\
\theta &=& \frac{1}{3}\cos^{-1}\left[ \frac{9x_{1}x_{2}-2x_{1}^{3}-27x_{3}}{2(x_{1}^{2}-3x_{2})^{3/2}}\right],
\end{eqnarray}
with
\begin{eqnarray}\label{x123}
x_{1}&\dot{=}&V_{A}+V_{B}+V_{C}+\Delta_{3}-2\Delta_{2}, \nonumber \\
x_{2}&\dot{=}& (V_{A}+V_{B}-\Delta_{2})(V_{C}+\Delta_{3}-\Delta_{2})+V_{B}(V_{A}-\Delta_{2}) \nonumber \\
&-& \kappa_{1}^{2} - \kappa_{2}^{2}, \nonumber \\
x_{3}&\dot{=}&V_{B}\left[(V_{A}-\Delta_{2})(V_{C}+\Delta_{3}-\Delta_{2})-\kappa_{2}^{2} \right] \nonumber \\
&-& \kappa_{1}^{2}(V_{C}+\Delta_{3}-\Delta_{2}).
\end{eqnarray}
Also, in the above relations we have defined
\begin{eqnarray}\label{vdefinition1}
\hspace{-0.5cm}V_{A} &\dot{=}&V(n_{1},n_{2}), \;\; V_{B}\dot{=}V(n_{1}+1,n_{2}), \;\; V_{C}\dot{=}V(n_{1},n_{2}+1), \nonumber \\
\hspace{-0.5cm}\kappa_{1}&\dot{=}&\lambda_{1}\;\sqrt{n_{1}+1}\;f_{1}(n_{1}+1), \;\; \kappa_{2}\dot{=}\lambda_{2}\;\sqrt{n_{2}+1}\;f_{2}(n_{2}+1), \nonumber \\
\hspace{-0.5cm} \Delta_{2} &=& \omega_{2}-\omega_{1}+\Omega_{1}, \;\;\; \Delta_{3}=\omega_{3}-\omega_{1}+\Omega_{2},
\end{eqnarray}
with
\begin{eqnarray}\label{vdefinition1}
V(n_{1},n_{2}) \dot{=} \chi\; n_{1} n_{2} g_{1}^{2}(n_{1})g_{2}^{2}(n_{2}).
\end{eqnarray}
Finally, by preparing the atom initially in the excited state, i.e. $A(0)=1$, $B(0)=C(0)=0$ or equivalently
\begin{eqnarray}\label{sayi}
|\psi(0)\rangle_{\mathrm{A-F}}=|1\rangle \otimes \sum_{n_{1}=0}^{+\infty}\sum_{n_{2}=0}^{+\infty}q_{n_{1}}q_{n_{2}} |n_{1}, n_{2} \rangle,
\end{eqnarray}
the following relation for $b_{j}$ may be found
\begin{eqnarray}\label{b123}
\hspace{-1cm} b_{j}=\frac{\mu_{k}+\mu_{l}+V_{A}+V_{B}-\Delta_{2}}{\mu _{jk} \mu _{jl}}, \; j\neq k\neq l=1,2,3,
\end{eqnarray}
where $ \mu _{jk}=\mu_{j}-\mu_{k} $. Hence, the wave function $|\psi(t)\rangle$ as given in (\ref{say}) is exactly derived.
It is valuable to state that the above formalism can be used for any physical system with arbitrary nonlinearity function. In this paper, we use the nonlinearity function $f(n) = 1/\sqrt{n} = g(n)$ which has been derived by Man'ko {\it et al} \cite{manko2} from the coherent states that have been named by Sudarshan as harmonious states \cite{harmonious}. This function is a popular nonlinearity function which has been usually used in the contents of deformation of bosonic operators in quantum optics literature \cite{example-harmonious1,example-harmonious2,example-harmonious3}.
\section{Investigating the physical properties of the introduced state}\label{phyproperties}
As we pointed out in the Introduction of the paper, in this section, we intend to examine the further physical properties with emphasising on some nonclassical features of the obtained state (in previous section) which are of special interest in the field of quantum optics and quantum information processing. To achieve this purpose, we check number-phase entropic squeezing, sub-Poissonian statistics, CSI and two different squeezing criteria, in detail. We will observe, the state has potential ability to show high nonclassicality behaviour in appropriate intervals of time. In this way, it may be noted that the unfinished work in \cite{usJOSA}, in our opinion, will be truly completed.
\subsection{Number-phase entropic uncertainty relation and entropy squeezing}\label{phase}
It is well-known that the investigation of the dynamical behavior of phase distribution of the field photons is in fact an analysis of the variation of the phase relation between field photons and atoms \cite{nakano}. Phase distribution and squeezing in number/phase operator of various physical systems with known discrete spectrum $e_{n}$ have been reported by one of us in \cite{honar}. Also, the number-phase entropic uncertainty relation and the number-phase Wigner function of generalized coherent states associated with a few solvable quantum systems that have non-degenerate spectra are examined \cite{honarasaphase}. Here, we intend to investigate the {\it entropy squeezing} for considered bipartite system in terms of entropies of {\it number} and {\it phase} operators, in which the phase properties of a photon field are evaluated by using the Pegg-Barnett approach \cite{pb}. According to this method, all observables corresponding to the phase properties are defined in an $(s+1)$-dimensional space, in which they constitute the eigenvalue equations with $(s+1)$ eigenstates (orthonormal phase states). Based on the Pegg-Barnett formalism, a complete set of $(s +1)$ orthonormal phase states (of a single-mode field) is defined by
\begin{eqnarray}\label{pb1mode}
|\theta_{p} \rangle = \frac{1}{\sqrt{s+1}}\sum_{n=0}^{s}\exp(in\theta_{p})| n \rangle,
\end{eqnarray}
where $\{| n \rangle\}_{n=0}^{s}$ are the number states and $\theta_{p}$ gets the following values:
\begin{eqnarray}\label{thetap}
\theta_{p}=\theta_{0}+\frac{2\pi p}{s+1},\;\;\;\;p=0,1,...,s,
\end{eqnarray}
with $\theta_{0}$ as an arbitrary value. It is useful to state that in calculating the required expectation values, the parameter $s$ tends finally to infinity. Moreover, since we are dealing with the two-mode field in the present atom-field interaction, it is necessary to generalize the orthonormal phase states. Thus, we propose the two-mode phase state $|\theta_{p},\theta_{q} \rangle$ which should be necessarily introduced as
\begin{eqnarray}\label{pq2mode}
\hspace{-1cm} |\theta_{p},\theta_{q} \rangle = \frac{1}{s+1}\sum_{n=0}^{s}\sum_{m=0}^{s} \exp\left(i n \theta_{p}\right) \exp\left( i m \theta_{q}\right) | n,m \rangle,
\end{eqnarray}
with
\begin{eqnarray}\label{thetapq}
\theta_{k}=\theta_{0}+\frac{2\pi k}{s+1},\;\;\;\;k=p,q,
\end{eqnarray}
and $\theta_{0}$ is an arbitrary value. In the same manner which stated in (\ref{thetap}), each of the indices, $p$ and $q$ can get the values $0,1,...,s$. By the way, we are able to define the two-mode Pegg-Barnett phase distribution function as follows
\begin{eqnarray}\label{pthetapq}
\mathcal{P}_{\theta}(\theta_{p},\theta_{q})=\lim_{s\rightarrow +\infty} \left( \frac{s+1}{2\pi} \right)^{2} \langle \theta_{p},\theta_{q}|\hat{\rho}_{F}|\theta_{p},\theta_{q} \rangle,
\end{eqnarray}
By putting the relation (\ref{pq2mode}) into (\ref{pthetapq}) and paying attention to the state vector of the whole system proposed in (\ref{say}), which all of its time-dependent coefficients have been explicitly obtained, one may obtain the phase distribution function associated with the two-mode cavity field as
\begin{eqnarray}\label{pthetapqfinal}
\hspace{-2cm}\mathcal{P}_{\theta}(\theta_{1},\theta_{2}) &=& \frac{1}{4 \pi^{2}} \left| \sum_{n_{1}=0}^{+\infty}\sum_{n_{2}=0}^{+\infty} q_{n_{1}} q_{n_{2}} A(n_{1},n_{2},t) \exp(- i n_{1} \theta_{1}) \exp(- i n_{2} \theta_{2}) \right|^{2} \nonumber \\
\hspace{-2cm}&+& \frac{1}{4 \pi^{2}} \left| \sum_{n_{1}=0}^{+\infty}\sum_{n_{2}=0}^{+\infty} q_{n_{1}} q_{n_{2}} B(n_{1}+1,n_{2},t) \exp(- i n_{1} \theta_{1}) \exp(- i n_{2} \theta_{2}) \right|^{2} \nonumber \\
\hspace{-2cm}&+& \frac{1}{4 \pi^{2}} \left| \sum_{n_{1}=0}^{+\infty}\sum_{n_{2}=0}^{+\infty} q_{n_{1}} q_{n_{2}} C(n_{1},n_{2}+1,t) \exp(- i n_{1} \theta_{1}) \exp(- i n_{2} \theta_{2}) \right|^{2}.
\end{eqnarray}
Now, based on the Shannon's idea which is related to classical information theory and following the path of \cite{pb}, we may define the (Shannon) entropies associated with the number and phase probability distribution by the following relations:
\begin{eqnarray}\label{shannonent}
\hspace{-1cm}R_{n}(t) &=& - \sum_{n_{1}=0}^{+\infty} \sum_{n_{2}=0}^{+\infty} \mathcal{P}_{n}(n_{1},n_{2}) \ln \mathcal{P}_{n}(n_{1},n_{2}), \nonumber \\
\hspace{-1cm}R_{\theta}(t) &=& - \int \limits_{\theta_{0}}^{\theta_{0}+2\pi} d\theta_{1} \int \limits_{\theta_{0}}^{\theta_{0}+2\pi} d\theta_{2} \; \mathcal{P}_{\theta}(\theta_{1},\theta_{2}) \ln \mathcal{P}_{\theta}(\theta_{1},\theta_{2}),
\end{eqnarray}
where $\mathcal{P}_{n}(n_{1},n_{2})=\langle n_{1},n_{2}|\hat{\rho}_{F}|n_{1},n_{2} \rangle$ and $\hat{\rho}_{F}=\mathrm{Tr}_{\mathrm{Atom}} \left( |\psi(t) \rangle \langle \psi(t)| \right)$. The sum of the entropic uncertainty relations for the number and phase in (\ref{shannonent}) can specify the lower bound for entropy which is given by $R_{n}+R_{\theta} \geq \ln 2\pi$. Considering this inequality, we suggest two quantities as follows:
\begin{eqnarray}\label{entsq}
S_{n}(t)&=&\frac{1}{\sqrt{2\pi}} \exp(R_{n}(t))-1, \nonumber \\
S_{\theta}(t)&=&\frac{1}{\sqrt{2\pi}} \exp(R_{\theta}(t))-1.
\end{eqnarray}
These quantities indicate that when $-1<S_{n}(t)<0$ ($-1<S_{\theta}(t)<0$), the number (phase) component of the field entropy is squeezed. It is worthwhile to declare that the negativity of $S_{n(\theta)}(t)$ and in other words, entropy squeezing is another expression of the fact that $R_{n(\theta)}(t)$ is below its minimum value.\\
Figure \ref{quant.ph-sq} shows the time evolution of the entropy squeezing in phase against the scaled time $\tau$ for initial mean number of photons fixed at $|\alpha_{1}|^{2} = 10=|\alpha_{2}|^{2}$. The upside plots concern with the absence of the intensity-dependent coupling, i.e. $f_{i} (n_{i}) = 1$ and the downside plots correspond to the intensity-dependent coupling regime by using the nonlinearity function $f_{i} (n_{i})=1/\sqrt{n_{i}}$.
In figure \ref{quant.ph-sq}(a) Kerr effect is absent ($\chi = 0$) and the exact resonant case is assumed ($\Delta_{2} = \Delta_{3} =0$).
Figure \ref{quant.ph-sq}(b) shows the effect of deformed Kerr medium ($\chi = 0.4 \lambda$), by considering the nonlinearity function $ g_{i}(n_{i}) = 1/\sqrt{n_{i}}$, in the absence of the detuning parameters. The influence of deformed Kerr medium in the presence of detuning parameters has been depicted in figures \ref{quant.ph-sq}(c).
From the upside plot of figure \ref{quant.ph-sq}(a) which corresponds to the constant coupling regime, no Kerr medium and in the resonance condition, it is observed that the state of the system possesses the entropy squeezing until it becomes positive after nearly $\tau=35$. By entering the effect of intensity-dependent coupling (downside plot of figure \ref{quant.ph-sq}(a)), remarkable constant amount of entropy squeezing ($\simeq -0.81954$) is appeared at all times. Considering both upside plots of figures \ref{quant.ph-sq}(a) and \ref{quant.ph-sq}(b) shows that, they are nearly the same, both qualitatively and quantitatively. Looking at the downside plot of figure \ref{quant.ph-sq}(b) indicates that, deformed Kerr medium can strengthen the entropy squeezing. In this case, the exact periodicity of oscillations is clearly observed. In detail, while upside plots of figures \ref{quant.ph-sq}(a) and \ref{quant.ph-sq}(b) are nearly the same, entering simultaneously the effects of nonlinearities in downside plot of figure \ref{quant.ph-sq}(b) results in an increase in the amount of the negativity of entropy squeezing in contrast with the downside plot of figure \ref{quant.ph-sq}(a). Figure \ref{quant.ph-sq}(c) shows the influences of the deformed Kerr medium and detuning parameters. It is observed that, in the presence of the detuning parameters with $f_{i}(n_{i}) = 1$ (the upside plot of figure \ref{quant.ph-sq}(c)), deformed Kerr medium generally increases the depth and time interval of the negativity of the phase entropic squeezing parameter. This is while in the intensity-dependent coupling regime, the appearance of entropy squeezing at all times is observed with fast oscillatory behavior (the downside one). \\
So, it may be inferred that, intensity-dependent coupling has a direct role in exhibiting the entropy squeezing in phase component at all times and strengthen this particular nonclassicality feature (see the downside plots). Also, deformed Kerr nonlinearity can improve this nonclassicality sign and detuning may enrich the negativity of this criterion (especially in the constant coupling). In addition, fast oscillatory behaviour of this quantity can be regarded as a consequence of the presence of the detuning parameters (in the intensity-dependent coupling regime).
\subsection{Sub-Poissonian statistics}\label{mandelparameter}
In order to determine photon statistics of the field we use the Mandel's $Q$ parameter which is defined by \cite{mandel}:
\begin{eqnarray}\label{mandel1}
Q = \frac{\langle (\Delta \hat{n})^2\rangle - \langle \hat{n} \rangle}{\langle \hat{n} \rangle},
\end{eqnarray}
where $\langle (\Delta \hat{n})^2\rangle=\langle \hat{n}^{2} \rangle-\langle \hat{n} \rangle^{2}$ and $\hat{n}=\hat{a}^{\dag}\hat{a}$. This parameter states that whenever $-1\leq Q <0 \;(Q>0)$ the statistics is sub-Poissonian (super-Poissonian) and $Q=0$ indicates the Poissonian statistics. By the way, the state vector of the system shows the nonclassical behavior when the photons statistics of field is sub-Poissonian. \\
Our presented results in figure \ref{mandel} show the time evolution of Mandel parameter versus the scaled time $\tau$, in which photon statistics of the cavity field (mode $1$) has been studied. In plotting this figure, the upside (downside) plots again correspond to the case $f_{i} (n_{i}) = 1$ ($f _{i}(n_{i}) = 1/\sqrt{n_{i}}$), also all parameters have been chosen like figure \ref{quant.ph-sq}.
Figure \ref{mandel}(a) is depicted for the case in which intensity-dependent coupling, deformed Kerr and detuning effects are all neglected, i.e., $(f_i(n_i)=1, \chi = 0 = \Delta_{2} = \Delta_{3})$. It is observed from the upside plot that the Mandel parameter varies between negative and positive values with collapse and revival behaviour. By entering the intensity-dependent coupling, it is seen that this parameter oscillates in negative region for all times. In figure \ref{mandel}(b), we have focused on the study of the effect of deformed Kerr medium on the time evolution of Mandel parameter. Comparing the upside plots of figures \ref{mandel}(a) and \ref{mandel}(b) indicates that the effect of deformed Kerr medium may enhance the negativity of Mandel parameter. Also, from the downside ones, the oscillatory behavior of the Mandel parameter is clearly observed, with increasing the amount of negativity of this quantity.
The effect of detuning parameters in the presence of deformed Kerr medium has been shown in figure \ref{mandel}(c).
According to the upside plots of figures \ref{mandel}(b) and \ref{mandel}(c), it is apparent that entering the detuning parameters in the presence of deformed Kerr medium reduces the average time intervals of the negativity of Mandel parameter (nonclassicality behaviour). While paying attention to the downside ones shows that due to entering the detuning, Mandel parameter will be always negative although the minima values of this nonclassicality sign is diminished. \\
From the numerical results depicted in figure \ref{mandel}, one can deduce that intensity-dependent coupling, as well as the deformed Kerr medium possesses a direct role in disappearing the classicality feature and so improving this nonclassicality indicator. Also, it is found that the depth of negativity of Mandel parameter in the presence of the detuning parameters may be decreased when intensity-dependent coupling is considered. At last, in the constant coupling regime, the typical collapse and revival exhibition as a nonclassical phenomenon can be observed.
\subsection{The Cauchy-Schwartz inequality}\label{CSInequality}
In this section we pay attention to CSI for checking another aspect of the nonclassicality of the state vector of the system. In the light of the Agarwal exploration \cite{agarwal} on the link between $P$-function \cite{GSPfunction1,GSPfunction2,GSPfunction3} and CSI, he showed that if the quantity $I_{0}$, as defined by
\begin{eqnarray}\label{CSinequality}
I_{0}=\frac{\left( \langle a_{1}^{\dag 2}a_{1}^{2}\rangle \langle a_{2}^{\dag 2} a_{2}^{2} \rangle \right)^{1/2}}{|\langle a_{1}^{\dag}a_{1} a_{2}^{\dag} a_{2} \rangle|}-1.
\end{eqnarray}
gets negative (positive) values, the corresponding $P$-function would be negative (positive), i.e. the nonclassical (classical) behavior is appeared. Notice that, commonly, one may expect that Glauber-Sudarshan $P$-distribution function, in the sense that it is a distribution function, have to be positive, while there exists quantum states for which $P$-function is negative or highly singular (nonclassical states).\\
In figure \ref{csi} the time evolution of the CSI against the scaled time $\tau$ is illustrated for different chosen parameters similar to figure \ref{quant.ph-sq}, where the upside (downside) plots refer to the constant (intensity-dependent) coupling regime.
The upside plot of figure \ref{csi}(a) shows that the quantum system is not sensitive to this nonclassical criterion when the atom-field coupling is constant, while in the presence of intensity-dependent coupling (downside plot), CSI always gets negative value with oscillatory behaviour.
In figure \ref{csi}(b) where the effect of deformed Kerr medium is examined it is seen that, the deformed Kerr medium has no remarkable effect on the negativity of CSI, either in the presence or absence of the intensity-dependent coupling.
In figure \ref{csi}(c) the effect of deformed Kerr medium together with the detuning parameters is shown. From the upside plot of this figure it is observed that, CSI varies between positive (classical behaviour) and negative (nonclassical behaviour) values. Also, from downside one, this quantity takes negative values in all time domain with fast oscillatory behaviour. Comparing the upside plots of figures \ref{csi}(b) and \ref{csi}(c) indicates that due to the presence of detuning, nonclassicality behaviour is revealed. \\
By considering the presented results depicted in figure \ref{csi}, one can deduce that intensity-dependent coupling (the downside plots) plays an effective role on appearing the negativity of CSI, as the detuning. Also, deformed Kerr medium has no notable effect on the negativity of CSI. In addition, for all cases concerning with constant coupling regime (upside plots) the (typical) collapse and revival is clearly observed.
\subsection{Squeezing}\label{squeezing}
Squeezing phenomenon, as the last nonclassicality feature which will be discussed, is one of the applied nonclassicality features which is described by decreasing the quantum fluctuation in one of the field quadratures below the value of vacuum or canonical coherent states. In the continuation, two kinds of squeezing parameters are introduced and in each case, the squeezing condition of the state vector of the whole system is studied.
\subsubsection{Two-mode squeezing}
In order to study the two-mode squeezing (since we are dealing with the two-mode cavity field) of the state vector of the system, the following (Hermitian) two-mode quadrature operators have been defined \cite{knight}
\begin{eqnarray}\label{lsp}
\hat{X}_{1} &=& \frac{1}{2\sqrt{2}}(\hat{a}_{1}+\hat{a}^{\dag}_{1}+\hat{a}_{2}+\hat{a}^{\dag}_{2}), \nonumber \\
\hat{X}_{2} &=& \frac{1}{2i\sqrt{2}}(\hat{a}_{1}-\hat{a}^{\dag}_{1}+\hat{a}_{2}-\hat{a}^{\dag}_{2}).
\end{eqnarray}
These definitions lead to the commutation relation $[\hat{X}_{1},\hat{X}_{2}]=i/2$. Consequently, uncertainty relation for such operators read as
$\langle (\Delta \hat{X}_{1})^{2} \rangle \langle (\Delta \hat{X}_{2})^{2} \rangle \geq 1/16$, where $\Delta \hat{X}_{1}$ and $\Delta \hat{X}_{2}$ are the uncertainties in $\hat{X}_{1}$ and $\hat{X}_{2}$, respectively. A state is squeezed in $\hat{X}_{1} (\hat{X}_{2})$ if $\langle (\Delta \hat{X}_{1})^{2} \rangle <0.25\; \left( \langle (\Delta \hat{X}_{2})^{2}\rangle <0.25\right)$, or equivalently if $S_{X_{1(2)}}=4 \langle (\Delta \hat{X}_{1(2)})^{2} \rangle -1$ satisfies the inequality $-1<S_{X_{1(2)}}<0$. The latter inequality implies that in order to exist squeezing, two-mode Glauber-Sudarshan $P$-function, $P(\alpha,\beta)= P(\alpha) P(\beta)$, requires to be nonpositive or singular in some regions of phase space \cite{knight}.
Anyway, the two-mode squeezing parameters can be written as:
\begin{eqnarray}\label{vsqdx2}
S_{X_{1}}&=&\mathrm{Re} \Big( \langle \hat{a}_{1}^{2} \rangle + \langle \hat{a}_{2}^{2} \rangle
+2 \langle \hat{a}_{1}^{\dag}\hat{a}_{2}\rangle+2\langle \hat{a}_{1}\hat{a}_{2} \rangle \Big)
+\langle \hat{a}_{1}^{\dag}\hat{a}_{1} \rangle \nonumber\\
&+& \langle \hat{a}_{2}^{\dag}\hat{a}_{2} \rangle - 2\Big[\mathrm{Re}\Big(\langle \hat{a}_{1} \rangle+\langle \hat{a}_{2} \rangle\Big)\Big]^{2}, \nonumber\\
S_{X_{2}}&=&\mathrm{Re} \Big( 2\langle \hat{a}_{1}^{\dag}\hat{a}_{2} \rangle
-2\langle \hat{a}_{1}\hat{a}_{2} \rangle
-\langle \hat{a}_{1}^{2} \rangle - \langle \hat{a}_{2}^{2} \rangle\Big)
+\langle \hat{a}_{1}^{\dag}\hat{a}_{1} \rangle \nonumber \\
&+& \langle \hat{a}_{2}^{\dag}\hat{a}_{2} \rangle - 2\Big[\mathrm{Im}\Big(\langle \hat{a}_{1} \rangle+\langle \hat{a}_{2} \rangle\Big)\Big]^{2}.
\end{eqnarray}
Figure \ref{two-modesq} shows the temporal behavior of two-mode squeezing in $X_{1}$ for some of the chosen parameters as considered in figure \ref{quant.ph-sq}. Also, the upside (downside) plots concern with the absence (presence) of the intensity-dependent coupling. In the upside plot of figure \ref{two-modesq}(a), where the Kerr effect and intensity-dependent coupling are both absent and the resonance case is considered, the state of the system does not exhibit two-mode squeezing except in some small intervals of time. However, entering the intensity-dependent coupling leads to the two-mode squeezing in all times (the downside one).
The upside plots of figures \ref{two-modesq}(b) (studying the effect of deformed Kerr medium) and \ref{two-modesq}(c) (exhibiting simultaneously the effects of deformed Kerr medium and detuning parameters) show a temporal behaviour similar to the upside plot of figure \ref{two-modesq}(a).
From the downside plot of figure \ref{two-modesq}(b), it is observed that the presence of deformed Kerr medium and intensity-dependent coupling effects brings about to appear deeply two-mode squeezing. In this case, it is valuable to state that the amount of the negativity of two-mode squeezing for this situation becomes nearly $20$ times greater than the downside plot of figure \ref{two-modesq}(a).
From the downside plot of figure \ref{two-modesq}(c), it seems that in the intensity-dependent coupling regime, adding the detuning parameters on the deformed Kerr medium reduces the amount of the negativity of this nonclassicality sign. \\
Finally, it is deduced that intensity-dependent coupling helps clearly to occur the two-mode squeezing in $X_{1}$ and particularly the deformed Kerr medium strengthens this nonclassicality feature. Also, the reduction in the negativity $S_{X_{1}}$ and fast oscillatory behaviour are the results of considering the detuning effect.
\subsubsection{Sum squeezing}
Here, we are going to investigate sum squeezing phenomenon. In order to study this nonclassicality indicator, two Hermitian operators $\hat{Y}_{1}$ and $\hat{Y}_{2}$ are given by \cite{hillery}
\begin{eqnarray}\label{ss1}
\hat{Y}_{1} = \frac{1}{2}(\hat{a}_{1}\hat{a}_{2}+\hat{a}^{\dag}_{1}\hat{a}^{\dag}_{2}), \;
\hat{Y}_{2} = \frac{1}{2i}(\hat{a}_{1}\hat{a}_{2}-\hat{a}^{\dag}_{1}\hat{a}^{\dag}_{2}),
\end{eqnarray}
where $[\hat{Y}_{1},\hat{Y}_{2}]=\frac{i}{2}(\hat{a}^{\dag}_{1}\hat{a}_{1}+\hat{a}^{\dag}_{2}\hat{a}_{2}+1)$ yields the uncertainly relation $\langle (\Delta \hat{Y}_{1})^{2} \rangle \langle (\Delta \hat{Y}_{2})^{2} \rangle \geq\frac{1}{16}|\langle \hat{a}^{\dag}_{1}\hat{a}_{1}+\hat{a}^{\dag}_{2}\hat{a}_{2}+1\rangle)|^{2}$.
Similar to the previous section, we can define the normalized sum squeezing parameters correspond to operators $Y_{1}$ and $Y_{2}$ respectively as
\begin{eqnarray}\label{ssqy2}
\hspace{-1cm}S_{Y_{1}}&=&\frac{2 \mathrm{Re} \langle \hat{a}_{1}^{2} \hat{a}_{2}^{2} \rangle+2 \langle \hat{a}_{1}^{\dag}\hat{a}_{1}\hat{a}_{2}^{\dag}\hat{a}_{2} \rangle -4 \left( \mathrm{Re} \langle \hat{a}_{1} \hat{a}_{2} \rangle \right)^{2} }{\langle \hat{a}_{1}^{\dag}\hat{a}_{1} \rangle+ \langle \hat{a}_{2}^{\dag}\hat{a}_{2} \rangle +1}, \nonumber\\
\hspace{-1cm}S_{Y_{2}}&=&\frac{ 2 \langle \hat{a}_{1}^{\dag}\hat{a}_{1}\hat{a}_{2}^{\dag}\hat{a}_{2} \rangle - 2 \mathrm{Re} \langle \hat{a}_{1}^{2} \hat{a}_{2}^{2} \rangle -4 \left( \mathrm{Im} \langle \hat{a}_{1} \hat{a}_{2} \rangle \right)^{2} }{\langle \hat{a}_{1}^{\dag}\hat{a}_{1} \rangle+ \langle \hat{a}_{2}^{\dag}\hat{a}_{2} \rangle +1}.
\end{eqnarray}
Figure \ref{sumsq} describes the sum squeezing for different chosen parameters assumed in figure \ref{quant.ph-sq} with constant (intensity-dependent) coupling relating to the upside (downside) plots. It is seen from the upside plot ($f_{i}(n_{i}) = 1$) of figure \ref{sumsq}(a) that in the absence of the Kerr medium and in the resonance condition, the state of the system does not have sum squeezing property. However, after entering the intensity-dependent coupling, sum squeezing will be appeared in all times (the downside one).
Figure \ref{sumsq}(b) indicates the effect of the deformed Kerr medium, from which it is found that the squeezing does not occur when the atom-field coupling is constant. On the contrary, in the intensity-dependent coupling regime, sum squeezing with oscillatory behavior is clearly observed at all times. Looking quantitatively at figures \ref{sumsq}(a) and \ref{sumsq}(b) implies the fact that, deformed Kerr medium has no remarkable effect in the negativity of sum squeezing. The effects of deformed Kerr medium and detuning parameters are simultaneously studied in figure \ref{sumsq}(c). As is seen, positive values of sum squeezing indicate the classical behaviour and so the state of quantum system does not possess this nonclassicality indicator. Comparing the downside plots of figures \ref{sumsq}(b) and \ref{sumsq}(c) shows that sum squeezing disappears when the detuning is considered. \\
At last, one can conclude that while intensity-dependent coupling plays the important role to reach this nonclassical behaviour (negativity of sum squeezing parameter), the existence of the detuning destroys the sum squeezing of the state of the system even in the presence of intensity coupling. In addition, deformed Kerr medium preserve the amount of the negativity of this nonclassicality feature.
\section{Summary and conclusion}\label{conclusion}
In this paper, we have studied further physical aspects of a $\Lambda$-type three-level atom interacting with a quantized two-mode radiation field in a cavity including deformed Kerr medium in the presence of the detuning parameters with and without atom-field intensity-dependent coupling regime.
We have demonstrated the importance and notability of the considered nonclassicality features (which are of special interest in the quantum optics field of research) in the Introduction of this paper.
In summary, using the analytical solution of the state vector of the considered bipartite (atom-field) system which has been very recently obtained by us in \cite{usJOSA}, first, the number-phase entropic uncertainty relation and subsequently entropy squeezing (in phase component) have been evaluated, by applying the two-mode Pegg-Barnett approach. In the continuation, other nonclassical properties of the state vector of the entire system namely, Mandel parameter, Cauchy-Schwartz inequality, Two-mode and sum squeezing have been numerically examined. In each case, we individually and simultaneously studied the effects of $` `$intensity-dependent coupling$"$ (entered by $f_{i}(n_{i})$), $` `$deformed Kerr medium$"$ (by the nonlinearity function $g_{i}(n_{i})$) and $` `$detuning parameters$"$ on the nonclassicality criteria.
Summing up, the main results of the paper are listed in what follows.
\begin{itemize}
\itemsep1em
\item {\it Tuning the nonclassicality indicators:} It is illustrated that the amount of considered nonclassicality features can be tuned by choosing the nonlinear parameters related to the atom-field system, suitably.
\item {\it Intensity-dependent coupling:} Presented results demonstrate that intensity-dependent coupling (which is considered by the function $f_{i}(n_{i}) = 1/\sqrt{n_{i}}$) has generally a direct role and striking effect in exhibiting and ameliorating the nonclassicality features. Even though, it seems that entering the intensity-dependent coupling in calculating Mandel parameter and CSI disappears the collapses and revivals phenomena as a nonclassicality feature.
\item {\it Deformed Kerr medium:} Paying attention to the related results implies that the deformed Kerr medium has an obvious role and noteworthy effect in appearing and improving the nonclassicality features. Also, it is worthwhile to accentuate the fact that deformed Kerr medium which is one of the new features of the presented work, is distinguished from a Kerr medium by the nonlinearity function $g_{i}(n_{i}) = 1/\sqrt{n_{i}}$.
\item {\it Detuning parameters:} Looking deeply at the obtained results shows that the detuning parameters quantitatively reduce the amount of the nonclassicality features. Also, as a result of the presence of the detuning parameters in the intensity-dependent coupling regime, fast oscillatory in the nonclassicality signs can be observed.
\end{itemize}
Finally, we would like to mention that this study can be performed for any physical system, either any nonlinear oscillator algebra with arbitrary nonlinearity function or any solvable quantum system with known discrete energy spectrum $e_{n}$ \cite{en1,en2,en3,en4,en5}, too.
\begin{flushleft}
{\bf Acknowledgements}\\
\end{flushleft}
The authors would like to express their thanks sincerely to Dr Mohammad Reza Hooshmandasl for his valuable help in the numerical results.
|
\section{Introduction}
\label{intro}
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}{Lemma}
\newtheorem{remark}{Remark}
\newtheorem{definition}{Definition}
\newtheorem{assumption}{Assumption}
\newtheorem{corollary}{Corollary}
The unknown control direction problem has attracted significant research interest over the last three decades. Nussbaum gains \cite{Nussbaum_paper}, \cite{Ye_Jiang98} originally introduced in \cite{Nussbaum_paper} have become the main theoretical tool for controller design for systems with unknown control directions.
Nussbaum functions (NFs) are continuous functions $N(\cdot)$ with the property
\begin{align}
\limsup_{\zeta\rightarrow +\infty}&\frac{1}{\zeta}\int_{0}^{\zeta}{N(s)ds}=+\infty\label{nussbaum propertyp}\\
\liminf_{\zeta\rightarrow +\infty}&\frac{1}{\zeta}\int_{0}^{\zeta}{N(s)ds}=-\infty\label{nussbaum propertym}.
\end{align}
Examples of NFs are $\zeta^2\cos(\zeta)$, $\exp(\zeta^2)\sin(\zeta)$ among many others.
For the simple integrator case $\dot{y}=bu$ with $b$ a nonzero constant of \emph{unknown sign}, standard analysis \cite{Nussbaum_paper} shows that the Nussbaum control law
\begin{align}\label{nussbaum controller}
u&=\zeta^2\cos(\zeta)y\nonumber\\
\dot{\zeta}&=y^2
\end{align}
ensures convergence of the output $y$ to the origin and boundedness of the Nussbaum parameter $\zeta$. However, Georgiou and Smith demonstrated in \cite{GS} that the proposed controller is nonrobust to fast parasitic unmodelled dynamics. Particularly, they considered the system
\begin{align}\label{G-S}
\dot{x}&=bu\nonumber\\
\dot{y}&=M(x-y)
\end{align}
and showed divergence for $M>1$ when the controller \eqref{nussbaum controller} is used.
An alternative nonlinear PI methodology was proposed by Ortega, Astolfi and Barabanov in \cite{Ortega_paper} to address the unknown control direction problem. For the simple integrator case, their controller takes the form
\begin{equation}\label{nonlinear PI}
u=z\cos(z)y
\end{equation}
with $z$ the PI square error defined by $z=(1/2)y^2+\lambda\int_0^t{y^2(s)ds}$. The main difference between the two controllers \eqref{nussbaum controller}, \eqref{nonlinear PI} is the existence of the proportional term in the control gain of \eqref{nonlinear PI} (see also p. 166 of \cite{AKO_book}). It was hinted in \cite{Ortega_paper},\cite{AKO_book}(no complete proof was given) that such a controller is robust to fast parasitic first-order perturbations and therefore can stabilize the Georgiou-Smith example system if $\lambda<M$. Their argument, however, was based on the fact that the related transfer function is positive real and cannot be carried over to the case of an unstable unforced linear system or even a nonlinear system. In fact, the introduction of a simple destabilizing pole in the system
\begin{align}\label{G-SL}
\dot{x}&=\alpha x+bu\nonumber\\
\dot{y}&=M(x-y)
\end{align}
($\alpha>0$) may result in instability of the closed-loop system with the controller \eqref{nonlinear PI} even if $\alpha+\lambda<M$ (see Section \ref{simulation}).
It remains therefore an open problem to design a nonlinear PI controller robust to fast parasitic dynamics when the plant to be controlled is originally unstable and nonlinear. To this end, we consider an extension of the Georgiou-Smith system. Particularly, we examine the overall dynamic behavior of the nonlinear system with first-order unmodelled dynamics given by
\begin{align}\label{G-SE}
\dot{x}&=f(x)+bu\nonumber\\
\dot{y}&=M(x-y)
\end{align}
when a nonlinear PI control law $u$ designed for the unperturbed system
\begin{equation}\label{unperturbed}
\dot{y}=f(y)+bu
\end{equation}
is applied.
\subsection{Nonlinear PI for the unperturbed system}
For system \eqref{unperturbed}, we assume that $f(\cdot)$
is a sector-bounded nonlinearity, i.e. $f(0)=0$, $f(y)=y\alpha(y)$ and there exist some constants $\alpha_1, \alpha_2\in\mathbb{R}$ such that $\alpha_1\leq \alpha(y)\leq \alpha_2$ $\forall y\in\mathbb{R}$.
A controllability assumption is also imposed, that is $b\neq 0$.
Let now a nonlinear PI controller of the form
\begin{align}
u&=\kappa(z)y\label{PI_o}\\
z&=(1/2)y^2+\lambda\int_0^t{y^2(s)ds}\label{z}
\end{align}
with PI gain $\kappa(z)=\alpha_0(z)\cos(z)$ where $\alpha_0(\cdot)$ is a class $\mathcal{K}_{\infty}$ function and $\lambda>0$.
For the $z$ derivative we have for the unperturbed system \eqref{unperturbed} that
\begin{equation}\label{zdot}
\dot{z}\leq \big[\max\{|\alpha_1|,|\alpha_2|\}+\lambda+b \kappa(z)\big]y^2.
\end{equation}
Note that whenever $z(t)=z_k$ with $z_k:=(\pi/2)[4k+1+\textrm{sgn}(b)]$ we have
\begin{equation}\label{zdot}
\dot{z}(t)\leq \big[\max\{|\alpha_1|,|\alpha_2|\}+\lambda-|b|\alpha_0(2k\pi)\big]y^2(t).
\end{equation}
Thus, $\dot{z}(t)\leq 0$ whenever $z(t)=z_k$ for every $k\geq k_0$
\begin{equation*}
k_0:=\bigg\lceil\frac{1}{2\pi}\alpha_0^{-1}\bigg[\frac{1}{|b|}(\max\{|\alpha_1|,|\alpha_2|\}+\lambda)\bigg]\bigg\rceil
\end{equation*}
($\lceil x\rceil$ denotes the largest integer not exceeding $x$) which in turn implies that $z$ is bounded by $z(t)\leq z_{k'}$ where $k':=\max\{k_0,\lceil y^2(0)/4\pi\rceil\}$. The fact that $z\in\mathcal{L}_{\infty}$ implies $y\in\mathcal{L}_{\infty}\cap\mathcal{L}_{2}$ and $u\in\mathcal{L}_{\infty}$ from \eqref{z} and \eqref{PI_o} respectively. Also, from \eqref{unperturbed} we have $\dot{y}\in{\mathcal{L}_{\infty}}$. Barbalat's lemma can now be invoked to prove that $\lim_{t\rightarrow\infty}y(t)=0$. This is a standard analysis in the spirit of \cite{AKO_book}.
Assume now the existence of parasitic first order unmodelled dynamics in the form of \eqref{G-SE}. Sufficient conditions are given in the next section for \emph{global boundedness and attractivity} for the closed-loop system comprised from \eqref{G-SE} and the nonlinear PI controller \eqref{PI_o} and \eqref{z}. A key property is that the nonlinear PI gain function $\kappa(z)$ should be a function of Nussbaum type.
\section{Extended Georgiou-Smith system with sector nonlinearity}
\label{}
In this section we consider system \eqref{G-SE} with a sector-bounded nonlinearity
\begin{align}
f(x)&=\alpha(x)x\label{sector}\\
\alpha_1\leq \alpha&(x)\leq \alpha_2 \qquad \forall x\in\mathbb{R}\label{sector bounds}
\end{align}
for some constants $\alpha_1,\alpha_2\in\mathbb{R}$. Note that $\alpha_1,\alpha_2$ can also take positive values rendering the unforced system unstable.
To simplify notation let us define the constant $\epsilon:=1/M$.
We have established the following theorem.
\begin{theorem}\label{main_theorem}
Let the closed-loop system described by \eqref{G-SE}, \eqref{PI_o}, \eqref{z} with sector-bounded nonlinearity given by \eqref{sector}, \eqref{sector bounds}. If
\begin{description}
\item[(i)] $\epsilon\lambda<1$, $\epsilon(\lambda+\alpha_2)<1$
\item[(ii)] $\alpha_2-\alpha_1\leq \frac{2\lambda}{\sqrt{1-\epsilon\lambda}}\Big[\sqrt{1-\epsilon(\lambda+\alpha_1)}+\sqrt{1-\epsilon(\lambda+\alpha_2)}\Big]$
\item[(iii)] $\kappa(z)$ has the Nussbaum property \eqref{nussbaum propertyp},\eqref{nussbaum propertym}
\end{description}
then, all closed-loop signals are bounded and $\lim_{t\rightarrow\infty}y(t)=\lim_{t\rightarrow\infty}x(t)=0.$
\end{theorem}
\begin{proof}
From the definition of the PI error $z$ in \eqref{z} and \eqref{G-SE} we have that
\begin{equation}\label{z_dynamics}
\dot{z}=Mxy-M(1-\epsilon\lambda)y^2.
\end{equation}
Let now the function
\begin{align}\label{S}
S:=\frac{\lambda}{2}x^2+\frac{1}{2}M(1-\epsilon\lambda)(x-y)^2+\epsilon c z- b\int_0^z{\kappa(s)ds}
\end{align}
with $c\in\mathbb{R}$ to be defined. Replacing from \eqref{G-SE}, \eqref{PI_o}, \eqref{z}, \eqref{z_dynamics} and canceling terms we have for its time derivative that
\begin{align}\label{dotS}
\dot{S}=\lambda\alpha(x)x^2-M^2&(1-\epsilon\lambda)(x-y)^2\nonumber\\
&+M(1-\epsilon\lambda)\alpha(x)x(x-y)+cxy-c(1-\epsilon\lambda)y^2.
\end{align}
Eq. \eqref{dotS} can be written in matrix notation as
\begin{align}\label{dotS1}
\dot{S}&=-M^2\left[
\begin{array}{cc}
x & y \\
\end{array}
\right]
\left[
\begin{array}{cc}
1-\epsilon(\lambda+\alpha(x)) & -\frac{1}{2}[c\epsilon^2+(1-\epsilon\lambda)(2-\epsilon\alpha(x))]\\
* & (1-\epsilon\lambda)(c\epsilon^2+1)
\end{array}\right]\left[
\begin{array}{c}
x \\
y \\
\end{array}
\right]
\nonumber\\
& := -M^2\left[
\begin{array}{cc}
x & y \\
\end{array}
\right]\Lambda(x) \left[
\begin{array}{c}
x \\
y \\
\end{array}
\right]
\end{align}
where $*$ denotes a symmetric w.r.t. the main diagonal element of $\Lambda(x)$. We claim that there is some constant $c\in\mathbb{R}$ such that $\Lambda(x)$ is positive definite for all $x\in\mathbb{R}$. Equivalently, we can prove that, for some $c\in\mathbb{R}$, the two principal minors of $\Lambda(x)$ given by
\begin{align*}
\Delta_1(x)&:=1-\epsilon(\lambda+\alpha(x))\\
\Delta_2(x)&:=[1-\epsilon(\lambda+\alpha(x))](1-\epsilon\lambda)(c\epsilon^2+1)-\frac{1}{4} [c\epsilon^2+(1-\epsilon\lambda)(2-\epsilon\alpha(x))]^2
\end{align*}
are positive $\forall x\in\mathbb{R}$. From assumption (i) of Theorem \ref{main_theorem}, it is obvious that $\Delta_1(x)>0$ $\forall x\in\mathbb{R}$. For $\Delta_2(x)$ we have that
\begin{equation}\label{Delta_2}
\Delta_2(x)=-(1/4)\big[A_c c^2+B_c(x) c+\Gamma_c(x)\big]
\end{equation}
with $A_c=\epsilon^4>0$, $B_c(x)=2\epsilon^3(1-\epsilon\lambda)[\alpha(x)+2\lambda]$, $\Gamma_c(x)=\epsilon^2(1-\epsilon\lambda)\alpha(x)[4\lambda+(1-\epsilon\lambda)\alpha(x)]$. $\Delta_2(x)$ is therefore a quadratic polynomial with respect to $c$ that is positive definite if
\begin{description}
\item[(a)] $\Delta_c(x):=B_c(x)^2-4A_c\Gamma_c(x)>0$
\item[(b)] there exists some constant $c\in \big(c_1(\alpha(x)),c_2(\alpha(x))\big)$ for all $x\in\mathbb{R}$ where $c_1(\alpha(x))$, $c_2(\alpha(x))$ are the two roots of $\Delta_2(x)$ given by
\begin{align*}
c_1(\alpha(x))&:=-M\Big[(1-\epsilon\lambda)(2\lambda+\alpha(x))+2\lambda\sqrt{(1-\epsilon\lambda)\big[1-\epsilon(\lambda+\alpha(x))\big]}\Big]\\
c_2(\alpha(x))&:=-M\Big[(1-\epsilon\lambda)(2\lambda+\alpha(x))-2\lambda\sqrt{(1-\epsilon\lambda)\big[1-\epsilon(\lambda+\alpha(x))\big]}\Big].
\end{align*}
\end{description}
If we carry out the calculations we have that
\begin{equation*}
\Delta_c(x)=16\epsilon^6\lambda^2(1-\epsilon\lambda)[1-\epsilon(\lambda+\alpha(x))]
\end{equation*}
and therefore the positivity condition for $\Delta_c(x)$ is satisfied if $\epsilon(\lambda+\alpha_2)<1$ and $\epsilon\lambda<1$. For condition (b) to be true, as $\alpha(x)$ varies in $[\alpha_1,\alpha_2]$, there must be some $c\in\mathbb{R}$ such that $c\in[c_1(\alpha),c_2(\alpha)]$ for all $\alpha\in[\alpha_1,\alpha_2]$. This holds true if $\max_{\alpha\in[\alpha_1,\alpha_2]}c_1(\alpha)<\min_{\alpha\in[\alpha_1,\alpha_2]}c_2(\alpha)$. Function $c_2(\cdot)$ is obviously decreasing with respect to $\alpha(x)$ with minimum value $c_2(\alpha_2)> c_2((1/\epsilon)(1-\epsilon\lambda))=-(1/\epsilon^2)(1-\epsilon^2\lambda^2)$. Function $c_1(\cdot)$ on the other hand is decreasing up to some point $\alpha_0=(1/\epsilon)[1-\epsilon\lambda/(1-\epsilon\lambda)]$ and then increasing up to $(1/\epsilon)(1-\epsilon\lambda)$ with value $c_1((1/\epsilon)(1-\epsilon\lambda))=-(1/\epsilon^2)(1-\epsilon^2\lambda^2)<c_2(\alpha_2)$. Thus, the second condition holds true if $c_1(\alpha_1)<c_2(\alpha_2)$ which is exactly assumption (ii) of the theorem. Thus, selecting $c:=\epsilon_0 c_1(\alpha_1)+(1-\epsilon_0)c_2(\alpha_2)$ for any $\epsilon_0\in(0,1)$ we have $\dot{S}\leq 0$.
Integrating now $\dot{S}\leq 0$ we have that $S(t)\leq S(0)$ or equivalently
\begin{align}\label{Sbound}
\lambda x^2(t)+M(1-\epsilon\lambda)(x(t)-y(t))^2 <2S(0)-2\epsilon cz(t)+ 2b\int_0^{z(t)}{\kappa(s)ds}.
\end{align}
The above inequality and the Nussbaum property of $\kappa(\cdot)$ ensure the boundedness of $z$. To prove this, let us assume the contrary. From the Nussbaum property (iii) of $\kappa(z)$ there exists a strictly increasing sequence $\{z_k\}_{k=1}^{\infty}$ such that $\lim_{k\rightarrow\infty}z_k=+\infty$ and
\begin{equation}\label{nussbaum_zk}
\lim_{k\rightarrow\infty}\frac{1}{z_k}\int_0^{z_k}{b\kappa(w)dw}=-\infty.
\end{equation}
Due to continuity of $z$, the PI square error $z$ will eventually pass from an infinite number of elements of $\{z_k\}_{k=1}^{\infty}$. From \eqref{Sbound}, we have for the times $t_k$ at which $z(t_k)=z_k$
\begin{align}\label{Sbound_tk}
\frac{\lambda}{2} x^2(t_k)+\frac{1}{2}M(1-\epsilon\lambda)(x(t_k)-y(t_k))^2 <S(0)-\epsilon cz_k+ b\int_{0}^{z_k}{\kappa(w)dw}.
\end{align}
If we divide all terms in \eqref{Sbound_tk} with $z_k$ and take into account the limiting property \eqref{nussbaum_zk} the left hand side (l.h.s.) of \eqref{Sbound_tk} should take negative values for all $k\geq k_0$ for some $k_0\in\mathbb{N}$. This yields the desired contradiction since the l.h.s. of \eqref{Sbound_tk} is a sum of squares which is always nonnegative. Thus, $z\in\mathcal{L}_{\infty}$ and therefore $y\in\mathcal{L}_{\infty}\cap\mathcal{L}_{2}$, $u\in\mathcal{L}_{\infty}$ and from \eqref{Sbound} $x\in\mathcal{L}_{\infty}$. Then, the system equations \eqref{G-SE} yield $\dot{x},\dot{y}\in\mathcal{L}_{\infty}$. Invoking now Barbalat lemma we obtain the desired property $\lim_{t\rightarrow\infty}x(t)=\lim_{t\rightarrow\infty}y(t)=0$.
\end{proof}
\begin{remark}
In the case of a linear system $f(x)=\alpha x$ condition (ii) of Theorem \ref{main_theorem} is no longer needed and (i) reduces to $\epsilon\lambda<1$ for $\alpha\leq 0$ (nominally stable system) and $\epsilon(\alpha+\lambda)<1$ for $\alpha>0$ (nominally unstable). Note that in the latter case the necessary condition for stabilization by simple output feedback (with known sign of $b$) is $\epsilon\alpha<1$.
\end{remark}
\begin{remark}
If $c_2(\alpha_2)\geq 0$ then the constant $c$ in the definition of $S$ can be nonnegative. This means that in \eqref{Sbound} $-c\epsilon z(t)\leq 0$ and the Nussbaum condition (iii) for $\kappa(z)$ in Theorem \ref{main_theorem} can be relaxed to
\begin{align}
\limsup_{z\rightarrow\infty}\int_{0}^{z}{\kappa(s)ds}=+\infty\\
\liminf_{z\rightarrow\infty}\int_{0}^{z}{\kappa(s)ds}=-\infty.
\end{align}
After calculations one can show that condition $c_2(\alpha_2)\geq 0$ holds true iff $\alpha_2\leq 0$.
Thus, if the unforced linear system is stable ($\alpha\leq 0$) and $\epsilon\lambda<1$ then, the controller \eqref{nonlinear PI} results in bounded and attractive closed-loop behavior. This also provides a strict proof for the integrator example of \cite{Ortega_paper}, \cite{AKO_book}.
\end{remark}
\begin{remark}\label{remark_sector}
Note that the r.h.s. of condition (ii) tends to $4\lambda$ in the limit $\epsilon\rightarrow 0$. Thus, for some sector bounded nonlinearity \eqref{sector bounds}, if we select $\lambda>(\alpha_2-\alpha_1)/4$ then there exists some $\epsilon_0>0$ such that for all $\epsilon<\epsilon_0$ the closed-loop system \eqref{G-SE}, \eqref{PI_o}, \eqref{z} is globally bounded and attractive.
\end{remark}
\section{Simulation results}
\label{simulation}
A simulation study was performed for the perturbed integrator (P-INT) and the perturbed linear system (P-LS) described by \eqref{G-S}, \eqref{G-SL} respectively with parameters $\alpha=1$, $b=1/2$, $\lambda=2.5$, $\epsilon=1/4$ and initial conditions $x(0)=y(0)=4$. For the specific parameters, condition (i) of Theorem \ref{main_theorem} holds true. We tested the case of a Nussbaum gain based (NG) controller \eqref{nussbaum controller} and a nonlinear PI controller \eqref{nonlinear PI} with gains $\kappa(z)=z\cos(z)$ (not a Nussbaum function) denoted as nPI and $\kappa(z)=z^2\cos(z)$ (Nussbaum function) denoted as nPI-N.
\begin{figure}[!t]
\centering
\includegraphics[width=4.1in]{comparisons}
\caption{Linear system: Time responses of $y(t)$ for the cases of a perturbed integrator (P-INT) and a perturbed linear system (P-LS) for the three controllers NG, nPI, nPI-N.}
\label{comparisons}
\end{figure}
The output response $y$ shown in Fig. \ref{comparisons} verifies our theoretical analysis. Particularly, for the P-INT system with the NG controller, $y$ is divergent as shown in \cite{GS}. If the nPI controller is used then $y$ remains bounded and converges to zero \cite{Ortega_paper}, \cite{AKO_book}. However, the nPI control fails to regulate the P-LS system. Convergent solutions are obtained for the P-LS system only when the nPI-N is employed.
Let now the perturbed nonlinear system \eqref{G-SE} with $f(x)=3[1+\sin^2(x)]x$ where $\alpha_1=3$, $\alpha_2=6$ and $b=1$. Selecting $\lambda=2.5$, we have that both conditions (i) and (ii) are satisfied for every $\epsilon<1/(\alpha_2+\lambda)=2/17$ (see Remark \ref{remark_sector}). For the control law \eqref{PI_o}, \eqref{z}, $\kappa(z)=z^2\sin(z)$ simulation results are shown in Fig. \ref{sector_example} with $\epsilon=0.1$ and initial conditions $x(0)=y(0)=4$.
\begin{figure}[!t]
\centering
\includegraphics[width=4.1in]{sector_example}
\caption{Nonlinear system: Time responses of system states $x,y$ and control input $u$.}
\label{sector_example}
\end{figure}
As expected, all $x,y,u$ are bounded and converge to the origin as time passes.
\section{Conclusions}
Sufficient conditions are derived for global boundedness and attractivity of a perturbed nonlinear system with sector-bounded nonlinearity under a nonlinear PI control action.
The results further demonstrate the superiority of the nonlinear PI controls compared to simple Nussbaum gain based schemes with respect to robustness to unmodelled dynamics.
|
\section{I Introduction}
The experimental study of nucleon excited states is fundamental for the understanding of its internal structure. Important differences are still observed today between the experimental nucleon spectrum and the predictions of the first Constituent Quark Models (CQM)\cite{Capstick:1986bm,Capstick:1992uc,Capstick:1993kb,Riska:2000gd} but also with the results of recent approaches like lattice QCD calculations\cite{Edwards:2011jj}, Dyson-Schwinger equation of QCD\cite{Roberts:2011cf}, harmonic oscillator CQM\cite{Klempt:2012fy} and hypercentral CQM\cite{Giannini:2015zia}. Recent reports on advances in the experimental studies of the excited nucleon state spectrum can be found in\cite{Crede:2013kia,Burkert:2013dia}
Several states predicted by these models have not been observed ({\it missing resonances}). The nucleon excited states decay strongly with meson emission; therefore meson photoproduction experiments off the nucleon are an ideal way of searching for missing resonances and complement the information obtained with pion-nucleon scattering experiments.
In pseudo-scalar meson photoproduction off the proton ($\gamma + p \rightarrow meson + p$) we have eight possible combinations of spin states. The scattering amplitude is thus described by eight matrix elements, only four of which are independent due to rotational invariance and parity transformations.
With these four complex amplitudes, 16 bilinear products can be constructed, corresponding to 16 observables: the differential cross section, three single polarisation observables and twelve double polarisation observables.
To determine the scattering amplitude thoroughly, the cross section, the
three single polarisation and four appropriately chosen double
polarisation observables must be measured \cite{Keaton:1996pe,Chiang:1996em}. These observables can be expressed in terms of helicity amplitudes and the following relations hold\cite{Drechsel:1998hk,Feuster:1998cj,Saghai:1996hn,Fasano:1992es}:
${d\sigma}/{d\Omega} \sim |H_{1}|^{2} + |H_{2}|^{2} + |H_{3}|^{2} + |H_{4}|^{2}$
$\Sigma \sim Re(H_{1}H_{4}^{*} - H_{2}H_{3}^{*})$
$T \sim Im(H_{1}H_{2}^{*} - H_{3}H_{4}^{*})$
$P \sim Re(H_{1}H_{3}^{*} - H_{2}H_{4}^{*})$ \newline
where ${d\sigma}/{d\Omega}$ is the differential cross section and $\Sigma$, $\it{T}$ and $\it{P}$ are the beam, target and recoil asymmetries respectively. From the above relations one can see that the
the amplitude phases can not be accessed from the data on differential cross
section alone and that the only source of information to determine them are the polarisation asymmetries.
The combined study of the unpolarised cross section and of the polarisation
asymmetries is critical in order to get access to the
production amplitudes from the experimental data, and the interference among the helicity amplitudes can play a crucial role in revealing subtle effects\cite{Arndt:1989ww}.
The pseudo-scalar nature of the $\eta^\prime$ meson ensures that only $N^*$ resonances contribute to the process. The production threshold at W = 1.896 GeV (corresponding to an incident photon energy of 1.447 GeV for a free proton target) is located just above the so-called $resonance~gap$, where many of the predicted, but so far unobserved, $N^*$ states should be located.
The first data on $\eta^{\prime}$ photoproduction cross section were produced in 1968
\cite{ABBHHM:1968aa}
, and confirmed in 1976
\cite{Struczinski:1975ik}. Over 20 years later, the SAPHIR collaboration\cite{Plotzke:1998ua} reported a more extended measurement, based on 250 events, from which the masses and widths of the dominating $S_{11}$ and $P_{11}$ resonances were extracted.
In more recent years, the CLAS experiment at Jlab and the CB-ELSA-TAPS in Bonn have produced a rich amount of precise total and differential cross section data on the proton\cite{Dugger:2005my,Williams:2009yj,Crede:2009zzb} in the energy region from threshold up to 2.84 GeV.
From the theoretical point of view, four approaches are available in the literature:
a relativistic meson-exchange model of hadronic interactions\cite{Nakayama:2005ts,Huang:2012xj};
a reggeized model for $\eta$ and $\eta^{\prime}$ photoproduction\cite{Chiang:2002vq};
a chiral quark-model\cite{Zhong:2011ht} and
an isobar model\cite{Tryasuchev:2013fua}.
As a consequence of this huge experimental and theoretical effort, it was established that three above-threshold resonances ($S_{11}, P_{11}, P_{13}$), and the four-star sub-threshold $P_{13}$(1720) resonance reproduce best all existing data for the $\eta^{\prime}$ production processes in the resonance-energy region\cite{Huang:2012xj}, and that above 2 GeV, where the process is dominated by the $\rho$ and $\omega$ exchange, the dynamics of $\eta^{\prime}$ photoproduction are similar to those of $\eta$ photoproduction\cite{Crede:2009zzb}.
All the abovementioned state-of-the-art theoretical calculations give a reasonable description of the data. In all cases the authors stress that the cross section data alone are unable to pin down the resonance parameters, while polarisation observables could be very helpful to better determine the partial wave contributions in this reaction and impose more stringent constraints on the parameter values of the different models.
In this letter, we present the first measurement of the single polarisation observable $\Sigma$ for $\eta^{\prime}$ photoproduction off the proton, at the incoming photon energies of 1.461 and 1.480 GeV, obtained with the Compton backscattered photon beam of the GrAAL experiment.
The GrAAL experiment was located at the European Synchrotron Radiation Facility (ESRF) in Grenoble
(France), where it took data from 1995 to 2008.
A linearly polarised photon beam impinged on a liquid H2 or D2 target, and the final products were detected by the large solid angle detector LAGRAN$\gamma$E (Large Acceptance GRaal-beam Apparatus for Nuclear $\gamma$ Experiments).
The photon beam was produced by the Compton back-scattering of low-energy polarised photons from an
Argon laser, against the 6.03 GeV electrons circulating inside the ESRF storage ring\cite{DAngelo:2000ce}. The UV laser line (3.53 eV) was used to produce a backscattered photon beam, covering the energy range up to 1.5 GeV. A tagging system, located inside the electron ring, provided an event-by-event measurement of the photon beam energy, with a resolution of 16 MeV (FWHM).
Since the electron involved in the Compton scattering is ultra-relativistic, its helicity is conserved in the
process at backward angles, and the outgoing photon retains the polarisation of the incoming laser beam (up to 96\% for the UV laser line). The correlation between photon energy and polarisation is calculated with QED \cite{Babusci:1995ji}. During data taking, the laser beam polarisation was rotated by 90$^{o}$ every 20 minutes approximately, and unpolarised data from the Bremsstrahlung of the electrons off the ESRF residual vacuum were collected as well.
A detailed description of the LAGRAN$\gamma$E apparatus can be found in \cite{Bartalini:2005wx}. For the purpose of this letter
we underline the excellent energy resolution of the BGO electromagnetic calorimeter (${\it Rugby~Ball}$)\cite{LeviSandri:1996tk} where
photons from the $\eta^{\prime}$ decay chain were measured, and the position and time resolution in the forward direction
(1.5$^{\circ}$ and 2$^{\circ}$ (FWHM) for polar and azimuthal angles, respectively and 300 ps for time-of-flight (TOF)) where the recoil protons
from the photoreaction were detected.
Data were collected during eight different stretches of the GrAAL experiment, from 1998 to 2002. As the threshold for $\eta^{\prime}$ photoproduction off the proton is $E_{th} = 1.447$~GeV, only the periods of measurement performed by using the UV laser line (351 nm wavelength) allow to reach $E_{th}$ and to explore the behaviour of the asymmetry as a function of the photon energy up to 1.5 GeV. The $\eta^{\prime}$ mesons were identified via $\gamma\gamma$, $\pi^{0}\pi^{0}\eta$ and $\pi^{+}\pi^{-}\eta$ decay modes and by requiring the fulfilment of the two-body kinematics for the recoil proton.
The initial event selection, common to all the $\eta^{\prime}$ decay modes, required:\newline
i) at least two photons measured in the ${\it Rugby~Ball}$ for the invariant mass reconstruction;\newline
ii) a tagging energy above $E_{th}$;\newline
iii) a proton detected in the forward TOF wall with polar angle $\theta_{p}$ lying in the acceptance region
shown in Fig. \ref{Fig1}~(a).
\begin{figure}[htbp]
\begin{center}
\resizebox{0.48\textwidth}{!}{\includegraphics{fig1a.eps}\includegraphics{fig1b.eps}}
\put(-183,50){(a)} \put(-95,50){(b)}\\
\resizebox{0.50\textwidth}{!}{\includegraphics{fig1c.eps}\includegraphics{fig1d.eps}}
\put(-233,50){(c)} \put(-105,50){(d)}\\
\resizebox{0.50\textwidth}{!}{\includegraphics{fig1e.eps}\includegraphics{fig1f.eps}}
\put(-233,50){(e)} \put(-105,50){(f)}
\caption{(Colour online) Panel a: energy of photon beam $\it{vs.}$ the proton polar angle $\theta_{p}$
for a simulated $\gamma p \rightarrow \eta^{\prime} p$. Panel b: missing mass spectrum from the recoil proton detection. The black dashed curve shows the effects of selection cuts i) and ii); while the solid blue curve is the result of all preliminary selection cuts i), ii) and iii). Panels c, d and e: Invariant mass spectrum
from photons (two photons in panels c and e, six photons in panel d) in the BGO calorimeter
$\it{vs.}$ the missing mass spectrum obtained from the measurement of
the recoil proton. There are no events in the white area. Panel f: Missing
mass spectra from the recoil proton measurement after the selection of
the events in panels c, d and e.\label{Fig1}}\vspace{-0.9cm}
\end{center}
\end{figure}
The distribution of Fig.~\ref{Fig1} (a) was produced with an upgraded version of the event generator described in \cite{Corvisiero:1994wz}. As we can see, for the photon energies available at GrAAL, the recoil proton is always detected in the forward direction ($\theta_{p}\leq 16^{\circ}$). Moreover, the momentum/energy ratio determined by the two-body kinematics is always below 0.4. We therefore detected non relativistic protons in the forward direction.
In these conditions, the resolution on the proton momentum for the $\eta^{\prime}$ photoproduction was estimated with a GEANT3\cite{Brun:1987ma} simulation to be about 2.5\%.
The $\eta^{\prime}$ missing mass calculated from the recoil proton is shown in Fig.~\ref{Fig1}~(b). The effects of the cuts i) and ii) are shown as a black dashed line. The inclusion of cut iii) gave as a result the blue solid line. The $\eta^{\prime}$ peak is clearly visible over a smooth background. This residual background was eventually suppressed by additional constraints on the decay products of the $\eta^{\prime}$ meson.
The cleanest decay channel for LAGRAN$\gamma$E is the decay $\eta^{\prime}\rightarrow\gamma\gamma$. The two final-state photons were detected in the ${\it Rugby~Ball}$ and give rise, together with the recoil proton, to the missing mass $\it{vs.}$ invariant mass distribution of Fig.~\ref{Fig1}~(c). This decay mode has a rather small branching ratio ($\simeq 2.20\%$\cite{Agashe:2014kda}) and the number of events collected (3400) did not allow for the extraction of the beam asymmetry with sufficiently good statistics. For this reason, the decay channels involving two pions and one $\eta$ meson were also included in the analysis. The $\eta^{\prime}\rightarrow\pi^{0}\pi^{0}\eta$ decay channel was included by requiring the detection of six photons in the ${\it Rugby~Ball}$ reconstructing the $\eta^{\prime}$ meson invariant mass (Fig. \ref{Fig1}~(d)). For the inclusion of the charged decay channel ($\eta^{\prime}\rightarrow\pi^{+}\pi^{-}\eta$) we required the invariant mass reconstruction from $\eta$ meson decay into two photons (Fig.~\ref{Fig1}~(e)) and two charged tracks in the whole detector, identified as charged pions. All events with extra spurious signals in the detector, charged or neutral, were rejected.
The influence on the missing mass calculated from the recoil proton of the selection on the decay products of the $\eta^{\prime}$ is shown in Fig. \ref{Fig1}~(f). The three missing mass distributions exhibit the same behaviour and the values of the resulting $\eta^{\prime}$ masses are in keeping with the literature\cite{Agashe:2014kda}. At the end of the data reduction, 12121 $\eta^{\prime}$ events are available for asymmetry determination with a residual background, estimated through simulation and mainly due to non-resonant multi-meson photoproduction, of less than 4\%. As the recoil proton angles are the best measured ones, the production angle of the meson in the center-of-mass frame $\theta^{\eta^{\prime}}_{c.m.}$ was calculated from the relevant proton angle $\theta^{p}_{c.m.}$. The angular resolution for $\theta^{p}_{c.m.}$ obtained with this procedure was $\simeq 2^{\circ}$ and no kinematical fit was used to improve it.
The selected $\eta^{\prime}$ events were grouped into two energy bins (the first bin is [1.447, 1.475] GeV with centroid 1.461 GeV; the second, with centroid 1.480 GeV, is [1.475, 1.490] GeV), seven angular bins for $\theta^{\eta^{\prime}}_{c.m.}$, and eight for the azimut angle $\phi$.
The beam asymmetry
$\Sigma(E_{\gamma}, \theta^{\eta ^{\prime}}_{c.m.})$ can be calculated by fitting the distribution defined by the following ratio:
\[\frac{N_V/F_V}{N_V/F_V + N_H/F_H} = \frac{1}{2}[1+P(E_\gamma)\cdot \Sigma \cdot cos(2\phi)]\]
\newline
where $N_V$ ($N_H$) and $F_V$ ($F_H$) are the number of events and the total $\gamma$ flux for vertical (horizontal) polarisation states and $P(E_\gamma)$ is the calculated degree of polarisation. Since the kinematics are the same for $H$ and $V$ photons, as is the photon energy distribution, this procedure significantly decreases the systematic errors of the extracted asymmetries, by minimizing the effect of the detection and reconstruction efficiencies. In Fig. \ref{Fig2} we give an example of this azimuthal distribution with the performed fit.
\begin{figure}[hbt]
\begin{center}\vspace{-0cm}
\resizebox{0.45\textwidth}{!}{\includegraphics{fig2.eps}}\vspace{-0.1cm}
\caption{(colour online) Azimuthal distribution at $E_\gamma$ = 1.461 GeV and $\theta^{\eta\prime}_{c.m.}$
= 41.82$^{\circ}$. \label{Fig2}}\vspace{-0.6cm}
\end{center}
\end{figure}
Two sources of systematic errors were considered: i) the possible deterioration of the laser light polarisation on the laser focusing system, with slightly different beam profiles on the target for each polarisation state, and ii) the residual hadronic background. The first error is characteristic of the GrAAL experiment and was established at $\Delta\Sigma$ = 0.02\cite{Bartalini:2007fg}. The second was estimated through determination of two large bins in $\theta^{\eta^{\prime}}_{c.m.}$ ([10,80]$^{\circ}$ and [100,170]$^{\circ}$), and extraction of asymmetry values from events in the peak of the distribution in Fig. \ref{Fig1}(f) $vs.$ the events belonging to the tails of the same distribution. Peak and tail regions were chosen so that they contain approximately the same number of events, and the results were fairly consistent, with a small decrease in the absolute value of $\Delta\Sigma \sim 0.01$ for the events in the tail regions. Moreover, a MonteCarlo closure test was performed, with trial asymmetry closely reproduced\cite{Mandaglio}. We therefore assumed a total systematic uncertainty $\Delta\Sigma$ = 0.03.
The stability of the results was verified in three alternate ways: i) extraction of the asymmetry with the same large bins in $\theta^{\eta^{\prime}}_{c.m.}$ separately for different stretches of the experiment;
ii) modification of the angular binning, and iii) separate analysis of the subsets of events resulting from neutral or charged decay modes. In all cases, results were satisfactorily stable\cite{Mandaglio}.
\begin{figure}[h]
\begin{center}\vspace{-0.2cm}
\resizebox{0.45\textwidth}{!}{\includegraphics{fig3.eps}}
\caption{(colour online) $\Sigma$ beam asymmetry for $\eta$ photoproduction obtained in this work (full dots) at $E_{\gamma}$ = 0.762 GeV compared with the results of \cite{Bartalini:2007fg} at $E_{\gamma}$ = 0.761 GeV (open triangles).
\label{Fig3}}\vspace{-0.8cm}
\end{center}
\end{figure}
Finally, with the same data set and the same analysis procedure, we extracted the events of the $\eta$ photoproduction process just above the threshold and compared the results with those of \cite{Bartalini:2007fg}. This is a particularly significant test, as the final state detected is exactly the same as with $\eta^\prime$ e.g. $2\gamma$, $6\gamma$ and $2\gamma\pi^{+}\pi^{-}$. Moreover, the data sample in this letter is different from \cite{Bartalini:2007fg}, as was the analysis procedure: in our previous work, only neutral decay channels were considered. The results are shown in Fig. \ref{Fig3} where we compare the values of the asymmetry extracted in this work at 0.762 GeV with the previous GrAAL results at 0.761 GeV. As one can see, the agreement is excellent.
\begin{figure}[hbt]
\begin{center}
\vspace{-0.4cm}
\resizebox{0.4\textwidth}{!}{\hspace{-1cm}\includegraphics{fig4u.eps}}\vspace{-0.cm}
\resizebox{0.4\textwidth}{!}{\hspace{-1cm}\includegraphics{fig4l.eps}}
\caption{(colour online) $\Sigma$ beam asymmetry at the incoming photon energies of 1.461 and 1.480 GeV
(corresponding to a total center-of-mass energy W of 1.903 and 1.912 GeV respectively)
as a function of the meson production angle in the center-of-mass system compared to theoretical calculations: red dotted line\cite{Chiang:2002vq}, blue dashed line\cite{Huang:2012xj} green dot-dashed\cite{Tryasuchev:2013fua}, orange long-dashed\cite{Zhong:2011ht}. The solid black line is the result of a fit performed with a function $f(\theta) = a \cdot sin^2(\theta) cos(\theta)$. The fit results for the free parameter are: $ a=0.321\pm 0.063$ at 1.461 GeV and $ a=0.096\pm0.051$ at 1.480 GeV. \label{Fig4}}\vspace{-0.6cm}
\end{center}
\end{figure}
The final results of the beam asymmetry $\Sigma$ for the $\eta^{\prime}$ photoproduction process are summarized in Fig. \ref{Fig4} together with the calculations of \cite{Huang:2012xj,Chiang:2002vq,Zhong:2011ht,Tryasuchev:2013fua}. As one can see, the asymmetry is positive at forward angles and negative at backward angles. Moreover, the data indicate a quite strong energy dependence, the effect being more evident at 1.461 GeV, closer to threshold. This behaviour is compatible with a $\sim sin^2(\theta^{\eta\prime}_{c.m.})cos(\theta^{\eta\prime}_{c.m.})$ function, typical of a P-wave D-wave (S-wave F-wave) interference\cite{Drechsel:1992pn,Sandorfi:2010uv}. The existing calculations, whilst providing a reasonable description of the measured cross section, cannot however reproduce these data, especially in the first energy bin ($E_{\gamma}$ = 1.461 GeV corresponding to a total center-of-mass energy of 1.903 GeV) where a change of sign in the asymmetry values around 90$^{\circ}$ for the meson center-of-mass production angle is clearly visible. A slightly better, but still not satisfactory, agreement between data and calculation is obtained at forward angles and at the highest energy bin ($E_{\gamma}$ = 1.480 GeV corresponding to a total center-of-mass energy of 1.912 GeV) in \cite{Huang:2012xj,Zhong:2011ht}. We must notice that the theoretical curves presented here are the result of interpolations of the existing models at low energies, and that none of these models contains D-wave or F-wave contributions. It is also important to underline that, in contrast with the conclusions of
\cite{Crede:2009zzb} for higher energies, at threshold the dynamics of $\eta$ and of $\eta^\prime$ photoproduction processes are clearly different.
These results prove once again that the polarisation degrees of freedom play an essential role in accessing the details of the interaction, and can lead to a better determination of the partial wave contributions and to a better comprehension of the reaction mechanism.
In conclusion, the $\Sigma$ beam asymmetry in the $\eta^{\prime}$ photoproduction was measured at the incoming photon energies of 1.461 and 1.480 GeV by using the highly linearly polarised GrAAL photon beam and the large solid angle LAGRAN$\gamma$E detector. This is the first measurement of this observable for this reaction. The values obtained indicate a P-wave D-wave (S-wave F-wave) interference, the closer to threshold the stronger. Available calculations fail to reproduce the observed behaviour, regardless of the intermediate resonance states involved in the models.
From the experimental point of view, new measurements with a finer energy binning as well as an extended energy range, would be highly desirable.
\\ \\
The authors would like to thank L. Tiator, F. Huang, H. Haberzettl, K. Nakayama, X-H Zhong, Q. Zhao and V. Tryasuchev for kindly providing the results of their models at the energies of this paper, and for helpful discussions. It is a pleasure to thank the ESRF for the reliable and stable operation of the storage ring and the technical staff of the contributing institutions for essential help in the realisation and maintenance of the apparatus.
|
\section{Introduction}
For the general background on matroid and oriented matroid theory we refer the reader to \cite{Ox-99} and \cite{Bjo-99}, respectively.
An (oriented) matroid is a finite ground set together with a (usually large) set of (oriented) circuits satisfying certain axioms. {But, how many of these circuits are actually needed to fully describe a given (oriented) matroid ?}
In~\cite[page 721]{Leh-64} Lehman shows, that the set $\mathrm{S}_e$ of circuits of a connected matroid $\mathrm{M}$ containing a fixed element $e$, distinguishes $\mathrm{M}$ from all other matroids on the same ground set, that is, if a matroid $\mathrm{M}'$ on the same ground contains all circuits from $\mathrm{S}_e$ then $\mathrm{M}$ and $\mathrm{M}'$ are the same.
\smallskip
Las Vergnas and Hamidoune~\cite{Ham-86} extend Lehman's result to {an oriented version}. They prove that a connected oriented matroid $\mathcal{M}$ is uniquely determined by the collection of signed circuits $\S_e$ containing a given element $e$, i.e., if an oriented matroid $\mathcal{M}'$ on the same ground set as $\mathcal{M}$ contains all circuits from $\S_e$ then $\mathcal{M}$ and $\mathcal{M}'$ are the same.
In view of Las Vergnas and Hamidoune's result, one may ask the following natural question:
\begin{quote}
{\em How many circuits are needed to determine a connected oriented matroid?}
\end{quote}
\subsection{Scope/general interest} It turns out that the above question can be interpreted in different ways. In this paper, we will investigate the number of circuits needed to determine an oriented matroid among all oriented matroids with the same underlying matroid. Let us introduce some notation in order to explain this more precisely. Generally, we represent matroids and oriented matroids as pairs of a ground set and a set of (signed) circuits. Throughout this paper we use calligraphic letters for sets of signed circuits and oriented matroids and non-italic roman letters for the non-oriented case. We say that two matroids $\mathrm{M}_1=(E_1,\mathrm{C}_1)$ and $\mathrm{M}_2=(E_2,\mathrm{C}_2)$ are the \emph{same}, i.e., $\mathrm{M}_1=\mathrm{M}_2$ if and only if $E_1=E_2$ and $\mathrm{C}_1=\mathrm{C}_2$.
Note that the equality is more restrictive than \emph{isomorphism} even when restricted to the same ground set, where the latter means that there is a permutation of the ground set which preserves circuits. This is illustrated in the following
\begin{example}\label{exmpl:isomorphic}
Let $\mathrm{M}(G_1),\mathrm{M}(G_2)$ and $\mathrm{M}(G_3)$ be the graphic matroids associated to the graphs $G_1,G_2$ and $G_3$ given in Figure~\ref{fig:ex1}.
\begin{figure}[htb]
\includegraphics[width=.5\textwidth]{G123}
\caption{} \label{fig:ex1}
\end{figure}
We clearly have that $\mathrm{M}(G_1)=\mathrm{M}(G_2)$ since $$\mathrm{C}(\mathrm{M}(G_1))=\{\{1,2,5\},\{3,4,5\},\{1,2,3,4\}\}=\mathrm{C}(\mathrm{M}(G_2)).$$
However, although $\mathrm{M}(G_1)$ and $\mathrm{M}(G_3)$ are isomorphic by taking the permutation
$\pi(1)=5,\pi(2)=2,\pi(3)=3,\pi(4)=4$ and $\pi(5)=1$ we have that
$\mathrm{M}(G_1)\neq \mathrm{M}(G_3)$ since $$\mathrm{C}(\mathrm{M}(G_1))=\{\{1,2,5\},\{3,4,5\},\{1,2,3,4\}\}\neq \{\{1,2,5\},\{1,3,4\},\{2,3,4,5\}\}=\mathrm{C}(\mathrm{M}(G_3)).$$
\end{example}
Similarly as for the non-oriented case, we say that two oriented matroids $\mathcal{M}_1=(E_1,\mathcal{C}_1)$ and $\mathcal{M}_2=(E_2,\mathcal{C}_2)$ are the {\em same}, i.e., $\mathcal{M}_1=\mathcal{M}_2$ if $E_1=E_2$ and $\mathcal{C}_1=\mathcal{C}_2$. For a signed set $\mathcal{X}$ we denote by $\underline{\mathcal{X}}$ its \emph{underlying unsigned set}. We extend this notation to sets of signed sets and furthermore denote by $\underline{\mathcal{M}}$ the \emph{underlying matroid} of the oriented matroid $\mathcal{M}$. In this case we say that $\mathcal{M}$ is an \emph{orientation} of $\underline{\mathcal{M}}$. For a subset of circuits $\mathrm{S}$ of a matroid $\mathrm{M}$ and an orientation $\mathcal{M}$ of $\mathrm{M}$ we denote by $\mathcal{S}_{\mathcal{M}}$ the (maximal) set of signed circuits of $\mathcal{M}$ such that $\underline{\mathcal{S}_{\mathcal{M}}}=\mathrm{S}$. We call $\mathcal{S}_{\mathcal{M}}$ the \emph{orientation of $\mathrm{S}$ corresponding to $\mathcal{M}$}.
\medskip
We say that a set $\S$ of signed circuits of $\mathcal{M}$ \emph{determines $\mathcal{M}$} if an orientation $\mathcal{M}'$ of $\mathrm{M}$ contains the set of signed circuits $\S$ if and only if $\mathcal{M}=\mathcal{M}'$. We say that a set of circuits $\mathrm{S}$ of $\mathrm{M}$ \emph{determines all orientations of $\mathrm{M}$} if for every orientation $\mathcal{M}$ of $\mathrm{M}$ the corresponding orientation $\S_{\mathcal{M}}$ of $\mathrm{S}$ determines $\mathcal{M}$.
\smallskip
The above mentioned result of Las Vergnas and Hamidoune can be restated as follows
\begin{eqnarray}\label{LVH1}
\text{the set $\mathrm{S}_e$ determines all orientations of $\mathrm{M}$.}
\end{eqnarray}
{In this spirit, we define three different quantities for a connected orientable matroid $\mathrm{M}$ in order to investigate the number of circuits needed to determine it.
\medskip
{Let {\boldmath $s(\mathrm{M})$} be the minimum size of a set $\mathrm{S}$ of circuits of $\mathrm{M}$ determining all orientations of $\mathrm{M}$. }
\medskip
We notice that $s(\mathrm{M})$ requires a {\em fixed} set of (non-oriented) circuits $\mathrm{S}$ that
will be used to determine any orientations of $\mathrm{M}$ ($\mathrm{S}$ is chosen independently of the orientation of $\mathrm{M}$). One may naturally ask whether the size of such a set of circuits can be improved for each fixed orientation.
\medskip
{Let {\boldmath $\widetilde{s}(\mathrm{M})$} be the smallest positive integer $k$ such that in any orientation $\mathcal{M}$ of $M$ there is a set $\S$ of signed circuits of $\mathcal{M}$ of size $k$ that determines $\mathcal{M}$.}
\medskip
{Finally, in this context, we introduce a third variant.}
\medskip
{Let {\boldmath $\overline{s}(\mathrm{M})$} be the smallest $k$ such that \emph{any} set $\mathrm{S}$ of circuits of $\mathrm{M}$ of size $k$ determines all orientations of $\mathrm{M}$.}
\smallskip
The parameter $s(\mathrm{M})$ has been investigated in~\cite{For-98,For-02,Cha-13} for uniform oriented matroids while $\widetilde{s}(\mathrm{M})$ has already been studied in~\cite{daS-07} in connection with a problem about realizability of rank $3$ matroids. As far as we are aware, {$\overline{s}$ has not been considered before. By the result of Las Vergnas and Hamidoune and simply by definition, respectively, we get
\begin{observation}\label{obs:three} Let $\mathrm{M}$ be a connected orientable matroid and $e$ an element. Then, $$s(\mathrm{M})\le |\mathrm{S}_e| \text{ and } \widetilde{s}(\mathrm{M})\leq s(\mathrm{M})\leq\overline{s}(\mathrm{M}).$$
\end{observation}
\subsection{Motivations and connections} The quantities $s(\mathrm{M}), \widetilde{s}(\mathrm{M})$ and $\overline{s}(\mathrm{M})$ are natural to be investigated. We expect that these quantities will provide useful and interesting insights in the following appealing and challenging subjects. We leave this for further future work.
\medskip
$\bullet$ The quantities $s(\mathrm{M}), \widetilde{s}(\mathrm{M})$ and $\overline{s}(\mathrm{M})$ may provide efficient ways for encoding oriented matroids by giving only a partial list of circuits. Counting the number of orientations of a given matroid and storing orientations of a matroid is non-trivial. The above quantities may help to simplify this process (and thus bounds on the values of $s(\mathrm{M}), \widetilde{s}(\mathrm{M})$ and $\overline{s}(\mathrm{M})$ would be worthwhile).
\medskip
$\bullet$ In \cite{For-98} the relation between $s(\mathrm{M})$ and both special {\em coverings} in Block Design Theory as well as Tur\'an systems (see Section \ref{sec:uniform}) is put forward and thus attractive to study on this context.
\medskip
$\bullet$ We shall see that some of our results do not rely on the Topological Representation Theorem for oriented matroids but only
on {\em invertible bases}. It might be of interest to investigate whether such results also hold in the wider context of {\em matroids with coefficients} \cite{Dress-91} or maybe for {\em complex matroids} \cite{And12}.
\medskip
$\bullet$ The reorientation classes of a matroid are an important concept in oriented matroid theory. They have a natural graphic (resp. geometric) interpretation for graphic (resp. representable) oriented matroids. In \cite{Gel-95} a characterization of reorientation classes of an oriented matroid by using {\em projective orientations} of $\mathrm{M}$ in terms of combinatorics of its circuits and cocircuits is given. This is of interest and significance in the study of stratifications of combinatorial Grassmannians. Therefore, it is attractive to understand the space of oriented matroids over a given matroid (and the quantities $s(\mathrm{M})$ and $\widetilde {s}(\mathrm{M})$).
\medskip
$\bullet$ An interesting class of oriented matroids are those having exactly one reorientation class. Indeed, it is known ~\cite{Bla-78} that any \emph{regular} oriented matroid has exactly one reorientation class. In~\cite{Rou-88} it is proved that regular matroids are characterized as those oriented matroids such that all restrictions have only one reorientation class. The following long standing conjecture is due to Las Vergnas ~\cite{Las-89}
\begin{conjecture}
The matroids $C_d$ obtained from the affine $d$-dimensional hypercube has a single reorientation class for all $d$.
\end{conjecture}
$C_d$ is a subclass of cubic matroids~\cite{daS-08}.
It is known \cite{Bok-96} that the conjecture is true when $d\leq 7$.
\medskip
Investigations on $s(\mathrm{M})$ and $\widetilde {s}(\mathrm{M})$ when $\mathrm{M}$ has one reorientation class deserve thus a particular attention. As noticed in Subsection \ref{subsec:reorie}
and at the beginning of Section~\ref{sec:regular} the quantities $s(\mathrm{M})$ and $\widetilde {s}(\mathrm{M})$ are closely related to the fact that determining a regular oriented matroid is equivalent to finding a
{\em connected covering} of its elements by circuits. The latter generalizes the widely studied problem of {\em circuit covers} in matroids~\cite{Sey-80,Lem-06,mcG-10} and {\em cycle covers} in graphs~\cite{Fan-02,Lai-03,Yan-11}.
\medskip
$\bullet$ The study of $\widetilde{s}(U_{r,n})$ naturally leads to consider the so-called {\em mutation} operation in uniform oriented matroids. A challenging question concerning mutations is the following famous simplex conjecture
of Las Vergnas
\begin{conjecture}
Every uniform oriented matroid has at least one mutation.
\end{conjecture}
This conjecture is known to be true only for \emph{realizable} oriented matroids~\cite{Sha-79}, oriented matroids of rank at most $3$~\cite{Lev-26}, and for rank $4$ oriented matroids with few elements~\cite{Bok-01}.
\subsection{Organization of the paper} In the next section, we recall some oriented matroid basics used throughout the paper.
\medskip
In Section~\ref{sec:gen-bound} we give general bounds by introducing the notion of {\em weak} and {\em connected covering} (Theorem~\ref{thm:generalbounds}). We present an upper bound for $\overline{s}(\mathrm{M})$ (Theorem~\ref{thm:generalboundssbar}) and also study the problem of determining an oriented matroid within its reorientation class (Theorem~\ref{thm:1reclass}).
\medskip
Section~\ref{sec:uniform} is devoted to the study of uniform oriented matroids. After recalling the relationship with Design Theory we recover an earlier results given in ~\cite{For-98} in a more general framework (Theorem~\ref{prop:motiv}).
We provide exact values for $\widetilde{s}(U_{n-2,n})$ with $n\ge 3$ (Theorem~\ref{Th:unif}) as well as a general lower bound for $\widetilde{s}(U_{n-r,n})$ with $3\leq r\leq n-2$ (Theorem~\ref{Th:unif1}). We finally present the exact value of $\overline{s}(U_{r,n})$ with $1\leq r\leq n-1$ (Theorem~\ref{thm:s(n,r)})
\medskip
In Section~\ref{sec:regular} we focus our attention to regular matroids. We first notice that if $\mathrm{M}$ is regular then $\widetilde{s}(\mathrm{M})=s(\mathrm{M})$ and both equal to the size of the smallest {\em connected element covering} of $\mathrm{M}$. We then give a different bounds for the latter in the case when the matroid $\mathrm{M}$ is regular (Lemma \ref{lem:makeconnected}, Theorem \ref{thm:edgevertexbound}) and in particular when $\mathrm{M}$ is graphic (Proposition \ref{prop:makeconnectedbestpossible}, Corollary \ref{cor:2conn}).
\medskip
Finally, in Section~\ref{sec:graph} we calculate the values for the graphic and cographic matroids associated to complete graphs and hypercubes. More precisely, we calculate
$\widetilde{s}(\mathrm{M})$ and $s(\mathrm{M})$ when
$\mathrm{M}$ is either $\mathrm{M}(K_n), \mathrm{M}^*(K_n), \mathrm{M}(Q_n)$ or $\mathrm{M}^*(Q_n)$ where $K_n$ is the complete graph on $n$ vertices and $Q_n$ is the hypercube graph of dimension $n$}
(Theorems~\ref{thm:Qn},~\ref{thm:Qn*},~\ref{thm:Kn} and~\ref{thm:Kn*}). These results show that some of our general bounds are tight.
\section{Basic definitions and general bounds}\label{sec:first}
Beside circuits we sometimes also use bases to represent a matroid. Given a basis $B$ of $\mathrm{M}$ and an element $e\notin B$, there is a unique circuit $C(B,e)$ of $\mathrm{M}$ contained in $B\cup\{e\}$ called the \emph{fundamental circuit} of $B$ with respect to $e$. In the oriented case two opposite orientations of this circuit appear. We denote by $C(B,e)$ any of them if no distinction is necessary.
A {\em basis orientation} of an oriented matroid $\mathcal{M}$ is a mapping $\chi_{\mathcal{M}}$ of the set of the ordered bases of $\mathrm{M}:=\underline{\mathcal{M}}$ to $\{-1,1\}$ satisfying the following properties :
\begin{itemize}
\item[(CH1)] $\chi_{\mathcal{M}}$ is alternating.
\item[(CH2)] for any two ordered bases $B, B'$ of $\mathrm{M}$ of the form $(e,b_2,\dots ,b_r)$ and $(e',b_2,\dots ,b_r)$, $e\neq e'$, we have $\chi_{\mathcal{M}}(e,b_2,\dots ,b_r)=-C(B',e)_{e'}C(B',e)_e\chi_{\mathcal{M}}(e',b_2,\dots b_r)$, where $C(B',e)_e$ and $C(B',e)_{e'}$ denote the sign corresponding to elements $e$ and $e'$ in $C(B',e)$ respectively.
\end{itemize}
We have that $\mathcal{M}_1=\mathcal{M}_2$ if an only if $\chi_{\mathcal{M}_1}=\pm \chi_{\mathcal{M}_2}$.
We say that a base $B'$ of an oriented matroid $\mathcal{M}$ with chirotope $\chi$ is {\em invertible} if
$$\chi^{B'}(B):=\left\{\begin{array}{rr}
-\chi_{\mathcal{M}}(B) & \hbox{if $B=B'$,}\\
\chi_{\mathcal{M}}(B) & \hbox{otherwise}\\
\end{array}\right.
$$
is also the chirotope of an oriented matroid $\mathcal{M}^{B'}$ (obtained thus from $\chi_{\mathcal{M}}$ by inverting only the sign of base $B'$). In the case of uniform oriented matroids invertible bases are called \emph{mutations}.
For every subset $A\subseteq E$ and every signed set $X$ of $E$, we denote by $_{\bar A}X$ the signed set obtained from $X$ by {\em reversing signs} on $A$, i.e., $(_{\bar A}X)^+=(X^+\setminus A)\cup (X^-\cap A)$ and
$(_{\bar A}X)^-=(X^-\setminus A)\cup (X^+\cap A)$. The set $\{_{\bar A}C \mid C\in\mathcal{C}\}$ is the set of signed circuits of an oriented matroid, denoted by $_{\bar A}\mathcal{M}$. Two oriented matroids $\mathcal{M}$ and $\mathcal{M}'$ are {\em related by sign-reversal} if $\mathcal{M}'=_{\bar A}\mathcal{M}$ for some $A\subseteq E$. The equivalence classes for this relation are called {\em reorientation classes}. Notice that our definition of reorientation classes differs from the definition that is common in the literature which applies to {\em unlabeled oriented matroids}, i.e., apart from sign-reversal also isomorphisms are allowed transformations.
\subsection{Topological representation : quick discussion}
The well-known Topological Representation Theorem due to Folkman and Lawrence~\cite{Fol-78} states that loop-free oriented matroids of rank $d+1$ (up to isomorphism) are in one-to-one correspondence with arrangements of pseudospheres in $S^{d}$ (up to topological equivalence) or equivalently to affine arrangements of pseudohyperplanes in
$\mathbb{R}^{d-1}$ (up to topological equivalence).
\medskip
Note that, as mentioned above, in the literature contrary to our definition, the term reorientation class is usually applied to \emph{unlabeled} oriented matroids. For instance the equivalence relation considered by the Topological Representation Theorem identifies two oriented matroids if they can be transformed via resignings, relabelings and reorientation into each other.
For example, $U_{2,n}$ has only one topological representation
but in our sense it admits exactly $\frac{(n-1)!}{2}$ reorientation classes~\cite{Cor-90}. Example~\ref{exmpl:U24} illustrates two of the three reorientation classes of $U_{2,4}$.
\section{General bounds}\label{sec:gen-bound}
We derive some necessary and sufficient conditions for a set of circuits $\mathrm{S}$ to determine all the orientations of $\mathrm{M}$ or to determine a specific $\mathcal{M}$. In order to do this, we introduce some matroid parameters which {will be used} as upper and lower bounds.
We say that a circuit $C$ of $\mathrm{M}$ \emph{covers} a basis $B$ if and only if there is an element $e\in E\setminus B$ such that $C$ is the fundamental circuit $C(B,e)$ of the basis $B$ with respect to $e$. A signed circuit $C$ of $\mathcal{M}$ \emph{covers} a basis $B$ of $\underline{\mathcal{M}}$ if $\underline{C}$ covers $B$.
\begin{proposition}\label{prop:orientedweak} Let $\mathcal{M}$ be an oriented matroid. If $\S$ determines $\mathcal{M}$ then $\S$ covers all invertible bases of $\mathcal{M}$.
\end{proposition}
\begin{proof}
Let $B$ be an invertible basis of $\mathcal{M}$ which is not covered by any signed circuit in $\S$. Let $\mathcal{M}'$ be the oriented matroid with chirotope $\chi^B$. By (CH2), the orientation of $\underline{\S}$ in $\mathcal{M}'$ depends only on signs of bases covered by $\underline{\S}$. Thus, the set $\S$ is a subset of the set of signed circuits of $\mathcal{M}'$. Therefore, $\S$ does not determine $\mathcal{M}$.
\end{proof}
We say that a set $\S$ of signed circuits of $\mathcal{M}$ is a \emph{weak covering} of $\mathcal{M}$ if it covers all the invertible bases of $\mathcal{M}$. Let {\boldmath $\widetilde{\mathrm{WC}}(\mathrm{M})$} be the smallest $k$ such that in each orientation $\mathcal{M}$ of $\mathrm{M}$ there is a weak covering of size $k$. Analogously, a set $\mathrm{S}$ of circuits $\mathrm{M}$ is called \emph{weak covering} of $\mathrm{M}$ if its orientation $\S_{\mathcal{M}}$ in any orientation $\mathcal{M}$ of $\mathrm{M}$ is a weak covering of $\mathcal{M}$. Let {\boldmath $\mathrm{WC}(\mathrm{M})$} be the size of a smallest weak covering of $\mathrm{M}$. The following results are an immediate consequence of Proposition~\ref{prop:orientedweak}.
\begin{corollary}\label{cor:weak}
For any orientable matroid $\mathrm{M}$ we have $\widetilde{\mathrm{WC}}(\mathrm{M})\leq \widetilde{s}(\mathrm{M})$ and $\mathrm{WC}(\mathrm{M})\leq s(\mathrm{M})$.
\end{corollary}
Given a set $\mathrm{S}$ of circuits of $\mathrm{M}$ define the graph $B_{\mathrm{S}}$ with vertex set the set of bases of $\mathrm{M}$ where $B, B'$ are adjacent if and only if $|B\Delta B'|=2$ and there is $C\in\mathrm{S}$ such that $C\subseteq B\cup B'$.
Note that $B_{\mathrm{C}(M)}$ is just the base graph of $\mathrm{M}$.
A {\em base covering} is a set of circuits covering all the bases of $M$. A base covering $\mathrm{S}$ of $\mathrm{M}$ is called \emph{connected} if the graph $B_{\mathrm{S}}$ is connected.
Let {\boldmath $\mathrm{CC}(\mathrm{M})$} be the size of a smallest connected base covering of $\mathrm{M}$.
\begin{theorem}\label{thm:generalbounds}
For every connected orientable matroid $\mathrm{M}$ we have $s(\mathrm{M})\leq \mathrm{CC}(\mathrm{M})$.
\end{theorem}
\begin{proof}
Let $\mathrm{S}$ be a connected base covering of $\mathrm{M}$. Let $B$ be a basis and $B'$ a neighbor of $B$ in $B_{\mathrm{S}}$, i.e., $|B\Delta B'|=2$ and there is $C\in\mathrm{S}$ such that $C\subseteq B\cup B'$. This means that there are $f\in B'$ and $e\in B$ such that $C=C(B',e)=C(B,f)$. By fixing the orientation of $\chi(B)$ using the signs of $C$ the orientation of $B'$ is determined via (CH2).
Therefore, if $\mathrm{S}$ is a connected base covering the choice of
the value for $\chi(B)$ as well as the signings of the circuits in $\mathrm{S}$ induce a unique oriented matroid. Moreover, both choices for $\chi(B)=1$ or $=-1$ determine the same oriented matroids, with opposite chirotopes.
\end{proof}
Let $\lambda(\mathrm{M})$ be the largest $k$ such that for any set $\mathrm{S}$ of $k-1$ circuits the graph $B_{\mathrm{C}\setminus\mathrm{S}}$ is connected. We denote by $r(\mathrm{M})$ the rank of $\mathrm{M}$.
\begin{theorem}\label{thm:generalboundssbar}
For every connected orientable matroid $\mathrm{M}=(E,\mathrm{C})$ we have $$\overline{s}(\mathrm{M})\leq |\mathrm{C}|+1-\min(\lambda(\mathrm{M}), |E|-r(\mathrm{M})).$$ If $\mathrm{M}$ has a base which is invertible in some orientation, then $|\mathrm{C}|+1-|E|+r(\mathrm{M})\leq\overline{s}(\mathrm{M})$.
\end{theorem}
\begin{proof}
We start by proving the first inequality. If $\mathrm{S}\subseteq\mathrm{C}$ has size $|\mathrm{C}|+1-\min(\lambda(\mathrm{M}), |E|-r(\mathrm{M}))$, then $\mathrm{C}-\mathrm{S}$ has cardinality $\min(\lambda(\mathrm{M}), |E|-r(\mathrm{M}))-1$. Therefore removing $\mathrm{C}-\mathrm{S}$ cannot disconnect $B_{\mathrm{C}}$, i.e., $B_{\mathrm{S}}$ is connected. Removing $\mathrm{C}-\mathrm{S}$ leaves no one basis uncovered, since each base is covered by exactly $|E|-r(\mathrm{M})$ circuits. Thus, $\mathrm{S}$ is a connected base covering of $\mathrm{M}$ and the result follows by Theorem~\ref{thm:generalbounds}.
For the second bound let $B$ be an invertible basis of $\mathrm{M}$. Now choosing $\mathrm{S}$ as all circuits except those covering $B$ yields a set of size $|\mathrm{C}|-|E|+r(\mathrm{M})$ which is not a weak covering of $\mathrm{M}$. The result follows by Corollary~\ref{cor:weak}.
\end{proof}
Indeed, we believe that the minimum in the upper bound in Theorem~\ref{thm:generalboundssbar} is always attained by $|E|-r(\mathrm{M})$. We will see this for uniform oriented matroids in Theorem~\ref{thm:s(n,r)}.
\subsection{Oriented matroids with one reorientation class}\label{subsec:reorie}
{In this subsection we give sufficient and necessary conditions for a set of circuits to determine a matroid \emph{within a given reorientation class}. More precisely, suppose that the oriented matroids, coinciding on a given set of circuits, lie in the same reorientation class. We shall study conditions yielding to a unique orientation in this class.
Although our main results in this subsection (Theorems \ref{thm:1reclass} and \ref{thm:1reclasssbar}) are stated in terms of matroids having a single reorientation class they yield to results for general matroids (Corollaries \ref{cor:generallb} and \ref{cor:generallbbar}).}
\smallskip
Given a matroid $\mathrm{M}$ an \emph{(element) covering} is a set $\mathrm{S}$ of circuits covering the ground set $E$.
An element covering $\mathrm{S}$ is said to be \emph{connected} if the \emph{(element) intersection graph} $I_{\mathrm{S}}$ of $\mathrm{S}$ is connected. Let {\boldmath $\mathrm{cc}(\mathrm{M})$} be the size of a smallest connected element covering of $\mathrm{M}$.
\begin{theorem}\label{thm:1reclass}
If a connected matroid $\mathrm{M}$ has a single reorientation-class, then we have $$\widetilde{s}(\mathrm{M})=s(\mathrm{M})=\mathrm{cc}(\mathrm{M}).$$
Moreover, these equalities are attained by the same fixed element covering $\mathrm{S}$ of $\mathrm{M}$.
\end{theorem}
\begin{proof}
We first show $\mathrm{cc}(\mathrm{M})\geq s(\mathrm{M})$. Let $\mathrm{S}$ cover $E$ and $I_{\mathrm{S}}$ be connected. Suppose there were two orientations $\mathcal{M}$ and $\mathcal{M}'$ of $\mathrm{M}$ coinciding on $\mathrm{S}$. By the preconditions $\mathcal{M}$ and $\mathcal{M}'$ differ by reorienting a set $X\subset E$.
We reorient $X$ in $\mathcal{M}$, but since all orientations of circuits of $\mathrm{S}$ shall be maintained, every circuit $C\in \mathrm{S}$ intersecting $X$ has to be reoriented entirely, i.e, $C\subseteq X$. Therefore all neighbors of $C$ in $I_{\mathrm{S}}$ are also contained in $X$. Iterating this argument \emph{all} circuits in $\mathrm{S}$ have to be completely reoriented. Since $\mathrm{S}$ covers $E$ all elements have to be reoriented, i.e., $X=E$. Thus, $\mathcal{M}=\mathcal{M}'$.
\medskip
We now show $\widetilde{s}(\mathrm{M})\geq\mathrm{cc}(\mathrm{M})$. If $\mathrm{S}$ does not cover some $e\in E$, then in any orientation $\mathcal{M}$ of $\mathrm{M}$ we can reorient $e$ independently of the rest, i.e., $\mathcal{M}$ and the reorientation of $\mathcal{M}$ at $e$ coincide on $\mathrm{S}$.
If $I_{\mathrm{S}}$ has two connected components corresponding to two sets of circuits $\mathrm{S}', \mathrm{S}''$, then in any orientation $\mathcal{M}$ of $\mathrm{M}$ we can reorient all elements covered by $\mathrm{S}'$. Since all signs of signed circuits in $\S'$ are reversed, the resulting orientation $\mathcal{M}'$ coincides with $\mathcal{M}$ on $\mathrm{S}'$ and thus on $\mathrm{S}$. Nevertheless, reorienting $\S'$ in particular changes the orientation of circuits containing an element covered by $\mathrm{S}'$ and one covered by $\mathrm{S}''$. Therefore $\mathcal{M}'\neq \mathcal{M}$,
Hence, if $\mathrm{S}$ is not a covering or $I_{\mathrm{S}}$ is disconnected then \emph{no} orientation $\mathcal{M}$ of $\mathrm{M}$ is determined by $\S_{\mathcal{M}}$. The result follows by Observation~\ref{obs:three}.
\end{proof}
The following result is an immediate consequence of Theorem~\ref{thm:1reclass} and gives an alternative lower bound for $s(\mathrm{M})$ to the one presented in Corollary~\ref{cor:weak}.
\begin{corollary}\label{cor:generallb}
For any connected orientable matroid $\mathrm{M}$ we have $\mathrm{cc}(\mathrm{M})\leq \widetilde{s}(\mathrm{M})$.
\end{corollary}
Theorem~\ref{thm:1reclass} allows us to say something about $\overline{s}$ for matroids with only one reorientation class. In order to prove the next result, we need the following definition. For a matroid $\mathrm{M}$ with set of circuits $\mathrm{C}$, denote by $\kappa(I_{\mathrm{C}})$ the vertex connectivity of the graph $I_{\mathrm{C}}$. If $\mathrm{S}_e$ is the set of circuits containing a given element $e$ of the ground set $E$, we set $\Delta(\mathrm{M}):=\max\{|\mathrm{S}_e| | e\in E\}$.
\begin{theorem}\label{thm:1reclasssbar}
Let $\mathrm{M}$ be a connected matroid with a single reorientation class, then
$$ \overline{s}(\mathrm{M})=|\mathrm{C}|+1-\min(\Delta(\mathrm{M}), \kappa(I_{\mathrm{C}})).
$$
\end{theorem}
\begin{proof}00
For any set $\mathrm{S}$ of circuits of $\mathrm{M}$ with $|\mathrm{S}|>|\mathrm{C}|-\Delta(\mathrm{M})$, we have that $\mathrm{S}$ is an element covering. Similarly, for any set $\mathrm{S}$ of circuits of $\mathrm{M}$ with $|\mathrm{S}|>|\mathrm{C}|-\kappa(I_{\mathrm{C}})$, the induced subgraph $I_{\mathrm{S}}$ of $I_{\mathrm{C}}$ is connected, otherwise there would be set $X\subset V(I_{\mathrm{S}})$ with $I_{\mathrm{S}}-X$ not connected and $|X|<\kappa(I_{\mathrm{C}})$ which is impossible. Thus, $\mathrm{S}$ is a connected element covering and with Theorem~\ref{thm:1reclass} we have $\overline{s}(\mathrm{M})\le |\mathrm{C}|+1-\min(\Delta(\mathrm{M}), \kappa(I_{\mathrm{C}}))$.
\medskip
On the other hand, we note that there exists a set $\mathrm{S}$ of circuits of $\mathrm{M}$ with $|\mathrm{S}|\le |\mathrm{C}|-\Delta(\mathrm{M})$ such that $\mathrm{S}$ is not an element covering. Similarly, there exists a set $\mathrm{S}$ of circuits of $\mathrm{M}$ with $|\mathrm{S}|\le |\mathrm{C}|-\kappa(I_{\mathrm{C}})$ such that the induced subgraph $I_{\mathrm{S}}$ of $I_{\mathrm{C}}$ is not connected. Hence with Theorem~\ref{thm:1reclass} $\overline{s}(\mathrm{M})\geq|\mathrm{C}|+1-\min(\Delta(\mathrm{M}), \kappa(I_{\mathrm{C}}))$.
\medskip
Together we get $\overline{s}(\mathrm{M})=|\mathrm{C}|+1-\min(\Delta(\mathrm{M}), \kappa(I_{\mathrm{C}}))$.
\end{proof}
The following result is an immediate consequence of Theorem~\ref{thm:1reclasssbar} and complements the upper bound in Theorem~\ref{thm:generalboundssbar}.
\begin{corollary}\label{cor:generallbbar}
For any connected orientable matroid $\mathrm{M}$ we have $|\mathrm{C}|+1-\min(\Delta(\mathrm{M}), \kappa(I_{\mathrm{C}}))\leq \overline{s}(\mathrm{M})$.
\end{corollary}
\section{Uniform Oriented Matroids}\label{sec:uniform}
Let us quickly describe the connection of $s(U_{r,n})$ with connected coverings.
\smallskip
Let $n,k,r$ be positive integers such that $n\ge k\ge r\ge 1$. An \textit{$(n,k,r)$-covering} is a family $\mathrm{B}$ of $k$-subsets of $\{1,\ldots,n\}$, called \textit{blocks}, such that each $r$-subset of $\{1,\ldots,n\}$ is contained in at least one of the blocks. The number of blocks is the covering's \textit{size}. The minimum size of such a covering is called the \textit{covering number} and is denoted by {\boldmath $\mathrm{C}(n,k,r)$}. Given an $(n,k,r)$-covering $\mathrm{B}$, its graph $G(\mathrm{B})$ has $\mathrm{B}$ as vertices and two vertices are joined if they have one $r$-subset in common. We say that an $(n,k,r)$-covering is \textit{connected} if the graph $G(\mathrm{B})$ is connected. The minimum size of a connected $(n,k,r)$-covering is called the \textit{connected covering number} and is denoted by {\boldmath $\mathrm{CC}(n,k,r)$}.
\begin{theorem}[\cite{For-98,For-02}]\label{prop:motiv}
$$\mathrm{C}(n,r+1,r)\leq s(U_{r,n})\leq \mathrm{CC}(n,r+1,r).$$
\end{theorem}
In~\cite{For-02} a \emph{disconnected} covering determining all orientations of a uniform matroid is presented. However, its size is larger than the size of a smallest connected covering.
\smallskip
We quickly recall some oriented matroids facts needed in the rest of this section. Let $\mathcal{M}$ be a uniform oriented matroid and let $\mathcal{A}_{\mathcal{M}^*}$ be the pseudosphere arrangement representing the dual oriented matroid $\mathcal{M}^*$ of $\mathcal{M}$. The signed circuits $\mathcal{C}$ of $\mathcal{M}$ correspond to the cocircuits of $\mathcal{M}^*$ which are represented by the set of vertices ($0$-dimensional cells) of the arrangement $\mathcal{A}_{\mathcal{M}^*}$. A pair of oppositely signed circuits of $\mathcal{M}$ corresponds to an $S_0$ in $\mathcal{A}_{\mathcal{M}^*}$. Let $R_{B^*}$ be a simplicial cell of codimension $1$ in $\mathcal{A}_{\mathcal{M}^*}$ where $B^*$ is a base of $\mathcal{M}^*$ whose elements correspond to the bounding pseudospheres of $R_{B^*}$. We notice that any of the circuits corresponding to the vertices of $R_{B^*}$ in
$\mathcal{A}_{\mathcal{M}^*}$ are circuits in $\mathcal{M}$ containing the base $B=E\setminus B^*$ because $\mathrm{M}$ is uniform. To see the latter, notice that the underlying set of each such circuit is formed by the pseudospheres not touching the corresponding vertex and so all the elements of $B$ will be included in such circuits. Finally, it is known that the mutations of $\mathcal{M}$ correspond to those bases corresponding to simplicial cells \cite{Rou-Sturm88}. Thus, in this section using Proposition~\ref{prop:orientedweak} we will encounter the problem of finding circuits touching all simplicial cells in an arrangement in order to obtain a weak covering of $\mathcal{M}$.
\medskip
Let us give an alternative proof of Theorem \ref{prop:motiv} in a more general framework.
\begin{proof}[Proof of Theorem \ref{prop:motiv}]
We shall show that $\mathrm{C}(n,r+1,r)=\mathrm{WC}(U_{r,n})$ and $\mathrm{CC}(n,r+1,r)=\mathrm{CC}(U_{r,n})$. The claimed inequalities then follow by Theorem~\ref{thm:generalbounds} and Corollary~\ref{cor:weak}.
First, note that the fundamental circuits of a base $B$ of $U_{r,n}$ are precisely the $(r+1)$-element sets containing $B$. Therefore the notions of $(n,r+1,r)$-covering and base covering of $U_{r,n}$ are the same.
For the first equality, it is enough to observe that for any base $B$ of $U_{r,n}$, there is an orientation with $B$ being invertible, i.e., a mutation and so the result will follow by Proposition~\ref{prop:orientedweak}.
So, let us take an $(n-r)$-simplex $R$ in $\mathbb{R}^{n-r}$. Define an affine hyperplane arrangement $\mathcal {A}_B$ consisting of the bounding hyperplanes of $R$ and $r$ further hyperplanes not intersecting $R$. We can label the bounding hyperplanes of $R$ with the elements of $B$. Since $R$ is a simplicial region of the arrangement $\mathcal {A}_B$, $B$ will be a mutation in any orientation of the hyperplanes of $\mathcal {A}_B$.
\smallskip
For the second equality we have to show that for a base covering $\mathrm{S}$ of $U_{r,n}$ we have that $G(\mathrm{\mathrm{S}})$ is connected if and only if $B_{\mathrm{S}}$ is connected. The crucial observation is that in $U_{r,n}$ a circuit $C$ covers $B$ and $B'$ if and only if $C=B\cup B'$. Therefore, there is a path from $C$ to $C'$ in $G(\mathrm{\mathrm{S}})$ if and only if there is a path from $B$ to $B'$ in $B_{\mathrm{S}}$ for all $B, B'$ covered by $C, C'$, respectively. Since $\mathrm{S}$ is a base covering, we obtain the result.
\end{proof}
\smallskip
Notice that by Observation~\ref{obs:three} we have $\widetilde{s}(\mathrm{M})\le s(\mathrm{M})$ and that Theorem~\ref{thm:1reclass} shows that both parameters are equal if the matroid has a single reorientation-class. It turns out that the inequality is strict for infinitely many matroids. Indeed, by~Theorem \ref{prop:motiv} and the fact that $\mathrm{C}(n,n-1,n-2)=\mathrm{CC}(n,n-1,n-2)=n-1$~\cite{Cha-13} we have that $s(U_{n-2,n})=n-1$ for every $n\ge 3$. {On the other hand, the following result shows that $\widetilde{s}(U_{n-2,n})$ is different from $s(U_{n-2,n})$ in general.}
\begin{theorem}\label{Th:unif} Let $n\ge 3$ be an integer. Then,
$\widetilde{s}(U_{n-2,n})=\lceil\frac{n}{2}\rceil$.
\end{theorem}
\begin{proof}
We start by proving $\widetilde{s}(U_{n-2,n})\le \lceil\frac{n}{2}\rceil$. Let $\mathcal{M}$ be a uniform oriented matroid
of rank $n-2$ and let $\mathcal{A}$ be the topological representation of its dual. This
is, $\mathcal{A}$ is an arrangement of oriented pairs of antipodal points on a circle, i.e., several copies of $S^0$,
on an $S^1$ each dividing $S^1$ in a positive and a negative half. Each point corresponds to a signed circuit of $\mathcal{M}$. The complement
of each edge, i.e., complement of a closed segment of $S^1$ between two
consecutive points, corresponds to a basis of $\mathcal{M}$. We will consider the following set $\S$ of signed circuits of $\mathcal{M}$. We choose points
from $\mathcal{A}$ to be part of $\S$ in an alternating way around $S^1$
starting at any point and continuing until $\S':=\S\cup -\S$ covers all edges. Clearly, $|\S|=\lceil\frac{n}{2}\rceil$. We prove that $\S$
determines $\mathcal{M}$, i.e., there is a unique arrangement $\mathcal{A}$ of
$n$ antipodal pairs yielding $\S$.
Clearly, $\S$ gives that also $-\S$ are circuits. So, let us show that $\S':=\S\cup -\S$ determines $\mathcal{M}$.
Take $\mathcal{A}'$ to be any arrangement having signed circuits $\S'$. First, observe that the subarrangement obtained by restricting to $\S'$ coincides with the restriction of $\mathcal{A}$ to $\S'$. (Both are representations of the same oriented matroid, corresponding to the restriction to the elements corresponding to $\S'$.)
Now, note that the signs
in $\S'$ determine the relative position of any point to points in $S'$. But
since $\S'$ covers all edges of $\mathcal{A}$ the relative position of a point
not in $\S'$ is between a unique pair of consecutive points of $\S'$ and no other
point is between them. Hence $\mathcal{A}'=\mathcal{A}$.
\smallskip
We now show that $\widetilde{s}(U_{n-2,n})\ge \lceil\frac{n}{2}\rceil$. We assume $|\S|<\lceil\frac{n}{2}\rceil$, then one edge of $\mathcal{A}$ is not incident to any element of $\S\cup -\S$. The oriented matroid arising by changing the order of the two copies of $S_0$ incident to that edge has different signs on the corresponding circuits, but does not differ on $\S$. This is a special case of Proposition~\ref{prop:orientedweak}.
\end{proof}
\begin{example}\label{exmpl:U24} Let $\mathcal{M}_1$ and $\mathcal{M}_2$ be the orientations of $U_{2,4}$ which are the duals of the oriented matroids $\mathcal{M}'_1$ and $\mathcal{M}'_2$ induced by the topological representations given in Figure~\ref{fig:ex3}.
\begin{figure}[htb]
\includegraphics[width=.7\textwidth]{exemp-corank2new.pdf}
\caption{Two 1-dimensional oriented arrangements representing two orientations of $U_{2,4}$.} \label{fig:ex3}
\end{figure}
We clearly have that $\underline{\mathcal{M}_1}=\underline{\mathcal{M}_2}$
since the set of circuits of $\underline{\mathcal{M}_1}$ and $\underline{\mathcal{M}_2}$ coincides. However, $\mathcal{M}_1\neq\mathcal{M}_2$ since for their sets of signed circuits $\mathcal{C}_1,\mathcal{C}_2$ we have:
$$\mathcal{C}_1=\{\{1,2,\bar{3}\},\{1,2,\bar{4}\},\{1,3,\bar{4}\},\{\bar{2},3,\bar{4}\}\}\neq \{\{1,\bar{2},3\},\{1,2,\bar{4}\},\{1,3,\bar{4}\},\{2,\bar{3},\bar{4}\}\}=\mathcal{C}_2.$$
Note that $1<\widetilde{s}(4,2)$. We may suppose that the circuit that had been chosen to determine $U_{2,4}$ was $\S=\{1,2,\bar{4}\}$ which clearly does not determine $U_{2,4}$ since $\{1,2,\bar{4}\}$ is a signed circuit of of $\mathcal{M}_1$ and $\mathcal{M}_2$.
Finally, it can be checked that there is no $A\subseteq \{1,2,3,4\}$ such that $_{\bar A}\mathcal{M}_1=\mathcal{M}_2$.
\end{example}
\begin{theorem}\label{Th:unif1}
Let $3\leq r\leq n-2$. We have $(\frac{1}{2}(\lfloor \frac{n}{r-1}\rfloor+1))^{r-1}\leq\widetilde{s}(U_{n-r,n})$.
\end{theorem}
\begin{proof}
We define a simple affine pseudo-hyperplane arrangement in $\mathbb{R}^{r-1}$ in which almost every vertex is contained in exactly one simplex. Start with the \emph{grid}, i.e., the set of translates of coordinate hyperplanes $\mathcal{H}:=(H_i^k)_{i\in [r-1],k\in[\ell]}$ where $H_i^k:=\{x\in\mathbb{R}^{r-1}\mid x_i=k\}$. Now, we add the \emph{diagonal} hyperplanes $\mathcal{D}:=(D^j)_{r-1\leq j \leq k(r-1)}$ given by equations of the form $\sum_{i\in[r-1]}x_i=j+\epsilon$ for $r-1\leq j\leq k(r-1)$ and $0<\epsilon<1$. This, is $\mathcal{D}:=(D^j)_{r-1\leq j \leq k(r-1)}$ are the diagonals intersecting the grid translated a little bit into direction $(1, \ldots, 1)$. See Figure~\ref{fig:grid} for the rank $3$ case.
\begin{figure}[htb]
\includegraphics[width=.3\textwidth]{grid2}
\caption{The construction of Theorem~\ref{Th:unif1} for the case of rank $3$ and $\ell=4$. The gray cells are of type $R_v$.}
\label{fig:grid}
\end{figure}
Note that in the resulting arrangement $\mathcal{H}\cup \mathcal{D}$ each vertex $v$ of $\mathcal{H}$ is incident to a unique simplex $R_v$ into direction $(1, \ldots, 1)$. Moreover, $R_v\cap R_w=\emptyset$ unless $v=w$.
Therefore, we need at least one vertex for each of these $\ell^{r-1}$ simplices.
We extend $\mathcal{H}\cup \mathcal{D}$ to an arrangement $\mathcal{A}$ representing an orientation of $U_{r,n}$ with $n=(\ell-1)(r-1)+1+(\ell)(r-1)=(2\ell-1)(r-1)+1$. Thus, to determine the dual of any oriented matroid arising from an orientation of $\mathcal{A}$ at least $(\frac{n+r-2}{2(r-1)})^{r-1}$ circuits are needed. So this is the lower bound if $n$ can be expressed as $(2\ell-1)(r-1)+1$. For general $n$ the argument of the lower bound is calculated by $(r-1)\lfloor \frac{n-1}{r-1}\rfloor$. A straight-forward computation leads to the claimed result.
\end{proof}
Even if we have shown that $s$ and $\widetilde{s}$ differ in general, one implication of the previous result is, that they are asymptotically the same for uniform oriented matroids.
A particular consequence of Theorem \ref{prop:motiv} shown in~\cite{Cha-13} is that $s(U_{r,n})$ behaves asymptotically as $\frac{1}{r+1} {n \choose r}$ for any fixed $r$.
We get that $s(U_{n-r,n})\in \Theta(n^{r-1})$. By combining Theorems~\ref{Th:unif} and~\ref{Th:unif1} and Observation~\ref{obs:three} we get
\begin{corollary}\label{cor:stilde}
For fixed $r\geq 1$ we have $\widetilde{s}(U_{n-r,n})=\Theta(n^{r-1})$.
\end{corollary}
Let us now consider $\overline{s}$ for uniform oriented matroids.
\begin{theorem}\label{thm:s(n,r)}
For any $1\leq r\leq n-1$ we have
$\overline{s}(U_{r,n})=\binom{n}{r+1}-n+r+1$.
\end{theorem}
\begin{proof}
Assume first that $n>r+1$. We will use Theorem~\ref{thm:generalboundssbar}. First, as argued in the proof of Theorem~\ref{prop:motiv} \emph{any} basis of $U_{r,n}$ is invertible in some orientation. Thus, we already have $$|\mathrm{C}|+1-|E|+r(M)\leq\overline{s}(\mathrm{M})\leq |\mathrm{C}|+1-\min(\lambda(U_{r,n}), |E|-r(M)).$$ As showed in the proof of Theorem~\ref{prop:motiv} for a set of $\mathrm{S}$ of circuits of $U_{r,n}$ we have that $B_{\mathrm{S}}$ is connected if and only if $G_{\mathrm{S}}$ is connected. Therefore $\lambda(U_{r,n})$ coincides with the vertex-connectivity $\kappa(G_{\mathrm{S}})$ of $G_{\mathrm{S}}$. But $G_{\mathrm{S}}$ is exactly the Johnson graph $J(n,r+1)$ whose vertex connectivity is well-known to be its degree, which is $(r+1)(n-r-1)$. On the other hand $|E|+r(M)=\binom{n}{r+1}-n+r$. We obtain the result.
If $n=r+1$ then $\binom{n}{r+1}-n+r+1=\binom{n}{r+1}$ is necessary and sufficient as well, because it indeed means taking all circuits and with one circuit less one could not cover one base.
\end{proof}
\section{Regular matroids}\label{sec:regular}
In~\cite{Bla-78} it is shown that binary orientable matroids are exactly the regular matroids and that regular matroids have exactly one reorientation class. This section relies on these two facts. In particular, the first one leads us to give some results not depending on orientability when considering a general setting of binary matroids. The second one together with Theorem~\ref{thm:1reclass} immediately gives
\begin{corollary}\label{cor:regular}
If $\mathrm{M}$ is regular and connected then $\widetilde{s}(\mathrm{M})=s(\mathrm{M})=\mathrm{cc}(\mathrm{M})$.
\end{corollary}
From now on we will focus on $\mathrm{cc}(\mathrm{M})$ rather than $s(\mathrm{M})$. We will compare $\mathrm{cc}(\mathrm{M})$ with the minimum size of a (not necessarily connected) element covering of $\mathrm{M}$ denoted by {\boldmath $\mathrm{c}(\mathrm{M})$}. Moreover, since we will only consider element coverings rather than base coverings in this section and the section after, we will simply refer to them by coverings. In the present section we will derive several general bounds on $\mathrm{cc}(M)$, which we will apply in the next section to some graphic and cographic matroids.
\smallskip
Let us first reformulate the parameters $\mathrm{c}(\mathrm{M})$ and $\mathrm{cc}(\mathrm{M})$ when $\mathrm{M}$ is either a graphic or cographic. By Corollary~\ref{cor:regular} finding the circuits needed to determine all the orientations of a graphic matroid $\mathrm{M}(G)$ is equivalent to finding a set of cycles $\mathrm{S}$ in $G$ such that
\begin{itemize}
\item every edge of $G$ is contained in some $C\in \mathrm{S}$,
\item the graph induced by $\mathrm{S}$ (having as set of vertices the cycles of $\mathrm{S}$ and where two vertices $C$ and $C'$ are joined by an edge if and only if $C\cap C'\neq\emptyset$) is connected.
\end{itemize}
Such a set is called \emph{connected cycle cover}. As for general matroids we denote the minimum size of such a set of cycles of $G$ by {\boldmath $\mathrm{cc}(G)$}$=\mathrm{cc}(\mathrm{M}(G))$. The size of a minimum (not necessary connected) cycle cover is denoted by {\boldmath$ \mathrm{c}(G)$}$=\mathrm{c}(\mathrm{M}(G))$.
\smallskip
A \emph{bond} $B$ in a connected graph $G$ is an edge-set which is inclusion-minimal with the property that $G\setminus B$ is disconnected. Finding the circuits needed to determine all the orientations of a cographic matroid $\mathrm{M}^*(G)$ is equivalent to finding a set of bonds $\mathrm{S}$ in $G$ such that
\begin{itemize}
\item every edge of $G$ is contained in some $B\in \mathrm{S}$,
\item the induced by $\mathrm{S}$ (having as set of vertices the bonds of $\mathrm{S}$ and where two vertices $B$ and $B'$ are joined by an edge if and only if $B\cap B'\neq\emptyset$) is connected.
\end{itemize}
Such a set is called \emph{connected bond cover}. We denote the minimum size of such a set of bonds of $G$ by {\boldmath $\mathrm{cbc}(G)$}. The size of a minimum (not necessary connected) bond cover is denoted by {\boldmath $\mathrm{bc}(G)$}. This is, $\mathrm{cbc}(G)=\mathrm{cc}(\mathrm{M}^*(G))$ and $\mathrm{bc}(G)=\mathrm{c}(\mathrm{M}^*(G))$.
\smallskip
\begin{lemma}\label{lem:makeconnected}
For any connected matroid $\mathrm{M}$ we have $\mathrm{c}(\mathrm{M})\leq \mathrm{cc}(\mathrm{M})\leq 2\mathrm{c}(\mathrm{M})-1$.
\end{lemma}
\begin{proof}
The first inequality is trivial and only stated for completeness. For the second one,
take a circuit cover $\mathrm{S}$ of $\mathrm{M}$ and let $C,C'$ be circuits not in the same component of $I_{\mathrm{S}}$. Since $\mathrm{M}$ is connected there is a circuit $C''$ incident to both $C$ and $C'$. Adding $C''$ to $\mathrm{S}$ reduces the number of components by at least one. This yields the claim.
\end{proof}
Indeed the upper bound in Lemma~\ref{lem:makeconnected} is best-possible as already shown by graphic matroids:
\begin{proposition}\label{prop:makeconnectedbestpossible}
For every even $n\geq 2$ we have $\mathrm{cc}(K_{2,n})=2\mathrm{c}(K_{2,n})-1$.
\end{proposition}
\begin{proof}
Clearly, in $K_{2,n}$ the longest cycles are of length $4$ and since $n$ is even a partition into $4$-cycles is possible. Thus, $\mathrm{c}(K_{2,n})=\frac{n}{2}$. Now, given some set of cycles $\mathrm{S}$ in $K_{2,n}$, adding another cycle $C$ it can be incident to at most two components of $I_{\mathrm{S}}$. Thus, the construction in Lemma~\ref{lem:makeconnected} is best-possible.
\end{proof}
On the other hand there are cases, were $\mathrm{c}$ and $\mathrm{cc}$ coincide.
\begin{theorem}\label{thm:3}
Let $\mathrm{M}$ be a binary and connected matroid. Denote by $\mathrm{C}^*_3$ the cocircuits of size at most $3$. If $\mathrm{C}^*_3$ covers $E$ and its intersection graph is connected, then any covering $\mathrm{S}$ of $\mathrm{M}$ is connected, i.e., $\mathrm{c}(\mathrm{M})=\mathrm{cc}(\mathrm{M})$.
\end{theorem}
\begin{proof}
Let $\mathrm{S}$ be an element covering. Note that the existence of a circuit covering implies that there are no cocircuits of size $1$. Let $G_3$ be the connected intersection graph of $\mathrm{C}^*_3$. Now, every $C\in \mathrm{S}$ intersects elements of $\mathrm{C}^*_3$. Given $C,C'\in \mathrm{S}$ denote by $d_3(C,C')$ the length of a shortest path between two elements of $X,X'\in\mathrm{C}^*_3$ in $G_3$ such that $X$ intersects $C$ and $X'$ intersects $C'$. We claim that between every $C,C'\in \mathrm{S}$ there is a path in $I_{\mathrm{S}}$. We proceed by induction on $d_3(C,C')$.
If $d_3(C,C')=0$, then there is $X\in\mathrm{C}^*_3$ intersecting both $C$ and $C'$. Since $\mathrm{M}$ is binary both $C$ and $C'$ intersect $X$ in an even number of elements. Since $|X|\leq 3$ both intersect $X$ in two elements and therefore they must intersect themselves. Thus, they are connected in $I_{\mathrm{S}}$.
If $d_3(C,C')>0$, then choose a shortest path in $G_3$ witnessing $d_3(C,C')$. Let $X$ be the first cocircuit on this path, i.e., $X\cap C\neq \emptyset$. As $\mathrm{M}$ is binary, We have that $|X\cap C|=2$. Since $X$ was the first member of a shortest path in $G_3$, the element $X\setminus C=\{e\}$ and $e$ must be the intersection with the next member $X'$. Since $\mathrm{S}$ is an element covering, there is $C''\in \mathrm{S}$ containing $e$. As $\mathrm{M}$ is binary, we have that $C''$ intersects $C$. Thus, $d_3(C,C'')=0$ and $d_3(C'',C')<d_3(C,C')$. By induction hypothesis $C$ and $C''$ as well as $C''$ and $C'$ are connected in $I_{\mathrm{S}}$. This yields the claim.
\end{proof}
It is not sufficient to require that $E$ be covered by $\mathrm{C}^*_3$, see the left side of Figure~\ref{fig:notconnected}.
Also, the converse of Theorem~\ref{thm:3} does not hold as demonstrated by the right side of Figure~\ref{fig:notconnected}.
\begin{figure}[htb]
\includegraphics[width=.8\textwidth]{notconnected.pdf}
\caption{A graph $G$ in which every edge is contained in a bond of size $3$, but the intersection graph of these bonds is not connected. Left: A disconnected cycle covering consisting of four black, one purple and one gray cycle. Right: A connected cycle covering consisting of $3$ cycles. Indeed, one can see $\mathrm{c}(G)=\mathrm{cc}(G)=3$.}
\label{fig:notconnected}
\end{figure}
Using Theorem~\ref{thm:3} and Lemma~\ref{lem:makeconnected} together with results on $\mathrm{c}$ for $2$-connected graphs, ($\mathrm{c}(G)\leq\lfloor\frac{2n-1}{3}\rfloor$, see~\cite{Fan-02}), and cubic graphs, ($\mathrm{c}(G)\leq\lceil\frac{n}{4}\rceil$, if $G$ is cubic and $n\geq 6$, see~\cite{Lai-03} and $\mathrm{c}(G)\leq\lceil\frac{n}{6}\rceil$, if $G$ is cubic, $3$-connected, $n\geq 8$, and $G$ is not one of five forbidden graphs, see~\cite{Yan-11}), we get some general bounds:
\begin{corollary}\label{cor:2conn}
Let $G$ be a $2$-connected graph with $n$ vertices. Then,
\begin{itemize}
\item $\mathrm{cc}(G)\leq2\lfloor\frac{2n-1}{3}\rfloor-1$,
\item $\mathrm{cc}(G)\leq\lceil\frac{n}{4}\rceil$, if $G$ is cubic and $n\geq 6$.
\item $\mathrm{cc}(G)\leq \lceil\frac{n}{6}\rceil$, if $G$ is cubic, $3$-connected, $n\geq 8$, and $G$ is not one of five forbidden graphs.
\end{itemize}
\end{corollary}
We can also find some bounds involving the size of the ground set, the rank, the \emph{circumference} $\mathrm{circ}(\mathrm{M})$, i.e., the size of the largest circuit of $\mathrm{M}$, and the \emph{cogirth} $g^*(\mathrm{M})$, i.e., the size of a smallest cocircuit of $\mathrm{M}$.
\begin{theorem}\label{thm:edgevertexbound}
For any regular matroid $\mathrm{M}$ we have $\frac{|E|-1}{\mathrm{circ}(\mathrm{M})}\leq\mathrm{cc}(\mathrm{M})\leq |E|-r(\mathrm{M})+2-g^*(\mathrm{M})$.
\end{theorem}
\begin{proof}
We start by proving the lower bound: The most optimistic way to find a connected covering is taking only circuits of maximal size, i.e., $\mathrm{circ}(\mathrm{M})$. Moreover, since their intersection graph is connected they can be ordered such that each of them (except the first) shares at least one element with some earlier chosen one. Thus with $s$ such chosen circuits we cover $\mathrm{circ}(\mathrm{M})+(s-1)(\mathrm{circ}(\mathrm{M})-1)$ elements. So, this value should be at least $|E|$. From this we compute $s\geq \frac{|E|-1}{\mathrm{circ}(\mathrm{M})}$.
Given a matroid $\mathrm{\mathrm{M}}=(E,\mathrm{C})$, following~\cite{Lem-06} we denote by $\theta_e(\mathrm{M})$ the size of a smallest set $\mathrm{S}'$ of circuits in $\mathrm{S}_e$ needed to cover $E$. Evidently, such $\mathrm{S}'$ is a connected element covering
of $\mathrm{M}$ and thus $\mathrm{cc}(\mathrm{M})\leq \theta_e(\mathrm{M})$ for all $e\in E$. Moreover, denote by $g^*_e(\mathrm{M})$ the size of a smallest cocircuit containing $e$ and by $r(\mathrm{M})$ the rank of $\mathrm{M}$. In~\cite[Corollary 1.5]{Lem-06} it is shown that if $\mathrm{M}$ is connected, regular and not a coloop, and $e\in E$ such that $\mathrm{M}/e$ is connected, then $\theta_e(\mathrm{M})+g^*_e(\mathrm{M})\leq |E|-r(\mathrm{M})+2$. This immediately gives the result.
\end{proof}
A binary matroid is called \emph{Eulerian} if all cocircuits are of even size.
\begin{lemma}\label{lem:Eulerian}
If $\mathrm{M}$ is an Eulerian matroid, then $\mathrm{cc}(\mathrm{M})\neq 2$.
\end{lemma}
\begin{proof}
Suppose $\mathrm{cc}(\mathrm{M})=2$ witnessed by circuits $C_1, C_2$ covering the entire ground set and $C_1\cap C_2=N\neq\emptyset$. Every cocircuit $X$ is even and since $\mathrm{M}$ is binary $X$ intersects both $C_1$ and $C_2$ in an even number of elements. This implies, that $|X\cap N|$ is even for all cocircuits $X$. It is a well-known fact that if $|X\cap N|\neq 1$ for all cocircuits $X$, then $N$ contains a circuit. This contradicts $C_1$ and $C_2$ being circuits.
\end{proof}
Even if Lemma~\ref{lem:Eulerian} seems relatively weak, it provides tight lower bounds in a large family of cographic matroids as we will see in the next section.
\section{The hypercube and the complete graph}\label{sec:graph}
In this section we determine $\mathrm{c}, \mathrm{cc}, \mathrm{bc}$, and $\mathrm{cbc}$ for the class of hypercubes and complete graphs. We will make use of some lemmas of the previous section and prove some bounds to be tight.
The next result for odd $n$ shows that the lower bounds in Theorem~\ref{thm:edgevertexbound} and Lemma~\ref{lem:makeconnected} can indeed be attained:
\begin{theorem}\label{thm:Qn}
For every $n\geq 3$ we have $$
\mathrm{cc}(Q_n)=\left\lceil\frac{n+1}{2}\right\rceil= \begin{cases}
\mathrm{c}(Q_n)+1 & \text{if }n\text{ even}, \\
\mathrm{c}(Q_n) & \text{if }n\text{ odd}.\\
\end{cases}
$$
\end{theorem}
\begin{proof}
For the upper bound if $n$ is even we use that by~\cite{Als-90} the edges of $Q_n$ can be partitioned into $\frac{n}{2}$ Hamiltonian cycles, which proves $\mathrm{c}(Q_n)=\frac{n}{2}$ in this case. Clearly, since this is a partition into Hamiltonian cycles no connected covering with $\frac{n}{2}$ cycles exists. Now take a bond $X$ corresponding to a change in one coordinate of $Q_n$. Since $X$ is a bond all Hamiltonian cycles have to intersect it. Now, since the coordinate corresponding to $X$ can be switched at any vertex of $Q_n$, $X$ is a perfect matching. Therefore $X$ can be extended to a Hamiltonian cycle, which intersects all the others, see~\cite{Fin-07}. This concludes the case $n$ even.
\medskip
If $n$ is odd, first note that $\mathrm{cc}(Q_n)\geq\lceil\frac{n+1}{2}\rceil$ results from just plugging the values into Theorem~\ref{thm:edgevertexbound}.
For the upper bound choose two copies of $Q_{n-1}$, where vertices $v,v'$ in different copies are identified in the natural way, i.e., adding the coordinate matching $X_n$ connecting identified pairs yields $Q_n$. Now, take the partition $P$ of $Q_{n-1}$ into Hamiltonian cycles and its copy $P'$ partitioning the copy $Q'_{n-1}$. The coordinate matching $X_{n-1}$ in $Q_{n-1}$ intersects every Hamiltonian cycle in $P$. Denote by $X\subseteq X_{n-1}$ a matching hitting each cycle in $P$ exactly once and by $X'$ its copy in $Q'_{n-1}$.
For every Hamiltonian cycle $C\in P$ and its copy $C'\in P'$ take the unique matching edges of $e=\{v,w\}\in X$ and $e'=\{v',w'\}\in X'$ intersecting precisely $C$ and $C'$, respectively. Delete them from $C$ and $C'$ and join both cycles by adding $\{v,v'\}$ and $\{w,w'\}$, to obtain a new cycle $C''$. We have obtained a set $\widetilde{\mathrm{S}}$ of $\frac{n-1}{2}$ cycles. The edges of $Q_n$ still not covered by $\widetilde{\mathrm{S}}$ are precisely $X\cup X'$ and all edges of $\{u,u'\}\in X_n$ with $u$ not incident to $X$. Note that this set of edges forms a perfect matching $\widetilde{X}$. We have to cover $\widetilde{X}$ by a Hamiltonian cycle $\widetilde{C}$ which additionally intersects all cycles in $\widetilde{\mathrm{S}}$. For the construction we contract $X_n$ and $X_{n-1}\cup X'_{n-1}$ in $Q_n$ obtaining $Q_{n-2}$ (with parallel edges). Every vertex in $Q_{n-2}$ corresponds to a square in $Q_n$, which contains either two edges of $X\cup X'$ or two edges of $X_n$, which still have to be covered respectively. We call a vertex of type $I$ and $II$, depending on this. Moreover, by construction both remaining edges in such a square belong to exactly one $C\in\widetilde{\mathrm{S}}$. We say that a vertex of $Q_{n-2}$ meets $C$. Note that there are $\frac{n-1}{2}$ vertices of type $I$ each meeting one of the $\frac{n-1}{2}$ cycles in $C\in\widetilde{\mathrm{S}}$. All remaining vertices are of type $II$.
If $\frac{n-1}{2}$ is even let $H$ be any Hamiltonian cycle in $Q_{n-2}$. We can blow $H$ up to the desired $\widetilde{C}$ in $Q_n$ by just locally prescribing how to behave in the resulting squares. See the left of Figure~\ref{fig:cube}. Filling a type $I$ square corresponds to changing the coordinate inside the square along $X_n$ and a type $I$ square along $X_{n-1}\cup X'_{n-1}$. By the parity assumption and since $Q_{n-2}$ has an even number of vertices the numbers of type $I$ and type $II$ vertices are even. And our construction closes nicely and gives a cycle in $Q_n$.
\begin{figure}[htb]
\includegraphics[width=.5\textwidth]{cube.pdf}
\caption{Blowing up $H$. Left: usual handling of vertices of type $I$ and $II$. Right: special treatment of the edge $e$. Purple edges are those needed to be covered by $\widetilde{C}$.}
\label{fig:cube}
\end{figure}
If $\frac{n-1}{2}$ is odd chose an edge $e$ in $Q_{n-2}$ connecting a type $I$ vertex $a$ meeting $C$ and a type $II$ vertex $b$, such that some other vertex in $Q_{n-2}$ different from $a,b$ meets $C$. Such an edge clearly exists; since otherwise all neighbors of type $II$ of $a$ different from $b$ meet a cycle different from $C$ and we can as well take an edge connecting $a$ with such a neighbor. Now, we choose a Hamiltonian cycle $H$ containing $e$. In order to obtain $\widetilde{C}$ we handle all vertices as in the case before except $a,b$, see the right of Figure~\ref{fig:cube}. Along $e$ we have to repair the parity in order to close to a cycle in $Q_n$. The choice of $e$ was complicated by the fact that $\widetilde{C}$ will not meet the cycle met by $a$ in the square resulting from $a$.
\end{proof}
\begin{theorem}\label{thm:Qn*}
For every $n\ge 2$ we have $\mathrm{cbc}(Q_n)=\mathrm{bc}(Q_n)+1=3$.
\end{theorem}
\begin{proof}
Let us prove first by induction on $n$ that $\mathrm{bc}(Q_n)=2$ for every $n\ge 2$. Clearly the proposition holds if $n=2$ and maybe a little less clearly also for $n=3$, see the right of Figure~\ref{fig:Q_n}. We suppose the result holds for $n-1$ and $n>2$. We can obtain the graph $Q_{n}$ as follows. Choose two copies of $Q_{n-1}$, $A$ and $B$, where vertices in different copies are identified in the natural way, i.e., adding a matching connecting identified pairs yields $Q_{n}$. By induction, there exists a partition of the edges of $A$ in two bonds $A_1$ and $A_2$. Let $B_i$ be the copy of $A_i$ in $B$ for $i=1,2$.
\medskip
Let us define as $[C,C']$, the sets of edges in a graph $G$ having one extreme in $C$ and the other in $C'$, for every $C,C'\subset V(G)$. Let $C_1^{A_j}$ and $C_2^{A_j}$ be the two components of $A-A_j$, for $j=1,2$.
Since $C_i^{A_j}$ has a copy in $B$, let $C_i^{B_j}$ be the copy of $C_i^{A_j}$ in $B$ for every $i,j=1,2$. Observe that $C_i^{B_j}$ is one component in $B-B_j$. We consider the following sets of edges:
\begin{center}
\begin{tabular}{ccc}
$E_1$ & $=$ & $E(C_1^{A_1})\cup [C_1^{A_1}\cap C_1^{A_2}, C_1^{B_1}\cap C_1^{B_2}]\cup E(C_1^{B_2})$\\
$E_2$ & $=$ & $E(C_2^{A_1})\cup [C_2^{A_1}\cap C_2^{A_2}, C_2^{B_1}\cap C_2^{B_2}]\cup E(C_2^{B_2})$\\
$E_3$ & $=$ & $E(C_2^{A_2})\cup [C_1^{A_1}\cap C_2^{A_2}, C_1^{B_1}\cap C_2^{B_2}]\cup E(C_1^{B_1})$\\
$E_4$ & $=$ & $E(C_1^{A_2})\cup [C_2^{A_1}\cap C_1^{A_2}, C_2^{B_1}\cap C_1^{B_2}]\cup E(C_2^{B_1})$\\
\end{tabular}
\end{center}
Notice that $(E_1\cup E_2)\bigcup(E_3\cup E_4)=E(Q_{n})$ and $(E_1\cup E_2)\bigcap (E_3\cup E_4)=\emptyset$ which means that $E_1\cup E_2$ and $E_3\cup E_4$ are a partition of the edges of $Q_{n}$. We will see that $E_1\cup E_2$ and $E_3\cup E_4$ are two bonds of $Q_{n}$. For, we can check that $C_1^{A_2}\cup C_2^{B_1}$ and $C_2^{A_2}\cup C_1^{B_1}$ are the two components of $Q_{n}-(E_1\cup E_2)$. Also, we can observe that each edge in $E_1\cup E_2$ is incident to a vertex in $C_1^{A_2}\cup C_2^{B_1}$ and incident to a vertex in $C_2^{A_2}\cup C_1^{B_1}$, which means that $E_1\cup E_2$ is a bond of $Q_{n}$, see Figure~\ref{fig:Q_n}. Similarly, one can see that $E_3\cup E_4$ is a bond of $Q_{n}$.
Then $\mathrm{bc}(Q_n)=2$ for every $n\ge 2$.
\begin{figure}[htb]
\includegraphics[width=.9\textwidth]{Q_n.pdf}
\caption{Left: a partition of $Q_3$ into two bonds. Right: Extending a partition of $Q_{n-1}$ to a partition of $Q_n$.}
\label{fig:Q_n}
\end{figure}
Let us prove now that $\mathrm{cbc}(Q_n)=3$ for every $n\ge 2$. We have proved above that there exist a partition of the edges of $Q_n$ in two bonds $A_1$ and $A_2$. Let $C$ be a bond in $Q_n$ different to $A_1$ and $A_2$. Then
$E(A_1)\cap E(C)\neq\emptyset$ and $E(A_2)\cap E(C)\neq\emptyset$ and hence $\mathrm{cbc}(Q_n)\le 3$ for every $n\ge 2$.
To prove that $\mathrm{cbc}(Q_n)\ge 3$, we only have to use Lemma~\ref{lem:Eulerian}. As the cographic matroid of $Q_n$ is Eulerian since $Q_n$ is bipartite,
by Lemma~\ref{lem:Eulerian} it follows that $\mathrm{cbc}(Q_n)\ge 3$ for every $n\ge 2$.
\end{proof}
For the complete graph $K_n$ the lower bounds of Theorem~\ref{thm:edgevertexbound} and Lemma~\ref{lem:makeconnected} are sharp if and only if $n$ is even. More precisely:
\begin{theorem}\label{thm:Kn}
For every $n\geq 4$ we have
$$
\mathrm{cc}(K_n)=\lceil\frac{n}{2}\rceil= \begin{cases}
\mathrm{c}(K_n) & \text{if }n\text{ even}, \\
\mathrm{c}(K_n)+1 & \text{if }n\text{ odd}.\\
\end{cases}
$$
\end{theorem}
\begin{proof}
We will view $K_n$ as the Cayley graph of $\mathbb{Z}_n$ with connecting set $\mathbb{Z}_n\setminus\{0\}$, where the double edges oriented into opposite directions are seen as undirected edges.
Our construction is based on a well-known partition $\mathcal{P}_n$ of the edges of $K_n$ for even $n$ into $\frac{n}{2}$ Hamiltonian zig-zag-paths \cite{Alsp-Gav01}. The path $P_i$ traverses the vertices as follows: $i, i+1,i+1-2,i+1-2+3,\dots$.
\medskip
Now, if $n$ is odd, take $\mathcal{P}_{n-1}$ and add an edge from the additional vertex $n$ to $P_i$ if it connects to one of the endpoints of $P_i$. This is a well-known construction for a partition of $K_n$ into Hamiltonian cycles. Thus, to obtain a connected cycle covering at least one additional cycle is needed. Indeed, such a cycle is easy to find. Take for instance $C=(0,1,\ldots, \frac{n+1}{2},0)$. By construction of $\mathcal{P}_{n-1}$ this cycle intersects all other
cycles in the partition.
\medskip
For $n$ even it is well-known that $\mathrm{c}(K_n)=\frac{n}{2}$, thus this lower bound on $\mathrm{cc}(K_n)$ follows from Lemma~\ref{lem:makeconnected}. We will show that there are indeed connected cycle covers of that size.
Now, if $n$ is even but not divisible by $4$, take $\mathcal{P}_{n}$ and add an edge to each $P_i$ connecting its end-vertices. This yields a cover $\mathcal{H}_{n}$ of $K_n$ by Hamiltonian cycles $H_i$, which is smallest possible. We show that it is indeed connected. Note that each $H_i$ contains two \emph{long diagonals}, i.e., edges labeled $\frac{n}{2}$, one being the $\frac{n}{2}$th or middle edge $e_i$ of $P_i$ and one being the newly added edge $f_i$. Each such edge is contained in another element of $\mathcal{H}_{n}$. More precisely, we have $e_i=f_{i+\frac{n+2}{4}}$ and therefore in the graph on $\mathcal{H}_{n}$ there is an edge between $H_i$ and $H_{i+\frac{n+2}{4}}$, but $i+\frac{n+2}{4}$ has to be taken modulo $\frac{n}{2}$. By the divisibility conditions on $n$ we get that $\frac{n}{2}$ and $\frac{n+2}{4}$ are coprime and therefore connecting $H_i$ and $H_{i+\frac{n+2}{4}}$ modulo $\frac{n}{2}$ for all $i$ yields a single connected component. This is, $\mathcal{H}_{n}$ is a connected cycle cover of size $\frac{n}{2}$.
The last case concerns $n$ divisible by $4$. If $n=4$ it is easy to find a connected cycle cover of size $2$. Otherwise we take the cycle cover constructed in the paragraph above for $K_{n-2}$ and modify it to cover the complete graph with two additional vertices $v,w$. In $K_{n-2}$, each long diagonal is covered twice. In each cycle $H\in\mathcal{H}_{n-2}$ replace the long diagonal by two consecutive edges passing through $v$ and $w$, respectively.
Denote the resulting set of cycles by $\mathcal{H}'$. It covers all edges but $\{v,w\}$ and the long diagonals of $K_{n-2}$, i.e., edges labeled $\frac{n-2}{2}$ connecting vertices different from $v,w$. We add one more cycle $C$ using all these edges and taking every other edge of the cycle $(0,1,2,\ldots, n-3,0)$ of $K_{n-2}$ except $\{0,1\}$ and $\{\frac{n-2}{2},\frac{n-2}{2}+1\}$. Instead $C$ includes $\{\frac{n-2}{2},v\}$ and $\{\frac{n-2}{2}+1,w\}$ (or $v$ and $w$ permuted) such that on these edges $C$ intersects the cycle arising from $H_0$. All the other $H_i$ are intersected by $C$ via every other edge of the cycle $(0,1,2,\ldots, n-3,0)$ of $K_{n-2}$ since $n-2$ is not divisible by $4$.
\end{proof}
\begin{theorem}\label{thm:Kn*}
For all $n\geq 2$ we have $\mathrm{cbc}(K_n)=\mathrm{bc}(K_n)=\lceil\log_2(n)\rceil$.
\end{theorem}
\begin{proof}
First it is easy to see that $\mathrm{M}^*(K_n)$ satisfies the preconditions of Theorem~\ref{thm:3}, i.e., every edge is contained in a triangle and the edge-intersection graph of triangles is connected. This implies $\mathrm{cbc}(K_n)=\mathrm{bc}(K_n)$. Now, any set of minimal cuts covering the edges of $K_n$ corresponds to a set of maximal bipartite subgraphs covering the edges. Note that this correspondence holds if and only if the graph is the complete graph. The minimum number of bipartite subgraphs to cover a graph $G$ is $\lceil\lg \chi(G)\rceil$, see~\cite{Mat-72,Had-75,Gar-76}. In our special case it yields the result.
\end{proof}
|
\section{Introduction}
There is a large literature concerning the definition of the Sobolev space $W^{1,p}(X,{\sf d},\mathfrak {m})$ of real valued functions defined on a metric measure space $(X,{\sf d},\mathfrak {m})$, we refer to \cite{Heinonen07} and \cite{AmbrosioGigliSavare11-3} for historical comments and a presentation of the various - mostly equivalent - approaches.
The definition of space $W^{1,p}(X,{\sf d},\mathfrak {m})$ comes with the definition of an object playing the role of the modulus of the distributional differential. More precisely, for $f\in W^{1,p}(X,{\sf d},\mathfrak {m})$ it is well defined a non-negative function $|{\rm D} f|_p\in L^p(X,\mathfrak {m})$, called minimal $p$-weak upper gradient, which, if $(X,{\sf d},\mathfrak {m})$ is a smooth space, coincides $\mathfrak {m}$-a.e.\ with the modulus of the distributional differential of $f$.
A key difference between the smooth and non-smooth case is that in the latter the minimal $p$-weak upper gradient may depend on $p$: say for simplicity that $\mathfrak {m}(X)=1$, then for $p<q\in(1,\infty)$ and $f\in W^{1,q}(X)$ one always has $f\in W^{1,p}(X)$ but in general only the inequality
\begin{equation}
\label{eq:1i}
|{\rm D} f|_p\leq |{\rm D} f|_q,\qquad\mathfrak {m}-a.e.,
\end{equation}
holds. The inequality above can be strict even on doubling spaces, see \cite{DiMarinoSpeight13} for an example and more details on the issue.
Worse than this, one might have
\begin{equation}
\label{eq:2i}
\text{a function $f\in W^{1,p}(X)$ with $f,|{\rm D} f|_p\in L^q(X)$ such that $f\notin W^{1,q}(X)$,}
\end{equation}
see \cite{AmbrosioGigliSavare11-3} for an example proposed by Koskela.
To have a $p$-weak upper gradients independent on $p$ is a regularity property of the metric measure space in question. For instance, as a consequence of the analysis done in \cite{Cheeger00} one has that on doubling space supporting a 1-1 weak local Poincar\'e inequality, equality always holds in \eqref{eq:1i}. In particular, this applies to ${\sf CD}(K,N)$ spaces with $N<\infty$.
In this note we show that on ${\sf RCD}(K,\infty)$ spaces not only \eqref{eq:1i} holds with equality, but also that the situation in \eqref{eq:2i} never occurs. The argument is based on some regularization properties of the heat flow proved in \cite{Savare13} and on the density in energy of Lipschitz functions in Sobolev spaces established in \cite{AmbrosioGigliSavare11-3}.
At least in the case of proper ${\sf RCD}(K,\infty)$ spaces, this identification extends to ${\rm BV}$ functions. The problem in non-proper spaces is the lack of an approximation result of ${\rm BV}$ functions with Lipschitz ones.
\bigskip
This result, beside its intrinsic usefulness in Sobolev calculus, has also the pleasant conceptual effect of somehow relieving the definition of ${\sf RCD}$ spaces from the dependence on the particular Sobolev exponent $p=2$. Recall indeed that one of the equivalent definitions of ${\sf RCD}(K,\infty)$ space is that of a ${\sf CD}(K,\infty)$ space such that $W^{1,2}(X)$ is Hilbert or equivalently such that
\[
|{\rm D} (f+g)|_2^2+|{\rm D}(f-g)|_2^2=2\big(|{\rm D} f|_2^2+|{\rm D} g|_2^2\big),\quad\mathfrak {m}\text{-a.e. }\qquad\forall f,g\in W^{1,2}(X).
\]
As a consequence of our result, a posteriori one could replace the minimal 2-weak upper gradients with $p$-weak upper gradients in the above.
\section{Preliminaries}
\subsection{Sobolev classes}
We assume the reader familiar with the basic concepts of analysis in metric measure spaces and recall here the definition of Sobolev class ${\rm S}^p(X)$. We fix a complete and separable space $(X,{\sf d},\mathfrak {m})$ such that $\mathfrak {m}$ is a non-negative Borel measure finite on bounded sets.
\begin{definition}[Test plans] Let ${\mbox{\boldmath$\pi$}}$ be a Borel probability measure on $C([0,1],X)$. We say that ${\mbox{\boldmath$\pi$}}$ has bounded compression provided
there exists $C =C({\mbox{\boldmath$\pi$}})>0$ such that
\[
({\rm e}_t)_\sharp {\mbox{\boldmath$\pi$}} \leq C\mathfrak {m},\qquad\forall t \in [0,1],
\]
where ${\rm e}_t:C([0,1],X)\to X$ is the evaluation map defined by ${\rm e}_t(\gamma):=\gamma_t$ for every $\gamma\in C([0,1],X)$.
For $q\in(1,\infty)$ we say that ${\mbox{\boldmath$\pi$}}$ is a $q$-test plan if it has bounded compression, is concentrated on $AC^q([0,1],X)$ and
\[
\int_0^1\int |\dot{\gamma}_t|^q \,\d{\mbox{\boldmath$\pi$}}(\gamma)\, \d t < +\infty.
\]
\end{definition}
The notion of Sobolev function is then introduced by duality with test plans.
\begin{definition}[Sobolev classes]
Let $p\in(1,\infty)$. The space ${\rm S}^p(X)$ is the space of all Borel functions $f : X \to \mathbb{R}$ for which there exists a non-negative function $G\in L^p(X)$ such that for any $q$-test plan ${\mbox{\boldmath$\pi$}}$ the inequality
\[
\int |f(\gamma_1)- f(\gamma_0)|\, \d{\mbox{\boldmath$\pi$}}(\gamma) \leq \iint_0^1 G(\gamma_s)|\dot{\gamma}_s|\, \d s\, \d{\mbox{\boldmath$\pi$}}(\gamma)
\]
holds, where $\frac1p+\frac1q=1$. Any such $G$ is called $p$-weak upper gradient.
\end{definition}
It is immediate to see that for $f\in{\rm S}^p(X)$ there is a unique $p$-weak upper gradient of minimal $L^p$-norm: we shall call such $G$ minimal $p$-weak upper gradient and denote it by $|{\rm D} f|_p$.
Basic important properties of minimal weak upper gradients are the \underline{locality}, i.e.:
\[
|{\rm D} f|_p=|{\rm D} g|_p,\quad\mathfrak {m}\text{-a.e. on }\{f=g\},\qquad\forall f,g\in{\rm S}^p(X),
\]
and the \underline{lower semicontinuity}, i.e.
\[
\left.\begin{array}{ll}
(f_n)\subset {\rm S}^p(X),\\
\sup_n\||{\rm D} f|_p\|_{L^p}<\infty,\\
f_n\to f\quad\mathfrak {m}\text{-a.e.}
\end{array}\right\}\qquad\Rightarrow\qquad\left\{\begin{array}{l}
f\in {\rm S}^p(X)\text{ and}\\
\text{for any weak limit $G$}
\text{ of $(|{\rm D} f_n|_p)$ in $L^p(X)$}\\
\text{it holds }|{\rm D} f|_p\leq G\qquad\mathfrak {m}\text{-a.e}.
\end{array}\right.
\]
The Sobolev space $W^{1,p}(X)$ is defined as $W^{1,p}(X):= {\rm S}^p\cap L^p(X)$ endowed with the norm
\[
\|f\|^p_{W^{1,p}(X)}:=\|f\|^p_{L^p(X)}+\||{\rm D} f|_p\|^p_{L^p(X)}.
\]
By $\Lip X$ we denote the space of Lipschitz functions on $X$ and for $f\in\Lip X$ the local Lipschitz constant $\lip f:X\to[0,\infty)$ is defined as
\[
\lip f(x):=\varlimsup_{y\to x}\frac{|f(y)-f(x)|}{{\sf d}(x,y)},
\]
if $x$ is not isolated, 0 otherwise.
In \cite{AmbrosioGigliSavare11-3} the following approximation property has been proved:
\begin{proposition}[Density in energy of Lipschitz functions]\label{prop.approx}
Let $(X,{\sf d},\mathfrak {m})$ be a complete separable metric space with $\mathfrak {m}$ being Borel non-negative and assigning finite mass to bounded sets. Let $p\in(1,\infty)$ and $f\in W^{1,p}(X)$.
Then there exists a sequence $(f_n)\subset W^{1,p}\cap\Lip X$ of functions with bounded support converging to $f$ in $L^p(X)$ and such that $\lip{ f_n}\to |{\rm D} f|_p$ in $L^p(X)$ as $n\to\infty$.
\end{proposition}
We conclude this introduction noticing that the locality property of $p$-weak upper gradients allows for a natural definition of the space of locally Sobolev functions. By $L^p_{\rm loc}(X)$ we shall intend the space of Borel functions $G:X\to\mathbb R$ whose $p$-power is integrable on every bounded set.
\begin{definition}[The spaces ${\rm S}^p_{\rm loc}(X)$]
We say that $f\in{\rm S}^p_{\rm loc}(X)$ provided for any Lipschitz function with bounded support ${\raise.3ex\hbox{$\chi$}}$ we have ${\raise.3ex\hbox{$\chi$}} f\in{\rm S}^p(X)$. In this case we define $|{\rm D} f|_p\in L^p_{\rm loc}(X)$ as
\[
|{\rm D} f|_p:=|{\rm D}({\raise.3ex\hbox{$\chi$}} f)|_p,\qquad\mathfrak {m}\text{-a.e. on }\{{\raise.3ex\hbox{$\chi$}}=1\},
\]
for every ${\raise.3ex\hbox{$\chi$}}$ as before.
\end{definition}
The role of the locality of the $p$-weak upper gradient is to ensure that the definition of $|{\rm D} f|_p$ is well posed. Also, it is not hard to check that ${\rm S}^p(X)\subset {\rm S}^p_{\rm loc}(X)$ and that a function $f\in {\rm S}^p_{\rm loc}(X)$ belongs to ${\rm S}^p(X)$ if and only if $|{\rm D} f|_p\in L^p(X)$.
For $p_1<p_2\in (1,\infty)$ and $q_1>q_2\in(1,\infty)$ such that $\frac1{p_i}+\frac1{q_i}=1$, the fact that the class of $q_1$-test plans is contained in the one of $q_2$-test plans grants that
\begin{equation}
\label{eq:easy}
{\rm S}^{p_2}_{\rm loc}(X)\subset {\rm S}^{p_1}_{\rm loc}(X)\qquad\text{and}\qquad|{\rm D} f|_{p_1}\leq |{\rm D} f|_{p_2}\quad\mathfrak {m}\text{-a.e.}\qquad\forall f\in {\rm S}^{p_2}_{\rm loc}(X).
\end{equation}
\subsection{Heat flow on $RCD(K,\infty)$ space}
In order to keep this preliminary part as short as possible, we shall assume the reader familiar with the definition of ${\sf RCD}(K,\infty)$ spaces and focus only on those properties they have which are relevant for our discussion. We refer to \cite{AmbrosioGigliSavare11-2}, \cite{AmbrosioGigliMondinoRajala12} and \cite{Savare13} for the throughout discussion.
From now on we shall assume that $(X,{\sf d},\mathfrak {m})$ is a ${\sf RCD}(K,\infty)$ space for some $K\in\mathbb R$ and that the support of $\mathfrak {m}$ is the whole $X$. Recall that in particular we have $\mathfrak {m}(B)<\infty$ for any bounded Borel set $B\subset X$.
In such space the 2-Energy functional $\sf E:L^2(X)\to[0,\infty]$ defined as
\[
\sf E_2(f):=\left\{\begin{array}{ll}
\frac12 \int_X |{\rm D} f|_2^2\,\d \mathfrak {m},&\qquad\text{ if }f\in W^{1,2}(X),\\
+\infty,&\qquad\text{ otherwise,}
\end{array}\right.
\]
is a strongly local and regular Dirichlet form. We shall denote by $(\H_t)$ the associated linear semigroup. Then it can be seen that for every $f\in L^2(X)$ and $p\in[1,\infty)$ we have
\[
\|\H_t(f)\|_{L^p(X)}\leq \| f\|_{L^p(X)},\qquad\forall t\geq0,
\]
and thus $(\H_t)$ can, and will, be extended to a linear non-expanding semigroup on $L^p(X)$.
On the other hand, there exists a unique ${\sf EVI}_K$-gradient flow of the relative entropy functional on $(\probt X,W_2)$ which we shall denote by $(\mathcal H_t)$ and provides a one parameter semigroup of continuous linear operators on $(\probt X,W_2)$, see \cite{AmbrosioGigliSavare11-2} and \cite{AmbrosioGigliMondinoRajala12} .
The non-trivial link between $(\H_t)$ and $(\mathcal H_t)$ is the fact that
\[
\begin{split}
&\text{for $\mu\in\probt X$ such that $\mu=f\mathfrak {m}$ for some $f\in L^2(X)$}\\
&\text{we have $\mathcal H_t(\mu)=\H_t(f)\mathfrak {m}$ for every $t\geq 0$,}
\end{split}
\]
and from the fact that $(\H_t)$ is self adjoint one can verify that for any $p\in[1,\infty)$ and every $t\geq 0$ it holds
\begin{equation}
\label{eq:version}
\H_t(f)(x)=\int f\,\d\mathcal H_t(\delta_x),\qquad\mathfrak {m}-a.e.\ x\qquad\forall f\in L^p(X).
\end{equation}
Moreover, for $f\in L^\infty(X)$ and $t>0$ the formula
\[
\tilde{\sf H}_t(f)(x):=\int f\,\d\mathcal H_t(\delta_x),
\]
is well defined for any $x\in X$ producing a pointwise version of the heat flow for which the $L^\infty\to{\rm Lip}$ regularization holds:
\begin{equation}
\label{eq:linftylip}
\Lip{\tilde{\sf H}_t(f)}\leq \frac1{\sqrt{2I_{2K}(t)}}\|f\|_{L^\infty(X)},\qquad\forall t>0,
\end{equation}
where $I_{2K}(t):=\int_0^te^{2Ks}\,\d s$.
Th crucial regularization property of the heat flow that we shall use to identify $p$-weak gradients is the following version of the Bakry-\'Emery contraction rate, proved in \cite{Savare13}:
\begin{equation}
\label{eq:BE}
\lip{\tilde{\sf H}_t(f)}\leq e^{-Kt}\tilde{\sf H}_t(\lip f)\qquad\text{pointwise on }X,
\end{equation}
valid for every Lipschitz function $f$ with bounded support and every $t\geq 0$.
We conclude recalling another useful regularity property of ${\sf RCD}$ spaces, this one concerning displacement interpolation of measures, see \cite{Rajala12-2} for a proof:
\begin{proposition}\label{prop:rajala}
Let $(X,{\sf d},\mathfrak {m})$ be a ${\sf RCD}(K,\infty)$ space and $\mu,\nu$ two Borel probability measures with bounded support and such that $\mu,\nu\leq C\mathfrak {m}$ for some $C>0$.
Then there exists a Borel probability measure ${\mbox{\boldmath$\pi$}}$ on $ C([0,1],X)$ such that $({\rm e}_t)_\sharp{\mbox{\boldmath$\pi$}}\leq C'\mathfrak {m}$ for every $t\in[0,1]$ for some $C'>0$ and for which the inequality
\[
\Lip{\gamma}\leq \sup_{x\in \supp(\mu)\atop y\in\supp(\nu)}{\sf d}(x,y),
\]
holds for every $\gamma$ in the support of ${\mbox{\boldmath$\pi$}}$.
\end{proposition}
\section{Proof of the main result}
The identification of $p$-weak gradients will come via a study of the regularization properties of the heat flow.
\begin{proposition}\label{prop:1}
Let $p\in(1,\infty)$, $f\in W^{1,p}(X)$ and $t\geq 0$. Then $\H_t(f)\in W^{1,p}(X)$ and
\[
|{\rm D} \H_tf|_p^p\leq e^{-pKt}\H_t(|{\rm D} f|_p^p),\qquad\mathfrak {m}-a.e..
\]
\end{proposition}
\begin{proof} The fact that $\H_t(f)\in L^p(X)$ follows from the fact that $f\in L^p(X)$.
By Proposition \ref{prop.approx} we can find a sequence $(f_n)\subset W^{1,p}\cap\Lip{X}$ converging to $f$ in $L^p(X)$ and such that $\lip{f_n}\to |{\rm D} f|_p$ in $L^p(X)$.
By the property \eqref{eq:BE} we know that
\begin{equation}
\label{eq:1}
\lip{\tilde{\sf H}_t(f_n)}^p \leq e^{-pKt}\tilde{\sf H}_t(\lip{f_n}^p),\qquad\text{pointwise on }X.
\end{equation}
The continuity in $L^1(X)$ of the heat flow grants that
\begin{equation}
\label{eq:2}
\tilde{\sf H}_t(|\lip{f_n}|^p)\to \tilde{\sf H}_t(|{\rm D} f|_p^p),\qquad \text{in }L^1(X),
\end{equation}
so that in particular \eqref{eq:1} grants that the sequence $(\lip{\tilde{\sf H}_t(f_n)})$ is bounded in $L^p(X)$. Therefore also $(|{\rm D} \tilde{\sf H}_t(f_n)|_p)$ is bounded in $L^p(X)$ and up to pass to a subsequence, not relabeled, we can assume that it weakly converges to some $G\in L^p(X)$. By \eqref{eq:1} and \eqref{eq:2} we have $G\leq \tilde{\sf H}_t(|{\rm D} f|_p^p)$ $\mathfrak {m}$-a.e.\ while the lower semicontinuity of $p$-weak upper gradients ensures that $\H_t(f)\in{\rm S}^p(X)$ with $|{\rm D} f|_p\leq G$ $\mathfrak {m}$-a.e.\ and the thesis follows.
\end{proof}
\begin{proposition}\label{prop:2}
Let $p\in(1,\infty)$, $f\in W^{1,p}(X)$ such that $f,|{\rm D} f|_p\in L^\infty(X)$ and $t> 0$. Then $\tilde{\sf H}_t(f)$ is Lipschitz and
\[
\lip{\tilde{\sf H}_t(f)}\leq e^{-Kt}\sqrt[p]{\tilde{\sf H}_t(|{\rm D} f|_p^p)},\qquad\text{pointwise on }X.
\]
\end{proposition}
\begin{proof}
The fact that $\tilde{\sf H}_t(f)$ is Lipschitz follows from \eqref{eq:linftylip}. To prove the thesis, pick $x,y\in X$, $r>0$, consider the measures $\mu_{0,r}:=\mathfrak {m}(B_r(x))^{-1}\mathfrak {m}\restr{B_r(x)}$, $\mu_{1,r}:=\mathfrak {m}(B_r(y))^{-1}\mathfrak {m}\restr{B_r(y)}$ and let ${\mbox{\boldmath$\pi$}}$ be given by Proposition \ref{prop:rajala}. Then we know that $({\rm e}_s)_\sharp{\mbox{\boldmath$\pi$}}\leq C\mathfrak {m}$ for some $C>0$ and every $s\in[0,1]$ and that $|\dot \gamma_s|\leq{\sf d}(x,y)+2r$ for ${\mbox{\boldmath$\pi$}}$-a.e.\ $\gamma$ and a.e.\ $s\in[0,1]$. In particular, ${\mbox{\boldmath$\pi$}}$ is a $q$-test plan, where $\frac 1p+\frac1q=1$, and since $\H_t(f)\in W^{1,p}(X)$ we know that
\[
\begin{split}
\Big|\int \tilde{\sf H}_t(f)\,\d(\mu_{1,r}- \mu_{0,r})\Big|&\leq \int |\tilde{\sf H}_t(f)(\gamma_1)-\tilde{\sf H}_t(f)(\gamma_0)|\,\d{\mbox{\boldmath$\pi$}}(\gamma)\\
&\leq\iint_0^1|{\rm D} \tilde{\sf H}_t(f)|_p(\gamma_s)|\dot\gamma_s|\,\d s\,\d{\mbox{\boldmath$\pi$}}(\gamma)\\
&\leq({\sf d}(x,y)+2r)\sqrt[p]{\iint_0^1|{\rm D} \tilde{\sf H}_t(f)|_p^p(\gamma_s)\,\d s\,\d{\mbox{\boldmath$\pi$}}(\gamma)}\\
&\leq({\sf d}(x,y)+2r)e^{-Kt}\sqrt[p]{\iint_0^1\tilde{\sf H}_t(|{\rm D} f|_p^p)(\gamma_s)\,\d s\,\d{\mbox{\boldmath$\pi$}}(\gamma)},
\end{split}
\]
having used Proposition \ref{prop:1} in the last step. Noticing that ${\sf d}(x,\gamma_s)\leq {\sf d}(x,y)+3r$ for ${\mbox{\boldmath$\pi$}}$-a.e.\ $\gamma$ and every $s\in[0,1]$ we deduce that
\[
\Big|\int \tilde{\sf H}_t(f)\,\d(\mu_{1,r}- \mu_{0,r})\Big|\leq ({\sf d}(x,y)+2r)e^{-Kt}\sqrt[p]{\sup_{B_{{\sf d}(x,y)+3r}}\tilde{\sf H}_t(|{\rm D} f|_p^p)},
\]
and letting $r\downarrow0$ and using the continuity of $\tilde{\sf H}_t(f)$ we deduce that
\[
\frac{|\tilde{\sf H}_t(f)(y)-\tilde{\sf H}_t(f)|}{{\sf d}(x,y)}\leq e^{-Kt}\sqrt[p]{\sup_{B_{{\sf d}(x,y)+\varepsilon}}\tilde{\sf H}_t(|{\rm D} f|_p^p)},\qquad\forall\varepsilon>0.
\]
Letting $y\to x$ using the continuity of $\tilde{\sf H}_t(|{\rm D} f|_p^p)$ (which follows from the hypothesis $|{\rm D} f|_p\in L^\infty(X)$ and \eqref{eq:linftylip}) and the arbitrariness of $\varepsilon>0$ we conclude.
\end{proof}
\begin{proposition}\label{prop:3}
Let $p,q\in(1,\infty)$ and $f\in \Lip X $. Then
\[
|{\rm D} f|_q=|{\rm D} f|_p,\qquad\mathfrak {m}-a.e..
\]
\end{proposition}
\begin{proof}Assume $p<q$. Then we already know by \eqref{eq:easy} that $|{\rm D} f|_p\leq |{\rm D} f|_q$ $\mathfrak {m}$-a.e.. Notice that by the locality property of the weak upper gradients it is not restrictive to assume that $f$ has bounded support, so that in particular $f\in L^\infty\cap W^{1,p}(X)$. Let $t>0$ and apply Proposition \ref{prop:2} to deduce that $\tilde{\sf H}_t(f)\in\Lip X$ with
\[
\lip{\tilde{\sf H}_t(f)}^q\leq e^{-qKt}\tilde{\sf H}_t(|{\rm D} f|_p^p)^{\frac qp}\leq e^{-qKt}\tilde{\sf H}_t(|{\rm D} f|_p^q),\qquad\text{pointwise},
\]
having used Jensen's inequality and formula \eqref{eq:version} in the last step and the fact that $|{\rm D} f|_p^q\in L^1(X)$, which follows from the fact that $f$ is Lipschitz bounded support. Since $|{\rm D} \tilde{\sf H}_t(f)|_q\leq \lip{\tilde{\sf H}_t(f)}$ $\mathfrak {m}$-a.e., it follows that
\[
\int|{\rm D} \tilde{\sf H}_t(f)|_q^q\,\d \mathfrak {m}\leq e^{-qKt}\int \tilde{\sf H}_t(|{\rm D} f|_p^q)\,\d\mathfrak {m},\qquad\forall t>0,
\]
and letting $t\downarrow 0$ and using the lower semicontinuity of $q$-weak upper gradients we conclude that
\[
\int|{\rm D} f|_q^q\,\d\mathfrak {m}\leq \int |{\rm D} f|_p^q\,\d\mathfrak {m},
\]
which is sufficient to get the thesis.
\end{proof}
\begin{theorem}[Identification of weak upper gradients]\label{thm:main}
Let $p,q\in(1,\infty)$ and $f\in{\rm S}^p_{\rm loc}(X)$ such that $|{\rm D} f|_p\in L^q_{\rm loc}(X)$. Then $f\in{\rm S}^q_{\rm loc}(X)$ and
\[
|{\rm D} f|_q=|{\rm D} f|_p,\qquad\mathfrak {m}-a.e..
\]
\end{theorem}
\begin{proof}
Assume that $p< q$ and notice that by \eqref{eq:easy} it is sufficient to prove that $|{\rm D} f|_p\geq |{\rm D} f|_q$ $\mathfrak {m}$-a.e.. Replacing if necessary $f$ with $\max\{\min\{f,n\},-n\}$ and using the locality property of weak upper gradients and the arbitrariness of $n\in\mathbb N$ we can assume that $f\in L^\infty(X)$. Similarly, with a cut-off argument we reduce to the case in which $f$ has bounded support and thus in particular $|{\rm D} f|_p\in L^p\cap L^q(X)$.
With these assumptions we have $f\in W^{1,p}(X)$ and thus for $t>0$ Proposition \ref{prop:1} gives
\[
|{\rm D} \H_tf|_p\leq e^{-Kt}\sqrt[p]{\H_t(|{\rm D} f|_p^p)},\qquad\mathfrak {m}-a.e..
\]
Moreover, the fact that $f$ is bounded grants, by \eqref{eq:linftylip}, that $\H_t(f)$ has a Lipschitz representative $\tilde{\sf H}_t(f)$ and thus Proposition \ref{prop:3} gives
\[
|{\rm D} \H_tf|_q\leq e^{-Kt}\sqrt[p]{\H_t(|{\rm D} f|_p^p)},\qquad\mathfrak {m}-a.e..
\]
Using the assumption that $|{\rm D} f|_p\in L^q(X)$ and Jensen's inequality in formula \eqref{eq:version} we deduce that $|{\rm D} \H_tf|^q_q\leq e^{-qKt} \H_t(|{\rm D} f|_p^q)$ $\mathfrak {m}$-a.e.\ and thus
\[
\int |{\rm D} \H_tf|^q_q\,\d\mathfrak {m}\leq e^{-qKt} \int\H_t(|{\rm D} f|_p^q)\,\d\mathfrak {m},\qquad\forall t>0.
\]
Letting $t\downarrow0$ and using the lower semicontinuity of $q$-weak upper gradients we conclude that
\[
\int |{\rm D} f|_q^q\,\d\mathfrak {m}\leq \int |{\rm D} f|_p^q\,\d\mathfrak {m},
\]
which is sufficient to prove the thesis.
\end{proof}
\begin{remark}[The case of $BV$ functions]{\rm Recalling the notation and results of \cite{Ambrosio-DiMarino14} about ${\rm BV}$ functions and denoting by $\bb f$ the total variation measure of $f\in {\rm BV}(X)$, assume for a moment that $(X,{\sf d},\mathfrak {m})$ is a \emph{proper} (=bounded closed sets are compact) ${\sf RCD}(K,\infty)$ space. Then the very same arguments just used allow to prove that
\begin{equation}
\label{eq:claimbv}
\begin{split}
&\text{if $f\in {\rm BV}(X)$ is such that $\bb f \ll\mathfrak {m}$ with $\frac{\d\bb f }{\d\mathfrak {m}}\in L^p_{\rm loc}(X)$ for some $p>1$,}\\
&\text{then $f\in{\rm S}^p_{\rm loc}(X)$ and $|{\rm D} f|_p=\frac{\d\bb f }{\d\mathfrak {m}}$ $\mathfrak {m}$-a.e.. }
\end{split}
\end{equation}
To see why, notice that the fact that $(X,{\sf d})$ is proper and the definition of ${\rm BV}(X)$ ensures that for $f\in {\rm BV}(X)$ there is a sequence $(f_n)$ of Lipschitz functions with bounded support such that $(f_n)\to f$ in $L^1(X)$ and $\lip{f_n}\mathfrak {m} \to \bb f $ weakly in duality with $C_c(X)$. Hence arguing as for Proposition \ref{prop:1} one gets by approximation that
\begin{equation}
\label{eq:mollbv}
f\in {\rm BV}(X)\qquad\Rightarrow\qquad \H_t(f)\in {\rm BV}(X)\qquad \bb{\H_t(f)}\leq e^{-Kt}\mathcal H_t(\bb f ).
\end{equation}
Then, using the a priori estimates on the relative entropy of $\mathcal H_t(\mu)$ in terms of the mass of $\mu$ (see \cite{AmbrosioGigliMondinoRajala12}) one obtains that for a sequence of non-negative measures $(\mu_n)$ weakly converging to some measure $\mu$ in duality with $C_b(X)$ and $t>0$, the sequence $n\mapsto g_n:=\frac{\d\mathcal H_t(\mu_n)}{\d\mathfrak {m}}$ converges to $g:=\frac{\d\mathcal H_t(\mu)}{\d\mathfrak {m}}$ weakly in duality with $L^\infty(X)$. Therefore, for ${\mbox{\boldmath$\pi$}}$ as in the proof of Proposition \ref{prop:2} and $(f_n)\subset \Lip X$ converging to $f\in {\rm BV}(X)$ and so that $\lip{f_n}\mathfrak {m}\to \bb f $ weakly in duality with $C_b(X)$, we can pass to the limit in the inequality
\[
\begin{split}
\int|\H_t(f_n)(\gamma_1)-\H_t(f_n)(\gamma_0)|\,\d{\mbox{\boldmath$\pi$}}(\gamma)&\leq\iint_0^1\lip{\H_t(f_n)}(\gamma_t)|\dot\gamma_t|\,\d t\,\d{\mbox{\boldmath$\pi$}}(\gamma)\\
&\leq e^{-Kt}\iint_0^1\H_t(\lip{f_n})(\gamma_t)|\dot\gamma_t|\,\d t\,\d{\mbox{\boldmath$\pi$}}(\gamma),
\end{split}
\]
to deduce that
\[
\int|\H_t(f)(\gamma_1)-\H_t(f)(\gamma_0)|\,\d{\mbox{\boldmath$\pi$}}(\gamma)\leq e^{-Kt}\iint_0^1\frac{\d\mathcal H_t(|Df|_w)}{\d\mathfrak {m}}(\gamma_t)|\dot\gamma_t|\,\d t\,\d{\mbox{\boldmath$\pi$}}(\gamma).
\]
In particular, arguing as in the proof of Proposition \ref{prop:2} we get that
\begin{equation}
\label{eq:bvlip}
f\in BV\cap L^\infty(X),\ |Df|_w\leq C\mathfrak {m}\qquad\Rightarrow\qquad\lip{\H_t(f)}\leq e^{-Kt}\mathcal H_t\Big(\frac{\d|D f|_w}{\d\mathfrak {m}}\Big).
\end{equation}
Then following the same lines of thought of Proposition \ref{prop:3} and Theorem \ref{thm:main} the claim \eqref{eq:claimbv} follows.
Notice also that from \eqref{eq:mollbv} and with a truncation and mollification argument we deduce that
\begin{quote}
for $f\in {\rm BV}(X)$ with $\bb f\ll\mathfrak {m}$ there is a sequence $(f_n)\subset \Lip X$ such that $f_n\to f$ and $\lip{f_n}\to\frac{\d\bb f }{\d\mathfrak {m}}$ strongly in $L^1(X)$ as $n\to\infty$.
\end{quote}
In particular, the three notions of space $W^{1,1}(X)$ discussed in \cite{Ambrosio-DiMarino14} all coincide.
All this if the space is proper. It is very natural to expect that the same results hold even without this further assumption, but in the general case it seems necessary to define ${\rm BV}$ functions taking limits of locally Lipschitz functions, rather than Lipschitz ones (see the proof of Lemma 5.2 in \cite{Ambrosio-DiMarino14}). The problem then consists in the fact that the property \eqref{eq:BE} is not available for locally Lipschitz functions with local Lipschitz constant in $L^1$.
}\fr\end{remark}
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
|
\section{Algorithms}
\begin{varalgorithm}{OptPrice($B,v,c,\epsilon$)}
\caption{}
\label{alg:optprices}
\begin{algorithmic}
\For {$k = 1$ to $n$}
\State $p^{(k)}_i = \min(v_i/v_k, 1)$ for all $i$
\State $x^{(k)} = \arg\max_{x \in X(u,p^{(k)},B)} x \cdot (p^{(k)} - c)$
\Comment{$O(n \log n)$ computation}
\State $\mathrm{Profit}(k) = x^{(k)} \cdot (p^{(k)} - c)$
\EndFor
\State $k_{\max} = \arg\max_k \mathrm{Profit}(k)$
\State $p^* = p^{(k_{\max})}$, $x^* = x^{(k_{\max})}$
\For {$i = 1$ to $n$}
\If {$x^*_i = 0$}
\State $\hat{p}_i = 1$
\ElsIf{$x^*_i = 1$}
\State $\hat{p}_i = p^*_i - \epsilon$
\Else
\State $\hat{p}_i = p^*_i - \epsilon/2$
\EndIf
\EndFor
\State \textbf{return} $\hat{p}$
\end{algorithmic}
\end{varalgorithm}
\begin{varalgorithm}{LearnVal($\delta$)}
\caption{}
\label{alg:learnval}
\begin{algorithmic}
\State $p_i = 1 \quad \forall i$
\Comment{First price query}
\State $x \leftarrow \texttt{consumer}^1(p)$
\While {$\neg \exists i \text{ such that } 0 < x_i < 1$}
\Comment{Find some fractionally bought good}
\State Choose $k$ such that $x_k = 0$
\State $p_k = p_k - \delta$
\State $x \leftarrow \texttt{consumer}(p)$
\EndWhile
\State $j = i$ \Comment {$j$ is least-preferred purchased good}
\For {$k = j+1$ to $n$} \Comment{Learn $s_k$ for $j+1, \ldots, n$}
\While {$x_k = 0$}
\State $p_k = p_k - \delta$
\State $x \leftarrow \texttt{consumer}(p)$
\EndWhile
\State $v_k/v_j = p_k$
\EndFor
\State $s_k = (v_k/v_j)/(v_n/v_j) \quad \forall k \geq j$
\Comment {Renormalize ratios}
\For {$k = j-1$ to $1$} \Comment{Learn $s_k$ for $j-1, \ldots, 1$}
\For {$\alpha = 1$ to $\delta$ (in increments of $\delta$)}
\State $p_i = 1 \quad i \leq j-1$
\State $p_i = \alpha s_i \quad \forall i > j-1$
\State $x \leftarrow \texttt{consumer}(p)$
\If {$x_k > 0$}
\State $s_k = 1/\alpha$
\State \textbf{break}
\EndIf
\EndFor
\If {$s_k$ is undefined}
\Comment{$k$ was always bought}
\State \textbf{break}
\EndIf
\EndFor
\State \textbf{return} $s$
\State
\State $^{1}$The notation $x \leftarrow \texttt{consumer}(p)$ specifies the bundle bought by the consumer at prices $p$.
\end{algorithmic}
\end{varalgorithm}
\begin{varalgorithm}{ProfitMax($B, c, \delta, \epsilon$)}
\caption{}
\label{alg:complete}
\begin{algorithmic}
\State $\text{Profit} = 0$
\State $s \leftarrow \texttt{LearnVal}(\delta)$
\Comment{Learn $s_i$ ratios}
\For{$i = 1$ to $n$}
\If{$s_i = 0$}
\State $p_i = 1$
\EndIf
\EndFor
\State $v = \{s_i \mid s_i \geq 0\}$
\State $p \leftarrow \texttt{OptPrice}(B, v, c, \epsilon)$
\Comment{Compute optimal prices}
\While {$t \leq T$}
\Comment{Set prices optimally}
\State $x^t \leftarrow \texttt{consumer}(p)$
\State $\text{Profit} = \text{Profit} + x^t \cdot (p - c)$
\EndWhile
\end{algorithmic}
\end{varalgorithm}
\begin{varalgorithm}{ExogLearnVal($B$, $\delta$)}
\caption{}
\label{alg:learnvalexg}
\begin{algorithmic}
\State $\textrm{Fixed} = \emptyset, w_i = 0$ for all $i$. \Comment{Initialize the fixed coordinates, initially none}
\State $C_0 = \{z \in [0,1]^n \mid 0 \leq z_i \leq 1 \; \forall i\}$ \Comment{Initialize the set of consistent hypotheses}
\For{$t = 0$ to $T$}
\State Observe $p^t$
\State $z^t \leftarrow \texttt{sample}^1(C_t)$ \Comment{Sample a valuation uniformly from consistent set}
\State $v^t = (z^t_{-\textrm{Fixed}}, w_{\textrm{Fixed}})$ \Comment{Combine sampled and fixed coefficients}
\State Predict $\hat{x}^t \in \arg \max_{x \cdot p^t \leq B} x \cdot v^t$ \Comment{Predict according to sampled valuation}
\State $x_t \leftarrow \texttt{consumer}(p^t)$
\State $C_{t+1} = C_t$
\For{$i,j \not\in \textrm{Fixed} \mid x^t_i > x^t_j$} \Comment{Update Constraints}
\State $C_{t+1} = C_{t+1} \cap \{z_i/p^t_i \geq z_j/p^t_j\}$
\EndFor
\For{$i \not\in \textrm{Fixed},j \in \textrm{Fixed} \mid x^t_i > x^t_j$} \Comment{Update Constraints}
\State $C_{t+1} = C_{t+1} \cap \{z_i/p^t_i \geq w_j/p^t_j\}$
\EndFor
\For{$j \not\in \textrm{Fixed},i \in \textrm{Fixed} \mid x^t_i > x^t_j$} \Comment{Update Constraints}
\State $C_{t+1} = C_{t+1} \cap \{w_i/p^t_i \geq z_j/p^t_j\}$
\EndFor
\If{There exists $i\not\in \textrm{Fixed}$ such that $\mathrm{width}_i(C_{t+1}) < \delta/2$} \Comment{Start a new Epoch}
\For{Each $i\not\in \textrm{Fixed}$ such that $\mathrm{width}_i(C_{t+1}) < \delta/2$} \Comment{Fix each determined coordinate}
\State $\mathrm{Fix}(i)$
\EndFor
$C_{t+1} = \{z \in [0,1]^{n-|\textrm{Fixed}|} \mid 0 \leq z_i \leq 1 \; \forall i\not\in \textrm{Fixed}\}$ \Comment{Re-initialize in the unfixed \\ \hfill coordinates}
\EndIf
\EndFor \\
\Procedure{Fix}{$i$}
\State $\textrm{Fixed} = \textrm{Fixed} \cup \{i\}$
\State $z \leftarrow \texttt{sample}(C_{t+1})$ \Comment{Sample a value for coordinate $i$}
\State $w_i = \texttt{round}^2(z_i/\delta) \cdot \delta$ \Comment{Round sampled value to nearest discrete value and fix it}\\
\EndProcedure
\State $^{1}$\texttt{sample} is any polynomial algorithm for sampling uniformly at random from a convex polytope.
\State $^{2}$\texttt{round}($x$) rounds to the nearest integer.
\end{algorithmic}
\end{varalgorithm}
\newpage
\clearpage
\section {Section 3.1 Lemmas}
\textbf{Lemma 1.} \emph{Let $p^{(k)}$ be the pricing vector such that $p_i^{(k)} = \min(v_i/v_k, 1)$. For any consumer utility parameterized by $(u,B)$, there exists a $k$, and an $x \in X(u,p^{(k)},B)$, such that $\mathrm{OPT} = x \cdot (p^{(k)} - c)$. }
\proof
Fix $u$ and $B$. Consider a profit-maximizing price $p^*$, and corresponding bundle $x^* \in X(u,p^*, B)$, so that $\mathrm{OPT} = x^* \cdot (p^* - c)$. Let $O = \{i : x^*_i > 0\}$ be the set of purchased goods in $x^*$. If there is a fractionally purchased good in $x^*$ ($0 < x_f^* < 1$), we denote its index by $f$.
We note that there must exist a $\tau$ such that $r_i(u,p^*) \leq \tau$ whenever $i\not\in O$ and $r_i(u,p^*) \geq \tau$ whenever $i \in O$. In other words, in order for the bundle $x^*$ to maximize the consumer's utility, the bang-for-buck for every purchased good must be at least as large as the bang-for-buck for every unpurchased good. If there is a fractional good in $x^*$, we take $\tau$ to be $r_f(u,p^*)$. Otherwise, we take it to be $\max_{i \not\in O} r_i(u, p^*)$, the most desirable, unpurchased, good.
Given $(x^*,p^*)$, we can write the following linear program. We claim that any solution to this LP is also a profit-maximizing price. More precisely, if $p^{(LP)}$ is a solution to the LP, there is an $x^{(LP)} \in X(u,p^{(LP)},B)$ such that $\mathrm{OPT} = x^{(LP)}(p^{(LP)}-c)$. In what follows we prove this claim.
\begin{equation*}
\label{eq:lp}
\begin{aligned}
& \max && \sum_{i \in O} p_i && \\
& \; \text{s.t.} && v_i/p_i \geq \tau && \forall i \in O \\
&&& v_i/p_i \leq \tau && \forall i \notin O \\
&&& p_i \leq 1 && \forall i \in [n]
\end{aligned}
\end{equation*}
We can straightforwardly characterize any solution $p^{(LP)}$ to the LP. Note that the constraints on $p_i$ are disjoint, and therefore for each $i \in O$, $p_i$ can increased separately until a constraint is saturated. Thus, the LP is optimized by setting $p_i^{(LP)} = \min\{v_i/\tau, 1\}$ for each $i \in O$, and $p_i^{(LP)} \geq v_i/\tau$ for each $i \not\in O$ (which is always possible since $p^*$ is a feasible solution).
The LP constraints imply that, under $p^{(LP)}$, the consumer prefers any item in $O$ to any item not in $O$. Moreover, if there is a fractional good, the definition of $\tau$ ensures that any item in $O \setminus \{f\}$ is preferred to $f$. Finally, we know that $\sum_{i \in O} p_i^* \leq \sum_{i \in O} p^{(LP)}_i$ since $p^*$ is a feasible solution to the LP. This, along with the previous preference ordering, tells us that the consumer can saturate it's budget at least as quickly under $p^{(LP)}$ as under $p^*$. In other words, under $p^{(LP)}$, the consumer might saturate its budget before purchasing all items in $O \setminus \{f\}$, or might require a smaller allocation of item $f$.
More precisely, there exists an $x^{(LP)} \in X(u,p^{(LP)},B)$ such that point-wise $x^{(LP)}_i \leq x^*_i$. Thus, $\mathrm{OPT} = x^* (p^* - c) = B - x^* \cdot c \leq B - x^{(LP)} \cdot c = x^{(LP)} \cdot (p^{(LP)} - c)$, and so $p^{(LP)}$ must be a profit-maximizing price.
We have established that $p^{(LP)}$, which sets $p_i = \min(v_i/\tau,1)$, is profit-maximizing. All that's left to show is that we can take $\tau = v_k$ for some $k$. Consider again a profit-maximizing price and bundle $(p^*,x^*)$. Notice that for any $i \not\in O$, we can modify $p_i^* = 1$ and still have $(p^*,x^*)$ be profit-maximizing. That is, we are only making the unpurchased goods more undesirable. Notice that with this modification $\max_{i \not\in O} r_i(u,p^*) = \max_{i \not\in O} v_i = v_{k^*}$ for some $k^*$. Next consider modifying the price of the fractional good. The price of $f$ can increased to $v_f/v_{k^*}$ (if $v_f/v_{k^*} < 1$) and $1$ otherwise. This increases the price of $f$ while keeping it the fractionally purchased item, thus reducing the merchant's cost. This will result in $r_f = v_{k^*}$ or $r_f = v_f$. In either case, there exists a $(p^*,x^*)$ such that $\tau = v_k$ for some $k$. Thus $p^{(LP)}$ derived from $(p^*,x^*)$ sets $p_i = \min(v_i/v_k, 1)$ for some $v_k$.
\qed(Lemma 1)\\
\textbf{Lemma 2.}\emph{
For any $p,u,B$, $\max_{x \in X(u,p,B)} x \cdot (p - c) = B - x \cdot c$ can be computed in $O(n \log n)$ time.}
\proof
Let $p$ be an arbitrary price vector. $\max_{x \in X(u,p,B)} B - x \cdot c$ can be computed as follows. Let $r_i(u,p) = v_i/p_i$. Sort the $r_i$ in decreasing order, so that $r_{i_1} \geq ... \geq r_{i_n}$. The consumer will buy items in this order until the budget $B$ is exhausted. Thus, we can simulate the consumer's behavior, iteratively buying items and decrementing the budget. The consumer's behavior is uniquely specified unless there is some run of items with $r_{i_j} = r_{i_{j+1}} = ... = r_{i_{j+d}}$, and $B'$ budget remaining, where $\sum_{l=0}^{d} p_{i_{j+l}} > B'$. In other words, the consumer is indifferent between these items, and can make different selections to exhaust the remaining budget $B'$.
In that case, we know that for any bundle in $X(u,p,B)$, $x_{i_l} = 1$ if $l < j$, and $x_{i_l} = 0$ if $l > j+d$. For the remaining items, the merchant's profit is maximized when $x \cdot c$ is minimized. This occurs when the consumer saturates the remaining budget $B'$ while minimizing the cost $c$ to the merchant. This is an instance of min-cost knapsack wherein the size of the items are $p_{i_j},...,p_{i_{j+d}}$ and the cost of the items are $c_{i_j},...,c_{i_{j+d}}$. A solution can be computed greedily. Thus the most profitable bundle for $p$ can be computed with at most two sorts (first for $r_i$ then for $p_i/c_i$).
\qed (Lemma 2)\\
\noindent \textbf{Lemma 3}
\emph{There exists a price vector $\hat{p}$ which uniquely specified a bundle $\hat{x}$ such that for any $\epsilon > 0$, $\hat{x} \cdot (\hat{p} −- c) \geq \emph{OPT} - \epsilon$.}
Recall that the merchant would like the consumer to purchase the bundle $x^*$, which is a member of the set $X(u, p^*, B)$. Even if the merchant sets prices at $p^*$, there is no guarantee that the consumer will purchase $x^*$ rather than some other bundle in the set. Our goal is to output a vector $\hat{p}$ that results from perturbing $p^*$ slightly so that the consumer will always purchase some $\hat{x}$ that is arbitrarily close to $x^*$. For any good $i$ such that $x^*_i = 0$ (a good that consumer should not buy at all), we simply set $\hat{p}_i = 1$. For any good $i$ such that $x^*_i = 1$ (a good that that the consumer should buy in its entirety), we set $\hat{p}_i = p^*_i - \epsilon_0$. Finally, for any good $i$ such that $0 < x^*_i < 1$ (a good that the consumer should buy fractionally), we set $\hat{p}_i = p^* - \epsilon_0/2$. These perturbations ensure that the consumer will buy goods in the order desired by the merchant.
We have decreased each price by at most $\epsilon_0$, so the consumer might have additional $n\epsilon_0$ budget to spend. Recall that prices are chosen by the algorithm to be $\min(v_i/v_k, 1)$ for some $i,k$. Because values are discretized and lower-bounded, the minimum price possible is therefore $\delta$. Consider setting $\epsilon_0 = \delta \epsilon/n$, which yields $\delta \epsilon$ additional budget. Then the consumer can afford to purchase at most an additional $\delta \epsilon/\delta = \epsilon$ fraction of a good. In the worst case, if this good is of maximum cost 1, the merchant will incur an additional cost of $\epsilon$.
\qed (Lemma 3)
\section{Proof of Theorem 4}
\begin{theorem4}
\label{thm:exogenous}
Algorithm \emph{\texttt{ExogLearnVal}} runs in polynomial time per round, and with probability $1-\beta$ makes at most $O\left(n^2\log(1/\delta) + n\sqrt{\log(1/\beta)\log(1/\delta)}\right)$ mistakes over any sequence of adaptively chosen price vectors.
\end{theorem4}
\proof
We prove a bound on the number of mistakes made by the algorithm in a single epoch (recall that each epoch ends when a new coordinate is fixed). Since there are only $n$ coordinates, and hence at most $n$ epochs, our final mistake bound is at most $n$ times the mistake bound per epoch, giving the theorem.
Consider an epoch $i$ in which there remain $d \leq n$ unfixed coordinates. We will track the $d$-dimensional volume of the sets $C_t$ during this epoch. Let $S_i$ be the first stage of the epoch, and let $F_i$ be the final stage of the epoch. Note that $\mathrm{Vol}(C_{S_i}) = 1$ (because $C_{S_i}$ is always initialized to be the $d$-dimensional hypercube). Note also that for $S_i \leq t < F_i$, any hypothesis in $C_t$ that leads to an incorrect prediction of $x^t$ is eliminated from $C_{t+1}$. Hence, if $M_t$ is the indicator random variable specifying whether our algorithm makes a mistake at round $t$, it is also the indicator random variable specifying whether our algorithm sampled a hypothesis that will be eliminated at the next round. Because we sample our hypothesis at round $t$ at random from $C_t$, we have:
$$\mathrm{E}[M_t] = 1 - \frac{\mathrm{Vol}(C_{t+1})}{\mathrm{Vol}(C_t)}$$
Note that we can write the volume of $C_{F_i}$ as a telescoping product:
$$\mathrm{Vol}(C_{F_i}) = \mathrm{Vol}(C_{S_i})\prod_{t=S_i}^{F_i-1}\frac{\mathrm{Vol}(C_{t+1})}{\mathrm{Vol}(C_{t})}=\prod_{t=S_i}^{F_i-1}\frac{\mathrm{Vol}(C_{t+1})}{\mathrm{Vol}(C_{t})}$$
Note also that before the end of an epoch, the width of $C_t$ in every coordinate is at least $\delta/2$. Hence we know $\mathrm{Vol}(C_{F_i}) \geq \delta^d \geq \delta^n$. Combining these facts, we can write:
$$\delta^n \leq \prod_{t=S_i}^{F_i-1}\frac{\mathrm{Vol}(C_{t+1})}{\mathrm{Vol}(C_{t})}
= \prod_{t=S_i}^{F_i-1} \left(1-\mathrm{E}[M_t]\right)
\le\prod_{t=S_i}^{F_i-1} \exp(-\mathrm{E}[M_t])
= \exp\left(-\mathrm{E}\left[\sum_{t=S_i}^{F_i-1} M_t\right]\right)$$
Solving for the expected number of mistakes made in a single epoch,
we find that:
$$\mathrm{E}\left[\sum_{t=S_i}^{F_i} M_t\right] \leq \mathrm{E}\left[\sum_{t=S_i}^{F_i-1} M_t\right] + 1 \leq 1 + n\ln\left(\frac{1}{\delta}\right) = O\left( n\ln\left(\frac{1}{\delta}\right)\right)$$
Now consider the expected number of mistakes made by our algorithm for its entire run, $t \in \{1,\ldots,T\}$. Since we can partition each time step into one of at most $n$ epochs, and have a bound on the expected number of mistakes in each epoch, by linearity of expectation, we have:
$$\mathrm{E}\left[\sum_{t=1}^{T} M_t\right] = \sum_{i=1}^n \mathrm{E}\left[\sum_{t=S_i}^{F_i} M_t\right] \leq n + n^2\ln\left(\frac{1}{\delta}\right) = O\left( n^2\ln\left(\frac{1}{\delta}\right)\right)$$
Note that each of the random variables $M_i$ is independent and bounded in $[0,1]$. We can therefore apply a multiplicative Chernoff bound. For any $\epsilon < 1$, we have:
$$\Pr\left[\sum_{t=1}^{T} M_t \geq (1+\epsilon)\mathrm{E}\left[\sum_{t=1}^{T} M_t\right]\right] \leq \exp\left(\frac{-\epsilon^2}{3}\mathrm{E}\left[\sum_{t=1}^{T} M_t\right]\right)$$
Setting the right hand side to be at most $\beta$, plugging in our bound on $\mathrm{E}\left[\sum_{t=1}^{T} M_t\right]$ and solving for $\epsilon$ allows us to take
$$\epsilon = O\left(\sqrt{\frac{\ln(1/\beta)}{n^2\ln(1/\delta)}}\right)$$
Plugging this into the Chernoff bound proves the theorem.
\qed (Theorem 4)
\subsection{Our Results}
We first consider the case of a monopolist merchant who has the ability to set prices arbitrarily, and is facing a consumer with an
unknown linear utility function. In this setting, we give an algorithm with bounded regret with respect to the optimal (profit-maximizing) set of prices in hindsight.
Our argument proceeds in two steps. We first show that, \emph{if we knew} the consumer's utility function $u$, then we could efficiently
compute the optimal profit-maximizing prices $p^*$:
\begin{theorem1}[Informal]
There is an efficient algorithm \emph{(}running in time $O(n^2\log n)$\emph{)}, which given as input the linear consumer utility function $u$, outputs the profit-maximizing prices $p^*$.
\end{theorem1}
The analysis of this algorithm first assumes we know only the {\em set\/} of goods purchased by the consumer under
optimal prices (but not the optimal prices themselves), and
introduces a family of linear programs with one free parameter. We then show there is a small set of
values for this parameter, one of which yields the optimal prices.
Note that although the \emph{consumer}'s optimization problem when selecting a bundle to purchase given prices
is simply a fractional knapsack problem, the problem of computing optimal prices is substantially more complex.
The optimal price vector $p^*$ is actually a subgame perfect Nash equilibrium strategy for the merchant in a two-stage extensive form
game between the merchant and the consumer (the merchant first picks prices, and then the consumer best responds).
Viewed in this way, the fractional knapsack problem that the consumer solves at the second stage is simply her best response function; what
we give is an algorithm for computing the merchant's subgame perfect equilibrium strategy in the first stage of this game. Note that doing this in polynomial time is non-trivial, because the merchant's strategy space is continuous (and even after discretization, is exponentially large).
We next give an algorithm that learns the consumer's unknown linear utility function by making price queries.
A price query specifies a price vector, and receives in return the bundle purchased by the consumer.
The key metric of interest here is the number of queries necessary to learn the utility function. We prove the following theorem:
\begin{theorem2}[Informal]
There is an efficient algorithm that learns, after at most $O(n)$ price queries, the utility coefficients for all goods except those that are so
preferred they will be bought regardless of prices.
\end{theorem2}
This algorithm has two phases. In the first phase, after setting all prices to 1 (the maximum possible price),
the algorithm gradually lowers the price of unpurchased items in succession until they are purchased,
thus learning the ratio of their utility coefficient to that of the least preferred good purchased under the price vector of all 1s. Note that learning a \emph{relative} rather than exact coefficient value is sufficient because the bundle bought by the consumer (i.e. the solution to the knapsack problem) is unchanged by a scaling of the coefficients. Indeed, because of this property, it is impossible for any algorithm to learn the exact values.
The harder coefficients to learn are those corresponding to goods purchased even when all prices are 1 --- these
are the consumer's most preferred goods. Some of these are learned by gradually lowering the prices of unpurchased goods
until a switch of purchased goods occurs;
for the ones that cannot be learned via this procedure, we prove that they are so favored they will be purchased
under any prices, so learning their coefficients is not necessary for price optimization.
Finally, we put these two algorithms together to prove our first main result:
\begin{theorem3}[Informal]
There is a price-setting algorithm that, when interacting with a consumer with an unknown linear utility
function for $T$ rounds, achieves regret $O(n^2/T)$ to the profit obtained by the optimal \emph{(}profit-maximizing\emph{)} price vector.
\end{theorem3}
In the last part of the paper, we consider the case of a commodity merchant who does not have the power to set prices.
The merchant wishes to predict the bundles that a consumer with an unknown linear utility function will buy, in the face of a stream of
arbitrary price vectors.
Here, the main quantity of interest is how many mistakes we make (by predicting the incorrect bundle)
in the worst case over both consumer utility functions and sequences of price vectors. We call this the {\em mistake
bound\/} of the algorithm (by analogy to the mistake bounded model of learning). Here we prove our second main result:
\begin{theorem4}[Informal]
There exists a polynomial time algorithm in the online exogenous price model that
has a mistake bound of $O(n^2)$ with high probability.
\end{theorem4}
\iffalse
This algorithm can be viewed as a continuous, randomized, iterative version of the classical halving algorithm, implemented efficiently using the ability to sample from convex bodies in polynomial time. We show, by carefully tracking the volume of a polytope representing a set of consistent hypotheses, that after $O(n)$ mistakes (in expectation), the algorithm is guaranteed to have learned at least one coefficient of the consumer's valuation function exactly. Once a coordinate is learned, we can fix it and restart the algorithm in a lower dimensional space. Since there are at most $n$ coefficients, this guarantees that the algorithm never makes more than $O(n^2)$ mistakes in expectation. The algorithm is implemented such that the mistake events are independent, and so we can apply concentration inequalities to get a high probability mistake bound.
\fi
\subsection{Related Work}
The work most directly related to our results is the recent paper of Balcan et al. \cite{BDMUV14},
which was conducted independently and concurrently.
They study the problem of learning from
revealed preferences in various settings, including in a query model
related to the model we study here. Our ``price queries'' differ
slightly from the queries in the Balcan et al. model, in that our
learner can only specify prices, whereas the learner from Balcan et
al. can specify prices, as well as the consumer's budget with each
query. However, the main distinction between our work and theirs is
that our goal is profit maximization (even if we do not exactly learn
the buyer's utility function), and the goal of Balcan et al \cite{BDMUV14} is to exactly
learn the buyer's utility function --- they do not consider the
profit maximization problem.
More broadly, there is a long line of work on the revealed preference problem, which was first introduced by Samuelson \cite{sam}. For a textbook introduction to the standard model, see \cite{mwg} and \cite{Rub12}, and for a survey of recent work, see \cite{varian}. Most previous efforts have focused on the construction of utility functions that explain a finite sequence of price/bundle observations. Afriat's Theorem \cite{afriat} is the seminal result in this field, and proves that a sequence of observations is \emph{rationalizable} (i.e. can be explained by a utility function) if and only if the sequence is rationalizable by a piecewise linear, monotone, concave utility function. The proof is by construction, and therefore yields a ``learning'' algorithm for this large class of utility functions. However, the hypothesis generated has description length proportional to the number of observations, and hence although it can \emph{explain} previous observations, it usually does not generalize to \emph{predict} the bundles purchased given new price vectors.
The problem of finding a utility function that is both consistent and predictive was first considered by Beigman and Vohra \cite{vohra}, who formalize the statement that ``Afriat's Theorem Learners'' do not generalize. \cite{vohra} studies a PAC-like learning model in which they assume a distribution over observations and seek to find a hypothesis that performs well on future observations drawn from the same distribution. Their results essentially show that it is only possible to find predictive hypotheses if we restrict the class of allowable utility functions beyond those that are rationalizable. Roth and Zadimoghaddam \cite{roth} extended this line of work by providing computationally efficient learning algorithms for two specific classes of utility functions --- linear and linearly separable and concave utility functions. Cummings, Echenique, and Wierman \cite{CEW14} consider the revealed preferences problem when the consumer can strategically choose bundles to subvert the merchant's learning. In this setting, they show that without assuming the consumer's utility function is linearly separable, the merchant is unable to learn anything. Like this previous work, we also seek to find predictive hypotheses for the class of linear utility functions, but we consider two new learning models: one in which prices are directly controlled, rather than observed (which corresponds to a query model of learning), and furthermore we wish to learn optimal prices;
and one in which prices are chosen adversarially and adaptively, and arrive online (which corresponds to online learning in the mistake bound model).
Our results in the second model are inspired by the classic halving algorithm for the online learning setting, which is credited to Littlestone \cite{littlestone}. The approach is simple but powerful: given a finite set of hypotheses and a sequence of observations, predict according to the majority of remaining hypotheses, and then discard all hypotheses that made a mistake. To implement the algorithm efficiently, we instead maintain a continuous hypothesis space from which we predict using a randomly sampled hypothesis (rather than predicting using a majority vote). We track the volume of the hypothesis space (rather than the number of consistent hypotheses), and show that after a bounded number of mistakes, we must have learned one coefficient of the consumer valuation function.
\section{Introduction}
\input{introduction}
\section{Preliminaries}
\input{preliminaries}
\section{Maximizing Profit in the Price-Setting Model}
We begin by considering the first model, in which the merchant controls prices, and seeks to maximize profit. First we show that, given the coefficients $v_i$ of the consumer's linear utility function,
we can efficiently compute the profit-maximizing prices. We will then combine this algorithm
with a query algorithm for learning the coefficients, thus yielding an online no-regret pricing algorithm.
\subsection{Computing Optimal Prices Offline}
\input{optprices}
\subsection{Learning Consumer Valuations}
\input{learnval}
\subsection{Putting It All Together}
\input{puttogether}
\section{Predicting Bundles in the Exogenous Price Model}
\input{exogenous}
\section{Conclusion and Future Work}
Our work provides efficient learning algorithms for two new preference learning models. We leave as an open question whether it is possible to remove our assumptions of discretized and lower-bounded valuations. In the price-setting model, one might also wish to devise an algorithm in which the merchant can approximately optimize profits \emph{during} the learning phase. Finally, there are several variants of our model to be considered in future work, such as multiple buyers with different utility functions, and stochastic budgets that vary daily.
\newpage
\clearpage
\bibliographystyle{plain}
|
\section{Introduction}
Models of epidemic spreadings as dynamical processes occurring on a graph appear in various contexts besides epidemiology~\cite{Hethcote_review,Dorogovtsev_review,Newman_review,Boccaletti_review,Barrat_book}; for instance social sciences study viral marketing campaigns aimed at propagating new social trends, and in economy it is crucial to understand cascading effects potentially leading to the bankrupt of financial institutions. In these models individual agents are located on the vertices of a graph, and their state (healthy or contaminated for instance) evolve in time according to the state of their neighbours, the edges of the graph representing the contacts between agents that can possibly transmit the illness from one contaminated agent to an healthy one.
There is a great diversity in the details of these models: the dynamics can occur in continuous (asynchronous) or discrete time, according to deterministic or random rules, the state of an agent can be boolean (healthy or contaminated) or describe several levels of contamination, and finally the dynamics can be monotonous or not. To precise this last point, a dynamics is said monotonous if the states of an agent always occur in the same order in time, for instance in the Susceptible-Infected-Recovered (SIR) model the only allowed transitions are S$\to$I and I$\to$R, a Recovered individual being immune forever, whereas in the SIS model an agent can become infected several times in a row. In this paper we will concentrate on a simple monotonous dynamics, that evolve deterministically in discrete time, with inactive (Susceptible) variables becoming active (Infected) when their number of active neighbours reach some threshold, and then remain active for ever. For this reason it is called the threshold model, see~\cite{thr_socio} for a version introduced in sociology with an underlying complete graph, and~\cite{bootstrap} for its first appearance in physics under the name of bootstrap percolation (on random regular graphs).
Given one specific dynamical model there are many different questions that can be asked. The first, a priori simplest, issue concerns the time evolution of the system from a random initial condition, taking the initial state of each agent as an independent random variable. For monotonous dynamics a stationary state is reached after some time, and one can wonder whether the epidemic has invaded the whole graph (in other words whether it percolates) in this final state. The probability of this event obviously depends on the fraction of infected vertices in the initial condition, and this may lead to phase transitions for certain class of graphs; see~\cite{AiLe88,Ho03,BaBoetal12} for such a study of the bootstrap percolation on finite-dimensional lattices, and~\cite{bootstrap,BaPi07,KaNe10,Lelarge2012,janson2012,ShMo14,SIR_Bohman,SIR_Janson} for various type of dynamics on random graphs. In particular one finds for the bootstrap percolation on random regular graphs a phase transition at some initial critical density ${\theta_{\rm r}}$ (dependent on the degree of the graph and the threshold of activation): with high probability initial conditions with a fraction $\theta$ of active vertices (without correlations between the sites) are percolating if and only if $\theta > {\theta_{\rm r}}$.
Besides these studies of the ``forward'' (or ``direct'') time evolution, which are somehow simplified by the independence assumption for the initial state variables, one can also formulate more difficult inference and optimization questions. An example of the former type is to infer some information on the initial state given a snapshot of the epidemic after some time evolution~\cite{inference_Shah,inference_Pinto,inference_Torino,inference_Orsay}; this ``inverse problem'' is particularly relevant in epidemiology in the search of the ``zero patient'' who triggered the spreading of an illness. For what regards the latter type of questions, the design of an efficient vaccination campaign can indeed be seen as an optimization problem: find the smallest set of nodes (to minimize the economical and social cost) whose vaccination will prevent the epidemic to reach a given fraction of the population~\cite{vaccination_Torino}. We shall actually consider in this paper the somehow reverse optimization problem, namely targeting a small set of initially active sites that lead to the largest possible propagation of the contagion. This obviously makes more sense in the perspective of viral marketing, in which it was first considered~\cite{Kempe} than in the epidemiological one; the initial adopters of a new product, that can be financially incited to do so, are expected to convince most of their acquaintances and progressively the largest possible part of the population. From this point of view the additional constraint that the propagation should be as fast as possible is also a relevant one.
More precisely, one can define two versions of this optimization problem: (i) given a fixed number of initially active agents, choose them in order to maximize the number of active agents at some fixed later time, or in the final state of the propagation; (ii) find the minimal number of initially active agents such that all the agents are active, again after some time or in the final state. We will concentrate on the latter version of the problem but part of our analysis applies to both. These optimization problems are known to be hard from a (worst-case) computational complexity point of view~\cite{Kempe,Chen,Dreyer09}, even to approximate. Exhibiting minimal percolating sets for bootstrap percolation on finite dimensional lattices is relatively easy thanks to their regular structures, but more refined extremal problems are also relevant in this case, see for instance~\cite{Morris09,maxtime2d}. The understanding of these optimization problems seems less advanced in the case of sparse random graphs. There exist upper and lower bounds on the size of minimal contagious sets~\cite{Dreyer09,Ackerman10,bounds12}, some based in particular on the expansion properties of such graphs~\cite{bounds_Amin}. One particular case of the optimization problem (when the threshold of activation is equal to the degree of the vertex minus one) is actually equivalent to the decycling number problem of graph theory~\cite{decycling_Beineke} (also known as minimal Feedback Vertex Set), which was settled rigorously for 3-regular random graphs in~\cite{decycling} (this paper also contains bounds for higher degrees). As this last point unveils the notion of minimal contagious sets is connected in some special cases to many other problems in graph theory; one way to see this connection is to picture the inactive sites of the initial condition as particles to be put on the graph. One wants to pack as many as possible of them (to obtain a contagious set of minimal size), yet they do have some kind of repulsive interactions because of the constraint of complete percolation at a later time. This is particularly clear when the threshold of activation is equal to the degree for all vertices: the problem is then exactly equivalent to the hard-core particle model, also known as independent set or vertex cover.
The strategy we shall follow to determine the minimal size of contagious sets of sparse random graphs will be the same as in~\cite{Torino1,Torino2}, namely a reformulation under the form of a statistical mechanics model which can be treated with the so-called cavity method~\cite{cavity,cavity_T0,MeZe,MeMo_book}. This (heuristic) method yields predictions for any interacting model defined on a sparse random graph; its use in the context of random constraint satisfaction problems led to the discovery of a very rich phenomenology of phase transitions~\cite{MeZe,KrMoRiSeZd}, with many of these predictions later confirmed rigorously~\cite{MoraMezard05b,clus_rig_Fede,AC08,molloy_col_freezing,amin_ksat,amin_cond_col}. Let us emphasise in particular the determination of the maximal size of independent sets of random regular graphs (which as we saw is a problem related to the present one), for which the predictions of the cavity method (see~\cite{is_japan} and references therein) have been recently rigorously confirmed (for graphs of large enough but finite degree) in~\cite{is_Sly}. Another example in the context of graph theory is the study of matchings in random graphs, where the cavity method~\cite{ZdMe_matchings} has also been proved to be correct~\cite{BoLeSa_matchings}. The main originality of our contribution with respect to~\cite{Torino1,Torino2} is the use of a more refined version of the cavity method (i.e. incorporating the effects of replica symmetry breaking), and an analytical study of the limit where the time at which the complete activation is required is sent to infinity.
The rest of the article is organized as follows.
In Sec.~\ref{sec_def_and_res} we define precisely the dynamics under study, recall briefly some known results for random initial conditions, formulate the optimization problem and propose various interpretations of it, and for the convenience of the reader we summarize the main results to be obtained in the following.
In Sec.~\ref{sec_cavity} we derive the cavity method equations, both at the replica symmetric and one step of replica symmetry breaking level.
The solution of these equations for random regular graphs is presented in Sec.~\ref{sec_results_rrg}, which contains the main analytical results of this work.
Sec.~\ref{sec_single_sample} is devoted to single sample numerical experiments, where we confront the analytical predictions with the optimized initial configurations obtained with two kind of algorithms (a simple greedy one and a more involved procedure based on message passing).
We finally draw our conclusions and present perspectives for future work in Sec.~\ref{sec_conclu}. The most technical parts of the computations are deferred to two Appendices.
\section{Definitions and main results}
\label{sec_def_and_res}
\subsection{Definition of the dynamics}
\label{sec_def_dyn}
Let us consider a graph on $N$ vertices (or sites), $G=(V,E)$, with the vertices labelled as $V=\{1,\dots,N\}$, and the number of edges denoted $|E|=M$. The dynamical process under study concerns the evolution of variables ${\sigma}_i^t$ on the vertices, ${\sigma}_i^t=0$ (resp. $1$) if the vertex $i$ is inactive (resp. active) at time $t$. We shall denote ${\underline{\sigma}}^t=({\sigma}_1^t,\dots,{\sigma}_N^t)$ the global configuration at time $t$. The latter is determined by the initial condition ${\underline{\sigma}}$ at the initial time, ${\underline{\sigma}}^0={\underline{\sigma}}$, and then evolves subsequently in a deterministic and parallel way, in discrete time, according to the rules:
\beq
{\sigma}_i^{t} = \begin{cases} 1 & \text{if} \ {\sigma}_i^{t-1}=1 \\
1 & \text{if} \ {\sigma}_i^{t-1}=0 \ \text{and} \ \underset{j \in {\partial i}}{\sum} {\sigma}_j^{t-1} \ge l_i \\
0 & \text{otherwise} \end{cases} \ ,
\label{eq_dynrules}
\end{equation}
where ${\partial i}$ is the set of neighbours of $i$ on the graph, and $l_i$ is a fixed threshold for each vertex; we will also use $d_i = |{\partial i}|$ to denote the degree of vertex $i$. The dynamics is monotonous (irreversible), an active site remaining active at all later times, an inactive site $i$ becoming active if its number of active neighbours at the previous time crosses the threshold $l_i$. Note that the configuration ${\underline{\sigma}}^t$ at time $t$ is a deterministic function of the initial condition ${\underline{\sigma}}={\underline{\sigma}}^0$, and that by monotonicity one can define the final configuration ${\underline{\sigma}}^{\rm f}=\underset{t \to \infty}{\lim} {\underline{\sigma}}^t$, this stationary configuration being reached in a finite number of steps for all finite graphs.
It turns out that the final configuration ${\underline{\sigma}}^{\rm f}$ is also the one reached by a sequential dynamics in which at each time step only one site $i$ with at least $l_i$ active neighbours is activated; a moment of thought reveals the independence of the final configuration with respect to the order of the updates. ${\underline{\sigma}}^{\rm f}$ is indeed the smallest configuration (considering the partial order ${\underline{\sigma}} \le {\underline{\sigma}}'$ if and only if ${\sigma}_i \le {\sigma}'_i$ for all vertices) larger than the initial condition ${\underline{\sigma}}$, such that no further site can be activated. It will sometimes be useful in the following to think of this process in a dual way, corresponding to the original presentation of bootstrap percolation in~\cite{bootstrap}, namely to consider that inactive sites are sequentially removed if they have less than a certain number of inactive neighbours. An equivalent definition of ${\underline{\sigma}}^{\rm f}$ is thus given by the inactive sites it contains, that form the largest set (with respect to the inclusion partial order) contained in the set of inactive sites of ${\underline{\sigma}}$, and such that in their induced graph the degree of site $i$ is larger or equal than $d_i-l_i+1$; they form thus a (generalized inhomogeneous version of the) core of the initially inactive sites.
\subsection{Reminder of the behaviour for random initial conditions on random regular graphs}
\label{sec_reminder_random}
To put in perspective the optimization problem to be studied in this paper it is instructive to first recall briefly some well-known results for the evolution from a random initial configuration~\cite{bootstrap,BaPi07}. For the sake of simplicity let us consider $G$ to be a $k+1$-random regular graph (i.e. a graph drawn uniformly at random among all graphs in which every vertex has degree $k+1$), with a uniform threshold for activation set to $l_i=l$ for all vertices. Suppose that the states of the vertices in the initial condition are chosen randomly, independently and identically for each vertex, with a probability $\theta$ (resp. $1-\theta$) for a vertex to be active (resp. inactive). The probability for one vertex $i_0$ to be active at some time $t+1$, denoted $x_{t+1}$, can be computed from the following equation:
\beq
x_{t+1} = \theta + (1-\theta) \sum_{p=l}^{k+1} \binom{k+1}{p} {\widetilde{x}}_t^p (1-{\widetilde{x}}_t)^{k+1-p} \ .
\label{eq_random_x}
\end{equation}
Indeed such a vertex was either active in the initial condition, or has seen at least $l$ of its neighbours activate themselves before time $t$, and without the participation of $i_0$. The probability ${\widetilde{x}}_t$ of this last event obeys the recursive equation
\beq
{\widetilde{x}}_{t+1} = \theta + (1-\theta) \sum_{p=l}^k \binom{k}{p} {\widetilde{x}}_t^p (1-{\widetilde{x}}_t)^{k-p} \ ,
\label{eq_random_tx}
\end{equation}
with a number of participating neighbours reduced from $k+1$ to $k$ as $i_0$ has to be supposed inactive here. The initial condition for these equations is $x_0={\widetilde{x}}_0=\theta$. In the limit $t \to \infty$ of large times ${\widetilde{x}}_t \to {\widetilde{x}}_\infty(\theta)$, the smallest fixed-point in $[0,1]$ of the recursion (\ref{eq_random_tx}). For each $k\ge 2$ and $l$ with $2\le l \le k$ there exists a threshold ${\theta_{\rm r}}(k,l)$ such that ${\widetilde{x}}_\infty(\theta)$ is equal to 1 for $\theta > {\theta_{\rm r}}$, strictly smaller than 1 for $\theta<{\theta_{\rm r}}$. From Eq.~(\ref{eq_random_x}) one realizes that the same statement applies to $x_\infty(\theta)$, hence ${\theta_{\rm r}}$ is the threshold for complete activation (percolation) from a Bernouilli random initial condition with probability $\theta$ for each active site. Studying more precisely Eq.~(\ref{eq_random_tx}) one realizes that for $l=k$ the transition is continuous ($x_\infty(\theta_{\rm r}^-)=1$), with an explicit expression for the threshold, ${\theta_{\rm r}}(k,k)=\frac{k-1}{k}$. For $2\le l \le k-1$ the transition is discontinuous ($x_\infty(\theta_{\rm r}^-)<1$), the threshold ${\theta_{\rm r}}$ is obtained as the solution of the equations:
\beq
\begin{cases}
{\widetilde{x}_{\rm r}}= {\theta_{\rm r}} + (1-{\theta_{\rm r}}) \overset{k}{\underset{p=l}{\sum}} \binom{k}{p} {\widetilde{x}_{\rm r}}^p (1-{\widetilde{x}_{\rm r}})^{k-p} \\
1 = (1-{\theta_{\rm r}}) l \binom{k}{l} {\widetilde{x}_{\rm r}}^{l-1} (1-{\widetilde{x}_{\rm r}})^{k-l}
\end{cases} \ ,
\label{eq_tandx_r}
\end{equation}
where ${\widetilde{x}_{\rm r}}={\widetilde{x}}_\infty(\theta_{\rm r}^-)$ is the value of the fixed-point of (\ref{eq_random_tx}) at the bifurcation where it disappears discontinuously. For $l=2$ these equations can be solved explicitly and yield
\beq
{\theta_{\rm r}}(k,l=2) = 1- \frac{(k-1)^{2k-3}}{k^{k-1}(k-2)^{k-2}}\ .
\end{equation}
For generic values of the parameters $k,l$ there is no explicit expression of ${\theta_{\rm r}}$, as (\ref{eq_tandx_r}) are algebraic equations of arbitrary degree; some numerical values of ${\theta_{\rm r}}$ will be given in Table \ref{table_res_Tinfty}. For a given value of $k$ the threshold ${\theta_{\rm r}}(k,l)$ is growing with $l$: if an initial condition leads to complete activation for some parameter $l$ it will also be activating under the less constrained dynamics with $l'<l$.
The relevant range for the threshold parameter $l$ in this study of random initial conditions is $2\le l \le k$. Indeed for $l=0$ after one step the configuration is completely active regardless of ${\underline{\sigma}}^0$, for $l=1$ a single active site (per connected component) in the initial configuration is enough to activate the whole graph, hence in these two cases ${\theta_{\rm r}}=0$. On the other hand if $l=k+1$ one has ${\theta_{\rm r}}=1$: any pair of adjacent inactive sites in the initial condition will remain inactive for ever, and the number of such pairs is linear in $N$ as soon as $\theta<1$.
Note that the recursion equations (\ref{eq_random_x},\ref{eq_random_tx}) are exact if the neighbourhood up to distance $t$ of the vertex $i_0$ is a regular tree of degree $k+1$. The limit $t\to\infty$ can be taken in this way only if the graph considered is an infinite regular tree. A rigorous proof that this reasoning is in fact correct also for the large size limit of random regular graphs (that converge locally to regular trees) can be found in~\cite{BaPi07}.
\subsection{Definition of the optimization problem over initial conditions}
\label{sec_def_optimization}
Let us now come back to a general graph $G$ with some thresholds $l_i$ for vertex activation, and consider the minimal fraction of active vertices in an initial configuration that activates the whole graph, i.e.
\beq
{\theta_{\rm min}}(G,\{l_i\}) = \frac{1}{N}\min_{\underline{\sigma}} \left\{ \sum_{i=1}^N {\sigma}_i \ | \ {\sigma}_i^{\rm f} = 1 \ \forall i \right\} \ .
\end{equation}
This corresponds to the minimal size of a contagious (or percolating) set, divided by the total number of vertices. Following~\cite{Torino1,Torino2} it will turn out useful to introduce another parameter $T$ (a positive integer) in this optimization problem, and impose now that the fully active configuration is reached within this time horizon $T$:
\beq
{\theta_{\rm min}}(G,\{l_i\},T) = \frac{1}{N}\min_{\underline{\sigma}} \left\{ \sum_{i=1}^N {\sigma}_i \ | \ {\sigma}_i^T = 1 \ \forall i \right\} \ .
\end{equation}
Obviously for any finite graph ${\theta_{\rm min}}(G,\{l_i\},T)$ decreases when $T$ increases and has ${\theta_{\rm min}}(G,\{l_i\})$ as its limit for $T\to\infty$. To turn the computation of ${\theta_{\rm min}}$ into a form more reminiscent of statistical mechanics problems we shall introduce a probability measure over initial configurations:
\beq
\eta({\underline{\sigma}}) = \frac{1}{Z(G,\{l_i\},T,\mu,\epsilon)}
e^{\underset{i=1}{\overset{N}{\sum}}[\mu {\sigma}_i - \epsilon (1-{\sigma}_i^T)]}
\ ,
\label{eq_eta_us}
\end{equation}
where ${\underline{\sigma}}^T$ is as above the configuration obtained after $T$ steps of the dynamics starting from the configuration ${\underline{\sigma}}={\underline{\sigma}}^0$, the $\mu$ and $\epsilon$ are for the time being arbitrary parameters, and the partition function $Z$ ensures the normalization of this law. The parameter $\mu$ is a ``chemical potential'' that controls the fraction of initially active vertices (if $\epsilon=0$ the measure $\eta$ reduces to the Bernouilli measure), while $\epsilon$ is the cost to be paid for each site $i$ inactive at the final time $T$. In particular if $\epsilon=+\infty$ one has
\beq
\eta({\underline{\sigma}}) = \frac{1}{Z(G,\{l_i\},T,\mu,\epsilon=+\infty)}
e^{\mu \underset{i=1}{\overset{N}{\sum}} {\sigma}_i} \prod_{i=1}^N {\mathbb{I}}({\sigma}_i^T=1)
\ ,
\end{equation}
with ${\mathbb{I}}(A)$ is the indicator function of the event $A$, the measure is thus supported by activating initial configurations (within the time horizon $T$). It is then obvious that the knowledge of $Z$ allows to deduce the sought-for minimal density ${\theta_{\rm min}}$, as
\beq
{\theta_{\rm min}}(G,\{l_i\},T) = \lim_{\mu \to -\infty}\frac{1}{\mu} \frac{1}{N} \ln Z(G,\{l_i\},T,\mu,\epsilon=+\infty) \ .
\end{equation}
Actually one can gain more information from the whole dependency of the partition function on $\mu$. Suppose indeed that the number of initial configurations with a fraction $\theta$ of active vertices that activate the whole graph in $T$ steps is, at the leading exponential order, $e^{N s(\theta)}$, with an entropy density $s(\theta)$ of order one with respect to $N$. Then this entropy density can be computed, in the large $N$ limit, as a Legendre transform of the logarithm of the partition function. More precisely, defining the free-entropy density $\phi$ as
\beq
\phi(G,\{l_i\},T,\mu,\epsilon=+\infty) = \frac{1}{N} \ln Z(G,\{l_i\},T,\mu,\epsilon=+\infty) \ ,
\end{equation}
the evaluation of the sum over configurations in the definition of $Z$ via the Laplace method yields in the large $N$ limit:
\beq
\phi(G,\{l_i\},T,\mu,\epsilon=+\infty) = \sup_{\theta \in [{\theta_{\rm min}},1]} \,[ \mu \, \theta + s(\theta)] \ ,
\end{equation}
hence $s(\theta)$ can be obtained by an inverse Legendre transform of $\phi(\mu)$, with $s(\theta) = \phi(\mu) - \mu \, \theta$ and $\theta = \phi'(\mu)$.
For completeness let us also make a similar statement when $\epsilon$ is finite, i.e. when one does not impose strictly the constraint of complete activation at time $T$. Denoting $s(\theta,\theta')$ the entropy density of initial configurations that have a fraction $\theta$ of initially active vertices and that lead after $T$ steps of evolution to a configuration with a fraction $\theta'$ of active sites, one has
\beq
\phi(G,\{l_i\},T,\mu,\epsilon) = \frac{1}{N} \ln Z(G,\{l_i\},T,\mu,\epsilon)= \underset{\theta,\theta'}{\sup}\,[ \mu \, \theta - \epsilon \, (1-\theta') + s(\theta,\theta') ] \ .
\label{eq_Legendre_both}
\end{equation}
Varying the parameters $\mu$ and $\epsilon$ thus allows to reconstruct the function $s(\theta,\theta')$, and hence to solve the optimization problem denoted (i) in the introduction, namely for a fixed value of $\theta$ find the maximal reachable $\theta'$. We will mainly concentrate in the following of the paper on the optimization problem denoted (ii) in the introduction, that is imposing the full activation of the graph at time $T$ ($\theta'=1$), which as explained above can be studied via the computation of $s(\theta)=s(\theta,\theta'=1)$ from the inverse Legendre transform of the free-entropy with $\epsilon=+\infty$.
The definitions above were valid for any graph and any choice of the activation thresholds; we shall however be particularly interested in the case of large random regular graphs with uniform thresholds, we thus define
\beq
{\theta_{\rm min}}(k,l) = \lim_{N \to \infty} \mathbb{E}[{\theta_{\rm min}}(G,\{l_i=l\}) ] \ , \qquad
{\theta_{\rm min}}(k,l,T) = \lim_{N \to \infty} \mathbb{E}[{\theta_{\rm min}}(G,\{l_i=l\},T) ] \ ,
\end{equation}
where the average is over uniformly chosen regular graphs of degree $k+1$ on $N$ vertices, with the same threshold for activation $l$ on every vertex. The fact that the limit in the definition of ${\theta_{\rm min}}(k,l,T)$ exists could actually be proven rigorously using the method developed in~\cite{Gamarnik_interpolation}, and it is expected that ${\theta_{\rm min}}(G,\{l_i=l\},T)$ is self-averaging (i.e. concentrates around its average in the large $N$ limit). The existence of ${\theta_{\rm min}}(k,l)$ might be a more difficult mathematical problem that we shall not discuss further; it is a reasonable conjecture that it coincides with the limit of ${\theta_{\rm min}}(k,l,T)$ when $T\to\infty$, i.e. that the large size and large time limits commute. We will see in Sec.~\ref{sec_kequall_Tinfty} one argument in favour of this conjecture. Let us emphasize that ${\theta_{\rm min}}(k,l) < {\theta_{\rm r}}(k,l)$, with a strict inequality. This is indeed a large-deviation phenomenon: even if most initial configurations with density smaller than ${\theta_{\rm r}}$ do not activate the whole graph some very rare ones (with a probability exponentially small in $N$ in the Bernouilli measure of parameter $\theta<{\theta_{\rm r}}$) are able to do so. Note also that ${\theta_{\rm min}}(k,l)$ is growing with $l$ at fixed $k$, for the same reasons as explained above in the discussion of ${\theta_{\rm r}}$. The computations of ${\theta_{\rm min}}$ we shall present will follow the strategy explained above on an arbitrary graph, namely the computation of a free-entropy density, that we define in the case of random regular graphs as the quenched average over the graph ensemble,
\beq
\phi(k,l,T,\mu,\epsilon) = \lim_{N \to \infty} \frac{1}{N} \mathbb{E}[\ln Z(G,\{l_i=l\},T,\mu,\epsilon)] \ .
\label{eq_def_phi}
\end{equation}
\subsection{Equivalence with other problems and bounds}
\label{sec_connections}
As mentioned in the introduction the problem of minimal contagious sets can be related, for appropriate choices of the threshold parameters $l_i$, to other standard problems in graph theory.
Consider first the case of an arbitrary graph where the thresholds $l_i$ are equal to the degrees $d_i$ for all vertices. An inactive site in the initial configuration will be activated only if it is surrounded by active vertices, and it will do so in a single step. In other words in any percolating initial condition, whatever the time horizon $T$, the inactive vertices must form an independent set (no two inactive vertices are allowed to be neighbours). For regular random graphs one has thus ${\theta_{\rm min}}(k,k+1,T)={\theta_{\rm min}}(k,k+1)$ for all $T$, and this quantity is equal to 1 minus the density of the largest independent sets of a $k+1$-regular random graph.
Another correspondance with previously studied models arises when $T=1$, for any choice of the thresholds $l_i$. Indeed in this case the vertex $i$ can be inactive in a percolating initial configuration only if its number of inactive neighbours is smaller than some value (namely, $\le d_i-l_i$). These generalized hard-core constraints (repulsion between inactive vertices) correspond exactly to the so-called Biroli-M\'ezard (BM) model~\cite{bm_prl,bm} (with the correspondance inactive vertex $\leftrightarrow$ vertex occupied by a particle in the BM model, and $d_i-l_i \leftrightarrow \ell_i$ of the BM model). Hence for $T=1$ the minimal density ${\theta_{\rm min}}$ is 1 minus the density of a close packing of the corresponding BM model. Further specializing this $T=1$ case by setting $l_i=1$ on each vertex leads to the constraint that every inactive site in a percolating initial configuration has to be adjacent with at least one active site, in other words that the active sites form a dominating set of $G$. The minimal density ${\theta_{\rm min}}$ is thus the domination number (divided by $N$) of $G$.
Consider now the thresholds of activation to be 1 less than the degrees, i.e. $l_i=d_i-1$ on all vertices, with no constraint on the time of activation ($T=\infty$). As explained at the end of Sec.~\ref{sec_def_dyn}, the inactive vertices in the final configuration form the 2-core of the inactive ones in the initial configuration. A percolating initial configuration must be such that this 2-core is empty, in other words the subgraph induced by the inactive sites of the initial configuration must be acyclic (a tree or a forest), i.e. the active sites have to form a decycling set~\cite{decycling_Beineke} (also known as a Feedback Vertex Set), and $N {\theta_{\rm min}}$ is the decycling number of $G$. This characterization leads to the following bound for every $k+1$-regular graph with thresholds $k$ of activation on every site,
\beq
{\theta_{\rm min}}(k,k) \ge \frac{k-1}{2 k} \ .
\label{eq_bound_kk}
\end{equation}
Indeed if $A$ denotes the number of active vertices in a percolating initial configuration, the $N-A$ other vertices induces a forest, the number of edges between inactive vertices is thus at most $N-A-1$. On the other hand this number is at least $\frac{k+1}{2} N - (k+1)A$ (the first term being the total number of edges, and the number of edges incident to at least one active site being at most $(k+1)A$). The decycling number of random regular graphs was studied in~\cite{decycling}, proving in particular that the bound (\ref{eq_bound_kk}) is actually tight for 3-regular large random graphs, i.e. ${\theta_{\rm min}}(2,2)=1/4$, and it was conjectured to be also the case for 4-regular ones (i.e. ${\theta_{\rm min}}(3,3)=1/3$). An asymptotic lowerbound on ${\theta_{\rm min}}(k,k)$ for large values of $k$ was worked out in~\cite{lb_klequal_rig} , we will come back on this result in Sec.~\ref{sec_kequall_Tinfty}. Note also that the decycling number of arbitrary sparse random graphs was studied with physics methods in~\cite{fvs_Zhou1,fvs_Zhou2}.
For general thresholds smaller than the degrees minus one the active sites of a percolating initial configuration must form a ``de-coring'' set instead of a ``de-cycling'' set (i.e. their removal has to provoke the disappearance of a $q$-core with $q>2$). A generalization of the lower bound (\ref{eq_bound_kk}) to any $k+1$-regular graph with uniform threshold $l$ was given in~\cite{Dreyer09}, and reads
\beq
{\theta_{\rm min}}(k,l) \ge \frac{2l-k-1}{2 l} \ .
\label{eq_bound_kl}
\end{equation}
Its proof goes as follows. Consider the sequential process explained at the end of Sec.~\ref{sec_def_dyn} in which at each time $t$ a single vertex gets activated, and denote $E(t)$ the number of edges between active and inactive vertices after $t$ steps of this process. By definition of the activation rule $E(t+1)-E(t) \le k+1-2l$. If as above $A$ denotes the number of active sites in a percolating initial configuration, by definition $E(N-A)=0$, hence $E(0)\ge (N-A) (2l-k-1)$. On the other hand $E(0) \le (k+1) A$, which gives the lower bound (\ref{eq_bound_kl}) on the possible values of $A$.
We should also mention an upper bound on the minimal sizes of contagious sets derived in~\cite{Ackerman10,bounds12} for graphs of arbitrary degree distributions, which yields in the case of $k+1$-regular graphs:
\beq
{\theta_{\rm min}}(k,l) \le \frac{l}{k+2} \ .
\end{equation}
To conclude this discussion let us mention that the ``de-coring'' perspective on the minimal contagious set problem is somehow reminiscent (even if not directly equivalent), to the Achlioptas processes~\cite{Achli_proc09,riordan2012} (more precisely of their offline version~\cite{Achli_offline}) where one looks for an extremal event avoiding the appearance of an otherwise typical structure (a giant component in the Achlioptas processes, a core in the minimal contagious set case).
\subsection{Main analytical results}
Let us draw here a more detailed plan of the rest of the paper to make its reading easier and faster for someone not interested in the technical details of the statistical mechanics method (who can browse quickly over the next section and jump to the results announced in Sec.~\ref{sec_results_rrg}). In order to compute the minimal density ${\theta_{\rm min}}$ of contagious sets we shall rephrase this problem as a statistical mechanics model and apply to it the cavity method. The latter is based on self-consistent assumptions of various degrees of sophistication, parametrized by the so-called level of replica symmetry breaking. We will study the first two levels of this hierarchy, named replica symmetric (RS) and one step of replica symmetry breaking (1RSB). These two approaches will lead to two predictions for ${\theta_{\rm min}}$, to be denoted respectively ${\theta_{\rm min,0}}(k,l,T)$ and ${\theta_{\rm min,1}}(k,l,T)$. From general bounds established in the context of disordered statistical mechanics models (first for the Sherrington-Kirkpatrick model~\cite{Gu03,Ta06,book_Panchenko} and later for some models on sparse random graphs~\cite{FrLe,FranzLeone03b,PaTa}) it is expected that the different levels of the cavity method provide improving lower bounds on ${\theta_{\rm min}}$, namely
\beq
{\theta_{\rm min,0}}(k,l,T) \le {\theta_{\rm min,1}}(k,l,T) \le {\theta_{\rm min}}(k,l,T) \ .
\end{equation}
Our computation of ${\theta_{\rm min,0}}(k,l,T)$ and ${\theta_{\rm min,1}}(k,l,T)$ relies on the resolution of a set of roughly $2T$ algebraic equations on $2T$ unknowns, explicit numbers will be given in Sec.~\ref{sec_results_rrg}. We managed to perform analytically the $T\to\infty$ limit and reduce the determination of ${\theta_{\rm min,0}}(k,l)$ and ${\theta_{\rm min,1}}(k,l)$ (their limit when $T\to\infty$) to a finite number of equations, that will also be presented along with numerical results in Sec.~\ref{sec_results_rrg}. We found four particular cases in which the predictions of the first two levels of replica symmetry breaking coincide when $T\to\infty$, which led us to conjecture that they are the exact ones, namely:
\beq
{\theta_{\rm min}}(2,2)=\frac{1}{4} \ , \qquad
{\theta_{\rm min}}(3,3)=\frac{1}{3} \ , \qquad
{\theta_{\rm min}}(4,3)=\frac{1}{6} \ , \qquad
{\theta_{\rm min}}(5,4)=\frac{1}{4} \ ,
\end{equation}
all these cases saturating the lower bounds of (\ref{eq_bound_kk},\ref{eq_bound_kl}). The first (resp. second) equality was actually proven (resp. conjectured) in~\cite{decycling}. We have also performed a large degree expansion of the decycling number of random regular graphs, yielding the conjecture
\beq
{\theta_{\rm min}}(k,k) = 1- \frac{2 \ln k}{k} - \frac{2}{k} + O\left(\frac{1}{k \ln k} \right) \ .
\end{equation}
\section{Cavity method treatment of the problem}
\label{sec_cavity}
\subsection{Factor graph representation}
As explained in Sec.~\ref{sec_def_optimization} the central quantity to compute is the free-entropy density defined from the partition function normalizing the probability law (\ref{eq_eta_us}), that for completeness we shall generalize to possibly site dependent chemical potentials $\mu_i$ and costs for non-activation $\epsilon_i$:
\beq
\eta({\underline{\sigma}}) = \frac{1}{Z(G,\{l_i\},T,\{\mu_i,\epsilon_i\})}
e^{\, \underset{i=1}{\overset{N}{\sum}}[\mu_i {\sigma}_i - \epsilon_i (1-{\sigma}_i^T)]}
\ .
\label{eq_eta_us_generalized}
\end{equation}
This expression is not very convenient to handle directly because the variables ${\sigma}_i$ have complicated interactions under this law: ${\sigma}_i^T$ is indeed a function of all variables ${\sigma}_j$ on the vertices $j$ at distance smaller than $T$ from $i$. A way to circumvent this difficulty and to turn the interactions of the model into local ones has been proposed in~\cite{Torino1,Torino2}, and we shall follow the same approach here.
Let us first define $t_i({\underline{\sigma}})$ as the time of activation of site $i$ in the dynamical process generated by the initial configuration ${\underline{\sigma}}$, i.e. $t_i({\underline{\sigma}})=\min \{ t : {\sigma}_i^t=1 \}$, with conventionally $t_i({\underline{\sigma}})=\infty$ if this time is strictly greater than the time horizon $T$. These variables obey the following equations:
\beq
t_i({\underline{\sigma}}) = f({\sigma}_i,\{t_j({\underline{\sigma}})\}_{j \in {\partial i}};l_i) \ \ \ \forall \, i \in V \ ,
\end{equation}
where the function $f$ is defined as
\beq
f({\sigma},t_1,\dots,t_n;l) = \begin{cases} 0 & \text{if} \ {\sigma}=1 \\
1+{\underset{l}{\min}}(t_1,\dots,t_n) & \text{if} \ {\sigma}=0 \ \text{and} \ 1+{\underset{l}{\min}}(t_1,\dots,t_n) \le T \\
\infty & \text{otherwise}
\end{cases} \ .
\end{equation}
Here ${\underset{l}{\min}}(t_1,\dots,t_n)$ is the $l$-th smallest $t_i$, i.e ordering the arguments as $t_1 \le t_2 \le \dots \le t_n$ one has ${\underset{l}{\min}}(t_1,\dots,t_n)=t_l$. This translates the dynamic rules (\ref{eq_dynrules}) in terms of the activation times, a site $i$ activating at the time following the first time where at least $l_i$ of its neighbours are active. In the following $f(0,t_1,\dots,t_n;l)$ will be abbreviated in $f(t_1,\dots,t_n;l)$. Reciprocally one can show that if a set of $\{t_i\}_{i\in V}$ verifies the condition that for all $i$ either $t_i=0$ or $t_i=f(\{t_j\}_{j \in {\partial i}};l_i)$, then they correspond to the activation times for the dynamics started from the initial condition ${\underline{\sigma}}$ such that ${\sigma}_i=1$ if and only if $t_i=0$. These two descriptions in terms of $({\sigma}_1,\dots,{\sigma}_N)$ and $(t_1,\dots,t_N)$ are thus equivalent, yet the great advantage of the latter is that the conditions to enforce among the $t_i$ are local along the graph, and that they contain in an obvious way the information on ${\sigma}_i^T$ that was lacking to deal with (\ref{eq_eta_us_generalized}).
Finally a last twist on Eq.~(\ref{eq_eta_us_generalized}) will be to ``duplicate'' the activation time $t_i$ on all edges connecting $i$ to one of its neighbour $j$, introducing redundant variables $t_{ij}$ to be finally constrained to be all equal to $t_i$. Let us denote ${\underline{t}}$ the collective configurations of all these $2M$ variables $t_{ij},t_{ji}$ on each edge ${\langle} i,j{\rangle}$ of the graph, that take values in $\{0,1,\dots,T,\infty\}$, and consider the following probability measure on $({\underline{\sigma}},{\underline{t}})$:
\beq
\eta({\underline{\sigma}},{\underline{t}}) = \frac{1}{Z(G,\{l_i\},T,\{\mu_i,\epsilon_i\})} \prod_{i=1}^N w_i({\sigma}_i,\{t_{ij},t_{ji}\}_{j \in {\partial i}}) \ ,
\label{eq_eta_ust}
\end{equation}
where the functions $w_i$ are defined by
\beq
w_i({\sigma}_i,\{t_{ij},t_{ji}\}_{j \in {\partial i}}) =
e^{\mu_i {\sigma}_i} e^{-\epsilon_i {\mathbb{I}}(f({\sigma}_i,\{t_{ki}\}_{k \in {\partial i}};l_i)=\infty)}
\prod_{j\in{\partial i}} {\mathbb{I}}(t_{ij}=f({\sigma}_i,\{t_{ki}\}_{k \in {\partial i}};l_i)) \ .
\end{equation}
The above observations imply that the marginal of ${\underline{\sigma}}$ under $\eta({\underline{\sigma}},{\underline{t}})$ is precisely the desired one from Eq.~(\ref{eq_eta_us_generalized}), and that in the support of the law the ${\underline{t}}$ are strictly constrained to be the activation times for the dynamics starting from ${\underline{\sigma}}$. This correspondance being one-to-one the partition function is the same in (\ref{eq_eta_us_generalized}) and (\ref{eq_eta_ust}). A portion of the factor graph~\cite{factorgraph} associated to the probability law (\ref{eq_eta_ust}) is sketched in Fig.~\ref{fig_factor}, with black squares representing the function nodes (interactions) $w_i$, black circles the variables ${\sigma}_i$, and white circles the variables $t_{ij},t_{ji}$. One notes that if the original graph $G$ is a tree (resp. is locally a tree) then the corresponding factor graph is a tree (resp. is locally a tree). This fact was the motivation for the ``duplication'' of the $t_i$ variables on the surrounding edge, without it short loops of interactions would still be present in the factor graph.
\begin{figure}
\includegraphics[width=4.5cm]{fig_factor_psf.eps}
\caption{A portion of the factor graph corresponding to the measure (\ref{eq_eta_ust}).}
\label{fig_factor}
\end{figure}
\subsection{Replica Symmetric (RS) formalism}
Let us now explain how the probability law (\ref{eq_eta_ust}) and its associated normalization $Z$ can be handled within the cavity formalism, first at the simplest, so called Replica Symmetric (RS), level.
\subsubsection{Single sample equations}
If the graph $G$ were a finite tree the factor graph associated to (\ref{eq_eta_ust}) would be a tree, hence $Z$ and the marginals of $\eta$ could be computed exactly via the recursive equations that we are about to write down. If the graph is only locally tree-like these equations are only approximate, they correspond to the (loopy) Belief Propagation equations, valid under some assumptions of long-range correlation decay under the measure $\eta$. This recursive computation of $Z$ amounts to introduce on each directed edge $i\to j$ of the graph a ``message'' $\eta_{i \to j}(t_{ij},t_{ji})$, which is a normalized probability distribution over a pair of activation times. These messages obey recursion relations of the form $\eta_{i \to j} = {\widehat{g}}(\{ \eta_{k \to i} \}_{k \in {\partial i \setminus j}};l_i,\epsilon_i,\mu_i)$, where the mapping $\eta={\widehat{g}}(\eta_1,\dots,\eta_k;l,\epsilon,\mu)$ is given by
\begin{multline}
\eta(t,t') =
\frac{1}{{\widehat{z}_{\rm iter}}(\eta_1,\dots,\eta_k;l,\epsilon,\mu)}
\left[ \delta_{t,0} e^\mu
\prod_{i=1}^k\left( \sum_{t''} \eta_i(t'',0)\right) \right. \\ \left.
+ e^{- \epsilon \delta_{t,\infty}} \sum_{t_1,\dots,t_k} \eta_1(t_1,t) \dots \eta_k(t_k,t) {\mathbb{I}}(t=f(t_1,\dots,t_k,t';l)) \right] \ .
\label{eq_recurs_eta}
\end{multline}
Here and in the following unprecised summations over a time index go along $\{0,\dots,T,\infty\}$. The function ${\widehat{z}_{\rm iter}}(\eta_1,\dots,\eta_k;l,\epsilon,\mu)$ is defined by normalization, in such a way that $\sum_{t,t'}\eta(t,t')=1$.
The knowledge of the messages $\eta_{i \to j}$ on all edges of the graph allows to compute the free-entropy density, according to the Bethe formula:
\beq
\phi = \frac{1}{N}\ln Z = \frac{1}{N} \sum_{i=1}^N \ln {\widehat{z}_{\rm site}}(\{\eta_{j \to i} \}_{j \in {\partial i}};l_i,\epsilon_i,\mu_i)
-\frac{1}{N} \sum_{{\langle} i,j {\rangle} \in E} \ln {\widehat{z}_{\rm edge}} (\eta_{i \to j},\eta_{j \to i}) \ ,
\label{eq_lnZ_eta}
\end{equation}
where the second sum runs over the (undirected) edges of the graph, and the local partition functions are
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{\widehat{z}_{\rm site}}(\eta_1,\dots,\eta_{k+1};l,\epsilon,\mu)
&=& e^\mu \prod_{i=1}^{k+1} \left(\sum_{t'} \eta_i(t',0)\right) +
\sum_{t=1}^T \sum_{t_1,\dots,t_{k+1}} \eta_1(t_1,t) \dots \eta_{k+1}(t_{k+1},t)
{\mathbb{I}}(t=1+{\underset{l}{\min}}(t_1,\dots,t_{k+1})) \nonumber \\
&&+ e^{-\epsilon} \sum_{t_1,\dots,t_{k+1}} \eta_1(t_1,\infty) \dots \eta_{k+1}(t_{k+1},\infty)
{\mathbb{I}}({\underset{l}{\min}}(t_1,\dots,t_{k+1}) \ge T)
\label{eq_hzsite} \\
{\widehat{z}_{\rm edge}} (\eta_1,\eta_2) &=& \sum_{t,t'} \eta_1(t,t') \eta_2(t',t) \ .
\label{eq_hzedge}
\eea
The marginals of the law (\ref{eq_eta_ust}) can also be deduced from the messages, for instance the probability distribution of the activation time $t_i$ for the vertex $i$ reads $\eta(t_i)={\widehat{\eta}_{\rm site}}(\{\eta_{j \to i} \}_{j \in {\partial i}};l_i,\epsilon_i,\mu_i)(t_i)$, where
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{\widehat{\eta}_{\rm site}} (\eta_1,\dots,\eta_{k+1};l,\epsilon,\mu) (t) &=&
\frac{1}{{\widehat{z}_{\rm site}}(\eta_1,\dots,\eta_{k+1};l,\epsilon,\mu)}
\left\{
\delta_{t,0} e^{\mu} \prod_{i=1}^{k+1} \left(\sum_{t'} \eta_i(t',0) \right)
\right. \nonumber \\ &+& \left.
(1-\delta_{t,0}-\delta_{t,\infty}) \sum_{t_1,\dots,t_{k+1}} \eta_1(t_1,t) \dots \eta_{k+1}(t_{k+1},t)
{\mathbb{I}}(t=1+{\underset{l}{\min}}(t_1,\dots,t_{k+1})) \right. \nonumber \\
&+ &\left. \delta_{t,\infty} e^{-\epsilon} \sum_{t_1,\dots,t_{k+1}} \eta_1(t_1,\infty) \dots \eta_{k+1}(t_{k+1},\infty)
{\mathbb{I}}({\underset{l}{\min}}(t_1,\dots,t_{k+1}) \ge T) \right\} \ .
\label{eq_hetasite}
\eea
The probability that the vertex $i$ is active in the initial condition is then deduced as $\eta({\sigma}_i=1)=\eta(t_i=0)$. As explained above in Eq.~(\ref{eq_Legendre_both}), one can deduce from the above results the entropy density $s(\theta,\theta')$ for initial configurations with a fraction $\theta$ of active sites leading to a fraction $\theta'$ of active sites after $T$ steps, taking $\mu_i=\mu$ and $\epsilon_i = \epsilon$ for all sites, with
\beq
s(\theta,\theta') = \phi(\mu,\epsilon) - \mu \theta + \epsilon (1-\theta') \ ,
\qquad \theta = \frac{1}{N} \sum_{i=1}^N \eta(t_i=0) \ , \qquad
\theta' = \frac{1}{N} \sum_{i=1}^N \eta(t_i \le T) \ .
\end{equation}
Note that the derivatives of $\phi$ with respect to $\mu$ and $\epsilon$ can be taken only on the explicit dependence in (\ref{eq_lnZ_eta}), the recursion equations on the messages $\eta_{i \to j}$ being precisely the stationarity condition of $\phi$ with respect to the $\eta$'s.
\subsubsection{A more compact parametrization of the messages}
Each probability distribution $\eta(t,t')$ is a priori described by $(T+2)^2-1$ independent real numbers (the times run over $T+2$ values, including $\infty$, and there is a global normalization constraint). We shall see however that a much more compact parametrization is possible, which will be very useful for the further analytical treatment of the model. From now on we shall assume that $\mu_i=\mu$ and $\epsilon_i =\epsilon$ for all vertices. To unveil these simplifications let us first rewrite Eq.~(\ref{eq_recurs_eta}) more explicitly:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
\eta(0,t') &=& \frac{1}{{\widehat{z}_{\rm iter}}} e^\mu \prod_{i=1}^k (\eta_i(0,0)+\eta_i(1,0)+\dots+\eta_i(T,0)+\eta_i(\infty,0)) \\
\eta(t,t') &=& \frac{1}{{\widehat{z}_{\rm iter}}} \sum_{t_1,\dots,t_k} \eta_1(t_1,t) \dots \eta_k(t_k,t) \ {\mathbb{I}}(t=1+{\underset{l}{\min}}(t_1,\dots,t_k,t')) \qquad \text{for} \ t\in\{1,\dots,T\} \\
\eta(\infty,t') &=& \frac{1}{{\widehat{z}_{\rm iter}}} e^{- \epsilon} \sum_{t_1,\dots,t_k} \eta_1(t_1,\infty) \dots \eta_k(t_k,\infty) \ {\mathbb{I}}({\underset{l}{\min}}(t_1,\dots,t_k,t') \ge T)
\eea
where in all the three cases $t'$ can take any value in $\{0,1,\dots,T,\infty\}$. Now the condition ``${\underset{l}{\min}}(t_1,\dots,t_k,t')=t-1$'' is easily seen to be equivalent to ``at least $l$ of the time arguments are $\le t-1$ and at most $l-1$ of them are $\le t-2$''. Similarly the condition ``${\underset{l}{\min}}(t_1,\dots,t_k,t') \ge T$'' is equivalent to ``at most $l-1$ times are $\le T-1$''. This observation allows to rewrite the above equations under the following form:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
\eta(0,t') &=& \frac{1}{{\widehat{z}_{\rm iter}}} e^\mu \prod_{i=1}^k (\eta_i(0,0)+\eta_i(1,0)+\dots+\eta_i(T,0)+\eta_i(\infty,0)) \label{eq_eta_0tp} \\
\eta(t,t') &=& \frac{1}{{\widehat{z}_{\rm iter}}} \sum_{\substack{ I,J,K \\ |I| + {\mathbb{I}}(t'\le t-2) \le l-1 \\ |I| + |J| + {\mathbb{I}}(t'\le t-1) \ge l } }
\prod_{i\in I} \left(\sum_{t''=0}^{t-2} \eta_i(t'',t) \right)
\prod_{i\in J} \eta_i(t-1,t)
\prod_{i \in K} \left( \sum_{t''\ge t} \eta_i(t'',t)\right) \label{eq_eta_ttp}
\\
\eta(\infty,t') &=& \frac{1}{{\widehat{z}_{\rm iter}}} e^{- \epsilon}
\sum_{\substack{ I,J \\ |I| + {\mathbb{I}}(t'\le T-1) \le l-1}}
\prod_{i\in I} \left(\sum_{t''=0}^{T-1} \eta_i(t'',\infty) \right)
\prod_{i \in J} \left(\sum_{t''\ge T} \eta_i(t'',\infty)\right)
\label{eq_eta_inftytp}
\eea
where the summation in the second (resp. third) line is over the partitions $I,J,K$ (resp. $I,J$) of $\{1,\dots,k\}$. These expressions reveal a first simplification, as already noticed in~\cite{Torino1,Torino2}: among the $(T+2)^2$ elements of $\eta(t,t')$ only $3T+2$ are distinct. Indeed $\eta(0,t')$ is independent of $t'$, for a given value of $t \in \{1,\dots,T\}$ $\eta(t,t')$ takes at most three distinct values, whether $t'\ge t$, $t'=t-1$, or $t' \le t-2$ and finally $\eta(\infty,t')$ takes two values whether $t'\le T-1$ or $t' \ge T$. There is however a further simplification to perform: in the right hand sides of the above equations the $\eta_i$'s always appear under the form of particular linear combinations. In particular the elements under the diagonal of the matrices $\eta_i$, i.e. $\eta_i(t,t')$ with $t\ge t'$, always intervene under the form $\sum_{t \ge t'}\eta(t,t')$. This allows to reduce further the number of relevant linear combinations of elements of the $\eta$'s. A convenient parametrization of the messages $\eta$ is thus provided by the numbers $a_t$ for $t \in \{0,1,\dots,T\}$ and $b_t$ for $t \in \{1,\dots,T\}$, defined
by:
\beq
e^{\mu a_t} = \frac{\eta(0,0)}{\sum_{t'} \eta(t',t)} \ , \qquad
e^{\mu b_t} = \frac{\eta(0,0)}{\sum_{t'=0}^t \eta(t',t)} = \frac{\eta(0,0)}{\sum_{t'=0}^t \eta(t',t'')} \ \forall t'' \ge t \ .
\label{eq_def_at}
\end{equation}
One can consistently extend these definitions with $b_0=0$, and it will be useful to adopt the convention $e^{-\mu b_{-1}}=0$ in order to simplify some expressions. Let us denote $h=(a_0,a_1,\dots,a_T,b_{T-1},\dots,b_1)$ the vector of $2T$ reals encoding in this way a matrix $\eta$; $h$ will be called a cavity field in the following (note that we excluded $b_T$ which disappears from the final expressions). The recursion relations (\ref{eq_eta_0tp}-\ref{eq_eta_inftytp}) should now be transformed into a relation between cavity fields, i.e. $h=g(h_1,\dots,h_k)$, with $h_i=(a_0^{(i)},a_1^{(i)},\dots,a_T^{(i)},b_{T-1}^{(i)},\dots,b_1^{(i)})$. Inserting the definitions (\ref{eq_def_at}) into the equations (\ref{eq_eta_0tp}-\ref{eq_eta_inftytp}) leads to the explicit form for $g$,
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
e^{- \mu a_t}
&=& 1 + e^{-\mu} \sum_{t'=1}^T \sum_{\substack{ I,J,K \\ |I| + {\mathbb{I}}( t' \ge t+2) \le l-1 \\ |I| + |J| + {\mathbb{I}}(t' \ge t+1) \ge l } }
{\cal P}_{t'}(h_1,\dots,h_k;I,J,K) +
e^{-\mu-\epsilon}
\sum_{\substack{ I,J,K \\ |I|+|J| + {\mathbb{I}}(t\le T-1) \le l-1}}
{\cal P}_T(h_1,\dots,h_k;I,J,K) \nonumber \\
e^{- \mu b_t} &=& 1 + e^{-\mu} \sum_{t'=1}^t
\sum_{\substack{I,J,K \\ |I| \le l-1 \\ |I|+|J| \ge l}}
{\cal P}_{t'}(h_1,\dots,h_k;I,J,K)
\label{eq_g}
\eea
where we defined
\beq
{\cal P}_t(h_1,\dots,h_k;I,J,K) =
e^{\mu \underset{i=1}{\overset{k}{\sum}} a_0^{(i)}}
\prod_{i \in I} e^{-\mu b_{t-2}^{(i)}}
\prod_{i \in J} ( e^{-\mu b_{t-1}^{(i)}} - e^{-\mu b_{t-2}^{(i)}} )
\prod_{i \in K} ( e^{-\mu a_t^{(i)}} - e^{-\mu b_{t-1}^{(i)}} ) \ .
\end{equation}
It can be checked that for $T=1$ and $\epsilon=+\infty$ these equations correspond, as they should, to the one of the Biroli-M\'ezard model (see Eqs.~(108,109) of~\cite{bm}). One can also express the partial partition functions ${\widehat{z}_{\rm site}}$ and ${\widehat{z}_{\rm edge}}$ in terms of these fields. It will be more convenient to factor out a common part in the site and edge contributions to the free-entropy. Denoting $r(\eta)=\sum_t \eta(t,0)$, we define ${z_{\rm edge}}$ as:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{z_{\rm edge}}(h_1,h_2) &=& \frac{{\widehat{z}_{\rm edge}} (\eta_1,\eta_2)}
{r(\eta_1)r(\eta_2)} \label{eq_zedge} \\
&=& e^{\mu (a_0^{(1)}+a_0^{(2)})} \left\{
e^{-\mu (a_T^{(1)}+a_T^{(2)}) } +
\sum_{t=0}^{T-1} \left[
\left( e^{-\mu a_t^{(1)}}-e^{-\mu a_{t+1}^{(1)}}\right)
e^{-\mu b_t^{(2)}}
+ e^{-\mu b_t^{(1)}}
\left( e^{-\mu a_t^{(2)}}-e^{-\mu a_{t+1}^{(2)}}\right)
\right]
\right\}\ ,
\nonumber
\eea
where the explicit expression is obtained from Eq.~(\ref{eq_hzedge}).
Similarly, exploiting Eq.~(\ref{eq_hzsite}), we get for the site term (factoring also a contribution from the chemical potential):
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{z_{\rm site}}(h_1,\dots,h_{k+1};l,\epsilon;\mu) &=&
\frac{e^{-\mu}{\widehat{z}_{\rm site}}(\eta_1,\dots,\eta_{k+1};l,\epsilon;\mu)}
{r(\eta_1) \dots r(\eta_{k+1}) } \label{eq_zsite} \\ &=&
1 + e^{-\mu}\sum_{t=1}^T
\sum_{\substack{I,J,K \\ |I| \le l-1 \\ |I|+|J| \ge l}}
{\cal P}_t(h_1,\dots,h_{k+1};I,J,K)
+ e^{-\mu- \epsilon} \sum_{\substack{I,J,K \\ |I|+|J|\le l-1}} {\cal P}_T(h_1,\dots,h_{k+1};I,J,K)
\nonumber
\eea
where as above in the summations $I,J,K$ denotes a partition of $\{1,\dots,k+1\}$.
To summarize the results of this reparametrization, on a given graph one has cavity fields $h_{i\to j}$ on each directed edge, obeying the Belief Propagation equations $h_{i\to j}=g(\{h_{k \to i}\}_{k \in {\partial i \setminus j}})$, with the $g$ defined in Eq.~(\ref{eq_g}), and the Bethe free-entropy density is computed from these cavity fields according to
\beq
\phi = \mu+ \frac{1}{N} \sum_i \ln {z_{\rm site}}(\{ h_{j \to i} \}_{j \in {\partial i}};l_i,\epsilon,\mu)
- \frac{1}{N} \sum_{{\langle} i,j {\rangle} \in E} \ln {z_{\rm edge}}(h_{i \to j},h_{j \to i}) \ ,
\label{eq_phi_h_BP}
\end{equation}
with ${z_{\rm site}}$ and ${z_{\rm edge}}$ defined in Eqs.~(\ref{eq_zsite}) and (\ref{eq_zedge}) respectively. Note that the factors $r$ introduced in the definitions of ${z_{\rm site}}$ and ${z_{\rm edge}}$ compensate because in the expression of the Bethe free-energy of Eq.~(\ref{eq_lnZ_eta}) the messages on each directed edge appear exactly once in the site term and once in the edge term. The marginals of the law $\eta({\underline{\sigma}},{\underline{t}})$ can also be computed from the cavity fields $h$, in particular from the expression (\ref{eq_hetasite}) one obtains the marginal of one activation time from the incident cavity fields as
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{\eta_{\rm site}}(h_1,\dots,h_{k+1};l,\epsilon;\mu)(t) = \frac{1}{{z_{\rm site}}(h_1,\dots,h_{k+1};l,\epsilon;\mu)} &&\left\{
\delta_{t,0} \,
+ (1-\delta_{t,0}-\delta_{t,\infty}) e^{-\mu}
\sum_{\substack{I,J,K \\ |I| \le l-1 \\ |I|+|J| \ge l}}
{\cal P}_t(h_1,\dots,h_{k+1};I,J,K)
\right. \nonumber \\
&&+ \left. \delta_{t,\infty} \, e^{-\mu-\epsilon}
\sum_{\substack{I,J,K \\ |I|+|J|\le l-1}} {\cal P}_T(h_1,\dots,h_{k+1};I,J,K)
\right\} \ .
\label{eq_etasite}
\eea
\subsubsection{Random (regular) graphs}
\label{sec_cavity_RS_rg}
The replica symmetric cavity method, for generic models defined on sparse random graphs, postulates the asymptotic validity of the above computations, exact on finite trees, thanks to the local convergence of random graphs to trees and an assumption of correlation decay at large distance. The order parameter is then a probability distribution over cavity fields, the randomness arising from the fluctuations of the degrees of the vertices in the graph and/or the randomness in the local interactions.
In the case of random regular graphs with no disorder in the coupling the situation is even simpler, as one can look for a ``factorized'' solution with all cavity fields equal. In particular for the model at hand on a $k+1$ random regular graph, with the same threshold of activation $l$ for all vertices, the RS prediction for the typical free-entropy density in the thermodynamic limit defined in Eq.~(\ref{eq_def_phi}) reads
\beq
\phi(k,l,T,\mu,\epsilon) = \mu + \ln \left( {z_{\rm site}}(h,\dots,h) \right)
- \frac{k+1}{2} \ln\left( {z_{\rm edge}}(h,h) \right) \ ,
\end{equation}
which is easily obtained from (\ref{eq_phi_h_BP}) noting that $2M=(k+1)N$ in a regular graph. The field $h$ is the fixed-point solution of the cavity recursion (\ref{eq_g}),
\beq
h=g(h,\dots,h) \ .
\label{eq_RS}
\end{equation}
The marginal law for the activation time is obtained from (\ref{eq_etasite}) by setting all the fields to $h$, which allows finally to compute the entropy density from the Legendre inverse transform discussed in (\ref{eq_Legendre_both}).
We shall discuss the results obtained from this RS prediction in the next Section, more explicit formulas for the RS equation in this case, along with some technical details on their resolution being displayed in the Appendix~\ref{sec_app}. One can however anticipate that in some regime of parameters the RS hypothesis will be violated. This is for instance known for $T=1$, $\epsilon=+\infty$, which corresponds to the Biroli-M\'ezard model; it was indeed shown in~\cite{bm} that for large negative values of $\mu$ the predictions of the RS ansatz are unphysical, the frustration arising from the contradictory constraints of putting as few active vertices in the initial condition as possible while imposing that all vertices become active at a latter time induces long-range correlations between variables that are incompatible with the RS ansatz. This limit $\mu \to -\infty$ being the interesting case for the computations of the minimal density of contagious sets, we shall now see how to include the effects of replica symmetry breaking in this model.
\subsection{One step of Replica Symmetry Breaking (1RSB) formalism}
The long-range correlation decay assumption underlying the RS cavity method breaks down for models with too much frustration. In this case one has to picture the configuration space as fractured into pure states, or clusters, that we shall index here by $\gamma$, such that the correlation decay assumption only holds for the Gibbs-Boltzmann probability law restricted to one pure-state. The partition function restricted to a given pure-state is denoted $Z_\gamma$, in such a way that $Z=\sum_\gamma Z_\gamma$. The replica symmetry breaking version of the cavity method then postulates some properties of this decomposition into pure states, which are compatible with the local convergence of the graph under study to a tree. In the first non-trivial version of the RSB formalism, so called one-step RSB (1RSB), one assumes the existence of a complexity function, also called configurational entropy in the context of glasses, $\Sigma(\phi)$, such that the number of pure states with an internal free-entropy density $\phi_\gamma = \frac{1}{N} \ln Z_\gamma$ close to some value $\phi$ is, at the leading exponential order, $e^{N \Sigma(\phi)}$. The computation of $\Sigma(\phi)$ is performed via the 1RSB potential with a parameter $m$ (known as the Parisi breaking parameter), related to $\Sigma$ through a Legendre transform structure~\cite{Mo95}:
\beq
\Phi(m) = \frac{1}{N} \ln \sum_\gamma Z_\gamma^m =
\sup_\phi \, [ \Sigma(\phi) + m \, \phi ] \ .
\label{eq_def_Phi}
\end{equation}
The function $\Sigma(\phi)$ can be reconstructed in a parametric way varying $m$, with
\beq
\Sigma(\phi_{\rm int}(m)) = \Phi(m) - m \phi_{\rm int}(m) \ , \qquad
\phi_{\rm int}(m) = \Phi'(m) \ ,
\label{eq_invLegendre_1RSB}
\end{equation}
$\phi_{\rm int}(m)$ denoting the internal free-entropy density of the clusters selected by the corresponding value of $m$. The value $m=1$ plays a special role in this approach, as it corresponds a priori to the original computation of the free-entropy density of the model. However a so-called condensation (or Kauzmann) transition can occur, signaled by the vanishing of the complexity $\Sigma$ associated to $m=1$. In this case the Gibbs-Boltzmann measure is dominated by a sub-exponential number of pure-states, corresponding to a parameter $m_{\rm s}<1$ with $\Sigma(m_{\rm s})=0$. In the following paragraphs we shall derive the 1RSB equations and the expression of the 1RSB potential for the model under study, before discussing the concrete results for random regular graphs in the next Section.
\subsubsection{Single sample equations}
Let us first discuss the 1RSB formalism with the basic messages represented in terms of the matrices $\eta(t,t')$. In the RS description one had a message $\eta_{i \to j}$ on each directed edge of the graph, solution of the recurrence equations $\eta_{i \to j} = {\widehat{g}}(\{ \eta_{k \to i} \}_{k \in {\partial i \setminus j}};l_i,\epsilon,\mu)$, see Eq.~(\ref{eq_recurs_eta}). At the 1RSB level one introduces instead a distribution ${\widehat{P}}_{i \to j}(\eta)$ on each directed edge, the randomness being over the choice of the pure-state $\gamma$ with a weight proportional to $Z_\gamma^m$. These distributions are thus found to obey the recurrence equations ${\widehat{P}}_{i \to j}={\widehat{G}}[\{ {\widehat{P}}_{k \to i} \}_{k \in {\partial i \setminus j}}]$, where ${\widehat{P}}={\widehat{G}}({\widehat{P}}_1,\dots,{\widehat{P}}_k)$ means
\beq
{\widehat{P}}(\eta) = \frac{1}{{\widehat{\cal Z}_{\rm iter}}({\widehat{P}}_1,\dots,{\widehat{P}}_k)}
\int {\rm d} {\widehat{P}}_1(\eta_1) \dots {\rm d} {\widehat{P}}_k(\eta_k) \
\delta(\eta - {\widehat{g}}(\eta_1,\dots,\eta_k)) \ {\widehat{z}_{\rm iter}}(\eta_1,\dots,\eta_k)^m \ ,
\end{equation}
with ${\widehat{g}}$ and ${\widehat{z}_{\rm iter}}$ defined in Eq.~(\ref{eq_recurs_eta}), and ${\widehat{\cal Z}_{\rm iter}}$ normalizes the distribution ${\widehat{P}}$. The 1RSB potential $\Phi(m)$ defined above is then computed from the solution of these equations, according to
\beq
\Phi(m)=
\frac{1}{N} \sum_{i=1}^N \ln {\widehat{\cal Z}_{\rm site}}(\{{\widehat{P}}_{j \to i} \}_{j \in {\partial i}};l_i,\epsilon_i,\mu_i)
- \frac{1}{N} \sum_{{\langle} i,j {\rangle} \in E} \ln {\widehat{\cal Z}_{\rm edge}} ({\widehat{P}}_{i \to j},{\widehat{P}}_{j \to i}) \ ,
\label{eq_Phi_single}
\end{equation}
where
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{\widehat{\cal Z}_{\rm site}}({\widehat{P}}_1,\dots,{\widehat{P}}_{k+1}) &=& \int {\rm d} {\widehat{P}}_1(\eta_1) \dots {\widehat{P}}_{k+1}(\eta_{k+1}) \ {\widehat{z}_{\rm site}}(\eta_1,\dots,\eta_{k+1})^m \ , \\
{\widehat{\cal Z}_{\rm edge}}({\widehat{P}}_1,{\widehat{P}}_2) &=& \int {\rm d} {\widehat{P}}_1(\eta_1) {\widehat{P}}_2(\eta_2) \ {\widehat{z}_{\rm edge}}(\eta_1,\eta_2)^m
\eea
are weighted averages, over the pure-states distribution, of the site and edge contributions to the free-entropy defined in (\ref{eq_hzsite},\ref{eq_hzedge}). Similarly the marginal distribution of an activation time can be computed as a weighted average of the RS expression in the various pure-states, i.e.
\beq
\eta(t) = \frac{1}{{\widehat{\cal Z}_{\rm site}}({\widehat{P}}_1,\dots,{\widehat{P}}_{k+1})}
\int {\rm d} {\widehat{P}}_1(\eta_1) \dots {\widehat{P}}_{k+1}(\eta_{k+1}) \
{\widehat{\eta}_{\rm site}}(\eta_1,\dots,\eta_{k+1})(t) \
{\widehat{z}_{\rm site}}(\eta_1,\dots,\eta_{k+1})^m \ .
\end{equation}
Note that the derivative $\Phi'(m)$, which plays an important role to compute the complexity from Eq.~(\ref{eq_invLegendre_1RSB}), can be taken in (\ref{eq_Phi_single}) on the explicit dependence on $m$ only, the recursion relations on the ${\widehat{P}}_{i\to j}$ being the stationarity conditions of (\ref{eq_Phi_single}) with respect to the ${\widehat{P}}$'s.
As we have seen in the discussion of the RS cavity method the matrices $\eta$ can be parametrized in a more economic way by the fields $h$ (vectors of $2T$ real numbers). The expressions of the 1RSB quantities can also be rewritten using this parametrization. After a few lines of algebra one finds that the potential $\Phi(m)$ reads
\beq
\Phi(m) = \mu m + \sum_{i=1}^N \ln {{\cal Z}_{\rm site}}(\{P_{j \to i} \}_{j \in {\partial i}};l_i,\epsilon,\mu)
- \sum_{{\langle} i,j {\rangle} \in E} \ln {{\cal Z}_{\rm edge}} (P_{i \to j},P_{j \to i}) \ ,
\end{equation}
with
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{{\cal Z}_{\rm site}}(P_1,\dots,P_{k+1}) &=& \int {\rm d} P_1(h_1) \dots P_{k+1}(h_{k+1}) \ {z_{\rm site}}(h_1,\dots,h_{k+1})^m \ , \\
{{\cal Z}_{\rm edge}}(P_1,P_2) &=& \int {\rm d} P_1(h_1) P_2(h_2) \ {z_{\rm edge}}(h_1,h_2)^m \ ,
\eea
the weighted averages of the quantities defined in (\ref{eq_zedge},\ref{eq_zsite}). The field distributions $P_{i \to j}(h)$ are solutions of the recurrence equations $P_{i \to j}=G(\{P_{k \to i} \}_{k \in {\partial i \setminus j}})$, where the mapping $P=G(P_1,\dots,P_k)$ is given explicitly by
\beq
P(h) = \frac{1}{{{\cal Z}_{\rm iter}}(P_1,\dots,P_k)}
\int {\rm d} P_1(h_1) \dots {\rm d} P_k(h_k) \
\delta(h - g(h_1,\dots,h_k)) \ {z_{\rm iter}}(h_1,\dots,h_k)^m \ .
\label{eq_1RSB_Ph}
\end{equation}
${{\cal Z}_{\rm iter}}$ is a normalizing factor ensuring that the left hand side is a probability distribution, $g$ is the function defined in Eq.~(\ref{eq_g}), and the reweighting factor reads
\beq
{z_{\rm iter}}(h_1,\dots,h_k) = \frac{e^{-\mu} {\widehat{z}_{\rm iter}}(\eta_1,\dots,\eta_k)r({\widehat{g}}(\eta_1,\dots,\eta_k)) }{r(\eta_1) \dots r(\eta_k)} = e^{-\mu a_0(h_1,\dots,h_k)} \ ,
\label{eq_ziter}
\end{equation}
the last equality following from Eqs.~(\ref{eq_eta_0tp},\ref{eq_def_at}).
\subsubsection{Random regular graphs}
For the reasons explained in the context of the RS ansatz in Sec.~\ref{sec_cavity_RS_rg} one can look for a simple factorized solution of the 1RSB equations in the case of a $k+1$ regular random graph with all thresholds of activation equal to $l$. In this case one has to find a distribution $P(h)$ solution of
\beq
P(h) = \frac{1}{{{{\cal Z}_{\rm iter}}}} \int {\rm d} P(h_1) \dots {\rm d} P(h_k) \ \delta(h-g(h_1,\dots,h_k)) \ {z_{\rm iter}}(h_1,\dots,h_k)^m \ ,
\label{eq_1RSB}
\end{equation}
where $m \in [0,1]$ is the Parisi breaking parameter and the functions $g$ and ${z_{\rm iter}}$ are the ones defined in Eqs.~(\ref{eq_g},\ref{eq_ziter}). The 1RSB potential is then computed as
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
\Phi(m) = \mu m &+& \ln \left( \int {\rm d} P(h_1) \dots {\rm d} P(h_{k+1})\ {z_{\rm site}}(h_1,\dots,h_{k+1})^m \right) \nonumber \\
&-& \frac{k+1}{2} \ln\left(\int {\rm d} P(h_1){\rm d} P(h_2) \ {z_{\rm edge}}(h_1,h_2)^m \right) \ ,
\label{eq_phi1RSB_m}
\eea
with the functions ${z_{\rm site}},{z_{\rm edge}}$ defined in Eqs.~(\ref{eq_zedge},\ref{eq_zsite}). As already mentioned above $\Phi'(m)$ can be computed by taking into account only the explicit dependence on $m$ of (\ref{eq_phi1RSB_m}).
\subsection{``Energetic'' 1RSB formalism}
\label{sec_1RSB_y}
Even within the simplified case of the factorized ansatz for regular graphs the 1RSB equations are relatively complicated, as they involve the resolution of a distributional equation on $P(h)$. However we are ultimately interested in a particular limit for the computation of the minimal density of contagious sets, namely the case where $\epsilon=+\infty$ (to take into account only the fully activating configurations), and in the limit $\mu \to -\infty$ (to select the initial configurations with the minimal number of active sites). It turns out that a simplified version of the 1RSB formalism can be devised in this case, corresponding to the ``energetic'' version of the 1RSB cavity method, first developed in~\cite{cavity_T0,MeZe}, see in particular Sec. 5 of~\cite{bm} for such a treatment of the related Biroli-M\'ezard model. This simplified treatment amounts to take simultaneously the limit $m\to 0$ and $\mu \to - \infty$, with a fixed finite value of a new parameter $y=-\mu m$. To explain the meaning of this limit let us rewrite more explicitly the expression of the 1RSB potential of Eq.~(\ref{eq_def_Phi}) in the case $\epsilon=+\infty$, introducing the complexity $\Sigma(s,\theta)$ counting the (exponential) number of clusters containing a number of order $e^{N s}$ of activating initial configurations with a fraction $\theta$ of active sites, hence with a free-entropy density $\phi=\mu \theta +s$:
\beq
\Phi(m) = \sup_{\theta,s}[\Sigma(s,\theta)+m (\mu \theta +s) ] \ .
\end{equation}
In the limit $m\to 0$, $\mu \to - \infty$ with $y=-\mu m$ this function becomes
\beq
\Phi_{\rm e}(y) = \sup_\theta [\Sigma_{\rm e}(\theta) - y \theta ] \ , \qquad
\Sigma_{\rm e}(\theta) = \sup_s \Sigma(s,\theta) \ .
\end{equation}
The ``energetic'' complexity $\Sigma_{\rm e}(\theta)$ can thus be computed via an inverse Legendre transform of the potential $\Phi_{\rm e}(y)$,
\beq
\Sigma_{\rm e}(\theta(y)) = \Phi_{\rm e}(y) + y \theta(y) \ , \qquad \theta(y) = - \Phi'_{\rm e}(y) \ .
\label{eq_Legendre_y}
\end{equation}
As we shall see the ``energetic'' 1RSB cavity equations leading to the computations of $\Phi_{\rm e}(y)$ are much simpler than the initial 1RSB ones at finite values of $\mu$ and $m$. The price to pay for this simplification is the loss of information on the entropy of the clusters when going from $\Sigma(s,\theta)$ to $\Sigma_{\rm e}(\theta)$. However this is not a problem for the determination of ${\theta_{\rm min}}$: its estimate at the 1RSB level, to be denoted ${\theta_{\rm min,1}}$, is the smallest value of $\theta$ with $\Sigma_{\rm e}(\theta) \ge 0$. Indeed the least dense activating configurations have to be in some pure states, whatever their entropy.
\subsubsection{Simplification of the cavity field recursion (Warning Propagation equations)}
We want to simplify the equations (\ref{eq_g}) giving $h=g(h_1,\dots,h_k)$ with $\epsilon=+\infty$ and in the limit $\mu \to -\infty$. First let us make some remarks, valid when $\epsilon=+\infty$ for any value of $\mu$. From the definition (\ref{eq_def_at}) of the fields $b_t$, or from their expressions in (\ref{eq_g}), it is obvious that
\beq
e^{-\mu b_T} \ge e^{-\mu b_{T-1}} \ge \dots \ge e^{-\mu b_1} \ge e^{-\mu b_0} = 1 \ .
\end{equation}
One can also notice that for $\epsilon=+\infty$ one has, for any $\mu$, the equality $a_T=b_T$: this appears both from the definition (\ref{eq_def_at}) of the fields, as $\eta(\infty,t)=0$ when $\epsilon=+\infty$, and from the recursion relations (\ref{eq_g}), the last term in the first line of (\ref{eq_g}) disappearing when $\epsilon=+\infty$. To continue the above chain of inequalities let us first compute from (\ref{eq_g})
\beq
e^{-\mu a_{T-1}} - e^{-\mu a_T} = e^{-\mu+\mu \overset{k}{\underset{i=1}{\sum}} a_0^{(i)} } \sum_{\underset{|I|=l-1}{I,J}}
\prod_{i \in I} e^{- \mu b_{T-1}^{(i)}}
\prod_{i \in J} \left(e^{-\mu b_T^{(i)}} - e^{- \mu b_{T-1}^{(i)}}
\right) \ ,
\label{eq_diff_aT}
\end{equation}
where $I,J$ forms a partition of $\{1,\dots,k\}$. This shows that $e^{-\mu a_{T-1}} \ge e^{-\mu a_T} = e^{-\mu b_T}$, because in the right-hand side $e^{-\mu b_T^{(i)}} \ge e^{- \mu b_{T-1}^{(i)}}$. These inequalities can then be continued by recurrence, as for $t \in \{0,\dots,T-2\}$ one obtains from (\ref{eq_g})
\beq
e^{-\mu a_t} - e^{-\mu a_{t+1}} = e^{-\mu +\mu \overset{k}{\underset{i=1}{\sum}} a_0^{(i)}} \sum_{\underset{|I|=l-1}{I,J}}
\prod_{i \in I} e^{-\mu b_{t}^{(i)}}
\left( \prod_{i \in J} (e^{- \mu a_{t+1}^{(i)}} - e^{-\mu b_t^{(i)}})
-\prod_{i \in J} (e^{- \mu a_{t+2}^{(i)}} - e^{-\mu b_t^{(i)}})
\right)
\ ,
\label{eq_diff_at}
\end{equation}
hence
\beq
e^{-\mu a_0} \ge a^{-\mu a_1} \ge \dots \ge e^{-\mu a_{T-1}} \ge e^{-\mu a_T} = e^{-\mu b_T} \ge e^{-\mu b_{T-1}} \ge \dots \ge e^{-\mu b_1} \ge e^{-\mu b_0} = 1 \ ,
\end{equation}
and for any $\mu \le 0$:
\beq
a_0 \ge a_1 \ge \dots \ge a_{T-1} \ge a_T = b_T \ge b_{T-1} \ge \dots \ge b_1
\ge b_0 = 0 \ .
\label{eq_ineq_ab}
\end{equation}
Let us now take the limit $\mu \to -\infty$ in the equations (\ref{eq_g}), assuming that $a_t$ and $b_t$ have finite limits. Treating (\ref{eq_g}) at the leading exponential order one obtains
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
a_t &=& \max\left(0, \max_{t'\in[1,T]} \max_{\substack{ I,J,K \\ |I| + {\mathbb{I}}( t' \ge t+2) \le l-1 \\ |I| + |J| + {\mathbb{I}}(t' \ge t+1) \ge l } } {\cal S}_{t'}(h_1,\dots,h_k;I,J,K) \right) \ , \label{eq_at_WP}\\
b_t &=& \max\left(0, \max_{t'\in[1,t]}
\max_{\substack{I,J,K \\ |I| \le l-1 \\ |I|+|J| \ge l}}
{\cal S}_{t'}(h_1,\dots,h_k;I,J,K) \right) \ , \label{eq_bt_WP}
\eea
where
\beq
{\cal S}_t(h_1,\dots,h_k;I,J,K) = 1 - \sum_{i \in I} (a_0^{(i)} - b_{t-2}^{(i)})
- \sum_{i \in J}(a_0^{(i)} - b_{t-1}^{(i)})
-\sum_{i \in K}(a_0^{(i)} - a_t^{(i)}) \ .
\end{equation}
Now from the inequalities (\ref{eq_ineq_ab}) it appears that ${\cal S}_t \le 1$, hence that the $a$'s and $b$'s belong to the interval $[0,1]$. It is however natural to assume that they are integers, as in the limit $\mu \to -\infty$ they can be interpreted as differences between number of particles in constrained groundstate configurations (see~\cite{bm,cavity_T0} for more details). Within this ansatz the $a$'s and $b$'s can only be equal to $0$ or $1$; using in addition the inequalities (\ref{eq_ineq_ab}) one realizes that the fields $h$ can only take $2T+1$ possible values, that we shall call $A_t$ for $t \in \{0,1,\dots,T-1\}$ and $B_t$ for $t\in\{0,1,\dots,T\}$. These are defined as follows; $A_t$ denotes the case where $a_0=\dots=a_t=1$, all the other $a$'s and $b$'s vanishing. For $t\in\{2,\dots,T\}$, $B_t$ means that $b_1=\dots=b_{t-1}=0$, all the other $a$'s and $b$'s being equal to 1. Finally $B_1$ corresponds to the case where all $a$'s and $b$'s are equal to 1, and $B_0$ to the case where they all vanish. Note that one can consistently extend these definitions to $A_T=B_T$, as by definition $a_T=b_T$.
It remains to determine the value of $h=g(h_1,\dots,h_k)$ in this $\mu \to -\infty$ limit, when all the fields $h_1,\dots,h_k$ belong to the set $\{A_0,A_1,\dots,A_{T-1},A_T=B_T,B_{T-1},\dots,B_1,B_0 \}$ of ``hard fields'', or Warning Propagation messages. Some algebra, sketched in Appendix \ref{app_eqs_WP}, leads to:
\beq
g(B_{t_1},\dots,B_{t_n},A_{t_{n+1}},\dots,A_{t_k}) = \begin{cases}
B_{1+{\underset{l}{\min}}(t_1,\dots,t_n)} & \text{if} \ n \ge l \ \text{and} \ \min(t_{n+1},\dots,t_k) \ge 1+ {\underset{l}{\min}}(t_1,\dots,t_n) \\
A_{\min(t_{n+1},\dots,t_k) -1 } &
\text{if} \ n \ge l-1 \ \text{and} \\
&1+ {\underset{l-1}{\min}}(t_1,\dots,t_n) \le \min(t_{n+1},\dots,t_k) \le {\underset{l}{\min}}(t_1,\dots,t_n)
\\
B_0 & \text{otherwise}
\end{cases}
\label{eq_g_ABt}
\end{equation}
where $t_1,\dots,t_n \in \{0,\dots,T-1\}$ and $t_{n+1},\dots,t_k \in \{0,\dots,T\}$. We assumed conventionally that ${\underset{l}{\min}}(t_1,\dots,t_{l-1})=\infty$.
The equation (\ref{eq_g_ABt}) can be given a very intuitive interpretation. The messages $h \in \{A_0,\dots,A_{T-1},B_0,\dots,B_T\}$ can be interpreted as ``warnings'' sent from one vertex of the graph to one of its neighbours, with the following meanings. A vertex $i$ sends a message $h_{i \to j}=B_t$ to one of its neighbour $j$ to say: ``if $j$ is kept inactive at all times the configuration of $i$ and of its sub-tree (the one rooted at $i$ and excluding $j$) leads to complete activation of the sub-tree within the time horizon $T$, and $i$ activates itself at time $t$''. In particular $h_{i \to j}=B_0$ means that $i$ is activated in the initial configuration. On the contrary $i$ sends the message $h_{i \to j}=A_t$ to $j$ to express: ``the complete activation of $i$ and its sub-tree requires that $j$ becomes activated at time $t$''. The rules of Eq.~(\ref{eq_g_ABt}) for the combination of these messages are then obtained by finding the configuration compatible with them, containing the minimal number of active sites in the initial configuration (because of the $\mu \to -\infty$ limit):
\begin{itemize}
\item if strictly less than $l-1$ incoming messages are of the type $B_{t_i}$, with $t_i \in \{0,\dots,T-1\}$, the central site $i$ will never have more than $l$ active neighbours (even with the participation of the receiving site $j$) if it is initially inactive, hence the only way for $i$ to be active at time $T$ is to be active in the initial configuration, which implies $h_{i \to j}=B_0$.
\item if at least $l$ of the incoming messages are of the type $B_{t_i}$, with $t_i \in \{0,\dots,T-1\}$, say $(B_{t_1},\dots,B_{t_n})$, the central site $i$ will become active at time $t=1+{\underset{l}{\min}}(t_1,\dots,t_n)$, without the ``help'' of the activation of the site $j$ receiver of the message. This situation thus leads to a message of type $B_t$, at the condition that all other incoming messages of type $\{A_0,\dots,A_T\}$ do not require the activation of the central site $i$ at a time strictly earlier than $t=1+{\underset{l}{\min}}(t_1,\dots,t_n)$.
\item the participation of the activation of the receiving site $j$ is required at some time $t$ when the above condition is not fulfilled, i.e. when the incoming messages $(A_{t_{n+1}},\dots,A_{t_k})$ require the activation of the central site at some time $t_{\rm act}=\min (t_{n+1}, \dots ,t_k) < 1+ {\underset{l}{\min}}(t_1,\dots,t_n)$. This mechanism is possible if at time $t_{\rm act} -1$ already $l-1$ of the neighbours sending messages of type $B$ are active, i.e. it requires ${\underset{l-1}{\min}}(t_1,\dots,t_n)\le t_{\rm act} -1$. The ``help'' needed from the receiving site is that it is active at some time before $t_{\rm act} -1$; in the limit $\mu \to -\infty$ the least dense configurations, and thus the least stringent constraint on the time of activation is privileged, hence the message sent in this case is $h_{i\to j}= A_{t_{\rm act} -1}$.
\item all cases not fulfilling one of the conditions above require that $i$ is active in the initial configuration to be active at time $T$, hence the message sent is $h_{i\to j}= B_0$.
\end{itemize}
\subsubsection{Energetic 1RSB single sample equations}
Within this ansatz the 1RSB distributions $P(h)$ greatly simplify, as they are supported on the discrete set $h \in \{A_0,A_1,\dots,A_{T-1},A_T=B_T,B_{T-1},\dots,B_1,B_0 \}$. We shall denote $p_t$ the weight in $P(h)$ of the event $h=A_t$, and similarly $q_t$ for $h=B_t$ (with again the convention $p_T=q_T$ to simplify notations), i.e.
\beq
P(h) = \sum_{t=0}^{T-1} p_t \ \delta(h - A_t )
+ \sum_{t=0}^T q_t \ \delta(h - B_t ) \ .
\label{eq_def_pqt}
\end{equation}
The 1RSB recursion relation (\ref{eq_1RSB_Ph}) now reduces to a recursion between these finite-dimensional vectors of probabilities; inserting the definition (\ref{eq_def_pqt}) in the right hand side of (\ref{eq_1RSB_Ph}) and exploiting the combination rule (\ref{eq_g_ABt}) between hard fields, one obtains the following limit for the recursion relation $P=G[P_1,\dots,P_k]$:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
p_t &=& \frac{1}{Z[P_1,\dots,P_k]} e^y {\widetilde{p}}_t \ , \quad
{\widetilde{p}}_t = \sum_{\substack{I,J,K \\ |I| = l-1 \\ |J| \ge 1 }}
\prod_{i \in I}\left(\sum_{t'=0}^t q_{t'}^{(i)}\right)
\prod_{i \in J} p_{t+1}^{(i)}
\prod_{i \in K}
\left(\sum_{t'=t+1}^T q_{t'}^{(i)} + \sum_{t'=t+2}^{T-1} p_{t'}^{(i)}
\right) \ \text{for} \ t \in \{0, \dots , T-1\} \nonumber \\
q_t &=& \frac{1}{Z[P_1,\dots,P_k]} e^y {\widetilde{q}}_t \ , \quad
{\widetilde{q}}_t = \sum_{\substack{I,J,K \\ |I| \le l-1 \\ |I| +|J| \ge l }}
\prod_{i \in I}\left(\sum_{t'=0}^{t-2} q_{t'}^{(i)}\right)
\prod_{i \in J} q_{t-1}^{(i)}
\prod_{i \in K}\left(\sum_{t'=t}^T q_{t'}^{(i)} + \sum_{t'=t}^{T-1} p_{t'}^{(i)} \right)\ \text{for} \ t \in \{1, \dots , T\}
\nonumber \\
q_0 &=& \frac{1}{Z[P_1,\dots,P_k]} \left[1-\sum_{t=0}^{T-1} {\widetilde{p}}_t - \sum_{t=1}^T {\widetilde{q}}_t \right] \ , \qquad \qquad Z[P_1,\dots,P_k] = 1 + (e^y -1) \left[\sum_{t=0}^{T-1} {\widetilde{p}}_t + \sum_{t=1}^T {\widetilde{q}}_t \right]
\label{eq_G_1RSBy}
\eea
the reweighting term of Eq.~(\ref{eq_ziter}) becoming indeed
${z_{\rm iter}}(h_1,\dots,h_k)^m = e^{y a_0(h_1,\dots,h_k)}$, hence the factor $e^y$
multiplying the probabilities of all warnings except $B_0$; this is indeed the only case where an active site has to be inserted in the initial configuration.
To compute the 1RSB potential we have to study the limit of the contribution of site and edge terms in the limit $\mu\to-\infty$, $m\to 0$. We have from Eq.~(\ref{eq_zsite})
\beq
{z_{\rm site}}(h_1,\dots,h_{k+1})^m \to \exp\left[y
\max\left(0, \max_{t\in[1,T]}
\max_{\substack{I,J,K \\ |I| \le l-1 \\ |I|+|J| \ge l}}
{\cal S}_t(h_1,\dots,h_{k+1};I,J,K) \right)
\right] \ ,
\end{equation}
which can be simplified following the same reasoning than the one which led to (\ref{eq_g_ABt}). This yields
\beq
{{\cal Z}_{\rm site}}(P_1,\dots,P_{k+1}) \to
1+(e^y-1) \sum_{t=1}^T \sum_{\substack{I,J,K \\ |I| \le l-1 \\ |I| +|J| \ge l }}
\prod_{i \in I}\left(\sum_{t'=0}^{t-2} q_{t'}^{(i)}\right)
\prod_{i \in J} q_{t-1}^{(i)}
\prod_{i \in K}\left(\sum_{t'=t}^T q_{t'}^{(i)} + \sum_{t'=t}^{T-1} p_{t'}^{(i)} \right) \ ,
\label{eq_Zsite_y}
\end{equation}
where $I,J,K$ is a partition of $\{1,\dots,k+1\}$. This expression can be interpreted intuitively in terms of the warnings defined above; the factor multiplying $(e^y-1)$ is indeed the probability of complete activation, at time $t \in \{1,\dots,T\}$, for an initially empty site receiving messages $(h_1,\dots,h_{k+1})$ from its neighbours, with their respective distributions $P_1,\dots,P_{k+1}$. As a matter of fact, for its activation to occur at time $t$ at least $l$ neighbours must have activated without any help from the central site at time $t-1$, no more than $l-1$ must be active at time $t-2$ (otherwise the activation time would be strictly less than $t$), and the neighbours sending messages of type $A_{t'}$ should not require activation at a time $t'<t$.
For the edge term we obtain from Eq.~(\ref{eq_zedge})
\beq
{z_{\rm edge}}(h_1,h_2)^m \to \exp\left[ - y \min_{t\in[0,T]} \min(
(a_0^{(1)}-b_t^{(1)})+ (a_0^{(2)}-a_t^{(2)}), (a_0^{(1)}-a_t^{(1)})+ (a_0^{(2)}-b_t^{(2)}))
\right] \ ,
\end{equation}
hence
\beq
{{\cal Z}_{\rm edge}}(P_1,P_2) \to e^{-y}+(1-e^{-y}) \left[
\left(\sum_{t=0}^T q_t^{(1)} \right)
\left(\sum_{t=0}^T q_t^{(2)} \right)
+ \sum_{t=0}^{T-1} p_t^{(1)} \sum_{t'=0}^t q_{t'}^{(2)}
+ \sum_{t=0}^{T-1} p_t^{(2)} \sum_{t'=0}^t q_{t'}^{(1)}
\right] \ .
\label{eq_Zedge_y}
\end{equation}
One can interpret the factor multiplying $(1-e^{-y})$ as the probability of complete activation when two messages $(h_1,h_2)$ drawn with the probabilities $P_1,P_2$ are sent in the two opposite directions of an edge.
Let us summarize the main findings of this subsection. In the limit $\mu \to -\infty$, $m\to 0$ with $y = -\mu m$ the 1RSB formalism simplifies in the following way. The cavity field distributions $P_{i \to j}(h)$ have now a discrete support with $2T$ possible values, each of them is thus described by a (normalized) vector of $2T$ probabilities denoted $\{p_t,q_t\}$. These vectors are solutions of recurrence equations of the form $P_{i\to j}=G(\{P_{k\to i} \}_{k\in{\partial i \setminus j}})$, the mapping $G$ being defined in Eq.~(\ref{eq_G_1RSBy}). The energetic limit of the 1RSB potential is then computed as
\beq
\Phi_{\rm e}(y) = - y +
\frac{1}{N} \sum_{i=1}^N \ln\left( {{\cal Z}_{\rm site}}(\{ P_{j\to i} \}_{j \in {\partial i}} )\right)
-\frac{1}{N}\sum_{{\langle} i,j{\rangle}} \ln\left( {{\cal Z}_{\rm edge}}(P_{i \to j},P_{j\to i}) \right)
\ ,
\label{eq_Phie_single_sample}
\end{equation}
with the expression of ${{\cal Z}_{\rm site}}$ and ${{\cal Z}_{\rm edge}}$ given in Eqs.~(\ref{eq_Zsite_y},\ref{eq_Zedge_y}). This expression of $\Phi_{\rm e}$ is variational, its derivative with respect to $y$ (which is needed in the computation of the inverse Legendre transform in (\ref{eq_Legendre_y})) can be taken on the explicit dependence only.
\subsubsection{Random regular graphs}
For the reasons already exposed in the context of the RS and of the full 1RSB cavity formalism a factorized solution of the energetic 1RSB equations can be searched for when dealing with random $k+1$ regular graphs with a constant threshold of activation $l$. One has thus a single vector of probabilities $P=(\{p_t,q_t\})$, fixed-point solution of Eq.~(\ref{eq_G_1RSBy}), from which the energetic 1RSB potential is obtained as
\beq
\Phi_{\rm e}(y) = - y + \ln\left( {{\cal Z}_{\rm site}}(P,\dots,P )\right)
-\frac{k+1}{2} \ln\left( {{\cal Z}_{\rm edge}}(P,P) \right) \ ,
\label{eq_Phie_reg}
\end{equation}
with ${{\cal Z}_{\rm site}}$ and ${{\cal Z}_{\rm edge}}$ defined in Eqs.~(\ref{eq_Zsite_y},\ref{eq_Zedge_y}).
\section{Results of the cavity method for random regular graphs}
\label{sec_results_rrg}
We shall present now the results of the resolution of the cavity equations for random regular graphs of degree $k+1$, with an activation threshold equal to $l$ for all vertices. In all this discussion it will be understood that $\epsilon=+\infty$, i.e. we only consider initial configurations that activate the whole graph in $T$ steps. We will first present in Sec.~\ref{sec_res_finiteT} the results for finite values of $T$, which are qualitatively the same for all values of $k,l$ and $T$; the behaviour of the replica symmetric cavity method are first displayed, then we turn to the effects of replica symmetry breaking, in particular in the ``energetic'' limit to compute the minimal density of initially active sites in activating configurations. In a second part (Sec.~\ref{sec_res_largeT}) we shall discuss the limit $T\to\infty$, in which some further analytical computations can be performed. In this case several qualitatively distinct phenomena emerge, depending on the values of $k$ and $l$.
\subsection{Finite $T$ results}
\label{sec_res_finiteT}
\subsubsection{Replica symmetric formalism}
The technical details of the resolution of the RS equation $h=g(h,\dots,h)$, where $g$ is given in Eq.~(\ref{eq_g}), and of the computation of the free-entropy density, are deferred to the Appendix \ref{sec_app}. From a numerical point of view it is an easy task, as it corresponds essentially to the resolution of a set of $2T$ equations on $2T$ unknowns. Let us discuss the numerical results obtained in this way. On the left panel of Fig.~\ref{fig_RS_k2l2_T3} we display the curve $\theta(\mu)$ of the average fraction of initially active sites as a function of the chemical potential $\mu$; the curve is for $k=l=2$ and $T=3$, the qualitative features are independent of these precise values. This function is increasing as it should, and reaches a finite limit when $\mu \to -\infty$, that would be the candidate value for ${\theta_{\rm min}}$ if the RS computation was correct in this limit. This however cannot be true, as revealed from the computation of the entropy, displayed in the right panel of Fig.~\ref{fig_RS_k2l2_T3}: for $\mu < \mu_{s=0}$ the RS entropy becomes negative, which is a certain indication of the inadequacy of the RS theory in this regime. In Fig.~\ref{fig_RS_k2l2_softheta} we display the results for the entropy $s(\theta)$ of the number of configurations with a fraction $\theta$ of initially active sites, for the regime of positive entropies where the RS prediction cannot be ruled out at once (for the cases $k=l=2$ and $k=3$, $l=2$). For increasing values of $T$ these curves converge to a limit, this will be further discussed in Sec.~\ref{sec_kequall_Tinfty}. The numerical values of the chemical potential and of the fraction of active sites at the point of entropy cancellation, which would be the best guess from the RS computation of the value of ${\theta_{\rm min}}$, denoted respectively $\mu_{s=0}$ and ${\theta_{\rm min,0}}$, can be found for various values of $T$ in the Tables \ref{tab_k2l2}, \ref{tab_k3l3} and \ref{tab_k3l2} for the cases $k=l=2$, $k=l=3$ and $k=3$, $l=2$ respectively. For $T=1$ they reproduce, as they should, the results of the Biroli-M\'ezard model given in~\cite{bm}.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{fig_k2l2_T3_thetaofmu.eps}
\hspace{1cm}
\includegraphics[width=7cm]{fig_k2l2_T3_sofmu.eps}
\end{center}
\caption{The density of initially active sites $\theta$ (left panel) and the entropy $s$ (right panel) as a function of the chemical potential $\mu$, computed from the replica symmetric cavity equations, for $k=l=2$ and $T=3$.}
\label{fig_RS_k2l2_T3}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{fig_k2l2_softheta.eps}
\hspace{1cm}
\includegraphics[width=7cm]{fig_k3l2_softheta.eps}
\end{center}
\caption{The RS entropy $s(\theta)$ of configurations with a fraction $\theta$ of initially active sites able to activate completely the graph within time $T$, for $k=l=2$ (left panel) and $k=3$, $l=2$ (right panel). The curve labelled ``random'' is the binary entropy function $-\theta \ln \theta - (1-\theta) \ln(1-\theta)$ that counts all configurations with such an initial density. The curves in the limit $T\to\infty$ are computed analytically, from Eq.~(\ref{eq_s_klequal_Tinfty}) for the left panel and (\ref{eq_lmk_Tinfty_RS}) for the right panel, see Sec.~\ref{sec_res_largeT} for a further discussion of this limit.}
\label{fig_RS_k2l2_softheta}
\end{figure}
\begin{table}
\begin{tabular}{|c||c|c||c|c|c|c||c|c|}
\hline
& \multicolumn{2}{c||}{RS} & \multicolumn{4}{c||}{1RSB} & \multicolumn{2}{c|}{energetic 1RSB} \\
\hline
$T$ & $\mu_{s=0}$ & ${\theta_{\rm min,0}}$ & ${\mu_{\rm d}}$ & ${\theta_{\rm d}}$ & ${\mu_{\rm c}}$ & ${\theta_{\rm c}}$ & ${y_{\rm s}}$ & ${\theta_{\rm min,1}}$ \\
\hline
1 & -7.403996 & 0.422251 & -6.49 & 0.4292 & -6.69 & 0.4275 & 5.563433 & 0.424257\\
\hline
2 & -11.374979 & 0.325742 & -9.89 & 0.3291 & -11.23 & 0.3260 & 10.826348 & 0.325882 \\
\hline
3 & -17.292682 & 0.289093 & -13.7 & 0.2922 &-17.28 & 0.2890 & 17.232166 & 0.289097 \\
\hline
4 & -24.936318 & 0.271564 & -20.9 & 0.2731 &-24.93 & 0.2715 & 24.933659 & 0.271564 \\
\hline
5 & -34.966263 & 0.262167 & -31.3 & 0.2628 &-34.63 & 0.2622 & 34.966225 & 0.262167\\
\hline
6 & -49.901175 & 0.256844 & & & & & 49.901175 & 0.256844 \\
\hline
7 & -74.984724 & 0.253779 & & & & & 74.984724 & 0.253779 \\
\hline
8 & -120.79085 & 0.252036 & & & & & 120.79085 & 0.252036 \\
\hline
10 & -378.44778 & 0.250553 & & & & & 378.44778 & 0.250553 \\
\hline
15 & $-1.069 \ 10^4$ & 0.250018 & & & & & $1.069 \ 10^4$ & 0.250018 \\
\hline
20 & $-3.4 \ 10^5$ & 0.250000 & & & & & $3.4 \ 10^5$ & 0.250000 \\
\hline
$\infty$ & $-\infty$ & $\frac{1}{4}$ & & & & & $+\infty$ & $\frac{1}{4}$ \\
\hline
\end{tabular}
\caption{Numerical results from the cavity computations at finite $T$ for $k=l=2$; the results in the limit $T\to\infty$ are explained in Sec.~\ref{sec_res_largeT}.}
\label{tab_k2l2}
\end{table}
\subsubsection{1RSB results}
As we have seen above the hypothesis underlying the RS computation must go wrong when $\mu$ is decreased towards $-\infty$, as the entropy computed within the RS scheme becomes negative for $\mu < \mu_{s=0}$; a 1RSB computation is thus required to investigate the limit $\mu \to -\infty$ and hence the properties of the least dense activating initial conditions, in particular their density ${\theta_{\rm min}}$.
We have thus solved numerically the 1RSB equations (\ref{eq_1RSB}) using population dynamics methods~\cite{cavity}, i.e. representing $P(h)$ as a weighted sample of fields $h_i$. This method has become fairly standard and we shall not give more details on the procedure, see for instance~\cite{cavity,MeMo_book} for detailed presentations. In the particularly important $m=1$ case we used a version of this procedure, inspired by the tree reconstruction problem, that allows to get rid of the reweighting terms in (\ref{eq_1RSB}) and is thus much more precise and efficient numerically, see~\cite{MeMo06,MRS08} for more technical details.
The results of these investigations follow the usual pattern encountered in constraint satisfaction problems~\cite{KrMoRiSeZd}: for large enough values of $\mu$ (i.e. for dense enough initial configurations) there is no non-trivial solution of the 1RSB equation with $m=1$; decreasing $\mu$ a non-trivial solution appears discontinuously at a threshold ${\mu_{\rm d}}$ (the ``dynamic'' transition). Its complexity (or configurational entropy) $\Sigma$ is positive in an interval $\mu \in [{\mu_{\rm c}},{\mu_{\rm d}}]$, which thus corresponds to a ``dynamic 1RSB phase'' with an exponential number of clusters contributing to the Gibbs measure, see Fig.~\ref{fig_k2l2T1m1} for an illustration in the case $T=1$. The numerical values of ${\mu_{\rm d}}$ and ${\mu_{\rm c}}$ (as well as the associated densities of initially active sites ${\theta_{\rm d}}$ and ${\theta_{\rm c}}$), can be found for several values of $T$ in the Tables \ref{tab_k2l2}, \ref{tab_k3l3} and \ref{tab_k3l2}. For the values of $\mu$ in the interval $[{\mu_{\rm c}},{\mu_{\rm d}}]$ the thermodynamic predictions of the RS computations are correct. Note that in all the cases we investigated ($k=2,3$, $2\le l \le k$ and $T\le 5$) we always found a discontinuous transition with ${\mu_{\rm c}} < {\mu_{\rm d}}$; we cannot rule out the possibility that for other values of the parameters the replica symmetry breaking transition is continuous with ${\mu_{\rm c}}={\mu_{\rm d}}$ (as happens for instance in the independent set problem at low degrees~\cite{is_japan}).
\begin{figure}
\includegraphics[width=8.5cm]{fig_k2l2_T1_Sigmaofmu_m1.eps}
\caption{The complexity at $m=1$ as a function of the chemical potential $\mu$, for $k=l=2$ and $T=1$. The function is defined for $\mu<{\mu_{\rm d}} \approx -6.49$, the complexity being positive for $\mu > {\mu_{\rm c}} \approx -6.69$.}
\label{fig_k2l2T1m1}
\end{figure}
Lowering further the chemical potential, i.e. in the regime $\mu<{\mu_{\rm c}}$, the complexity at $m=1$ becomes negative. This is thus a true replica symmetry breaking phase with only a sub-exponential number of clusters contributing to the Gibbs measure; ${\mu_{\rm c}}$ corresponds to the ``condensation'' transition. In this phase the thermodynamic properties of the model differ from the RS prediction and are given by the properties of the clusters selected by the static value of the Parisi parameter, ${m_{\rm s}}(\mu)$, for which the complexity vanishes. This value can be determined by computing the complexity as a function of $m$, for a fixed value of $\mu$, see left panel of Fig.~\ref{fig_k2l2T1condensed} for an example.
To compute the minimal density ${\theta_{\rm min}}(T)$ one has to take the limit $\mu\to-\infty$; we have introduced above in Sec.~\ref{sec_1RSB_y} a simplifying ansatz in this limit, assuming in particular a finite value of $-\mu m$. To check the consistency of this ansatz we solved the complete 1RSB equations for $T=1$ and several values of $\mu$ large and negative. The Parisi parameter ${m_{\rm s}}$ is plotted as a function of $-1/\mu$ in the right panel of Fig.~\ref{fig_k2l2T1condensed}; in the limit $\mu \to -\infty$ one obtains indeed a linear behaviour, corresponding to a finite limit of $-\mu {m_{\rm s}}$.
\begin{figure}
\includegraphics[width=8cm]{fig_k2l2_T1_Sigmaofm_mu-7.5.eps}\hspace{1cm}
\includegraphics[width=8cm]{fig_k2l2_T1_mstatofmu.eps}
\caption{Study of the condensed phase for $k=l=2$ and $T=1$.
Left panel: complexity as a function of $m$ for $\mu=-7.5<{\mu_{\rm c}}$, the complexity vanishes for ${m_{\rm s}} \approx 0.84$. Right panel: Parisi parameter ${m_{\rm s}}$ as a function of $-1/\mu$, departing from 1 for $\mu<{\mu_{\rm c}}$; the dashed line corresponds to the linear behaviour $-\mu {m_{\rm s}}=5.56$ that fits the $\mu \to -\infty$ limit.}
\label{fig_k2l2T1condensed}
\end{figure}
\subsubsection{Energetic 1RSB results}
We turn now to the results obtained via the energetic 1RSB cavity method, i.e. taking simultaneously the limits $\mu \to -\infty$ and $m \to 0$ with a finite value for $y=-\mu m$. The equations to solve in this case amounts to find the fixed point of Eq.~(\ref{eq_G_1RSBy}), from which one obtains the 1RSB potential (\ref{eq_Phie_reg}) and the energetic complexity $\Sigma_{\rm e}(\theta)$ from the Legendre transform structure explained in (\ref{eq_Legendre_y}), as a parametric plot varying the parameter $y$. The computational complexity of this problem is drastically reduced compared to the complete 1RSB equations: as in the RS case one has a set of (roughly) $2T$ equations on $2T$ real unknowns, instead of an equation on a probability distribution of fields. More technical details on the procedure to solve these equations can be found in Appendix \ref{sec_app}.
Fig.~\ref{fig_1RSBy_k2l2_Sigma} displays the energetic complexity $\Sigma_{\rm e}(\theta)$ for a few values of $T$, in the cases $k=l=2$ and $k=3$, $l=2$. The expert reader will notice that we restricted the range of $y$ used in this plot to the so-called physical branch, in such a way that $\Sigma_{\rm e}$ is a concave function of $\theta$. The most important characteristics of these curves are the values of ${\theta_{\rm min,1}}$ where the complexity vanishes, and the corresponding values ${y_{\rm s}}$ of the parameter $y$; these are reported for several values of $T$ in the last columns of the Tables \ref{tab_k2l2}, \ref{tab_k3l3} and \ref{tab_k3l2}. Indeed ${\theta_{\rm min,1}}$ is the 1RSB prediction for ${\theta_{\rm min}}$, as it corresponds to the smallest density of active sites in initial configurations belonging to clusters with a non-negative complexity. For $T=1$ these values can be successfully cross-checked with the results of the Biroli-M\'ezard model~\cite{bm}, and the parameter ${y_{\rm s}}$ agrees with the fit of $-\mu {m_{\rm s}}(\mu)$ in the limit $\mu \to -\infty$ obtained from the full 1RSB equations (cf. right panel of Fig.~\ref{fig_k2l2T1condensed}).
\begin{figure}
\includegraphics[width=8cm]{fig_k2l2_sigmaeoftheta.eps}
\hspace{1cm}
\includegraphics[width=8cm]{fig_k3l2_sigmaeoftheta.eps}
\caption{The complexity $\Sigma_{\rm e}(\theta)$ obtained from the energetic 1RSB cavity formalism, for $k=l=2$ (left panel) and $k=3$, $l=2$ (right panel); see Sec.~\ref{sec_res_largeT} for explanations on the $T\to\infty$ result.}
\label{fig_1RSBy_k2l2_Sigma}
\end{figure}
\begin{table}
\begin{tabular}{|c||c|c||c|c|c|c||c|c|}
\hline
& \multicolumn{2}{c||}{RS} & \multicolumn{4}{c||}{1RSB} & \multicolumn{2}{c|}{energetic 1RSB} \\
\hline
$T$ & $\mu_{s=0}$ & ${\theta_{\rm min,0}}$ & ${\mu_{\rm d}}$ & ${\theta_{\rm d}}$ & ${\mu_{\rm c}}$ & ${\theta_{\rm c}}$ & ${y_{\rm s}}$ & ${\theta_{\rm min,1}}$ \\
\hline
1 & -6.113951 & 0.479455 & -5.35 & 0.4906 & -5.39 & 0.4900 & 4.644980 & 0.482712 \\
\hline
2 & -8.175902 & 0.397326 & -7.38 & 0.4027 & -7.95 & 0.3988 & 7.485437 & 0.397922 \\
\hline
3 & -10.381917 & 0.366187 & -8.63 & 0.3725 & -10.33 & 0.3663 & 10.077681 & 0.366291 \\
\hline
4 & -13.140888 & 0.351221 & -9.59 & 0.3583 & -13.11 & 0.3513 & 13.037666 & 0.351234 \\
\hline
5 & -17.249334 & 0.343205 & -10.3 & 0.3507 & -17.36 & 0.3432 & 17.232334 & 0.343206 \\
\hline
6 & -24.322138 & 0.338721 & & & & & 24.321721 & 0.338721 \\
\hline
7 & -35.739653 & 0.336191 & & & & & 35.739653 & 0.336191 \\
\hline
8 & -54.198587 & 0.334760 & & & & & 54.198587 & 0.334760 \\
\hline
$\infty$ & $-\infty$ & $\frac{1}{3}$ & & & & & $+\infty$ & $\frac{1}{3}$ \\
\hline
\end{tabular}
\caption{Numerical results from the cavity computations at finite $T$ for $k=l=3$; the results in the limit $T\to\infty$ are explained in Sec.~\ref{sec_res_largeT}.}.
\label{tab_k3l3}
\end{table}
\begin{table}
\begin{tabular}{|c||c|c||c|c|c|c||c|c|}
\hline
& \multicolumn{2}{c||}{RS} & \multicolumn{4}{c||}{1RSB} & \multicolumn{2}{c|}{energetic 1RSB} \\
\hline
$T$ & $\mu_{s=0}$ & ${\theta_{\rm min,0}}$ & ${\mu_{\rm d}}$ & ${\theta_{\rm d}}$ & ${\mu_{\rm c}}$ & ${\theta_{\rm c}}$ & ${y_{\rm s}}$ & ${\theta_{\rm min,1}}$ \\
\hline
1 & -7.730059 & 0.362794 & -7.06 & 0.3681 & -7.38 & 0.3654 & 6.778540 & 0.363813 \\
\hline
2 & -10.21534 & 0.236821 & -9.16 & 0.2416 & -10.12 & 0.2372 & 9.873120 & 0.237009 \\
\hline
3 & -11.90150 & 0.182272 & -10.38 & 0.1875 & -11.85 & 0.1824 & 11.72892 & 0.182338 \\
\hline
4 & -13.03158 & 0.151659 & -11.45 & 0.1563 & -13.00 & 0.1517 & 12.92114 & 0.151693\\
\hline
5 & -13.80059 & 0.132014 & -12.47 & 0.1354 &-13.78 & 0.1321 & 13.71834 & 0.132036\\
\hline
6 & -14.33193 & 0.118324 & & & & & 14.26439 & 0.118341\\
\hline
7 & -14.70251 & 0.108237 & & & & & 14.64332 & 0.108251 \\
\hline
8 & -14.96150 & 0.100498 & & & & & 14.90729 & 0.100510 \\
\hline
10 &-15.26375 & 0.089415& & & & & 15.21429 & 0.089425\\
\hline
15 & -15.42086 & 0.074242 & & & & & 15.37163 & 0.074251 \\
\hline
20 & -15.27922 & 0.066569 & & & & & 15.22489 & 0.066579 \\
\hline
30 & -14.85174 & 0.058995 & & & & & 14.78367 & 0.059008 \\
\hline
$\infty$ & -12.72072 & 0.046283 & & & & & 12.54796 & 0.046328 \\
\hline
\end{tabular}
\caption{Numerical results from the cavity computations at finite $T$ for $k=3$, $l=2$; the results in the limit $T\to\infty$ are explained in Sec.~\ref{sec_res_largeT}.}
\label{tab_k3l2}
\end{table}
\subsection{The large $T$ limit}
\label{sec_res_largeT}
The limit case $T\to\infty$ is particularly interesting as it corresponds to the original influence maximization problem with no constraint on the time taken to activate the whole graph. This limit can be performed analytically for the RS and energetic 1RSB formalism; the technical details of these computations can be found in Appendix \ref{sec_app_Tinfty}, we present here the results of these analytical simplifications. It turns out that the case $k=l$ is qualitatively different from the case $k>l$, we shall thus divide this section according to this distinction.
\subsubsection{The case $k=l$}
\label{sec_kequall_Tinfty}
Let us first recall that when $k=l$ the dynamics from a random initial configuration of density $\theta$ has a continuous transition at ${\theta_{\rm r}}(k,k)=\frac{k-1}{k}$ (see Sec.~\ref{sec_reminder_random}); we also saw in Sec.~\ref{sec_connections} that minimal contagious sets (with no constraint on the activation time) correspond to minimal decycling sets, which led to the bound ${\theta_{\rm min}}(k,k) \ge \frac{k-1}{2 k} =\frac{{\theta_{\rm r}}(k,k)}{2}$. In the rest of this subsection we shall for simplicity abbreviate ${\theta_{\rm r}}(k,k)$ by ${\theta_{\rm r}}$.
As suggested by the left panel of Fig.~\ref{fig_RS_k2l2_softheta} in the case $k=l=2$, the RS entropy $s(\theta)$ converges to a limit curve when $T\to\infty$. This limit curve can actually be computed analytically for all $k$; we defer the details of the computation to App.~\ref{sec_app_Tinfty_klequal} and only state here the properties of this limit curve. For $\theta \ge {\theta_{\rm r}}$ it coincides with the binary entropy function $-\theta \ln \theta - (1-\theta) \ln (1-\theta)$; this is a posteriori obvious. Indeed by definition of ${\theta_{\rm r}}$ typical configurations in this density range do activate the whole graph, hence the number of activating initial configurations coincide (at the leading exponential order) with the total number of configurations of this density. A non-trivial portion of the limit curve arises in the density range $[{\theta_{\rm r}}/2,{\theta_{\rm r}}]$, where it is given by
\beq
s(\theta) = - \frac{k}{2} \left(2 \theta - {\theta_{\rm r}} \right) \ln (2 \theta - {\theta_{\rm r}}) + k \theta \ln \theta
+(1-\theta) \ln(k-1) - \frac{k+1}{2} \ln\left(\frac{k-1}{k} \right) \ .
\label{eq_s_klequal_Tinfty}
\end{equation}
This function has the same value and the same first derivative than the binary entropy function in ${\theta_{\rm r}}$, while at the lower limit ${\theta_{\rm r}}/2$ of its range of definition it has an infinite derivative with a finite value
\beq
s({\theta_{\rm r}}/2) = \ln k - \frac{k-1}{2k} \ln (k-1) - \frac{k-1}{2} \ln 2 \ .
\label{eq_s_trovertwo}
\end{equation}
The parametric plot of $s(\theta)$ also contains a vertical segment for $\theta = {\theta_{\rm r}}/2$, from $-\infty$ to the value given in (\ref{eq_s_trovertwo}).
The complexity $\Sigma_{\rm e}(\theta)$ of the energetic 1RSB formalism also converges to a limit curve when $T\to\infty$, as shown in Fig.~\ref{fig_1RSBy_k2l2_Sigma} and obtained analytically in App.~\ref{sec_app_Tinfty_klequal}. This limit curve has the same vertical segment in ${\theta_{\rm r}}/2$ from $-\infty$ to the value (\ref{eq_s_trovertwo}); the non-trivial part of the curve is given in a parametrized form as follows:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
\Sigma_{\rm e}({\widetilde{\lambda}})&=& \ln {{\cal Z}_{\rm site}}({\widetilde{\lambda}}) - \frac{k+1}{2} \ln {{\cal Z}_{\rm edge}}({\widetilde{\lambda}}) - y({\widetilde{\lambda}}) (1-\theta({\widetilde{\lambda}})) \ , \label{eq_1RSBy_klequal_Tinfty} \\
\theta({\widetilde{\lambda}}) &=& 1
- \frac{e^{y({\widetilde{\lambda}})}}{e^{y({\widetilde{\lambda}})}-1} \frac{{{\cal Z}_{\rm site}}({\widetilde{\lambda}}) -1}{{{\cal Z}_{\rm site}}({\widetilde{\lambda}}) }
- \frac{k+1}{2} \frac{1}{e^{y({\widetilde{\lambda}})}-1} \frac{1-{{\cal Z}_{\rm edge}}({\widetilde{\lambda}})}{{{\cal Z}_{\rm edge}}({\widetilde{\lambda}}) } \ ,
\eea
where ${\widetilde{\lambda}}$ is the positive parameter along the curve, the Parisi parameter
\beq
y({\widetilde{\lambda}})= \ln \left(\frac{(1+{\widetilde{\lambda}})^k - k \, {\widetilde{\lambda}}^{k-1} - {\widetilde{\lambda}}^k}{(k-1)\, {\widetilde{\lambda}}^k} \right) \ ,
\end{equation}
is the slope of the tangent to the curve $\Sigma_{\rm e}(\theta)$, and
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{{\cal Z}_{\rm site}}({\widetilde{\lambda}}) &=& 1+ \frac{(k+1+{\widetilde{\lambda}})((1+{\widetilde{\lambda}})^{k-1}-k\, {\widetilde{\lambda}}^{k-1})}{(k-1)(1+{\widetilde{\lambda}})^k} \ ,\\
{{\cal Z}_{\rm edge}}({\widetilde{\lambda}}) &=&\frac{{\widetilde{\lambda}}}{1+{\widetilde{\lambda}}} \left(1+\frac{(1+{\widetilde{\lambda}})^{k-1}-{\widetilde{\lambda}}^{k-1}}{(1+{\widetilde{\lambda}})^k - k \, {\widetilde{\lambda}}^{k-1} - {\widetilde{\lambda}}^k} \right) \ .
\eea
When ${\widetilde{\lambda}} \to 0^+$ this part of the curve connects with the vertical segment in ${\theta_{\rm r}}/2$. The large values of ${\widetilde{\lambda}}$ yield a non-concave branch of $\Sigma_{\rm e}$ that has to be discarded.
Depending on the value of $k$ qualitatively different behaviours emerge from the analysis of the RS entropy and 1RSB energetic complexity:
\begin{itemize}
\item
For $k=l=2$ the entropy of the endpoint in ${\theta_{\rm r}}/2$ given in (\ref{eq_s_trovertwo}) is strictly positive (it is equal to $(\ln 2)/2$); moreover the energetic complexity curve converges, in the $T\to\infty$ limit, to a vertical segment (the non-trivial part parametrized by ${\widetilde{\lambda}}$ is convex and has thus to be discarded). This leads to the conclusion that ${\theta_{\rm min}}={\theta_{\rm r}}/2=1/4$ in this case, saturating the lowerbound of (\ref{eq_bound_kk}), and recovering the rigorous result of~\cite{decycling} on the decycling number of 3-regular graphs. This is a reassuring evidence in favour of the validity of the approach, in particular on the interversion of the $T\to\infty$ and $N\to\infty$ limit. It would be an even more challenging computation to determine the limit of ${\theta_{\rm d}}$ and ${\theta_{\rm c}}$ as $T$ diverges; we are however tempted to conjecture that they both go to $1/4$ and that the effects of replica symmetry breaking are irrelevant in this limit. A numerical argument in favour of this conjecture will be presented in Sec.~\ref{sec_single_sample}, where it is shown that a simple greedy algorithm is able to find contagious sets of these densities. Assuming this is true, the expression (\ref{eq_s_klequal_Tinfty}) would give for $k=2$ the typical (quenched) entropy of the decycling sets of 3-regular random graphs in their non-trivial regime of densities $[1/4,1/2]$. Note that the coincidence of the RS entropy and 1RSB energetic complexity at ${\theta_{\rm min}}$ is reminiscent of the phenomenology discussed for the matching problem in~\cite{ZdMe_matchings}, which might suggest that the minimal density activating configurations are at a large Hamming distance in configuration space one from the other.
\item
For $k=l=3$ the expression (\ref{eq_s_trovertwo}) of the entropy in ${\theta_{\rm r}}/2$ is still positive (equal to $\ln 3 - (4/3) \ln 2$), hence the endpoint of the non-trivial part of both the RS entropy and the 1RSB complexity curves occurs in ${\theta_{\rm min,0}}={\theta_{\rm min,1}}={\theta_{\rm r}}/2=1/3$, saturating again the bound (\ref{eq_bound_kk}). This leads to the conclusion that ${\theta_{\rm min}}=1/3$ in this case, as was also conjectured in~\cite{decycling}. However, at variance with the previous case, the energetic complexity curve has a non-trivial part for $\theta > {\theta_{\rm min}}$, as shown in the left panel of Fig.~\ref{fig_Tinfty_klequal}. We thus expect that the limits of ${\theta_{\rm d}}$ and ${\theta_{\rm c}}$ when $T\to\infty$ are strictly greater than $1/3$, hence that simple algorithms would have difficulties to find the minimal contagious sets (see Sec.~\ref{sec_single_sample} for a numerical check of this statement), and that the RS entropy (\ref{eq_s_klequal_Tinfty}) is incorrect for some regime of densities close to $1/3$.
\item
Finally when $k=l\ge 4$ the entropy in (\ref{eq_s_trovertwo}) is negative, the cancellation of $s$ occurs at a value ${\theta_{\rm min,0}}$ strictly between ${\theta_{\rm r}}/2$ and ${\theta_{\rm r}}$, see the right panel of Fig.~\ref{fig_Tinfty_klequal}. The energetic complexity vanishes on its non-trivial part parametrized by ${\widetilde{\lambda}}$, at a value ${\theta_{\rm min,1}}$ slightly larger than ${\theta_{\rm min,0}}$, see Table~\ref{table_res_Tinfty} for some numerical values. Whether ${\theta_{\rm min,1}}$ should be taken as a conjectured exact value for ${\theta_{\rm min}}$ or simply as a lowerbound is dubious and might depend on the value of $k$. Indeed one should test the stability of the 1RSB ansatz against further levels of replica symmetry breaking. This computation is in principle doable along the lines of~\cite{stab1,stab2,bm}, but has not been performed yet. It is however relatively easy to set up an asymptotic expansion at large $k$ of the thresholds ${\theta_{\rm min,0}}$ and ${\theta_{\rm min,1}}$ from the expressions (\ref{eq_s_klequal_Tinfty},\ref{eq_1RSBy_klequal_Tinfty}). One finds that the first terms of the expansion are equal at the RS and 1RSB level, it is thus natural to conjecture that they are indeed the correct expansion of ${\theta_{\rm min}}$, namely
\beq
{\theta_{\rm min}}(k,k) = 1- \frac{2 \ln k}{k} - \frac{2}{k} + O\left(\frac{1}{k \ln k} \right) \ .
\end{equation}
This conjecture is in agreement with the rigorous lowerbound proven in~\cite{lb_klequal_rig},
\beq
{\theta_{\rm min}}(k,k) \ge 1- \frac{2 \ln k}{k} - \frac{4-2\ln 2}{k} + o\left(\frac{1}{k} \right) \ .
\end{equation}
It can also be compared with the asymptotic expansion in the case $l=k+1$~\cite{is_Frieze} where the inactive vertices have to form an independent set of the graph:
\beq
{\theta_{\rm min}}(k,k+1) = 1- \frac{2 \ln k}{k} + \frac{2 \ln\ln k}{k} + \frac{2 \ln 2 -2}{k} + o\left(\frac{1}{k} \right) \ .
\end{equation}
The third term of this expansion is of a larger order; indeed the condition imposed on the graph induced by the inactive vertices is much more stringent when $l=k+1$ (it has to be made of isolated vertices) with respect to the case $l=k$ (it only has to be acyclic).
\end{itemize}
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{fig_k3l3_Tinf_sandsigmaeoftheta.eps}
\hspace{1cm}
\includegraphics[width=7cm]{fig_k4l4_Tinf_sandsigmaeoftheta.eps}
\end{center}
\caption{The RS entropy $s(\theta)$ and energetic 1RSB complexity $\Sigma_{\rm e}(\theta)$ in the $T\to \infty$ limit, for $k=l=3$ (left panel) and $k=l=4$ (right panel). The binary entropy function is also plotted for comparison (the RS entropy coincides with it for $\theta \ge {\theta_{\rm r}}$). The physical part of the complexity extends on a small range of $\theta$, on which it is only slightly smaller than the RS entropy, the inset allows to see this small difference at the end of the domain of definition of $\Sigma_{\rm e}$.}
\label{fig_Tinfty_klequal}
\end{figure}
Let us mention at this point that ${\theta_{\rm min}}(T)$, the minimal density of initial configuration percolating within $T$ steps of the dynamics, reaches its asymptotic value ${\theta_{\rm min}}$ as $T\to\infty$ with different finite $T$ corrections in the various cases listed above. The analysis of App.~\ref{sec_app_Tinfty_klequal} shows that for $k=l=2$ (resp. $k=l=3$) these corrections are of order $2^{-T}$ (resp. $3^{-T}$), which is in agreement with a numerical fit of the data in Table \ref{tab_k2l2} (resp. Table \ref{tab_k3l3}). On the contrary for $k=l\ge 4$ these corrections are only polynomially small in $T$.
Finally, we could also compute analytically the distribution of activation times, within the RS formalism, for the initial configurations with a non-trivial density $\theta$ of active vertices in the interval $[{\theta_{\rm r}}/2,{\theta_{\rm r}}]$. Their cumulative distribution function $P_t=\eta(t_i\le t)$ obtained from the marginals of the law (\ref{eq_eta_ust}) reads in the $T\to\infty$ limit with $t$ kept fixed:
\beq
P_{t+1}=\theta + \frac{(2\theta-{\theta_{\rm r}})(1-{\theta_{\rm r}})}{{\theta_{\rm r}}} w_t^{k+1} + (1-{\theta_{\rm r}})(k+1) w_t^k \left(\frac{\theta}{{\theta_{\rm r}}}-\frac{2\theta-{\theta_{\rm r}}}{{\theta_{\rm r}}} w_t \right) \ ,
\label{eq_Pt_klequal_Tinfty}
\end{equation}
where $w_t$ is a series defined recursively by
\beq
w_0 = {\theta_{\rm r}} \ , \qquad w_{t+1}={\theta_{\rm r}} + (1-{\theta_{\rm r}}) w_t^k \ .
\label{eq_wt}
\end{equation}
Examples of this cumulative distribution are displayed in Fig.~\ref{fig_k2l2_tact_analytic}. As explained above the predictions of the RS cavity method are not expected to be correct for $\theta < {\theta_{\rm c}}$; in the particular case $k=l=2$ we however expect this result to be true down to $\theta={\theta_{\rm min}}=1/4$. Note that $P_t$ goes to $1$ when $t\to\infty$, in other words in the limit $T\to\infty$ the support of the distribution of activation times does not scale with $T$ and remains of order 1. One can also check that when $\theta={\theta_{\rm r}}$, the prediction $P_t$ of (\ref{eq_Pt_klequal_Tinfty}) coincides, as it should, with the distribution of activation times for random initial conditions of density ${\theta_{\rm r}}$ given in Eq.~(\ref{eq_random_x}); to see this one can notice that $w_t$ is equal to the series ${\widetilde{x}}_t$ defined in Eq.~(\ref{eq_random_tx}) for the study of random initial conditions, when $k=l$ and $\theta={\theta_{\rm r}}$. At the lower limit of the interval of density, $\theta={\theta_{\rm r}}/2$, one obtains instead a simple expression,
\beq
P_{t+1} = \frac{{\theta_{\rm r}}}{2} + (k+1) \frac{1-{\theta_{\rm r}}}{2} w_t^k \ .
\label{eq_Pt_trsudue}
\end{equation}
A straightforward analysis of (\ref{eq_Pt_klequal_Tinfty},\ref{eq_wt}) reveals that for all $\theta < {\theta_{\rm r}}$ the cumulative distribution $P_t$ reaches 1 with corrections of order $1/t$, in other words the probability $P_t-P_{t-1}$ that a vertex activates precisely at time $t$ has a power-law tail with exponent $-2$. On the contrary the random initial conditions of density ${\theta_{\rm r}}$ have $1-P_t$ of order $1/t^2$, hence the exponent of the tail is $-3$; random initial conditions with $\theta>{\theta_{\rm r}}$ have instead an exponentially decaying tail for their distribution of activation times.
\begin{figure}
\includegraphics[width=8cm]{fig_k2l2_tact.eps}
\caption{The integrated distribution of activation times (\ref{eq_Pt_klequal_Tinfty}) for percolating initial conditions of density $\theta \in [{\theta_{\rm r}}/2,{\theta_{\rm r}}]$. The curves are presented in the case $k=l=2$.}
\label{fig_k2l2_tact_analytic}
\end{figure}
\subsubsection{The case $k>l$}
\label{sec_lmk_Tinfty}
We shall now turn to a description of the limit as $T\to\infty$ of the RS and energetic 1RSB results when $k>l$, with again the technical details relegated in the Appendix \ref{sec_app_Tinfty_lmk}.
The RS entropy $s(\theta)$ coincides with the binary entropy function for $\theta \ge {\theta_{\rm r}}$, for exactly the same reasons as explained above in the case $k=l$ (here and in the rest of this subsection we denote ${\theta_{\rm r}}$ the threshold ${\theta_{\rm r}}(k,l)$). The non-trivial part of $s(\theta)$ and $\Sigma_{\rm e}(\theta)$ are obtained in a parametric way, with unfortunately rather long expressions that we shall now progressively describe. We keep implicit below the dependency of all quantities on $k$ and $l$ when there is no risk of confusion.
This parametrization is given in terms of a real $\lambda$ in the range $]0,\lambda_{\rm r}]$, where this upper limit is expressed in terms of the threshold ${\theta_{\rm r}}$ for activation from a random initial condition as $\lambda_{\rm r}=(1-{\theta_{\rm r}})\theta_{\rm r}^{k-1}$. We need first to introduce some auxiliary functions ${\widehat{u}}(\lambda)$, ${\widehat{v}}(\lambda)$, $u_*(\lambda)$ and $v_*(\lambda)$. The first two are given explicitly as
\beq
{\widehat{u}}(\lambda) = \left(\frac{1-{\theta_{\rm r}}}{\lambda}\right)^{\frac{1}{k-1}} \ , \qquad
{\widehat{v}}(\lambda) = {\widetilde{x}_{\rm r}} \left(\frac{1-{\theta_{\rm r}}}{\lambda}\right)^{\frac{1}{k-1}} \ ,
\label{eq_uvhat}
\end{equation}
where we recall that ${\widetilde{x}_{\rm r}}$ is the fixed-point of Eq.~(\ref{eq_random_tx}) at the bifurcation ${\theta_{\rm r}}$, see also (\ref{eq_tandx_r}). The last one, $v_*(\lambda)$, is defined as the smallest positive solution of
\beq
v=1+\lambda \sum_{p=l}^k \binom{k}{p} \left(\lambda l \binom{k}{l} \right)^{-\frac{k-p}{k-l}} v^{\frac{p(k-1)-k(l-1)}{k-l}} \ , \qquad
\label{eq_vstar}
\end{equation}
then $u_*(\lambda)$ can be deduced as the solution of
\beq
1 = \lambda l \binom{k}{l} v_*(\lambda)^{l-1} (u_*(\lambda)-v_*(\lambda))^{k-l}
\qquad \text{with} \ \ u_*(\lambda)\ge v_*(\lambda) \ .
\label{eq_ustar}
\end{equation}
One can check that $u_*(\lambda) \ge {\widehat{u}}(\lambda) \ge {\widehat{v}}(\lambda) \ge v_*(\lambda)$ on the interval $\lambda \in ]0,\lambda_{\rm r}]$, and that $u_*={\widehat{u}}=1/{\theta_{\rm r}}$ and $v_*={\widehat{v}}={\widetilde{x}_{\rm r}}/{\theta_{\rm r}}$ in $\lambda=\lambda_{\rm r}$. We then define two functions ${F_{\rm site}}(\lambda)$ and ${F_{\rm edge}}(\lambda)$ through
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
{F_{\rm site}}(\lambda) &=& \frac{\lambda}{u_*}\left[
{\widehat{u}}^{k+1} + (k+1) \sum_{p=l}^k \binom{k}{p} \left[
\frac{l-1}{k-l} I_{p-1} - I_p \right]\right]\label{eq_Fsite_lmk} \\
{F_{\rm edge}}(\lambda) &=& \frac{1}{u_*}\left[({\widehat{u}}-{\widehat{v}})^2 + 2 u_* v_* - v_*^2 + 2 \lambda l \binom{k}{l} I_{l-1} \right] \label{eq_Fedge_lmk}
\eea
where for clarity we kept implicit the $\lambda$ dependency of ${\widehat{u}},{\widehat{v}},u_*$ and $v_*$, and we introduced
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
I_p &=& \left(\lambda l \binom{k}{l} \right)^{-\frac{k-p}{k-l}}
\int_{v_*}^{{\widehat{v}}} {\rm d} v \ v^{\frac{p(k-1)-k(l-1)}{k-l}} \ , \\
&=& \left(\lambda l \binom{k}{l} \right)^{-\frac{k-p}{k-l}}
\times
\begin{cases}
\ln \left(\frac{{\widehat{v}}}{v_*} \right) \qquad \text{if} \ p=l-1 \ \text{and} \ k=2l-1 \ , & \\
\frac{k-l}{(p+1)(k-1)-(k+1)(l-1)} \left( {\widehat{v}}^{\frac{(p+1)(k-1)-(k+1)(l-1)}{k-l}} - v_*^{\frac{(p+1)(k-1)-(k+1)(l-1)}{k-l}} \right) & \text{otherwise} \ .
\end{cases} \nonumber
\eea
We can finally give the parametric form of the RS entropy $s(\theta)$:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
s(\lambda) &=& \ln(1+{F_{\rm site}}(\lambda)) - \frac{k+1}{2} \ln \left(\frac{{F_{\rm edge}}(\lambda)}{u_*(\lambda)} \right) +\mu(\lambda) (1-\theta(\lambda)) \ , \nonumber \\
\theta(\lambda) &=& \frac{1}{1+{F_{\rm site}}(\lambda)} \ , \nonumber \\
\mu(\lambda) &=& - \ln(\lambda \, u_*(\lambda)^k)\ ,
\label{eq_lmk_Tinfty_RS}
\eea
where $\mu(\lambda)$ is the opposite of the derivative of $s(\theta)$ in the point $\theta(\lambda)$. Thanks to the values ${\widehat{u}},{\widehat{v}},u_*$ and $v_*$ assume in $\lambda_{\rm r}$ this curve joins the binary entropy function in ${\theta_{\rm r}}$ with a continuous slope.
Similarly the 1RSB entropic complexity $\Sigma_{\rm e}(\theta)$ is obtained parametrically as
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
\Sigma_{\rm e}(\lambda) &=& \ln \left(1+\left(1-\frac{1}{\lambda \, u_*(\lambda)^{k-1}} \right) {F_{\rm site}}(\lambda) \right)
-\frac{k+1}{2} \ln \left(\frac{1+ (\lambda \, u_*(\lambda)^{k-1} - 1)\, {F_{\rm edge}}(\lambda)}{\lambda \, u_*(\lambda)^k - u_*(\lambda) +1} \right) - y(\lambda) (1-\theta(\lambda))
\ , \nonumber \\
\theta(\lambda) &=&
\frac{1-\frac{1}{\lambda \, u_*(\lambda)^k} {F_{\rm site}}(\lambda)}{1 + \left(1 - \frac{1}{\lambda \, u_*(\lambda)^{k-1}} \right) {F_{\rm site}}(\lambda)}
- \frac{k+1}{2} \frac{1-\frac{1}{u_*(\lambda)} {F_{\rm edge}}(\lambda)}{1+(\lambda\, u_*(\lambda)^{k-1}-1) {F_{\rm edge}}(\lambda)}
\ , \nonumber \\
y(\lambda) &=& \ln(\lambda \, u_*(\lambda)^k - u_*(\lambda) +1 ) \ ,
\label{eq_lmk_Tinfty_1RSB}
\eea
with $y(\lambda)$ giving the slope of the tangent of $\Sigma_{\rm e}(\theta)$ in the point $\theta(\lambda)$.
An example of the limit for the RS entropy can be found in the right panel of Fig.~\ref{fig_RS_k2l2_softheta} for $k=3$, $l=2$, along with some finite $T$ curves, and a similar plot for the energetic complexity is displayed in the right panel of Fig.~\ref{fig_1RSBy_k2l2_Sigma}. The entropy and energetic complexity for this case in the limit are compared in Fig.~\ref{fig_Tinf_k3l2}. The values ${\theta_{\rm min,0}}$ and ${\theta_{\rm min,1}}$ where $s(\theta)$ and $\Sigma_{\rm e}(\theta)$ vanish are easily determined numerically from the above representation, and are collected in Table~\ref{table_res_Tinfty} for various values of $k$ and $l$. For most of the cases one finds ${\theta_{\rm min,1}}$ to be slightly larger than ${\theta_{\rm min,0}}$; as explained above the exactness of this 1RSB prediction has still to be assessed from a computation of the stability with respect to further replica symmetry breaking.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig_k3l2_Tinf_sandsigmaeoftheta.eps}
\end{center}
\caption{The RS entropy $s(\theta)$ and energetic 1RSB complexity $\Sigma_{\rm e}(\theta)$ in the $T\to \infty$ limit, for $k=3$, $l=2$, from the analytical formulas given in (\ref{eq_lmk_Tinfty_RS},\ref{eq_lmk_Tinfty_1RSB}).}
\label{fig_Tinf_k3l2}
\end{figure}
\begin{table}
\begin{tabular}{|c|c|c||c|c||c|c|}
\hline
$k$ & $l$ & ${\theta_{\rm r}}$ & $\mu_{s=0}$ & ${\theta_{\rm min,0}}$ & $y_{\rm s}$ & ${\theta_{\rm min,1}}$ \\
\hline
$2$ & $2$ & $\frac{1}{2}$ & $-\infty$ & $\frac{1}{4}$ & $\infty$ &$\frac{1}{4}$ \\
\hline
\hline
$3$ & $2$ & $0.111111$ & -12.720727 & 0.046283 & 12.547960 & 0.046328 \\
\hline
$3$ & $3$ & $\frac{2}{3}$ & $-\infty$ & $\frac{1}{3}$ & $\infty$ & $\frac{1}{3}$ \\
\hline
\hline
4 & 2 & 0.050781 & -9.633812 & 0.013108 & 9.125975 & 0.013258\\
\hline
4 & 3 & 0.275158 & $-\infty$ & $\frac{1}{6}$ & $\infty$ & $\frac{1}{6}$ \\
\hline
$4$ & $4$ & $\frac{3}{4}$ & -14.904539 & 0.378463 & 14.883293 & 0.378465 \\
\hline
\hline
5 & 2 & 0.029096 & -9.499859 & 0.005715 & 8.891066 & 0.005820 \\
\hline
5 & 3 & 0.165116 & -12.395257 & 0.076228 & 12.333754 & 0.076247 \\
\hline
5 & 4 & 0.397212 & $-\infty$ & $\frac{1}{4}$ & $\infty$ & $\frac{1}{4}$ \\
\hline
$5$ & $5$ & $\frac{4}{5}$ & -9.786306 & 0.422619 & 9.647302 & 0.422695 \\
\hline
\hline
6 & 2 & 0.018854 & -9.675930 & 0.003098 & 9.026488 & 0.003166 \\
\hline
6 & 3 & 0.112870 & -10.396651 & 0.042825 & 10.234248 & 0.042894 \\
\hline
6 & 4 & 0.269022 & -16.484079 & 0.150054 & 16.480311 & 0.150055 \\
\hline
6 & 5 & 0.486312 & -40.532392 & 0.300090 & 40.532392 & 0.300090 \\
\hline
$6$ & $6$ & $\frac{5}{6}$ & -8.403727 & 0.460014 & 8.191036 & 0.460228 \\
\hline
\end{tabular}
\caption{The predictions of the RS and energetic 1RSB cavity method in the $T\to \infty$ limit.}
\label{table_res_Tinfty}
\end{table}
There are however two special cases which stand on a different footing, namely $(k,l)=(4,3)$ and $(k,l)=(5,4)$. Indeed in these two cases one has the same phenomenology than for $k=l=3$, namely a coincidence of ${\theta_{\rm min,0}}$ and ${\theta_{\rm min,1}}$ due to a vertical segment in the curves $s(\theta)$ and $\Sigma_{\rm e}(\theta)$ extending to positive values. This phenomenon can be understood by studying the limit $\lambda \to 0$ of the above representation of these curves. After some algebra one finds indeed that for $k<2l-1$,
\beq
\underset{\lambda \to 0}{\lim} \ \theta(\lambda) = \frac{2l-k-1}{2l} \ , \qquad
\underset{\lambda \to 0}{\lim} \ s(\lambda)
=\underset{\lambda \to 0}{\lim}\ \Sigma_{\rm e}(\lambda) =
\frac{k+1}{2l} \ln \left( \frac{l^l}{(l-1)^{l-1}} \binom{k}{l} \right)
- \frac{k-1}{2} \ln \left( \frac{2l}{2l-k-1}\right) \ ,
\end{equation}
the limiting value for $\theta$ being valid both for the RS (\ref{eq_lmk_Tinfty_RS}) and 1RSB (\ref{eq_lmk_Tinfty_1RSB}) expressions. It turns out that for $k=4$, $l=3$ and $k=5,l=4$, the latter expression for the entropy $s$ and complexity $\Sigma_{\rm e}$ is strictly positive, hence the simple predictions $1/6$ and $1/4$ for ${\theta_{\rm min}}$ in these two cases respectively, that saturate the lowerbound of (\ref{eq_bound_kl}). We did not find any other values of $k,l$ that produce the same phenomenon.
Finally the distribution of activation times in the RS formalism exhibits a very different pattern with respect to the case $k=l$ (see Fig.~\ref{fig_k3l2_tact} for an illustration). As a matter of fact, in the limit $T\to\infty$ the activation times $t$ of the vertices have to be divided in three categories, each of them comprising a finite fraction of the $N$ vertices: (i) $t=O(1)$ (ii) $t=O(T)$ (iii) $t=T-O(1)$. The category (ii) of vertices can be described by a scaling function for the cumulative distribution, $P(s)=P_{t=sT}$, with $s\in]0,1[$ a reduced time. One has $P(s=0^+)>0$ and $1-P(s=1^-)>0$, these two numbers representing the fractions of vertices of type (i) and (iii) respectively. They can be computed following the techniques of the Appendix \ref{sec_app_Tinfty_lmk}, yielding for initial configurations with a fraction $\theta(\lambda) < {\theta_{\rm r}}$ of active vertices:
\begin{eqnarray}} \newcommand{\eea}{\end{eqnarray}
P(s=0^+)&=& \theta + \theta \frac{\lambda}{u_*} \sum_{p=l}^{k+1} \binom{k+1}{p} v_*^p\, (u_*-v_*)^{k+1-p} \ , \nonumber \\
1-P(s=1^-) &=& \theta \frac{\lambda}{u_*} \left(\frac{1-{\theta_{\rm r}}}{\lambda} \right)^{\frac{k+1}{k-1}}\left[1-\sum_{p=l}^{k+1} \binom{k+1}{p} {\widetilde{x}_{\rm r}}^p \, (1-{\widetilde{x}_{\rm r}})^{k+1-p} \right] \ .
\label{eq_Pslimits}
\eea
\begin{figure}
\includegraphics[width=8cm]{fig_k3l2_tact.eps}
\caption{An example of the cumulative distribution of activation times for $k=3$, $l=2$, obtained with the numerical resolution of the RS equations for a large but finite value of $T=400$, with a parameter $\lambda=0.005$, corresponding to an initial density of active sites of $0.089$. The two horizontal lines corresponds to $P(s=0^+)$ and $P(s=1^-)$ from Eq.~(\ref{eq_Pslimits}), delimiting the fraction of vertices that activate within a finite time after the beginning of the process (resp. before its end).}
\label{fig_k3l2_tact}
\end{figure}
\section{Algorithmic results}
\label{sec_single_sample}
We shall present in this Section the results of numerical experiments performed on finite size random regular graphs, for which we have constructed explicitly some activating initial configurations. We have used two strategies to do so, one based on a simple greedy heuristic, the other inspired by the results of the cavity method. Both of them build iteratively a percolating initial configuration, starting from the configuration with all vertices inactive, and adding one active vertex at a time (another route would be to start from the all active configuration and sequentially reduce the number of active vertices, but we did not investigate this alternative strategy). We shall denote $\tau$ the number of addition steps performed by the algorithm, and ${\underline{\sigma}}(\tau)$ the initial configuration considered at this point (that contains by definition $\tau$ active vertices). The configuration denoted ${\underline{\sigma}}^T(\tau)$ (resp. ${\underline{\sigma}}^{\rm f}(\tau)$) is thus the configuration obtained after $T$ (resp. an infinite) number of steps of the dynamics defined in (\ref{eq_dynrules}) from the initial configuration ${\underline{\sigma}}(\tau)$; we will denote $|{\underline{\sigma}}^T(\tau)|$ the number of active vertices in this configuration. The algorithm stops when this number reaches $N$, as ${\underline{\sigma}}(\tau)$ is then the first percolating initial configurations encountered. The difference in the two algorithms to be presented below lies in the rule used to choose which additional active vertex to add in the initial configuration in a step $\tau \to \tau +1$.
\subsection{A greedy algorithm}
\label{sec_single_sample_greedy}
Let us first consider the case of a finite time horizon $T$, i.e. the problem of finding an initial configuration ${\underline{\sigma}}$ with ${\underline{\sigma}}^T$ the fully active configuration and ${\underline{\sigma}}$ containing the smallest possible number of active vertices. The simplest strategy is to choose at each time step $\tau\to\tau+1$ the inactive vertex of ${\underline{\sigma}}(\tau)$ whose activation leads to the largest possible value of $|{\underline{\sigma}}^T(\tau+1)|$, and stop at the first time $\tau$ such that ${\underline{\sigma}}^T(\tau)$ is the fully active configuration. This can be immediately generalized to the case $T=\infty$ by including at each time step the vertex whose activation increases most $|{\underline{\sigma}}^{\rm f}(\tau+1)|$; this version of the greedy procedure was actually a tool in the rigorous bounds on ${\theta_{\rm min}}$ for graphs with good expansion properties of~\cite{bounds_Amin}. If several vertices lead to the same increase the ties can be broken arbitrarily. The time complexity of the greedy algorithm is a priori cubic in the number $N$ of vertices: a linear number of steps $\tau\to\tau +1$ have to be performed before finding a percolating initial configuration. For each of these steps a number of order $N$ of candidate new configurations ${\underline{\sigma}}(\tau+1)$ have to be considered, the computation of ${\underline{\sigma}}^T(\tau+1)$ requiring itself a linear number of operations for each configuration. It is however easy to reduce significantly this complexity when $T=\infty$. As explained at the end of Sec.~\ref{sec_def_dyn} , in this case the final configuration of the dynamical process can be obtained sequentially, regardless of the order of the activations. By monotonicity the configuration ${\underline{\sigma}}^{\rm f}(\tau+1)$ can be computed by adding one active vertex to ${\underline{\sigma}}^{\rm f}(\tau)$ (instead of ${\underline{\sigma}}(\tau)$) and determining the number (of order 1) of additional activations that can be triggered by this addition. This reduces the total complexity to a quadratic scaling with $N$.
In Fig.~\ref{fig_greedy} we plot the fraction of active vertices in the configuration ${\underline{\sigma}}^T(\tau)$ as a function of the density $\tau/N$ of the active vertices in the initial configuration obtained after $\tau$ steps of this greedy procedure; when the curve reaches 1 we have thus obtained an initial configuration that percolates within $T$ steps (note that the part of the curve for smaller $\tau$ corresponds to the alternative optimization problem labelled (i) in the introduction). The density of the contagious sets reached in this way are summarized in Table~\ref{tab_single_sample_finiteT}; as expected these densities are strictly greater than the prediction ${\theta_{\rm min,1}}$ of the 1RSB cavity method, and also than the ones reached by more involved message-passing algorithms (see the discussion in next subsection).
\begin{figure}
\includegraphics[width=8cm]{fig_greedy_k2.eps}
\hspace{1cm}
\includegraphics[width=8cm]{fig_greedy_k3.eps}
\caption{The density of active vertices in the configuration ${\underline{\sigma}}^T(\tau)$ after $\tau$ steps of the greedy algorithm, for $k=l=2$ (left panel) and $k=3$, $l=2$ (right panel). Each curve corresponds to a single run of the algorithm on a graph of $N=10^4$ vertices.}
\label{fig_greedy}
\end{figure}
\begin{table}
\begin{tabular}{|c|c|c|c|c||c|c|c|c|}
\hline
& \multicolumn{4}{c||}{$k=l=2$} & \multicolumn{4}{c|}{$k=3$, $l=2$} \\
\hline
$T$ & ${\theta_{\rm min,1}}$ & $\theta_{\rm sp}$ & $\theta_{\rm maxsum}$~\cite{private} & $\theta_{\rm greedy}$ & ${\theta_{\rm min,1}}$ & $\theta_{\rm sp}$ & $\theta_{\rm maxsum}$~\cite{private} & $\theta_{\rm greedy}$ \\
\hline
1 & 0.424257 & 0.426 & 0.427 & 0.482 & 0.363813 & 0.366 & 0.370 & 0.426 \\
\hline
2 & 0.325882 & 0.328 & 0.330 &0.376 & 0.237009 & 0.240 & 0.243 & 0.291 \\
\hline
3 & 0.289097 & 0.291& 0.293 &0.335 & 0.182338 & 0.185 & 0.190 & 0.233 \\
\hline
4 & 0.271564 & 0.273 & 0.275 &0.311 & 0.151693 & 0.156 & 0.164 & 0.197 \\
\hline
5 & 0.262167 & 0.263 & 0.266 &0.296 & 0.132036 & 0.142 & 0.146 & 0.174 \\
\hline
7 & 0.253779 & & 0.257 &0.278 & 0.108251 & 0.127 & 0.125 & 0.144 \\
\hline
10 & 0.250553 & & 0.251 &0.265 & 0.089425 & & 0.108 &0.119 \\
\hline
\end{tabular}
\caption{The density of (finite time) contagious sets reached by the greedy and message-passing algorithms, compared to the predictions of the cavity method for their minimal size. The data for the algorithmic results correspond to averages over ten graphs of size $N=10^4$.}
\label{tab_single_sample_finiteT}
\end{table}
One can clearly see a qualitative difference between the cases $k=l$ and $k>l$ in the two panels of Fig.~\ref{fig_greedy}: in the latter case as $T$ gets larger the last active vertices added in the initial configuration before finding a percolating one provoke a very steep increase in the final size of the activated set.
As said above the greedy procedure can easily be generalized to $T=\infty$; the density of the smallest contagious sets constructed in this way are presented in Table~\ref{tab_single_sample_Tinfty} for various values of $k$ and $l$. As these results demonstrate the greedy algorithm is able, in all cases we investigated, to find contagious sets with a density strictly smaller than ${\theta_{\rm r}}$, the density above which typical uncorrelated configurations are percolating. However in general the density reached by this simple procedure is strictly greater than the prediction ${\theta_{\rm min,1}}$ of the cavity method for their minimal size; this is in agreement with the interpretation of the replica symmetry breaking creating metastable states that trap simple local search procedures and prevent them from reaching global optima of the cost function landscape in which the search moves. The only exception is the case $k=l=2$, for which the minimal density $1/4$ (corresponding to the decycling number of 3-regular random graphs~\cite{decycling}) is actually reached by the greedy procedure; this result is in line with the analysis of Sec.~\ref{sec_kequall_Tinfty}, which revealed a disappearance of the RSB phase in the large $T$ limit for this peculiar case.
\begin{table}
\begin{tabular}{|c|c|c|c|c|}
\hline
$k$ & $l$ & ${\theta_{\rm r}}$ & ${\theta_{\rm min,1}}$ & $\theta_{\rm greedy}$ \\
\hline
2 & 2 & $\frac{1}{2}$ & $\frac{1}{4}$ & 0.250 \\
\hline
3 & 2 & 0.111111 & 0.046328 & 0.070 \\
\hline
3 & 3 & $\frac{2}{3}$ & $\frac{1}{3}$ & 0.387 \\
\hline
4 & 4 & $\frac{3}{4}$ & 0.378465 & 0.482 \\
\hline
5 & 5 & $\frac{4}{5}$ & 0.422695 & 0.551 \\
\hline
\end{tabular}
\caption{The density of (infinite time) contagious sets reached by the greedy algorithm, compared to the predictions of the cavity method. The algorithm was run on ten graphs of size $N=10^4$, the last column is the average over these repetitions. Experiments with graphs of different sizes revealed a very clear $1/N$ dependency of the finite-size corrections of $\theta_{\rm greedy}$ in the cases with $k=l$. We could not get such a clear dependency when $k>l$, slower finite-size corrections might be at play in these cases.}
\label{tab_single_sample_Tinfty}
\end{table}
Further information on the minimal contagious sets produced by the greedy algorithm with $T=\infty$ can be obtained from the distribution of the activation times of the vertices they induce, which are plotted in Fig.~\ref{fig_tact_compa}. Of course as the graphs under study are finite the support of these distributions is bounded; in all cases we investigated we found that the time to reach total activation from these initial configurations scales logarithmically with the number of vertices of the graph (see also Fig.~\ref{fig_tact_compa} for a comparison between two different sizes of the graph). The qualitative difference between the cases $k=l$ and $k>l$ expected from the discussion of the $T\to\infty$ limit of Sec.~\ref{sec_res_largeT} is indeed apparent on these curves; in the latter case a finite fraction of the vertices are activated at the very end of the dynamical process. However the activation time distributions induced by the configurations produced by the greedy algorithm are not in quantitative agreement with the RS analytical predictions (with a value of $T$ and $\theta$ chosen to fit the numerical ones). A possible explanation for this discrepancy is that the greedy algorithm is a very ``out-of-equilibrium'' algorithm, hence the configurations it reaches are not the typical ones of the ``equilibrium'' measure (\ref{eq_eta_us}).
\begin{figure}
\includegraphics[width=8cm]{fig_k2l2_tact_compa.eps}
\hspace{1cm}
\includegraphics[width=8cm]{fig_k3l2_tact_compa.eps}
\caption{The ``numerical'' curves represent the distribution of activation times for the least dense activating initial configurations found by the greedy algorithm for $T=\infty$, for $k=l=2$ (left panel) and $k=3$, $l=2$ (right panel). In both cases the graph studied contained $N=8 \cdot 10^4$ vertices, in the left panel the complete activation is reached in 93 steps, in the right one it takes 367 steps. For comparison in the left panel the analytical prediction is plotted both for $T=\infty$ (see Eq.~(\ref{eq_Pt_trsudue})) and for $T=93$, in the right panel the analytical curve corresponds to $T=367$.}
\label{fig_tact_compa}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{fig_k3l2_tact_num.eps}
\caption{The distribution of activation times for the least dense activating initial configurations found by the greedy algorithm for $T=\infty$, for $k=3$, $l=2$, and two different sizes $N$ of the graph. For $N=4 \cdot 10^4$ the complete activation took place after $T=248$ steps, while for $N=8 \cdot 10^4$ it occured at $T=367$.}
\label{fig_tact_num}
\end{figure}
\subsection{Survey propagation}
\label{sec_single_sample_mp}
The second algorithmic procedure we investigated is based on the insight provided by the statistical mechanics analysis on the structure of the configuration space of the problem; it corresponds indeed to the Survey Propagation algorithm introduced in~\cite{MeZe} for the analysis of random satisfiability problem (and more precisely to its variant introduced in~\cite{SPy} for the energy minimization in the unsatisfiable phase of such problems). An idealized thought experiment for the construction of minimal contagious sets would be to sequentially assign the values of the ${\sigma}_i$ according to their marginal probabilities in the law (\ref{eq_eta_us}), with $\epsilon=+\infty$ and $\mu=-\infty$; the exact determination of such marginals is in general a very hard computational tasks, and in practice one has to content oneself with approximations provided for instance by message passing procedures. This is the road we have followed here, by implementing the single-sample energetic 1RSB equations (\ref{eq_G_1RSBy}), i.e. assigning to each directed edge $i\to j$ of the graph under study a vector $P_{i\to j}$ of $2T$ probabilities. At each step $\tau$ of the algorithm the equations (\ref{eq_G_1RSBy}) are iterated several times to look for a global solution of these equations; the presence of $\tau$ active (decimated) vertices in the current configuration ${\underline{\sigma}}(\tau)$ is implemented as a boundary condition in these equations, easily seen to be $P_{i\to j}(h)=\delta(h-B_0)$ for the outgoing messages from an activated vertex $i$. The information contained in such a solution of the 1RSB equations can be a priori exploited in several ways; we chose to compute, for each vertex $i$ not yet activated, the quantity
\beq
W_i = 1 - \frac{\partial}{\partial y}\ln {{\cal Z}_{\rm site}}(\{ P_{j\to i} \}_{j \in {\partial i}} )+ \frac{1}{2} \sum_{j \in {\partial i}} \frac{\partial}{\partial y} \ln {{\cal Z}_{\rm edge}}(P_{i \to j},P_{j\to i}) \ ,
\end{equation}
i.e. the contribution of the site $i$ to the derivative of the potential $\Phi_{\rm e}$ given in Eq.~(\ref{eq_Phie_single_sample}). This number measures indeed the tendency of $i$ to be active in all configurations belonging to the clusters considered in the energetic 1RSB formalism. Accordingly we choose the vertex $i$ with the largest value of $W_i$ to be the new active vertice to be added to ${\underline{\sigma}}(\tau)$ in order to form ${\underline{\sigma}}(\tau+1)$. For simplicity we fixed the value of $y$ in the whole procedure to the value $y_{\rm s}$ determined analytically, that leads to a vanishing complexity before the decimation; we also tried to recompute this value of $y$ during the course of the decimation but did not obtain significant improvement of the performances in the cases considered.
The values of the density of the percolating initial configurations we managed to construct in this way are presented in Table \ref{tab_single_sample_finiteT} for the two cases $k=l=2$ and $k=3$, $l=2$, for several (relatively small) values of $T$. The results are better than the simple greedy algorithm, and in most of the cases also than the maxsum replica-symmetric algorithm~\cite{Torino1,Torino2,private}, but in some cases deviate significantly from the prediction ${\theta_{\rm min,1}}$ for the density of minimal contagious sets. An analytical understanding of the performances of such decimation procedures is actually a challenging open problem (see~\cite{RiSe09,Coja11} for partial results in the simpler case of the Belief-Propagation guided decimation). We did not study much larger values of $T$ because we faced in this case convergence issues for the iterations of the equations (\ref{eq_G_1RSBy}), that a simple damping did not seem to alleviate efficiently. A pragmatic, even if not completely satisfactory, position we adopted for the results at $T\ge 4$ for the case $k=3$, $l=2$, was to ignore somehow the convergence problems, stopping the iterations of (\ref{eq_G_1RSBy}) after a time fixed beforehand, and computing the value of $W_i$ from these unconverged messages. As Table \ref{tab_single_sample_finiteT} demonstrates this attitude is not unreasonable as the densities reached are still better than the one of the greedy algorithm (yet can get worse than the maxsum procedure~\cite{Torino1,Torino2,private}).
\section{Conclusions and perspectives}
\label{sec_conclu}
In this paper we have continued the study initiated in~\cite{Torino1,Torino2} of the minimal contagious sets for the bootstrap percolation (or threshold model) dynamics on random graphs. We have shown the importance of taking into account the phenomenon of replica symmetry breaking in the determination of the minimal density ${\theta_{\rm min}}$ of active vertices in percolating initial conditions, and could simplify analytically the equations determining ${\theta_{\rm min}}$ in the limit $T\to\infty$ where the constraint on the time to reach a complete activation of the graph disappears. Reformulating the problem as the minimal number of vertices to be removed in a graph in order to destroy some specific subgraphs (its cycles or more generically its $q$-core) we recovered a previously known result for the decycling number of 3-regular random graphs~\cite{decycling} as well as a conjecture for 4-regular ones~\cite{decycling}, and proposed new quantitative conjectures for the sizes of the minimal ``de-coring'' sets for all pairs of degree of the graph and minimal degree of the targeted core. These take a particularly simple rational form for the removal of the 3-core in 5- and 6- regular random graphs.
Let us sketch now some possible directions for future study. A first project would be to test the stability of the 1RSB ansatz we used to compute ${\theta_{\rm min,1}}$, to assess for which values of $(k,l)$ this number should be expected to be the exact value ${\theta_{\rm min}}$ and not only a lowerbound. This computation should be doable following the techniques of~\cite{stab1,stab2,bm} for all finite $T$, and might even be simplified in the large $T$ limit. By analogy with the independent set problem which is a marginal case of the problem investigated here one could surmise to find that the 1RSB ansatz is stable for large enough values of the degree $k$ (and maybe also of the threshold $l$). This is also the regime where one can hope to see a mathematically rigorous proof of these predictions, as recently obtained for the independent sets in~\cite{is_Sly}. Asymptotic expansions of ${\theta_{\rm min,0}}(k,l)$ and ${\theta_{\rm min,1}}(k,l)$ in the large $k$ limit for $k>l$ should also be performed, considering either $l$ fixed in this limit, $l$ proportional to $k$, or $k-l$ fixed.
For the sake of concreteness and simplicity we presented explicit results only for regular random graphs, however we gave the intermediate equations of the RS and 1RSB cavity method under a form that can be directly applied to any sparse random graph ensembles with arbitrary prescribed degree distribution, and possibly fluctuating thresholds for activation. The latter could naturally be correlated with the degree of the vertices, triggering for instance the activation if the fraction of active neighbours reaches some fixed proportion (instead of a fixed number). It would be interesting to see how the results presented here are qualitatively modified by the local fluctuations in the graph structure, which would be particularly severe in the case of power-law tails in the degree distribution.
We also concentrated exclusively in this paper on the problem of optimizing the number of initially active vertices, imposing that all vertices are active at a later time. The variant of this problem where one puts a constraint on the maximal number of active vertices allowed in the initial configuration and try to maximize the level of activation at a later time is also relevant, in particular for applications to real-world situations. At the RS level we have sketched how to do this by controlling the parameter $\epsilon$ (the cost to be paid for finally inactive vertices) that we kept arbitrary in the first steps of the computations, a systematic study and the inclusion of the effects of replica symmetry breaking remains to be done.
Finally we believe that the message passing procedure inspired by the energetic 1RSB equations presented in Sec.~\ref{sec_single_sample_mp} would be worth investigated further. One should try to study (and cure) the convergence issues that arise for larger values of $T$, maybe changing the way the information provided by the messages is used. One could in particular exploit them in a softer way by implementing a reinforcement technique~\cite{Torino1,Torino2} instead of a direct decimation. A more extensive comparison with the maxsum message passing procedure studied in~\cite{Torino1,Torino2} could also be interesting.
\acknowledgments
We warmly thank
Fabrizio Altarelli,
Victor Bapst,
Alfredo Braunstein,
Amin Coja-Oghlan,
Luca Dall'Asta,
Svante Janson,
Marc Lelarge and
Riccardo Zecchina
for useful discussions, and in particular FA, AB, LDA and RZ for sharing with us the unpublished numerical results~\cite{private} on their maxsum algorithm, and SJ for a useful correspondence and for pointing out the reference \cite{decycling}.
The authors acknowledge the support of the French Agence Nationale de la Recherche (ANR) under reference ANR-11-JS02-005-01 (GAP project) and of the People
Programme (Marie Curie Actions) of the European Union's Seventh Framework Programme
FP7/2007-2013/ under REA grant agreement no 290038.
|
\section{Introduction}
The main theme in the dynamics of linear operators is the notion of
hypercyclicity which plays an important role in the study of the
invariant subset problem in Banach spaces. This notion was initiated
by S. Rolewicz \cite{Rolewicz} in the setting of infinite
dimensional Banach spaces in 1969, though the examples of
translation and differential operators on the space of entire
functions equipped with the compact-open topology were known to be
hypercyclic in an earlier work of G. D. Birkhoff \cite{Birkhoff} and
G. R. MacLane \cite{MacLane}. Now a vast literature dealing with
hypercyclicity of operators as well as other related notions in
linear dynamics is available in \cite{Bayart Matheron},
\cite{Erdmann Peris2}, and \cite{Erdmann}.
In 2006, F. Bayart and S. Grivaux \cite{Bayart Grivaux} further
strengthened this concept to frequent hypercyclicity, which
quantifies the frequency with which the iterates of a given linear
operator at a point visit each non-empty open set. After the
appearance of this work, several results on frequently hypercyclic
operators have been established, for instance one may refer to
\cite{Blasco Bonilla Erdmann}, \cite{Bonilla Erdmann1},
\cite{Bonilla Erdmann2}, \cite{Erdmann Peris1} and \cite{Shkarin}.
In this paper we introduce $q$-frequent hypercyclicity which lies
between hypercyclicity and frequent hypercyclicity, where $q$ is a
fixed natural number. The case $q=1$ coincides with frequent
hypercyclicity. We prove a sufficient criterion for a continuous
linear operator to be $q$-frequently hypercyclic on a locally convex
space and give applications to obtain $q$-frequently hypercyclic
operators with respect to the norm on Banach spaces, the $F$-norm on
$F$-spaces and the weak*-topology on dual of Banach spaces. We also
provide examples of hypercyclic operators that are not
$q$-frequently hypercyclic for any $q\in \mathbb{N}$.
\section{Preliminaries}
Let $X$ be a separable topological vector space, and
$\mathcal{L}(X)$ denote the space of all continuous linear operators
on $X$. An operator $T\in \mathcal{L}(X)$ is said to be
\emph{hypercyclic} if there exists a vector $x\in X$ such that the
orbit $\{T^n(x):n\geq 0\}$ is dense in $X$. Such a vector $x$ is
called a \emph{hypercyclic vector} for $T$. As mentioned in the
previous section, Birkhoff's translation operator
$T_a(f)(z)=f(z+a)$, for nonzero $a\in \mathbb{C}$ and MacLane's
differentiation operator $D(f)=f'$ on the space $H(\mathbb{C})$ are
hypercyclic. Also, Rolewicz proved the hypercyclicity of the
operator $\lambda B$ on $\ell^p$ or $c_0$ for $1\leq p <\infty$ and
$|\lambda|
> 1$, where $B$ is the unweighted backward shift defined by
$B(e_n)=e_{n-1}, n\geq1$, with $e_0=0$ and $e_n=\{0,0,..,1,0..\}$,
$1$ being placed at the $n$th coordinate. Generalizing this result,
H.N. Salas \cite{Salas} proved that the weighted shift $B_w$
associated to a weight sequence $(w_n)$ of positive reals, given by
$B_w(e_n)=w_ne_{n-1}, n\geq1$ is hypercyclic on $\ell^p$ or $c_0$ if
and only if $\limsup_{n\to \infty}(w_1w_2...w_n)=\infty$.
\\
For testing the hypercyclicity of a linear operator, a sufficient
criterion known as the hypercyclicity criterion, initially obtained
by Kitai \cite{Kitai}, has appeared in different forms and the one
which is given below is due to H. Petterson \cite{Petterson}. This
is useful even for linear operators defined on non-metrizable
topological vector spaces. For the definition of $F$-norm, we refer
to \cite{Bonilla Erdmann2}, p. 385.
\begin{theorem} [Hypercyclicity criterion] \label{thm2.1} Let $(X,\tau)$ be
a separable topological vector space. Suppose further that $X$
carries an $F$-norm $||.||$ with respect to which it is complete and
that $\|.\|$-topology is stronger than $\tau$. If $T$ is an operator
continuous with respect to the $F$-norm, $D\subset X$ a countable
$\tau$-dense set and $S_n:D\rightarrow X$ maps such that, for all
$x\in D$,
\begin{itemize}
\item[1.]$||T^{n}(x)||\rightarrow 0$ and $||S_n(x)||\rightarrow 0$ as $n\rightarrow \infty$; and
\item[2.]$T^nS_n(x)=x$, for each $n\in \mathbb{N}$,
\end{itemize}
then the operator $T$ is $\tau$-hypercyclic.
\end{theorem}
An operator $T\in \mathcal{L}(X)$ is called \emph{frequently
hypercyclic} if there exists an $x$, called a \emph{frequently
hypercyclic vector} for $T$ such that for every nonempty open set
$U$ in $X$, the set $\mathbf{N}(x,U)=\{ n \in \mathbb{N}: T^n(x) \in
U\}$ has positive lower density; where the \emph{lower density} of
a subset $A$ of $\mathbb{N}$, the set of natural numbers, is defined
as
\begin{equation}
\begin{gathered}
\underline{\text{dens}}(A) $ = $\displaystyle \liminf_{N\to \infty}
\dfrac{\text{card}\{n\in A: n\leq N\}}{N},
\end{gathered}
\end{equation}
the symbol $\text{card}(B)$ being used to denote the cardinality of
the set $B$.
Let us note that $\underline{\text{dens}}(B)\in [0,1]$ for any
subset $B$ of $\mathbb{N}$. Clearly, the lower density of any finite
set is zero and that of $\mathbb{N}$ is 1. If $A$ is a strictly
increasing sequence $(n_k)$, the lower density of $A$ is
characterized as \cite{Bonilla Erdmann2}
\begin{equation}
\underline{\text{dens}}(n_k) = \liminf_{k\to \infty}
\frac{k}{n_k}.
\end{equation}
Alternatively, an operator $T\in L(X)$ is frequently hypercyclic if
there is some $x\in X$ such that for every nonempty open subset $U$
of $X$, there exist a strictly increasing $(n_k)$ of natural numbers
and a constant $C>0$ such that
\begin{center}
$T^{n_k}(x)\in U$ $\text{and}$ $n_k \leq Ck,$ $\forall$ $k\in
\mathbb{N}.$
\end{center}
Analogous to the hypercyclicity criterion, we have the following
criterion, proved in \cite{Bayart Grivaux} and \cite{Bonilla
Erdmann2}.
\begin{theorem}[Frequent hypercyclicity criterion]\label{thm2.2}
Let $X$ be a separable $F$-space and $T\in L(X)$. If there exist a
dense subset $D\subset X$ and a map $S:D\rightarrow D$ such that
\begin{itemize}
\item[1.]$\sum T^n(x)$ and $\sum S^n(x)$ are unconditionally convergent
for each $x\in D$; and
\item[2.] $TS=I$, the identity on $D$,
\end{itemize}
then the operator $T$ is frequently hypercyclic.
\end{theorem}
The operators of Birkhoff and MacLane satisfy the above criterion,
cf. \cite{Bayart Grivaux} and \cite{Bayart Matheron}, and so they
are frequently hypercyclic. In fact, any continuous operator, except
a scalar multiple of the identity, that commutes with all
translations on $H(\mathbb{C})$, has been shown to be frequently
hypercyclic \cite {Bonilla Erdmann1}. We also recall the hypercyclic
comparison principle from \cite{Bayart Grivaux} and \cite{Bayart
Matheron}, which says how to transfer the hypercyclicity via a
linear quasi-conjugacy.
\begin{proposition}[Hypercyclic comparison principle]\label{prop2.3} Let $T$ and $S$
be continuous linear operators on two topological vector spaces $X$
and $Y$ respectively and $A:X\rightarrow Y$ be a continuous linear
map with dense range such that $SA=AT$. If $T$ is hypercyclic (or
frequently hypercyclic) on $X$, then $S$ is hypercyclic (or
frequently hypercyclic) on $Y$.
\end{proposition}
\section {$q$-Frequently hypercyclic operators
} We first introduce the $q$-lower density of a subset of natural
numbers, for $q\in \mathbb{N}$ and determine a useful
characterization.
\begin{definition}
Let $A \subset \mathbb{N}$ and $q\in \mathbb{N}$. The $q$-lower
density of $A$ is defined as
\begin{center}
$q$-$\underline{\text{dens}}(A) = \displaystyle \liminf_{N\to
\infty} \dfrac{\text{card}\{n\in A: n\leq N^q \}}{N}.$
\end{center}
\end{definition}
Let us note that the lower density of a set is always finite and
lies in the interval $[0,1]$, but the $q$-lower density can vary in
$[0,\infty]$ for $q\geq 2$. As in the case of the lower density, we
have the following.
\begin{proposition}\label{lm2}
Let $(n_k)$ be a strictly increasing sequence of natural numbers.
Then
\begin{enumerate}
\item $q$-$ \emph{\underline{dens}}(n_k) = \displaystyle \liminf_{k\to
\infty} \dfrac {k}{{n_k}^{1/q}}$.
\item $q$-$\emph{\underline{dens}}(n_k)>0$ if and only if there exists a
constant $C>0$ such that $n_k \leq Ck^q$ for all $k\in \mathbb{N}$.
\end{enumerate}
\end{proposition}
\begin{proof}
Fix a number $k\in \mathbb{N}$. For any $N\in \mathbb{N}$ with
$n_k\leq N^q<n_{k+1}$, we have that
\begin{center}
$p_N=\dfrac{\text{card}\{k\in \mathbb{N}: n_k\leq N^q \}}{N}=k/N$.
\end{center}
Thus the inequality
\begin{center}
$\dfrac {k}{{n_{k+1}}^{1/q}}< p_N \leq \dfrac {k}{{n_k}^{1/q}}$
\end{center}
implies the first part in the theorem.
Part $(2)$ follows immediately from the fact that $\liminf a_k>0$ if
and only if $\frac {1} {a_k} \leq C$ for some $C>0$, where $(a_k)$
is a sequence of positive numbers.
\end{proof}
We now define the notion of $q$-frequent hypercyclicity of linear
operators on topological vector spaces.
\begin{definition}
Let $q \in \mathbb{N}$. A continuous linear operator $T$ on a
separable topological vector space $X$ is said to be $q$-frequently
hypercyclic if there exists an $x \in X$ such that for any nonempty
open subset $U$ of $X$, the set $\mathbf{N}(x,U)=\{ n \in
\mathbb{N}: T^n(x) \in U\}$ has positive $q$-lower density. Such a
vector is called a $q$-frequently hypercyclic vector for $T$.
\end{definition}
Alternatively, a continuous linear operator $T$ on a separable
topological vector space $X$ is \emph{$q$-frequently hypercyclic}
if there exists an $x \in X$ such that for any nonempty open subset
$U$ of $X$, we can find a strictly increasing sequence $(n_k)$ of
natural numbers and a constant $C>0$ such that
\begin{center}
$T^{n_k}(x)\in U$ and $n_k \leq Ck^q,$ for all $k\in \mathbb{N}.$
\end{center}
Such a vector is called a \emph{$q$-frequently hypercyclic vector}
for $T$.
Obviously, every frequently hypercyclic operator is $q$-frequently
hypercyclic for any natural number $q$, and the two notions are the
same for the case $q=1$. Also this new property of linear operators
is stronger than hypercyclicity; however, none of the converse
implications is true, e.g. consider
\begin{example}
Here we show that there exists a hypercyclic operator on $\ell^1$
that is not 2-frequently hypercyclic with respect to the weak
topology of $\ell^1$. Indeed, the weighted backward shift $B_w$ with
weights $w_n=\sqrt{\frac{n+1}{n}}$ is hypercyclic by the result of
Salas, but was shown to be non-frequently hypercyclic on $\ell^2$ in
\cite{Bayart Grivaux}. For showing the non-2-frequent hypercyclicity
of $B_w$, choose the weakly open set $U=\{(y_n)\in
\ell^1:|y_1|>1\}$. Let $x=(x_n)$ be a $2$-frequently hypercyclic
vector for $B_w$. Enumerate the set $N(x,U)=\{ n \in \mathbb{N}:
B_w^n(x) \in U\}$ as $(n_k)$. Thus we have a constant $c>0$ such
that $n_k \leq ck^2$, and hence
\begin{center}
$\displaystyle \sum_{k\geq 1}\frac {1}{\sqrt{n_k}}=\infty$.
\end{center}
On the other hand, $B_w^{n_k}(x) \in U$ implies that
\begin{center}
$ \sqrt{n_k+1}$ $ |x_{n_k+1}|> 1$, for all $k\geq 1$
\end{center}As $(x_n)\in \ell^1$, we get
\begin{center}
$\sum_{k\geq 1}\frac {1}{\sqrt{n_k+1}} <\infty$,
\end{center}
which is a contradiction. Hence $B_w$ is not 2-frequently
hypercyclic on $\ell^1$ for the norm topology.
\end{example}
\begin{example}\label{example3.4} We show the existence of a hypercyclic operator that is not
$q$-frequently hypercyclic with respect to the weak topology, for
any $q\in \mathbb{N}$. Let us consider the unilateral shift $B_w$ on
$\ell^1$ with weights $w_k=\displaystyle \frac
{\ln(k+2)}{\ln(k+1)}$, $k\in \mathbb{N}$. By the result of
H.N.Salas, $B_w$ is hypercyclic since $w_1w_2...w_k= \frac
{\ln(k+2)}{\ln 2} \rightarrow \infty$ as $k \rightarrow \infty$. Let
$x=(x_n)$ be a $q$-frequently hypercyclic vector for $B_w$ for some
$q\in \mathbb{N}$. Enumerate the set $N(x,U)=\{ n \in \mathbb{N}:
B_w^n(x) \in U\}$ as $(n_k)$, where $U=\{(y_n)\in \ell^1:|y_1|>1\}$.
Thus we have a constant $c>0$ such that $n_k \leq ck^q$. This
implies that $\ln(n_k)\leq C \ln k$,
for some constant $C>0$ and for all $k\in \mathbb{N}$. Hence
\begin{center}
$\displaystyle \sum_{k\geq 1}\frac {1}{\ln{n_k}}=\infty$.
\end{center}
On the other hand, $B_w^{n_k}(x) \in U$ implies that
\begin{center}
$\displaystyle \frac {\ln({n_k}+2)}{\ln 2} |x_{n_k+1}|> 1$.
\end{center} Consequently,
\begin{center}
$\displaystyle \sum_{k\geq 1}\frac {1}{\ln{(n_k+2)}} <\infty$,
\end{center} which is a contradiction. Hence $B_w$ is not
$q$-frequently hypercyclic in the norm topology, for any $q\in
\mathbb{N}$.
\end{example}
We now prove a criterion, similar to the frequent hypercyclicity
criterion, which works even for operators defined on certain
non-metrizable locally convex spaces. Using this, we obtain a
2-frequently hypercyclic operator that is not frequently
hypercyclic. Before stating the result, let us recall that a series
$\sum {x_n}$ in a topological vector space is
\emph{unconditionally convergent} if $\sum {x_{\sigma (n)}}$ is
convergent for every permutation $\sigma$ of $\mathbb{N}$. In any
topological vector space, this mode of convergence is equivalent to
the unordered convergence of $\sum {x_n}$, cf. \cite{Kamthan Gupta},
p.154. Thus a series $\sum {x_n}$ is unconditionally convergent if
and only if for every non-empty open set $U$ of $0$, there
corresponds an $N\in \mathbb{N}$ such that $\sum _{n\in F}{x_n}\in
U$ for every finite set $F\subset [N,\infty)$. Also, for the proof
of our criterion we need the following lemma from \cite{Bayart
Matheron}.
\begin{lemma} \label{lem3.4} Let $(N_k)$ be a strictly increasing sequence of
natural numbers. Then there exists a pairwise disjoint sequence
$(J_k)$ of subsets of $\mathbb{N}$ such that
\begin{enumerate}
\item $\emph{\underline{dens}}(J_k)>0$ for each $k\geq 1$
\item $|n-m|\geq N_k+N_p$ for $n\not=m$ and $(n,m)\in
J_k\times J_p$.
\item $n\geq N_k$, for each $n\in J_k$ and $k\geq 1$.
\end{enumerate}
\end{lemma}
We now state and prove the main result of the paper.
\begin{theorem}[$q$-frequent hypercyclicity criterion]
\label{thm3.5} Let $(X,\tau)$ be a separable locally convex space
and $q\in \mathbb{N}$. Suppose that $X$ is equipped with an
$F$-norm $\|.\|$ such that the F-norm topology is stronger than
$\tau$ and $(X,\|.\|)$ is complete. If $T$ is an operator continuous
with respect to the $F$-norm, $D$ is a subset of $X$ containing a
countable $\tau$-dense subset and $S:D\rightarrow D$ is a map such
that
\begin{enumerate}
\item $\sum T^{n^q}(x)$ and $\sum S^{n^q}(x)$ are unconditionally
convergent with respect to the $F$-norm, for each $x\in D$; and
\item $TS=I$, the identity on $D$,
\end{enumerate}
then the operator $T$ is $q$-frequently hypercyclic with respect to
$\tau$.
\end{theorem}
\begin{proof}
Our proof is inspired by that of the frequent hypercyclicity
criterion given in \cite{Bayart Matheron}. However, we outline the
proof for the sake of completeness. Let $Y=\{x_1,x_2,... \}\subset
D$ be a countable $\tau$-dense set. Consider a summable sequence
$(\epsilon _k)$ of positive real numbers, which are to be chosen
later. By the hypothesis, corresponding to $\epsilon_k$, we can find
$N_k\in \mathbb{N}$ such that
\begin{equation}\label{eq1}
\left\|\sum_{n\in F} T^{n^q}(x_i)\right\|+ \left\| \sum_{n\in F}
S^{n^q}(x_i)\right\|<\epsilon_k, 1\leq i\leq k,
\end{equation}
for any finite set $F\subset [N_k,\infty)$ of natural numbers. We
may now assume that $(N_k)$ is strictly increasing so that by Lemma
\ref{lem3.4}, we get a sequence $(J_k)$ of subsets of $\mathbb{N}$
with the properties mentioned therein. We now set
\begin{center}
$\displaystyle x=\sum_{k\geq 1}\sum_{n\in J_k} S^{n^q}(x_k)$.
\end{center}
Since unconditional convergence implies subseries convergence
\cite{Kamthan Gupta} p.154, the series $\sum_{n\in J_k}
S^{n^q}(x_k)$ converges for each natural number $k$. It follows by
\eqref{eq1} that
\begin{center}
$\displaystyle \sum_{k=1}^{\infty}\left\|\sum_{n\in J_k} S^{n^q}(x_k)\right\| \leq
\sum_{k\geq 1}
{\epsilon_k}< \infty$
\end{center}
Thus $x\in X$.
Let us now fix $k\in \mathbb{N}$ and $m\in J_k$. Then
\begin{equation*}
\displaystyle T^{m^q}(x)=\sum_{l\geq 1}\sum_{n\in J_l}
T^{m^q}S^{n^q}(x_l).
\end{equation*} and so
\begin{equation}\label{eq2}
\displaystyle \left\|T^{m^q}(x)-x_k\right\|\leq \sum_{l\geq
1}\left\|\sum_{n\in J_l,m>n} T^{m^q-n^q}(x_l)\right\|+\displaystyle
\sum_{l\geq 1}\left\|\sum_{n\in J_l,m<n} S^{n^q-m^q}(x_l)\right\|.
\end{equation}
Let us now consider the first sum on the right hand side of
\eqref{eq2}. Indeed, writing the first term as
\begin{equation*}
\begin{gathered}
\displaystyle \sum_{l=1}^{k}\left\|\sum_{n\in J_l,m>n}
T^{m^q-n^q}(x_l)\right\| +\displaystyle \sum_{l\geq
k+1}\left\|\sum_{n\in J_l,m>n} T^{m^q-n^q}(x_l)\right\|,
\end{gathered}
\end{equation*}
we have
\begin{equation*}
\begin{gathered}
\displaystyle \sum_{l\geq 1}\left\|\sum_{n\in J_l,m>n}
T^{m^q-n^q}(x_l)\right\| \leq k\epsilon_k +\sum_{j\geq k+1}
\epsilon_j.
\end{gathered}
\end{equation*}
The last inequality arrives because whenever $m\in J_k$, we have
$m^q-n^q>$max$(N_k,N_l)$ for any $n\in J_l$ with $m>n$. Similarly,
we evaluate the second term to get
\begin{equation*}
\displaystyle \sum_{l\geq 1}\left\|\sum_{n\in J_l,m<n}
S^{n^q-m^q}(x_l)\right\| \leq k\epsilon_k +\sum_{j\geq k+1}
\epsilon_j.
\end{equation*} Set $\alpha_k=k\epsilon_k +\sum_{j\geq
k+1}\epsilon_j.$ So we arrive at the inequality,
\begin{equation}\label {eq3}
\displaystyle \left\|T^{m^q}(x)-x_k\right\|\leq 2\alpha_k<3
\alpha_k, \forall m\in J_k, k\geq 1.
\end{equation}
Choose $\epsilon_k$ such that $\alpha_k \rightarrow 0$. We now show
that $x$ is a $q$-frequently hypercyclic vector for the operator $T$
with respect to the topology $\tau$. Let $G$ be a nonempty
$\tau$-open set and $y+U\subset G$ for some $\tau$-neighborhood $U$
of the origin. Then we find a balanced neighborhood $V$ of the
origin such that $V+V\subset U$. Since $Y$ is $\tau$-dense in $X$,
we find an increasing sequence of natural numbers $(n_k)$ such that
$x_{n_k}-y\in V$ for all $k\geq 1$. Since the $||.||$-topology is
finer than $\tau$, it follows from \eqref{eq3} that, for some $N\in
\mathbb{N}$, $T^{m^q}(x)-x_k \in V$ for every $m\in J_k$ and $k\geq
N$. Thus from the facts that $x_{n_N}-y\in V$ and
$T^{m^q}(x)-x_{n_N} \in V$, we obtain that for all $m\in J_{n_N}$,
$T^{m^q}(x)-y \in V+V \subset U$. Our conclusion now follows because
$T^{m^q}(x)\in G $ for all $m\in J_{n_N}$, which has positive lower
density.
\end{proof}
\begin{remark}
It is evident from the proof of Theorem \ref{thm3.5} that the
sequence $(T^{n^q})$ is frequently hypercyclic and thus $T$ is
$q$-frequently hypercyclic. It would be interesting to know whether
the converse is true or not, i.e., is $(T^{n^q})$ frequently
hypercyclic whenever $T$ is $q$-frequently hypercyclic?
\end{remark}
\begin{remark} \label{rem3.6} Let us also
note that in the above theorem, the countability assumption on a
subset of $D$ may be waived in case the topology $\tau$ is generated
by an $F$-norm. Indeed, if $(x_n)$ is a $\tau$-dense sequence in
$X$, choose a countable set $\{y_{n,m}:n,m\geq1\}$ where $y_{n,m}\in
D\cap B(x_n,\frac {1} {m})$, $B(x_n,\frac {1} {m})$ being the open
ball of radius $\frac {1} {m}$ centered at $x_n$. It is now easy to
verify that $\{y_{n,m}:n,m\geq1\}$ is a $\tau$-dense set in $X$.
\end{remark}
\section{Applications}
In this section, we consider some applications of the $q$-frequent
hypercyclicity criterion for obtaining $q$-frequently hypercyclic
operators on spaces equipped with linear topologies which are not
necessarily metrizable. Besides, we prove the frequent
hypercyclicity of a non-convolution operator on $H(\mathbb{C})$, at
the end of the section. Let us begin with the results on sequence
spaces with metrizable topologies.
\begin{proposition}\label{prop4.1}
Let $\lambda$ be a sequence space equipped with an $F$-norm and let
$\{e_n\}$ be an unconditional basis in $\lambda$. If for some $q\in
\mathbb{N}$,
\begin{center}
$\displaystyle \sum_{n\in\mathbb{N}} \frac {1}
{w_1w_2...w_{n^q+j}}e_{n^q+j}$
\end{center}
converges unconditionally for each $ j\in \mathbb{N}$, then the
backward shift $B_w$ associated to the weight sequence $(w_n)$, is
$q$-frequently hypercyclic.
\end{proposition}
\begin{proof}
Since $\lambda$ is an $F$-space, in view of Remark \ref {rem3.6}
choose $D$ to be the dense set spanned by $\{e_n:n\geq1\}$. Define
$S_w$ on $D$ by $S_w(e_n)= \frac {1} {w_{n+1}} e_{n+1}$. To apply
our criterion, we are only required to prove the unconditional
convergence of $\sum S_w^{n^q}(x)$ for each $x\in D$. Indeed, for a
given $k\in \mathbb{N}$, we have
\begin{center}
$\displaystyle S_w^n(e_k)= \frac {1} {w_{k+1}...w_{k+n}} e_{k+n}$.
\end{center}
Thus, by the hypothesis, the series
\begin{center}
$\displaystyle \sum _{n\geq 1} S_w^{n^q}(e_k)=w_1w_2...w_k
\sum_{n\in\mathbb{N}} \frac {1} {w_1...w_{k+n^q}} e_{k+n^q}$.
\end{center}
converges unconditionally in $\lambda$. Hence $B_w$ is
$q$-frequently hypercyclic.
\end{proof}
As a consequence of the above result, we obtain a $q$-frequently
hypercyclic operator that is not frequently hypercyclic. This is the
Bergman shift, considered in \cite {Bayart Grivaux}.
\begin{corollary}
Let $B_w$ be the unilateral shift on $\ell^2$ given by the weights
$w_n=\sqrt{\frac{n+1}{n}}$, $n\geq 1$. Then $B_w$ is $2$-frequently
hypercyclic and is not frequently hypercyclic; a fortiori, $B_w$ is
$q$-frequently hypercyclic for any $q\geq 2$.
\end{corollary}
\begin{proof}
Since $w_1w_2...w_{n^2+j}=\sqrt{n^2+j+1}$, the result follows.
\end{proof}
\begin{remark}
One may apply Proposition \ref{prop4.1} to obtain the $q$-frequent
hypercyclicity of shift operators defined on Fr\'{e}chet spaces, for
example the space $H(\mathbb{C})$ of entire functions equipped with
the compact-open topology, the space of all sequences with the
topology of co-ordinate convergence and the classical $\ell^p$
spaces.
\end{remark}
Before we move on to another application, let us see an example.
\begin{example} Let $p\in \mathbb{N}$. Then there
exists an operator which is not $p$-frequently hypercyclic, but it
is $q$-frequently hypercyclic for all $q\geq p+1$. We consider the
unilateral shift $B_w$ on $\ell^2$ with weights $w_k=\displaystyle
{\left(\frac {k+2}{k+1}\right)}^{1/2p}$, $k\in \mathbb{N}$. Then the
proof similar to that of Example \ref{example3.4} shows that $B_w$
is not $p$-frequently hypercyclic on $\ell^2$. We can also conclude
that $B_w$ is $(p+1)$-frequently hypercyclic by applying Proposition
\ref{prop4.1} to the dense set of finite sequences.
\end{example}
Our next application of Theorem \ref{thm3.5} is for the bilateral
backward shift operators. Let $X$ be an $F$-sequence space over the
set $\mathbb{Z}$ of integers such that the unit sequences
$(e_n)_{n\in \mathbb{Z}}$ form an unconditional basis in $X$. For
$w=(w_n)\in \ell^\infty(\mathbb{Z})$, the operator
$T_w(e_n)=w_ne_{n-1}$ is the bilateral backward shift. The content
of the following proposition is the $q$-frequent hypercyclicity of
$T_w$.
\begin{proposition}\label{prop4.5}
Let $(e_n)_{n\in \mathbb{Z}}$ form a unconditional basis in an
$F$-sequence space $X$ and let $q\in \mathbb{N}$. If
\begin{center}
$\displaystyle \sum_{n\in\mathbb{N}} \frac {1}
{w_1w_2...w_{n^q+j}}e_{n^q+j}$ and $\displaystyle
\sum_{n\in\mathbb{N}} w_jw_{j-1}...w_{-n^q+j+1}e_{-n^q+j}$
\end{center}
converge unconditionally for each $ j\in \mathbb{N}$, then $T_w$ is
$q$-frequently hypercyclic on $X$.
\end{proposition}
\begin{proof}
Let $D$ be the set spanned by the sequence $(e_n)_{n\in\mathbb{Z}}$.
Consider the map $S_w(e_k)=\frac {1} {w_{k+1}}e_{k+1}$ on the dense
set $D$ of $X$, so that $T_wS_w$ is the identity operator on $D$.
Then,
\begin{center}
$T_w^{n^q}(e_j)=w_jw_{j-1}..w_{j-n^q+1}e_{j-n^q}$
\end{center} and
\begin{center}
$S_w^{n^q}(e_j)=\frac {1} {w_{j+1}w_{j+2}...w_{j+n^q}}e_{j+n^q}$
\end{center} for each $j\in \mathbb{Z}$.
From the hypothesis, we obtain that the series $\sum T_w^{n^q}(e_j)$
and $\sum S_w^{n^q}(e_j)$ converge unconditionally in $X$. The
desired result now follows since $T_w$ and $S_w$ are linear and $D$
is the span of $(e_n)_{n\in\mathbb{Z}}$.
\end{proof}
A particular case of Proposition \ref{prop4.5} is when $T_w$ is the
bilateral backward shift defined on the sequence space
$\ell^p(\mathbb{Z})$ or $c_0(\mathbb{Z})$ for $1\leq p<\infty$.
Recall that hypercyclicity of $T_w$ was characterized by H. N. Salas
in \cite{Salas}. Also a series $\sum_{n\in \mathbb{Z}} a_ne_n$
converges unconditionally in $\ell^p(\mathbb{Z})$ if and only if the
sequence $(a_n)\in \ell^p(\mathbb{Z})$. Thus we derive the following
result from Proposition \ref{prop4.5}.
\begin{corollary}\label{cor4.6}
Let $q\in \mathbb{N}$ and $1\leq p<\infty$. Assume that for each
$j\in \mathbb{Z}$,
\begin{center}
$\sum_{n\in\mathbb{N}} \frac {1} {(w_1w_2...w_{n^q+j})^p}<\infty$
and $\displaystyle \sum_{n\in\mathbb{N}}
(w_jw_{j-1}...w_{-n^q+j+1})^p<\infty$.
\end{center}
Then $T_w$ is $q$-frequently hypercyclic on $\ell^p(\mathbb{Z})$. If
$\lim_{n\rightarrow \infty} (w_1w_2...w_{n^q+j})= \infty$ and
$\lim_{n\rightarrow \infty} (w_jw_{j-1}...w_{-n^q+j+1})=0$ for each
$j\in \mathbb{Z}$, then $T_w$ is $q$-frequently hypercyclic on
$c_0(\mathbb{Z})$.
\end{corollary}
We have yet another application of Theorem \ref{thm3.5}. Let $X$ be
a Banach space having a Schauder basis $\{x_n,f_n\}$. Then the dual
$X^*$ is weak*-separable and the weak*-topology is not metrizable,
when $X$ is infinite dimensional. We assume that $\{x_n,f_n\}$ is a
\emph{symmetric Schauder basis} for $X$, which means that $ \sum
_{n\geq 1} f_{\mu(n)}(x)x_{\sigma(n)}$ converges for each $x\in X$
and every pair $(\mu,\sigma)$ of permutations of $\mathbb{N}$. A
symmetric base in a Banach space is regular $(\inf ||x_n||>0)$ and
bounded $(\sup ||x_n||< \infty)$, see for example \cite {Kamthan
Gupta1}, p. 133. Corresponding to a weight sequence $(w_n)$ and a
symmetric Schauder basis $\{x_n,f_n\}$ (where the index starts from
1), we define the backward shift $B_w$ by $B_w(f_n)=w_nf_{n-1}$,
$n\geq1$ with $f_0=0$. This operator is continuous with respect to
the norm as well as the weak*-topology of $X^*$, cf. \cite
{Petterson}. We now prove:
\begin{theorem} \label{thm4.4}
Let $X$ be a Banach space with a symmetric Schauder basis
$\{x_n,f_n\}$. Then for $q\in \mathbb{N}$, the backward shift
operator on $X^*$ is $q$-frequently hypercyclic with respect to the
weak*-topology of $X^*$ if $\displaystyle \sum_{n\in\mathbb{N}}
\frac {1} {w_{1}...w_{j+n^q}}$ converges for each $j\in \mathbb{N}$.
In particular, if $\displaystyle \sum_{n\in\mathbb{N}} \frac {1}
{w_{1}w_{2}...w_{n}}$ converges, then $B_w$ is weak*-frequently
hypercyclic on $X^*$.
\end{theorem}
\begin{proof}
In order to apply Theorem \ref{thm3.5}, consider $D$ to be the span
of $\{f_n:n\geq 1\}$. Then $D$ contains a countable weak$^*$-dense
subset. The forward shift $S_w(f_n)= \frac {1} {w_{n+1}} f_{n+1}$,
$n\geq 1$ maps $D$ to itself. Thus
\begin{center}
$\displaystyle S_w^n(f_k)= \frac {1} {w_{k+1}...w_{k+n}} f_{k+n}$
\end{center} and
\begin{center}
$\displaystyle \sum_{n\in \mathbb{N}} S_w^{n^q}(f_k)=w_1w_2...w_k
\sum_{n\in\mathbb{N}} \frac {1} {w_{1}...w_{k+n^q}} f_{k+n^q}$.
\end{center}
Since a symmetric Schauder basis is regular, we have that
$||f_n||<K$ for some constant $K>0$ and for all $n\in \mathbb{N}$,
cf. \cite{Kamthan Gupta1}, p. 261 and \cite{Singer}, p. 25. Hence by
our hypothesis, the series $\sum_{n\in \mathbb{N}} S_w^{n^q}(f_k)$
is absolutely convergent and so unconditionally convergent.
Consequently, the shift $B_w$ is $q$-frequently hypercyclic with
respect to the weak*-topology on $X^*$ by Theorem \ref{thm3.5}.
\end{proof}
As a consequence of the above result, we derive:
\begin{corollary}\label{cor4.5}
The backward shift $B_w$ is weak*-$q$-frequently hypercyclic on
$\ell^\infty$ if $\displaystyle \sum_{n\in\mathbb{N}} \frac {1}
{w_{1}...w_{j+n^q}}$ converges for each $j\in \mathbb{N}$.
\end{corollary}
\begin{proof}
Immediate since $\{e_n:n\geq 1\}$ is a symmetric Schauder basis for
$\ell^1$.
\end{proof}
Thus if $\displaystyle \sum_{n\in\mathbb{N}} \frac {1}
{w_{1}w_{2}...w_{n}}$ converges, $B_w$ is weak*-frequently
hypercyclic on $\ell^\infty$. In fact, the following stronger result
holds.
\begin{proposition}\label{prop4.6}
(1)If $\displaystyle \lim_{n\to \infty}(w_1w_2...w_n)=\infty$, then
the unilateral backward shift $B_w$ is weak*-frequently hypercyclic
on $\ell^\infty$. \\
(2)If $\displaystyle \lim_{n\to \infty}(w_1w_2...w_n)=\infty$ and
$\displaystyle \lim_{n\to \infty}(w_{-1}w_{-2}...w_{-n})=0$, then
the bilateral backward shift $T_w$ is weak*-frequently hypercyclic
on $\ell^\infty(\mathbb{Z})$.
\end{proposition}
\begin{proof}
Since $\displaystyle \lim_{n\to \infty}(w_1w_2...w_n)=\infty$, the
weighted shift $B_w$ is frequently hypercyclic on $c_0$, cf. \cite
{Bayart Grivaux} or \cite {Bayart Matheron}. Also, it is easy to see
that the identity operator from $c_0$ to $\ell^\infty$ is norm to
weak* continuous and has weak$^*$-dense range. Hence $B_w$ is
weak*-frequently hypercyclic on $\ell^\infty$, by Proposition
\ref{prop2.3}. Similarly, the identity map takes $c_0(\mathbb{Z})$
into $\ell^\infty(\mathbb{Z})$ continuously and densely. Thus $T_w$
is weak*-frequently hypercyclic on $\ell^\infty(\mathbb{Z})$.
\end{proof}
\begin{remark}
We would like to mention here that hypercyclicity on $\ell^\infty$
which is weak$^*$-separable, was studied in \cite {Bes Chan Sanders}
and \cite {Petterson}. It was proved that a backward shift $B_w$ is
weak$^*$-hypercyclic on $\ell^\infty$ if and only if $\limsup_{n\to
\infty}(w_1w_2...w_n)=\infty$. However, the condition
$\lim_{n\rightarrow \infty} (w_1w_2..w_n)=\infty$ is not necessary
for $B_w$ to be weak$^*$-frequently hypercyclic on $\ell^\infty$.
Indeed, there exists a frequently hypercyclic $B_w$ on $c_0$ (and
thus weak$^*$-frequently hypercyclic on $\ell^\infty$) such that
$w_1w_2..w_n\nrightarrow \infty$. cf. \cite{Bayart Griv}, p. 205.
\end{remark}
\begin{remark}
In view of Theorem \ref{thm3.5}, a weakly $q$-frequently hypercyclic
operator on a separable Banach space, which satisfies the
$q$-frequent hypercyclicity criterion with respect to a weakly dense
set is necessarily norm-$q$-frequently hypercyclic; for the closed
convex sets are the same in the weak and norm topologies .
\end{remark}
Finally, we consider the frequent hypercyclicity of a
non-convolution operator. It is known that any convolution operator
on $H(\mathbb{C})$ (a continuous linear operator that commutes with
all translations) that is not a multiple of the identity operator is
frequently hypercyclic \cite {Bonilla Erdmann1}. The operator
$T_\mu(f)(z)=f'(\mu z)$ on the space $H(\mathbb{C})$ is a
non-convolution for $\mu\not=1$ and was shown to be hypercyclic for
$|\mu|\geq 1$, cf. \cite{Gustavo Hallack} and \cite{Aron Dinesh}. In
fact, this operator is a weighted backward shift with weights
$w_n=n\mu^{n-1}$. So, our result can be derived using the
Proposition \ref{prop4.1}. We rather prove this in the following
way.
\begin{proposition}
Let $H(\mathbb{C})$ be equipped with the compact-open topology. Then
the operator $T_\mu$ is frequently hypercyclic on $H(\mathbb{C})$
for $|\mu|\geq 1$.
\end{proposition}
\begin{proof}
Let $D$ be the set of all polynomials. Define the map $S_\mu$ by,
\begin{center}
$\displaystyle S_\mu(f)(z)=\mu \int_0^{\frac {1} {\mu} z} f(u)\,du
$.
\end{center}
It is easy to see that $\sum T_\mu ^n(f)$ is absolutely convergent
in $H(\mathbb{C})$ and that $T_\mu S_\mu = I$, the identity on the
set $D$. We fix a $k\geq0$ and consider the function $z^k$. Then
\begin{center}
$\displaystyle S_\mu ^n(z^k)= \frac {k!z^{k+n}} {(k+n)!
\mu^{nk+n(n-1)/2}}.$
\end{center}
Since $|\mu|\geq 1$, the series $\sum S_\mu ^n(f)$ is absolutely
convergent for any polynomial $f$ and we conclude that $T_\mu$ is
frequently hypercyclic by Theorem \ref{thm2.2}.
\end{proof}
We ask the following question. Can one say that the operators
$T_{\mu,b}(f)(z)=f'(\mu z+b)$ on $H(\mathbb{C})$ for $|\mu|\geq 1$
and $b\in \mathbb{C}$, $b\not=0$, are frequently hypercyclic? These
have been shown to be hypercyclic in \cite{Aron Dinesh} and
\cite{Gustavo Hallack}.
\section{Rotation and Powers}
We now remark that rotations and powers of a $q$-frequently
hypercyclic operator on an arbitrary separable topological vector
space remain $q$-frequently hypercyclic for any $q\in \mathbb{N}$.
They also share the same set of $q$-frequently hypercyclic vectors.
These results have been proved for the case $q=1$ in \cite{Bayart
Matheron}. The hypercyclicity of powers and rotations have been
considered by S. I. Ansari \cite{Ansari} and F. Leon-Saavedra and V.
M\"{u}ller \cite{Leon Muller} respectively. For establishing the
following theorem on powers and rotations of $q$-frequently
hypercyclic operators, we need a lemma stated as
\begin{lemma}
Let $A\subset \mathbb{N}$ have positive $q$-lower density and
$\displaystyle \bigcup_{j=1}^{k}I_j=\mathbb{N}$. If $n_1,...,n_k$
are finitely many natural numbers, then
\begin{center}
$\displaystyle \bigcup_{j=1}^{k}(n_j+A\cap I_j)$
\end{center}
has positive $q$-lower density.
\end{lemma}
\begin{proof}
Omitted as it follows on the same lines as given in \cite{Bayart
Matheron}, p. 148.
\end{proof}
Let us denote by $qFHC(T)$, the set of all $q$-frequently
hypercyclic vectors for $T$ and $S^1$, the unit circle in the
complex plane. Then we have
\begin{theorem}
Let $T$ be a $q$-frequently hypercyclic operator on a complex
topological vector space $X$. Then $\lambda T$ and $T^p$ are
$q$-frequently hypercyclic for each $\lambda \in S^1$ and $p\in
\mathbb{N}$. Also $qFHC(T)=qFHC(\lambda T)=qFHC(T^p)$.
\end{theorem}
\begin{proof}
To get this result, we proceed on similar lines as in the case of
frequent hypercyclicity, \cite{Bayart Matheron}, p. 148.
\end{proof}
Finally, we would like to mention that the notion of $q$-frequent
hypercyclicity is a particular case of $(m_k)$-hypercyclicity
studied in \cite{Bayart Mathero}. Indeed, for a strictly increasing
sequence $(m_k)$ of natural numbers, an element $x\in X$ is called
$(m_k)$-hypercyclic for an operator $T$ on $X$ if for every
non-empty open set $U\subset X$, there exists a strictly increasing
$(n_k)=O(m_k)$ such that $T^{n_k}(x)\in U$ for all $k$. Thus the
case $m_k=k^q$, for all $k$ coincides with the notion of
$q$-frequent hypercyclicity; however, the results in our paper have
no overlap with the results of \cite{Bayart Mathero} except that the
$q$-frequent hypercyclicity $(q\geq 2)$ of the Bergman shift has
been proved in \cite{Bayart Mathero} using the notion of a
hypercyclicity set; see Example 5.3, \cite{Bayart Mathero}.
\textbf{Acknowledgement.} The authors are thankful to the referee for
his/her careful reading of the paper, and pointing out the
references \cite{Bayart Griv} and \cite{Bayart Mathero}, which
respectively helped them to answer the converse of Proposition
\ref{prop4.6}(1) and mention the $(m_k)$-hypercyclicity. The second
author acknowledges a financial support from the Council of
Scientific and Industrial Research India for carrying out research
at IIT Kanpur.
\bibliographystyle{amsplain}
|
\section{Introduction}
Molecular clouds are believed to be among the coolest regions in a galaxy where low-mass star formation can take place. Because star formation involves an enormous range of complex physical processes, it is difficult to construct exhaustive theoretical models which include every aspect of the problem (\cite{Krumholtz}).
It is generally found that systems consisting of multiple protostars emerge as a result of the gravitational collapse of rotating molecular cloud cores (\cite{Raghavan,Janson}).
A significant fraction of these multiple star systems are binary systems.
In fact, infrared surveys of nearby star formation regions reveal that around $~75\%$ of the newly formed protostars are part of multiple stellar systems (\cite{Duchêne}). Furthermore, an interesting population of loosely bound very low-mass (VLM) binaries with large separation $d$ ($d >$ 100 AU) is also found in very young star forming regions( e.g. \cite{Béjar,Bouy,Luhman}).
Theories which try to explain the formation processes of binary star systems include disk fragmentation, fission of a protostar(s), and dynamical capture of protostars into bound systems (\cite{Tobin,Moeckel}).
Among these possible scenarios, the disk fragmentation model is the subject of the investigation reported in this paper. The basic purpose of this paper is to reexamine numerically certain aspects of the gravitational collapse and the subsequent fragmentation of rotating molecular cloud cores. We focus on the dependence of disk fragmentation as a low-mass binary and multiple star formation mechanism on the initial temperature as well as the critical density which marks the transition from isothermal to adiabatic thermodynamic behaviour of the collapsing gas.
A significant number of simulation results have so far been reported in the literature in which binary systems emerging from disk fragmentation have been investigated (\cite{Fumitaka,Machida,Forgan}). The availability of advanced computational resources and improved versions of state-of-the-art star formation codes have significantly contributed to our understanding of the physical processes involved in star formation (\cite{Frank,Loughnane,Hubber,Duffin}). Among the recent efforts, we mention numerical models of the collapse of molecular clouds to protostellar densities from an initial Gaussian and Plummer density distribution (\cite{Gomez-Ramirez}). The impact of magnetic fields on
protostellar collapse and fragmentation has also been investigated in detail (for example, \cite{Bürzle,Donati,Peters,Boss}).
Based on the results of his collapse calculations including radiative transfer in the SPH framework, \cite{Stamatellos} found that for densities below 10\superscript{-13} g/cm\superscript{3}, the temperature of the gas can be approximated by the following expression:
\begin{equation} \label{Tdependence}
T=5 K\left( \frac{\rho}{10\superscript{-13} g/cm\superscript{3}}\right)^{0.08}.
\end{equation}
The above relation describes the slow heating process in collapsing cores within the density range 10\superscript{-18} g/cm\superscript{3} $< \rho <$ 10\superscript{-13} g/cm\superscript{3}. Within this range the effective equation of state of the gas is almost isothermal. Beyond this density range the temperature creeps up much more rapidly due to a rise in the opacity of the gas. As a result, the model equation \ref{Tdependence} can not be used any more to estimate the temperature of the cloud.
Furthermore, in a recent paper, \cite{Launhardt} revealed the decisive role which is played by physical quantities such as the initial temperature and the initial density structure of the cloud in determining the final end products of gravitationally collapsing cores. Molecular gas with a slightly enhanced metallicity compared to the metallicity which is normally found in the solar neighborhood could provide an efficient cooling mechanism in collapsing clouds. Metal rich gas allows molecular cloud cores to remain isothermal for a longer period of time because a cloud with a relatively higher metallicity would be subject to more efficient radiative cooling, hence prolonging its isothermal state during collapse compared to clouds in which metal poor conditions prevail (see for example \cite{Omukai}). To mimic the effect of varying the metallicity, we have used a barotropic equation of state and analyzed the evolution of rotating collapsing molecular cloud cores by varying the value of the transition density at which the gas switches its thermodynamic behaviour from isothermal to adiabatic. A similar strategy has been adopted by \cite{Gomez-Ramirez}. The main difference of their setup with our work is that we consider initial conditions with uniform density rather than Gaussian and Plummer models.
On the other hand, observations of molecular cloud cores have revealed values for the lowest temperatures down to at least 8K with typical values of 10K (\cite{BensonMyers,WardThompson}).Therefore we explored a slightly unorthodox but yet possible range of initial temperatures in star-forming molecular clouds between 8K to 12K, keeping in mind that several physical mechanisms can cause prestellar gas to heat up or to cool down from the commonly adopted value of 10 K. For example, cosmic rays can ionize the medium causing the temperature to climb up from 10K to a certain limit or line radiation from molecules can cool down the gas below this value(\cite{Padovani,Indriolo,Wiener}). This leads us to consider the possible role of the initial thermal state of star-forming clouds in determining the properties of binary protostar systems. We focus on the relation between the binary separation of the protostars formed through gravitational collapse and the initial temperature of the cloud, as well as the possible emergence of protostellar objects as a result of secondary fragmentation.
The numerical models of cloud collapse recently reported in (\cite{Gomez-Ramirez}) demonstrate that there exists a tendency for models with initial Plummer and Gaussian density distributions to switch from the isothermal to the adiabatic regime at slightly lower critical densities. Models with a lower critical density show enhanced fragmentation compared to models with higher critical densities.
Keeping this in mind, we also examine the relation between the critical density and
the number of fragments resulting from the gravitational collapse. We also present an analysis of the extreme values of density and temperature within the fragments.
The structure of the paper is as follows. In section 2, we provide details on the SPH code GRADSPH which has been used in this work. We describe the setup of the numerical models in section 3. Section 4 gives an overview of the results of our simulations. Finally, our conclusions are presented in section 5.
\section{Numerical Method}
For the hydrodynamical models of molecular could collapse considered in this work we use GRADSPH, a tree-based, parallel particle-based hydrodynamics code based on the Smoothed Particle Hydrodynamics (SPH) method written in FORTRAN 90(\cite{GRADSPH}). The code has several features which enable us to model gravitational cloud collapse, including the treatment of gravitational forces using a tree-based gravity (TCG) method combined with a variable gravitational softening length formalism(for details see \cite{GRADSPH} and \cite{PriceMon}). The SPH formalism implemented in GRADSPH is based on deriving the SPH equations from a variational principle(\cite{Pricereview}). The reliability of the code has been established through an extensive series of tests on standard benchmark problems and(\cite{GRADSPH}) and the code has recently been extended to MHD(\cite{GRADSPMHD}).
According to (\cite{Pricereview}),(see also \cite{Gingold}),the basic equations of motion used for evolving the SPH particles in GRADSPH are given as follows:
\begin{equation} \label{GRADSPHeom}
\begin{split}
& \frac{d\vec{v}_{i} }{dt} = \\ &
-\sum _{j=1}^{N} m_{j} \left[\frac{P_{i} }{\rho _{i}^{2} \Omega_{i}}\nabla _{i} W(r_{ij} ,h_{i})
+ \frac{P_{j} }{\rho_{j}^{2} \Omega_{j} } \nabla _{i} W(r_{ij} ,h_{j}) \right],
\end{split}
\end{equation}
in which the coefficients $\Omega_{i}$ are defined as
\begin{equation} \label{omega}
\Omega_{i}=1-\frac{\partial h_{i}}{\partial \rho_{i}}\sum_{j=1}^{N} m_{j} \frac{\partial W(r_{ij},h_{j})}{\partial h_{j}}
\end{equation}
In the above equations, $N$ is the total number of particles in the simulation, $P_{i}$ and $\rho_{i}$ denote the density and pressure of a particle labeled $i$, $h_{i}$ is its smoothing length and $m_{i}$ denotes its mass.
The position and velocity vectors for the particles are denoted by $\vec{r}_{i}$ and $\vec{v}_{i}$, respectively, and $\vec{r}_{ij}=\vec{r}_{i}-\vec{r}_{j}$.
The coefficients $\Omega_{i}$ self-consistently take into account the effect of a variable smoothing length (\cite{Pricereview}). In this paper, we close the system of equations by specifying a barotropic equation of state $P=P\left(\rho\right)$ for the gas which will be specified in section 3. In this way we eliminate the need to solve an energy equation. The code uses time-dependent artificial viscosity terms in order to capture shock waves and prevent particle penetration. In order to maintain hydrodynamic stability, we use the signal-velocity approach introduced by Price and Monaghan (\cite{PriceMon}) to calculate the artificial viscosity terms and the Courant time step. The Courant number is set to 0.1 in our simulations. This choice is found to be sufficient to maintain stability in our calculations. In the artificial viscosity terms, we set the parameters $C_{a}$ and $\alpha_{min}$ in Eqns. (34) and (35) in (\cite{GRADSPH}) to 0.2 and 0.5, respectively.
The particle densities are computed from the standard SPH summation equation (\cite{Pricereview,GRADSPH})
by summing the contribution from all the particles which overlap with the position of particle $i$,
using a weighting function $W\left(\vec{r}_{i}-\vec{r}_{j},h_{i}\right)$:
\begin{equation} \label{density}
\rho _{i} =\sum _{j=1}^{N} m_{j} W\left(\vec{r}_{i}-\vec{r}_{j},h_{i}\right).
\end{equation}
In the above expression, $W\left(\vec{r}_{i}-\vec{r}_{j},h_{i}\right)$ is a smooth differentiable function, referred to as the smoothing kernel or the interpolating kernel. GRADSPH uses the standard M4-kernel or cubic spline kernel (\cite{Pricereview}).
The smoothing length is updated at each time step by iteratively solving the following equation for each particle:(\cite{Pricereview}):
\begin{equation} \label{smoothinglength}
\frac{4\pi}{3}\left(2 h_{i}\right)^{3}\rho_{i}=m_{i}N_{opt},
\end{equation}
in which $N_{opt}$ is the number of neighbours contained within the smoothing sphere of each particle.
$N_{opt}$ is set to 50 in our calculations.
In this way we ensure that a constant mass is contained within the smoothing sphere for every SPH particle
so that the code adapts its resolution to keep track of density changes that may occur during the dynamical evolution of the fluid. The system of ordinary differential equations \ref{GRADSPHeom}, which updates the positions and velocities of the particles, is solved using a predictor-corrector scheme combined with an individual particle time stepping method.
As mentioned before, the self-gravity of the gas is treated using the TCG method. We compute the gravitational acceleration of each particle using a Barnes-Hut tree algorithm. The opening angle for the tree is an important parameter which we set to $\theta$ = 0.7.
The same Barnes-Hut tree algorithm is also used to update the list of neighbours of the SPH particles. We use the cubic spline kernel to soften gravitational forces and include the correction terms derived by (\cite{PriceMon}) to ensure the conservation of energy when dealing with variable particle smoothing lengths. The gravitational acceleration of particle $i$ is thus given by
\begin{equation} \label{gravaccel}
\begin{split}
& \vec{g}_{i}=-G \sum_{j=1}^{N}\left[\Phi^{'}\left(r_{ij},h_{i}\right)+
\Phi^{'}\left(r_{ij},h_{j}\right)\right]\vec{e}_{ij}- \\ &
\frac{G}{2}\sum_{j=1}^{N}\left[\frac{\zeta_{i}}{\Omega_{i}}\nabla_{i}W\left(r_{ij},h_{i}\right)
+ \frac{\zeta_{j}}{\Omega_{j}}\nabla_{i}W\left(r_{ij},h_{i}\right)\right]
\end{split}
\end{equation}
in which the quantities $\zeta_{i}$ are defined as
\begin{equation} \label{zeta}
\zeta_{i}=\frac{\partial h_{i}}{\partial \rho_{i}}\sum_{j=1}^{N} m_{j} \frac{\partial \Phi(r_{ij},h_{i})}{\partial h_{i}},
\end{equation}
$\Phi\left(r,h\right)$ is the softened gravitational potential of a particle, and $\frac{\partial \Phi\left(r,h\right)}{\partial h}$ is the derivative of the potential with respect to the smoothing length. Expressions for the softened potential and its derivatives are tabulated in (\cite{PriceMon}).
\section{Initial conditions}
\begin{table*}
\centering
\begin{minipage}{140mm}
\caption{Summary of the physical parameters and the final outcome for the models considered in this paper. The initial mass, radius and density for each model are given by the constant values 5 x 10\superscript{16} cm, 1 $M_{\bigodot}$ and 3.8 x 10\superscript{-18} g/cm\superscript{3},respectively. }
\begin{center}
\begin{tabular}{cccccc}
\hline
Model & Temperature(K) & $c_{0}$(cm/s) & $\rho_{crit}$(g/cm\superscript{3}) & Final outcome & Binary separation(AU) \\
A & 8 & 1480.0 & 5 x 10\superscript{-14} & Binary & 278.685 \\
B & 9 & 1570.0 & 5 x 10\superscript{-14} & Binary & 343.235 \\
C & 10 & 1650.0 & 5 x 10\superscript{-14} & Binary & 378.364 \\
D & 11 & 1730.0 & 5 x 10\superscript{-14} & Single & --- \\
E & 12 & 1810.0 & 5 x 10\superscript{-14} & Single & --- \\
F & 10 & 1650.0 & 5 x 10\superscript{-15} & None & --- \\
G & 10 & 1650.0 & 5 x 10\superscript{-13} & Triple & --- \\
H & 11 & 1730.0 & 5 x 10\superscript{-14} & Binary & 228.438 \\
I & 12 & 1810.0 & 5 x 10\superscript{-14} & Binary & 197.608 \\
\hline
\end{tabular}
\end{center}
\end{minipage}
\end{table*}
The initial conditions for the cloud collapse calculations considered in this a paper are variants of the Boss and Bodenheimer collapse test with initial conditions described in (\cite{Burkert}). We used this test previously to validate the GRADSPH code on collapse calculations(see section 8.2.3 in \cite{GRADSPH}).
The initial condition that we take as our standard model is a solar mass cloud with uniform density and radius R = 5 x 10\superscript{16} cm. The cloud is assumed to be in solid body rotation around the z-axis of the coordinate system with an angular velocity equal to $\omega$ = 7.2 x 10\superscript{-13} rad/s and is rotating counter clockwise. The mean initial density of the cloud is
$\rho_{0}$ = 3.8 x 10\superscript{-18} g/cm\superscript{3}. The chemical composition of the cloud is assumed to be a mixture of hydrogen and helium gas with mean molecular weight $\mu$ = 3. The initial condition is characterized by the parameters $\alpha$ and $\beta$, which correspond to the ratio of thermal and kinetic energy with respect to the gravitational potential energy of the cloud. These parameters are defined as follows:
\begin{equation} \label{alpha}
\alpha=\frac{5 R k T}{2 G M \mu m_{h}},
\end{equation}
\begin{equation} \label{beta}
\beta=\frac{R^{3}\omega^{2}}{3 G M},
\end{equation}
where $G$ is the gravitational constant, $k$ is the Boltzmann constant, and $m_{h}$ denotes the mass of the hydrogen atom. For the standard initial temperature of 10 K, the initial values
for $\alpha$ and $\beta$ are 0.26 and 0.16, respectively. In our models, the value of $\alpha$ changes by modifying the initial temperature, whereas $\beta$ will be kept fixed at its standard value. The mean free-fall time of the initial condition is given by
\begin{equation} \label{freefalltime}
t_{ff}=\sqrt{\frac{3 \pi}{32 G \rho_{0}}}
\end{equation}
and is 33968 years for the standard initial condition defined above. The initial setup is implemented in our SPH code by placing equal-mass particles on a hexagonal closely packed lattice and retaining only the particles within the initial cloud radius. The code uses internal dimensionless units which are defined by setting G=M=R=1.
To initiate fragmentation of the cloud models, we add an azimuthal density perturbation to the uniform initial condition with a certain mode number m and amplitude A which has the following form:
\begin{equation} \label{densityperturbation}
\rho=\rho_{0}\left(1+A sin\left(m \phi \right)\right],
\end{equation}
where $\phi$ is the azimuthal angle in spheriçal coordinates $\left(r,\phi,z\right)$. We implement this perturbation by perturbing the azimuthal angle of the SPH particles to a new angle $\phi^{*}$ which is determined by solving the equation
\begin{equation} \label{phiperturbed}
\phi=\phi^{*}+\frac{A sin\left(m \phi^{*}\right)}{m}.
\end{equation}
A common mode number m = 2 is adopted for all the models listed in table I while azimuthal density perturbations of amplitude A = 0.1 and A = 0.25 are considered for models (A, B, C, D, E, F, G) and models (H and I), respectively. The thermodynamic behaviour of the gas during cloud collapse is approximated by adopting a barotropic equation of the state as suggested by (\cite{Tohline}) and (\cite{Matsunaga}).
The pressure and sound velocity of the gas are given by the expressions
\begin{equation} \label{Pressure}
P=\rho c_{0}^{2}\left[1+\left(\frac{\rho}{\rho_{crit}}\right)^{\gamma-1}\right],
\end{equation}
and
\begin{equation} \label{soundvelocity}
c_{s}=c_{0}\left[1+\left(\frac{\rho}{\rho_{crit}}\right)^{\gamma-1}\right]^{1/2},
\end{equation}
where $c_{0}$ is the initial sound velocity, $\rho_{crit}$ is the critical density which determines the transition point from isothermal to adiabatic behaviour of the gas, and
$\gamma$ is the adiabatic exponent which we set to 5/3. Our standard value for the critical density is $\rho_{crit}$ = 5 x 10\superscript{-14} g/cm\superscript{3}.
The initial models are allowed to evolve under the action of the short-range hydrodynamical forces and the self-gravity of the cloud. No external radiation feedback is taken into consideration during evolution of our models.
In order to avoid artificial fragmentation in our models because of insufficient resolution, we need to satisfy the resolution criterion derived by (\cite{BateBurkert}) in our models. This is achieved by keeping the minimum resolvable mass of the code, which equals $M_{min}=2 N_{opt}m $, below the local Jeans mass. Since the Jeans mass decreases during the isothermal collapse stage and increasing again when the gas becomes optically thick and adiabatic, this condition defines an upper limit to the mass of an SPH particle following (\cite{Arreaga}) and (\cite{GRADSPH}):
\begin{equation} \label{particlemass}
m < \frac{\pi^{3/2} c_{0}^{3}}{2 N_{nopt}\rho_{crit}^{1/2} G^{3/2}}.
\end{equation}
The present numerical study is conducted by taking a total number of SPH particles equal to N=250025 in the initial condition and setting the number of neighbours to $N_{opt}$=50.
Table I provides an overview of the models which have been used to examine the evolution of the binary separation as a function of the initial temperature of the clouds as well as the impact of the changes in the critical density along with changes in perturbation amplitude. The table gives, for the 9 models labeled A-I, the values of the initial temperature, the sound velocity, the critical density, the resulting binary separation when applicable, as well as the final outcome of the simulations which will be discussed in section 4. It can be seen that models A-E explore the effect of changes in the temperature with the critical density kept at its standard value. In models F and G, on the other hand, the initial temperature is set to 10 K and $\rho_{crit}$ takes on different values to examine the effect of this parameter on the evolution of the cloud models. The last two models labeled H and I reveal the effects of strength of perturbation amplitude on evolution of the collapsing cloud.
In our models, fragmentation proceeds in two stages. In the first stage, a rotating disk like structure forms in the center of the cloud as a result of the conservation of angular momentum. In the second stage, this disk fragments into protostellar fragments once the gas has become optically thick and evolves adiabatically because of non-axisymmetric gravitational instabilities. The growth of this kind of gravitational instabilities is governed by the Toomre parameter for a disk with surface density $\Sigma$ and epicyclic frequency $\kappa$(\cite{Toomre}). This parameter is defined as
\begin{equation} \label{ToomreQ}
Q=\frac{c_{s}\kappa}{\pi G \Sigma},
\end{equation}
and the value of Q must be smaller than unity to enable non-axisymmetric instabilities to grow within the disk. Plots of the radial behaviour of the Toomre parameter at various stages of the collapse will be discussed in section 4. For visualization of the results of our simulations, we use the visualization tool SPLASH developed and made publicly available to the community by Daniel Price(\cite{SPLASH}).
\section{Results and discussion}
\begin{table*}
\centering
\begin{minipage}{140mm}
\begin{center}
\begin{tabular}{ccccc}
\hline
Model & A & B & C & D \\
$t_{max}$(yrs) & 47913.746 & 47913.746 & 47913.746 & 47913.746 \\
$\rho_{max}$(g/cm\superscript{3}) & 4.767 x 10\superscript{-11} & 2.380 x 10\superscript{-11} & 1.883x10\superscript{-11} & 7.551x10\superscript{-12} \\
$T_{max}$(K) & 1298 & 922.5 & 878.4 & 530.2 \\
\hline
\end{tabular}
\begin{tabular}{cccccc}
\hline
Model & E & F & G & H & I \\
$t_{max}$(yrs) & 47913.746 & 47913.746 & 41443.078 & 47913.746 & 47913.746 \\
$\rho_{max}$(g/cm\superscript{3}) & 4.643x10\superscript{-12} & 8.118x10\superscript{-14} & 8.909 x 10\superscript{-10} & 7.787 x 10\superscript{-12} & 5.042 x 10\superscript{-12}\\
$T_{max}$(K) & 421.5 & 116.7 & 2456 & 540.965 & 444.7 \\
\hline
\end{tabular}
\caption{Summary of the maximum evolution time, as well as the maximum density and the maximum temperature obtained from the evolution of the models. }
\end{center}
\end{minipage}
\end{table*}
Figure 1 shows successive column density maps of the evolution of models A-E at 5 different times indicated to the left of the columns. The initial temperature of the models is indicated at the top of each column. The column density maps are top down views in the xy plane where the density distribution has been integrated along the rotational axis of the cloud. Figure 2 shows the corresponding edge-on column density maps in the xz plane orthogonal to the rotational axis. Every single square panel in figure 1 has a physical dimension of approximately 401 x 401 AU. In Figure 2 every single panel has physical dimensions of approximately 401 x 167 AU, respectively. The outcome of each simulation model and the separation of the binary at the end of each run are summarized in table I. Table II contains the maximum time, $t_{max}$, expressed in years, during which we have been able to evolve the model before the timestep became prohibitively small, as well as the maximum density $\rho_{max}$ and the maximum temperature $T_{max}$ attained during the evolution of each model.
The rise in initial temperature going from model A to model E has a drastic impact on the evolution of the cloud collapse. Models with initial temperatures up to standard value of 10 K develop binary systems, whereas molecular cores with temperatures above 10 K fail to develop binaries yielding only single protostars instead. Changing the initial temperature also has a big impact on the separation of the binaries in the range $8K < T <= 10 K$. For example, raising the temperature from 8 to 9 K leads to change in the final binary separation of around $\Delta d$ = 64 AU, while a further increase from 9 to 10 K leads to a change around $\Delta d$ = 35 AU. The panels in Figure 2 also clearly show the formation of the intermediate disk structure which fragments into protostars for $T <=10 K$.
The phase in which the collapsing fragments are still connected by a bar-like structure is comparatively longer in molecular cloud cores with temperatures below the conventional value of 10 K. This is illustrated in Figure 1 by the snapshots a5, b5, and c5. It can be seen that this connecting filament starts to disappear earlier in model C than in models A and B.
There is also an important difference in evolution of the orbital elements of the models with increasing temperature for models A,B and C which produce binary systems as the final products.
The left panel of Figure 3 shows the evolution of the binary separation with time for these models. The time is indicated in units of the initial freefall time of the cloud. Each model shows a qualitatively similar evolution showing a minimum in the binary separation, after which the separation increases up to the end of the simulation. The physical reason for this behavior is likely connected to the accretion of high angular momentum gas by the protostars from the surrounding envelope. However, there are important differences in terms of the speed of this evolution and the minimum separation of the binaries. Colder molecular cloud evolve slower than hotter ones in the sense that the minimum separation is attained later in the course of the evolution. Note that at the end of the evolution, the faster evolving hotter cores have reached larger binary separations than the colder ones.
We approximately estimate the masses of the evolving fragments by locating the particle with the highest density $\rho_{max}$ within each fragment and include all particles whose density is higher than $f \rho_{max}$, where $f$ is set to 0.001. The velocity and position of each fragment is determined from the position and velocity of the center of mass of the clump, respectively. A similar procedure has been adopted by \cite{Arreaga}. Finally, the orbital elements of the binaries are determined from the positions, velocities and masses of the fragments.
Figure 3 shows that the binary separation in models (A, B, C) mainly follows the variations in semi-major axis and is less affected by the evolution of the orbital eccentricities which remain high ($\sim 0.5-0.6$) throughout of the evolution. Figure 4, on the other hand, shows that the initial thermal state of the clouds in decisive in determining the final mass fraction included in the binary system. The cooler the core, the more mass is involved in the binary fragmentation. The coolest model A has more than
50 $\%$ of its mass included at t=1.35 $t_{ff}$. Since about half of the mass of the cloud is at most included in the binaries at the end of our simulations, we expect that the orbital elements of the binaries reported after 1.4 $t_{ff}$ are unlikely to represent the final orbital elements but are rather indicative of the trend of their further evolution. In particular, we expect that the increasing trend of the binary separation will continue despite the fact that we were unable to prolong the simulations for much longer than 1.4 $t_{ff}$. We also note that the decline of the mass fraction which can be seen in the lower left panel in Figure 4, but is only temporary for model C because this model evolves faster, is likely caused by mass exchange between the fragments and their surrounding disk like structure which could be the result of the tidal forces exerted by the companions. The symmetry of the initial conditions ($m=2$) implies that the binary mass ratio is always close to unity.
Figure 5 shows the time evolution of the maximum temperature for models A-E (left panel) as well as the time evolution of the maximum of the density in the right panel. Comparison of the models shows that for colder molecular cloud cores, the clouds are subject to an earlier increase in temperature along with a faster transition to the adiabatic regime. It is also evident that the final temperature of the models is highly dependent on their initial thermal state with colder models reaching higher final temperatures.
The logarithmic density profiles shown in the right panel indicate a similar trend for the maximum density. Here, the colder models reach higher final densities because self-gravitating fragments which have reached the adiabatic heating phase are formed earlier during the course of their evolution.
Figures 6 to 8 respectively show top down views of the evolution of models H and I as well as the evolution of their maximum density and temperature, respectively. In these models, the temperature is raised above the nominal value of 10 K( 11 K for model H and 12 K for model I) and a bigger perturbation amplitude is applied(A=$25 \%$). Because of the increased strength of the perturbation, models H and I form binary systems at the end of the simulation, and thus clearly indicate that the fragmentation process is a function of both the thermal energy of the core and the strength of the initial perturbation.
Comparison of the results of models A-C and H and I leads to several conclusions. Firstly, cold molecular cores ($T_{0}=$8 K, 9 K, 10 K) with a smaller amplitude of the initial azimuthal density perturbation (A = 10 $\%$) give birth to protobinary systems that evolve relatively quickly and include protostellar fragments that reach higher densities and temperatures compared to those in models H and I. The binaries in models A-C attain minimum separations of $\sim 100$ A.U. and then increase in separation due to the accretion of high angular momentum material.
However, upon examining the evolution of the relatively warm molecular cores in models H and I(11 K and 12 K, respectively) with a larger amplitude of the initial azimuthal density perturbation(A=25 \%), we find that the overall speed of the evolution of these fragmenting models is much slower than for models A-C. At the end of the simulation, models H and I have not yet experienced pericenter passage, and the resulting densities and temperatures are lower than for models A-C, as can be seen by comparing the left and right panels of Figures 5 and 8, respectively. Over the course of the simulation, models H and I undergo a monotonic decline in excentricity from $\sim 0.7$ to $\sim 0.5$.
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=5in]{Fig-1revised}
\caption{Simulation results for models A, B, C, D and E. Each column of plots shows face-on views of the column density integrated along the z axis for each particular model. The corresponding initial temperature is shown at the top of each column. The times corresponding to each row are given in units of $t_{ff}$. The binary separations for each snapshot are respectively given by (a1) d= 313.184 AU, (a2) d = 224.943 AU,(a3) d = 88.758 AU, (a4) d = 199.348 AU, (a5) d = 278.685 AU,
(b1) d = 282.382 AU, (b2) d = 200.013 AU,(b3) d = 151.047 AU, (b4) d = 251.859 AU, (b5) d = 343.235 AU, (c1) d = 262.515 AU,
(c2) d = 166.739 AU,(c3) d = 173.062 AU, (c4) d = 266.667 AU, (c5) d = 378.364 AU. The horizontal and vertical
dimensions of each plot in the xy-plane are 0.12 x 0.12 in dimensionless units. The color bar on the right shows $log\left(\Sigma \right)$ in
dimensionless units. Each calculation was performed with 250025 SPH particles. }
\label{fig1}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 20mm 0mm 0mm, clip, width=5in]{Fig-2revised}
\caption{Same as in Figure 1, but here we show edge-on views of the column densities integrated along the direction orthogonal to the z-axis. The horizontal and vertical dimensions in the xz-plane in each panel are 0.1 x 0.05 in dimensionless units, respectively. }
\label{fig2}
\end{figure*}
Let us now take a closer look at the evolution of the protostellar disks in models A,B and C. The time evolution of the Toomre parameter is an important indicator of the growth of gravitational
instabilities in the disks which form during the gravitational collapse of our cloud models with values below 1 indicating the onset of gravitational instabilities.
Figures 9a to 9e contain plots of the radial evolution of the Toomre parameter, averaged with respect to the azimuthal angle in the midplane of the disk, for models A-E at various instants of time.
It can be observed that the time evolution of the Toomre parameter closely follows the development of fragments during the collapse of the cloud models. At the earlier stages of the evolution, all models exhibit gravitational instability within a radial range of 167 AU.
As the clouds evolve further, the gravitational instabilities in models A,B and C, which start their evolution from temperatures at or below 10 K, gradually sweep larger parts of the disks, almost doubling their active radial range up to 334 AU. On the other hand, The hotter models D and E do not show this kind of behaviour.
For the relatively cold gas models, the part of the initial molecular core which is subject to gravitational fragmentation is about 10 percent of the initial radius. Beyond this range the disk seems fairly stable and remains free of fragmentation processes.
The time evolution of Q for each model also shows that molecular cloud cores with higher initial temperatures take more time to undergo fragmentation than models which start their evolution with lower initial temperature values. The threshold value of unity at or below which fragmentation should happen
never seems to be reached outside of the inner regions by models D and E, whereas models A to C, which start their evolution from temperatures ranging from 8 K to 10 K, show a drop of the Q values well below unity and hence give rise to self-gravitating fragments which are well under way to reach protostellar densities.
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=5in]{Fig-3revised}
\caption{Evolution of the binary separation for models A, B, and C(left panel) and for models H and I(right panel)
as a function of time up to 1.4 $t_{ff}$. The binary separation for the two
protostars is determined by searching for the SPH particles with maximum density and their corresponding distance in simulation space. }
\label{fig3}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=3in]{Fig-4revised}
\caption{Evolution of the total relative mass fraction for the protostellar binaries in
models A, B, and C as a function of time up to 1.4 $t_{ff}$. }
\label{fig4}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=5in]{Fig-5revised}
\caption{Time evolution of the maximum temperature within
the collapsing cloud for models A, B, C, D, E (left panel),and the time evolution
of the maximum density for the same set of models (right panel).}
\label{fig5}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=5in]{Fig-6revised}
\caption{Simulation results for model H. Each plot shows a face-on view the column density integrated along the z axis. The times corresponding to each row are given in units of $t_{ff}$. The binary separations for each snapshot are respectively given by (h1) d= 590.130 AU, (h2) d = 540.667 AU,(h3) d = 479.196 AU, (h4) d = 410.897 AU, (h5) d = 328.628 AU, (h6)= 228.438 AU. The horizontal and vertical dimensions of each plot in the xy-plane are 0.15 x 0.15 in dimensionless units. The color bar on the right shows $log\left(\Sigma \right)$ in dimensionless units. Each calculation was performed with 250025 SPH particles. }
\label{fig6}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=5in]{Fig-7revised}
\caption{Simulation results for model I. Each plot shows a face-on view the column density integrated along the z axis. The times corresponding to each row are given in units of $t_{ff}$. The binary separations for each snapshot are respectively given by (i1) d= 531.065 AU, (i2) d = 522.757 AU,(i3) d = 435.9223 AU, (i4) d = 364.861 AU, (i5) d = 289.241 AU, (i6)= 197.608 AU. The horizontal and vertical dimensions of each plot in the xy-plane are 0.15 x 0.15 in dimensionless units. The color bar on the right shows $log\left(\Sigma \right)$ in dimensionless units. Each calculation was performed with 250025 SPH particles. }
\label{fig7}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=5in]{Fig-8revised}
\caption{Time evolution of the maximum temperature within
the collapsing cloud for models H and I(left panel),and the time evolution
of the maximum density for the same set of models (right panel). }
\label{fig8}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=4.0in]{Fig-9revised}
\caption{Traces of the azimuthal average of the Toomre parameter values as a function of the radial distance r in the cloud midplane for models A to E at times $t=1.47 \, t_{ff}, t=1.49 \, t_{ff},t=1.51 \, t_{ff}, t=1.53 \, t_{ff}$, and $t=1.56 \, t_{ff}$, respectively. }
\label{fig9}
\end{figure*}
We also made an attempt to understand the sensitivity of the structural evolution of collapsing molecular cores
on the threshold density that marks the beginning of the adiabatic heating phase.
At that moment the collapsing core becomes optically thick enough to trap heat inside leading to a transition from isothermal to adiabatic behaviour of the collapsing gas.
For this purpose, we varied the value of the critical density $\rho_{crit}$ by an
order of magnitude around 5 x 10\superscript{-14} g/cm\superscript{3} and investigated the role of adiabatic heating in determining the evolution of the models. The outcome of this investigation is illustrated by the results for models F and G, for which the critical density is set to values 10 times lower or 10 times larger than the standard value, respectively. The evolution for model G is illustrated in Figures 10 and 11. Model G follows a delayed isothermal phase during which a thin but very dense bar is formed connecting the two primary fragments. This bar-like structure subsequently fragments into multiple fragments. This interesting phenomenon also happened in a collapse model test described in (\cite{Kitsionas}), which even further prolonged the isothermal phase by setting $\rho_{crit}$= 5 x 10\superscript{-12} g/cm\superscript{3}.
At the later stages of evolution, the secondary fragments combine with each other and produce a third fragment that starts to interact gravitationally with one of the primary fragments hence giving rise to the possibility of a hierarchical system of protobinaries, with one binary as a component of another already formed binary system.
This final state of the evolution is shown in panel g6 of Figure 10, after which we unfortunately had to stop the calculations because of too small time steps. The emerging protobinary system in model G is found to be the result of a combination of both primary and secondary fragmentation. Note that this evolution is different from that seen in Models A, B, C which resulted in protobinary systems consisting only of primary-primary fragments. For comparison, we refer to panels a5, b5, c5, and g6 of Figure 1 and Figure 10, respectively. Figure 12 shows a detailed view of the protobinary system formed as a result of the secondary fragmentation of the bar. The density evolution of this model shows a quicker shift of eight orders of magnitude in density of the fragments that may evolve up to protostellar densities. This evolution happens in roughly half of the time needed for the rest of the models that yielded a binary system to reach densities near 10\superscript{-10} g/cm\superscript{3}, at which point hydrogen molecules in the collapsing gas experience molecular dissociation and the second collapse phase starts. The sideway views in Figure 13 also clearly show the extent of the vertical dimension of the intermediate disk structure, which is found to be much thinner than for any other collapsing model studied in the present work. Although model G could not evolved to the extent of models of A, B, and C, we suggest that the model is well on its way to becoming a stable triple stellar system.
\newpage
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=7in]{Fig-10revised}
\caption{Snapshots of column density integrated along the rotational axis during the evolution of model G. The horizontal and vertical dimensions in the xy-plane in each panel are 0.15 x 0.15 in dimensionless units.
The colour bar shows $log\left(\Sigma\right)$ in dimensionless units.
The times in each row are in units of $t_{ff}$. }
\label{fig10}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 50mm 0mm 0mm, clip, width=7in]{Fig-11revised}
\caption{Same as in figure 10, but here we show the column densities integrated along the direction orthogonal to the rotational axis. The horizontal and vertical dimensions in the xz-plane in each panel are 0.1 x 0.05 in dimensionless units. }
\label{fig11}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 40mm 0mm 0mm, clip, width=7in]{Fig-12revised}
\caption{Detailed snapshots of
the column density integrated along the z-axis during the evolution of model G.
The first panel on the left has dimensions 0.25 x 0.25, while the other two panels are successive enlargements of the first one.
The colour bar on the right shows $log\left(\Sigma \right)$ in dimensionless units.
The time indicated in each snapshot is given in units of $t_{ff}$. }
\label{fig12}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=7in]{Fig-13revised}
\caption{Snapshots of column density integrated along the rotational axis during the evolution of model F. The horizontal and vertical dimensions in the xy-plane in each panel are 0.15 x 0.15 in dimensionless units.
The color bar shows $log\left(\Sigma\right)$ in dimensionless units.
The times in each row are in units of $t_{ff}$. Each calculation was performed with 250025 SPH particles. }
\label{fig13}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[trim = 0mm 0mm 0mm 0mm, clip, width=7in]{Fig-14revised}
\caption{Same as in figure 13, but here we show the column densities integrated along the direction orthogonal to the rotational axis. The horizontal and vertical dimensions in the xz-plane in each panel are 0.1 x 0.05 in dimensionless units. }
\label{fig14}
\end{figure*}
Figures 13 and 14 describe the time evolution of model F in the equatorial and vertical planes, respectively. The evolution of the core is free of any process of fragmentation for this model. This mainly happens because the inward gravitational pull is challenged by the quicker rise of the temperature of the core so that the gas can not collapse to the stage where clumps could be formed. Instead model F initially collapses to a bar that keeps on gaining thickness and eventually gives birth to a defused spiral structure that surrounds the dense bar as can be seen in panel f6 of Figure 13. Similarly, we also see in Figure 14 that such a warm molecular core fails to develop a thin disk structure as in model G.
The evolution of the surface density and Toomre parameter values in models G and F are shown in Figures 15 and 16 as well as Figures 17 and 18, respectively. The surface density evolution seen in models F and G is a clear manifestation of how conditions leading to fragmentation play a key role in transferring angular momentum of the disk outwards resulting in infall of gas that gives rise to increased surface densities of the disk and hence a drop in the Toomre parameter values. This inverse relationship between Toomre parameter value and the surface density is obvious when comparing the corresponding panels in Figures 16 and 17 and Figures 17 and 18.
Values of the Toomre parameter well above unity also suggest that the thick disk resulting from model F never experiences a phase of gravitational instability.
A comparison of models C, F, G yields further interesting results on the temperature and density
evolution of the molecular core models which are illustrated in Figure 19.
The left and right panels respectively show the evolution of the maximum temperature and density for these three models. It can be
seen that changing the critical density over an order of magnitude range results in a direct relation
between the critical density and the resulting maximum temperature and density values reached during the evolution with higher critical densities corresponding to higher temperature and density values. This
is explained by the fact that a little longer isothermal phase during the collapse allows the molecular gas to get compressed to a denser state, which initially gives rise to a thin filament structure along with multiple fragments which later on in the adiabatic phase become full-fledged protostellar objects. On
the other hand, if the adiabatic phase takes over the initial isothermal regime a little earlier,
the core might end up as a mere reservoir of gas that never reaches densities corresponding to protostellar systems.
Finally, Figure 20 shows snapshots of the temperature integrated along the rotational
axis for the 6 models A-C(the first row) and for models D,E and G(second row) at the final state of evolution of the models. Hardly any temperature difference is found for the two fragments reaching the state of protostars in models A, B, and C.
The single protostellar objects arising in models D and E correspond to the slightly hotter regions in the two leftmost columns in the second row. However, Model G (the bottom right panel) represents a significant
difference in temperature of the resulting triple system mainly because the system hosts both primary and
secondary fragments which have delayed isothermal phases. Moreover, the two primary fragments spend a longer part of their evolution within the isothermal phase compared to the third secondaryy fragment. This may explain the low temperature which is associated with the secondary fragment compared to the primary fragments in the resulting triple stellar system.
\newpage
\newpage
\section{Conclusions}
In this paper, we have investigated the influence of the initial thermal state of molecular cloud cores on the formation of protostellar binary systems through gravitational fragmentation. We find that in the relatively low temperature cores in which binaries are formed(with $T_{0}$ in the range 8-10 K), the binary separation is a function of the initial thermal conditions prevailing in the molecular cloud cores. Darker and colder molecular cloud cores at $T_{0}$ = 8 K evolve into protobinary systems with small separation, whereas for slightly higher initial temperatures of 9 K or 10 K, an increase in the binary separation is observed based on the results of our models. Our quantitative analysis has therefore revealed a strong thermal sensitivity of the separation of evolving protobinary systems. Molecular cloud cores with temperature above 10 K and with small amplitude of initial azimuthal density perturbation(A = 10 $\%$)(models E and F) do not develop binary systems but form single protostars instead. Further investigation of such cores by introducing a stronger amplitude of perturbation (A = 25 $\%$), however, has shown that warm cores can indeed be forced to fragment. The evolution of the resulting binaries is apparently much slower than in the colder case, although we have not been able to follow their evolution until after pericenter passage.
We also investigated the impact of an effective cooling environment that may prolong the isothermal phase of core collapse. We find that higher values for the critical density which separates the regime of isothermal and adiabatic cloud collapse significantly affects the binary fragmentation process, in that additional secondary fragmentation has been observed in a bar connecting the primary prostellar fragments in the central part of the models. This process happens in a fairly short period of time compared to the freefall time of the cores and can give rise to the formation of multiple hierarchical protostellar systems. On the other hand, a reduction of the critical density, hence a less efficient cooling environment, suppresses fragmentation and may even lead to stable disk structures without star formation in the first place.
\section{Acknowledgements}
We thank the Institute of Space and Planetary Astrophysics (ISPA) for financially supporting this research
project. We are also grateful to the Pakistani National Center for Physics (NCP) for providing access to the computational facility that was used in this study.
We also thank the anonymous referee, whose comments substantially improved the contents of the final version of the manuscript.
\newpage
|
\section{Lattice Models for Quantum Contacts}
For studying the 2D interferometer formed in a 2DEG between a QPC and the depletion region induced by a charged tip,
we use lattice models where two semi-infinite square lattices (leads) are connected by a small contact region of length
$2L_x+1$ and maximum width $2L_y+1$. $c_{{\bf i}\sigma}$ ($c^{\dagger}_{{\bf i}\sigma}$) being the destruction (creation) operator of an
electron of spin $\sigma$ at site ${\bf i}$ of coordinates $(i_x,i_y)$, and $n_{{\bf i}\sigma}=c^{\dagger}_{{\bf i}\sigma}c_{{\bf i}\sigma}$,
the Hamiltonians of the left ($i_x\leq -L_x$) and right ($i_x\geq L_x$) leads read
\begin{equation}
H_{leads}= \sum_{{\bf i},\sigma} \left( -4t n_{{\bf i}\sigma} + t \sum_{\bf j} c^{\dagger}_{{\bf i}\sigma} c_{{\bf j}\sigma} \right) + H.C.
\label{lead-Hamiltonian}
\end{equation}
The hopping terms are non-zero between nearest neighbors sites ${\bf i,j}$ only. The energy scale is
defined by taking $t=-1$ (the conduction bands of the leads are in the energy interval $[0,8]$ when the site
potentials are equal to $-4t$). Hereafter, we study the continuum limit (energy $E \ll 1$). The contact Hamiltonian reads
\begin{equation}
H_{contact}= \sum_{{\bf i},\sigma} \left( (V_{\bf i}-4t) n_{{\bf i}\sigma} + t \sum_{\bf j} c^{\dagger}_{{\bf i}\sigma}
c_{{\bf j}\sigma} \right) + H.C.
\label{contact-Hamiltonian}
\end{equation}
The summations are restricted to $-L_x\leq i_x \leq L_x$ and to $-L_y\leq i_y \leq L_y$. Moreover, the site potentials $V_{\bf i}$ are
taken infinite inside the contact region if $|i_y| \geq (L_y-k)+k \left(i_x/L_x \right)^2$, where $k$ is a parameter. This restricts
the electron motion inside a smoothly opening region known to favor a sharp opening of the conductance channels as one increases the
energy. A smooth opening also reduces the interference effects induced by the back-scattering of electron waves leaving the contact region,
effects which induce oscillations in the transmission function $T(E)$. The Hamiltonians describing the coupling between the contact and the
two leads read
\begin{eqnarray}
H_c^l &=& t_c \sum_{i_y=-L_y,\sigma}^{L_y} \left(c^{\dagger}_{{(-L_x,i_y)}\sigma} c_{{(-L_x-1,i_y)}\sigma}+H.C.\right), \\
H_c^r &=& t_c \sum_{i_y=-L_y,\sigma}^{L_y} \left(c^{\dagger}_{{(L_x,i_y)}\sigma} c_{{(L_x+1,i_y)}\sigma}+H.C.\right).
\label{coupling-Hamiltonian}
\end{eqnarray}
The QPC Hamiltonian reads $H_{0}=H_{contact}+\sum_{\alpha=l,r} (H_c^{\alpha}+H_{leads}^{\alpha})$. The QPC transmission $T_0(E)$ is
a staircase function, each stair taking an integer value which counts the number of open channels. To describe the depletion region
induced by the charged tip, a term $H_{tip}(x,y)=\sum_{\sigma} V n_{{\bf T}\sigma}$ is added to $H_{0}$, which modifies by an amount $V$
the potential $-4t$ of a single site ${\bf T}$ of coordinates $(x,y)$ located at a distance $r=\sqrt{x^2+y^2}$ from the contact.
The interferometer Hamiltonian reads $H=H_{0}+H_{tip}(x,y)$. Fig.~\ref{fig2QPC} shows such an interferometer when $L_x=L_y=3$ and $k=2$.
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig2QPC.pdf}
\caption{(Color online) Lattice model for the SGM of a QPC: Two semi-infinite square lattices (leads) are contacted by a region of length
$2L_x+1$ and width $2L_y+1$. The site potentials are equal to $-4t$, excepted in the contact region ($-L_x\leq i_x \leq L_x$) where the
potential at a site ${\bf i}$ of coordinates $(i_x,i_y)$ is taken infinite (green sites) if $|i_y| \geq (L_y-k)+k \left(i_x/L_x \right)^2$
($L_x=L_y=3$ and $k=2$) and in a single site (red) of coordinates $(x,y)$ and of potential $V-4t$ which describes the depletion region
induced by the charged tip.}
\label{fig2QPC}
\end{figure}
Before studying a QPC when $L_x$ and $L_y$ are large in section~\ref{QPC}, it is very instructive to study the limit shown
in Fig.~\ref{fig2RCM} where the contact is reduced to a single site ${\bf I}$ of coordinates $(0,0)$ and potential $-4t+V_{\bf I}$.
This defines the resonant contact model (RCM) which can be solved analytically~\cite{alp} when the width of the leads becomes infinite.
As one varies the energy $E$ for $t_c \ll 1$ and $V=0$, the RCM transmission $T_0(E)$ exhibits a single spin-degenerate Breit-Wigner
resonance, and not the usual staircase function which characterizes the QPC conductance quantization.
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig2RCM.pdf}
\caption{(Color online) RCM model for the SGM of a quantum contact: Two semi-infinite square lattices are contacted via a single
site ${\bf I}$ of coordinates $(0,0)$, potential $-4t+V_{\bf I}$ and coupling term $t_c$. Taking a potential $-4t+V$ at another
single site ${\bf T}$ of coordinates $(x,y)$ gives rise to an electron interferometer of size $r=\sqrt{x^2+y^2}$.}
\label{fig2RCM}
\end{figure}
\section{SGM of the Resonant Contact Model}
\subsection{RCM model: Spin-degenerate case at ${\cal T}=0$}
For the RCM contact without tip, the transmission of an electron of spin $\sigma$ and energy $E$ is given by
the Fisher-Lee formula \cite{Datta:book97}:
\begin{equation}\label{eq:FL}
T_0^{\sigma}(E) = \text{Tr} \left [\Gamma_l(E) G_{0}^R(E) \Gamma_r(E) G_{0}^A(E) \right ] \; ,
\end{equation}
where $G_{0}^R$ is the retarded Green's function of the contact dressed by the right (r) and left (l) leads:
\begin{equation}\label{eq:GF0}
G_{0}^R (E) = \underset{\eta \rightarrow 0^+}{\text{lim}} (E+i \eta - 4 - V_{\bf I} - \Sigma_l^R - \Sigma_r^R)^{-1} \; .
\end{equation}
The contact being reduced to a single site ${\bf I}$ coupled to another single site per lead, the lead self-energies
$\Sigma_{l,r}(E)$ are only two complex numbers $\Sigma_{l,r}(E)=R_{l,r}(E)+iI_{l,r}(E)=t_c^2<\pm 1,0|G_{l,r}^R(E)|\pm 1,0>$
when $t=-1$ and where $G^R_{l,r}(E)$ are the retarded Green's function of the left and right leads evaluated at the sites
directly coupled to ${\bf I}$. The coupling rates to the right and left leads verify: $\Gamma_{r,l} = i (\Sigma_{r,l}^R - \Sigma_{r,l}^A )$.
Using the method of mirror images~\cite{Molina:PRB06}, $G^R_{l,r}(E)$ can be expressed in terms of the Green's function
$G^R_{2D}(E)$ of the infinite 2D square lattice~\cite{Economou:book06}. One gets:
\begin{equation}
T_0^{\sigma}(E)= \frac{4 I_rI_l}{(E-4-V_{\bf I}-R_r-R_l)^2+(I_r+I_l)^2}.
\label{transmission_without_tip}
\end{equation}
If the variation of $\Sigma_{l,r}(E)$ can be neglected when $E$ varies inside the resonance (typically $t_c<0.5$ in the continuum limit
where the Fermi momentum $k_F \ll 1$), one gets a Lorentzian of width $\Gamma = -2I$ and center $4+V_{\bf I}+2R$ since $R_l=R_r \equiv R$
and $I_r=I_l\equiv I$.\\
\indent If one adds a tip potential $V \neq 0$ in the right lead, the effect of the tip can be included by adding an amount
$\Delta \Sigma_r(E)=\Delta R_{r}(E)+i\Delta I_{r}(E)$ to $\Sigma_r(E)$. The interferometer transmission $T^{\sigma}(E)$ is still given
by Eq.~\eqref{transmission_without_tip}, once $R_r+\Delta R_r$ and $I_r+\Delta I_r$ have been substituted for $R_r$ and $I_r$
(see Refs.~\cite{alp,Darancet:PRB10}). When the effect of the charged tip is restricted to a single site ${\bf T}$,
$\Delta \Sigma_r$ can be obtained from Dyson's equation for $G_{r+V}(E)$ the Green's function of the right lead with the tip potential:
\begin{eqnarray}
<1,0| G_{r+V}^R(E)|1,0>= <1,0| G_{r}^R(E)|1,0> \nonumber \\
+ \frac{<1,0| G_{r}^R(E)|{\bf T}> V <{\bf T}| G_{r}^R(E)|1,0>}{1-V <{\bf T}| G_{r}^R(E)|{\bf T}>}
\end{eqnarray}
In the continuum limit and for distances $r \gg k_F^{-1}$, one finds:
\begin{equation}
\frac{\Delta \Sigma_r}{t_c^2 \rho} \approx -\frac{k_F x^2}{2 \pi r^3} \exp[i(2k_Fr+\pi/2+\phi)] + O(\frac{x^{3/2}}{r^3}),
\label{self-energy}
\end{equation}
where $\rho$ and $\phi$ are the modulus and the phase of the amplitude of $V/(1-V \langle 0,0|G^R_{2D}(E)|0,0\rangle )$.
In the continuum limit, $I \approx -t_c^2 k_F^2/4$. Therefore, at sufficiently large distances
$r \gg \lambda_F/2$, $\Delta \Sigma_r^R(E) \ll I$ and one can expand $T^{\sigma}(E)$ to the leading order $\propto x^2/r^3$ in $\Delta \Sigma_r$:
\begin{eqnarray}
\dfrac{T^{\sigma}(E)-T^{\sigma}_{0}(E)}{T^{\sigma}_{0}(E)} \approx & -s \sqrt{T^{\sigma}_{0}(E) (1-T^{\sigma}_{0}(E))} \; \frac{\Delta R_r(E)}{I} \nonumber \\
& + (1- T^{\sigma}_{0}(E)) \; \frac{\Delta I_r(E)}{I} \; ,
\end{eqnarray}
where $s= \text{sign}[V_{\bf I}^{res}-V_{\bf I}]$ and $V^{res}_{\bf I}\equiv E-4-2R$ is the value of $V_{\bf I}$ where
$T^{\uparrow}_{0}=T^{\downarrow}_{0}=1$. This leads to the simple prediction:
\begin{eqnarray}
\dfrac{T^{\sigma}(E)-T^{\sigma}_{0}(E)}{T^{\sigma}_{0}(E)} \approx A_0 \cos (2 k_F r + \Phi_0) + O\left( \dfrac{x^{3/2}}{{r}^{3}}\right)\;
\label{SGM-T=0}
\end{eqnarray}
where the amplitude $A_0=\frac{2 {\rho}}{\pi k_F} \frac{x^2}{r^3} \sin \zeta_0 $ decreases as $\frac{x^2}{r^3}$, the
phase $\Phi_0= \pi/2 + {\phi} - \zeta_0$ and $\sin \zeta_0 = - s \sqrt{1- T^{\sigma}_{0}(E)}$. Eq.~\eqref{SGM-T=0} describes
Fabry-P\'erot fringes spaced by $\lambda_F/2$ and their decay with $r$, assuming $T^{\sigma}_{0}(E)<1$. One needs to take into account
corrections of higher order~\cite{alp} when $T^{\sigma}_{0}(E) \to 1$, a limit which we will not consider in this work.
\subsection{RCM model: Spin-degenerate case at ${\cal T} \neq 0$}
Let us now study how the effect of the tip upon the conductance $g$ in units of $e^2/h$ depends on the temperature ${\cal T}$:
\begin{equation}
\Delta g=g-g_0= \sum_{\sigma}\int dE (T^{\sigma}(E)-T^{\sigma}_{0}(E)) (-\frac{\partial f}{\partial E}),
\label{conductance-change}
\end{equation}
where $f$ is the Fermi-Dirac distribution. Let us consider the case of a sharp Lorentzian resonance ($t_c <0.5$) of $T^{\sigma}_{0}(E)$.
Then, $R$ and $I$ do not vary rapidly inside the resonance. In the same way, $\rho$ and $\phi$ vary slowly inside the resonance and thus
can be considered as constants. In order to calculate the integral \eqref{conductance-change} analytically, we make the approximation~\cite{Heller:NL05}
$-\partial f/ \partial E \approx (1/4k_B{\cal T}) \exp-[\sqrt{\pi}(E-E_F)/(4k_B{\cal T})]^2$, where $E_F$ and $k_B$ are
the Fermi energy and the Boltzmann constant. One gets $\Delta g = D_1+ D_2$ where:
\begin{widetext}
\begin{eqnarray}\label{eq:Temp-D1-D2-hors-res}
D_1 &\approx& \frac{2 \rho}{\pi^{3/2} k_F } \dfrac{x^2}{r^3}\frac{l_{\cal T}}{l_\Gamma} \;\; \Re \left[ e^{i (2 k_F r+ \phi+\pi/2)}
\int_{-\infty}^{\infty} \dfrac{q+v}{\left[1+(q+v)^2\right]^2} e^{-\left(\frac{l_{\cal T}}{l_\Gamma} q\right)^2} e^{i \frac{r}{l_\Gamma} q} \; \mathrm{d} q \right] \; , \\
D_2 &\approx & \frac{2 \rho_T}{\pi^{3/2} k_F } \dfrac{x^2}{r^3}\frac{l_{\cal T}}{l_\Gamma} \;\; \Im \left[ e^{i (2 k_F r+ \phi_T+\pi/2)}
\int_{-\infty}^{\infty} \dfrac{(q+v)^2}{\left[1+(q+v)^2\right]^2} e^{-\left(\frac{l_{\cal T}}{l_\Gamma} q\right)^2} e^{i \frac{r}{l_\Gamma} q} \; \mathrm{d} q \right] \; .
\end{eqnarray}
\end{widetext}
$v \equiv (V^{res}_{\bf I}-V_{\bf I})/\Gamma$ gives the energy shift of $V_{\bf I}$ from the resonance $V^{res}_{\bf I}$ in units of
$\Gamma$, and $q=(E-E_F)/\Gamma$. $l_{\cal T}$ and $l_{\Gamma}$ are the two length scales respectively associated to ${\cal T}$ (Fermi-Dirac
statistics) and to $\Gamma$ (resonant transmission):
\begin{eqnarray}
\label{scales}
l_{\cal T}&=&\frac{\sqrt{\pi}k_F}{4k_B{\cal T}} \\
l_{\Gamma}&=& \frac{k_F}{\Gamma}.
\end{eqnarray}
Calculating the Fourier transforms $D_1$ and $D_2$, we eventually obtain:
\begin{equation}\label{eq:deltaghorsres}
\Delta g ({\cal T})\approx 2 A({\cal T}) \cos (2k_Fr+\Phi({\cal T})) \;,
\end{equation}
where at large distance $r > r^*\equiv 2 l_{\cal T}[1+l_{\cal T}(1+|v|)/l_{\Gamma}]$:
\begin{equation}
A({\cal T})=\frac{\rho x^2 l_{{\cal T}}}{\sqrt{\pi}k_F r^3} \exp -[(1+v^2)(\frac{l_{\cal T}}{l_{\Gamma}})^2 +\frac{r}{l_{\Gamma}}]
\label{conductance1}
\end{equation}
\begin{equation}
\Phi({\cal T})=\phi+v\frac{r}{l_{\Gamma}}-2v(\frac{l_{\cal T}}{l_{\Gamma}})^2.
\label{conductance2}
\end{equation}
The factor $2$ in $\Delta g ({\cal T})$ comes from the spin degeneracy.
\subsection{RCM model: Effect of a parallel magnetic field applied everywhere}
Let us consider first the case where a uniform parallel magnetic field is applied everywhere. The spin degeneracy is broken
and the electrons of opposite spin have energies which are shifted ($E^{\sigma} \to E(h=0)\pm h$). They contribute to transport
with different wave-vectors in the 2DEG and a somewhat trivial beating effect is induced in the SGM images, the Fabry-P\'erot fringes
having different wavelengths ($\lambda^{\uparrow} \neq \lambda^{\downarrow}$). When $E_F$ is small enough, we can use the continuum dispersion
relation $k^{\sigma}=\sqrt{E^{\sigma}}$. If the Zeeman energy $h$ remains small compared to $E_F$, the 2DEG is not fully polarized. In
Fig.~\ref{fig4}, the conductance $g_0$ of the RCM contact is given as a function of the contact potential $V_I$. One can see how the spin
degenerate peak of conductance ($h=0$) of width $\Gamma$ is split by the field ($h=8 \Gamma$) for increasing values of ${\cal T}$.
Hereafter, we study the SGM image when the contact is open between the peaks (symmetric point indicated by an arrow in Fig.~\ref{fig4}
where $V_{\bf I}=V_{\bf I}^{res}(h=0)$).
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig4.pdf}
\caption{(Color online) Conductance $g_0$ (in units of $e^2/h$) of the RCM model without field ($h=0$, solid line) and
with a uniform parallel magnetic field ($h=8\Gamma$, dashed line) as a function of $V_{\bf I}$ for ${\cal T}=0$ (blue)
${\cal T}=\Gamma/2$ (green) and ${\cal T}=\Gamma$ (red). There is no tip, $\Gamma=0.003$ ($t_c=0.2$ and $E_F=0.15$)
and the field is applied everywhere. The arrow gives the symmetric point where the SGM images are studied.}
\label{fig4}
\end{figure}
The relative effect of the tip upon the transmission of an electron of spin $\sigma$ at an energy $E$ becomes,
\begin{eqnarray}
\dfrac{T^{\sigma}(E)-T^{\sigma}_{0}(E)}{T^{\sigma}_{0}(E)} \approx A_0^{\sigma} \cos (2 k_F^{\sigma} r + \Phi_0^{\sigma})
+ O\left( \dfrac{x^{3/2}}{{r}^{3}}\right)\;
\label{deltaT}
\end{eqnarray}
where $A_0^{\sigma}= (2 {\rho}^{\sigma})/(\pi k_F^{\sigma}) (x^2/r^3) \sin \zeta_0^{\sigma}$,
$\Phi_0^{\sigma}= \pi/2 + {\phi}^{\sigma} - \zeta_0^{\sigma}$,
$\sin \zeta_0^{\sigma} = - s^{\sigma} \sqrt{1- T^{\sigma}_{0}(E)}$ and $s^{\sigma}=\text{sign}(V_I^{res}-V_I)$.
If the contact is opened in the middle between the two transmission peaks, $s^{\uparrow}=-s^{\downarrow}$,
$\sin \zeta_0^{\uparrow}=-\sin \zeta_0^{\downarrow}$. The contribution of electrons of opposite spins to
$\sum_{\sigma} (T^{\sigma}(E)-T^{\sigma}_{0}(E))/T^{\sigma}_{0}(E)$ have opposite signs when the tip is put at a distance
\begin{eqnarray}
r^{\cal D}(n)=\frac{2\pi n +({\phi}^{\downarrow}-{\phi}^{\uparrow})+({\rho}^{\uparrow}-{\rho}^{\downarrow})}{2(k_F^{\uparrow}-k_F^{\downarrow})}\;,
\end{eqnarray}
where $n$ is integer ($0,1,2,\ldots$).
At a temperature ${\cal T}=0$, this means that the SGM image of a contact opened between its two transmission resonances exhibits a pattern
of rings of radii $r^{\cal D}$, where the beating between the contribution of opposite spin is destructive. In contrast, the beating becomes
constructive on rings of radii $r^{\cal C}$. Neglecting the small spin dependence of $T^{\sigma}_{0}$ at the symmetric point, these radii
become independent of the temperature ${\cal T}$ and reads
\begin{equation}
r^{\cal D}(n) \approx \frac{\pi n + \arcsin (\sqrt{1-T_0})}{\sqrt{E_F+h}-\sqrt{E_F-h}} \\
\label{radii}
\end{equation}
\begin{equation}
r^{\cal C}(n) \approx \frac{(n+1/2)\pi+ \arcsin (\sqrt{1-T_0})}{\sqrt{E_F+h}-\sqrt{E_F-h}}\;.
\label{radii-c}
\end{equation}
In Fig.~\ref{fig5}, one can see a SGM image taken with a parallel magnetic field at ${\cal T}=0$: We can see the three first rings
at the expected radii $r^{\cal D}(n)$ with $n=0,1,2$ (dashed lines), where the effect of the tip upon $g$ is suppressed, separated
by regions centered around rings of radii $r^{\cal C}(n)$ where this effect is enhanced by the applied magnetic field.
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig5.pdf}
\caption{(Color online) (Color online) RCM model with a parallel magnetic field applied everywhere ($h=4\Gamma\approx 0.0119$) at a temperature
${\cal T}=0$. The relative effect $\Delta T/T_0$ (upper color scale) of the tip ($V=-2$) upon the RCM transmission $T_0$ has been
numerically calculated and is given as a function of the tip coordinates $(x,y)$ when $t_c=0.2$ and $E_F=0.15$. The contact potential
has the value $V_I$ indicated by the arrow in Fig.~\ref{fig4} ($T_0=0.12$ and $\Gamma\approx 0.003$). The dashed lines give the rings
of radii $r^{\cal D}(n)$ predicted by Eq.~\eqref{radii}.}
\label{fig5}
\end{figure}
When the temperature ${\cal T} \neq 0$ but satisfies the condition $k_B{\cal T}/\Gamma \ll v$, the expressions can be simplified
if $r \ll 2 l_{\cal T} \left[1+(l_{\cal T}/l_{\Gamma})(1+|v|)\right]$. One finds that the SGM images are roughly identical to those
described by Eq.~\eqref{deltaT} for ${\cal T}=0$ within a circle of radius $l_{\cal T}$, and are suppressed outside (see Fig.~\ref{fig6})
\begin{equation}
\label{deltagwithT}
\Delta g ({\cal T},h) \approx \Delta g ({\cal T}=0,h) \exp-(\frac{r}{2l_{\cal T}})^2 \;.
\end{equation}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig6.pdf}
\caption{(Color online)
RCM contact with a parallel magnetic field applied everywhere ($h=4\Gamma\approx 0.0119$): The numerical
values of $\Delta g$ are given as a function of the tip coordinates $(x,y=0)$ for $E_F=0.15$, $V=-2$, $t_c=0.2$ and
$\Gamma\approx 0.003$. The figures (a) ${\cal T}=0$, (b) $k_B{\cal T}=10^{-4}$, (c) $k_B{\cal T}=8\times 10^{-4}$ are in the
regime $k_B{\cal T}/{\Gamma} \ll v$. The location $r^{\cal D}(n)$ of the destructive interferences is independent of the temperature
${\cal T}$, in agreement with Eq.~\eqref{deltagwithT}.}
\label{fig6}
\end{figure}
\subsection{RCM model: Effect of a parallel magnetic field applied upon the contact}
Let us consider now the case where a parallel magnetic field is applied upon the contact only.
This removes the spin degeneracy in the contact by a local Zeeman term $\pm h$.
In contrast to the previous case, $k_F$ and hence $l_{\cal T}$ and $l_{\Gamma}$ remain independent of $\sigma$, while $\Delta \phi=0$.
There is nevertheless a beating effect between the interference fringes of opposite spins, which exhibits a more unusual
temperature dependence than before: It can be observed only when the temperature becomes of the order of the Zeeman splitting,
but vanishes as ${\cal T}\to 0$.\\
\indent Let us consider the value of $V_{\bf I}$ (symmetric point) where there is the resonance without field
(total transmission $T_0(h=0)=2$), one has $v_{\uparrow}=-v_{\downarrow}$ and Eq.~\eqref{eq:deltaghorsres} gives
\begin{equation}
r^{\cal D}(n)=\frac{2k_F}{\Gamma}(\frac{l_{\cal T}}{l_{\Gamma}})^2+(n+\frac{1}{2}) \frac{\pi k_F}{h}
\label{radius-ring}
\end{equation}
($n=0,1,\ldots$) for the radii of the rings where the effect of the tip is suppressed by the field. Conversely, the oscillations of
$\Delta g^{\uparrow}({\cal T})$ and $\Delta g^{\downarrow}({\cal T})$ add if the distance $r$ is given by $r^{\cal C}(n)=r^{\cal D}(n)+\pi k_F/(2h)$.
The SGM image is characterized by a first ring at a distance $r^{\cal D}(n=0)$ followed by other rings spaced by $\pi k_F/h$ where
$\Delta g ({\cal T})=0$. To optimize the contrast in the images, we calculate for a given value of $h$ the temperature ${\cal T}^*$ and the width
$\Gamma^*$ for which $\Delta g({\cal T},r=r^{\cal C}(n=0))$ is maximum. The extrema are given by the conditions $\partial A / \partial l_{\Gamma}=0$
and $\partial A / \partial l_{\cal T} =0$. This gives two coupled non-linear algebraic equations which can be solved numerically, yielding
\begin{equation}
k_B{\cal T}^* \approx 0.73 h \ \ \ \Gamma^*\approx 0.25 h.
\end{equation}
To observe the rings, their spacing must exceed $\lambda_F$ ($h<(\pi/\lambda_F)^2$). In Fig.~\ref{fig8}, a numerical calculation
of an SGM image is shown when ${\cal T}={\cal T}^*$ and $\Gamma=\Gamma^*$. In the presence of a Zeeman term $h$ in the contact,
the image exhibits the ring pattern predicted by the theory.
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig8.pdf}
\caption{(Color online)
RCM model with a Zeeman term $h= \pm 0.0136$ in the contact only at a temperature ${\cal T}^*=0.0099/k_B$ ($l_{\cal T}=17.34$).
The numerically calculated values of $\Delta g/g_0 ({\cal T}^*,\Gamma^*)$ are plotted as a function of the coordinates $(x,y)$
of the tip (potential $V=-2$). $\Gamma^*=0.0035$. The parameters have been chosen such that the radius (Eq.~\eqref{radius-ring}) of
the first ring $r^D(n=0)=50$. The dashed lines give the circles of radii $r^{\cal D}(n)$ predicted by the theory (Eq.~\ref{radius-ring}).}
\label{fig8}
\end{figure}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig11.pdf}
\caption{(Color online)
Radii $r^{\cal D}(n)$ of the successive rings ($n=0, 1, 2$ and $3$) when a local field $h$ is applied in the contact and
$\Gamma=0.003$: Left figure: Radii $r^{\cal D}(n)$ as a function of $k_B{\cal T}$ when $h=0.01$. Right figure: $r^{\cal D}(n)$ as a function of
$h$ when $k_B{\cal T}=0.0028$. The dots give the successive radii where the numerically calculated values of $\Delta g/g_0 \approx 0$, their colors
corresponding to a visibility scale indicated at the right ($0$ without contrast, $1$ for the best contrast). The solid lines give the analytical
expression~\eqref{radius-ring} of the radii $r^{\cal D}(n)$ which was derived assuming $r^{\cal D}(n)>r^*$.}
\label{fig11}
\end{figure}
Fig.~\ref{fig11} shows us how the radii $r^{\cal D}(n)$ of the successive rings where the tip does not change the conductance depend on the
temperature ${\cal T}$ for a given local field $h$, or on the field $h$ for a given temperature ${\cal T}$. For a resonance
width $\Gamma=0.003$, one can see how the $r^{\cal D}(n)$ increase when ${\cal T} \to 0$ or $h \to 0$, making impossible the observation
of a beating effect in those limits. In Fig.~\ref{fig9}, the relative change $\Delta g(x,y=0)/g_0$ of the conductance is shown in the presence
of a Zeeman term in the contact as one varies the tip coordinate $x$ (keeping $y=0$). This change is given for three different values of $h$,
when the temperature ${\cal T}$ and the resonance width $\Gamma$ take their optimal values ${\cal T}^*$ and $\Gamma^*$.
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig9.pdf}
\caption{(Color online) $\Delta g/g_0 ({\cal T}^*,\Gamma^*,y=0)$ as a function of tip coordinate $x$ (keeping $y=0$) for three values $h$
of a Zeeman term added in the RCM contact only: $V=-2$, $E_F=0.1542$ and $\lambda_F/2=8$. (a): $h=0.0136$, $k_B{\cal T}^*=0.0099$, $\Gamma^*=0.0035$
(b): $h=0.0091$, $k_B{\cal T}^*=0.0066$, $\Gamma^*=0.0023$; (c): $h=0.0068$, $k_B{\cal T}^*=0.005$, $\Gamma^*=0.0017$.}
\label{fig9}
\end{figure}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig10-a.pdf}
\caption{(Color online) At ${\cal T}=0$, difference between the effect of a local (a) or global (b) field $h=0.0136$:
$\Delta g (x,y=0)$ is plotted as a function of $x$ for a tip potential $V=-2$, $E_F=0.1542$, $\Gamma^*=0.0035$ and $\lambda_F/2=8$.}
\label{fig10-a}
\end{figure}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig10-b.pdf}
\caption{(Color online) RCM contact with parallel magnetic field $h=0.0136$, $E_F=0.1542$, $\Gamma^*=0.0035$, $V=-2$, and $\lambda_F/2=8$:
At ${\cal T}={\cal T}^*=0.0099/k_B$, $\Delta g (x,y=0)$ is plotted as a function of the tip coordinate $x$ keeping $y=0$. (a): Magnetic field
inside the contact only ($k_F^{\uparrow}= k_F^{\downarrow}$). (b): Magnetic field everywhere ($k_F^{\uparrow} \neq k_F^{\downarrow}$).
The difference of periodicity and frequency are mainly due to the field dependence of $k_F^{\sigma}$.}
\label{fig10-b}
\end{figure}
Let us underline the difference between the effect of a field restricted to the contact or applied everywhere as ${\cal T}\to 0$.
In the first case (see Eq.~\eqref{radius-ring}), the first ring has a radius $r^{\cal D}(n=0) \to \infty$ and cannot be seen.
It is only when one increases ${\cal T} \to {\cal T}^*$ that $r^{\cal D}(n=0)$ becomes small enough and that the rings can be seen.
In the second case, the rings do not depend on ${\cal T}$ when $r \ll l_{\cal T}$ (see Eq.~\eqref{deltagwithT})
and remains visible as ${\cal T}\to 0$. Numerical illustrations of the differences between the SGM images obtained when a field
is either applied everywhere or restricted to the contact are given in Fig.~\ref{fig10-a} when ${\cal T} = 0$ and in Fig.~\ref{fig10-b}
when ${\cal T}={\cal T}^*$: In Fig.~\ref{fig10-a}, one can notice the absence of a beating effect when the field is applied only inside
the contact, while it can be seen if the field is applied everywhere. In Fig.~\ref{fig10-b}, the beating effect is visible in the two
cases.
\section{SGM of a double-dot setup}
Instead of using a magnetic field for breaking the spin degeneracy of a single resonance, let us now show that
a contact having a double-peak structure of its transmission without magnetic field gives rise also to a similar
beating pattern when it is opened between the peaks. Let us take a contact made of two sites of potentials $V_{I}-4t$
coupled by an hopping term $t_d$ (see Fig.~\ref{fig12_1}). This gives rise to a two-level system which is often
used~\cite{revmodphys} to describe electron transport through double quantum dots. As before, the depletion region
induced by the charged tip is described by a single scattering site and we take again $t=-1$.
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig12_1.pdf}
\caption{(Color online) Scheme of the SGM of a contact made of a double-dot setup.}
\label{fig12_1}
\end{figure}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig12_2.pdf}
\caption{(Color online) Double-dot setup without tip ($V=0$): Conductance $g_0$ (in units of $2e^2/h$) with $t_c=0.2$ and $\Gamma=0.0015$
as a function of the dot potentials $V_{\bf I}$ for $t_d=0.012$ (blue) $t_d=0.006$ (green) and $t_d=0.002$ (red).
${\cal T}=0$ (solid line) and $k_B{\cal T}=\Gamma/2$ (dashed line). The arrow gives the potential $V_{\bf I}$ for the SGM study.}
\label{fig12_2}
\end{figure}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig13.pdf}
\caption{(Color online) Double-dot setup - Main Fig: $\Delta g/g_0$ as a function of the tip coordinates $(x,y)$ at a temperature
${\cal T}={\cal T}^*=0.0022/k_B$ where $l_{\cal T}=78$ ($g_0=0.3064$, $\Gamma=0.0015$, $V=-2$, $t_d=0.006$, $t_c=0.2$, $E_F=0.1542$ and $\lambda_F/2=8$).
The dashed lines give the circles of radii $r^{\cal D}(n)$ predicted by the theory (Eq.~\eqref{radius-ring} after making the changes
$h \to t_d$ and $\Gamma \to \Gamma/2$). Above: Color code giving the magnitude of the relative effect.
Below: $\Delta g/g_0(y=0)$ as a function of $x$ (same parameters as in the main figure).}
\label{fig13}
\end{figure}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig14.pdf}
\caption{(Color online) Double-dot setup with $\Gamma=0.003$: Radii $r^{\cal D}(n)$ of the interference rings for $n=0, 1, 2$ and $3$
as a function of $k_B{\cal T}$ (left fig. $t_d=0.01$) and of the interdot coupling $t_d$ (right fig. $k_B{\cal T}=0.0028$). The dots
give the successive values of the radii where the numerically calculated values of $\Delta g/g_0 \approx 0$, the colors corresponding
to a visibility scale indicated at the right. The solid lines are the analytical values of $r^{\cal D}(n)$ derived assuming $r^{\cal D}(n)>r^*$.}
\label{fig14}
\end{figure}
The Green's function of this model is now given by a $2 \times 2$ matrix which reads
\begin{equation*}
G(E)=
\begin{bmatrix}
E-4-V_{I}-\Sigma_l(E) & -t_d \\
-t_d & E-4-V_{I}-\Sigma_r(E)-\Delta\Sigma_r \\
\end{bmatrix}
\end{equation*}
where $\Sigma_{r,l}=R+iI$ are the lead self-energies and $\Delta\Sigma_r$ the change induced on $\Sigma_r$ by the tip.
Without tip, the transmission of an electron of spin $\sigma$ reads
\begin{equation}
T_0^{\sigma}(E)= \frac{t_d}{{\tilde E}}\left(\frac{I^2}{(t_d-{\tilde E})^2+I^2}-\frac{I^2}{(t_d-{\tilde E})^2+I^2}\right)
\label{2QT-transmission_without_tip}
\end{equation}
where ${\tilde E}=E-4-V_{\bf I}-R$.
This gives two peaks of equal height and width $\Gamma/2=-I$ (instead of $\Gamma$ for the RCM model)
which are spaced by a ``Zeeman energy'' $2 t_d$ (instead of $2h$ for the RCM model with a field restricted to the contact).
Of course, the transmission of the double-dot setup is suppressed when the inter-dot coupling $t_d \to 0$, in contrast to the
RCM model through which the electrons are transmitted when $h \to 0$.\\
\indent These similarities are a consequence of the inversion symmetry of the double-dot model. If the two sites of the contact have
respective coordinates $(0,0)$ and $(1,0)$, one can rewrite the Hamiltonian of the double-dot setup in terms of fermion operators
which destroy/create an electron of spin $\sigma$ in an even/odd combination of two symmetric orbitals of the original model.
For instance, for the destruction operators of a particle with spin $\sigma$ and pseudospin $e,o$, one has
\begin{eqnarray}
a^{\sigma,o}_{(x,y)}=(c^{\sigma}_{(-x+1,-y)}-c^{\sigma}_{(x,y)})/\sqrt{2}\\
a^{\sigma,e}_{(x,y)}=(c^{\sigma}_{(-x+1,-y)}+c^{\sigma}_{(x,y)})/\sqrt{2}.
\end{eqnarray}
This allows us to map the original model (electrons with spins free to move on two semi-infinite square lattices coupled by
two sites of potential $V_{I}$) onto a transformed model of electrons with spins and pseudo-spins (even or odd states)
free to move on a single semi-infinite square lattice coupled by a single hopping term $t_c$ to a single site of potential
$V_{I} \pm t_d$ ($+ t_d$ for the odd states, $-t_d$ for the even states). In that sense, $t_d$ can be seen
indeed as a ``pseudo Zeeman energy'' which removes the pseudo-spin degeneracy. For more details about this mapping, we
refer the reader to Ref.~\cite{fp1}, where an inversion symmetric one dimensional model was studied, the extension to two dimensions
being straightforward.\\
\indent The similarity between the two models is also evident when one compares the conductance $g_0$ (in units of $2e^2/h$)
of the double-dot setup shown in Fig.~\ref{fig12_2} for different values of $t_d$ and ${\cal T}$ with the conductance
$g_0$ (in units of $e^2/h$) of the RCM model for different values of $h$ and ${\cal T}$. This leads us to expect that their SGM
images must be also similar: When $t_d \neq 0$, the SGM images of the double-dot setup should also exhibit rings where the effect
of the tip does not change the conductance of the contact, the radii of these rings being given by Eq.~\eqref{radius-ring},
after making the changes $h \to t_d$ and $\Gamma \to \Gamma/2$. This replacement implies also that the ring spacing is equal to
$\pi k_F/t_d$ for the double-dot setup. A numerical check of such a prediction is given in Fig.~\ref{fig13} (which shows that the
SGM of the double-dot setup gives rise to a similar pattern of rings) and in Fig.~\ref{fig14} (which shows that the location of
the rings is indeed given by Eq.~\eqref{radius-ring} when one puts $t_d$ instead of $h$ and $\Gamma/2$ instead of $\Gamma$).
\section{SGM of a QPC with quantized conductance plateaus}
\label{QPC}
\begin{figure}
\includegraphics[keepaspectratio,width=\columnwidth]{Nfig16.pdf}
\caption{(Color online) Lattice model (see Fig.~\ref{fig2QPC}) of a QPC without tip ($V=0$) with $L_y=20$, $L_x=16$ and $k=2$:
Transmission $T_{\sigma}^0$ of an electron of spin $\sigma$ and total transmission $T^{0}=\sum_{\sigma}T_{\sigma}^0$ as a function
of $E$ when a parallel magnetic field $h=0.0025$ is applied everywhere (case 1) and only upon the contact region $-L_x \leq x \leq L_x$
(case 2). The arrows give energies around which beating effects can be seen.}
\label{fig16}
\end{figure}
\begin{figure*}
\includegraphics[keepaspectratio,width=\textwidth]{Nfig17.pdf}
\caption{(Color online). SGM images of a QPC at the opening of the second transmission channel ($E_F=0.027$, left arrow of Fig.~\ref{fig16})
at ${\cal T}=0$ (top) and $k_B{\cal T}=0.006$ (bottom). A parallel magnetic field $h=0.0025$ is applied everywhere (left) or only upon the contact
region (right). For a better visibility, we have plotted as a function of the tip coordinates $x,y$, $\frac{\delta (\Delta g)}{\delta x} \times r$
when ${\cal T}=0$ and $\frac{\delta (\Delta g)}{\delta x} \times r^2$ when $k_B{\cal T}=0.006$.}
\label{fig17}
\end{figure*}
We now consider larger contacts, able to have more than one open transmission channel. The energy dependence of their transmission
is given by a staircase function, in contrast to the single Breit-Wigner resonance of the RCM model without field. Such staircase
functions with quantized conductance plateaus are observed in the setups sketched in Fig.~\ref{fig1} where the gate potentials give
rise to a smooth saddle-point QPC potential. These staircase functions characterize also the lattice model for a QPC sketched in
fig.~\ref{fig2QPC}, as one can see in Fig.~\ref{fig16} in the cases where a parallel magnetic field is applied either everywhere,
or only inside the contact region.
An analytical approach being more difficult, we numerically study these larger contacts. The self-energies
$\Sigma_{l,r}(E)$ are now two matrices of size $(2 L_y+1) \times (2 L_y+1)$, where $L_y$ defines the width of
the contact region (see Fig.~\ref{fig2QPC}). Their analytical expressions without tip can be found in
Ref.~\cite{Datta:book97}. To include the effect of the tip which modifies the potential of a single site in the right
lead, we use again Dyson equation ($V$ playing the role of a perturbation), extending the method used for the RCM model.
The usual recursive numerical method for calculating the Green function is only used for the contact region of width
$\leq 2 L_y+1$. With this method, we study leads of very large transverse width ${\cal L}_y \approx 2. 10^4$
and a contact region of size $L_x=16$ and $L_y=20$ where the potential $V_{\bf i}$ of a contact site ${\bf i}$ of coordinates
$(i_x,i_y)$ is taken infinite if $|i_y| \geq (L_y-k)+k \left(i_x/L_x \right)^2$. Once the self energies $\Sigma_{l,r}(E)$
of the leads are obtained, the total interferometer transmission $T(E)$ is calculated.
The channel openings of these QPCs play the role of the resonances of the RCM contact~\cite{alp} and one can also use a
parallel magnetic field to split the QPC channel openings. The effect of a parallel magnetic field upon the SGM images
of a QPC is shown in Fig.~\ref{fig16}, both when it is applied everywhere or restricted to the contact region. For a QPC
opened in the energy interval where a new channel is opened for the electrons with parallel spin, but not yet for those
with antiparallel spins, the effect of the tip upon the QPC conductance should exhibit a beating between the two spin
contributions of this new channels. This is indeed what can be seen in Fig.~\ref{fig17} when the contact is
open around the second channel opening (see the arrow in Fig.~\ref{fig16}). The effect of the tip has the V-shape
which characterizes this second channel. To make the SGM images clearer, we have plotted the effect of the tip over
the conductance derivatives with respect to $x$. In panels (a) and (b) of Fig.~\ref{fig17}, the SGM images are taken at
${\cal T}=0$. When the field is applied everywhere (a), one can see a beating effect between the two spin contributions of
the second channel. When the field is restricted to the contact region (b), there is no beating effect at ${\cal T}=0$.
In panels (c) and (d), the SGM images are taken at a temperature where $k_B{\cal T}=0.006$, and one recovers a beating effect
when the field is applied only in the contact region: The behaviors which were previously explained using the RCM contact
can be qualitatively extended to a QPC with quantized conduction modes. We observed similar beating patterns near the
opening of the other channels.
\section{Conclusion}
In summary, we have studied contacts open around spin-degenerate resonances or channel openings.
Breaking the spin degeneracy by a magnetic field $h$ (or the pseudo-spin degeneracy by an inter-dot
coupling $t_d$), we have shown that the SGM images of contacts open between split resonances
or channel openings exhibit a beating pattern of interference fringes where the effect of the tip
is suppressed. The spacing of the Fabry-P\'erot fringes being $\lambda_F/2$, SGM provides a method for
measuring by electron interferometry both $\lambda_F$ and either the magnetic field $h$ (RCM-contact)
or the inter-dot coupling $t_d$ (double-dot setup). We have shown that the spacing between the rings
is given respectively by $\pi k_F/h$ and $\pi k_F/t_d$.\\
\indent If the magnetic field is applied everywhere, these rings can be observed at zero temperature. If the field is
applied only upon the contact, the rings are sufficiently close to the contact for being observable only if the
temperature is close to an optimum temperature ${\cal T}^*$ where $k_B{\cal T}^*$ is of order of the peak splitting.\\
\indent These beating effects can be seen in the SGM images of the RCM model (with a single resonance) or of a QPC (with
quantized conductance plateaus), when the spin degeneracy is removed by a field. The difference between a local or
a global Zeeman effect which was analytically shown using the RCM model can be also seen in the numerical sudies of QPCs.\\
\acknowledgments{This research has been supported by the EU Marie Curie network ``NanoCTM'' (project no.234970).
Discussions with B. Brun, K. Ensslin, M. Sanquer and H. Sellier about SGM experiments and
with C. Gorini, R. Jalabert and D. Weinmann about SGM theory are gratefully acknowledged.}
|
\section{Introduction}
\label{s:intro}
Following the overly optimistic expectations of early research, after several decades of intense research on the one hand and relatively disappointing practical results on the other, face recognition technology has been finally started to enjoy some success in the consumer market. An example can be found within the online photo sharing platform in Google Plus which automatically recognizes individuals in photographs based on previous labelling by the user. It is revealing to observe that the recent success of face recognition has been in the realm of data and image retrieval \cite{Shan2010}, rather than security, contrasting the most often stated source of practical motivation driving past research \cite{ChelWilsSiro1995}. This partial paradigm shift is only a recent phenomenon. In hindsight, that success would first be achieved in the domain of retrieval should not come as a surprise considering the relatively low importance of type-II errors in this context: the user is typically interested in only a few of the retrieved matches and the consequences of the presence of false positives is benign, more often being mere inconvenience.
Nevertheless, the appeal of face recognition as a biometric means that the interest in its security applications is not waning. Face recognition can be performed from a distance and without the knowledge of the person being recognized, and is more readily accepted by the general public in comparison with other biometrics which are regarded as more intrusive. In addition, the acquisition of data for face recognition can be performed readily and cheaply, using widely available devices. However, although the interest in security uses of face recognition continues, it has become increasingly the case that research has shifted from the use of face recognition as a sole biometric. Instead, the operational paradigm is that of face recognition as an element of a multi-biometric system \cite{MuraIwamMakiYagi2013}.
Of particular interest to us is the use of infrared imaging as a modality complementary to the `conventional', visible spectrum-based face recognition. Our focus is motivated by several observations. Firstly, in principle an IR image of a face can be acquired whenever a conventional image of a face can. This may not be the case with other biometrics (gait, height etc). Secondly, while certainly neither as cheap nor as widely available as conventional cameras, in recent years IR imagers have become far more viable for pervasive use. It is interesting to note the self-enforcing nature of this phenomenon: the initial technology advancement-driven drop in price has increased the use of IR cameras thereby making their production more profitable and in turn lowering their cost even further.
In a distal sense, the key challenges in IR-based face recognition remain to be pose invariance, robustness to physiological conditions affecting facial IR emissions, and occlusions due to facial hair and accessories (most notably eyeglasses). In a more proximal sense, as argued in a recent review \cite{GhiaAranBendMald2013b}, the main challenge is to formulate a framework which is capable of dealing with all of the aforementioned factors affecting IR `appearance' in a unified manner. While a large number of IR-based face recognition algorithms have been described in the literature, without exception they all constrain their attention to a few, usually only a single, extrinsic factor (e.g.\ facial expression or pose). None of them can cope well with a concurrent variability in several extrinsic factors. Yet, this is the challenge encountered in practice.
In this paper our aim is to describe what we believe to be the first IR-based face recognition method which is able to deal with all of the major practical challenges. Specifically, our method explicitly addresses (i) the variability in the user's physiological state, (ii) pose changes, (iii) facial expressions, (iv) partial occlusion due to prescription glasses, and (v) quasi-occlusion due to facial hair.
\section{Unified treatment of extrinsic variability}\label{s:main}
In this section we detail different elements of our system. We start by describing the Dual Dimension Active Appearance Model Ensemble framework and then demonstrate how it can be extended to perform model selection which allows for the handling of partial occlusion due to prescription glasses, and quasi-occlusion due to facial hair.
\subsection{Dual Dimension AAM Ensemble (DDAE)}
At the coarsest level, the Dual Dimension Active Appearance Model Ensemble algorithm comprises three distinct components. These are: (i) a method for fitting an active appearance model (AAM) \cite{GrosMattBake2006} particularly designed for fast convergence and reliable fitting in the IR spectrum, (ii) a method for selecting the most appropriate AAM from a trained ensemble, and (iii) the underlying extraction of person-specific discriminative information. The ultimate functions of these elements within the system as a whole are respectively pose normalization within a limited yaw range, invariance across the full range of yaw, and invariance to physiological changes of the user.
\vspace{-10pt}
\subsubsection{AAM fitting}
There are two crucial design aspects of the design and deployment of AAMs that need to be considered in order to ensure their robustness. These are: (i) the model initialization procedure, and (ii) the subsequent iterative refinement of model parameters. In the context of the problem considered in this paper, the former is relatively simple. Given that we are using thermal imaging background clutter is virtually non-existant applying simple thresholding to the input image allows the face to be localized and its spatial extent estimated. Reliable initialization of the AAM is then readily achieved by appropriately positioning its centroid and scale. A much greater challenge in the use of the AAM model for the normalization of pose and facial expression concerns the subsequent convergence -- the model is notoriously prone to convergence towards a local mininum, possibly far from the correct solution. This problem is even more pronounced when fitting is performed on face images acquired in the IR spectrum; unlike in the visible spectrum, in thermal IR human faces lack face-specific detail that is crucial in directing iterative optimization. This is the likely explanation for the absence of published work on the use of AAMs for faces in IR until the work of Ghiass \textit{et al.}~\cite{GhiaAranBendMald2013}. Their key idea was to perform fitting by learning and applying an AAM not on raw IR images themselves but rather on images automatically processed in a manner which emphasizes high-frequency detail. Specifically the detail-enhanced image $I_e$ is computed by anisotropically diffusing the input image $I$:
{\small\begin{align}
\frac{\partial I}{\partial t} = \nabla.\left( c(\| \nabla I\|)~\nabla I\right) = \nabla c.\nabla I + c(\| \nabla I\|)~\Delta I,
\end{align}}
using a spatially varying and image gradient magnitude-dependent parameter $c(\| \nabla I\|) = \exp \left\{ -\|\nabla I\|/400 \right\}$, subtracting the result from the original image and applying histogram equalization $I_e = \text{histeq}(I - I_d)$. Warped examples are shown in Fig.~\ref{f:aamExamp}.
\begin{SCfigure
\vspace{-20pt}
\centering
\includegraphics[width=0.4\textwidth]{aam_examples.png}
\caption{Examples of automatically pre-processed thermal images warped to the canonical geometric frame using the described AAM fitting method. }
\label{f:aamExamp}
\vspace{-10pt}
\end{SCfigure}
\subsubsection{Person and pose-specific model selection}
Applied to faces, the active appearance model comprises a triangular mesh with each triangle describing a small surface patch. To use this model for the normalization of pose, it is implicitly assumed that the surface patches are approximately planar. While this approximation typically does not cause fitting problems when pose variation is small this is not the case when fitting needs to be performed across a large range of poses (e.g.\ ranging from fully frontal to fully profile orientation relative to the camera's viewing direction). This is particularly pronounced when the applied AAM is generic (rather than person-specific) and needs to have sufficient power to describe the full scope of shape and appearance variability across faces. The DDAE method overcomes these problems by employing an ensemble of AAMs. Each AAM in an ensemble `specializes' to a particular region of IR face space. The space is effectively partitioned both by pose and the amount of appearance variability, making each AAM specific to a particular range of poses and individuals of relatively similar appearance (in the IR spectrum). In training, this is achieved by first dividing the training data corpus into pose-specific groups and then applying appearance clustering on IR faces within each group. A single AAM in an ensemble is trained using a single cluster. On the other hand, when the system is queried with a novel image containing a face in an arbitrary and unknown pose, the appropriate AAM from the ensemble needs to be selected automatically. This can be readily done by fitting each AAM from the ensemble and then selecting the best AAM as the one with the greatest maximal likelihood, that is, the likelihood achieved after convergence.
\vspace{-15pt}
\subsubsection{Discriminative representation}
A major challenge to IR-based face recognition in practice emerges as a consequence of thermal IR appearance dependence on the physiological state of the user. Factors which affect the sympathetic or parasympathetic nervous system output (e.g.\ exercise or excitement) or peripheral blood flow (e.g.\ ambient temperature or drugs) can have a profound effect. This is illustrated in Fig.~\ref{f:phys}. With the notable exception of the work by Buddharaju \textit{et al.} \cite{BuddPavlTsiaBaza2007} there has been little research done on developing an IR-based person-specific representation invariant to these changes.
\begin{figure}[thp]
\centering
\subfigure[Usr~1, seq~1]{~~\includegraphics[width=0.14\textwidth]{user1_before.png}~~}
\subfigure[Usr~1, seq~2]{~~\includegraphics[width=0.14\textwidth]{user1_after.png}~~}~~~
\subfigure[Usr~2, seq~1]{~~\includegraphics[width=0.14\textwidth]{user2_before.png}~~}
\subfigure[Usr~2, seq~2]{~~\includegraphics[width=0.14\textwidth]{user2_after.png}~~}
\vspace{-5pt}
\caption{Thermal IR appearance before (`seq 1') and after (`seq 2') mild exercise. }
\label{f:phys}
\vspace{-10pt}
\end{figure}
We adopt the use of a representation which is dependent on the distribution of superficial blood vessels. Unlike the vessel network based representation \cite{BuddPavlTsiaBaza2007} our representation is not binary i.e.\ a particular image pixel is not merely classified as a blood vessel or not. Rather, the extracted pseudo-probability map allows for the level of confidence regarding the saliency of a particular pixel to be encoded. Additionally, unlike the representation extracted by various blood perfusion methods \cite{SealNasiBhatBasu2011}, our representation is based on temperature gradients, rather than the absolute temperature. As such, it is less affected by physiological changes which influence the amount of blood flow, and is rather a function of the invariant distribution of underlying blood vessels.
Our representation is extracted using the so-called vesselness filter \cite{FranNiesVincVier1998}, originally developed for use on 3D MRI data for the extraction of tubular structures from images. Just as in 3D, in 2D this is achieved by considering the of the Hessian matrix $H(I,x,y,s)$ computed at the image locus $(x,y)$ and at the scale $s$:
{\small\begin{align}
H(I,x,y,s)=\begin{bmatrix}
L_x^2(x,y,\sigma) & L_x L_y (x,y,\sigma)\\
L_x L_y (x,y,\sigma) & L_y^2(x,y,\sigma) \\
\end{bmatrix}
\end{align}}
where $L_x^2(x,y,s)$, $L_y^2(x,y,s)$, and $L_x L_y (x,y,\sigma)$ are the second derivatives of $L(x,y,s)$, resulting from the smoothing the original image $I(x,y)$ with a Gaussian kernel $L(x,y,s) = I(x,y) \ast G(s)$. If the two eigenvalues of $H(I,x,y,s)$ are $\lambda_1$ and $\lambda_2$ and if without loss of generality we take $|\lambda_1| \leq |\lambda_2|$, two statistics which characterize local appearance and which can be used to quantify the local vesselness of appearance are $\mathcal{R}_\mathcal{A} = |\lambda_1|/|\lambda_2|$ and $\mathcal{S} = \sqrt{\lambda_1^2 + \lambda_2^2}$. The former of these measures the degree of local 'blobiness' which should be low for tubular structures, while $\mathcal{S}$ rejects nearly uniform, uninformative image regions which are characterized by small eigenvalues of the Hessian. For a particular scale of image analysis $s$, the two measures, $\mathcal{R}_\mathcal{A}$ and $\mathcal{S}$, are unified into a single measure $\mathcal{V}(s) = (1-e^{-\frac{\mathcal{R}_\mathcal{A}}{2\beta^2}}) \times (1-e^{-\frac{\mathcal{S}}{2c^2}})$ for $\lambda_2 > 0$ and $\mathcal{V}(s) =0$ otherwise, where $\beta$ and $c$ are the parameters that control the sensitivity of the filter to $\mathcal{R}_\mathcal{A}$ and $\mathcal{S}$. The overall vesselness of a particular image locus can be computed as the maximal vesselness across scale $\mathcal{V}_0 = \max_{s_{min} \leq s \leq s_{max}} \mathcal{V}(s)$.
\subsection{Robustness to eye-wear and facial hair changes}
A significant challenge posed to face recognition algorithms, conventional and IR-based ones alike, is posed by occlusions \cite{GrosMattBake2006,Mart2002}. Of particular interest to us are specific commonly encountered occlusions -- these are occlusions due to prescription glasses \cite{HeoKongAbidAbid2004} (for practical purposes nearly entirely opaque to the IR frequencies in the short, medium and long wave sub-bands and facial hair. To prevent them having a dramatic effect on intra-personal matching scores, it is of paramount importance to detect and automatically exclude from comparison the corresponding affected regions of the face. The DDAE framework can be extended to achieve precisely this.
In particular, we extend the existing AAM ensemble with additional models which are now occlusion-specific too. In the training stage this can be achieved by including AAMs which correspond to the existing ones but which are geometrically truncated. Note that this means that the new AAMs do not need to be re-trained -- it is sufficient to adopt the already learnt appearance and shape modes and truncate them directly. In particular, we created two truncated models -- one to account for the growth of facial hair (beard and moustache) and one to account for the presence of eye-wear. These two can also be combined to the produce the third truncated model for the handling of differential facial hair and eye-wear between two images hypothesized to belong to the same person. We created the two baseline truncated models manually; this is straightforward to do using high-level domain knowledge, as the nature of the specific occlusions in question constrains them to very specific parts of the face. An example of a fitted geometrically truncated AAM is shown in Fig.~\ref{f:mask}.\vspace{-12pt}
\begin{SCfigure
\centering
\includegraphics[width=0.233\textwidth]{mask2.png}
\caption{A geometrically truncated AAM. The truncated portion (red) is not used for fitting -- here it is shown using the average shape after the fitting of the remainder of the mesh. }
\label{f:mask}
\end{SCfigure}
\vspace{-12pt}The application of the new ensemble in the classification of a novel image of a face, and in particular the process of model selection needed to achieve this, is somewhat more involved. In particular, the strategy whereby the highest likelihood model is selected is unsatisfactory as it favours smaller models (in our case the models which describe occluded faces). This is a well-known problem in the application of models which do not seek to describe jointly the entire image, i.e.\ both the foreground and the background, but the foreground only. Thus, we overcome this by not selecting the highest likelihood model but rather the model with the highest log-likelihood normalized by the model size \cite{AranCipo2006c}. Recall that the inverse compositional AAM fitting error is given by $e_{icaam} = \sum_{\text{All pixels } \mathbf{x}} \left[ I_e(\mathbf{W}(\mathbf{x};\mathbf{p})) - A_0(\mathbf{W}(\mathbf{x};\mathbf{p})) \right]^2$, where $\mathbf{x}$ are pixel loci, $A_i$ the retained appearance principal components and $\mathbf{W}(\mathbf{x};\mathbf{p})$ the location of the pixel warped using the shape parameters $\mathbf{p}$. Noting that by design of the model the error contributions of different pixels are de-correlated, in our case the highest normalized log-likelihood model is the one with the lowest mean pixel error $\bar{e_{icaam}}$.
\section{Evaluation and results}
In this section we describe the experiments we conducted with the aim of assessing the effectiveness of the proposed framework, and analyze the empirical results thus obtained. We start by summarizing the key features of our data set, follow that with a description of the evaluation methodology, and finally present and discuss our findings.
\vspace{-12pt}\subsubsection{Evaluation data}
In Sec.~\ref{s:intro} we argued that a major limitation of past research lies in the \textit{ad hoc} approach of addressing different extrinsic sources of IR appearance variability. Hand in hand with this goes the similar observation that at present there are few large, publicly available data sets that allow a systematic evaluation of IR-based face recognition algorithms and which contain a gradation of variability of all major extrinsic variables. We used the recently collected Laval University Thermal IR Face Motion Database \cite{GhiaAranBendMald2013a}. The database includes 200 people, aged 20 to 40, most of whom attended two data collection sessions separated by approximately 6 months. In each session a single thermal IR video sequence of the person was acquired using FLIR's Phoenix Indigo IR camera in the 2.5--5$\mu$m wavelength range. The duration of all video sequences in the database is 10~s and they were captured at 30~fps, thus resulting in 300 frames of $320 \times 240$ pixels per sequence. The imaged subjects were instructed to perform head motion that covers the yaw range from frontal ($0^\circ$) to approximately full profile ($\pm 90^\circ$) face orientation relative to the camera, without any special attention to the tempo of the motion or the time spent in each pose. The pose variability in the data is thus extreme. The subjects were also asked to display an arbitrary range of facial expressions. Lastly, a significant number of individuals were imaged in different physiological conditions in the two sessions. In the first session all individuals were imaged in a relatively inactive state, during the course of a sedentary working day and after a prolonged stay indoors, for the second session some users were asked to come straight from the exposure to cold ($<0^\circ$C) outdoors temperatures, alcohol intake, and/or exercise; see Fig.~\ref{f:phys}. In addition, individuals who wore prescription glasses in the first session were now asked to take them off. Several participants were also asked to allow for the growth of facial hair (beard, moustache) between the two sessions.
\vspace{-12pt}\subsubsection{Evaluation methodology}\label{ss:evalMethod}
We evaluated the proposed algorithm in a setting in which the algorithm is trained using only a single image in an arbitrary pose and facial expression. The querying of the algorithm using a novel face is also performed using a single image, possibly in a different pose and/or facial expression. Pose, facial expression changes and partial occlusion due to eye-wear or hair all present a major challenge to the current state of the art, and the consideration of all of the aforementioned in concurrence make our evaluation protocol extremely challenging (indeed, more so than any attempted by previous work), and, importantly, representative of the conditions which are of interest in a wide variety of practical applications. In an attempt to perform a comprehensive comparative evaluation we contacted a number of authors of previously proposed approaches. However none of them was able or willing to provide us with the source code or a working executable of their methods. Thus herein we constrain ourselves to the comparison of the proposed method with the DDAE algorithm which was compared with a thermal minutia points \cite{BuddPavlTsia2006} and vascular networks \cite{BuddPavl2009} methods in \cite{GhiaAranBendMald2013a}.
\subsection{Results}
The key results evaluation are summarized in Table~\ref{t:recognition} and Fig.~4. These show respectively the recognition rates achieved by our system in different experiments we conducted, and the receiver-operator characteristic curves corresponding to the recognition experiments in the presence of occlusion (in all subjects) due to facial hair or prescription glasses.
\begin{figure}[thp]
\centering
\vspace{-15pt}
\begin{tabular}{ccccc}
\rotatebox{90}{~~~Facial hair}~~~~&
\includegraphics[width=0.19\textwidth]{rocBeard.pdf} & \includegraphics[width=0.19\textwidth]{rocBeard_00_30.pdf} & \includegraphics[width=0.19\textwidth]{rocBeard_30_60.pdf} & \includegraphics[width=0.19\textwidth]{rocBeard_60_90.pdf}\\
& \scriptsize Overall & \scriptsize $\Delta$ 0--30$^\circ$ & \scriptsize $\Delta$ 30--60$^\circ$ & \scriptsize $\Delta$ 60--90$^\circ$
\end{tabular}
\begin{tabular}{ccccc}
\rotatebox{90}{~~~~~~Eye-wear}~~~~&
\includegraphics[width=0.19\textwidth]{rocGlasses.pdf} & \includegraphics[width=0.19\textwidth]{rocGlasses_00_30.pdf} & \includegraphics[width=0.19\textwidth]{rocGlasses_30_60.pdf} & \includegraphics[width=0.19\textwidth]{rocGlasses_60_90.pdf}\\
& \scriptsize Overall & \scriptsize $\Delta$ 0--30$^\circ$ & \scriptsize $\Delta$ 30--60$^\circ$ & \scriptsize $\Delta$ 60--90$^\circ$
\end{tabular}
\vspace{-5pt}
\caption{Performance in the presence of occlusion across different extents of pose changes. }
\label{f:roc}
\vspace{-30pt}
\end{figure}
\begin{SCtable}
\centering
\small
\vspace{-10pt}
\caption{Average recognition rate. In experiments with partial occlusion the occlusion was differential (e.g.\ if a training image was acquired with eye-wear on, the test image was acquired with it off, and \textit{vice versa}. }
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{l|ccc}
\noalign{\hrule height 4\arrayrulewidth}
& Unoccluded & Facial hair & Eye-wear\\
\hline
Rank 1 & 100\% & 87\% & 74\% \\
Rank 2 & 100\% & 94\% & 84\% \\
Rank 3 & 100\% & 95\% & 92\% \\
\hline
\noalign{\hrule height 4\arrayrulewidth}
\end{tabular}
\label{t:recognition}
\vspace{-20pt}
\end{SCtable}
\vspace{-10pt}To start with, consider the results in Table~\ref{t:recognition}. It is interesting to notice that performance deterioration is greater when occlusion is caused by eye-wear, rather than facial hair growth. This may be particularly surprising considering that in our data the area occluded by facial hair was larger in extent. One possible explanation of this finding may be rooted in different discriminative abilities of different facial components. Further work is needed to ascertain this; although the eye region appears to be highly informative in the visible spectrum \cite{CampFeriCesa2000} this may not be the case in the IR spectrum as suggested by evidence from some previous work \cite{AranHammCipo2010}. However, there are alternative possibilities which may explain or contribute to the explanation of the observed performance differential. For example, the choice to grow a beard (say) is not arbitrary but rather a conscious decision made with aesthetic considerations in mind. It is possible that individuals who choose to grow facial hair have more characteristic faces. It is also possible that the explanation is of a more practical nature -- perhaps the accuracy of AAM fitting is more affected by the absence of the information around the eyes, rather than those areas of the face typically covered by facial hair. We could not examine this quantitatively as it was prohibitively laborious to obtain the ground truth AAM parameters for the entire database. More research is certainly needed to establish the contribution of each of the aforementioned factors.
As Table~\ref{t:recognition} shows, both types of occlusions, those due to eye-wear and those due to facial hair, have a significant effect on recognition accuracy. However, what is interesting to observe is that already at rank-3 the correct recognition rate in all cases is at least 92\%. This exceeds the performance of the vascular networks based method which used thermal minutia points \cite{BuddPavlTsia2006} and is competitive with the iteratively registered networks approach \cite{BuddPavl2009}, even though the aforementioned algorithms employ several images per person for training and do not consider occlusions.
\section{Summary and conclusions}
We described what we believe to be the first attempt at addressing all major challenges in practical IR-based face recognition in a unified manner. In particular, our system explicitly handles changes in a person's pose, mild facial expression, physiological changes, partial occlusion due to eye-wear, and quasi-occlusion due to facial hair. Our future work will focus on the extension of the described framework to recognition from video and the utilization of partial information available in the regions of the face covered by facial hair.
\tiny
\bibliographystyle{unsrt}
|
\section{Introduction}
There has long been an interest in planar defects or stacking faults (SFs) in
crystals~\cite{Hend42a,Wils42a}. With the recent realization of the
technological import of many materials prone to
SFs---graphene~\cite{Neto09a,Geim13a} and SiC~\cite{Zeke11a} being but two
well known examples---that interest has only grown. Since SFs shift an entire
layer of atoms, it is not unexpected that they can profoundly affect material
properties. Many of these materials have more than one stable stacking
configuration and additionally many metastable ones can exist~\cite{Seba94a},
as well as many stacking configurations that show varying amounts of disorder.
Thus, understanding SFs, perhaps as a prelude to controlling their kind and
placement, presents a significant, but compelling challenge.
Disordered layered materials are often studied via the pairwise correlations
between layers, as these correlations are experimentally accessible from the
Fourier analysis of a diffraction pattern
(DP)~\cite{Este01a,Este01b,Varn02a,Varn13a} or directly from simulation
studies~\cite{Kabr88a,Kabr95a,Shre96a,Shre96b,Shre96c,Shre97a,Varn04a}. Such
studies yield important insights into the structural organization of
materials. For example, Kabra \& Pandey~\cite{Kabr88a} were able to show
that a model of the 2H $\to$ 6H{\footnote{We will use the Ramsdell
notation~\cite{Orti13a,Seba94a} to describe well known crystalline stacking
structures.}} transformation in SiC could retain long-range order even as
the short-range order was reduced. Tiwary \& Pandey~\cite{Tiwa07a}
calculated the size of domains in a model of randomly faulted close-packed
structure (CPSs) by calculating the (exponential) decay rate of pairwise
correlation functions between layers. Recently Estevez-Rams
\etal~\cite{Este08a} derived analytical expressions for the correlation
functions for CPSs that contained both random growth and deformation faults,
and Beyerlein \etal~\cite{Beye11a} demonstrated that correlation functions
in finite-sized FCC crystals depend not only on the kind and amount of
faulting, but additionally on their placement.
Beyond the study of layered materials, pairwise correlation information, in the
form of {\emph {pair distribution functions}} (PDFs), has recently attracted
significant attention~\cite{Egam13a}. However, as useful as the study of
pairwise correlation information is, it does \emph{{not}} provide a complete
description of the specimen. Indeed, it has long been known that very different
atomic arrangements of atoms can reproduce the same PDF, although there has
been recent progress in reducing this degeneracy~\cite{Clif10a}. Nor are they
in general suitable for calculating material properties, such as conductivities
or compressibilities.
For crystalline materials, a complete description of the specimen comes in the
form of its crystal structure, \ie, the specification of the placement of all
the atoms within the unit cell, as well as the description of how the unit
cells are spatially related to each other, commonly referred to as the lattice.
Determining these quantities for specimens and materials is of course the
traditional purview of crystallography. For disordered materials, a similar
formalism is required that provides a unified platform not only to calculate
physical quantities of interest but also to give insight into their physical
structure. For layered materials, where there is but one axis of interest,
namely the organization along the stacking direction, such a formalism has been
identified~\cite{Varn02a,Varn07a,Varn13a}, and that formalism is {\em
{computational mechanics}}~\cite{Crut89a,Crut12a}. The mathematical entity
that gives a compact, statistical description of the disordered material (along
its stacking direction) is its \emph{\eM}, a kind of hidden Markov model
(HMM)~\cite{Rabi89a,Elli95a}. Computational mechanics also has the advantage
of encompassing traditional crystal structures, so both ordered and disordered
materials can be treated on the same footing in the same formalism.
It is our contention that an \eM\ describing a specimen's stacking includes all
of the structural information necessary to calculate physical quantities that
depend on the stacking statistics~\cite{Varn14a}. In the following, we
demonstrate how pairwise correlation functions can be either calculated {\em
analytically} or to a high degree of numerical certainty for an arbitrary
HMM and, thus, for an arbitrary \eM. Previous researchers often calculated pairwise correlation functions for particular realizations of stacking
configurations~\cite{Berl86a,Kabr88a,Shre97a,Este01a} or from analytic
expressions constructed for particular models~\cite{Tiwa07a,Este08a}. The
techniques developed here, however, are the first generally applicable methods
that do not rely on samples of a stacking sequence. The result delivers both an
analytical solution and an efficient numerical tool. And while we will
specialize to the case of CPSs for concreteness, the methods developed are
extendable to other materials and stacking geometries.
Our development is organized as follows: In \S \ref{Stacking Notations} we
introduce nomenclature. In \S \ref{sec:HaggToABC} we develop an algorithm to
change between different representations of stacking sequences. In \S
\ref{sec:CFsFromHMMs} we derive expressions, our main results, for the pairwise
correlation functions between layers in layered CPSs. In \S \ref{Examples} we
consider several examples; namely, (i) a simple stacking process that
represents the 3C crystal structure or a completely random stacking depending
on the parameter choice, (ii) a stacking process that represents random growth
and deformation faults, and (iii) a stacking process inspired by recent
experiments in 6H-SiC. And, in \S \ref{Conclusions} we give our conclusions and
directions for future work.
\section{Definitions and Notations}
\label{Stacking Notations}
We suppose the layered material is built up from identical sheets called {\em
{modular layers}} (MLs)~\cite{Pric83a,Ferr08a}. The MLs are completely ordered
in two dimensions and assume only one of three discrete positions, labeled $A$,
$B$, or $C$~\cite{Ashc76a,Seba94a}. These represent the physical placement of
each ML and are commonly known as the $ABC$-notation~\cite{Orti13a}. We define the set of possible
orientations in the $ABC$-notation as $\AlphabetABC = \{A, B, C\}$. We further
assume that the MLs obey the same stacking rules as CPSs, namely that two
adjacent layers may not have the same orientation; \ie, stacking sequences
$AA$, $BB$ and $CC$ are not allowed. Exploiting this constraint, the stacking
structure can be represented more compactly in the \Hagg-notation: one takes
the transitions between MLs as being either cyclic, ($A \to B, B \to C,$ or $C
\to A$), and denoted as `+'; or anticyclic, ($A \to C, C \to B,$ or $B \to A$),
and denoted as `-'. The \Hagg-notation then gives the relative orientation of
each ML to its predecessor. It is convenient to identify the usual
\Hagg-notation `+' as `1' and `-' as `0'. Doing so, we define the set of
possible relative orientations in the \Hagg-notation as $\AlphabetHagg =
\{0,1\}$. These two notations---$ABC$ and \Hagg---carry an identical message,
up to an overall rotation of the specimen. Alternatively, one can say that there is
freedom of choice in labeling the first ML.
\subsection{Correlation functions}
\label{CorrelationFunctions}
Let us define three
statistical quantities, $\Qcn$, $\Qan$, and $\Qsn$~\cite{Yi96a}: the
pairwise \emph{correlation functions} (CFs) between MLs, where c, a, and
s stand for \emph{cyclic}, \emph{anticyclic}, and \emph{same},
respectively. $\Qcn$ is the probability that any two MLs at a separation of
$n$ are cyclically related. $\Qan$ and $\Qsn$ are defined in a similar
fashion.\footnote{As yet, there is no consensus on notation for these
quantities. Warren~\cite{Warr69a} uses $P_m^0, P_m^+$, and $P_m^-$, Kabra
\& Pandey~\cite{Kabr88a} call these $P(m), Q(m)$, and $R(m)$, and
Estevez~\etal~\cite{Este08a} use $P_0(\Delta), P_{\rm f}(\Delta)$, and $P_{\rm
b}(\Delta)$. Since we prefer to reserve the symbol `$P$' for other
probabilities previously established in the literature, here and elsewhere we
follow the notation of Yi \& Canright~\cite{Yi96a}, with a slight
modification of replacing `$Q_{\rm{r}}(n)$' with `$\Qan$'.} Since these are probabilities: $0
\leq \Qxin \leq 1$, where $\xi \in \{\rm{c,a,s} \}$. Additionally, at each $n$ it is
clear that $\sum_{\xi} \Qxin = 1$. Notice that the CFs are defined in terms of
the $ABC$-notation.
\subsection{Representing layer stacking as a hidden process}
\label{TransitionMatrices}
We chose to represent a stacking sequence as the output of discrete-step,
discrete-state hidden Markov model (HMM). A HMM $\Gamma$ is an ordered tuple
$\Machine = (\Alphabet, \States, \mu_0, \TransMatSet)$, where $\Alphabet$ is
the set of symbols that one observes as the HMM's output, often called an
alphabet, $\States$ is a finite set of $M$ internal states, $\mu_0$ is an
initial state probability distribution, and $\TransMatSet$ is a set of matrices
that give the probability of making a transition between the states while
outputting one of the symbols in $\Alphabet$. These transition probability
matrices or more simply \emph{transition matrices} (TMs)~\cite{Paz71a,Karl75a}
are usually written:
\begin{align}
\TransMat^{[\symbolHagg]} =
\begin{bmatrix} \nonumber
\Pr(\symbolHagg, \StateHagg_1|\StateHagg_1 ) & \Pr(\symbolHagg, \StateHagg_2|\StateHagg_1 ) & \cdots & \Pr(\symbolHagg, \StateHagg_M|\StateHagg_1 ) \\
\Pr(\symbolHagg, \StateHagg_1|\StateHagg_2 ) & \Pr(\symbolHagg, \StateHagg_2|\StateHagg_2 ) & \cdots & \Pr(\symbolHagg, \StateHagg_M|\StateHagg_2 ) \\
\vdots & \vdots & \ddots & \vdots \\
\Pr(\symbolHagg, \StateHagg_1|\StateHagg_M ) & \Pr(\symbolHagg, \StateHagg_2|\StateHagg_M ) & \cdots & \Pr(\symbolHagg, \StateHagg_M|\StateHagg_M ) \\
\end{bmatrix} ,
\end{align}
where $\symbolHagg \in \Alphabet$ and $\StateHagg_1, \StateHagg_2, \dots, \StateHagg_M \in \States$.
For a number of purposes it is convenient to work directly with the internal
state TM, denote it $\TransMat$. This is the matrix of state transition
probabilities regardless of symbol, given by the sum of the symbol-labeled TMs:
$\TransMat = \sum_{x \in \Alphabet} \TransMat^{[x]}$. For example, the
internal state distribution evolves according to $\bra{\boldsymbol{\mu_1}} =
\bra{\boldsymbol{\mu_0}} \TransMat$.
Or, more generally, $\bra{\boldsymbol{\mu_L}} = \bra{\boldsymbol{\mu_0}} \TransMat^L$.
(In this notation, state distributions are row vectors.)
In another use, one finds the stationary state probability distribution:
\begin{align}
\bra \Dist
&= \begin{bmatrix} \nonumber
\Pr({{\StateHagg}_1}) & \Pr({{\StateHagg}_2}) & \dotsm &
\Pr({{\StateHagg}_{M}})
\end{bmatrix},
\end{align}
as the left eigenvector of $\TransMat$ normalized in probability:
\begin{align}
\bra \Dist = \bra \Dist \TransMat ~.
\label{eq:ProbabilityDistribution}
\end{align}
The probability of any finite-length sequence of symbols can be computed exactly from these
objects using linear algebra.
In particular, a length-$L$ `word'
$w = \symbolHagg_0 \symbolHagg_1 \dots \symbolHagg_{L-1} \in \Alphabet^L$,
where $\Alphabet^L$ is the set of length-$L$ sequences, has the stationary probability:
\begin{align*}
\Pr(w) &= \bra \Dist \TransMat^{[w]} \DistOne \\
& = \bra \Dist \TransMat^{[\symbolHagg_0]} \TransMat^{[\symbolHagg_1]} \dotsm \TransMat^{[\symbolHagg_{L-1}]} \DistOne~,
\end{align*}
where $\DistOne$ is the column-vector of all ones.
As a useful convention, we will use bras $\bra{\cdot}$ to denote row vectors
and kets $\ket{\cdot}$ to denote column vectors. On the one hand, any object
closed by a bra on the left and ket on the right is a scalar and commutes as a
unit with anything. On the other hand, a ket--bra $\ket{\cdot} \bra{\cdot}$ has
the dimensions of a square matrix.
To help make these ideas concrete, let us consider a CPS stacked according to the \emph{Golden Mean
Process} (GMP) represented in the \Hagg-notation. This process has been
previously treated in the context of CPSs in Varn~\cite{Varn13b}. Any
sequence is allowed as long as there are no consecutive 0s. This is
accomplished by examining the previous observed symbol: if it is 1, then the
next symbol in the sequence is either 0 or 1 with equal probability; if it is
0, then the next symbol is necessarily 1. Thus, there are two states in the
\Hagg-machine, corresponding to the above two conditions. Let us call these
states $\mathcal{U}$ (next symbol is a 0 or 1 with equal probability) and
$\mathcal{V}$ (next symbol is a 1). And so, we say $\States
=\{\mathcal{U},\mathcal{V}\}$. The two 2-by-2 TMs for this process---one for
each symbol in the alphabet---are given by:
\begin{align*}
\begin{array}{r@{\mskip\thickmuskip}l}
\TransMatHagg^{[0]}
&=
\begin{bmatrix}
0 & \OneHalf \\
0 & 0
\end{bmatrix}
\end{array}
{\rm {and}}
\begin{array}{r@{\mskip\thickmuskip}l}
\TransMatHagg^{[1]}
&=
\begin{bmatrix}
\OneHalf & 0 \\
1 & 0
\end{bmatrix}~.
\end{array}
\end{align*}
The GMP has internal-state TM:
\begin{align}
\TransMatHagg
&= \TransMatHagg^{[0]} + \TransMatHagg^{[1]} \nonumber \\
&=
\begin{bmatrix} \nonumber
\OneHalf & \OneHalf \\
1 & 0
\end{bmatrix}~.
\end{align}
The asymptotic state probabilities are given by
$\bra{\Dist_{\textrm{H}}} = \Big[ \tfrac{2}{3} ~~ \tfrac{1}{3} \Big]$.
In this way, the \MachineHagg\ for the GM Process is defined as
$\Machine^{({\rm H})}_{\rm {GM}} = (\Alphabet, \States, \mu_0,
\TransMatSet) = (\{0,1\}, \{\mathcal{U},\mathcal{V}\},
\left[ \tfrac{2}{3} ~~ \tfrac{1}{3} \right] , \{\TransMatHagg^{[0]},
\TransMatHagg^{[1]}\})$.\footnote{Here and in the examples of \S\ref{Examples},
we take the stationary state probability distribution $\Dist$ as the initial probability state distribution $\mu_0$,
as we are interested for now in the the long term behavior.}
HMMs are often conveniently depicted using labeled directed graphs.
As an example, the GM Process's HMM is shown in Fig.~\ref{FigGMProcessHagg}.
\begin{figure}
\begin{center}
\includegraphics[width=0.35\textwidth]{images/tikz/GMProcessHagg.pdf}
\end{center}
\caption{The GM Process written as a \MachineHagg. The circles indicate states,
and the arcs between them are transitions, labeled by $s|p$, where $s$ is the
symbol emitted upon transition and $p$ is the probability of making such a
transition.
}
\label{FigGMProcessHagg}
\end{figure}
\section{Expanding the \MachineHagg\ to the \MachineABC}
\label{sec:HaggToABC}
While simulation studies~\cite{Varn04a} and $\epsilon$-Machine Spectral
Reconstruction (\eMSR)~\cite{Varn02a,Varn07a,Varn13a,Varn13b} express stacking
structure in terms of the \MachineHagg, for some calculations it is more
convenient to represent the stacking process in terms of the \MachineABC. Here,
we give a graphical procedure for expanding the \MachineHagg\ into the
\MachineABC\ and then provide an algebraically equivalent algorithm. We note
that this expansion procedure is \emph{not} unique and can vary up to an
overall rearrangement of the columns and rows of the resulting \MachineABC\ TM.
This difference, of course, does not alter the results of calculations of
physical quantities.
\subsection{Graphical expansion method}
\label{GraphicalExpansion}
\begin{figure}
\begin{center}
\includegraphics[width=0.22\textwidth]{images/tikz/GMProcessABC1.pdf}
\end{center}
\caption{The first step in expanding a \MachineHagg\ into a \MachineABC\ is to
treble the number of states.
}
\label{FigGMProcessABC1}
\end{figure}
Recall that the \Hagg-notation and the $ABC$-notation are equivalent
representations of the stacking structure,
up to an overall rotation of the crystal. Stated alternatively, in
the \Hagg-notation, there is an ambiguity concerning the orientation of each
ML---it could be either $A$, $B$ or $C$. To account for this degeneracy, when
we transform to the $ABC$ representation, we triple the size of the
\MachineHagg. As a first step, one writes down three states for each state
found in the \MachineHagg, but not the transitions between them. To
distinguish among these new states of the triplet, label each with a superscript
($A$, $B$ or $C$) indicating the last ML added to arrive at that state.
Symbolically, this is stated:
\begin{align}
\left\{\StateHagg_i \right\}
\xRightarrow{\text{\Hagg\ to } ABC}
\left\{ \StateABC_i^{[A]}, \text{ } \StateABC_i^{[B]}, \text{ } \StateABC_i^{[C]} \right\}~. \nonumber
\end{align}
Transitions between states on the \MachineABC\ still respect the same state
labeling scheme as on the \MachineHagg\ (explained below), but now they store the ML information.
Transitions between states on the \MachineHagg\ that were labeled with $1$
advance the stacking sequence cyclically (\ie, $A \to B \to C \to A$) and the
corresponding transitions on the \MachineABC\ reflect this by taking the ML
label on the initial state and advancing it cyclically. In a completely
analogous way, transitions labeled 0 on the \MachineHagg\ advance the states on the
\MachineABC\ in an anticyclical fashion (\ie, $A \to C \to B \to A$).
Continuing our GM Process example, let us write out the six ($= 3 \times 2$)
states labeled with superscripts to distinguish them. This is done in
Fig.~\ref{FigGMProcessABC1}. (It does not matter in what order these states are
labeled. The scheme chosen in Fig.~\ref{FigGMProcessABC1} turns out to be
convenient given the state-to-state transition structure of the final
\MachineABC, but any arrangement is satisfactory.) The transitions between the
states on the \MachineABC\ preserve the labeling scheme of the original
\MachineHagg. That is, if in the original \MachineHagg\ there is transition
$\StateHagg_i \xrightarrow{\symbolHagg|p} \StateHagg_j$, then there must be
\emph{three} similar transitions on the \MachineABC\ of the form
$\StateABC_i^{[\symbolABC]} \xrightarrow{\symbolABC^{\prime}|p}
\StateABC_j^{[\symbolABC^{\prime}]}$, with $\symbolABC, \symbolABC^{\prime} \in
\{A,B,C\}$. Additionally, the transitions on the \MachineABC\ corresponding to
the transitions on the \MachineHagg\ have the same probability.
\begin{figure}
\begin{center}
\includegraphics[width=0.25\textwidth]{images/tikz/GMProcessABC2.pdf}
\end{center}
\caption{The second step in expanding the \MachineHagg\ into the \MachineABC\ is to
add the transitions. Here, a single transition on the example \MachineHagg,
$\mathcal{U} \xrightarrow{1|\OneHalf} \mathcal{U}$, is expanded into {\emph {three}}
transitions on the \MachineABC.}
\label{FigGMProcessABC2}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.24\textwidth]{images/tikz/GMProcessABC3.pdf}
\end{center}
\caption{The completely expanded six-state \MachineABC\ that corresponds to the two-state
\MachineHagg\ shown in Fig.~\ref{FigGMProcessHagg}.}
\label{FigGMProcessABC3}
\end{figure}
Let us consider the self-state transition on the \MachineHagg\ shown in
Fig.~\ref{FigGMProcessHagg}: $\mathcal{U} \xrightarrow{1|\OneHalf}
\mathcal{U}$. Since the corresponding transitions on the \MachineABC\ still
respect the state labeling scheme, the self-loop on $\mathcal{U}$ only induces transitions among the
{${\mathcal{U}}^{[\symbolABC]}$}. Since a $1$ advances the stacking sequence cyclically, the appropriate transitions are:
\begin{align}
{\mathcal{U}}^{[A]} & \xrightarrow{B|\OneHalf} {\mathcal{U}}^{[B]}, \nonumber \\
{\mathcal{U}}^{[B]} & \xrightarrow{C|\OneHalf} {\mathcal{U}}^{[C]}, \nonumber \\
{\mathcal{U}}^{[C]} & \xrightarrow{A|\OneHalf} {\mathcal{U}}^{[A]} ~. \nonumber
\end{align}
This is illustrated in Fig.~\ref{FigGMProcessABC2}. Applying the same procedure to the other
transitions on the \MachineHagg, \ie, $\mathcal{U} \xrightarrow{0|\OneHalf} \mathcal{V}$
and $\mathcal{V} \xrightarrow{1|1} \mathcal{U}$,
results in the completely expanded \MachineABC,
and this is shown in Fig.~\ref{FigGMProcessABC3}.
We are now able to write down the stacking process for the GM Process from its
expanded graph, Fig.~\ref{FigGMProcessABC3}. First, we note that the alphabet
is ternary: $\AlphabetABC = \{A, B, C\}$. Second, there are six states on the
\MachineABC, \ie, $\StateABC = \{ {\mathcal{U}}^{[A]}, {\mathcal{U}}^{[B]},
{\mathcal{U}}^{[C]}, {\mathcal{V}}^{[A]}, {\mathcal{V}}^{[B]},
{\mathcal{V}}^{[C]} \}$. Ordering the states as above, the TMs may be directly
constructed from the expanded graph, and are given by:
\begin{align}
\TransMatABC^{[A]} &=
\begin{bmatrix} \nonumber
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & \OneHalf & 0 & 0 \\
\OneHalf & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0
\end{bmatrix},
\TransMatABC^{[B]} =
\begin{bmatrix} \nonumber
0 & \OneHalf & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & \OneHalf & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0
\end{bmatrix},
\end{align}
and
\begin{align}
\TransMatABC^{[C]}
&=
\begin{bmatrix} \nonumber
0 & 0 & 0 & 0 & 0 & \OneHalf \\
0 & 0 & \OneHalf & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0
\end{bmatrix}~.
\end{align}
As before, the internal-state TM is simply the sum of the symbol-specific TMs,
given by $\TransMatABC = \TransMatABC^{[A]} +\TransMatABC^{[B]} +
\TransMatABC^{[C]}$. For the GM Process this turns out to be:
\begin{align}
\TransMatABC
&=
\begin{bmatrix} \nonumber
0 & \OneHalf & 0 & 0 & 0 & \OneHalf \\
0 & 0 & \OneHalf & \OneHalf & 0 & 0 \\
\OneHalf & 0 & 0 & 0 & \OneHalf & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0
\end{bmatrix}~.
\end{align}
For completeness, the HMM for the GM Process in terms of the physical stacking of MLs is
$\Machine^{({\rm P})}_{\rm {GM}} = (\Alphabet, \States, \mu_0, \TransMatSet) = \bigl( \{A,B,C\}, \bigr.$
$\{{\mathcal{U}}^{[A]}, {\mathcal{U}}^{[B]}, {\mathcal{U}}^{[C]},
{\mathcal{V}}^{[A]}, {\mathcal{V}}^{[B]}, {\mathcal{V}}^{[C]}\},$
$\tfrac{1}{9} \left[ 2~~2~~2~~1~~1~~1 \right],$
$\bigl. \{\TransMatABC^{[A]}, \TransMatABC^{[B]}, \TransMatABC^{[C]}\} \bigr)$.
\subsection{Mixing and Nonmixing State Cycles}
Observe Fig.~\ref{FigGMProcessABC3}'s directed graph is \emph{strongly
connected}---any state is accessible from any other state in a finite number of
transitions. It should be apparent that this need not have been the case. In
fact, in this example connectivity is due to the presence of the self-state
transition $\mathcal{U} \xrightarrow{1} {\mathcal{U}}$. The latter guarantees a
strongly connected expanded graph. Had this transition been absent on the
\MachineHagg, such that there were only transitions of the form $\mathcal{U}
\xrightarrow{0} \mathcal{V}$ and $\mathcal{V} \xrightarrow{1} \mathcal{U}$, the
expansion would have yielded a graph with three distinct, unconnected
components. Only one of these graphs would be physically manifest. It is
sufficient to take just one component, arbitrarily assign a $A,B$ or $C$ to an
arbitrary state on that component, and then replace all of the $\{0,1\}$
transitions with the appropriate $\{A,B,C\}$ transitions, as done above.
To determine whether the expansion process on a \MachineHagg\ results in a
strongly connected graph, one can examine the set of \emph{simple state cycles}
(SSCs) and calculate the \emph{winding number} for each. A SSC is defined
analogous to a \emph{causal state cycle} (CSC)~\cite{Varn13a} on an \eM\ as a
``finite, closed, nonself-intersecting, symbol-specific path" along the graph.
The winding number $W$ for a SSC on a \MachineHagg\ is similar to the
parameter $\Delta$ previously defined by Yi \& Canright~\cite{Yi96a} and the
{\em cyclicity} ($C$)~\cite{Dorn71a} for a polytype of a CPS. $W$ differs from $C$ as the former is
not divided by the period of the cycle. We define the winding number for a SSC
as:
\begin{align}
W^{\rm {SSC}} = n_1 - n_0~, \nonumber
\end{align}
where $n_1$ and $n_0$ are the number of 1s and the number of $0$s encountered
traversing the SSC, respectively. We call those SSCs \emph{mixing} if $W^{\rm
{SSC}} \pmod{3} \neq 0$, and \emph{nonmixing} if $W^{\rm {SSC}} \pmod{3} = 0$.
If there is at least one mixing SSC on the \MachineHagg, then the expanded
\MachineABC\ will be strongly connected. For example, there are two SSCs on the
\MachineHagg\ for the GM Process: $[\mathcal{U}]$ and
$[\mathcal{U}\mathcal{V}]$.\footnote{We use the same nomenclature to denote a
SSC as previously used to denote a CSC: The state sequence visited traversing
the cycle is given in square brackets~\cite{Varn13a}. For those cases where an
ambiguity exists because the transition occurs on more than one symbol, we
insert a subscript in parentheses denoting that symbol.} The winding number for each is given
by $W^{[\mathcal{U}]} = 1 - 0 = 1$ and $W^{[\mathcal{U}\mathcal{V}]} = 1 - 1 =
0$. Since $W^{[\mathcal{U}]} \neq 0$ and $[\mathcal{U}]$ is thus a mixing SSC,
the \MachineHagg\ for the GM Process will expand into a strongly connected
\MachineABC. Let us refer to those \MachineHagg s with at least one mixing SSC
as \emph{mixing \MachineHagg s} and those that do not as \emph{nonmixing
\MachineHagg s} and similarly for the corresponding \MachineABC s. We find that
mixing \MachineHagg s, and thus mixing \MachineABC s, are far more common than
nonmixing ones and that the distinction between the two can have profound
effects on the calculated quantities, such as the CFs and the
DP~\cite{Riec14c}.
\subsection{Rote expansion algorithm}
\label{AlgebraicExpansion}
To develop an algorithm for expansion, it is more convenient to change notation
slightly. Let us now denote $\StateABC$ as the set of hidden recurrent states
in the \MachineABC, indexed by integer subscripts: $\StateABC = \{\StateABC_i:
i=1, \dots, \NstatesABC\}$, where $\NstatesABC = |\StateABC|$. Define the
probability to transition from state $\StateABC_i$ to state $\StateABC_j$ on
the symbol $\symbolABC \in \AlphabetABC$ as $\TransMatABC^{[\symbolABC]}_{i,
j}$. Let's gather these state-to-state transition probabilities into a
$\NstatesABC \times \NstatesABC$ matrix, referring to it as the
\emph{$\symbolABC$-transition matrix} ($\symbolABC$-TM)
$\TransMatABC^{[\symbolABC]}$. Thus, there will be as many $\symbolABC$-TMs as
there are symbols in the alphabet of the \MachineABC, which is always
$|\AlphabetABC| = 3$ for CPSs.
As before, transitioning on symbol $1$ has a threefold degeneracy in the $ABC$
language, as it could imply any of the three transitions $(A \to B$, $B \to C$,
or $C \to A$), and similarly for 0. Thus, each labeled edge of the Hagg-machine
must be split into three distinct labeled edges of the $ABC$-machine.
Similarly, each state of the \MachineHagg\ maps onto three distinct states of
the $ABC$-machine. Although we have some flexibility in indexing states in the
resulting $ABC$-machine, we establish consistency by committing to the
following construction.\footnote{Alternative constructions merely swap the
labels of different states, but this choice of indexing affects the particular
form of the TMs and how they are extracted from the \MachineHagg\ TMs. We
choose the construction here for its intuitive and simple form.}
If $\NstatesHagg$ is the number
of states in the \MachineHagg, then $\NstatesABC = 3 \NstatesHagg$ for mixing \MachineHagg s.
(The case of nonmixing \MachineHagg s is treated afterward.)
Let the $i^{\rm {th}}$ state of the \MachineHagg\ split into
the ${\left( 3i-2 \right) }^{\rm {th}}$ through the ${\left( 3i \right) }^{\rm
{th}}$ states of the
corresponding \MachineABC. Then, each labeled-edge transition from
the $i^{\rm {th}}$ to the $j^{\rm {th}}$ states of the \MachineHagg\ maps into a
3-by-3 submatrix for each of the three labeled TMs of the
\MachineABC\ as:
\begin{align}
\label{eq:T0toABC}
\left\{\TransMatHagg_{ij}^{[0]} \right\}
\xRightarrow{\text{\Hagg\ to } ABC}
\left\{ \TransMatABC_{3i-1, 3j-2}^{[A]}, \text{ } \TransMatABC_{3i, 3j-1}^{[B]}, \text{ } \TransMatABC_{3i-2, 3j}^{[C]} \right\}
\end{align}
and
\begin{align}
\label{eq:T1toABC}
\left\{\TransMatHagg_{ij}^{[1]} \right\}
\xRightarrow{\text{\Hagg\ to } ABC}
\left\{ \TransMatABC_{3i, 3j-2}^{[A]}, \text{ } \TransMatABC_{3i-2, 3j-1}^{[B]}, \text{ } \TransMatABC_{3i-1, 3j}^{[C]} \right\} .
\end{align}
We can represent the mapping of Eq.\ \eqref{eq:T0toABC} and Eq.\ \eqref{eq:T1toABC} more
visually with the following equivalent set of statements:
\begin{align}
\begin{pmatrix}
\TransMatABC_{3i-2, 3j-2}^{[A]} & \TransMatABC_{3i-2, 3j-1}^{[A]} & \TransMatABC_{3i-2, 3j}^{[A]} \\
\TransMatABC_{3i-1, 3j-2}^{[A]} & \TransMatABC_{3i-1, 3j-1}^{[A]} & \TransMatABC_{3i-1, 3j}^{[A]} \\
\TransMatABC_{3i, 3j-2}^{[A]} & \TransMatABC_{3i, 3j-1}^{[A]} & \TransMatABC_{3i, 3j}^{[A]}
\end{pmatrix}
=
\begin{pmatrix}
0 & 0 & 0 \\
\TransMatHagg_{ij}^{[0]} & 0 & 0 \\
\TransMatHagg_{ij}^{[1]} & 0 & 0
\end{pmatrix} ,
\end{align}
\vspace{-0.1in}
\begin{align}
\begin{pmatrix}
\TransMatABC_{3i-2, 3j-2}^{[B]} & \TransMatABC_{3i-2, 3j-1}^{[B]} & \TransMatABC_{3i-2, 3j}^{[B]} \\
\TransMatABC_{3i-1, 3j-2}^{[B]} & \TransMatABC_{3i-1, 3j-1}^{[B]} & \TransMatABC_{3i-1, 3j}^{[B]} \\
\TransMatABC_{3i, 3j-2}^{[B]} & \TransMatABC_{3i, 3j-1}^{[B]} & \TransMatABC_{3i, 3j}^{[B]}
\end{pmatrix}
=
\begin{pmatrix}
0 & \TransMatHagg_{ij}^{[1]} & 0 \\
0 & 0 & 0 \\
0 & \TransMatHagg_{ij}^{[0]} & 0
\end{pmatrix} ,
\end{align}
\vspace{-0.1in}
and:
\begin{align}
\begin{pmatrix}
\TransMatABC_{3i-2, 3j-2}^{[C]} & \TransMatABC_{3i-2, 3j-1}^{[C]} & \TransMatABC_{3i-2, 3j}^{[C]} \\
\TransMatABC_{3i-1, 3j-2}^{[C]} & \TransMatABC_{3i-1, 3j-1}^{[C]} & \TransMatABC_{3i-1, 3j}^{[C]} \\
\TransMatABC_{3i, 3j-2}^{[C]} & \TransMatABC_{3i, 3j-1}^{[C]} & \TransMatABC_{3i, 3j}^{[C]}
\end{pmatrix}
=
\begin{pmatrix}
0 & 0 & \TransMatHagg_{ij}^{[0]} \\
0 & 0 & \TransMatHagg_{ij}^{[1]} \\
0 & 0 & 0
\end{pmatrix} ,
\end{align}
which also yields the 3-by-3 submatrix for the \emph{unlabeled} $ABC$ TM in
terms of the \emph{labeled H\"{a}gg TMs}:
\begin{align}
\begin{pmatrix}
\TransMatABC_{3i-2, 3j-2} & \TransMatABC_{3i-2, 3j-1} & \TransMatABC_{3i-2, 3j} \\
\TransMatABC_{3i-1, 3j-2} & \TransMatABC_{3i-1, 3j-1} & \TransMatABC_{3i-1, 3j} \\
\TransMatABC_{3i, 3j-2} & \TransMatABC_{3i, 3j-1} & \TransMatABC_{3i, 3j}
\end{pmatrix}
=
\begin{pmatrix}
0 & \TransMatHagg_{ij}^{[1]} & \TransMatHagg_{ij}^{[0]} \\
\TransMatHagg_{ij}^{[0]} & 0 & \TransMatHagg_{ij}^{[1]} \\
\TransMatHagg_{ij}^{[1]} & \TransMatHagg_{ij}^{[0]} & 0
\end{pmatrix} .
\end{align}
Furthermore, for mixing \MachineHagg s, the probability from the stationary
distribution over their states maps to a triplet of probabilities for the
stationary distribution over the \MachineABC\ states:
\begin{align}
\left\{ {p^{\text{H}}}_{i} \right\}
\xRightarrow{\text{\Hagg\ to } ABC}
\left\{ 3p_{3i-2}, \text{ } 3p_{3i-1}, \text{ } 3p_{3i} \right\}
\end{align}
such that:
\begin{align}
{\bra \Dist } &= \;
\begin{bmatrix} p_1 & p_2 & p_3 & p_4 & \dots & p_{\NstatesABC - 1} & p_{\NstatesABC} \end{bmatrix} \nonumber \\
& =\tfrac{1}{3} \begin{bmatrix} p^{\rm {H}}_1 & p^{\rm {H}}_1 & p^{\rm {H}}_1
& p^{\rm {H}}_2
& \dots & p^{\rm {H}}_{\NstatesHagg} & p^{\rm {H}}_{\NstatesHagg}
\end{bmatrix}~.
\label{eq:PiExpansion}
\end{align}
The reader should check that applying the rote expansion method given here
results in the same HMM for the GM Process as we found in \S\ref{GraphicalExpansion}.
\section{Correlation Functions from HMMs}
\label{sec:CFsFromHMMs}
At this point, with the process expressed as an \MachineABC, we can derive expressions for the CFs.
We introduce the family of cyclic-relation functions ${\xoperator} (\symbolABC)
\in \{\coperator(\symbolABC), \aoperator(\symbolABC),
\soperator(\symbolABC)\}$, where, for example:
\begin{equation}
\coperator(\symbolABC) = \left\{
\begin{array}{l l}
B & \quad \textrm{if } \, \symbolABC = A\\
C & \quad \textrm{if } \, \symbolABC = B\\
A & \quad \textrm{if } \, \symbolABC = C\\
\end{array} \right..
\end{equation}
Thus, $\coperator(\symbolABC)$ is the cyclic permutation function.
Complementarily, $\aoperator(\symbolABC)$ performs anticyclic permutation among
$\symbolABC \in \{ A, B, C \}$; $\soperator(\symbolABC)$ performs the identity
operation among $\symbolABC \in \{ A, B, C \}$ and is suggestively denoted with
an `s' for \emph{sameness}. In terms of the absolute position of the MLs---\ie,
$\AlphabetABC = \{A, B, C\}$---the CFs directly relate to the products of
particular sequences of TMs. This perspective suggests a way to uncover the
precise relation between the CFs and the TMs. Using this, we then give a
closed-form expression for $\Qxin$ for any given HMM.
\subsection{CFs from TMs}
\label{subsec:CFsFromTMs}
To begin, let us first consider the meaning of $\QcThree$. In words, this is
the probability that two MLs separated by two intervening MLs are cyclically
related. Mathematically, we might start by writing this as:
\begin{align}
\QcThree = \Pr(A**B) + \Pr(B**C) + \Pr(C**A),
\end{align}
where $*$ is a wildcard symbol denoting an indifference for the symbol observed
in its place \footnote{While it is tempting to add the stipulation that no two
consecutive symbols can be the same, this will fall out naturally from $\QsOne
= 0$ via the transition-constraints built into the $ABC$-machine construction.}.
That is, $*$s denote marginalizing over the intervening MLs such that, for
example:
\begin{align}
\Pr(A**B) = \sum_{\symbolABC_1 \in \AlphabetABC}
\sum_{\symbolABC_2 \in \AlphabetABC} \Pr(A \symbolABC_{1} \symbolABC_{2} B).
\end{align}
Making use of the TM-formalism discussed previously, this becomes:
\begin{align}
\Pr(A**B) &= \sum_{\symbolABC_{1} \in \AlphabetABC} \sum_{\symbolABC_{2} \in \AlphabetABC}
\Pr(A \symbolABC_{1} \symbolABC_{2} B) \nonumber \\
&= \sum_{\symbolABC_{1} \in \AlphabetABC} \sum_{\symbolABC_{2} \in \AlphabetABC}
\bra \Dist \TransMatABC^{[A]} \TransMatABC^{[\symbolABC_{1}]}
\TransMatABC^{[\symbolABC_{2}]} \TransMatABC^{[B]} \DistOne \nonumber \\
&= \bra \Dist \TransMatABC^{[A]} \Bigl( \sum_{\symbolABC_{1} \in \AlphabetABC}
\sum_{\symbolABC_{2} \in \AlphabetABC} \TransMatABC^{[\symbolABC_{1}]}
\TransMatABC^{[\symbolABC_{2}]} \Bigr) \ \TransMatABC^{[B]} \DistOne \nonumber \\
&= \bra \Dist \TransMatABC^{[A]} \Bigl( \underbrace{\sum_{\symbolABC_{1} \in \AlphabetABC}
\TransMatABC^{[\symbolABC_{1}]} }_{= \TransMatABC} \Bigr)
\Bigl( \underbrace{\sum_{\symbolABC_{2} \in \AlphabetABC}
\TransMatABC^{[\symbolABC_{2}]} }_{= \TransMatABC} \Bigr) \TransMatABC^{[B]} \DistOne \nonumber \\
&= \bra \Dist \TransMatABC^{[A]} (\TransMatABC) (\TransMatABC) \TransMatABC^{[B]} \nonumber \DistOne \\
&= \bra \Dist \TransMatABC^{[A]} \TransMatABC^2 \TransMatABC^{[B]} \DistOne~, \nonumber
\end{align}
where $\DistOne$ is a column vector of 1s of length $\NstatesABC$.
Hence, we can rewrite $\QcThree$ as:
\begin{align}
\QcThree &= \Pr(A**B) + \Pr(B**C) + \Pr(C**A) \nonumber \\
&= \bra \Dist \TransMatABC^{[A]} \TransMatABC^2 \TransMatABC^{[B]} \DistOne + \nonumber
\bra \Dist \TransMatABC^{[B]} \TransMatABC^2 \TransMatABC^{[C]} \DistOne \\
& \qquad + \bra \Dist \TransMatABC^{[C]} \TransMatABC^2 \TransMatABC^{[A]} \DistOne \nonumber \\
&= \sum_{\symbolABC \in \AlphabetABC}
\bra \Dist \TransMatABC^{[\symbolABC]} \TransMatABC^2
\TransMatABC^{[\coperator(\symbolABC)]} \DistOne. \nonumber
\end{align}
For \emph{mixing} {\MachineABC}s,
$\Pr(A**B) = \Pr(B**C) = \Pr(C**A) = \frac{1}{3} \QcThree$,
in which case the above reduces to:
\begin{align}
\QcThree = 3 \bra \Dist \TransMatABC^{[\symbolABC_{0}]} \TransMatABC^2
\TransMatABC^{[\coperator(\symbolABC_{0})]} \DistOne, \text{ where }
\symbolABC_{0} \in \AlphabetABC. \nonumber
\end{align}
The generalization to express any $\Qxin$ in terms of TMs may already be
obvious by analogy. Nevertheless, we give a brief derivation for completeness,
using similar concepts to those developed more explicitly above. For all $\xi
\in \{\rm{c}, \rm{a}, \rm{s} \}$ and for all $n \in \{1, 2, 3, \dots \}$, we can write the CFs
as:
\begin{align}
\Qxin &= \Pr(A \underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (A))
+ \Pr(B \underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (B)) \nonumber \\
& \qquad + \Pr(C \underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (C)) \nonumber \\
&= \sum_{\symbolABC_{0} \in \AlphabetABC} \Pr(\symbolABC_{0}
\underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (\symbolABC_{0})) \nonumber \\
&= \sum_{\symbolABC_{0} \in \AlphabetABC} \sum_{\w \in \AlphabetABC{^{n-1}}}
\Pr(\symbolABC_{0} \, \w \, \xoperator (\symbolABC_{0})) \nonumber \\
&= \sum_{\symbolABC_{0} \in \AlphabetABC} \sum_{\w \in \AlphabetABC^{n-1}}
\bra \Dist \TransMatABC^{[\symbolABC_{0}]} \TransMatABC^{[\w]}
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \nonumber \\
&= \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]} \Bigl(\sum_{\w \in \AlphabetABC^{n-1}}
\TransMatABC^{[\w]} \Bigr) \TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \nonumber \\
&= \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist \TransMatABC^{[\symbolABC_{0}]}
\Bigl(\prod_{i=1}^{n-1}
\underbrace{\sum_{\symbolABC_{i} \in \AlphabetABC}
\TransMatABC^{[\symbolABC_{i}]} }_{ = \TransMatABC} \Bigr)
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \nonumber \\
&= \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]} \TransMatABC^{n-1}
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne ,
\label{Eq:TMsToCFs}
\end{align}
where the stationary distribution $\bra{\Dist}$ over states of the
$ABC$-machine is found from Eq.~(\ref{eq:ProbabilityDistribution}).
\emph{The most general connection between CFs and TMs is given by Eq. (\ref{Eq:TMsToCFs}).}
As before, we might assume on physical grounds that:
\begin{align}
\label{eq:EqualProbs}
\Pr(A \underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (A)) =
\Pr(B \underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (B)) =
\Pr(C \underbrace{* \dots *}_{n-1 \text{ } * \text{s}} \xoperator (C)).
\end{align}
For example, Eq.~\eqref{eq:EqualProbs} is always true of mixing {\MachineABC}s.
This special case yields the more constrained set of equations:
\begin{align}
\label{eq:ThreeTimesBraKetSimplification}
\Qxin = 3 \bra \Dist \TransMatABC^{[\symbolABC_{0}]} \TransMatABC^{n-1}
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne ~,
\end{align}
where $\symbolABC_{0} \in \AlphabetABC$.
\subsection{CFs from Spectral Decomposition}
\label{subsec:SpectralResults}
Although Eq.~\eqref{Eq:TMsToCFs} is itself an important result, we can also
apply a spectral decomposition of powers of the TM to provide a closed-form
that is even more useful and insightful. Ameliorating the computational burden,
this result reduces the matrix powers in the above expressions to expressions
involving only powers of scalars. Also, yielding theoretical insight, the
closed-forms reveal what types of behaviors can ever be expected of the CFs
from stacking processes described by finite HMMs.
The most familiar case occurs when the TM is diagonalizable. Then,
$\TransMatABC^{n-1}$ can be found via diagonalizing the TM, making use of the
fact that $\abcT^L = C D^L C^{-1}$, given the eigen-decomposition $\abcT = C D
C^{-1}$, where $D$ is the diagonal matrix of eigenvalues. However, to
understand the CF behavior, it is more appropriate to decompose the matrix in
terms of its projection operators.
Moreover, an analytic expression for $\TransMatABC^{n-1}$ can be found in terms
of the projection operators even when the TM is not diagonalizable. Details are
given elsewhere~\cite{Crut13a,Riec14a}. By way of summarizing, though, in the
general case the $L^{\text{th}}$ iteration of the TM follows from:
\begin{align}
\abcT^L = \mathcal{Z}^{-1} \left\{ \left( \IdentMat - z^{-1} \abcT \right) ^{-1} \right\}
\label{eq:MatrixPowersViaZtransform}
~,
\end{align}
where $\IdentMat$ is the $\NstatesABC \times \NstatesABC$ identity matrix, $z
\in \mathbb{C}$ is a continuous complex variable, and $\mathcal{Z}^{-1} \{ \cdot \}$ denotes the inverse $z$-transform~\cite{Oppe75a} defined to operate elementwise:
\begin{align}
\mathcal{Z}^{-1}\left( g(z) \right) & \equiv
\frac{1}{2 \pi i} \oint_{C} z^{L-1} g(z) \, dz
\end{align}
for the $z$-dependent matrix element $g(z)$ of $\left( \IdentMat - z^{-1}
\abcT \right)^{-1}$. Here, $\oint_C$ indicates a counterclockwise contour
integration in the complex plane enclosing the entire unit circle.
For nonnegative integers $L$, and with the allowance that $0^L = \delta_{L,0}$
for the case that $0 \in \Lambda_{\abcT}$, Eq.\ \eqref{eq:MatrixPowersViaZtransform} becomes:
\begin{align}
\abcT^L
& = \sum_{\lambda \in \Lambda_{\abcT} }
\sum_{m=0}^{\nu_\lambda - 1}
\lambda^{L-m} \binom{L}{m}
\abcT_\lambda
\left( \abcT - \lambda \IdentMat \right)^m
~,
\label{eq: T^L spectral decomp for positive integer L}
\end{align}
where $\Lambda_{\abcT} = \{ \lambda \in \mathbb{C}: \text{det}(\lambda
\IdentMat - \abcT) = 0 \}$ is the set of $\abcT$'s eigenvalues, $\abcT_\lambda$
is the projection operator associated with the eigenvalue $\lambda$ given by
the elementwise residue of the resolvent $\left( z \IdentMat - \abcT
\right)^{-1}$ at $z \to \lambda$, the index $\nu_\lambda$ of the eigenvalue
$\lambda$ is the size of the largest Jordan block associated with $\lambda$,
and $\binom{L}{m} = \frac{L!}{m! (L-m)!}$ is the binomial
coefficient.\footnote{Recall, \eg,
that $\binom{L}{0} = 1$, $\binom{L}{1} = L$, $\binom{L}{2} = \tfrac{1}{2!} L (L-1)$,
and $\binom{L}{L} = 1$.}
In terms of elementwise contour integration, we have:
\begin{align}
\abcT_\lambda = \frac{1}{2 \pi i} \oint_{C_\lambda} \left( z \IdentMat - \abcT \right)^{-1} \, dz,
\label{eq:GeneralProjOpEqn}
\end{align}
where $C_\lambda$ is any contour in the complex plane enclosing the point $z_0
= \lambda$---which may or may not be a singularity depending on the particular
element of the resolvent matrix---but encloses no other singularities.
As guaranteed by the Perron--Frobenius theorem, all eigenvalues of the
stochastic TM $\abcT$ lie on or within the unit circle. Moreover, the
eigenvalues on the unit circle are guaranteed to have index one. The indices
of all other eigenvalues must be less than or equal to one more than the
difference between their algebraic $a_\lambda$ and geometric $g_\lambda$
multiplicities. Specifically:
\begin{align*}
\nu_\lambda - 1 \leq a_\lambda - g_\lambda \leq a_\lambda - 1 ~\text{and}~
\nu_\lambda = 1 , \text{ if } |\lambda| = 1
~.
\end{align*}
Using Eq.~\eqref{eq: T^L spectral decomp for positive integer L} together with
Eq.~\eqref{Eq:TMsToCFs}, the CFs can now be expressed as:
\begin{align}
\Qxin & = \sum_{\lambda \in \Lambda_{\abcT}}
\sum_{m=0}^{\nu_\lambda - 1} \Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}} \binom{n-1}{m} \lambda^{n-m-1}
~,
\label{eq:GenCFdecomp}
\end{align}
where $\Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}}$ is a complex-valued scalar:\footnote{$\Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}}$ is
constant with respect to the relative layer displacement $n$. However, $\Big\{
\Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}} \Big\}$ can be a function of a process's parameters.}
\begin{align}
\Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}} \equiv
\sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]}
\abcT_\lambda \left( \abcT - \lambda \IdentMat \right)^m
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne
~.
\end{align}
Evidently, the CFs' mathematical form Eq.~\eqref{eq:GenCFdecomp} is strongly
constrained for any stacking process that can be described by a finite HMM.
Besides the expression's elegance, we note that its constrained form is very
useful for the so-called ``inverse problem'' of discovering the stacking
process from CFs~\cite{Varn02a,Varn07a,Varn13a,Varn13b}.
When $\abcT$ is diagonalizable, $\nu_\lambda = 1$ for all $\lambda$ so that
Eq.~\eqref{eq: T^L spectral decomp for positive integer L} simply reduces to:
\begin{align}
\abcT^L
& = \sum_{\lambda \in \Lambda_{\abcT} }
\lambda^L \abcT_\lambda
~,
\end{align}
where the projection operators can be obtained more simply as:
\begin{align}
\label{eq: T_lambda algorithm}
\abcT_{\lambda} & = \prod_{\substack{\zeta \in \Lambda_{\abcT} \\ \zeta \neq \lambda }}
\frac{\abcT - \zeta \IdentMat }{\lambda - \zeta}
~.
\end{align}
In the diagonalizable case, Eq.~\eqref{eq:GenCFdecomp} reduces to:
\begin{align}
\Qxin &= \sum_{\lambda \in \Lambda_{\abcT}} \lambda^{n-1}
\sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]} \abcT_\lambda
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \nonumber \\
& = \sum_{\lambda \in \Lambda_{\abcT}}
\Braket{\abcT_\lambda ^{\xi(\Alphabet)}} \lambda^{n-1}
~,
\label{eq:DiagCFdecomp}
\end{align}
where
$\Braket{\abcT_\lambda ^{\xi(\Alphabet)}} \equiv \Braket{\abcT_{\lambda, 0} ^{\xi(\Alphabet)}}$ is again a constant:
\begin{align}
\Braket{\abcT_\lambda ^{\xi(\Alphabet)}} & = \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]} \abcT_\lambda
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne
~.
\label{eq:DiagonalizableCorrelationAmplitudes}
\end{align}
\subsection{Asymptotic behavior of the CFs}
From the spectral decomposition, it is apparent that the CFs converge to
some constant value as $n \to \infty$, unless $\abcT$ has eigenvalues
on the unit circle besides unity itself.
If unity is the sole eigenvalue with a magnitude of one, then all
other eigenvalues have magnitude less than unity and their contributions decay to
negligibility for large enough $n$. Explicitly, if $\argmax_{\lambda \in
\Lambda_{\abcT}} |\lambda| = \{ 1 \}$, then:
\begin{align*}
\lim_{n \to \infty} \Qxin
& = \lim_{n \to \infty}
\sum_{\lambda \in \Lambda_{\abcT}}
\sum_{m=0}^{\nu_\lambda - 1}
\Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}} \binom{n-1}{m} \lambda^{n-m-1} \\
& = \Braket{\abcT_1 ^{\xi(\Alphabet)}} \\
& = \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]}
\abcT_1
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \\
& = \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]}
\DistOne \bra \Dist
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \\
& = \sum_{\symbolABC_{0} \in \AlphabetABC}
\Pr(\symbolABC_{0}) \Pr(\xoperator (\symbolABC_{0}))
~.
\end{align*}
Above, we used the fact that $\nu_1 = 1$ and that---for an ergodic
process---$\abcT_1 = \DistOne \bra \Dist$.
For mixing $ABC$-machines, $\Pr(\symbolABC) = 1 / 3$
for all $\symbolABC \in \AlphabetABC$. That this is so should
be evident from the graphical expansion method of \S\ref{GraphicalExpansion}.
Therefore, mixing processes with $\argmax_{\lambda \in
\Lambda_{\abcT}} |\lambda| = \{ 1 \}$ have CFs that all converge to $1/3$:
\begin{align*}
\lim_{n \to \infty} \Qxin
& = \sum_{\symbolABC_{0} \in \AlphabetABC}
\Pr(\symbolABC_{0}) \Pr(\xoperator (\symbolABC_{0})) \\
& = 3 ( \tfrac{1}{3} \times \tfrac{1}{3} ) \\
& = \tfrac{1}{3}
~.
\end{align*}
Non-mixing processes with $\argmax_{\lambda \in
\Lambda_{\abcT}} |\lambda| = \{ 1 \}$ can have their CFs converging to
constants other than $1/3$, depending on $\{ \Pr(\symbolABC) : \symbolABC \in \AlphabetABC \}$,
although they are still constrained by $\sum_{\xi} \Qxin = 1$.
If other eigenvalues in $\Lambda_{\TransMatABC}$ beside unity exist on the unit circle, then the CFs approach a periodic sequence as $n$ gets large.
\subsection{Modes of Decay}
Since $\TransMatABC$ has no more eigenvalues than its dimension (\ie,
$|\Lambda_{\TransMatABC}| \leq \NstatesABC$), {\emph{Eq.~\eqref{eq:GenCFdecomp}
implies that the number of states in the \MachineABC\ for a stacking process
puts an upper bound on the number of modes of decay.}} Indeed, since unity is
associated with stationarity, the number of modes of decay is strictly less
than $\NstatesABC$. It is important to note that these modes do not always
decay strictly exponentially: They are in general the product of a decaying
exponential with a polynomial in $n$, and the CFs are sums of these products.
Even if---due to diagonalizability of $\TransMatABC$---there were only strictly
exponentially decaying modes, it is simple but important to understand that
there is generally more than one mode of exponential decay present in the CFs.
And so, ventures to find \emph{the} decay constant of a process are misleading
unless it is explicitly acknowledged that one seeks, \eg, the slowest decay
mode. Even then, however, there are cases when the slowest decay mode only acts
on a component of the CFs with negligible magnitude. In an extreme case, the
slowest decay mode may not even be a large contributor to the CFs before the
whole pattern is numerically indistinguishable from the asymptotic value.
In analyzing a broad range of correlation functions, nevertheless, many authors have been led to consider \emph{correlation lengths}, also known as \emph{characteristic lengths}~\cite{Tiwa07a,Varn13b}. The form of Eq.~\eqref{eq:GenCFdecomp} suggests that this perspective will often be a clumsy oversimplification for understanding CFs. Regardless, if one wishes to assign a correlation length associated with an index-one mode of CF decay, we observe that the reciprocal of the correlation length is essentially the negative logarithm of the magnitude of the eigenvalue for that mode. We find that the typically reported correlation length $\ell_{\text{C}}$ derives from the
second-largest contributing magnitude among the eigenvalues:
\begin{align}
\ell_{\text{C}}^{-1} & = -\log |\zeta| ~,
& \text{for } \zeta \in \argmax_{\lambda \in \boldsymbol{\rho} } |\lambda|~,
\label{eq:CorrLen}
\end{align}
where $\boldsymbol{\rho} = \left\{ \lambda \in \Lambda_{\abcT} \setminus \{ 1 \} : \Braket{\abcT_\lambda ^{\xi(\Alphabet)}} \neq 0 \right\}$.
Guided by Eq.~\eqref{eq:GenCFdecomp}, we suggest that a true understanding of CF behavior
involves finding $\Lambda_{\TransMatABC}$ with the corresponding eigenvalue indices and the amplitude of each mode's contribution $\Big\{ \Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}} \Big\}$.
This now completes our theoretical development, and in the next section we apply these
techniques to three examples.
\section{Examples}
\label{Examples}
\subsection{3C Polytypes and Random ML Stacking: IID Processes}
\label{RandomStacking}
Although not often applicable in practice, as a pedagogical exercise the random
ML stacking has often been treated~\cite{Guin63a}. This stacking process is the
simplest stacking arrangement that can be imagined,\footnote{This is not mere
hyperbole. It is possible to quantify a process's structural organization in
the form of its \emph{statistical complexity} $\Cmu$, which measures the
internal information processing required to produce the
pattern~\cite{Crut89a,Crut12a,Varn13a}. In the present case $\Cmu = 0$ bits,
the minimum value.} and there are previous analytical results that can be
compared to the techniques developed here. In statistics parlance, this process
is an \emph{independent and identically distributed} (IID)
process~\cite{Cove06a}.
\begin{figure}
\begin{center}
\includegraphics[width=0.25\textwidth]{images/tikz/RandomProcessHagg.pdf}
\end{center}
\caption{\MachineHagg\ for the IID Process. When $q=1$, the IID Process
generates a string of $1$s, which is physically the 3C$^+$ stacking
structure. Conversely, when $q=0$, the structure corresponds to the 3C$^-$
structure. For $q = 1/2$, the MLs are stacked as randomly as possible.
}
\label{FigRandomProcessHagg}
\end{figure}
Let us assume that the placement of MLs is independent of the previous MLs
scanned, except that it of course must obey the stacking constraints. The
\MachineHagg\ that describes this process is shown in
Fig.~\ref{FigRandomProcessHagg}. We allow for the possibility that there might
be a bias in the stacking order, and we assign a probability $q$ that the next
layer is cyclically related to its predecessor. Thus, the 1-by-1 symbol-labeled
TMs for the \MachineHagg\ are:
\begin{align*}
\begin{array}{r@{\mskip\thickmuskip}l}
\HaggTone & = \begin{bmatrix} q \end{bmatrix}
\end{array}
{\rm {and}}
\begin{array}{r@{\mskip\thickmuskip}l}
\HaggTzero = \begin{bmatrix} \overline{q} \end{bmatrix} ,
\end{array}
\end{align*}
where $\overline{q} \equiv 1-q$, with $q \in [0,1]$.
The physical interpretation of the IID Process is straightforward. In the case
where $q = 1$, the process generates a stacking sequence of all $1$s, giving a
physical stacking structure of $\dots ABCABCABC \dots$. We recognize this as
the 3C$^{+}$ crystal structure. Similarly, for $q = 0$, the process generates
stacking sequence of all 0s, which is the 3C$^{-}$ crystal structure. For those
cases where $q$ is near but not quite at its extreme values, the stacking
structure is 3C with randomly distributed deformation faults. When $q=\OneHalf$, the
MLs are stacked in a completely random fashion.
Now, we must determine whether this is a mixing or nonmixing \MachineHagg. We
note that there are two SSCs, namely [$\StateHagg_{(0)}$] and
[$\StateHagg_{(1)}$]. The winding numbers for each are
$W^{[\StateHagg_{(1)}]} = 1$ and $W^{[\StateHagg_{(0)}]} = 2$, respectively.
Since at least one of these is not equal to zero, the \MachineHagg\ is mixing,
and we need to expand the \MachineHagg\ into the \MachineABC. This is shown in
Fig.~\ref{FigRandomProcessABC}.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{images/tikz/RandomProcessABC.pdf}
\end{center}
\caption{\MachineABC\ for the IID Process. The single state of the
\MachineHagg\ has expanded into three.
}
\label{FigRandomProcessABC}
\end{figure}
The \MachineABC\ TMs can either be directly written down from inspecting
Fig.~\ref{FigRandomProcessABC} or by using the rote expansion algorithm of
\S\ref{AlgebraicExpansion}, using Eqs.~\eqref{eq:T0toABC}
and~\eqref{eq:T1toABC}. By either method we find the 3-by-3 TMs to be:
\begin{align*}
\begin{array}{r@{\mskip\thickmuskip}l}
\abcTA
=
\begin{bmatrix}
0 & 0 & 0 \\
\overline{q} & 0 & 0 \\
q & 0 & 0
\end{bmatrix} ,
\end{array}
\begin{array}{r@{\mskip\thickmuskip}l}
\abcTB
=
\begin{bmatrix}
0 & q & 0 \\
0 & 0 & 0 \\
0 & \overline{q} & 0
\end{bmatrix}
\end{array}
\end{align*}
and
\begin{align}
\abcTC
=
\begin{bmatrix} \nonumber
0 & 0 & \overline{q} \\
0 & 0 & q \\
0 & 0 & 0
\end{bmatrix} .
\end{align}
The internal state TM then is their sum:
\begin{align}
\abcT
=
\begin{bmatrix} \nonumber
0 & q & \overline{q} \\
\overline{q} & 0 & q \\
q & \overline{q} & 0
\end{bmatrix}
~.
\end{align}
The eigenvalues of the $ABC$ TM are
\begin{align*}
\Lambda_{\abcT} = \{ 1, \, \Omega, \, \Omega^* \}~,
\end{align*}
where:
\begin{align*}
\Omega \equiv - \frac{1}{2} + i \frac{\sqrt{3}}{2} (4q^2 - 4q + 1 ) ^{1/2}
\end{align*}
and $\Omega^*$ is its complex conjugate.
Already, via Eq.~\eqref{eq:CorrLen}, we can identify what the characteristic length of the CFs will be.
In particular, $\ell_{\text{C}}^{-1} = -\log |\Omega| = -\tfrac{1}{2} \log (1 - 3q + 3q^2)$ yields:
\begin{align*}
\ell_{\text{C}} = -\frac{2}{\log(1 - 3q + 3q^2)}~.
\end{align*}
If we identify $q$ with the deformation faulting parameter $\alpha$ in the model introduced
by Estevez \etal~\cite{Este08a} (see the next example in \S\ref{GrowthDeformationStacking}, the RGDF Process),
this is identical to the result obtained there
in Eq. (35). There is much more structural information in the CFs, however,
than a single characteristic length would suggest. This fact will become
especially apparent as our examples become more sophisticated.
According to Eq.~\eqref{Eq:TMsToCFs}, we can obtain the CFs via:
\begin{align}
\Qxin &= \sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]} \TransMatABC^{n-1}
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne . \nonumber
\end{align}
The stationary distribution over the $ABC$-machine states is found from
Eq.~\eqref{eq:ProbabilityDistribution}:
\begin{align*}
\bra{\Dist} = \left[ \tfrac{1}{3} \quad \tfrac{1}{3} \quad \tfrac{1}{3} \right]
~.
\end{align*}
Furthermore, an analytic expression for $\abcT^{n-1}$ follows
from the $z$-transform as given in Eq.~\eqref{eq:MatrixPowersViaZtransform}.
As a start, we find:
\begin{align*}
\IdentMat - z^{-1} \abcT =
\begin{bmatrix}
1 & -q/z & -\overline{q}/z \\
-\overline{q}/z & 1 & -q/z \\
-q/z & -\overline{q}/z & 1
\end{bmatrix}
\end{align*}
and its inverse:
\begin{align*}
\left( \IdentMat - z^{-1} \abcT \right) ^{-1} & =
\frac{1}{(1-z^{-1})(1- \Omega z^{-1})(1- \Omega ^*z^{-1})} \\
\times & \begin{bmatrix}
1- q \overline{q} z^{-2} & qz^{-1}+ \overline{q}^2 z^{-2} & \overline{q} z^{-1}+ q^2 z^{-2} \\
\overline{q} z^{-1}+ q^2 z^{-2} & 1- q \overline{q} z^{-2} & qz^{-1}+ \overline{q}^2 z^{-2} \\
qz^{-1}+ \overline{q}^2 z^{-2} & \overline{q} z^{-1}+ q^2 z^{-2} & 1- q \overline{q} z^{-2}
\end{bmatrix} .
\end{align*}
Upon partial fraction expansion, we obtain:
\begin{align}
& \left( \IdentMat - z^{-1} \abcT \right) ^{-1} \nonumber \\
& = \frac{1}{3} \frac{1}{(1-z^{-1})}
\begin{bmatrix}
1 & 1 & 1 \\
1 & 1 & 1 \\
1 & 1 & 1
\end{bmatrix} \nonumber \\
& +
\tfrac{1}{(\Omega -1) (\Omega - \Omega ^*)}
\tfrac{1}{(1- \Omega z^{-1})}
\begin{bmatrix}
\Omega ^2 - q \overline{q} & q \Omega + \overline{q}^2 & \overline{q} \Omega + q^2 \\
\overline{q} \Omega+ q^2 & \Omega ^2 - q \overline{q} & q \Omega + \overline{q}^2 \\
q \Omega + \overline{q}^2 & \overline{q} \Omega + q^2 & \Omega ^2 - q \overline{q}
\end{bmatrix} \nonumber \\
& +
\tfrac{1}{(\Omega^* -1) (\Omega^* - \Omega)}
\tfrac{1}{(1- \Omega^* z^{-1})}
\begin{bmatrix}
{\Omega^*} ^2 - q \overline{q} & q \Omega^* + \overline{q}^2 & \overline{q} \Omega^* + q^2 \\
\overline{q} \Omega^*+ q^2 & {\Omega^*} ^2 - q \overline{q} & q \Omega^* + \overline{q}^2 \\
q \Omega^* + \overline{q}^2 & \overline{q} \Omega^* + q^2 & {\Omega^*} ^2 - q \overline{q}
\end{bmatrix}
,
\label{eq:RandomResolvent}
\end{align}
for $q \neq 1/2$. (The special case of $q = 1/2$ is discussed in the next
subsection.) Finally, we take the inverse $z$-transform of
Eq.~\eqref{eq:RandomResolvent} to obtain an expression for the $L^{\text{th}}$
iterate of the TM:
\begin{align*}
\abcT^L & = \mathcal{Z}^{-1}
\left\{ \left( \IdentMat - z^{-1} \abcT \right) ^{-1} \right\} \\
& =
\frac{1}{3}
\begin{bmatrix}
1 & 1 & 1 \\
1 & 1 & 1 \\
1 & 1 & 1
\end{bmatrix} \\
& ~~+
\frac{\Omega^L}{(\Omega -1) (\Omega - \Omega ^*)}
\begin{bmatrix}
\Omega ^2 - q \overline{q} & q \Omega + \overline{q}^2 & \overline{q} \Omega + q^2 \\
\overline{q} \Omega+ q^2 & \Omega ^2 - q \overline{q} & q \Omega + \overline{q}^2 \\
q \Omega + \overline{q}^2 & \overline{q} \Omega + q^2 & \Omega ^2 - q \overline{q}
\end{bmatrix} \\
& ~~+
\frac{{\Omega^*}^L}{(\Omega^* -1) (\Omega^* - \Omega)}
\begin{bmatrix}
{\Omega^*} ^2 - q \overline{q} & q \Omega^* + \overline{q}^2 & \overline{q} \Omega^* + q^2 \\
\overline{q} \Omega^*+ q^2 & {\Omega^*} ^2 - q \overline{q} & q \Omega^* + \overline{q}^2 \\
q \Omega^* + \overline{q}^2 & \overline{q} \Omega^* + q^2 & {\Omega^*} ^2 - q \overline{q}
\end{bmatrix}
.
\end{align*}
These pieces are all we need to calculate the CFs. Let's start with $\Qsn$.
First, we find:
\begin{align}
\bra \Dist \abcT^{[A]} =
\begin{bmatrix} \nonumber
\frac{1}{3} & 0 & 0
\end{bmatrix}
\end{align}
and:
\begin{align}
\abcT^{[\soperator(A)]} \DistOne = \abcT^{[A]} \DistOne =
\begin{bmatrix} \nonumber
0 \\
\overline{q} \\
q
\end{bmatrix} .
\end{align}
Then:
\begin{align}
\bra \Dist & \abcT^{[A]} \abcT^{n-1} = \tfrac{1}{9}
\begin{bmatrix}
1 & 1 & 1
\end{bmatrix} \nonumber \\
+ &
\tfrac{1}{3}
\tfrac{\Omega^{n-1}}{(\Omega -1) (\Omega - \Omega ^*)}
\begin{bmatrix}
\Omega ^2 - q \overline{q} & q \Omega + \overline{q}^2 & \overline{q} \Omega + q^2
\end{bmatrix} \nonumber \\
+ &
\tfrac{1}{3}
\tfrac{{\Omega^*}^{n-1}}{(\Omega^* -1) (\Omega^* - \Omega)}
\begin{bmatrix}
{\Omega^*} ^2 - q \overline{q} & q \Omega^* + \overline{q}^2 & \overline{q} \Omega^* + q^2
\end{bmatrix}
\label{eq:RandomPiATnBra}
\end{align}
and:
\begin{align}
\bra \Dist \abcT^{[A]} \abcT^{n-1} \abcT^{[A]} \DistOne & =
\tfrac{1}{9}
+ \tfrac{1}{3}
\tfrac{\Omega^{n-1}}{(\Omega -1) (\Omega - \Omega ^*)}
\left( 2 q \overline{q} \Omega + \Omega \Omega^* \right) \nonumber \\
+ &
\tfrac{1}{3}
\tfrac{{\Omega^*}^{n-1}}{(\Omega^* -1) (\Omega^* - \Omega)}
\left( 2 q \overline{q} \Omega^* + \Omega \Omega^* \right) .
\end{align}
One can verify that Eq.~\eqref{eq:ThreeTimesBraKetSimplification} can be applied in lieu of
Eq.~\eqref{Eq:TMsToCFs}, which saves some effort in finding the final result,
which is:
\begin{align}
\Qsn = 1/3 & +
2 \text{Re} \left\{ {
\frac{\Omega^{n}}{(\Omega -1) (\Omega - \Omega^*)}
\left( 2 q \overline{q} + \Omega^* \right)
} \right\}
.
\label{eq:generalIIDQs}
\end{align}
The cyclic and anticyclic CFs can also be calculated from
Eq.~\eqref{eq:ThreeTimesBraKetSimplification} using the result we have already
obtained in Eq.~\eqref{eq:RandomPiATnBra} and a quick calculation yields:
\begin{align}
\abcT^{[\coperator(A)]} \DistOne = \abcT^{[B]} \DistOne =
\begin{bmatrix} \nonumber
q \\
0 \\
\overline{q} \\
\end{bmatrix}
\end{align}
and
\begin{align}
\abcT^{[\aoperator(A)]} \DistOne = \abcT^{[C]} \DistOne =
\begin{bmatrix} \nonumber
\overline{q} \\
q \\
0
\end{bmatrix} .
\end{align}
Then, we have:
\begin{align}
\Qcn &= 3 \bra \Dist \abcT^{[A]} \abcT^{n-1} \abcT^{[B]} \DistOne \nonumber \\
&=
1/3
+
2 \text{Re} \left\{ {
\frac{\Omega^{n}}{(\Omega -1) (\Omega - \Omega^*)}
\left( \overline{q}^2 + q \Omega \right)
} \right\}
\label{eq:RandomStackingQcn}
\end{align}
and:
\begin{align}
\Qan &= 3 \bra \Dist \abcT^{[A]} \abcT^{n-1} \abcT^{[C]} \DistOne \nonumber \\
&=
1/3
+
2 \text{Re} \left\{ {
\frac{\Omega^{n}}{(\Omega -1) (\Omega - \Omega^*)}
\left( q^2 + \overline{q} \Omega \right)
} \right\} .
\label{eq:RandomStackingQan}
\end{align}
All of this subsection's results hold for the whole range of $q \in [0,
\tfrac{1}{2}) \cup (\tfrac{1}{2}, 1]$, where all $\abcT$'s eigenvalues are
distinct. However, for $q = 1 / 2$, the two complex conjugate
eigenvalues, $\Omega$ and $\Omega^*$, lose their imaginary components, becoming
repeated eigenvalues. This requires special treatment.\footnote{Indeed, the
straightforward $z$-transform approach yielding the CF equations given in this
section appears to need special treatment for $q = 1/2$. However, a
more direct spectral perspective as developed in \S
\ref{subsec:SpectralResults} shows that since $\abcT$ is diagonalizable for all
$q$, all eigenvalues have index of one and so yield CFs of the simple form of
Eq.~\eqref{eq:DiagCFdecomp}.} We address the case of $q = 1 / 2$ in the
next subsection, which is of interest in its own right as being the most random
possible stacking sequence allowed.
\subsubsection{A Fair Coin?}
When a close-packed structure has absolutely no underlying crystal order in
the direction normal to stacking, the stacking sequence is as random as it
possibly can be. This is the case of $q = 1/2$, where spins are effectively
assigned by a fair coin, which yields a symmetric TM with repeated eigenvalues.
Due to repeated eigenvalues, the CFs at least superficially obtain a special
form.
To obtain the CFs for the Fair Coin IID Process, we follow the procedure of the previous subsection, with all of the same results through Eq.~\eqref{eq:RandomResolvent}, which with $q = 1/2$ and $\left. \Omega \right|_{q = 1/2} = \left. \Omega^* \right|_{q = 1/2} = -1/2$ can now be written as:
\begin{align}
\left( \IdentMat - z^{-1} \abcT \right) ^{-1}
& =
\frac{1}{(1-z^{-1})(1+ \frac{1}{2} z^{-1})^2} \nonumber \\
\times
&
\begin{bmatrix} \nonumber
1- \frac{1}{4} z^{-2} & \frac{1}{2} z^{-1} + \frac{1}{4} z^{-2} & \frac{1}{2} z^{-1} + \frac{1}{4} z^{-2} \\
\frac{1}{2} z^{-1} + \frac{1}{4} z^{-2} & 1- \frac{1}{4} z^{-2} & \frac{1}{2} z^{-1}+ \frac{1}{4} z^{-2} \\
\frac{1}{2} z^{-1}+ \frac{1}{4} z^{-2} & \frac{1}{2} z^{-1}+ \frac{1}{4} z^{-2} & 1- \frac{1}{4} z^{-2}
\end{bmatrix} .
\end{align}
However, the repeated factor in the denominator yields a new partial fraction
expansion. Applying the inverse $z$-transform gives the $L^{\text{th}}$ iterate
of the TM\footnote{By inspection, we see from Eq.~\eqref{eq:DiagCFdecomp} that
$\abcT^0$ is the identity matrix and $\abcT^1 = \abcT$, as must be the case.
More interestingly, the decaying deviation from the asymptotic matrix is
oscillatory.} as:
\begin{align*}
\abcT^L = \mathcal{Z}^{-1}
&
\left\{ \left( \IdentMat - z^{-1} \abcT \right) ^{-1} \right\} \\
= \frac{1}{3}
&
\begin{bmatrix} \nonumber
1 & 1 & 1 \\
1 & 1 & 1 \\
1 & 1 & 1
\end{bmatrix}
+ \frac{1}{3}
\left(- \frac{1}{2}\right)^L
\begin{bmatrix}
2 & -1 & -1 \\
-1 & 2 & -1 \\
-1 & -1 & 2
\end{bmatrix}
~.
\end{align*}
Then, we find:
\begin{align}
\bra \Dist \abcT^{[A]} \abcT^{n-1} =
\frac{1}{9}
\begin{bmatrix} \nonumber
1 & 1 & 1
\end{bmatrix}
+
\frac{1}{9}
\left(- \frac{1}{2}\right)^{n-1}
\begin{bmatrix}
2 & -1 & -1
\end{bmatrix}
.
\end{align}
With the final result that:
\begin{align}
\Qsn &= 3 \bra \Dist \abcT^{[A]} \abcT^{n-1} \abcT^{[A]} \DistOne \nonumber \\
&=
\frac{1}{3}
+
\frac{2}{3} \left(- \frac{1}{2}\right)^{n} ,
\label{eq:IIDQs}
\end{align}
\begin{align}
\Qcn &= 3 \bra \Dist \abcT^{[A]} \abcT^{n-1} \abcT^{[B]} \DistOne \nonumber \\
&=
\frac{1}{3}
-
\frac{1}{3} \left(- \frac{1}{2}\right)^{n} ,
\end{align}
and
\begin{align}
\Qan &= 3 \bra \Dist \abcT^{[A]} \abcT^{n-1} \abcT^{[C]} \DistOne \nonumber \\
&=
\frac{1}{3}
-
\frac{1}{3} \left(- \frac{1}{2}\right)^{n} .
\end{align}
For $q = 1/2$, we see that $\Qcn$ and $\Qan$ are identical, but this is not
generally the case as one can check for other values of $q$ in
Eqs.~\eqref{eq:RandomStackingQcn} and~\eqref{eq:RandomStackingQan}.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/BernoulliABCTeigs.pdf}
\end{center}
\caption{TM's eigenvalues in the complex plane for the IID Process as $q$
is varied. Note that there is always an eigenvalue at $1$.
}
\label{fig:BernoulliABCTeigs}
\end{figure}
Figure~\ref{fig:BernoulliABCTeigs} shows a graph of the TM's eigenvalues in the
complex plane as $q$ is varied. Notice that there is an eigenvalue at $1$ for
all values of $q$. This is generic feature, and we always find such an
eigenvalue. The other two eigenvalues start at the other two cube roots of
unity for $q \in \{0,1\}$ and, as $q \to 1/2$, they migrate to the point $-1/2$
and become degenerate at $q = 1/2$. It is this degeneracy that requires the
special treatment given in this section.
It is interesting that even the Fair Coin \MachineHagg\ produces structured
CFs. This is because---even though the allowed transitions of the underlying
$ABC$-machine are randomized---not all transitions are allowed. For example, if
we start with an $A$ ML, the next ML has a zero probability of being an $A$, a
$1/2$ probability of being a $B$, and a $1/2$ probability of being a $C$. Then,
the \emph{next} ML has a rebounding $1/2$ probability of being an $A$ while the
probability of being either a $B$ or $C$ is each only $1/4$. So, we see that the
underlying process has structure, and there is nothing we can do---given the
physical constraints---to make the CFs completely random.
When we can compare our expressions for CFs at $q = 1/2$ to those derived
previously by elementary means~\cite{Guin63a,Varn01b}, we find agreement. Note
however that unlike in these earlier treatments, here there was no need to
assume a recursion relationship.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/Bernoulli_Qs_0p1q.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ for $q=0.1$ the IID Process.}
\label{fig:BernoulliQs0p1q}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/Bernoulli_Qs_0p3q.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ for $q=0.3$ the IID Process.}
\label{fig:BernoulliQs0p3q}
\end{figure}
Figures~\ref{fig:BernoulliQs0p1q} and~\ref{fig:BernoulliQs0p3q} show $\Qsn$
versus $n$ for the IID Process with $q=0.1$ and $q=0.3$, respectively, as
computed from Eq.~\eqref{eq:generalIIDQs}. In each case the CFs decay to an
asymptotic value of $1/3$, although this decay is faster for $q=0.3$. This is
not surprising, as one interpretation for the IID Process with $q=0.1$ is that
of a 3C$^+$ crystal interspersed with 10\% random deformation faults.
\subsection{Random Growth and Deformation Faults in Layered 3C and 2H CPSs: The RGDF Process}
\label{GrowthDeformationStacking}
Estevez \etal~\cite{Este08a} recently showed that simultaneous random
growth and deformation SFs in 2H and 3C CPSs can be modeled for all values of
the fault parameters by a simple HMM, and this is shown in
Fig.~\ref{fig:EstevezHagg}. We refer to this process as the \emph{Random
Growth and Deformation Faults} (RGDF) Process.\footnote{ Estevez
\etal~\cite{Este08a} give a thorough and detailed discussion of the RGDF
process, and readers interested in a comprehensive motivation and derivation of
the RGDF process are urged to consult that reference.} As has become
convention~\cite{Warr69a,Este08a}, $\alpha$ refers to deformation faulting and
$\beta$ refers to growth faults.
\begin{figure}
\begin{center}
\includegraphics[width=0.50\textwidth]{images/tikz/HaggM_of_Estevez08.pdf}
\end{center}
\caption{RGDF Process, first proposed by Estevez~\etal~\cite{Este08a} and
adapted here from Panel (c) of their Fig. (2). There is a slight change in
notation. We relabeled the states given as `f' and `b' by
Estevez~\etal~\cite{Este08a} as `$\mathcal{U}$' and `$\mathcal{V}$' and,
instead of drawing an arc for each of the possible eight transitions, we took
advantage of the multiple transitions between the same states and labeled
each arc with two transitions. There is, of course, no change in meaning;
this instead provides for slightly tidier illustration. Additionally, we
correct a typographical error in Estevez~\etal~\cite{Este08a} when we
relabel the transition ${\rm{b}} \xrightarrow{0|\overline{\alpha}\beta}
{\rm{b}}$ with $\mathcal{V} \xrightarrow{0|\overline{\alpha}\overline{\beta}}
\mathcal{V}$.
}
\label{fig:EstevezHagg}
\end{figure}
The HMM describing the RGDF Process is unlike any of the others considered here
in that on emission of a symbol from a state, the successor state is {\emph
{not}} uniquely specified. For example, $\mathcal{U} \xrightarrow{0}
\mathcal{U}$ and $\mathcal{U} \xrightarrow{0} \mathcal{V}$; \ie, being in state
$\mathcal{U}$ and emitting a 0 does not uniquely determine the next state. Such
a representations were previously called {\emph
{nondeterministic}}~\cite{Hopc79a}, but to avoid a conflict in terminology we
prefer the term {\emph {nonunifilar}}~\cite{Ephr02a,Elli09a}. Since \eMs\ are
unifilar~\cite{Crut89a,Shal01b}, the HMM representing the RGDF model is not an
\eM. Nonetheless, the techniques we have developed are applicable: CFs do not
require unifilar HMMs for their calculations, as do other properties such as
the entropy density.
Inspecting Fig.~\ref{fig:EstevezHagg}, the RGDF \MachineHagg's TMs are seen to
be (Eqs. (1) and (2) of Estevez \etal~\cite{Este08a}):
\begin{align*}
\begin{array}{r@{\mskip\thickmuskip}l}
\HaggTzero & =
\begin{bmatrix}
\alpha \overline{\beta} & \overline{\alpha} \beta \\
\alpha \beta & \overline{\alpha} \overline{\beta}
\end{bmatrix}
\end{array}
{\rm {and}}
\begin{array}{r@{\mskip\thickmuskip}l}
\HaggTone & =
\begin{bmatrix}
\overline{\alpha} \overline{\beta} & \alpha \beta \\
\overline{\alpha} \beta & \alpha \overline{\beta}
\end{bmatrix} ,
\end{array}
\end{align*}
where $\alpha \in [0, 1]$ and $\overline{\alpha} \equiv 1 - \alpha$, such that
$\alpha + \overline{\alpha} = 1$, and $\beta \in [0, 1]$ and $\overline{\beta}
\equiv 1 - \beta$, such that $\beta + \overline{\beta} = 1$. There are eight
SSCs and, if at least one of them has $W^{\rm {SSC}} \neq 0$, the \Hagg-machine
is mixing. The self-state transitions each generate a nonvanishing $W^{\rm
{SSC}}$, so for the \MachineHagg\ to be nonmixing, these transitions must be
absent. Indeed, there are only two SSCs that have vanishing winding numbers,
and these are $[\mathcal{U}_{(0)}\mathcal{V}_{(1)}]$ and
$[\mathcal{U}_{(1)}\mathcal{V}_{(0)}]$. These, and only these, SSCs can exist
if $\overline{\beta} = 0$ and $\alpha \in \{0,1\}$. Thus, the \MachineHagg\ is
nonmixing only for the parameter settings $\beta = 1$ and $\alpha \in \{0,1\}$,
which corresponds to the 2H crystal structure.
From the \Hagg-machine, we obtain the corresponding TMs of the $ABC$-machine for
$\alpha, \beta \in (0, 1)$ by the rote expansion method (\S \ref{AlgebraicExpansion}):
\begin{align}
\abcTA & =
\begin{bmatrix} \nonumber
0 & 0 & 0 & 0 & 0 & 0 \\
\alpha \overline{\beta} & 0 & 0 & \overline{\alpha} \beta & 0 & 0 \\
\overline{\alpha} \overline{\beta} & 0 & 0 & \alpha \beta & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
\alpha \beta & 0 & 0 & \overline{\alpha} \overline{\beta} & 0 & 0 \\
\overline{\alpha} \beta & 0 & 0 & \alpha \overline{\beta} & 0 & 0 \\
\end{bmatrix} , \\ \vspace{.05in}
\abcTB & =
\begin{bmatrix} \nonumber
0 & \overline{\alpha} \overline{\beta} & 0 & 0 & \alpha \beta & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & \alpha \overline{\beta} & 0 & 0 & \overline{\alpha} \beta & 0 \\
0 & \overline{\alpha} \beta & 0 & 0 & \alpha \overline{\beta} & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & \alpha \beta & 0 & 0 & \overline{\alpha} \overline{\beta} & 0 \\
\end{bmatrix} ,
\intertext{and}
\abcTC & =
\begin{bmatrix} \nonumber
0 & 0 & \alpha \overline{\beta} & 0 & 0 & \overline{\alpha} \beta \\
0 & 0 & \overline{\alpha} \overline{\beta} & 0 & 0 & \alpha \beta \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & \alpha \beta & 0 & 0 & \overline{\alpha} \overline{\beta} \\
0 & 0 & \overline{\alpha} \beta & 0 & 0 & \alpha \overline{\beta} \\
0 & 0 & 0 & 0 & 0 & 0 \\
\end{bmatrix} ,
\end{align}
and the orientation-agnostic state-to-state TM:
\begin{align}
\abcT & = \abcTA + \abcTB + \abcTC . \nonumber
\end{align}
Explicitly, we have:
\begin{align}
\abcT & =
\begin{bmatrix} \nonumber
0 & \overline{\alpha} \overline{\beta} & \alpha \overline{\beta} & 0 & \alpha \beta & \overline{\alpha} \beta \\
\alpha \overline{\beta} & 0 & \overline{\alpha} \overline{\beta} & \overline{\alpha} \beta & 0 & \alpha \beta \\
\overline{\alpha} \overline{\beta} & \alpha \overline{\beta} & 0 & \alpha \beta & \overline{\alpha} \beta & 0 \\
0 & \overline{\alpha} \beta & \alpha \beta & 0 & \alpha \overline{\beta} & \overline{\alpha} \overline{\beta} \\
\alpha \beta & 0 & \overline{\alpha} \beta & \overline{\alpha} \overline{\beta} & 0 & \alpha \overline{\beta} \\
\overline{\alpha} \beta & \alpha \beta & 0 & \alpha \overline{\beta} & \overline{\alpha} \overline{\beta} & 0 \\
\end{bmatrix} .
\end{align}
$\abcT$'s eigenvalues satisfy det$(\abcT - \lambda \IdentMat) = 0$. Here,
with $a \equiv \alpha \beta$, $b \equiv \alpha \overline{\beta}$, $c \equiv \overline{\alpha} \beta$, and $d \equiv \overline{\alpha} \overline{\beta}$,
we have:
\begin{align*}
\text{det}&(\abcT - \lambda \IdentMat) \\
& =
\bigl[ \bigl( \lambda - (b+d) \bigr)^2 - (a+c)^2 \bigr] \\
& \quad \times \bigl[ \lambda^2 + \lambda (b+d) + ac - bd - a^2 - c^2 + b^2 + d^2 \bigr]^2 \\
& = 0 ,
\end{align*}
from which we obtain the eigenvalues:
$\lambda = b+d \pm (a+c)$
and
$\lambda = \OneHalf(b+d) \pm \OneHalf\left[ 4(a+c)^2 - 3(b+d)^2 + 12(bd - ac) \right]^{\OneHalf} $.
To get back to $\alpha$s and $\beta$s, we note that
$a+c = \beta$, $b+d = \overline{\beta}$,
$ac = \beta^2 \alpha \overline{\alpha}$, and $bd = \overline{\beta}^2 \alpha \overline{\alpha}$.
It also follows that $b+d+a+c = 1$, $b+d - (a+c) = \overline{\beta} - \beta = 1 - 2\beta$, and $bd-ac = \alpha \overline{\alpha}(\overline{\beta}^2 - \beta^2) = \alpha \overline{\alpha}(1 - 2\beta) = \alpha \overline{\alpha}(\overline{\beta} - \beta) $.
Hence, after simplification, the set of $\abcT$'s eigenvalues can be written as:
\begin{align}
\Lambda_{\abcT}
\label{eq:EstevezEigsWSigma}
& = \left\{ 1, \, 1 - 2\beta, \, -\tfrac{1}{2} (1 - \beta) \pm \tfrac{1}{2} \sqrt{\sigma} \right\} ,
\end{align}
with
\begin{align}
\label{eq:ExpansionOfSigma}
\sigma &\equiv
4\beta^2 - 3\overline{\beta}^2 + 12 \alpha \overline{\alpha}(\overline{\beta} - \beta) \\
& = -3 + 12\alpha + 6\beta - 12\alpha^2 + \beta^2 - 24\alpha \beta + 24\alpha^2 \beta .
\end{align}
Except for measure-zero submanifolds along which the eigenvalues become extra
degenerate, throughout the parameter range the eigenvalues' algebraic
multiplicities are: $a_1 = 1$, $a_{1-2\beta} = 1$, $a_{- \tfrac{1}{2} (1 -
\beta + \sqrt{\sigma})} = 2$, and $a_{- \tfrac{1}{2} (1 - \beta -
\sqrt{\sigma})} = 2$.
Moreover, the \emph{index} of all eigenvalues is 1 except along $\sigma = 0$.
Immediately from the eigenvalues and their corresponding indices, we know all
possible characteristic modes of CF decay. All that remains is to find the
contributing amplitude of each characteristic mode. For comparison, note that
our $\sigma$ turns out to be equivalent to the all-important $-s^2$ term
defined in Eq.~(28) of Estevez \etal~\cite{Este08a}.
Eqs.~\eqref{eq:EstevezEigsWSigma} and \eqref{eq:ExpansionOfSigma} reveal an
obvious symmetry between $\alpha$ and $\overline{\alpha}$ that is \emph{not}
present between $\beta$ and $\overline{\beta}$. In particular, $\abcT$'s
eigenvalues are invariant under exchange of $\alpha$ and
$\overline{\alpha}$---the CFs will decay in the same manner for $\alpha$-values
symmetric about $1/2$. There is no such symmetry between $\beta$ and
$\overline{\beta}$. Parameter space organization is seen nicely in Panel (c) of
Fig.~6 from Estevez \etal~\cite{Este08a}. Importantly, in that figure
$\sigma = 0$ should be seen as the critical line organizing a phase transition in
parameter space. Here, we will show that the $\sigma = 0$ line actually
corresponds to nondiagonalizability of the TM and, thus, to
the qualitatively different polynomial
behavior in the decay of the CFs predicted by our Eq.~\eqref{eq:GenCFdecomp}.
Note that since $\abcT$ is doubly-stochastic (\ie, all rows sum to one
\emph{and} all columns sum to one), the all-ones vector is not only the right
eigenvector associated with the eigenvalue of unity, but also the left
eigenvector associated with unity. Moreover, since the stationary distribution
$\bra{\Dist}$ is the left eigenvector associated with unity (recall that $\bra
\Dist \abcT = \bra \Dist$), the stationary distribution is the uniform
distribution: $\bra \Dist = \tfrac{1}{6} \begin{bmatrix} 1&1&1&1&1&1
\end{bmatrix}$, \ie, $\bra \Dist = \tfrac{1}{6}
\DistOneBackward$, for $\alpha, \beta \in (0, 1)$. Hence, throughout this
range, the projection operator associated with unity is $\abcT_1 = \tfrac{1}{6}
\DistOne \DistOneBackward$.
It is interesting to note that the eigenvalue of $1-2\beta$ is associated with
the decay of out-of-equilibrium probability density between the \Hagg\ states
of $\mathcal{U}$ and $\mathcal{V}$---or at least between the $ABC$-state
clusters into which each of the \Hagg\ states have split. Indeed, from the
\Hagg\ machine: $\Lambda_{\HaggT} = \{ 1, \, 1-2\beta \}$. So, questions about
the relative occupations of the \Hagg\ states themselves are questions invoking
the $1-2\beta$ projection operator. However, due to the antisymmetry of output
orientations emitted from each of these \Hagg\ states, the $1-2\beta$
eigenvalue will not make any direct contribution towards answering questions
about the process's output statistics. Specifically, $\Braket{\abcT_{1-2\beta}
^{\xi(\Alphabet)}} = 0$ for all $\xi \in \{\rm{c}, \rm{a}, \rm{s} \}$.
Since $a_{1-2\beta} = 1$, the projection operator is simply the matrix
product of the right and left eigenvectors associated with $1-2\beta$. With
proper normalization, we have:
\begin{align}
\abcT_{1-2\beta} & =
\tfrac{1}{6}
\ket{\boldsymbol{1-2\beta}} \bra{\boldsymbol{1-2\beta}} \nonumber
\end{align}
with $\ket{\boldsymbol{1-2\beta}} = \begin{bmatrix} 1&1&1&-1&-1&-1 \end{bmatrix} ^\top$
and $\bra{\boldsymbol{1-2\beta}} = \begin{bmatrix} 1&1&1&-1&-1&-1
\end{bmatrix}$ where $\top$ denotes matrix transposition. Then, one can easily
check via Eq.~\eqref{eq:DiagonalizableCorrelationAmplitudes} that indeed
$\Braket {\abcT_{1-2\beta}^{\xi(\Alphabet)}} = 0$ for all $\xi \in \{\rm{c}, \rm{a}, \rm{s}
\}$.
To obtain an explicit expression for the CFs, we
must obtain the remaining projection operators. We can
always use Eq.~\eqref{eq:GeneralProjOpEqn}. However, to draw attention to
useful techniques,
we will break the remaining
analysis into two parts: one for $\sigma = 0$ and the other for $\sigma \neq
0$. In particular, for the case of $\sigma = 0$, we show that
nondiagonalizablity need not make the problem harder than the diagonalizable
case.
\subsubsection{$\sigma=0$:}
As mentioned earlier, the $\sigma = 0$ line is the critical line that organizes
a phase transition in the ML ordering. We also find that $\abcT$ is
nondiagonalizable only along the $\sigma = 0$ submanifold. For $\sigma = 0$,
the $\tfrac{1}{2} (1 - \beta) \pm \tfrac{1}{2} \sqrt{\sigma}$ eigenvalues of
Eq.~\eqref{eq:EstevezEigsWSigma} collapse to a single eigenvalue so that the
set of eigenvalues reduces to: $\left. \Lambda_{\abcT} \right|_{\sigma = 0} =
\left\{ 1, \, 1 - 2\beta, \, -\tfrac{1}{2} (1 - \beta) \right\} $ with
corresponding indices: $\nu_1 = 1$, $\nu_{1-2\beta} = 1$, and $\nu_{-
\overline{\beta}/2} = 2$.
In this case, the projection operators are simple to obtain.
As in the general case, we have:
\begin{align}
\abcT_1 &=
\tfrac{1}{6} \DistOne \DistOneBackward \nonumber \\
& =
\tfrac{1}{6}
\begin{bmatrix} \nonumber
1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1
\end{bmatrix}
\end{align}
and
\begin{align}
\abcT_{1-2\beta} & =
\tfrac{1}{6}
\ket{\boldsymbol{1-2\beta}} \bra{\boldsymbol{1-2\beta}} \nonumber \\
& =
\tfrac{1}{6}
\begin{bmatrix} \nonumber
1 & 1 & 1 & -1 & -1 & -1 \\
1 & 1 & 1 & -1 & -1 & -1 \\
1 & 1 & 1 & -1 & -1 & -1 \\
-1 & -1 & -1 & 1 & 1 & 1 \\
-1 & -1 & -1 & 1 & 1 & 1 \\
-1 & -1 & -1 & 1 & 1 & 1
\end{bmatrix} .
\end{align}
Recall that the projection operators sum to the identity: $\IdentMat =
\sum_{\lambda \in \Lambda_{\abcT}} \abcT_\lambda = \abcT_1 + \abcT_{1-2\beta} +
\abcT_{-\overline{\beta}/2} $. And so, it is easy to obtain the
remaining projection operator:
\begin{align}
\abcT_{-\overline{\beta}/2} & =
\IdentMat - \abcT_1 - \abcT_{1-2\beta} \nonumber \\
& =
\tfrac{1}{3}
\begin{bmatrix} \nonumber
2 & -1 & -1 & 0 & 0 & 0 \\
-1 & 2 & -1 & 0 & 0 & 0 \\
-1 & -1 & 2 & 0 & 0 & 0 \\
0 & 0 & 0 & 2 & -1 & -1 \\
0 & 0 & 0 & -1 & 2 & -1 \\
0 & 0 & 0 & -1 & -1 & 2
\end{bmatrix} .
\end{align}
Note that $3 \bra \Dist \abcTA = \tfrac{1}{2} \DistOneBackward \abcTA =
\tfrac{1}{2} \begin{bmatrix} 1 & 0 & 0 & 1 & 0 & 0 \end{bmatrix}$ and
that:
\begin{align}
\abcTA \DistOne = &
\begin{bmatrix}
0 \\
\alpha \overline{\beta} + \overline{\alpha} \beta \\
\alpha \beta + \overline{\alpha} \overline{\beta} \\
0 \\
\alpha \beta + \overline{\alpha} \overline{\beta} \\
\alpha \overline{\beta} + \overline{\alpha} \beta \\
\end{bmatrix} ,
\abcTB \DistOne =
\begin{bmatrix}
\alpha \beta + \overline{\alpha} \overline{\beta} \\
0 \\
\alpha \overline{\beta} + \overline{\alpha} \beta \\
\alpha \overline{\beta} + \overline{\alpha} \beta \\
0 \\
\alpha \beta + \overline{\alpha} \overline{\beta} \\
\end{bmatrix} , \nonumber \\
\text{ and } & \;
\abcTC \DistOne =
\begin{bmatrix}
\alpha \overline{\beta} + \overline{\alpha} \beta \\
\alpha \beta + \overline{\alpha} \overline{\beta} \\
0 \\
\alpha \beta + \overline{\alpha} \overline{\beta} \\
\alpha \overline{\beta} + \overline{\alpha} \beta \\
0 \\
\end{bmatrix} ~. \nonumber
\end{align}
Then,
according to Eq.\ \eqref{eq:GenCFdecomp}, with
$\Braket{\abcT_1 ^{\xi(\Alphabet)}} = \tfrac{1}{3}$,
$\Braket{\abcT_{1-2\beta} ^{\xi(\Alphabet)}} = 0$,
$\Braket{\abcT_{-\overline{\beta}/2} ^{s(\Alphabet)}} = -\tfrac{1}{3}$,
$\Braket{\abcT_{-\overline{\beta}/2} ^{c(\Alphabet)}} =
\Braket{\abcT_{-\overline{\beta}/2} ^{a(\Alphabet)}} = \tfrac{1}{6}$,
$\Braket{\abcT_{-\overline{\beta}/2, 1} ^{s(\Alphabet)}} =
\tfrac{1}{6} (\sigma + \beta - \beta^2) = \tfrac{1}{6} \beta \overline{\beta}$, and
$\Braket{\abcT_{-\overline{\beta}/2, 1} ^{c(\Alphabet)}} =
\Braket{\abcT_{-\overline{\beta}/2, 1} ^{a(\Alphabet)}} =
-\tfrac{1}{12} (\sigma + \beta - \beta^2) = -\tfrac{1}{12} \beta \overline{\beta}$, the CFs are:
\begin{align}
\Qxin & =
\sum_{\lambda \in \Lambda_{\abcT}}
\sum_{m=0}^{\nu_\lambda - 1}
\Braket{\abcT_{\lambda, m} ^{\xi(\Alphabet)}} \binom{n-1}{m} \lambda^{n-m-1} \nonumber \\
& = \Braket{\abcT_1 ^{\xi(\Alphabet)}} +
\sum_{m=0}^{1}
\Braket{\abcT_{-\overline{\beta}/2, m} ^{\xi(\Alphabet)}} \binom{n-1}{m} {\left(-\overline{\beta}/2 \right)}^{n-m-1} \nonumber \\
& = \tfrac{1}{3} + \left[ \Braket{\abcT_{-\overline{\beta}/2} ^{\xi(\Alphabet)}} - \frac{2}{\overline{\beta}} \Braket{\abcT_{-\overline{\beta}/2, 1} ^{\xi(\Alphabet)}} (n-1) \right] {\left(-\overline{\beta}/2 \right)}^{n-1} \nonumber
~.
\end{align}
Specifically:
\begin{align}
\Qsn & = \tfrac{1}{3} \left[ 1 + 2 \left( 1 + \frac{\beta}{\overline{\beta}} \, n \right) {\left(-\overline{\beta}/2 \right)}^{n} \right]~,
\end{align}
and
\begin{align}
\Qcn =
\Qan = \tfrac{1}{3} \left[ 1 - \left( 1 + \frac{\beta}{\overline{\beta}} \, n \right) {\left(-\overline{\beta}/2 \right)}^{n} \right]~.
\end{align}
\subsubsection{$\sigma \neq 0$:}
For any value of $\sigma$, excluding of course $\sigma = 0$, we can obtain
the projection operators via Eq.~\eqref{eq:GeneralProjOpEqn}.
In addition to those quoted above and, in terms of the former $\abcT_{-\overline{\beta}/2}$, the remaining projection operators turn out to be:
\begin{align*}
\abcT_{\frac{-\overline{\beta} \pm \sqrt{\sigma}}{2}}
= \pm \frac{1}{\sqrt{\sigma}} \abcT_{-\overline{\beta}/2}
\left[ \abcT + \left( \frac{\overline{\beta} \pm \sqrt{\sigma} }{2} \right) \IdentMat \right]~.
\end{align*}
Since the $1-2\beta$ eigen-contribution is null and since:
\begin{align*}
\Braket{T_{1}^{\xi(\Alphabet)}} & = 1/3 ~, \\
\Braket{T_{\frac{-\overline{\beta} \pm \sqrt{\sigma}}{2}}^{s(\Alphabet)}}
& = \frac{1}{6} \left[ -1 \pm \left( \sqrt{\sigma}
+ \frac{\beta \overline{\beta}}{\sqrt{\sigma}} \right) \right] \\
& = \pm \frac{1}{6} \left( 1 \mp \frac{\beta}{\sqrt{\sigma}} \right)
\left( \sqrt{\sigma} \mp \overline{\beta} \right) ~, \text{~and} \\
\Braket{T_{\frac{-\overline{\beta} \pm \sqrt{\sigma}}{2}}^{c(\Alphabet)}}
& = \Braket{T_{\frac{-\overline{\beta} \pm \sqrt{\sigma}}{2}}^{a(\Alphabet)}}
\\
& = \frac{1}{12} \left[ 1 \mp \left( \sqrt{\sigma} + \frac{\beta
\overline{\beta}}{\sqrt{\sigma}} \right) \right] \\
& = \mp \frac{1}{12} \left( 1 \mp \frac{\beta}{\sqrt{\sigma}} \right)
\left( \sqrt{\sigma} \mp \overline{\beta} \right)
~,
\end{align*}
the CFs for $\sigma \neq 0$ are:
\begin{align}
\Qxin &=
\sum_{\lambda \in \Lambda_{\abcT}}
\lambda^{n-1}
\sum_{\symbolABC_{0} \in \AlphabetABC} \bra \Dist
\TransMatABC^{[\symbolABC_{0}]} \abcT_\lambda
\TransMatABC^{[\xoperator (\symbolABC_{0})]} \DistOne \nonumber \\
& =
\tfrac{1}{3} +
\sum_{\lambda \in \{
\frac{-\overline{\beta} \pm \sqrt{\sigma}}{2}
\}}
\Braket{\abcT_\lambda ^{\xi(\Alphabet)}} \lambda^{n-1} ~.
\label{eq:RandomStackingQxinWith2ProjOps}
\end{align}
Specifically, for $\xi = s$:
\begin{align}
\Qsn = \tfrac{1}{3}
&+ \tfrac{1}{6} \left( 1 - \tfrac{\beta}{\sqrt{\sigma}} \right) \left( \sqrt{\sigma} - \overline{\beta} \right)
\left( \tfrac{-\overline{\beta} + \sqrt{\sigma}}{2} \right)^{n-1} \nonumber \\
&- \tfrac{1}{6} \left( 1 + \tfrac{\beta}{\sqrt{\sigma}} \right) \left( \sqrt{\sigma} + \overline{\beta} \right)
\left( \tfrac{-\overline{\beta} - \sqrt{\sigma}}{2} \right)^{n-1} \\
\nonumber \\
=
\tfrac{1}{3} \Bigl[ 1 + &
\left( 1 - \tfrac{\beta}{\sqrt{\sigma}} \right) \left( \tfrac{-\overline{\beta} + \sqrt{\sigma}}{2} \right)^{n}
+ \left( 1 + \tfrac{\beta}{\sqrt{\sigma}} \right) \left( \tfrac{-\overline{\beta} - \sqrt{\sigma}}{2} \right)^{n}
\Bigr]~, \nonumber
\end{align}
and we recover Eq.~(29) of
Estevez~\etal~\cite{Este08a}.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/Estevez_Qs_0p01a_0b.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ with $\alpha = 0.01$ and $\beta = 0$ for the RGDF
Process. This should be compared to Panel (b) of Fig.~8 in
Estevez~\etal~\cite{Este08a}. Although different means were used to make
the calculations, they appear to be identical.
}
\label{fig:Estevez_Qs_0p01a_0b}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/Estevez_Qs_0p01a_0p01b.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ with $\alpha = 0.01$ and $\beta = 0.01$ for the RGDF
Process. Comparison with Panel (d) of Fig.~8 in
Estevez~\etal~\cite{Este08a} shows an identical result.
}
\label{fig:Estevez_Qs_0p01a_0b01}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/Estevez_Qs_0p1a_0p2b.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ with $\alpha = 0.1$ and $\beta = 0.2$ for the RGDF
Process.
}
\label{fig:Estevez_Qs_0p1a_0b2}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/Estevez_Qs_0p2a_0p1b.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ with $\alpha = 0.2$ and $\beta = 0.1$ for the RGDF
Process.
}
\label{fig:Estevez_Qs_0p2a_0b1}
\end{figure}
Estevez \etal~\cite{Este08a} recount the embarrassingly long list of recent
failures of previous attempts to analyze organization in RGDF-like processes.
These failures resulted from not obtaining all of the terms in the CFs, which in turn
stem primarily from not using a sufficiently clever ansatz in their
methods, together with not knowing how many terms there should be. In contrast,
even when casually observing the number of HMM states, our method gives
immediate knowledge of the number of terms. Our method is generally applicable
with straightforward steps to actually calculate all the terms once and for
all.
Figures~\ref{fig:Estevez_Qs_0p01a_0b},~\ref{fig:Estevez_Qs_0p01a_0b01},~\ref{fig:Estevez_Qs_0p1a_0b2}
and~\ref{fig:Estevez_Qs_0p2a_0b1} show plots of $\Qsn$ versus $n$ for the
RGDF Process at different values of $\alpha$ and $\beta$. The first two graphs,
Figs.~\ref{fig:Estevez_Qs_0p01a_0b} and~\ref{fig:Estevez_Qs_0p01a_0b01}, were
previously produced by Estevez~\etal~\cite{Este08a} and appear to be
identical to our results. The second pair of graphs for the IID Process,
Figs.~\ref{fig:Estevez_Qs_0p1a_0b2} and~\ref{fig:Estevez_Qs_0p2a_0b1} show the
behavior of the CFs for larger values of $\alpha$ and $\beta$, but with the
numerical values of each exchanged ($0.1 \Leftrightarrow 0.2$). The CFs are
clearly sensitive to the kind of faulting present, as one would expect.
However, each does decay to $1/3$, as they must.
\subsection{Shockley--Frank Stacking Faults in 6H-SiC: The SFSF Process}
\label{ShockleyFrankStacking}
While promising as a material for next generation electronic components,
fabricating SiC crystals of a specified polytype remains challenging. Recently
Sun~\etal~\cite{Sun12a} reported experiments on 6H-SiC epilayers ($\sim 200
\, \mu$m thick) produced by the fast sublimation growth process at 1775 $^{\circ}$C.
Using high resolution transmission electron microscopy (HRTEM), they were able
to survey the kind and amount of particular SFs present. In the \Hagg\ notation 6H-SiC is
specified by 000111, and this is written in the Zhdanov notation as (3,3)~\cite{Orti13a}.
Thus, unfaulted 6H-SiC can be thought of as alternating blocks of size-three domains.
Ab initio super-cell calculations by
Iwata~\etal~\cite{Iwat03a} predicted that the Shockley defects (4,2),
(5,1), (9,3), and (10,2) should be present, with the (4,2) defect having the
lowest energy and, thus, it presumably should be the most common. Of these, however,
Sun~\etal~\cite{Sun12a} observed only the (9,3) defect (given there
as (3,9)) and, at that, only once. Instead, the most commonly observed
defects were (3,4), (3,5), (3,6), and (3,7), appearing nine, two, two, and three
times respectively, with isolated instances of other SF sequences. They
postulated that combined Shockley--Frank defects~\cite{Hirt68a} could produce
these results. The (3,4) stacking sequences could be explained as external
Frank SFs, and the other observed faults could result from further Shockley
defects merging with these (3,4) SFs. We call this process the
\emph{Shockley-Frank Stacking Fault} (SFSF) Process.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/tikz/SunHaggMachine.pdf}
\end{center}
\caption{\MachineHagg\ for the SFSF Process, inspired by the observations of
Sun \etal~\cite{Sun12a}. We observe that there is one faulting parameter
$\gamma \in [0,1]$ and three SSCs. Or, equivalently three CSCs, as this graph
is also an \eM. The three SSCs are [$\sf{S}_7$], [$\sf{S}_0]$ and [$\sf{S}_7
\sf{S}_6 \sf{S}_4 \sf{S}_0 \sf{S}_1 \sf{S}_3$]. The latter we recognize as
the 6H structure if $\gamma = 0$. For large values of $\gamma$, \ie, as
$\gamma \to 1$, this process approaches a twinned 3C structure, although the
faulting is {\emph {not}} random. The causal state architecture prevents the
occurrence of domains of size-three or less.
}
\label{fig:SunHaggMachine}
\end{figure}
Inspired by these observations, we ask what causal-state structure could
produce such stacking sequences. We suggest that the \eM\ shown in
Fig.~\ref{fig:SunHaggMachine} is a potential candidate, with $\gamma \in [0,1]$
as the sole faulting parameter. (Here, we must insist that only a thorough
analysis, with significantly more HRTEM data or a DP, can properly reveal the
appropriate causal-state structure. The SFSF Process is given primarily to
illustrate our methods.) For weakly faulted crystals ($\gamma \approx 0$), as
seems to be the case here, there must be a CSC that gives the 6H structure, and we
see that the causal-state sequence [$\sf{S}_7 \sf{S}_6 \sf{S}_4 \sf{S}_0
\sf{S}_1 \sf{S}_3$] does that. Indeed, if the fault parameter $\gamma$ were
identically zero, then this \eM\ would give only the 6H structure.
Sun~\etal~\cite{Sun12a}'s observations suggest that deviations from 6H
structure occur (almost) always as {\emph {additions}} to the size-three 0 or 1
domains. The self-state transitions on $\sf{S}_7$ and $\sf{S}_0$ have just this
effect: After seeing three consecutive 1s (0s), with probability $\gamma$ the
current domain will increase in size to four. And likewise, with probability
$\gamma$, size-four domains will increase to size-five domains. Thus, with
decreasing probability, the faults (3,4), (3,5) $\dots$ can be modeled by this
\eM. Notice that the causal state architecture prevents domains of any size
less that three, which is consistent with the bulk of the observations by
Sun~\etal~\cite{Sun12a}.\footnote{They did observe a single (3,2) sequence
(see their Table I), and the SFSF Process cannot reproduce that structure.
Additional causal states and/or transitions would be needed to accommodate this
additional stacking structure. One obvious and simple modification that would
produce domains of size-two would be to allow the transitions $\sf{S}_3
\xrightarrow{0} \sf{S}_6$ and $\sf{S}_4 \xrightarrow{1} \sf{S}_1$ with some
small probability. However, in the interest maintaining a reasonably clear
example, we neglect this possibility.} Also, this \eM\ does predict (4,4)
sequences, which Sun~\etal~\cite{Sun12a} observed once. Thus,
qualitatively, and approximately quantitatively, the proposed \eM\ largely
reproduces the observations of Sun~\etal~\cite{Sun12a}.
We begin by identifying the SSCs on the HMM, the \eM\ shown in
Fig.~\ref{fig:SunHaggMachine}. We find that there are three, [$\sf{S}_7$],
[$\sf{S}_0]$ and [$\sf{S}_7 \sf{S}_6 \sf{S}_4 \sf{S}_0 \sf{S}_1 \sf{S}_3$]. We
calculate the winding numbers to be $W^{[\sf{S}_7]} = 1$, $W^{[\sf{S}_0]} = 2$,
and $W^{[\sf{S}_7 \sf{S}_6 \sf{S}_4 \sf{S}_0 \sf{S}_1 \sf{S}_3]} = 0$. The
first two of these SSCs vanish if $\gamma = 0$, giving a nonmixing
\MachineHagg. Thus, for $\gamma \neq 0$ the \MachineHagg\ is mixing and we
proceed with the case of $\gamma \in (0,1]$.
By inspection we write down the two 6-by-6 TMs of the \MachineHagg\ as:
\begin{align}
\HaggTzero & =
\begin{bmatrix} \nonumber
\gamma & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & \overline{\gamma} & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 & 0 & 0 \\
\end{bmatrix}
\intertext{and:}
\HaggTone & =
\begin{bmatrix} \nonumber
0 & \overline{\gamma} & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & \gamma & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
\end{bmatrix} ,
\end{align}
where the states are ordered $\sf{S}_0$, $\sf{S}_1$, $\sf{S}_3$, $\sf{S}_7$,
$\sf{S}_6$, and $\sf{S}_4$. The internal state TM is their sum:
\begin{align*}
\HaggT & =
\begin{bmatrix} \nonumber
\gamma & \overline{\gamma} & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & \gamma & \overline{\gamma} & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 & 0 & 0 \\
\end{bmatrix} .
\end{align*}
Since the six-state \MachineHagg\ generates an ($3 \times 6 = $) eighteen-state
\MachineABC, we do not explicitly write out the TMs of the \MachineABC.
Nevertheless, it is straightforward to expand the \MachineHagg\ to the
\MachineABC\ via the rote expansion method of \S\ref{AlgebraicExpansion}. It
is also straightforward to apply Eq.~\eqref{eq:ThreeTimesBraKetSimplification}
to obtain the CFs as a function of the faulting parameter $\gamma$. To use
Eq.~\eqref{eq:ThreeTimesBraKetSimplification}, note that the stationary
distribution over the \MachineABC\ can be obtained via
Eq.~\eqref{eq:PiExpansion} with:
\begin{align*}
\bra{\Dist_{\textrm{H}}} &=
\tfrac{1}{6 - 4 \gamma}
\begin{bmatrix} 1 & \overline{\gamma} & \overline{\gamma} & 1 & \overline{\gamma} & \overline{\gamma} \end{bmatrix}
\end{align*}
as the stationary distribution over the \MachineHagg.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/TeigsColored.pdf}
\end{center}
\caption{The six eigenvalues of the \Hagg-machine as they evolve from
$\gamma = 0$ (thickest blue markings) to $\gamma = 1$ (thinnest red
markings). Note that the eigenvalues at $\gamma = 0$ are the six roots of
unity. Unity is a persistent eigenvalue. Four of the eigenvalues approach 0
as $\gamma \to 1$. Another of the eigenvalues approaches unity as $\gamma \to
1$. The eigenvalues are nondegenerate throughout the parameter range except
for the transformation event where the two eigenvalues on the right collide
and scatter upon losing their imaginary parts.
}
\label{fig:RealImagPartsHagg}
\end{figure}
The eigenvalues of the \Hagg-TM can be obtained as the solutions of
$\det(\HaggT - \lambda \IdentMat) = (\lambda - \gamma)^2 \lambda^4 -
\overline{\gamma}^2 = 0$. These include $1$, $-\tfrac{1}{2} \overline{\gamma}
\pm \sqrt{\gamma^2 + 2\gamma - 3}$, and three other eigenvalues involving cube
roots. Their values are plotted in the complex plane
Fig.~\ref{fig:RealImagPartsHagg} as we sweep through $\gamma$.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/ABCTeigsColored.pdf}
\end{center}
\caption{The eighteen eigenvalues of the \MachineABC\ as they evolve from
$\gamma = 0$ (thickest blue markings) to $\gamma = 1$ (thinnest red
markings). Note that the eigenvalues at $\gamma = 0$ are still the six roots
of unity. The new eigenvalues introduced via transformation to the
\MachineABC\ all appear in degenerate (but diagonalizable) pairs. In terms of
increasing $\gamma$, these include eigenvalues approaching zero from $\pm 1$,
eigenvalues taking a left branch towards zero as they lose their imaginary
parts, and eigenvalues looping away and back towards the nontrivial
cube-roots of unity.
}
\label{fig:RealImagPartsABC}
\end{figure}
The eigenvalues of the $ABC$-TM are similarly obtained as the solutions of
$\det(\abcT - \lambda \IdentMat) = 0$. The real and imaginary parts of these
eigenvalues are plotted in Fig.\ \ref{fig:RealImagPartsABC}. Note that
$\Lambda_{\abcT}$ inherits $\Lambda_{\HaggT}$ as the backbone for its more
complex structure, just as $\Lambda_{\HaggT} \subseteq \Lambda_{\abcT}$ for all
of our previous examples. The eigenvalues in $\Lambda_{\abcT}$ are, of course,
those most directly responsible for the structure of the CFs.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/SFrank_Qs_0p01gamma.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ for the SFSF Process with $\gamma = 0.01$. This specimen is only
very weakly faulted and, hence, there are small decay constants giving
a slow decay to $1/3$.
}
\label{fig:SFrankQs0p01gamma}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/SFrank_Qs_0p1gamma.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ for the SFSF Process with $\gamma = 0.1$.
With increasing $\gamma$, the CFs approach their asymptotic value of
$1/3$ much more quickly.
}
\label{fig:SFrankQs0p1gamma}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/SFrank_Qs_0p5gamma.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ for the SFSF Process with $\gamma = 0.5$.
Here, the specimen is quite disordered, and the CFs decay quickly.
}
\label{fig:SFrankQs0p5gamma}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{images/graphs/SFrank_Qs_0p9gamma.pdf}
\end{center}
\caption{$\Qsn$ vs. $n$ for the SFSF Process with $\gamma = 0.9$.
The slower CF decay suggests that the process is now less disordered
than the $\gamma = 0.5$ case.
Notice that this CF is large for $n \mod(3) = 0$, indicating strong correlation
between MLs separated by a multiple of three MLs. This is the kind of
behavior that one expects from a twinned 3C crystal.
}
\label{fig:SFrankQs0p9gamma}
\end{figure}
The SFSF Process's CFs are shown for several example parameter values of
$\gamma$ in
Figs.~\ref{fig:SFrankQs0p01gamma},~\ref{fig:SFrankQs0p1gamma},~\ref{fig:SFrankQs0p5gamma},
and~\ref{fig:SFrankQs0p9gamma} calculated directly from numerical
implementation of Eq.~\eqref{eq:ThreeTimesBraKetSimplification}. As the
faulting parameter is increased from $0.01 \to 0.5$, the CFs begin to decay
more quickly. However, for $\gamma = 0.9$, the correlation length increases as
the eigenvalues, near the nontrivial cube-roots of unity, loop back toward the
unit circle. The behavior near $\gamma = 0.9$ suggests a longer ranged and
more regularly structured specimen, even though there are fewer significant
eigen-contributions to the specimen's structure. Indeed, the bulk of the
structure is now more apparent but less sophisticated.
\section{Conclusion}
\label{Conclusions}
We introduced a new approach to exactly determining CFs directly from HMMs that
describe layered CPSs. The calculation can be done either with high numerical
accuracy and efficiency, as we have shown in the CF plots for each example, or
analytically, as was done for the IID and RGDF Processes.
The mathematical representation that assumes central importance here is the
HMM. While we appreciate the value that studying CFs and, more generally, PDFs
brings to understanding material structure, pairwise
correlation information is better thought of as a consequence of a more
fundamental object (\ie, the HMM)
than one of intrinsic importance. This becomes clear when we consider that the
structure is completely contained in the very compact HMM representation. More
to the point, all of the correlation information is directly calculable from
it, as we demonstrated. In contrast, the task of inverting correlation
information to specify the underlying organization of a material's structure,
\ie, its HMM, is highly nontrivial. Over the past century considerable effort
has been expended to invert DPs, the Fourier transform of the CFs, into these
compact structural models.\footnote{We have not explicitly made the connection
here, but almost all previous models of planar disorder can generically be
expressed as HMMs.} The work of Warren~\cite{Warr69a}, Krishna and
coworkers~\cite{Seba84a,Seba87a,Seba87b,Seba87c,Seba87d}, Berliner \&
Werner~\cite{Berl86a} and that of our own
group~\cite{Varn02a,Varn07a,Varn13a,Varn13b}, to mention a few, all stand in
testament to this effort.
Although the presentation concentrated on CFs in layered CPSs, the potential
impact of the new approach is far wider. First, we note that it was necessary
to make some assumptions about the geometry of the stacking process, \ie, the
number of possible orientations of each ML and how two MLs can be stacked, in
order to demonstrate numerical results and make contact with previous work.
These assumptions however in no way limit the applicability: Any set of
stacking rules over a finite number of possible positions is amenable to this
treatment. Second, it may seem that starting with a HMM is unnecessarily
restrictive. It is not. Given a sample of the stacking process from (say) a
simulation study, there are techniques that now have become standard for
finding the \eM, a kind of HMM, that describes the process. The subtree-merging
method~\cite{Crut89a} and causal-state splitting reconstruction~\cite{Shal02a}
are perhaps the best known, but recently a new procedure based on Bayesian
inference has been developed~\cite{Stre14a}. Finally, a HMM may be proposed on
theoretical grounds, as done with the RGDF and SFSF HMMs in our second and
third examples. And, for the case when a DP is available, there is \eM\
spectral reconstruction theory~\cite{Varn02a,Varn07a,Varn13a,Varn13b}. We
anticipate that HMMs will become the standard representation for describing
layered structures.
The approach presented here should also be viewed in the larger context of our
recent research thrusts. While crystallography has historically struggled to
settle on a formalism to describe disordered structures, we propose that such a
framework has been identified, at least for layered materials. Based in
computational mechanics, {\emph {chaotic crystallography}}~\cite{Varn14a}
employs information theory as a key component to characterize disordered
materials. Although the use of information theory in crystallography has been
previously proposed by Mackay and coworkers~\cite{Mack86a,Mack02a,Cart12a},
chaotic crystallography realizes this goal. Additionally, using spectral
methods in the spirit of \S \ref{subsec:SpectralResults}, information- and
computation-theoretic measures are now directly calculable from
\eMs~\cite{Crut13a,Riec14a}. And importantly, a sequel will demonstrate how
spectral methods can give both a fast and efficient method for calculating the
DP of layered CPSs or analytical expressions thereof~\cite{Riec14c}.
\section{Acknowledgment}
The authors thank the Santa Fe Institute for its hospitality during visits.
JPC is an External Faculty member there. This material is based upon work
supported by, or in part by, the U. S. Army Research Laboratory and the U. S.
Army Research Office under contract number W911NF-13-1-0390.
|
\section{\bf Introduction}
Let $G$ be a group. An element $g \in G$ is called {\it reversible} in $G$ if it is conjugate to its inverse in $G$;
there exists $h \in G$ such that \begin{equation} g^{-1} = hgh^{-1}. \end{equation}
We say that $g$ is reversed by $h$. An element $h$ in $G$ is called an \emph{involution} if $h = h^{-1}$.
If the conjugating element $h$ can be chosen to be an involution then $g$
is called {\it strongly reversible}. Any strongly reversible element
can be expressed as a product of two involutions. So involutions are
strongly reversible and strongly reversible elements are reversible.
For $n\in \mathbb{N}^{*}$, define $$I_n(G) =
\{\tau_1 \tau_2\dots \tau_n \ \ : \ \ \forall \ 1 \leq i \leq n, \ \tau_i \ \hbox{is an involution in} \ G \};$$
$$R_n(G) = \{\ g_1 g_2\dots g_{n} \ \ : \ \ \forall \ 1 \leq i \leq n, \ g_i \
\hbox{is a reversible element in} \ G\}.$$ \noindent For each
integer $n \in \mathbb{N}^{*}$, it is clear that $I_n(G) \subseteq
I_{n+1}(G)$ and $R_n(G) \subseteq R_{n+1}(G)$. The set $I_1(G)$
(resp. $I_2(G)$) consists of the involutions (resp. the strongly
reversible elements) in $G$. Denote by
\textbullet \ $\mathbb{S}^{1} = \{z \in \mathbb{C} \ | \ \left|z\right| = 1\}$ the circle, it is a multiplicative group.
\textbullet \ $\textrm{Homeo}(\mathbb{S}^{1})$ (resp.
$\textrm{Homeo}^+(\mathbb{S}^1)$) the group of all homeomorphisms
(resp. orientation-preserving homeomorphisms) of $\mathbb{S}^1$.
\textbullet \ $\textrm{id}$ the identity map of $\mathbb{S}^1$.
\textbullet \ $s: \mathbb{S}^{1}\longrightarrow \mathbb{S}^{1}, \ z\mapsto \overline{z}$ the reflection.
\textbullet \ $\textrm{Fix}(f)$ the set of fixed points of $f$.
\noindent Two elements $f$ and $g$ of $\textrm{Homeo}(\mathbb{S}^1)$ are called \textit{conjugate }in $\textrm{Homeo}(\mathbb{S}^1)$ if
there exists $h \in
\textrm{Homeo}(\mathbb{S}^1)$ such that $g = hfh^{-1}$. It is well
known that for every homeomorphism $f: \mathbb{S}^{1}\rightarrow
\mathbb{S}^1$ there exists a unique
(up to a translation by an integer) homeomorphism $\tilde{f} : \mathbb{R} \rightarrow \mathbb{R}$ such that
\begin{equation}f(e^{i2 \pi x}) = e^{i2 \pi \tilde{f}(x)} \noindent \\ \textrm{ and} \ \tilde{f}(x+1) = \tilde{f}(x)+k,
\textrm{ for all } x \in \mathbb{R}\end{equation} where $k \in \{-1,
1\}$. Such a homeomorphism $\tilde{f}$ is called a \emph{lift} of
$f$. We call $f$ orientation-preserving if $k = 1$; resp.
orientation-reversing if $k = -1$; which is equivalent to the fact
that $\tilde{f}$ is increasing, resp. $\tilde{f}$ is decreasing.
\bigskip
\medskip
\begin{defn}\label{d1} A homeomorphism $f$ of $\mathbb{S}^{1}$ is said to be {\it piecewise
linear} ($\mathrm{PL}$) if it is
derivable except at finitely or countably many points $(c_{i})_{i\in \mathbb{N}}$ called
\textit{break points} of $f$ at which $f$ admits left and right
derivatives (denoted, respectively, by Df$_{-}$ and Df$_{+}$)
and such that the derivative $\mathrm{Df}: \ S^{1} \longrightarrow
\mathbb{R}^{*}$ is constant on each connected component of $\mathbb{S}^{1}\backslash \{c_{i}: \ i\in \mathbb{N}\}$.
\end{defn}
\medskip
Let $f$ be a {\it piecewise linear} ($\textrm{PL}$) homeomorphism
$f$ of $\mathbb{S}^{1}$. Define $\sigma_f(x) = {\textrm{Df}_-(x)
\over \textrm{Df}_+(x)}$ the $f$-jump in $x\in \mathbb{S}^1$. Denote by $B(f) = \{c_{i}: \ i\in \mathbb{N}\}$ the set of
\emph{break points} of $f$. So $B(f)= \{x\in \mathbb{S}^1: \sigma_f(x)\neq
1\}$.\\
\medskip
\noindent The homeomorphism $f$ is $\textrm{PL}$ (resp. $\textrm{PL}^+$, $\textrm{PL}^-$) if and only if $\tilde{f}$
is a piecewise linear (resp. piecewise linear increasing, piecewise linear decreasing) homeomorphism of the real
line $\mathbb{R}$.
Denote by
\textbullet \ \ $\textrm{PL}(\mathbb{S}^{1})$ the group of all
piecewise linear homeomorphisms of $\mathbb{S}^1$,
\textbullet \ $\textrm{PL}^+(\mathbb{S}^{1})$ the group of orientation-preserving elements of $\textrm{PL}(\mathbb{S}^{1})$,
\textbullet \ \ $\textrm{PL}^-(\mathbb{S}^{1})$ the set of
orientation-reversing elements of $\textrm{PL}(\mathbb{S}^{1})$.
If $f\in \textrm{Homeo}^+(\mathbb{S}^1)$, we denote by $\rho(f)$ its
rotation number.
\medskip
\emph{In the sequel we identify $\rho(f)$ to its lift in $[0, 1[$.}
\medskip
It is known (see for instance \cite{H}) that if an element $f \in
\textrm{PL}^+(\mathbb{S}^{1})$ is reversed by $h \in
\textrm{PL}^+(\mathbb{S}^{1})$, then by equality (1.1), $\rho(f)=0$
or $\frac{1}{2}$.
\vskip 3 mm
We refer the reader to the book \cite{aN10} and papers (
\cite{G}, \cite{Ma}) for a thorough account on groups of circle
homeomorphisms.
The object of this paper is to characterize reversible elements
(resp. strongly reversible elements) in the groups
$\textrm{PL}^+(\mathbb{S}^1)$ and $\textrm{PL}(\mathbb{S}^1)$. Our
main results are the following.
\medskip
\begin{thm}[Reversibility in $\mathrm{PL}^+(\mathbb{S}^{1})$] \label{t11}
Let $f\in \mathrm{PL}^+(\mathbb{S}^1)$. Then $f$ is reversible in
$\mathrm{PL}^+(\mathbb{S}^1)$ if and only if one of the following
holds:
\begin{enumerate}
\item[(1)] $\rho(f) = 0$, and $f$ is strongly reversible in $\mathrm{PL}^+(\mathbb{S}^1)$.
\item[(2)] $\rho(f) = \frac{1}{2}$, and $f$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of
$\mathrm{PL}^-(\mathbb{S}^{1})$.
\end{enumerate}
\end{thm}
\medskip
\begin{rem} \label{r1} If instead of the group $\mathrm{PL}^+(\mathbb{S}^{1})$ we take the group
$\mathrm{Homeo}^+(\mathbb{S}^1)$, the theorem \ref{t11} is false: In
\cite{GOS}, Gill et al. gave an example of a homeomorphism $f\in
\mathrm{Homeo}^+(\mathbb{S}^{1})$ with rotation number $\rho(f) =
\frac{1}{2}$ that is reversible in $\mathrm{Homeo}^+(\mathbb{S}^1)$
but not strongly reversible in $\mathrm{Homeo}(\mathbb{S}^1$).
\end{rem}
\smallskip
\begin{thm}[Reversibility in
$\mathrm{PL}(\mathbb{S}^{1})$]\label{t2} In
$\mathrm{PL}(\mathbb{S}^{1})$ reversibility and strong reversibility
are equivalent.
\end{thm}
\medskip
We denote by $n_f$ the smallest positive integer $n$ such that $f^n$
has fixed points and by $\Delta_f$ the signature of $f$ (see
definition in Section 2).
\medskip
\begin{thm}\label{t52}
\begin{itemize}
\item[(1)] Let $f \in \mathrm{PL}^+(\mathbb{S}^1)$. Then $f$ is strongly reversible in $\mathrm{PL}^+(\mathbb{S}^1)$ if and
only if one of the following holds.
\item[(i)] $f^2 = \mathrm{id}$.
\item[(ii)] $\mathrm{Fix}(f) \neq \emptyset$ and there exists $h\in \mathrm{PL}^+(\mathbb{S}^1)$ such that
$\rho(h) = \frac{1}{2}$ and $\Delta_f = -\Delta_f \circ h$.\\
\item[(2)] Let $f\in \mathrm{PL}^+(\mathbb{S}^1)$.
\subitem(i) If $\rho(f)\in \Bbb Q$ then $f$ is strongly reversible
in $\mathrm{PL}(\mathbb{S}^1)$ by an element of $\mathrm{PL}^-(\mathbb{S}^1)$ if and only if
there exists an involution $h \in \mathrm{PL}^-(\mathbb{S}^1)$ such that $\Delta_{f^{n_f}} = \Delta_{f^{n_f}} \circ h$.
\subitem(ii) If $\rho(f)\in \mathbb{R}\backslash \Bbb Q$ then $f$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^1)$ by
an element of $\mathrm{PL}^-(\mathbb{S}^1)$ if and only if $f$ is conjugate to the rotation $r_{\rho(f)}$ through a
homeomorphism $h$ such that $hrsh^{-1}\in \mathrm{PL}^-(\mathbb{S}^1)$, for some rotation $r$ of $\mathbb{S}^1$.
\end{itemize}
\end{thm}
\medskip
The next theorem is about composition of reversible (resp.
involution) maps.
\medskip
\begin{thm}\label{t91}
We have
\begin{itemize}
\item[(i)] $\mathrm{PL}^+(\mathbb{S}^{1}) = R_2(\mathrm{PL}^+(\mathbb{S}^{1})) =
I_3(\mathrm{PL}^+(\mathbb{S}^{1})\neq I_2(\mathrm{PL}^+(\mathbb{S}^{1}))$ and \\
$R_1(\mathrm{PL}^+(\mathbb{S}^{1}))\subset\neq I_2(\mathrm{PL}(\mathbb{S}^{1}))$.\\
\item[(ii)] $\mathrm{PL}(\mathbb{S}^1) = R_2(\mathrm{PL}(\mathbb{S}^{1})) = I_3(\mathrm{PL}(\mathbb{S}^{1}))
\neq I_2(\mathrm{PL}(\mathbb{S}^{1}))$ \\ and $R_1(\mathrm{PL}(\mathbb{S}^{1})) = I_2(\mathrm{PL}(\mathbb{S}^{1}))
\neq I_1(\mathrm{PL}(\mathbb{S}^{1}))$.
\end{itemize}
\end{thm}
\medskip
The structure of the paper is as follows. In Section 2 we give some
notations and preliminaries results that are needed for the rest of the paper. In
Section 3 we study reversibility in $\textrm{PL}^+(\mathbb{S}^{1})$
of elements $f$ of $\textrm{PL}^+(\mathbb{S}^{1})$ by proving
Theorem \ref{t11}. In Section 4, we study reversibility in
$\textrm{PL}(\mathbb{S}^{1})$ by proving Theorem \ref{t2}. Section 5
is devoted to the characterisation of strong reversibility in
$\textrm{PL}(\mathbb{S}^{1})$ of elements of
$\textrm{PL}^{+}(\mathbb{S}^{1})$. Finally, Section 6 is devoted to
the proof of Theorem \ref{t91}.
\bigskip
\section{\bf Notations and some results}
\medskip
Denote by
\textbullet \ $T$ the translation of $\mathbb{R}$ defined by $T(x) =
x+1$, and for each $a\in \mathbb{R}$ let $T_a$ be the translation
defined by $T_a(x) = x+a$. So $T = T_{1}$.
\textbullet \ \ $r_\pi: z\mapsto -z$ the rotation of
$\mathbb{S}^{1}$ by $\pi$.
\textbullet \ For points $x$ and $y$ in $\mathbb{S}^1$, we denote by
$(x,y)$ the open anticlockwise interval from $x$ to $y$, and by
$[x,y]$ the closure of $(x,y)$. We say that $x<y$ in a proper open
interval $I$ in $\mathbb{S}^1$ if $(x,y)\subset I$.
\textbullet \ For $f\in \textrm{Homeo}(\mathbb{S}^{1})$ we denote
by deg$(f)= \begin{cases}
1, & \textrm{if } f\in \textrm{Homeo}^{+}(\mathbb{S}^{1})\\
-1, & \textrm{if } f\in \textrm{Homeo}^{-}(\mathbb{S}^{1})
\end{cases}$ the degree of $f$.
\textbullet \ For $f\in \textrm{Homeo}^{+}(\mathbb{S}^{1})$ which
has a fixed point, then each point $x\in \mathbb{S}^{1}$ is either
in $\textrm{Fix}(f)$ or it lies in an open interval component $I$ in
$\mathbb{S}^{1}\backslash \textrm{Fix}(f)$. The \textit{signature} of $f$ (see \cite{GOS}) is
a map $\Delta_f: \mathbb{S}^{1}\longrightarrow \{-1,0,1\}$ given by
$$\Delta_f(x) = \left\{
\begin{array}{ll}
1, & \textrm{if } f(x)> x \\
0, & \textrm{if } f(x)=x \\
-1, & \textrm{if } f(x)<x.
\end{array}
\right.$$
\vskip 2 mm
We have the following lemma.
\bigskip
\begin{lem} \cite{GOS}\label{l100} Let $f\in\mathrm{Homeo}^{+}(\mathbb{S}^{1})$ with a fixed point
and $h \in\mathrm{Homeo}(\mathbb{S}^{1})$. Then\\
$(i)$ $\Delta_{hfh^{-1}} = \mathrm{deg}(h)(\Delta _{f} \circ h^{-1}).$\\
$(ii)$ $\Delta_{ f^{-1}} = -\Delta _{f}.$
\end{lem}
\medskip
\medskip
\begin{lem}\label{l28}
Let $f, \ h\in \mathrm{Homeo}(\mathbb{S}^1)$ be such that $f^{-1} = hfh^{-1}$ and let $n\in \mathbb{Z}$.
Then we have $f^{(-1)^{n}} = h^n f h^{-n}$.
\end{lem}
\begin{proof} The proof is down by induction, which is straighforward.
\end{proof}
\medskip
The following lemma shows that any reversible element of
$\textrm{PL}^+(\mathbb{R})$ must have a fixed point.
\begin{lem}\label{l22}
Let $f, \ h\in \mathrm{PL}^+(\mathbb{R})$ such that
$hfh^{-1} = f^{-1}$. Assume that $f$ is not the identity.
Then $\mathrm{Fix}(f) \neq \emptyset$ and $\mathrm{Fix}(h)
=\emptyset$.
\end{lem}
\medskip
\begin{proof}
In fact we show that $f$ has a fixed point in any subinterval $I$ of $\mathbb{R}$ such that $f(I) = I= h(I)$: Otherwise, either $f(x) > x$ for all $x \in I$ or $f(x) < x$ for all $x \in I$.
Assume that $f(x) > x$ for all $x\in I$. Then for all $x\in I$, $f^{-1}(x) = hf(h^{-1}(x)) > h(h^{-1}(x))= x$
since $f$ and $h$ are increasing, which means that $f(x) < x$; a contradiction. Thus $\textrm{Fix}(f_{|I})\neq\emptyset$ and
in particular, we have $\textrm{Fix}(f)\neq\emptyset$. As $f\neq \textrm{id}$, so it is not an involution and hence
$h\neq \textrm{id}$. We will show that $\textrm{Fix}(h)
=\emptyset$. Otherwise, suppose that $\mathrm{Fix}(h)
\neq\emptyset$. We will prove that in this case $f = \textrm{id}$, this leads to a contradiction.
Let $I = (a,b)$ be a connected component of $\mathbb{R} \backslash \textrm{Fix}(h)$.
Then, either $a$ or $b$ is a real number. Let us assume that $a \in \mathbb{R}$.
By ([2], Lemma 2.4) $f$ fixes each point of $\textrm{Fix}(h)$. Then $f(I) = I = h(I)$.
Now, by the first step, $f$ has a fixed point $s \in I$.
So for each integer $n \in \mathbb{Z}$, $h^n(s)\in I$ and by Lemma \ref{l28}, $f((h^n(s)) = h^n(s)$.
Since $\textrm{Fix}(h_{|I}) = \emptyset$, we can assume, by swapping $h$ and $h^{-1}$ if necessary, that
$h(s) < s < h^{-1}(s)$. Then the points $h^n(s)\in [a, s]$ for $n\in \mathbb{N}$, and accumulate at $a$.
So, $f$ has infinitely many fixed points in the interval $[a, s]$.
Since $f \in \textrm{PL}^+(\mathbb{R})$, there is an integer $N$ such that $f_{|[a, h^N(s)]}
= \textrm{id}$. Thus for any $y\in [a, s]$, one has $f(y)= fh^{-N}(x)$ where $x= h^{N}(y)\in [a,h^{N}(s)]$. By Lemma \ref{l28},
$f(y)=
h^{-N}f^{(-1)^{N}}(x)=h^{-N}(x)=y$. Therefore $f_{|[a, s]}
= \textrm{id}$. Similarily, by considering the points $h^{-n}(s) \in [s, b]$, $n\in \mathbb{N}$, we get as above $f_{|[s, b]}
= \textrm{id}$. Therefore $f = \textrm{id}$ on $\mathbb{R} \backslash \textrm{Fix}(h)$. As
$f = \textrm{id}$ on $\textrm{Fix}(h)$, thus $f = \textrm{id}$ on $\mathbb{R}$.
\end{proof}
\medskip
\begin{prop}\label{p24}
\begin{itemize}
\item[(1)] If $\tau$ is an involution in $\mathrm{PL}^+(\mathbb{S}^{1})$ which is
not the identity then $\tau$ is conjugate in
$\mathrm{PL}^+(\mathbb{S}^{1})$ to the rotation $r_{\pi}$.
\item[(2)] If $\tau$ is an involution in $\mathrm{PL}^-(\mathbb{S}^1)$, then it is conjugate in $\mathrm{PL}(\mathbb{S}^1)$ to
the reflection $s$.
\end{itemize}
\end{prop}
\medskip
\begin{proof} Let $a$ be the point on $\mathbb{S}^1$ with coordinates $(1, 0)$ and let $b$ be the point with
coordinates $(-1, 0)$. (1) Let $x \in S^1$. Since $\tau \in
\textrm{PL}^+(\mathbb{S}^{1})$ and $\tau^2 = \textrm{id}$, we have
$\tau([x, \tau(x)]) = [\tau(x), x]$. Let $v : [a, b] \longrightarrow
[x, \tau(x)]$ be a piecewise linear homeomorphism, and let $\psi$ be
the map of $S^1$ defined by $$ \psi(x) = \left\{
\begin{array}{ll}
v(x), & \hbox{if} \ x \in [a, b]; \\
\cr \cr \tau v r_\pi(x), & \hbox{if } \ x \in [b, a].
\end{array}
\right.$$ \noindent Then $\psi$ is a well defined piecewise
linear homeomorphism of $S^1$, and it satisfies $\psi^{-1} \tau \psi
= r_\pi$. We conclude that $\tau = \psi r_\pi \psi^{-1}$ is
conjugate in $\textrm{PL}(\mathbb{S}^{1})$ to $r_\pi$.
\smallskip
\noindent (2) Let $\{c, d\} = \textrm{Fix}(\tau)$. We have $\tau([c, d]) = [d, c]$.
Let $u: [a, b] \longrightarrow [c, d]$ be a piecewise
linear homeomorphism and let $\varphi$ be the map of $\mathbb{S}^1$
defined by
$$ \varphi(x) = \begin{cases}
u(x), & \textrm{ if }\ x \in [a, b] \\
\cr \cr \tau u s(x), & \textrm{ if } \ x \in [b, a]
\end{cases}$$
\noindent Then $\varphi\in \textrm{PL}(\mathbb{S}^1)$. If $x \in [a, b]$ then
$\varphi (s(x)) = \tau us(s(x)) = \tau u(x)= \tau \varphi(x)$. If $x \in [b, a]$ then
$\varphi (s(x)) = u(s(x))= \tau \varphi(x)$. Therefore $\tau = \varphi
s\varphi^{-1}$.\end{proof}
\vskip 3 mm
\vskip 3 mm
\begin{thm}[\cite{Her}]\label{t26} Let $h\in \mathrm{PL}^+(\mathbb{S}^{1})$ with rotation number $\rho(h)$ irrational.
Then $h$ is conjugate in $\mathrm{Homeo}^+(\mathbb{S}^{1})$ to the rotation $r_{\rho(h)}$.
\end{thm}
\vskip 3 mm
\begin{lem}\label{l27}
Let $f \in \mathrm{Homeo}^+(\mathbb{S}^{1})$ and $g \in \mathrm{Homeo}(\mathbb{S}^{1})$ such that $fg = gf$ and
$\mathrm{Fix}(f) \neq \emptyset \neq \mathrm{Fix}(g)$.
Then $\mathrm{Fix}(f) \cap \mathrm{Fix}(g)\neq \emptyset$.
\end{lem}
\begin{proof} Let $x \in \textrm{Fix}(g)$ and let $\omega_{f}(x)$ be the $\omega$-limit set of $x$ under $f$. It is well known that
$\omega_{f}(x)$ is a periodic orbit (see \cite{H}). Since
$\textrm{Fix}(f) \neq \emptyset$, $\rho(f) = 0$ and any periodic
orbit under $f$ is a fixed point of $f$. It follows that $w_{f}(x) =
\{a\}$; where $a \in \textrm{Fix}(f)$. As $\mathbb{S}^{1}$ is
compact, so $(f^n(x))_{n}$ converges
to $a$. Now, we have $g(f^n(x)) = f^n(g(x)) = f^n(x)$ since $g$ commutes with $f$. So, $f^n(x)$ converges to $g(a) = a$.
Hence $a\in \textrm{Fix}(f)\cap \textrm{Fix}(g)$. \\
\end{proof}
\vskip 3 mm
\medskip
\section{\bf Reversibility in $\mathrm{PL}^+(\mathbb{S}^{1})$}
\subsection{\bf Reversibility in $\textrm{PL}^{+}(\mathbb{S}^1)$ of
elements $f\in \textrm{PL}^+(\mathbb{S}^{1})$ with $\rho(f)=0$.} The
aim of this subsection is to prove the following proposition.
\medskip
\begin{prop}\label{p21}
Let $f\in \mathrm{PL}^+(\mathbb{S}^{1})$ such that $\rho(f) = 0$. Then $f$ has a lift $\tilde{f}$ in $\mathrm{PL}^+(\mathbb{R})$,
which is strongly reversible in $\mathrm{PL}(\mathbb{R})$ by an element of
$\mathrm{PL}^-(\mathbb{R})$.
\end{prop}
\medskip
\begin{proof} Since $\rho(f) = 0$, $f$ has a fixed point $x_0 = e^{it} \in \mathbb{S}^1$. We can assume that $1 \in \textrm{Fix}(f)$
(by taking $r_{-t} f r_t$ instead of $f$, where $r_t(z)= e^{it}z$ is the rotation by angle $t$). Let $\tilde{f}$ be the lift for $f$
such that $\tilde{f}(0) = 0$. Then for all $n \in \mathbb{Z}$, $(T\tilde{f})^n(0) = n$. Let $\alpha_0 : [0, 1] \longrightarrow [0,1]$
be an orientation preserving piecewise linear homeomorphism ($\alpha_0 \in
\textrm{PL}^+([0,1])$) and for each $n \in \mathbb{Z}$, let $\alpha_{n}: [n, n+1] \rightarrow [n, n+1]$ be the homeomorphism defined as:
$\alpha_n = T^n \alpha_0 (T\tilde{f})^{-n}$. Define $\alpha:
\mathbb{R}\longrightarrow \mathbb{R}$ as $\alpha_{\mid \ [n, n+1]} =
\alpha_n$, for all $n\in \mathbb{Z}$. Then $\alpha \in
\textrm{PL}^+(\mathbb{R})$ and $T\tilde{f} = \alpha^{-1} T \alpha$.
Moreover, $(T\tilde{f})$ is a lift of $f$. On the other hand, the
translation $T$ satisfies $T^{-1} = iTi$; where $i$ is the
involution of $\mathbb{R}$ defined by $x \longmapsto 1-x$. Then
$(T\tilde{f})^{-1} = \tau (T\tilde{f}) \tau$; where $\tau =
\alpha^{-1}i \alpha$ is an involution in
$\textrm{PL}^-(\mathbb{R})$. The proof is complete.
\end{proof}
\vskip 3 mm
\begin{prop}\label{p23}
Let $f\in \mathrm{PL}^+(\mathbb{S}^{1})$ such that $\rho(f) = 0$. If f is reversed by $h \in \mathrm{PL}^+(\mathbb{S}^{1})$, then there exists a lift
$\tilde{f} \in \mathrm{PL}^+(\mathbb{R})$ of $f$ which is reversed in $\mathrm{PL}^+(\mathbb{R})$ by any lift
$\tilde{h}$ of $h$.
\end{prop}
\medskip
\begin{proof}
Since $\rho(f) = 0$, $\textrm{Fix}(f) \neq \emptyset$ and there is a
lift $\tilde{f} \in \textrm{PL}^+(\mathbb{R})$ of $f$ such that
$\textrm{Fix}(\tilde{f}) \neq\emptyset$. We have \begin{equation}
f^{-1} = h f h^{-1} \end{equation} \noindent Let $\tilde{h} \in
\textrm{PL}^+(\mathbb{R})$ be a lift of $h$. Then
$\tilde{h}\tilde{f}\tilde{h}^{-1}$ is a lift of $h f h^{-1}$. From
equality (3.1), it follows that $\tilde{f}^{-1} =
\tilde{h}\tilde{f}\tilde{h}^{-1} T_{-q}$ for some integer $q \in
\mathbb{Z}$. As $T \tilde{f} = \tilde{f} T$ and $T \tilde{h} =
\tilde{h} T$, then $\tilde{f}^{-1} = T_{-q}
\tilde{h}\tilde{f}\tilde{h}^{-1} = \tilde{h} (T_{-q} \tilde{f})
\tilde{h}^{-1}$. Therefore, since $\tilde{f}$ has a fixed point,
$(T_{-q} \tilde{f})$ has also a fixed point $a\in \mathbb{R}$. Then
$(T_{-q} \tilde{f})(a) = a$, equivalently to $\tilde{f}(a) =
T_q(a)$.
So, $$\tilde{f}^{n}(a) = {(T_q)}^n(a) = a+nq, \ \ \ \ \ \ \ \forall n \in \mathbb{Z}.$$
\noindent Now, we show that $q = 0$. Suppose that $q \neq 0$ (say $q
> 0$). Then we have $\mathbb{R} = \underset{n \in \mathbb{Z}}\cup\ [a+nq,
a+(n+1)q]$. Let $\alpha_0: [a, a+q] \longrightarrow [0,1]$ be a
piecewise linear orientation preserving homeomorphism such that
$\alpha_0(a) = 0$ and $\alpha_0(a+q) = 1$. For $n \in \mathbb{Z}$,
let $\alpha_n: [a+nq, a+(n+1)q] \rightarrow [n, n+1]$ be a
homeomorphism defined as $\alpha_n= T^n \alpha_0 \tilde{f}^{-n}$.
Define $\alpha: \mathbb{R} \longrightarrow \mathbb{R}$ as
$\alpha_{\mid [a+nq, a+(n+1)q]} = \alpha_n$ for all $n \in \mathbb{Z}$.
Then $\alpha\in \textrm{PL}^+(\mathbb{R})$ and we see that
$\tilde{f} = \alpha^{-1}T \alpha$, which is impossible since
$\textrm{Fix}(\tilde{f}) \neq\emptyset$. We conclude that $q=0$ and
so $\tilde{f}^{-1} = \tilde{h}\tilde{f}\tilde{h}^{-1}$.
\end{proof}
\vskip 3 mm
\begin{prop}\label{t24}
Let $f\in \mathrm{PL}^+(\mathbb{S}^{1})$ which is not the identity and such that $\rho(f)=0$. Then $f$ is reversible in $\mathrm{PL}^+(\mathbb{S}^{1})$ if and only
if some lift $\tilde{f}$ of $f$ is reversed by a homeomorphism $\tilde{h}\in
\mathrm{PL}^+(\mathbb{R})$ satisfying $\tilde{h}T = T\tilde{h}$.
\end{prop}
\begin{proof} The if part follows from Proposition \ref{p23}. The only if part: let $f \in \textrm{PL}^+(\mathbb{S}^{1})$ with $\rho(f) = 0$
and let $\tilde{f}$ be a lift for $f$ such that $\tilde{f}^{-1} =
\tilde{h}\tilde{f}\tilde{h}^{-1}$; where $\tilde{h}\in
\textrm{PL}^+(\mathbb{R})$ satisfying $\tilde{h}T = T\tilde{h}$.
Then $\tilde{h}$ is a lift for the homeomorphism $h:
\mathbb{S}^{1}\longrightarrow \mathbb{S}^1$ defined by \ $h(e^{i2
\pi x}) = e^{i2 \pi\tilde{h}(x)}, \ \ \forall \ x \in \mathbb{R}$.
Then $h \in \textrm{PL}^+(\mathbb{S}^1)$ and for all $x \in
\mathbb{R}$, we have $f^{-1} = hfh^{-1}$.
\end{proof}
\vskip 3 mm
\begin{prop}\label{p25} Let $f\in \mathrm{PL}^+(\mathbb{S}^{1})$ such that $\rho(f)=0$.
If $f$ is a reversible element in $\mathrm{PL}^+(\mathbb{S}^{1})$ then some lift $\tilde{f}$ of $f$ is
conjugate to a homeomorphism $\tilde{g}\in \mathrm{PL}^+(\mathbb{R})$ which is the lift of a homeomorphism $g\in \mathrm{PL}^+(\mathbb{S}^{1})$ that is reversible by the rotation $r_\pi$.
\end{prop}
\begin{proof} From Proposition \ref{p23}, there exists $\tilde{h} \in \textrm{PL}^+(\mathbb{R})$ such that
$\tilde{f}^{-1} = \tilde{h}\tilde{f}\tilde{h}^{-1}$. Then by Lemma
\ref{l22}, $\textrm{Fix}(\tilde{h}) =\emptyset$. So $\tilde{h}$ is
conjugate in $\textrm{PL}^+(\mathbb{R})$ to either the translation
$T: x \longmapsto x+1$ or $T_{-1}: x \longmapsto x-1$, say $T$ for
example (see \cite{GS}). It follows that $\tilde{f}$ is conjugate to
a homeomorphism $\tilde{g} \in
\textrm{PL}^+(\mathbb{R})$ satisfying $\tilde{g}^{-1} = T\tilde{g}T^{-1}$. Therefore $T^2\tilde{g} = \tilde{g}T^2$ and we can define
a homeomorphism $g \in \textrm{PL}^+(\mathbb{S}^{1})$ by \ $g(e^{i \pi x}) = e^{i \pi \tilde{g}(x)}, \ \forall x \in \mathbb{R}$.
We have $g^{-1}(e^{i \pi x}) = e^{i \pi \tilde{g}^{-1}(x)} = e^{i \pi T\tilde{g}T^{-1}(x)} = r_\pi g r_{-\pi}(e^{i \pi x}), \
\forall \ x \in \mathbb{R}$, which means that $g^{-1} = r_\pi g r_\pi$.
\end{proof}
\vskip 3 mm
\begin{proof}[Proof of the part (1) of Theorem \ref{t11}]
\noindent Let $f$ be a reversible homeomorphism in $\textrm{PL}^+(\mathbb{S}^{1})$ such that $\rho(f) =
0$. If $f$ is the identity, assertion (1) is clear. So suppose that $f$ is not the identity.
Then there exists $h \in PL^+(\mathbb{S}^{1})$ such that $f^{-1} =
hfh^{-1}$. Let us prove that $\rho(h) \in \mathbb{Q}$. Otherwise,
$h$ is conjugate to an irrational rotation $r$ by Theorem \ref{t26};
there exists $\alpha \in \textrm{Homeo}^+(\mathbb{S}^{1})$ such that
$h = \alpha r\alpha^{-1}$. It follows that $f^{-1} = \alpha r
\alpha^{-1} f \alpha r^{-1} \alpha^{-1}$, equivalently $g^{-1} = r g
r^{-1}$; where $g = \alpha^{-1} f \alpha$. Then by Lemma \ref{l28}
for each integer $n \in \mathbb{Z}$,
\begin{equation} r^{2n}g = g r^{2n}. \end{equation}
\noindent Since $\rho(f) = 0$, $\textrm{Fix}(g) \neq \emptyset$. So let $a \in \textrm{Fix}(g)$. By equality (3.2), for
each $n \in \mathbb{Z}$, $r^{2n}(a) \in \textrm{Fix}(g)$. As the
rotation $r$ is irrational, we have $\mathbb{S}^{1}=
\overline{\{r^{2n}(a): \ n \in \mathbb{Z}\}}$. Therefore
$\mathbb{S}^{1}\subset \textrm{Fix}(g)$ and $g = \textrm{id}$. So,
$f = \textrm{id}$, a contradiction.
\medskip
Let then $\rho(h) = \frac{p}{q}$; where $p$ and $q$ are coprime
positive integers. Then $h^q$ has a fixed point. Let us prove that
$q$ is even.
Otherwise, by Lemma \ref{l28}, $f^{-1} = h^qfh^{-q}$. Then as in the
proof of Proposition \ref{p23}, there is a lift $\tilde{h^{q}}$ of
$h^q$ such that $\textrm{Fix}(\tilde{h^{q}})\neq \emptyset$ and a
lift $\tilde{f}$ for $f$ such that $\tilde{f}^{-1} =
\tilde{h^{q}}\tilde{f}\tilde{h^{q}}^{-1}$; this contradicts Lemma
\ref{l22}. We conclude that $q$ cannot be odd.
Now, one can take $q = 2i$ for some integer $i$. Since $\rho(h^{2i})
= p$, $\textrm{Fix}(h^{2i}) \neq\emptyset$. Since $f h^{2i}=
h^{2i}f$ and $\textrm{Fix}(f) \neq\emptyset$, so
by Lemma \ref{l27}, there exists $a \in \mathbb{S}^{1}$ such that $f(a) = a = h^{2i}(a)$. It follows that $f(h^j(a)) = h^j(a)$
for each integer $j$ since $f^{-1} = hfh^{-1}$. Let $\mu: \mathbb{S}^{1}\longrightarrow \mathbb{S}^1$ be the homeomorphism defined by
$$\mu(z) = \begin{cases}
h^{i}(z), & \textrm{if } z \in [a, h^i(a)]\\
h^{-i}(z), & \textrm{if } z \in [h^i(a), a]
\end{cases}$$
\vskip 3 mm
\noindent It is clear that $\mu$ is an involution in $\textrm{PL}^+(\mathbb{S}^{1})$. We have
\vskip 2 mm
\begin{equation}\mu f \mu(z) = \begin{cases}
h^{-i}f h^i(z), & \textrm{if} \ z \in [a, h^i(a)]\\
h^i f h^{-i}(z), & \textrm{if} \ z \in [h^i(a), a]
\end{cases}
\end{equation}
\noindent We prove that the integer $i$ is odd: Otherwise, suppose
that $i$ is even. Then $f h^i = h^i f$, and by (3.3) we obtain
that $\mu f \mu = f$. Moreover, as $f^{-1} = h f h^{-1}$, then we
have $f^{-1} = h \mu f \mu h^{-1}$.
Let $a = e^{i2 \pi y}$; where $y\in [0, 1[$, and let $\tilde{h}$ be a lift of $h$.
As $\rho(h) = \frac{p}{2i}$, then $n = 2i$ is the smallest integer
such that $h^n$ has a fixed point. Therefore, $\tilde{h}^i(y)\neq
y$. Assume that $y < \tilde{h}^i(y)$. Since $\mu(a) = h^i(a)$
and $\tilde{h}^i$ is a lift for $h^i$, there is a lift $\tilde{\mu}$ for $\mu$ such that $\tilde{\mu}(y) = \tilde{h}^i(y)$.
By Proposition \ref{p23}, there exists a lift $\tilde{f}$ of $f$
such that $\tilde{f}(y) = y$ and $\tilde{f}^{-1} = \tilde{h}
\tilde{\mu} \tilde{f} \tilde{\mu} \tilde{h}^{-1} = \tilde{h}
\tilde{f} \tilde{h}^{-1}$. It follows that $\tilde{\mu} \tilde{f}
\tilde{\mu} = \tilde{f}$. Therefore
${\tilde\mu}\tilde{f}\tilde{\mu}(y) =
{\tilde\mu}\tilde{f}(\tilde{h}^i(y)) = \tilde{f}(y) = y$. Since
$\tilde{\mu} \tilde{f}$ is an increasing homeomorphism, the
inequality $y< \tilde{h}^i(y)$ implies that $\tilde{\mu}
\tilde{f}(y) \leq y$; equivalently $\tilde{h}^i(y) \leq y$, which is
a contradiction since $y < \tilde{h}^i(y)$. Now the equality $f^{-1}
= hfh^{-1}$ implies that $f^{-1} = h^ifh^{-i}$, equivalently $f^{-1}
= h^{-i}fh^i$. From (3.3), we deduce that $\mu f \mu = f^{-1}$.
Therefore $f$ is strongly reversible by $\mu$ in
$\textrm{PL}^+(\mathbb{S}^{1})$.
This completes our proof.
\end{proof}
\vskip 3 mm
\begin{lem}\label{l299}
Let $I = (a, b)$ be an open interval in $\mathbb{R}$ or in $\mathbb{S}^{1}$. Then the following statements hold.
\begin{itemize}
\item[(1)] Let $f\in \mathrm{PL}^-(I)$. Then $f$ is reversible in
$\mathrm{PL}(I)$ if and only if $f$ is an involution.
\item[(2)] Let $f\in \mathrm{PL}^+(I)$. Then $f$ is reversible in $\mathrm{PL}(I)$ by an element of
$\mathrm{PL}^-(I)$ if and only if $f$ is strongly
reversible in $\mathrm{PL}(I)$ by an element of $\mathrm{PL}^-(I)$.
\end{itemize}
\end{lem}
\begin{proof} \textbullet \ First, assume that $I$ is an open interval in the real line $\mathbb{R}$.
\noindent Assertion (1). If $f$ is an involution then it is reversible in $\textrm{PL}(I)$ by the identity map.
Conversely, let $f\in \textrm{PL}^-(I)$, $h \in \textrm{PL}(I)$ such that $hfh^{-1} = f^{-1}$. By replacing $h$ with $hf$ if necessary,
we can assume that $h \in \textrm{PL}^+(I)$. One can extend $f$ and $h$ on $\mathbb{R}$ as follows:
$$\hat{f}(x) = \begin{cases}
f(x), & \textrm{if } x \in (a, b)\\
a+b-x, & \textrm{if } x \in \mathbb{R} \setminus (a, b)
\end{cases} \ \ ; \ \ \hat{h}(x) = \begin{cases}
h(x), & \textrm{if } x \in (a, b)\\
x, & \textrm{if } x \in \mathbb{R} \setminus (a, b)
\end{cases}$$
\vskip 3 mm
\noindent Then, clearly $\hat{f} \in \textrm{PL}^-(\mathbb{R})$, $\hat{h} \in \textrm{PL}^+(\mathbb{R})$ and
$\hat{h}\hat{f}\hat{h}^{-1} = \hat{f}^{-1}$. So, by ([2], Proposition 2.5), $\hat{f}^2 = \textrm{id}$ on $\mathbb{R}$.
It follows that $f^2 = \textrm{id}$ on $I$.
\vskip 3 mm
\noindent Assertion (2). Let $f\in \textrm{PL}^+(I)$ and $h\in \textrm{PL}^-(I)$ such that $hfh^{-1} = f^{-1}$.
Let $\{p\} = \hbox{Fix}(h)$, $p\in I$. We define the involution $\tau \in \textrm{PL}^-(I)$ as follows:
$$ \tau(x) = \begin{cases}
h^{-1}(x), & \textrm{if } x \geq p\\
h(x), & \textrm{if } x \leq p
\end{cases}$$
\vskip 3 mm
\noindent If $f(p) = p$, then clearly $\tau f\tau = f^{-1}$. If $f(p)\neq p$, let $(c, d)$ be the connected component of
$I\backslash \hbox{Fix}(f)$ containing $p$ (when $\hbox{Fix}(f) = \emptyset$, \ $(c, d)= I$).
By the fact that $h f h^{-1} = f^{-1}$, we have $h(\hbox{Fix}(f)) = \hbox{Fix}(f)$. Therefore $h((c, d)) = (c, d) = f((c, d))$
(since $h(p) = p$). Let $$ \widehat{f}(x) = \begin{cases}
f(x), & \textrm{if } x\in (c,d)\\
x, & \textrm{if } x\in \mathbb{R}\backslash (c,d)
\end{cases}$$ and $$ \widehat{h}(x) = \begin{cases}
h(x), & \textrm{if } x\in (c,d)\\
x, & \textrm{if } x\in \mathbb{R}\backslash (c,d)
\end{cases}.$$ Then $\widehat{h}$ is one bump function and satisfies $\hat{h}\hat{f}\hat{h}^{-1} = \hat{f}^{-1}$.
From ([2], Lemma 4.2), $h^2 = \textrm{id}$ on $(c, d)$ and hence $h(x) = h^{-1}(x)$, for each $x\in (c, d)$. We conclude that
the equality $\tau f \tau = f^{-1}$ is satisfied.
\vskip 3 mm
\textbullet \ Now, let us assume that $I = (a, b)$ is an open interval in the circle $\mathbb{S}^{1}$.
Then, there exists an open interval $\widehat{I} = (t_{1}, t_{2})$ in $\mathbb{R}$ such that the map
$\varphi: (t_{1}, t_{2})\longrightarrow (a, b)$ given by $\varphi(t) = e^{2i\pi t}$, is a homeomorphism. For
$f \in \textrm{PL}(I)$, set $g = \varphi^{-1}f\varphi$.
We have $g(t) = \varphi^{-1}f\varphi (t) = \varphi^{-1}f(e^{i 2 \pi t}) =
\varphi^{-1}(e^{i 2\pi\tilde{f}(t)}) = \tilde{f}(t)$, for each $t\in (t_{1}, t_{2})$. Hence $g\in \textrm{PL}(\widehat{I})$.
\vskip 3 mm
Assertion (1). If $f\in \textrm{PL}^-(I)$ then $g\in \textrm{PL}^{-}(\widehat{I})$.
If $f$ is reversed by $h\in \textrm{PL}(I)$ i.e. $hfh^{-1} = f^{-1}$
then $g$ is reversed by $k= \varphi^{-1} h \varphi\in \textrm{PL}(\widehat{I})$. Hence by above, $g$ is an involution and so is
$f$.
\vskip 3 mm
Assertion (2). If $f\in \textrm{PL}^+(I)$ and $h \in \textrm{PL}^-(I)$ such that $f^{-1} = hfh^{-1}$ then
$g$ is reversed in $\textrm{PL}(\widehat{I})$ by an element of $k = \varphi^{-1}h \varphi\in \textrm{PL}^-(\widehat{I})$. By the above,
$g$ is strongly reversible in $\textrm{PL}(\widehat{I})$ by an element of $\textrm{PL}^{-}(\widehat{I})$. Hence there exists an
involution $\tau\in \textrm{PL}^-(\widehat{I})$ such that $\tau g \tau = g^{-1}$. So $(\varphi \tau \varphi^{-1})f
(\varphi \tau \varphi^{-1}) = f^{-1}$. As $(\varphi \tau \varphi^{-1}) \in \textrm{PL}^-(I)$ and is an involution,
so $f$ is strongly reversible in $\textrm{PL}(I)$ by an element of $\textrm{PL}^-(I)$.
\end{proof}
\medskip
\begin{lem}\label{l29}
Let $f \in \mathrm{PL}^+(\mathbb{S}^{1})$ such that $\rho(f) = 0$. If $f$ is strongly reversible in $\mathrm{PL}^+(\mathbb{S}^{1})$,
then it is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^-(\mathbb{S}^{1})$.
\end{lem}
\begin{proof} If $f$ is strongly reversible in $\textrm{PL}^+(\mathbb{S}^{1})$, then by Proposition \ref{p25},
there exist $\alpha \in \textrm{PL}(\mathbb{S}^{1})$ and $g \in
\textrm{PL}^+(\mathbb{S}^{1})$ such that $f = \alpha g \alpha^{-1}$
and $g^{-1} = r_\pi g r_\pi$. By Proposition \ref{p23}, there exists
a lift $\widetilde{g} \in \textrm{PL}^+(\mathbb{R})$ of $g$ such
that
$$\widetilde{g}^{-1} = T_{\frac{1}{2}}\widetilde{g}T_{\frac{1}{2}}; \ \ \ \ \ (\ast)$$
\noindent Recall that $T_{\frac{1}{2}}$ defined by
$T_{\frac{1}{2}}(t) = t + \frac{1}{2}$ for all $t\in \mathbb{R}$, is
a lift of $r_\pi$.
Let $\widetilde{s}$ be the involution in $\textrm{PL}^-(\mathbb{R})$ defined
by $\widetilde{s}(t) = -t$. Then $\widetilde{s}T_{\frac{1}{2}}$ is
an involution in $\textrm{PL}^-(\mathbb{R})$, and by equality
$(\ast)$, we have $\widetilde{s} {\widetilde{g}}^{-1} =
(\widetilde{s}T_{\frac{1}{2}})(\widetilde{g}\widetilde{s})(\widetilde{s}T_{\frac{1}{2}})$.
Therefore $(\widetilde{g}\widetilde{s})\in
\textrm{PL}^-(\mathbb{R})$ which is strongly reversible in $\textrm{PL}(\mathbb{R})$ by an element of
$\textrm{PL}^{-}(\mathbb{R})$. So by Lemma \ref{l299},
$(\widetilde{g}\widetilde{s})^2 = \textrm{id}$; equivalently
${\widetilde{g}}^{-1} = \widetilde{s}\widetilde{g}\widetilde{s}$.
The involution $\widetilde{s}$ satisfies $\widetilde{s}(t+1) =
\widetilde{s}(t) - 1$ for all $t\in \mathbb{R}$. Therefore
$\widetilde{s}$ is a lift for the involution $\sigma$ of
$\textrm{PL}^-(\mathbb{S}^{1})$ defined by
$$\sigma(e^{2i\pi t}) = e^{2i\pi \widetilde{s}(t)}; \ \ \ \ \forall \ t \in
\mathbb{R}.$$ \noindent It follows that for each $t \in \mathbb{R}$,
$g^{-1}(e^{2i\pi t}) = e^{2i\pi
\widetilde{s}\widetilde{g}\widetilde{s}(t)} = \sigma g \sigma(e^{2i
\pi t})$. So $g^{-1} = \sigma g \sigma$. We deduce that $f^{-1} =
\tau f\tau$; where $\tau = \alpha\sigma \alpha^{-1}$ is an
involution in $\textrm{PL}^-(\mathbb{S}^1)$.
\end{proof}
\medskip
\medskip
\subsection{\bf Proof of the part (2) of Theorem \ref{t11}}
We need the following lemma.
\begin{lem}\label{l211}
Let $f\in \mathrm{PL}^+(\mathbb{S}^{1})$ such that $\mathrm{Fix}(f^n) \neq \emptyset$ for some
integer $n\in \mathbb{N}^*$.
Then the following are equivalent.
\begin{itemize}
\item[(1)] $f^n$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^-(\mathbb{S}^{1})$.
\item[(2)] $f$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of
$\mathrm{PL}^-(\mathbb{S}^{1})$.
\end{itemize}
\end{lem}
\medskip
\begin{proof} If $f$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an involution $\mu \in \textrm{PL}^-(\mathbb{S}^{1})$ then $f^{-1} = \mu f \mu$,
which implies that $f^{-n} = \mu f^n \mu$. Conversely, assume that, for some integer $n\in \mathbb{N}^*$,
$f^n$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an involution $\tau \in
\textrm{PL}^-(\mathbb{S}^{1})$, that is $f^{-n} = \tau f^n \tau$. We will show that $f$ is strongly
reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$. We have
\begin{equation} (f^{n-1} \tau) f^n (\tau f^{1-n} ) = f^{-n},
\end{equation}
\begin{equation} (f \tau) f^n (\tau f^{-1}) = f^{-n}. \end{equation}
The equality (3.4) implies that $(f^{n-1} \tau)^2 f^n = f^n
(f^{n-1} \tau)^2$. Therefore from Lemma \ref{l27}, there exists $a
\in \textrm{Fix}(f^n) \cap \textrm{Fix}((f^{n-1} \tau)^2)$. Thus
$f^n(a) = a$, $f^n (\tau(a)) = \tau(a)$ and $f^{n-1} (\tau(a)) =
\tau f^{1-n}(a) = \tau f(a)$. In particular, we have $f
\tau\left((\tau(a), f(a))\right) = (\tau(a), f(a))$. Then the
restriction of $f^n/(\tau(a), f(a))$ is an element of
$\textrm{PL}^+((\tau(a), f(a)))$, which is reversed by $(f \tau)_{|
(\tau(a), f(a))} \in \textrm{PL}^-((\tau(a), f(a)))$ by equality
(3.5). Then by Lemma \ref{l299}, $f^n_{| (\tau(a),
f(a))}$ is strongly reversible in $\textrm{PL}((\tau(a), f(a)))$ by an involution $\sigma \in
\textrm{PL}^-((\tau(a), f(a)))$; that is,
$f^{-n}_{| (\tau(a),
f(a))} = \sigma f^n_{| (\tau(a), f(a))} \sigma$. The point $f(a)$ is
either in $(a, \tau(a))$ or in $(\tau(a), a)$. We can assume that
$f(a) \in (\tau(a), a)$. Since $f$ is orientation-preserving, we can
easily see that
$$\mathbb{S}^{1} = \bigcup_{p=1}^n \ [f^p(a), f^p \tau(a)] \cup \bigcup_{p=2}^{n+1}\ [f^p \tau(a), f^{p+1}(a)].$$
\noindent Let $\mu: \mathbb{S}^{1}\longrightarrow \mathbb{S}^{1}$ be
the map of $\mathbb{S}^{1}$ defined by
$$\mu (x) = \begin{cases}
f^{n-p+1}\tau f^{n-p}(x), & \textrm{if } x \in [f^p(a), f^p\tau(a)], \ \ \forall \ 1 \leq p \leq n; \\
f^{n-p} \sigma f^{n-p}(x), & \textrm{if } x \in (f^p\tau(a), f^{p+1}(a)), \ \ \forall \ 2 \leq p \leq n+1
\end{cases}$$
\noindent The map $\mu$ is a well defined homeomorphism of $\mathbb{S}^1$. Moreover
$\mu\in \textrm{PL}^-(\mathbb{S}^{1})$ that satisfies
$\mu^2 = \textrm{id}$ and $\mu f \mu = f^{-1}$. This completes the proof.
\end{proof}
\vskip 3 mm
\noindent \textit{Proof of the part (2) of Theorem \ref{t11} } Let
$f\in \textrm{PL}^+(\mathbb{S}^{1})$ be reversible in
$\textrm{PL}^+(\mathbb{S}^{1})$. Then the rotation number $\rho(f)$
is equal to either $0$ or $\frac{1}{2}$. The first case $\rho(f) =
0$ corresponds to the first part of Theorem \ref{t11}. In the second
case, $\rho(f) = \frac{1}{2}$ we have $\rho(f^2) = 0 (\textrm{mod }
1)$. Let us prove that $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$. By hypothesis, there exists a
homeomorphism $h \in \textrm{PL}^+(\mathbb{S}^{1})$ such that
$f^{-1} = hfh^{-1}$ and so $f^{-2} = h f^2 h^{-1}$. Then by the
proof of the part (1) of Theorem \ref{t11}, we know that either
$\rho(h) \in \mathbb{R} \setminus \mathbb{Q}$ or $\rho(h) =
\frac{1}{2i}$ (mod 1); where $i$ is an odd integer. It follows that
in the first case, $f^2 = \textrm{id}$, and in the second case,
$f^2$ is strongly reversible in $\textrm{PL}^+(\mathbb{S}^{1})$ (see
the proof of the part (1) of Theorem \ref{t11}). By Lemma \ref{l29},
$f^2$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$. As $\textrm{Fix}f^{2}\neq \emptyset$, we conclude by Lemma \ref{l211} that $f$ is
strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$. \ \qed \\
\medskip
\section{\bf Reversibility in $\textrm{PL}(\mathbb{S}^{1})$}
\medskip
\subsection{\bf Reversibility in $\textrm{PL}(\mathbb{S}^{1})$ of
elements of $\textrm{PL}^+(\mathbb{S}^{1})$} The aim of this section
is to prove the following proposition.
\medskip
\begin{prop}\label{p13}
Let $f \in \mathrm{PL}^+(\mathbb{S}^{1})$. Then $f$ is reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of
$\mathrm{PL}^-(\mathbb{S}^{1})$ if and only if it is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$.
\end{prop}
\medskip
\begin{lem}\label{l31} \begin{itemize}
\item[(1)] Let $I = (a, b)$ be an open interval in $\mathbb{R}$ or in $\mathbb{S}^{1}$.
Then every fixed point free element $v\in \mathrm{PL}^+(I)$ is strongly
reversible in $\mathrm{PL}(I)$ by an element of $\mathrm{PL}^-(I)$.
\item[(2)] Let $f \in \mathrm{PL}^+(\mathbb{S}^{1})$. If $f$ has exactly one
fixed point, then $f$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^-(\mathbb{S}^{1})$.
\end{itemize}
\end{lem}
\begin{proof} (1) \textbullet \ Let $I$ be an open interval in $\mathbb{R}$ and $v\in \textrm{PL}^+(I)$ be a
fixed point free homeomorphism. A similar construction as in the proof of (\cite{J}, Theorem 1)
prove that there exists an involution $\alpha \in \textrm{PL}^-(I)$ satisfying $v^{-1} = \alpha v\alpha$.
\textbullet \ Assume that $I = (a, b)$ is an open interval in $\mathbb{S}^{1}$.
Then, there exists an open interval $\widehat{I} = (t_{1}, t_{2})$ in $\mathbb{R}$ such that the map
$\varphi: (t_{1}, t_{2})\longrightarrow (a, b)$ given by $\varphi(t) = e^{2i\pi t}$, is a homeomorphism. If $v$ is a fixed point free element in
$\textrm{PL}^+(I)$, then $\varphi^{-1}v\varphi$ is a fixed point free element in $\textrm{PL}^+(\widehat{I})$. Then, by the above,
there exists an involution $\alpha \in \textrm{PL}^-(\widehat{I})$ satisfying
$\varphi^{-1} v^{-1} \varphi = \alpha \varphi^{-1}v\varphi \alpha $.
It follows that $v^{-1} = (\varphi \alpha \varphi^{-1}) v (\varphi \alpha \varphi^{-1})$. As
$\tau = \varphi \alpha \varphi^{-1} \in \textrm{PL}^-(I)$, we conclude that $v$ is strongly reversible in $ \textrm{PL}(I)$
by an element of $\textrm{PL}^-(I)$.
\medskip
(2) Let $\{a\} = \textrm{Fix}(f)$. By (1), the restriction $f_{| \ \mathbb{S}^{1}\setminus \{a\}}$
is strongly reversible in $\textrm{PL}(\mathbb{S}^{1}\setminus \{a\})$ by an involution $\sigma\in
\textrm{PL}^-(\mathbb{S}^{1}\setminus \{a\})$. Then we extend
$\sigma$ to a map $\hat{\sigma}: \mathbb{S}^{1}\longrightarrow
\mathbb{S}^{1}$
given by $$\hat{\sigma}(x) = \begin{cases} \sigma(x), & \textrm{if } \ x \in \mathbb{S}^{1}\setminus \{a\},\\
a, & \textrm{if }\ x = a
\end{cases}$$
\vskip 3 mm
\noindent We see that $\hat{\sigma}$ is an involution in $\textrm{PL}^-(\mathbb{S}^{1})$ which satisfies
$f^{-1} = \hat{\sigma}f\hat{\sigma}$.
\end{proof}
\medskip
\begin{lem}\label{l32} Let $f\in \textrm{PL}^+(\mathbb{S}^{1})$ such that $\rho(f) = 0$. If $f$ is reversible in
$\mathrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^-(\mathbb{S}^{1})$ then it is strongly reversible in
$\mathrm{PL}(\mathbb{S}^{1})$ by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$.
\end{lem}
\medskip
\begin{proof} Assume that there exists $h \in \textrm{PL}^-(\mathbb{S}^{1})$ such that $f^{-1} = h f h^{-1}$.
Let us show that $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$.
If $f$ has exactly one fixed point, then the conclusion follows from Lemma \ref{l31}. Now, assume that $f$ has more than one fixed point. Since $h \in
\textrm{PL}^-(\mathbb{S}^{1})$, so $h$ has exactly two fixed points
$a$ and $b$ which divides the circle $\mathbb{S}^{1}$ onto two
connected components $A=(a,b)$ and $B=(b,a)$ satisfying $h(A) = B$
and $h(B) = A$. Moreover we have always $\textrm{Fix}(f) \cap A \neq
\emptyset$ and $\textrm{Fix}(f) \cap B \neq \emptyset$. Let $c$ be
the nearest point of $\textrm{Fix}(f) \cap A$ to the point $a$, and
let $d$ be the nearest point of $\textrm{Fix}(f) \cap A$ to the
point $b$. From the equality $f^{-1} = h f h^{-1}$, we see that
$h(c)$ is the nearest point of $\textrm{Fix}(f) \cap B$ to $a$ and
that $h(d)$ is the nearest point of $\textrm{Fix}(f) \cap B$ to $b$.
The restrictions $f_{|(h(c), c)}$ and $f_{|(d, h(d))}$ are fixed
point free piecewise linear homeomorphisms of open arcs of the
circle $\mathbb{S}^{1}$. Then by
Lemma \ref{l31}, (1), they are reversed respectively in $\textrm{PL}((h(c), c))$ and $\textrm{PL}((d, h(d)))$ by involutions
$\sigma_1 \in \textrm{PL}^-((h(c), c))$ and $\sigma_2 \in
\textrm{PL}^-((d, h(d)))$. Define $\tau: \mathbb{S}^{1}\longrightarrow \mathbb{S}^{1}$ as follows
\hskip 53 mm $$\ \ \tau(x)= \begin{cases}
h(x), & \textrm{if } x \in [c,d],\\
h^{-1}(x), & \textrm{if } x \in [h(d), h(c)]\\
\sigma_1(x), & \textrm{if } x\in (h(c), c)\\
\sigma_2(x), & \textrm{if } x\in (d, h(d))
\end{cases}$$
\vskip 3 mm
\noindent We can easily see that $\tau$ is an involution in $\textrm{PL}^-(\mathbb{S}^{1})$ that satisfies $f^{-1} = \tau f \tau$.
\end{proof}
\noindent \emph{\it Proof of Proposition \ref{p13}}. Assume that $f$ is reversible in $\textrm{PL}(\mathbb{S}^{1})$
by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$.
We distinguish two cases:
\noindent Case 1: $\rho(f) = 0$. Then $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$ by Lemma \ref{l32}.
\noindent Case 2: $\rho(f) \in \mathbb{R} \setminus \mathbb{Q}$. Then by Theorem \ref{t26}, there exists
$\alpha \in \textrm{Homeo}(\mathbb{S}^{1})$ such that $f = \alpha r \alpha^{-1}$; where $r$ is the rotation of $\mathbb{S}^{1}$
by $\rho(f)$. On the other hand, there exists $h \in \textrm{PL}^-(\mathbb{S}^{1})$ such that $f^{-1} = h f h^{-1}$, which implies that $r^{-1} = g r g^{-1}$, where $g = \alpha^{-1}h \alpha$. Then \begin{equation}g^2r = rg^2. \end{equation}
\noindent Since $g$ is an orientation-reversing element of $\textrm{Homeo}(\mathbb{S}^{1})$,
$\textrm{Fix}(g) \neq\emptyset$. Let $a \in \textrm{Fix}(g) \subset \textrm{Fix}(g^2)$. The equality
(4.1) implies that for each $n\in \mathbb{Z}$, $g^2 r^n = r^n g^2$.
It follows that $r^n(a) \in \textrm{Fix}(g^2)$, for each $n\in
\mathbb{Z}$ and by the fact that $\mathbb{S}^{1} =
\overline{\{r^n(a): \ n\in \mathbb{Z}\}}$, we obtain that
$\mathbb{S}^{1} = \textrm{Fix}(g^2)$. Thus $g^2 = \textrm{id}$ and
so $h^2 = \textrm{id}$. We conclude that $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by the involution $h\in \textrm{PL}^-(\mathbb{S}^{1})$.
\noindent Case 3. $\rho(f) = \frac{p}{q}\in \mathbb{Q}\backslash \{0\}$. In this case, $\rho(f^q) = 0$ and by the case 1,
$f^q$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$. So
by Lemma \ref{l211}, $f$ is strongly reversible in
$\textrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$. \qed \vskip 3 mm
\medskip
\subsection{\bf Reversibility in $\textrm{PL}(\mathbb{S}^{1})$ of
elements of $\textrm{PL}^-(\mathbb{S}^{1})$} \ \ In this paragraph
we study reversibility in $\textrm{PL}(\mathbb{S}^{1})$ of elements of
$\textrm{PL}^-(\mathbb{S}^{1})$ by
proving the following proposition.
\begin{prop}\label{p12}
Let $f \in \mathrm{PL}^-(\mathbb{S}^{1})$. Then the following statements are equivalent.
\begin{enumerate}
\item $f$ is reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^+(\mathbb{S}^{1})$.
\item $f$ is reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^-(\mathbb{S}^{1})$.
\item $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^-(\mathbb{S}^{1})$.
\item $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\mathrm{PL}^+(\mathbb{S}^{1})$.
\end{enumerate}
\end{prop}
\medskip
\begin{lem}\label{l61}
Let $f \in \mathrm{PL}^-(\mathbb{S}^{1})$. Then the following statements are equivalent.
\begin{enumerate}
\item $f$ is reversible by an element $h \in \mathrm{PL}(\mathbb{S}^{1})$ that fixes each of the fixed points of $f$.
\item $f^2 = \mathrm{id}$.
\end{enumerate}
\end{lem}
\begin{proof} $(1)\Longrightarrow (2)$: Since $f$ is orientation-reversing, so it has exactly two fixed points
$a$ and $b$. Set $I = \mathbb{S}^{1} \backslash \{a\}$, it is an open interval in $\mathbb{S}^{1}$.
As $f^{-1} = h f h^{-1}$ and $h(a) = a$, the restriction $f_{| I} \in \hbox{PL}^-(I)$ and is reversed by
$h_{| I} \in \hbox{PL}(I)$. Then by Lemma \ref{l299}, (1),
$f_{| I}$ is an involution and so is $f$.
$(2)\Longrightarrow (1)$: is clear.
\end{proof}
\vskip 3 mm
\noindent \emph{\it Proof of Proposition \ref{p12}.} \noindent $(1) \Longrightarrow (2)$: Let $f\in
\textrm{PL}^-(\mathbb{S}^{1})$, $h \in \textrm{PL}^+(\mathbb{S}^{1})$ such that $f^{-1} = h f
h^{-1}$. So $f^{-1} = (fh) f(h^{-1}f^{-1})$. Hence $f$ is reversed in $\textrm{PL}(\mathbb{S}^{1})$ by
$fh\in \textrm{PL}^-(\mathbb{S}^{1})$.
\noindent $(2) \Longrightarrow (3)$: Let $f\in \textrm{PL}^-(\mathbb{S}^{1})$ and
$h \in \textrm{PL}^-(\mathbb{S}^{1})$ be such that $$f^{-1} = h f h^{-1}. \ \ \ \ \ (\ast)$$
Since $f$ is orientation-reversing, so it has exactly two fixed points
$a$ and $b$. We have $h(\textrm{Fix}(f)) = \textrm{Fix}(f)$. So, either $h$ fixes each of $a$ and $b$ or it
interchanges them.
In the first case, we have $f^2 = \textrm{id}$ by Lemma \ref{l61}. So
$f$ is an involution in $\textrm{PL}^-(\mathbb{S}^{1})$ and hence
it is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of $\textrm{PL}^-(\mathbb{S}^{1})$. In the second case; that is $h(a)
= b$ and $h(b) = a$, we have $h((a,b)) = (a,b)$. So by
equality $(\ast)$, the restriction $f^2_{|(a,b)}$ is an element of $\textrm{PL}^+((a,b))$ that is reversed by
$h_{|(a,b)}$. Thus, by Lemma \ref{l299}, (2), $f^2_{|(a,b)}$ is strongly reversible in $\textrm{PL}((a,b))$
by an involution $\tau\in
\textrm{PL}^-((a,b))$; that is, $$f^{-2}_{|(a,b)} = \tau f^2_{|(a,b)} \tau. \ \ \ \ \ \ \ \ (\ast \ast)$$
\noindent Let $\mu: \mathbb{S}^{1} \longrightarrow \mathbb{S}^{1}$ be the map defined by $$\mu(x) = \left\{
\begin{array}{ll}
\tau(x), & \textrm{if } \ x \in [a, b] \\
f^{-1} \tau f^{-1}(x), & \textrm{if
}
\
x \in [b, a]
\end{array}
\right.$$
\noindent Clearly $\mu \in \textrm{PL}^-(\mathbb{S}^{1})$ and $\mu f \mu = f^{-1}$. Moreover,
by equality $(\ast \ast)$, we have $\mu^2 = \textrm{id}$. This
implies that $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$.
\noindent $(3) \Longrightarrow (4)$. Assume that $f^{-1} = \tau f \tau$, where $\tau$ is an involution in
$\textrm{PL}^-(\mathbb{S}^{1})$. Then $(f \tau)^2 = \textrm{id}$ and so $f^{-1} = (f\tau) f (\tau
f^{-1})$. Hence $f$ is also strongly reversible in $\textrm{PL}(\mathbb{S}^{1})$ by the
involution $(f\tau)$ in $\textrm{PL}^+(\mathbb{S}^{1})$.
\noindent $(4) \Longrightarrow (1)$ is clear. \qed
\bigskip
{\it Proof of Theorem \ref{t2}}. This follows from Theorem
\ref{t11}, Propositions \ref{p13} and \ref{p12}. \qed
\medskip
\begin{rem} \label{r3} In Proposition
\ref{p12}, we showed that any reversible element in $\textrm{PL}(\mathbb{S}^{1})$ by an element of
$\textrm{PL}^-(\mathbb{S}^{1})$ must be strongly reversible. This
does not hold for elements of $\textrm{Homeo}^-(\mathbb{S}^1)$ (see
\cite{GOS}).
\end{rem}
\vskip 3 mm
\section{\bf Strong reversibility in $\textrm{PL}(\mathbb{S}^{1})$ of elements of $\textrm{PL}^+(\mathbb{S}^{1})$}
\medskip
\subsection{Strong reversibility of elements of
$\textrm{PL}^+(\mathbb{S}^{1})$}
\medskip
The aim of this subsection is to prove the part (1) of Theorem
\ref{t52}.
\medskip
\begin{lem}\label{l53}
Let $f, \ g \in \mathrm{PL}^+(\mathbb{S}^1)$ such that $\mathrm{Fix}(f) \neq \emptyset \neq \mathrm{Fix}(g)$.
If $\Delta_f = \Delta_g$, then there exists $v \in \mathrm{PL}^+(\mathbb{S}^1)
$ such that $g = vfv^{-1}$ and $v = \mathrm{id}$ on $\mathrm{Fix}(f)$.
\end{lem}
\begin{proof} Since $\Delta_f = \Delta_g$, we have $\textrm{Fix}(f) = \textrm{Fix}(g)$.
For each open interval component $(a,b)$ of $S^1 \backslash \textrm{Fix}(f)$,
there exists an orientation preserving piecewise linear homeomorphism
$u: (a,b) \longrightarrow \mathbb{R}$. Then $u f u^{-1}$ and $u g u^{-1}$ are two fixed point free elements
of $\textrm{PL}^+(\mathbb{R})$. Since $\Delta_f = \Delta_g$, $u f u^{-1}$ and $u g u^{-1}$
are conjugate in $\textrm{PL}^+(\mathbb{R})$ by (\cite{GS}, Proposition 2.6). Let $v_0 \in PL^+((a,b))$ such that $g(x) = v_0 f v_0^{-1}(x)$ for $x \in (a, b)$. Then the map $v$ defined by
$v(x) = v_0(x)$ for $x \in (a, b)$ and $v(x) = x$ for $x \in
\textrm{Fix}(f)$, is the required homeomorphism.
\end{proof}
\medskip
\begin{lem}\label{l54}
Let $f \in \mathrm{PL}^+(\mathbb{S}^1)$ such that $\rho(f) = \frac{1}{2}$. Then $f$ is strongly reversible in
$\mathrm{PL}^+(\mathbb{S}^1)$ if and only if $f^2 = \mathrm{id}$.
\end{lem}
\medskip
\begin{proof} Lemma \ref{l54} is a particular case of (\cite{GOS}, Theorem 3.3).
\end{proof}
\medskip
\noindent \emph{Proof of the part (1) of Theorem \ref{t52}}.
Assume that $f$ is a strongly reversible element of $\textrm{PL}^+(\mathbb{S}^1)$. We know that either $\rho(f) = 0$ or $\rho(f) = {1 \over 2}$.
If $\rho(f) = {1 \over 2}$, then by Lemma \ref{l53}, $f^2 = \textrm{id}$. If $\rho(f) = 0$, then
$\textrm{Fix}(f)\neq \emptyset$, and since $f$ is strongly reversible in $\textrm{PL}^+(\mathbb{S}^1)$, there exists
an involution $h \in \textrm{PL}^+(\mathbb{S}^1)$ such that $f^{-1} = h^{-1} f h$. Therefore,
$\rho(h) = \frac{1}{2}$ and by (\cite{GOS}, Lemma 2.1), $\Delta_f =
-\Delta_f \circ h$.
Conversely, assume that $f \in
\textrm{PL}^+(\mathbb{S}^1)$ such that $\textrm{Fix}(f)\neq
\emptyset$ and there exists $h\in \textrm{PL}^+(\mathbb{S}^1)$ with
$\rho(h) = \frac{1}{2}$ satisfying $\Delta_f = -\Delta_f \circ h$. Then $\Delta_{f^{-1}} = \Delta_{h^{-1} f h}$.
By Lemma \ref{l53}, there exists $v \in \textrm{PL}^+(\mathbb{S}^1)$ such that $f^{-1} = v^{-1}h^{-1} f h v$;
which means that $f$ is reversible by $hv\in \textrm{PL}^+(\mathbb{S}^1)$. Since $\textrm{Fix}(f)\neq \emptyset$,
$\rho(f) = 0$ and by Theorem \ref{t11}, $f$ is strongly reversible in $\textrm{PL}^+(\mathbb{S}^1)$.
If $f^2 = \textrm{id}$, then it is clear that $f$ is strongly reversible in $\textrm{PL}^+(\mathbb{S}^1)$
by the identity map. \qed \\
\subsection{Strong reversibility of elements of
$\textrm{PL}^-(\mathbb{S}^{1})$}
\medskip
The aim of this subsection is to prove the part (2) of Theorem
\ref{t52}.
\medskip
\begin{lem}\label{l55}
Let $f \in \mathrm{PL}^+(\mathbb{S}^1)$ with $\rho(f) = 0$. Then $f$ is strongly reversible in $\mathrm{PL}(\mathbb{S}^1)$
by an element of $\mathrm{PL}^-(\mathbb{S}^1)$ if and only if there exists $h \in
\mathrm{PL}^-(\mathbb{S}^1)$ such that $\Delta_f = \Delta_f \circ
h$.
\end{lem}
\noindent \emph{Proof.} Let $f \in \textrm{PL}^+(\mathbb{S}^1)$ such that $\rho(f) = 0$. If $f$
is strongly reversible in $\mathrm{PL}(\mathbb{S}^1)$ by an element of $\mathrm{PL}^-(\mathbb{S}^1)$ then there exists an involution $h \in \textrm{PL}^-(\mathbb{S}^1)$ such that
$f^{-1} = h^{-1}fh$. Thus by Lemma \ref{l100}, $\Delta_f = -\textrm{deg}(h) \Delta_f \circ h = \Delta_f \circ h$.
Conversely, if $\Delta_f = \Delta_f \circ h$ for some element $h \in \textrm{PL}^-(\mathbb{S}^1)$ then
$\Delta_f = -\textrm{deg}(h) \Delta_f \circ h$ and $\Delta_{f^{-1}}
= \Delta_{h^{-1} f h}$. So, by the proof of the part (1) of Theorem
\ref{t52}, $f$ is reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of $\textrm{PL}^-(\mathbb{S}^1)$. By
Proposition \ref{p13}, $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of
$\textrm{PL}^-(\mathbb{S}^1)$. \qed
\bigskip
\noindent \emph{Proof of the part (2) of Theorem \ref{t52}}. (i): Let $f \in \textrm{PL}^+(\mathbb{S}^1)$
such that $\rho(f) \in \Bbb Q$. If $f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of
$\textrm{PL}^-(\mathbb{S}^1)$, then there exists an involution $h \in \textrm{PL}^-(\mathbb{S}^1)$ such that
$f^{-1} = h^{-1}fh$, which implies that $f^{-n_f} = h^{-1} f^{n_f} h$. So by Lemma \ref{l55}, we have
$\Delta_{f^{n_f}} = \Delta_{f^{n_f}} \circ h$. Conversely, assume that there exists
$h \in \textrm{PL}^-(\mathbb{S}^1)$ such that $\Delta_{f^{n_f}} = \Delta_{f^{n_f}} \circ h$. Then by Lemma \ref{l55},
$f^{n_f}$ is strongly reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of $\textrm{PL}^-(\mathbb{S}^1)$ and by
Lemma \ref{l211},
$f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of $\textrm{PL}^-(\mathbb{S}^1)$.
(ii): Let $f \in \textrm{PL}^+(\mathbb{S}^1)$ such that $\rho(f)\in \mathbb{R}\setminus \Bbb Q$. If $f$ is
strongly reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of $\textrm{PL}^-(\mathbb{S}^1)$ then
there exists an involution $\tau\in \textrm{PL}^-(\mathbb{S}^1)$ such that
$f^{-1} = \tau f\tau$. On the other hand, by Theorem \ref{t26}, there is $h\in \textrm{Homeo}^+(\mathbb{S}^{1})$
such that $f = hr_{\rho(f)}h^{-1}$. Therefore $hr_{\rho(f)}^{-1}h^{-1}= \tau hr_{\rho(f)} h^{-1}\tau$. As $r_{\rho(f)}^{-1} = sr_{\rho(f)}s$, where
$s: z\mapsto \overline{z}$ is the reflection, so
$h^{-1}\tau h sr_{\rho(f)}= r_{\rho(f)}h^{-1}\tau h s$. Hence $h^{-1}\tau h s = r_{t}$ for some $t \in \mathbb{R}$.
It follows that
$\tau = hr_{t}sh^{-1}\in \textrm{PL}^-(\mathbb{S}^1)$.
Conversely, if there is $h\in \textrm{Homeo}^+(\mathbb{S}^{1})$
such that $f = hr_{\rho(f)}h^{-1}$ where $h$ satisfies $hrsh^{-1}\in \textrm{PL}^-(\mathbb{S}^1)$, for some rotation $r$ of
$\mathbb{S}^1$, then $\tau= hrsh^{-1}$ is an involution in $\textrm{PL}^-(\mathbb{S}^1)$ and satisfies $f^{-1} = \tau f\tau$
(since $rs = sr^{-1}$). Therefore
$f$ is strongly reversible in $\textrm{PL}(\mathbb{S}^1)$ by an element of $\textrm{PL}^-(\mathbb{S}^1)$. \hfill \qed \\
\vskip 3 mm
\section{\bf Proof of Theorem \ref{t91}}
\noindent \emph{Proof of $(i)$}. If $f^2 = \textrm{id}$, there is nothing to prove.
If $f^2 \neq id$, from Theorem \ref{t52}.(1), it suffices to find two involutions $\tau$ and $h$ in $\textrm{PL}^+(\mathbb{S}^1)$
such that Fix$(\tau f) \neq \emptyset$ and $\Delta_{\tau f} = -\Delta_{\tau f} \circ h$ since in that case, $f$ is a composition of three involutions of
$\textrm{PL}^+(\mathbb{S}^1)$. There is a point $x$ in $\mathbb{S}^1$
such that $x\neq f^2(x)$. We can assume that the points $x, \ f(x)$ and $f^2(x)$ occur in that order anticlockwise around
$\mathbb{S}^1$. Choose a point $y$ in $(x, f(x))$ such that $f^{-1}(y)$ be in $(f^2(x), x)$.
Let $u : [x, f(x)] \longrightarrow [f(x), x]$ be an orientation-preserving piecewise linear homeomorphism such that:\\
$u(y) = f(y)$,
$$\begin{array}{ll}
u(t) < \min(f(t), f^{-1}(t)), & \textrm{for} \ t \in (x, y); \\
f(t)<u(t)<f^{-1}(t), & \textrm{for }
\
t\in (y, f(x)).
\end{array}$$
\noindent Then, let $\tau$ be the involution in PL$^+(\mathbb{S}^1)$ defined by $$\tau(t) = \left\{
\begin{array}{ll}
u(t), & \textrm{if } \ t \in [x, f(x)] \\
u^{-1}(t), & \textrm{if
}
\
t \in [f(x), x]
\end{array}
\right.$$
\noindent We have $\tau f(t) = t$ if and only if $t=x$ or $t=y$. So, $\textrm{Fix}(\tau f) = \{x, y\}$. Moreover, we have:
$$\begin{array}{ll}
\forall \ t \in (x,y), & f(t) > u(t) \Longleftrightarrow \tau f(t) > t \\
\forall \ t \in (y, f(x)), & f(t) < u(t) \Longleftrightarrow \tau f(t) < t \\
\forall \ t \in (f(x), x), & u\left(u^{-1}(t)\right) < f^{-1}\left(u^{-1}(t)\right) \Longleftrightarrow \tau f(t) < t.
\end{array}$$
\noindent Therefore: $$\Delta_{\tau f}(t) = \left\{
\begin{array}{ll}
0, & \textrm{if } \ t = x, y \\
1, & \textrm{if }
\
t \in (x, y) \\
-1, & \textrm{if } \ t \in (y, x).
\end{array}
\right.$$
Now, let $v: [x,y] \longrightarrow [y,x]$ be any orientation-preserving piecewise linear homeomorphism, and let $h$ be
the involution in $\textrm{PL}^+(\mathbb{S}^1)$ defined by $$h(t) = \left\{
\begin{array}{ll}
v(t), & \textrm{if } \ t \in [x, y] \\
v^{-1}(t), & \textrm{if
}
\
t \in [y, x]
\end{array}
\right.$$
\noindent It is easy to see that $h$ satisfies $\Delta_{\tau f} \circ h = - \Delta_{\tau f}$.
We conclude that each member
of $\textrm{PL}^+(\mathbb{S}^{1})$ can be expressed as a composite of three involutions of
$\textrm{PL}^+(\mathbb{S}^{1})$.
So, $\textrm{PL}^+(\mathbb{S}^{1}) =
I_3(\textrm{PL}^+(\mathbb{S}^{1})) =
R_2(\textrm{PL}^+(\mathbb{S}^{1}))$.
There are elements in $\textrm{PL}^+(\mathbb{S}^{1})$ which are not strongly reversible in $\textrm{PL}^+(\mathbb{S}^{1})$; one can choose,
for example, a homeomorphism $f \in \textrm{PL}^+(\mathbb{S}^{1})$ which is not an involution and
with rotation number $\rho(f) = \frac{1}{2}$ (such map $f$ is not strongly reversible in
$\textrm{PL}^+(\mathbb{S}^{1})$ by Lemma \ref{l53}).
The fact that $R_1(\textrm{PL}^+(\mathbb{S}^{1})) {\subset}_{\neq} I_2(\textrm{PL}(\mathbb{S}^{1}))$
follows from Theorem \ref{t11}.
\medskip
\textit{ Proof of $(ii)$}. If $f\in \textrm{PL}^+(\mathbb{S}^{1})$, then by (i), $f\in I_3(\textrm{PL}(\mathbb{S}^{1}))$.
If $f\in \textrm{PL}^-(\mathbb{S}^{1})$, then
$\textrm{Fix}(f)\neq \emptyset$. Let $a\in \textrm{Fix}(f)$ and let $I = \mathbb{S}^{1}\setminus \{a\}$.
Then, there exists an open interval $\widehat{I}$ in $\mathbb{R}$ such that the map
$\varphi: \widehat{I}\longrightarrow I$ given by $\varphi(t) = e^{2i\pi t}$, is a homeomorphism. Set $g = \varphi^{-1} f \varphi$.
We have $g \in \textrm{PL}^-(\widehat{I})$. Choose an involution $\sigma \in \textrm{PL}^-( \widehat{I})$ such that,
for each $x\in \widehat{I}$, $\sigma(x) > g(x)$. Then $g(\sigma(x)) < x$, for each $x\in \widehat{I}$. Therefore, $g\sigma$ is a
fixed point free element in $\textrm{PL}^+( \widehat{I})$. By Lemma 4.2, (1), it is strongly reversible
in $\textrm{PL}( \widehat{I})$ by an element of $\textrm{PL}^-( \widehat{I})$; which means that there exist three involutions $u$, $v$ in $\textrm{PL}^-(\widehat{I})$ such that
$g\sigma = u v$. Thus $g = uv\sigma$. It follows that $f_{| I} = \varphi g \varphi^{-1} =
(\varphi u \varphi^{-1})(\varphi v \varphi^{-1})(\varphi \sigma \varphi^{-1})$. By extending $\varphi u \varphi^{-1}$,
$\varphi v \varphi^{-1}$ and $\varphi \sigma \varphi^{-1}$ to $\mathbb{S}^{1}$ by fixing $a$, we get three involutions $\tau_1, \ \tau_2$ and $\tau_3$
in $\textrm{PL}^-(\mathbb{S}^{1})$ satisfying
$f = \tau_1\tau_2\tau_3$. Hence $f\in I_3(\textrm{PL}(\mathbb{S}^{1}))$. We conclude that
$\textrm{PL}(\mathbb{S}^{1}) = I_3(\textrm{PL}(\mathbb{S}^{1}))=
R_2(\textrm{PL}(\mathbb{S}^{1}))$. Moreover,
$\textrm{PL}(\mathbb{S}^{1})\neq I_2(\textrm{PL}(\mathbb{S}^{1}))$
as in the proof of (i). From Theorem \ref{t2}, we have
$R_1(\textrm{PL}(\mathbb{S}^{1})) =
I_2(\textrm{PL}(\mathbb{S}^{1}))$. Now to show that
$I_2(\textrm{PL}(\mathbb{S}^{1})) \neq
I_1(\textrm{PL}(\mathbb{S}^{1}))$, it suffices to choose a
nontrivial reversible element $f$ in $\textrm{PL}^+(\mathbb{S}^{1})$
which is not the identity and with rotation number $\rho(f) = 0$.
\qed
\medskip
\begin{rem} \label{r4} Contrarily to $\mathrm{PL}^-(\mathbb{R)}$ (cf. Lemma \ref{l299}), there exists an element of $\mathrm{PL}^-(\mathbb{S}^{1})$ which is strongly
reversible in $\mathrm{PL}(\mathbb{S}^{1})$ but not an involution.
\end{rem}
\begin{proof} Indeed, Suppose that the remark is not true. We will prove in this case that any element
$f\in \textrm{PL}^-(\mathbb{S}^{1})$ is an involution, this leeds to a contradiction since there are elements in
$\textrm{PL}^-(\mathbb{S}^{1})$ which are not involutions.
Indeed, let $\sigma$ be any involution in $\textrm{PL}^-(\mathbb{S}^{1})$. Then $\sigma f\in \textrm{PL}^+(\mathbb{S}^{1})$ and from Theorem \ref{t91}, (i), there exist three involutions
$\tau_1, \ \tau_2$ and $\tau_3$ in $\textrm{PL}^+(\mathbb{S}^{1})$ such that $\sigma f = \tau_1 \tau_2 \tau_3$. This implies that
$f = \sigma \tau_1 \tau_2 \tau_3$. By assumption, $\sigma \tau_1$ is an involution in
$\textrm{PL}^-(\mathbb{S}^{1})$ and then so is $(\sigma \tau_1) \tau_2$. We conclude that $f$ is an involution.
\end{proof}
|
\section{Introduction}
Iron (Fe) and carbon (C) are the basic components of any steel and particularly
of the low alloy banitic steels used in the reactor pressure vessels (RPV) of most commercial nuclear power plants. Neutron irradiation induces the creation of point defects, \textit{i.e.} self-interstitial atoms (SIA) and vacancies (V), isolated or in clusters. These, over the life time of the RPV, will have significant impact on the integrity of the material. Computer simulation models are an effective way to understand these kind of degradation processes and a first step in building a physical model for RPV steels is to address the Fe-C system.
In \cite{jansson2013simulation}, we presented an Object Kinetic Monte Carlo
(OKMC) model to describe the nanostructural evolution under irradiation in Fe-C alloys at temperatures $<$350 K, as well as post-irradiation annealing at temperatures up to 700 K. In the model, the effect of C was introduced by means of effective traps for vacancy and SIA type defects. The parameters used were strictly derived from physical considerations or atomistic calculations, or determined by values that reflected the reference experimental conditions. Only one parameter was used to guarantee the best reproduction possible of experimental data via fitting. In this paper, we examine the sensitivity of the model to the variation of this, as well as other parameters, such as the concentration of carbon in terms of traps, other features of the traps, and an environmental parameter such as the dose-rate.
The simulation results are compared for convenience with the low temperature ($<$350 K) irradiation experiment by Eldrup, Singh and Zinkle
\cite{zinkle2006microstructure,eldrup2002dose}, which was also used as a
reference experiment in \cite{jansson2013simulation}. This is one of the most
complete irradiation experiments for pure iron because both the
evolution with dose of vacancies and SIA clusters has been traced, using respectively positron annihilation spectroscopy (PAS) and transmission electron
microscopy (TEM). We have also included experimental data from
other TEM studies from similar experiments, such as
\cite{singh1999effects,bryner1966study,robertson1982low,horton1982tem,
takeyama1981,eyre1965electron}. A more thorough overview of the available irradiation and
annealing experiments in Fe-C will be published in a separate paper \cite{malerba2011review}.
This paper is organized as follows: In Sec. \ref{sec:methods} our computational method is briefly overviewed. In Sec. \ref{sec:traps}, we study the nanostructure evolution when no traps are present. In Sec. \ref{sec:C_density}, we study the effect of the trap concentration, which correlates with the carbon content. In Sec. \ref{sec:threshold} and \ref{sec:etlarge}, the sensitivity of variables
giving the size dependence of the SIA traps are studied.
In Sec. \ref{sec:flux}, we study the model's sensitivity to the dose rate. Finally, we
discuss the results and present our conclusions in Sec. \ref{sec:discussion}.
\section{Computation method}\label{sec:methods}
For our OKMC simulations, we use the code {\sc LAKIMOCA}, thoroughly described in \cite{domain2004simulation}. Our methodology is described in \cite{jansson2013simulation}, of which this paper is a
direct follow-up. Here we point out only the main features of the method.
Following our work in \cite{jansson2013simulation}, most of the simulations in this paper have a set-up corresponding to the
experiment of Eldrup, Zinkle and Singh \textit{et al.}
\cite{zinkle2006microstructure,eldrup2002dose}. Accordingly, we use as reference conditions a dose rate of
$7\cdot10^{-7}$ dpa/s (except in Sec \ref{sec:flux}) and reach 0.73 dpa, we introduce 100 appm of spherical traps for vacancies, corresponding to
the amount of carbon and nitrogen in the iron used in the aforementioned experiment (except in Sec. \ref{sec:C_density}), and we set the temperature to 343 K in all cases. Unless otherwise stated, we use a simulation box size of $350\times400\times450\times a_0^3$.
For simulating neutron irradiation, debris of vacancy and SIA objects of different sizes are randomly chosen from a database of displacement cascades \cite{stoller1996point,stoller1997primary,stoller2000statistical,
stoller2000evaluation,stoller2004secondary,stoller2000role} and randomly introduced into the system at a certain rate per time and volume. The displacement cascades were simulated using the Finnis-Sinclair potential \cite{finnis1984simple} and the considered cascade energies are 5 keV, 10 keV, 20 keV, 30 keV, 40 keV, 50 keV and 100 keV. The cluster size distribution per cascade, as used in our simulations, are shown for the SIA clusters in Fig. \ref{cascade_sia_size_distribution.pdf} and for the vacancy clusters in Fig. \ref{cascade_vac_size_distribution.pdf}. Vacancy clusters and SIA clusters smaller than size 150 have spherical shapes defined by a capture radius that depends on the cluster type and size (See \cite{jansson2013simulation} for details). The capture radius for a single vacancy is 4.315 Å and for a single SIA 6.396 Å. If two clusters overlap, they will recombine. The accumulated dpa is calculated using the NRT formula \cite{domain2004kinetic,norgett1975proposed}:
\begin{equation}
dpa = \frac{0.8 E_{MD}}{2E_D},
\end{equation}
where $E_{MD}$ is the cascade energy and $E_D = 40$ eV is the displacement threshold energy for Fe.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{cascade_sia_size_distribution.pdf}
\caption{SIA cluster mean size distribution per cascade for different cascade energies in the used database \cite{stoller1996point,stoller1997primary,stoller2000statistical,
stoller2000evaluation,stoller2004secondary,stoller2000role}.}
\label{cascade_sia_size_distribution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{cascade_vac_size_distribution.pdf}
\caption{Vacancy cluster mean size distribution per cascade for different cascade energies in the used database \cite{stoller1996point,stoller1997primary,stoller2000statistical,
stoller2000evaluation,stoller2004secondary,stoller2000role}.}
\label{cascade_vac_size_distribution.pdf}
\end{figure}
Vacancy and SIA clusters have a mobility that depends on their cluster size, as described in detail in \cite{jansson2013simulation}. Vacancy clusters migrate randomly in 3D. SIA clusters change direction according to a rotation energy, derived in terms of a Boltzmann expression. SIA clusters of size 1 have a rotation energy $\eta=0$ and will change direction at every jump, like the vacancy clusters. Larger SIA clusters will have a rotation energy that increases gradually with size and their migration direction will change less frequently. SIA clusters of size 12 and above have a rotation energy of 1 eV and will migrate fully in 1D along their Burgers vectors, at the simulation temperature considered in this work. If two SIA clusters meet, the resulting cluster will assume the Burgers vector and migration direction of the larger parent cluster. The rotational energy of the new cluster will depend on the cluster size. Only $\langle 111 \rangle$ SIA clusters are considered.
Spherical sinks with a sink strength corresponding to a dislocation
density of $10^{12}$ m$^{-2}$ are introduced as well. Defects are removed from the system when the
capture radii of the defect and the sink overlap. The effect of grain boundaries
is allowed for using the algorithm described in \cite{malerba2007object}.
It is known that C forms complexes with vacancies and that these complexes, in turn, trap SIA clusters \cite{anento2013carbon}. These effects are introduced in the
system using generic spherical traps for SIA or vacancies. When the capture
radii of a defect and a trap overlap, the two will be bound together with a
specified trapping energy $E^\delta_t$, which depends on whether the defect is of vacancy
($\delta=v$) or SIA ($\delta=i$) type and on the size of the defect, as seen in Table
\ref{SIA_trapping_energies}. Above a threshold size $N_{th}$, we use a strong binding
energy for SIA clusters, $E^i_t=1.2$ eV, that can be associated with the binding energy of a CV$_2$ complex bound to the centre of a SIA cluster, and comparable to MD results \cite{anento2013carbon}. Below $N_{th}$, we use the binding energy with a single C atom. This choice allows for the fact that the vacancies in CV$_2$ complexes will only recombine if interacting with the edge of an SIA cluster and this type of interaction is the more likely, the smaller the cluster.
\begin{table}
\centering
\caption{Trapping energy for the SIA traps at $T=343 K$. The values
for SIA clusters of sizes 1--4
and the values for vacancy cluster of sizes 1--6 are from DFT calculations
\cite{becquart2011p60}.}
\label{SIA_trapping_energies}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}} l c c}
\toprule
$N^i$ & SIA $E^i_t$ [eV] & Vac. $E^v_t$ [eV]\\
\midrule
1 & 0.17 & 0.65\\
2 & 0.28 & 1.01\\
3 & 0.36 & 0.93\\
4 & 0.34 & 0.96\\
5 & 0.60 & 1.23\\
6 & 0.60 & 1.20\\
7--$N_{th}$ & 0.60 & 0.00\\
$N_{th}<$& 1.20 & 0.00\\
\bottomrule
\end{tabular*}
\end{table}
\section{Results}\label{sec:results}
\subsection{The effect of traps}\label{sec:traps}
Traps for vacancy and SIA clusters were shown in \cite{jansson2013simulation} to play a key role in the nanostructure evolution, where simulations of irradiation of Fe-C at 343 K obtained good agreement with experiments \cite{zinkle2006microstructure,eldrup2002dose}. In this study the traps were removed, which results in significantly lower vacancy cluster density, as shown in Fig. \ref{R20131230_vac.pdf}, even though the vacancy cluster mean size evolution is not significantly affected, as shown in Fig. \ref{R20131230_vac_mean_size_evolution.pdf}. Without traps, no visible SIA clusters appear, contrary to the experimental data \cite{eldrup2002dose} and the results in \cite{jansson2013simulation}, where traps were used.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20131230_vac.pdf}
\caption{The vacancy cluster number density versus dpa with and without traps for vacancies and SIA clusters. The experimental PAS data are from \cite{eldrup2002dose}.}
\label{R20131230_vac.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20131230_vac_mean_size_evolution.pdf}
\caption{The mean vacancy cluster size versus dpa. The experimental PAS data are from \cite{eldrup2002dose}.}
\label{R20131230_vac_mean_size_evolution.pdf}
\end{figure}
\subsection{The effect of carbon content}\label{sec:C_density}
We studied the model's sensitivity to C content by varying the concentration of SIA and
vacancy traps from 1 appm to 200 appm, 100 appm corresponds to the reference concentration \cite{zinkle2006microstructure,eldrup2002dose}, which the model was fine-tuned to reproduce. It is unlikely that the C content in the matrix would exceed these concentrations, even in steels. The SIA and vacancy trapping energies have a size dependency as described in Table \ref{SIA_trapping_energies} and the threshold $N_{th}=29$.
In Fig. \ref{R20120905-0_vac_dens.pdf}, the vacancy cluster density evolution for the
different trap concentrations is shown. It is observed that the higher the trap concentration, the higher the density of clusters, although this effect is strong for low trap contents and tends to saturate for higher contents. This saturation is even more clear for the visible SIA cluster density evolution shown in Fig, \ref{R20120905-0_visible_SIA.pdf}: it appears that the correct trend is reached already with only 10 appm traps and higher trap densities gives no higher SIA density, although there is a shift to higher dose for the appearance of the first visible clusters. It should be noted, however, that the fact that other experimental points than those used here as reference lie on curves corresponding to lower C content should not be ascribed to less C in those experiments. Instead, it is probably the consequence of the fact that the resolution of microscopes has been increasing over the years \cite{malerba2011review}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-c_vac_dens.pdf}
\caption{Sensitivity of the results to trap concentration: the vacancy cluster density versus dpa. The case with 100 appm traps corresponds to the carbon content in the reference experiment \cite{eldrup2002dose}.}
\label{R20120905-0_vac_dens.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-c_visible_SIA.pdf}
\caption{Sensitivity of the results to trap concentration: the visible SIA density versus
dpa. The reference experimental data are denoted with triangles
\cite{zinkle2006microstructure}. Included in the graph are also data from
other
comparable irradiation experiments in Fe-C (bullets)
\cite{singh1999effects,eyre1965electron,bryner1966study,robertson1982low,
horton1982tem,takeyama1981}. See \cite{malerba2011review} for full
details. The dotted line gives the density for one visible cluster in the simulation
box.}
\label{R20120905-0_visible_SIA.pdf}
\end{figure}
In Fig. \ref{R20130129-c_321841_vac_mean_size_evolution.pdf} and \ref{R20130129-c_sia_mean_size_evolution.pdf} the cluster mean size evolution versus dpa for vacancy and SIA clusters, respectively, are shown. The general trend for both kinds of clusters is that the mean size increase with decreased C concentration, even though the effect is very small for vacancy clusters. The
vacancy mean sizes do not differ more than 0.2 nm at most from the experimental data \cite{eldrup2002dose} for any trap concentration. The main effect of reducing the trap concentrations is that the negative curvature in the size evolution is anticipated to lower dose. For 100 appm traps, the marginal is even less than that and thus in good agreement with the experiment. For the SIA cluster sizes, it can be observed that the sizes do not change much for C concentrations above 5 appm. Above 80 appm, the mean size evolutions are almost identical. The model overestimates the SIA cluster mean size, about a factor of 5 higher than the experimental TEM data, but, on the other hand, the experimental data give surprisingly small mean sizes for the SIA clusters. Zinkle and Singh report a TEM resolution of $\sim$0.5 nm \cite{zinkle2006microstructure}, wheras 1.5 nm is a more common values and also the one we have choosen to use as our limit. Using a lower limit does not improve the results.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-c_vac_mean_size_evolution.pdf}
\caption{Sensitivity of the results to trap concentration: vacancy cluster mean size versus dpa. The reference experimental data are for $\sim$100 appm C \cite{eldrup2002dose}.}
\label{R20130129-c_321841_vac_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-c_sia_mean_size_evolution.pdf}
\caption{Sensitivity of the results to trap concentration: SIA cluster mean size versus dpa. The reference experimental data are for $\sim$100 appm C \cite{zinkle2006microstructure}.}
\label{R20130129-c_sia_mean_size_evolution.pdf}
\end{figure}
\subsection{The effect of the threshold} \label{sec:threshold}
The threshold, $N_{th}$, makes the energy with which SIA clusters are trapped
depend on the size of the trapped cluster. SIA clusters of size $5\leq N^i\leq
N_{th}$ are trapped with $E^i_t=0.6$ eV, which is the binding energy
between a C atom and an SIA cluster \cite{terentyev2011interaction}. For $N^i > N_{th}$ they are trapped with 1.2 eV, which corresponds to the strong binding energy between a CV$_2$ complex and the centre of a SIA cluster. This is summarized in Table \ref{SIA_trapping_energies}. $N_{th}$ is the only parameter of the model used for fine-tuning to fit the experimental data. It is therefore important to verify up to what extent the results depend on it.
We used the simulation set-up corresponding to \cite{zinkle2006microstructure} with 100
appm traps for SIA and vacancy clusters, respectively.
In Fig. \ref{R20120423_visible_SIA_av.pdf}, the scaling with dose of the density of
visible SIA clusters is shown for different values of $N_{th}$. For comparison, the experimental data from \cite{singh1999effects,bryner1966study,robertson1982low,
horton1982tem,takeyama1981,eyre1965electron} are also included.
The best fit to match the reference experiment from \cite{zinkle2006microstructure} is given by $N_{th}=29$, as already reported in
\cite{jansson2013simulation}. Higher values give lower densities, which remain acceptable, as compared to other experiments, up to 50. Very high values provide negligible densities, as a consequence of the fact that most clusters are very mobile and only a few remain trapped stably enough to manage to grow by absorbing the smaller ones. On the other hand, low values of $N_{th}$ shoot the number density by having the opposite effect: too many clusters remain stably trapped and act as nuclei of clusters growing above the visibility threshold. The vacancy cluster density follows the same trend with $N_{th}$, likely because of enhanced recombination with more mobile SIA clusters, while the mean vacancy cluster sizes are larger with smaller $N_{th}$ (\textit{Cf.} \ref{R20130107-a_vac_mean_size_evolution.pdf}). The mean SIA cluster sizes are larger with smaller $N_{th}$ (\textit{Cf.} \ref{R20130107-a_sia_mean_size_evolution.pdf}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130107-a_visible_SIA_av.pdf}
\caption{Number density of visible SIA versus dpa for different values of the
threshold parameter, $N_{th}$. The reference experimental data are denoted with
triangles
\cite{zinkle2006microstructure}. Included in the graph are also data from
other
comparable irradiation experiments in Fe-C (bullets)
\cite{singh1999effects,bryner1966study,robertson1982low,
horton1982tem,takeyama1981,eyre1965electron}. See \cite{malerba2011review} for full details.}
\label{R20120423_visible_SIA_av.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130107-a_vac_mean_size_evolution.pdf}
\caption{Sensitivity of the results to the threshold parameter, $N_{th}$: vacancy cluster mean size versus dpa. The experimental data are
from \cite{eldrup2002dose}.}
\label{R20130107-a_vac_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130107-a_sia_mean_size_evolution.pdf}
\caption{Sensitivity of the results to the threshold parameter, $N_{th}$: SIA cluster mean size versus dpa. The experimental data are
from \cite{zinkle2006microstructure}.}
\label{R20130107-a_sia_mean_size_evolution.pdf}
\end{figure}
\subsection{Effect of the trapping energy of large SIA sizes}\label{sec:etlarge}
We have also analyzed the sensitivity of the results to the trapping energy, $E^i_t$ for
SIA clusters above the threshold size $N_{th}=29$. In principle such energy is fixed by the type of traps (C-vacancy clusters) with which the SIA clusters interact and by where the interaction occurs (centre versus periphery). However, since the value for this interaction energy comes from atomistic simulations, it is worth verifying which values are acceptable. We thus considered different values
for the $E_t^i$ above the threshold, from 0.8 eV to 2.0 eV.
We see that higher $E^i_t$ gives higher densities both for visible SIA (\textit{Cf.} Fig. \ref{R20120905-e_visible_SIA.pdf}) and for
vacancies (\textit{Cf.} Fig. \ref{R20120905-e_vac.pdf}). For visible SIA, however, we see no significant difference when $E^i_t$ exceeds 1.2 eV. The effect on vacancy cluster density is at that point totally negligible.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-e_visible_SIA.pdf}
\caption{Sensitivity of the visible SIA number density to the value of the trapping energy above $N_{th}$. The reference experimental
data are denoted with triangles
\cite{zinkle2006microstructure}. Included in the graph are also data from
other comparable irradiation experiments in Fe-C (bullets)
\cite{singh1999effects,bryner1966study,robertson1982low,
horton1982tem,takeyama1981,eyre1965electron}. See \cite{malerba2011review} for full
details. The dotted line gives the density for one visible cluster in the simulation
box.}
\label{R20120905-e_visible_SIA.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-e_vac.pdf}
\caption{Sensitivity of the results to $E_t$ for large SIA clusters: the vacancy cluster density
versus dpa. The experimental data are from \cite{eldrup2002dose}.}
\label{R20120905-e_vac.pdf}
\end{figure}
Considering the mean cluster sizes, the evolutions for both vacancy and SIA clusters (Fig. \ref{R20130129-e_321841_vac_mean_size_evolution.pdf} and \ref{R20130129-e_sia_mean_size_evolution.pdf}, respectively) are very similar for $E_t^\delta=$ 1.0--2.0 eV and show good agreement with the experimental data (\cite{eldrup2002dose} and \cite{zinkle2006microstructure}, respectively) in the case of vacancy clusters and fair agreement for the SIA clusters. With $E_t^\delta=0.8$ eV, the trend differs considerably from the experimental data for both kinds of defects. Essentially, the effect of lowering the trapping energy is that, after the initial buildup of cluster populations, above a certain dose ($\sim$0.01 dpa), massive recombination starts occurring, thereby reducing both size and density of clusters.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-e_vac_mean_size_evolution.pdf}
\caption{Sensitivity of the result to the trapping energy above $N_{th}$: vacancy cluster mean size versus dpa. The experimental data are
from \cite{eldrup2002dose}.}
\label{R20130129-e_321841_vac_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-e_sia_mean_size_evolution.pdf}
\caption{Sensitivity of the results to the trapping energy above $N_{th}$: SIA cluster mean size versus dpa. The experimental data are
from \cite{zinkle2006microstructure}.}
\label{R20130129-e_sia_mean_size_evolution.pdf}
\end{figure}
\subsection{Study of the effect of the dose rate}\label{sec:flux}
To conclude this study, we investigated the effect of an environmental, rather than model, parameter, namely the flux, by varying the dose rate from $1.0\cdot 10^{-11}$ to $1.0\cdot 10^{-3}$ dpa/s. The former value corresponds to a typical flux on the RPV walls in commercial nuclear power plants, whereas the latter value corresponds to ion irradiation. The importance of this parameter resides in the fact that as yet it is not established whether or not one should expect significant differences between irradiation in materials test reactors as opposed to surveillance specimens irradiated during operation. The box size was here $150\times200\times250\times a_0^3$ and the temperature was 343 K. Traps for SIA were introduced with a density of 100 appm and a trapping energy of 1.0 eV. The set-up is thus similar to the experimental set-up in \cite{eldrup2002dose,zinkle2006microstructure}. The simulation was stopped after 0.23 dpa.
The results for the vacancy and visible SIA cluster density evolution are shown in
Fig.~\ref{R20111202-1_vac_dens.pdf} and
\ref{R20120520-00_SIA_dens.pdf}, respectively. It is observed that the density of
both the vacancy and the visible SIA clusters in general are lower with lower flux. The mean size evolution for vacancy and SIA clusters are shown versus dpa in Figs. \ref{R20130129-a1_sia_mean_size_evolution.pdf} and \ref{R20130129-a1_vac_mean_size_evolution.pdf}, respectively. The general trend observed is that lower dose rates increase the growth of clusters, at the expenses of density.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{R20130129-a1_vac_dens.pdf}
\caption{Effect of dose on the vacancy density evolution. The experimental
data are from \cite{eldrup2002dose} and correspond to a dose rate of $7\cdot10^{-7}$ dpa/s. The black line is the trap density.}
\label{R20111202-1_vac_dens.pdf}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{R20130129-a1_SIA_dens.pdf}
\caption{Effect of dose rate on the visible SIA cluster density evolution. The dotted line gives the density for one visible cluster in the simulation
box. The experimental data are from
\cite{zinkle2006microstructure,singh1999effects,bryner1966study,
robertson1982low,horton1982tem,takeyama1981,eyre1965electron}. See \cite{malerba2011review} for
full details.}
\label{R20120520-00_SIA_dens.pdf}
\end{center}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-a1_sia_mean_size_evolution.pdf}
\caption{Sensitivity of the results to the dose rate: SIA cluster mean size versus dpa. The experimental data are
from \cite{zinkle2006microstructure}.}
\label{R20130129-a1_sia_mean_size_evolution.pdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{R20130129-a1_vac_mean_size_evolution.pdf}
\caption{Sensitivity of the results to the dose rate: vacancy cluster mean size versus dpa. The experimental data are
from \cite{eldrup2002dose}.}
\label{R20130129-a1_vac_mean_size_evolution.pdf}
\end{figure}
\section{Discussion}\label{sec:discussion}
The effect of the trap concentration, which translates the C content, on the evolution of the density of both vacancy and visible SIA clusters, is found to be moderate. An increase of two orders of magnitude in
the trap density, from 1 appm to 100 appm, only increases the vacancy cluster density by less than one order of magnitude after 0.2 dpa (\textit{Cf}. Fig. \ref{R20120905-0_vac_dens.pdf}). Moreover, there is clearly a saturation of the effect that is quite quickly reached above 50 appm, because at that point many traps remain unused. The overall effect is even less for visible SIA clusters, where no significant change is observed when varying the trap density between 5 appm and 200 appm. To see a significant decrease in cluster density, less than 5 appm traps need to be used.
The SIA cluster mean sizes are overestimated by about half an order of magnitude, which is a general problem for our model. However, the experimental data by Zinkle and Singh \cite{zinkle2006microstructure} report suprisingly small SIA cluster sizes. Their reported TEM resolution is 0.5 nm for SIA clusters, which is much smaller than the more common value of 1.5 nm (e.g. in \cite{hernandez2010transmission} the minimum size reported in the given size distributions is 2 nm). We chose to use the latter value as the lower size limit of visible SIA clusters, as we could not see much difference in our results with either value. Mean sizes of visible clusters of 1 nm are, as a matter of fact, never reported: with the exception of this experiment, the smallest mean size found to be reported in the literature is 3--4 nm \cite{malerba2011review}. We therefore suspect that the sizes reported in \cite{zinkle2006microstructure} might be an underestimation, maybe they are in fact the smallest sizes observed. At any rate, even though the SIA cluster mean sizes are systematically too large, the trends generally agree with the experimental data in all sensitivity studies reported in this paper.
Traps replace in our model the effect of C and CV$_2$ complexes and we have used
in \cite{jansson2013simulation} 100 appm traps, corresponding to the reported
amount of carbon and nitrogen in the material of the reference experiment
\cite{zinkle2006microstructure,eldrup2002dose}. The results from Sec.
\ref{sec:C_density} show, therefore, that the exact knowledge of the amount of C in the matrix is not crucial to model the cluster density evolution: It can vary even up to 50 \% without any significant change
of the end result. So, at least at this irradiation temperature, variations in the actual C content (in the matrix) are not expected to have any strong influence, so long as the concentration is sufficiently high. Our results also show that, in order to remove the effect of C, it is necessary to reduce its concentration very significantly, down to a level of purity rarely reached in iron.
Our results in \cite{jansson2013simulation} revealed that the fundamental ingredient to reproduce experimental data correctly is that large SIA clusters should be trapped more strongly than small ones. This ingredient makes sure that larger clusters can grow by absorbing more easily de-trapped small clusters, thereby reaching sizes comparable with the experimental ones, and without in the meantime coalescing into a single large cluster. The latter situation would be produced sooner or later if only one trapping energy was used for all sizes \cite{lee2009kinetic}. In our model this size dependence is introduced using the threshold, $N_{th}$.
Physically, the idea is that large SIA clusters are more likely to interact strongly with the CV$_2$ complexes, as strong interactions only occur with the centre of the cluster. Small SIA clusters are more likely to interact via the edge, in which case the vacancies recombine and only a single C atoms interacts with the SIA cluster, which corresponds to a binding energy of 0.6 eV, independently of the size. The $N_{th}$ parameter fine-tunes these probabilities in a rather rough way, \textit{i.e.} by introducing a step-like function when one should expect a gradual transition. Nonetheless, this approximation proves satisfactory. We see in Sec. \ref{sec:threshold} that a higher value of $N_{th}$ gives a lower density of visible SIA clusters.
In \cite{jansson2013simulation}, $N_{th}=29$ was fitted to the experimental values for the experiment material in \cite{zinkle2006microstructure}. A value in this range is actually not meaningless if considering the cross-sections for the interactions of C atoms with the centre or the edge of the SIA clusters. SIA clusters are made up of dumbbells in a hexagonal configuration and SIA clusters with a certain number of SIA will form a perfectly symmetric hexagon: $N^i =$ 7, 19, 37$\ldots$ These numbers can be derived from the formula
\begin{equation}
N^i_j = 3j(j+1)+1, \quad j \in \mathbb{N}_0.
\end{equation}
The number of edge SIA would be $6j$ for a cluster with $j$ layers of SIA around the central SIA. From this formula it can be seen that a cluster with half of the SIA at the edge has a size somewhere between 19 and 37. A C atom or C-V cluster interacting with a SIA cluster with more than 37 SIA, \textit{i.e.} $N^i>37$, will be more likely to interact with the centre, than with the edge of the cluster. For the dominating C-V clusters at 343 K, this means that they are likely to be bound with a strong binding energy of $\sim$1.4 eV, according to MD calculations \cite{anento2013carbon}. For smaller clusters, the C-V complexes are most likely to interact with the edge and thus with a weak binding energy, 0.6--0.7 eV \cite{anento2013carbon}.
The actual value of the trapping energy for SIA clusters above the threshold size, $N_{th}$, is not important, as long as it is higher than $E^i_t=1.2$ eV (at the irradiation temperature of 343 K). This happens to be, according to atomistic calculation, the order of the binding energy of the CV$_2$ complexes with the centre of the SIA cluster \cite{anento2013carbon}. If the actual energy was higher, the effect would be the same because what is needed is that large SIA clusters, when trapped, must have a sufficiently small rate of de-trapping to be able to grow. Lower values, however, would not have the same effect. If the trapping energy is set too low, no visible SIA clusters form, as
happens with $E^i_t=0.8$ eV.
Finally, we have explored the effect of dose rate, which is an important parameter a priori in order to know whether the nanostructural evolution under irradiation leading to embrittlement in the vessel of a power reactor would be the same as in a high flux materials test reactor (or under ion irradiation). The results show that there is a clear dose rate effect on both the vacancy and the visible SIA cluster density evolution, at least at the temperature considered here, which is much lower than the operation temperature of RPV steels. However, a change of the dose rate by eight orders of magnitude only gives a difference of one order of magnitude for both the vacancy and the visible SIA densities, so the effect can be considered limited. It can be summarized by saying that, the lower the dose rate, the more time defects have to cluster, also by coalescence, before new defects are nucleated; or to disappear at sinks before new defects come to make them grow. The overall effect is that the density decreases, while the size increases. The gap to decide whether this will have or not an effect also on mechanical property
changes is wide. However, one can speculate that the effect on radiation hardening of the increased size with lower dose rate might be offset by decreased density of obstacles to dislocation motion, with overall limited effect on not only hardening, but also embrittlement.
Another important environmental parameter is, of course, irradiation temperature. This, however, changes also the nature of the damage, as the loops observed tend to have $\langle100\rangle$ Burgers vector with increasing temperature. This requires a revision of the model, which is the topic of another paper \cite{jansson2013nanostructure}.
\section{Conclusions}
In conclusion, the present sensitivity study shows that a model for nanostructural evolution under irradiation in iron based on traps translating the effect of C is effective and physically solid, as the variation of the key parameters lends itself to physically consistent interpretations. Moreover, the only fitting parameter used allows the reproduction of experimental results by taking physically reasonable values. Finally, the model allows an assessment of the dose-rate effect that points to a relatively weak one, although the gap between the current model and the embrittlement of RPV steels is too wide to be able to make any definitive conclusion.
\section*{Acknowledgement}
This work was carried out as part of the PERFORM60 project of the 7th Euratom
Framework Programme, partially supported by the European Commission, Grant
agreement number FP7-232612. The authors wish to thank N. Anento, C. S. Becquart,
A. De Backer, C. Domain, A. Serra, and D. Terentyev for advice, assistance and
fruitful discussions during the performance of this work.
\bibliographystyle{model1a-num-names}
|
\section{Introduction}
Compact groups of galaxies provide an excellent natural laboratory for the
study of galaxy interactions and evolution. The group environment is
conducive to such interactions, with low velocity dispersions
($\lesssim$500\ensuremath{\km \ps}), small galaxy separations, and galaxy densities
comparable to those seen in galaxy clusters. The repeated tidal
interactions between member galaxies in compact groups are likely to play a
role in a variety of processes involved in galaxy evolution, including gas
stripping, the triggering or quenching of star formation and galactic
winds, and the feeding and growth of the central supermassive black holes.
As most galaxies in the local universe reside in groups \citep{Ekeetal04}
and it is thought that many galaxies in clusters have been processed
through group environments in the past
\citep{Cappellarietal11,Mahajanetal12}, it is clearly important to
understand the role of the group environment in driving galaxy evolution .
The gas content and galaxy population of groups appears to be linked, with spiral-rich groups typically containing large reservoirs of \ensuremath{\mathrm{H}\textsc{i}}\ and other cold gas, while elliptical dominated groups often have extended halos of hot, X-ray emitting gas, with neutral hydrogen restricted to spiral galaxies in the group outskirts \citep{Kilbornetal06}. Examination of X-ray faint, spiral-rich compact groups has led to the suggestion of an evolutionary sequence, with galaxy interactions stripping the \ensuremath{\mathrm{H}\textsc{i}}\ from spiral galaxies to form intergalactic clouds and filaments or even a diffuse cold IGM \citep{VerdesMontenegroetal01,Johnsonetal07,Konstantopoulosetal10}. The redistribution of the \ensuremath{\mathrm{H}\textsc{i}}\ component is accompanied by the transformation of some member galaxies from late to early-type, and in some cases by star formation. HCG~16, also known as Arp~318, is one of the best studied examples of a compact group in the late stages of this evolutionary sequence.
HCG~16 was originally identified \citep{Hickson82} as a compact group of
four spiral galaxies, NGC~833 (HCG~16B), NGC~835 (A), NGC~838 (C) and
NGC~839 (D). Figure~\ref{fig:Xopt} includes a Digitized Sky Survey image
showing the relative positions of the group members. Later observations
increased the number of members to seven, including one large spiral
galaxy, NGC~848, of comparable luminosity to the original four
\citep{Ribeiroetal98}. A search in the NASA/IPAC Extragalactic Database
(NED\footnote{The NASA/IPAC Extragalactic Database (NED) is operated by the
Jet Propulsion Laboratory, California Institute of Technology, under
contract with the National Aeronautics and Space Administration.}) finds
three additional galaxies within 30\hbox{$^\prime$}\ and in the velocity range
3800-4100\ensuremath{\km \ps}, suggesting a more dispersed halo of dwarf galaxies
surrounding the core. All five major galaxies host AGN and/or starbursts
\citep[e.g.,][]{Martinezetal10,DeCarvalhoCoziol99}, and an ongoing or
recent interaction between NGC~833 and NGC~835 has warped their galaxy
disks and produced a dusty tidal arm extending east from NGC~835 toward
NGC~838 \citep{Konstantopoulosetal13}. Both \citet{Mulchaeyetal03} and
\citet{OsmondPonman04} find spiral fractions for the group $<$1 ($f_{\rm
sp}$=0.86, and $f_{\rm sp}$=0.83 respectively), but in both cases this
primarily reflects the classification of NGC~839 as an S0, which seems
misleading given its current high rate of star formation. We consider all
five major galaxies in the group to be late-type.
\begin{figure*}
\centerline{
\includegraphics[width=1.05\columnwidth,bb=36 175 576 616]{HCG16_5gals_DSS_HI.pdf
\includegraphics[width=1.05\columnwidth,bb=36 175 576 616]{allfive_0_5_2_smth_v2.pdf
}
\caption{\label{fig:Xopt} \textit{Left:} Digitized Sky Survey 2 (DSS2) $R$-band image of the five largest galaxies in HCG~16, with the four galaxies originally identified as a compact group to the northwest. VLA \ensuremath{\mathrm{H}\textsc{i}}\ contours from Verdes-Montenegro et al. (2014, in prep.) are overlaid, with levels N(\ensuremath{\mathrm{H}\textsc{i}})$\simeq$10,20,40,65,85,110,140,160,200,250,350,450,570$\times$10$^{-19}$\ensuremath{\cm^{-2}}. \textit{Right:} Adaptively smoothed \emph{Chandra}\ 0.5-2~keV image using data from the S3 CCD in all five observations. Contours are overlaid in red to help elucidate the distribution of diffuse emission. Dashed ellipses indicate the \ensuremath{D_{25}}\ contours of the four main galaxies. Cyan regions indicate the active sections of the S3 CCD in ObsID 923 (roughly square) and 10394 (rectangular). The two images have the same orientation, but different scales, as indicated by the scalebars.}
\end{figure*}
Neutral hydrogen mapping of the group (see Figure~\ref{fig:Xopt}) revealed a
$\sim$20\hbox{$^\prime$}\ long complex filament of cold gas surrounding the four
original members of the group and linking them to NGC~848
\citep{VerdesMontenegroetal01}, almost certainly as the result of tidal
interactions between group members. The total mass of \ensuremath{\mathrm{H}\textsc{i}}\ in the group is
$>$2.63$\times$10$^{10}$\ensuremath{\mathrm{~M_{\odot}}}, and Verdes-Montenegro et al. estimate that
the group is $<$30\% \ensuremath{\mathrm{H}\textsc{i}}-deficient. The four original member galaxies are
$\sim$50-80\% deficient, while NGC~848 is only $\sim$7\% deficient. This
suggests that the majority of the intergalactic \ensuremath{\mathrm{H}\textsc{i}}\ originated in the four
main galaxies, perhaps being transported out into the IGM by interactions
among them, and then drawn into its current morphology by a close passage
of NGC~848. \citet{Borthakuretal10} show that the \ensuremath{\mathrm{H}\textsc{i}}\ velocity
distribution covers the range $\sim$3650-4100\ensuremath{\km \ps}, confirming its
association with the major member galaxies.
HCG~16 was first detected in X-rays using the \textit{Einstein} observatory \citep{Bahcalletal84}. More detailed studies with \emph{ROSAT}\ were able to separate emission from the galaxies and diffuse inter-galactic gas \citep{Ponmanetal96,DosSantosMamon99} with the brightest emission around and between the four main galaxies. First light data from \emph{XMM-Newton}\ were used to examine the four main galaxies, providing evidence of a combination of star formation and AGN emission from NGC~833, NGC~835 and NGC~839, and classifying NGC~838 as a pure starburst \citep[][hereafter TRP01]{Turneretal01special}. A short (12.5~ks) \emph{Chandra}\ observation in cycle~1 was unable to improve on the characterization of the galaxy emission \citep{GonzalezMartinetal06} but did provide evidence of a bridge of diffuse emission linking NGC~833 and NGC~835 \citep[][see also the unpublished image previously produced from this data by Mamon \& Lima Neto\footnote{
http://dx.doi.org/10.6084/m9.figshare.971366}]{Jeltemaetal08}.
Optical spectroscopic studies of the starburst galaxies NGC~838 and NGC~839 have provided detailed characterization of these two galaxies and their outflowing galactic winds. In NGC~839 the wind has formed a biconical polar outflow with velocity $\sim$250\ensuremath{\km \ps}, while the galaxy is dominated by a rapidly rotating $\sim$400~Myr old stellar component \citep{Richetal10}. The winds appear to be shock excited, and Rich et al. argue that emission probably arises from these shocks rather than from an AGN. NGC~838 appears to have undergone a galaxy-wide starburst episode $\sim$500~Myr ago, although rapid star formation is now found only in the galaxy core \citep{Vogtetal13}. The asymmetric photoionized winds have inflated bubbles north and south of the galaxy, above and below the galactic disk, with sizes of $\sim$5~kpc and 7~kpc ($\sim$18\mbox{\arcm\arcm}\ and 25\mbox{\arcm\arcm}) for the north and south bubbles respectively, and likely ages of 5-50~Myr. Highly redshifted gas emission in the outer part of the larger southern bubble indicates that it is probably leaking material.
In this paper we use new, deep \emph{Chandra}\ observations to examine the major
galaxies of HCG~16, with the goal of studying their point source
populations, star formation and nuclear activity. A study of the diffuse
emission in the group, and its relation to the major galaxies is presented
in \citet[hereafter paper II]{OSullivanetal14c_special}. We adopt a Galactic
hydrogen column density of \NH=2.56$\times$10$^{20}$\ensuremath{\cm^{-2}}\ for the four
original group member galaxies and the surrounding diffuse emission
\citep[taken from the Leiden/Argentine/Bonn survey,][]{Kalberlaetal05}. For
NGC~848 we adopt a hydrogen column of \NH=2.75$\times$10$^{20}$\ensuremath{\cm^{-2}}. All
fluxes and luminosities are corrected for Galactic absorption. We adopt a
redshift of $z$=0.0132 for the group \citep{Hicksonetal92}. A
redshift-independent distance measurement is available for one of the five
major galaxies, a Tully-Fisher distance of 56.5~Mpc for NGC~848
\citep{Theureauetal07}. This is consistent, within errors, with
redshift-based estimates for all five of the galaxies, correcting for
infall toward the Virgo cluster, great attractor and Shapley Supercluster,
for a cosmology with $H_0$=70\ensuremath{\km \ps \pMpc\,}. We therefore adopt this distance
estimate for the group as a whole, which gives an angular scale of
1\mbox{\arcm\arcm}=273~pc.
\section{Observations and Data Analysis}
\subsection{Chandra}
\label{sec:obs}
\begin{deluxetable*}{lccccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:obs}Summary of \emph{Chandra}\ observations of HCG~16}
\tablehead{
\colhead{ObsID} & \colhead{P.I.} & \colhead{Observation date} & \colhead{mode} & \colhead{Frame time} & \colhead{subarray} & \colhead{Roll angle} & \colhead{Cleaned exposure} \\
\colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{(s)} & \colhead{} & \colhead{(\hbox{$^\circ$})} & \colhead{(s)}
}
\startdata
\dataset[ADS/Sa.CXO#obs/923]{923} & Mamon & 2000 Nov 16 & F & 3.2 & N & 325.98 & 12565 \\
\dataset[ADS/Sa.CXO#obs/10394]{10394} & Alexander & 2008 Nov 23 & VF & 1.5 & 1/2 & 319.47 & 13824 \\
\dataset[ADS/Sa.CXO#obs/15181]{15181} & Vrtilek & 2013 Jul 16 & VF & 3.1 & N & 111.35 & 49457 \\
\dataset[ADS/Sa.CXO#obs/15666]{15666} & Vrtilek & 2013 Jul 18 & VF & 3.1 & N & 111.35 & 29714 \\
\dataset[ADS/Sa.CXO#obs/15667]{15667} & Vrtilek & 2013 Jul 21 & VF & 3.1 & N & 111.35 & 58335
\enddata
\end{deluxetable*}
HCG~16 has been observed five times by the \emph{Chandra}\ ACIS instrument,
briefly in cycles~1 and 10, and most recently in cycle~14 for a total of
137.5~ks. Table~\ref{tab:obs} summarises the observational setup of the
five exposures. A summary of the \emph{Chandra}\ mission and instrumentation
can be found in \citet{Weisskopfetal02}. In all five observations the S3
CCD was placed at the focus of the telescope. ObsID~10394 was performed
using a 1/2 subarray, with a short (1.5~s) frame time, and with only the
ACIS-S3 CCD switched on. The three observations in cycle~14 were performed
with the same roll angle. The four original group members are located on
the S3 CCD in four of the five observations, but ObsID~10394 covers only NGC~835 and NGC~838. NGC~848 is only visible on the S1 CCD of ObsID~923.
We have reduced the data from all five pointings using CIAO 4.6.1
\citep{Fruscioneetal06} and CALDB 4.5.9 following techniques similar to
those described in \citet{OSullivanetal07} and the \emph{Chandra}\ analysis
threads\footnote{http://asc.harvard.edu/ciao/threads/index.html}. The
level 1 event files were reprocessed, bad pixels and events with \emph{ASCA}\
grades 1, 5 and 7 were removed, and the cosmic ray afterglow correction was
applied. Very Faint mode cleaning was applied to all observations except
ObsID~923. The data were corrected to the appropriate gain map, the
standard time-dependent gain and charge-transfer inefficiency (CTI)
corrections were made, and background light curves were produced. None of
the observations suffered from significant background flaring. A comparison
of 0.5-0.7~keV and 2.5-5~keV light curves shows no significant difference
between the two bands, indicating that the observations are not affected by
solar wind charge exchange emission. In general all five observations were
combined for imaging analysis, but spectra were extracted separately and
fitted simultaneously.
Point source identification was performed using the \textsc{ciao} task
\textsc{wavdetect}, with a detection threshold of 10$^{-6}$, chosen to
ensure that the task detects $\lesssim$1 false source in the S3 field of
view, working from a 0.3-7.0 keV image and exposure map from the five
observations combined. We defined source apertures based on the position of
each source as reported by \textsc{wavdetect}. In the case of confused or
extended sources we adjusted these centroids based on the peak of the
emission in the 2-7~keV band. Ellipticities and position angles from
\textsc{wavdetect} were retained, but the axes of each source region were
reduced to approximately the 90\% encircled energy radius, with a minimum
radius of 1\mbox{\arcm\arcm}. For sources outside the bodies of the galaxies we used
the regions reported by \textsc{wavdetect}, which satisfied the $>$90\%
encircled energy limit. The resulting regions were used to examine source
properties, and to exclude them from spectral fits to the diffuse emission.
Spectra were extracted from each dataset using the \textsc{specextract}
task. When examining diffuse emission, background spectra were drawn from
blank-sky event lists, scaled to match the data in the 9.5-12.0 keV band.
For point sources, local background spectra were used, typically from an
elliptical annulus with radius 1.5-3 times that of the source extraction
region. Where the source was partially surrounded by diffuse emission (as
in NGC~838), background regions were chosen to include a representative
fraction of that diffuse emission. Spectral fitting was performed in XSPEC
12.8.1. Abundances were measured relative to the abundance ratios of
\citet{GrevesseSauval98}. 1$\sigma$ uncertainties for one interesting
parameter are reported for all fitted values.
\subsection{Very Large Array and Giant Metrewave Radio Telescope}
\begin{deluxetable*}{lcccccccc}[!t]
\tablewidth{0pt}
\tablecaption{\label{tab:Robs}Summary of radio observations of HCG~16}
\tablehead{
\colhead{Observatory} & \colhead{project} & \colhead{Observation date}& \colhead{Frequency} & \colhead{Bandwidth} & \colhead{Integration time} & \colhead{HPBW} & \colhead{Pos. Angle} & \colhead{rms} \\
\colhead{} & \colhead{} & \colhead{} & \colhead{(MHz)} & \colhead{(MHz)} & \colhead{(hr)} & \colhead{(\mbox{\arcm\arcm}$\times$\mbox{\arcm\arcm})} & \colhead{(\hbox{$^\circ$})} & \colhead{($\mu$Jy beam$^{-1}$)}
}
\startdata
VLA & AW500 & 1999 Jan 13-14 & 1402 & 6.3 & 5 & 25.0$\times$18.1 & -6.64 & 200 \\
GMRT & 17\_026 & 2009 Nov 21 & 610 & 32 & 6 & 5.6$\times$5.4 & -3.0 & 60
\enddata
\end{deluxetable*}
We analysed 1.4~GHz data retrieved from the \textit{Very Large Array} (VLA) public
archive. The observations were performed in C array, in spectral line mode
with 64 channels centred at 1402~MHz for a total on-source exposure of
$\sim$5~hrs. Further details of the observation are shown in
Table~\ref{tab:Robs}. We used the NRAO Astronomical Image Processing System
(\textsc{aips}) package for the data reduction and analysis. Data
calibration and imaging were carried out following the standard procedure
(Fourier transform, clean and restore). Phase-only self-calibration was
applied to remove residual phase variations and improve the quality of the
image. The final image has an angular resolution of 25$\times$18.1\mbox{\arcm\arcm}\
and an rms noise level (1$\sigma$) of 0.2~mJy beam$^{-1}$.
We also observed the group with the \textit{Giant Metrewave Radio Telescope} (GMRT) in dual-frequency 610/235~MHz mode (project 17$\_$026, P.I. E. O'Sullivan). The observations were again analysed using \textsc{aips}, with phase self-calibration, following the methods described in \citet{Giacintuccietal11}. The 235~MHz data suffer from calibration problems, and we therefore excluded them from further analysis. The 610~MHz data are essentially free of radio frequency interference (RFI), but are dynamic range limited owing to a powerful nearby source whose side lobes affect the region of the group. This problem becomes more severe in lower resolutions, and we therefore use only the full resolution (5.6$\times$5.4\mbox{\arcm\arcm}) image which has an rms noise level (1$\sigma$) of 0.06~mJy~beam$^{-1}$. Further details of the 610~MHz observation are shown in Table~\ref{tab:Robs}.
\section{Point Sources}
\label{sec:PS}
Given its distance, we only expect to be able to detect relatively bright
point sources in HCG~16. We use the \emph{Chandra}\ Portable Interacting
Multi-Mission Simulator \citep[PIMMS;][]{Mukai93} to estimate the detection
limit of the combined 2013 observations, requiring a minimum of 10 counts
(0.5-7~keV) for detection and adopting a powerlaw model with $\Gamma$=1.7
and Galactic absorption, and find a limit of
$L_{0.5-7}\geq$2.1$\times$10$^{38}$\ensuremath{\erg \ps}\
($F_{0.5-7}\geq$5.5$\times$10$^{-16}$\ensuremath{\erg \ps \pcmsq}). This is only about a
factor of five below our adopted threshold for ultra--luminous X--ray point
sources (ULXs), $L_{0.5-8}\geq$10$^{39}$\ensuremath{\erg \ps}\
\citep[e.g.,][]{Swartzetal04}. Since these are star forming galaxies, we
expect most sources in the galaxies (excluding AGN) to be high mass X-ray
binaries (HMXBs), formed from massive, short-lived stars.
Of the 30 point sources identified in the ACIS-S3 field of view, 18 fall
within the \ensuremath{D_{25}}\ ellipses of the four major galaxies, or in the tidal
structures around NGC~833 and NGC~835. Table~\ref{tab:sources} summarises
the locations of these 18 sources, the regions used to measure their
fluxes, and the number of net counts in each source in a co-added image of
all five observations. As expected, many of the sources in the starburst
galaxy NGC~838 are clustered in the galaxy core, where star formation is
ongoing. The other three galaxies contain only 1-4 sources each, and there
is no clear correlation between source position and galaxy structure. We
initially extracted spectra for each source, and estimated their background
subtracted count rates using the \textsc{dmextract} task. The three sources
with the highest fluxes ($>$700 net 0.5-7~keV counts) are associated with
the nuclei of NGC~833, NGC~835 and NGC~839 (HCG~16B, A, and D). Detailed
fits to these sources and their surrounding diffuse emission will be
described in \textsection~\ref{sec:AGNdiff}. A further four sources had a
sufficient number of counts in individual exposures to allow fitting of
simple absorbed powerlaw and absorbed APEC thermal models, to help
determine their origin. These fits are described later in
\textsection~\ref{sec:PSfits}.
The remaining eleven sources were too faint for individual spectral
fitting. In order to estimate luminosities from their count rates we
required a conversion factor, including corrections for Galactic absorption
and for the fraction of flux scattered outside the extraction region by the
\emph{Chandra}\ point spread function (PSF). As the responses calculated for each
source during spectral extraction include a correction for PSF scattering,
this conversion factor can be found by folding a standard source spectral
model through each set of responses and determining the expected flux for a
fixed count rate. Extragalactic X-ray binaries are commonly modelled using
spectrally hard models such as a $\Gamma$=1.7 powerlaw
\citep[e.g.,][]{Smithetal12}. To test the suitability of such a model we
co-added the spectra and responses of the eleven faint sources using the
\textsc{combine\_spectra} task to create three stacked spectra, one each
for ObsID 923, 10394, and the 2013 observations (15181, 15666 and 15667).
These spectra were then simultaneously fitted using an absorbed powerlaw
model with hydrogen column fixed at the Galactic value. The best fit has
$\Gamma$=1.69$\pm$0.14, in agreement with expectations, and reduced
$\chi^2$=1.245 for 22 degrees of freedom. We therefore adopt the
$\Gamma$=1.7 powerlaw to calculate the counts-to-flux conversion factors
for each source in each observation. The resulting fluxes are listed in
Table~\ref{tab:variable}. Where sources are detected at less than 3$\sigma$
significance, we calculate an upper limit based on the integer number of
counts required to produce a 3$\sigma$ detection, taking into account the
scaled local background.
We also estimate the mean flux in the three 2013 observations combined;
individual sources can be detected in a stacked image yet be undetected in
each individual observation, and combining the three most recent
observations allows us to estimate a flux without trying to combine counts
from the earlier observations which have very different conversion factors.
Three of the sources show evidence of variability between observations at
$>$3$\sigma$ significance. Figure~\ref{fig:var} shows the
0.5-7~keV fluxes for these sources in each observation. We note
that when classifying sources as potential ULXs, we estimate a factor of
1.075 increase in flux between the 0.5-7~keV and 0.5-8~keV bands, so
sources with $L_{0.5-7}$$\geq$9.32$\times$10$^{38}$\ensuremath{\erg \ps}\ are bright
enough to be ULXs if they are located within HCG~16.
\begin{deluxetable*}{lcccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:sources}List of point sources and their properties}
\tablehead{
\colhead{Source} & \colhead{R.A.} & \colhead{Dec.} & \colhead{Radii$^a$} & \colhead{p.a.$^a$} & \colhead{net counts$^b$} & \colhead{Notes} \\
\colhead{} & \colhead{(J2000)} & \colhead{(J2000)} & \colhead{(\mbox{\arcm\arcm})} & \colhead{(\hbox{$^\circ$})} & \colhead{(cnt.~s$^{-1}$)} & \colhead{}
}
\startdata
\multicolumn{7}{l}{\textit{NGC~839}} \\
1 & 2$^{\text{h}}$09${\text{m}}$42.754$^{\text{s}}$ & -10$^\circ$11\hbox{$^\prime$} 02.40\mbox{\arcm\arcm} & 2.09,1.65 & 24.7 & 735.3$\pm$29.4 & nucleus\\
\hline
\multicolumn{7}{l}{\textit{NGC~838}} \\
2 & 2$^{\text{h}}$09${\text{m}}$37.062$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 55.91\mbox{\arcm\arcm} & 2.19,1.00 & 50.3 & 27.8$\pm$5.5 & \\
3 & 2$^{\text{h}}$09${\text{m}}$38.531$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 48.26\mbox{\arcm\arcm} & 1.45,1.30 & 139.9 & 274.0$\pm$24.3 & ULX or nuclear source\\
4 & 2$^{\text{h}}$09${\text{m}}$37.984$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 48.83\mbox{\arcm\arcm} & 1.61,1.00 & 49.2 & 93.3$\pm$11.0 & ULX\\
5 & 2$^{\text{h}}$09${\text{m}}$38.128$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 45.71\mbox{\arcm\arcm} & 1.19,1.14 & 155.5 & 24.4$\pm$12.5 & ULX, 1.44\mbox{\arcm\arcm}\ from CXO J020938.1-100847\\
6 & 2$^{\text{h}}$09${\text{m}}$38.449$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 44.87\mbox{\arcm\arcm} & 1.27,1.20 & 151.3 & 101.4$\pm$19.0 & probably thermal, 1.75\mbox{\arcm\arcm}\ from SN 2005H \\
7 & 2$^{\text{h}}$09${\text{m}}$39.237$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 44.41\mbox{\arcm\arcm} & 1.57,1.31 & 15.4 & 36.2$\pm$7.4 & \\
8 & 2$^{\text{h}}$09${\text{m}}$38.340$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 34.93\mbox{\arcm\arcm} & 1.32,1.10 & 25.0 & 21.0$\pm$5.7 & \\
9 & 2$^{\text{h}}$09${\text{m}}$40.591$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 26.72\mbox{\arcm\arcm} & 1.12,1.00 & 24.8 & 15.5$\pm$4.4 & \\
11 & 2$^{\text{h}}$09${\text{m}}$38.118$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 19.64\mbox{\arcm\arcm} & 1.26,1.05 & 140.4 & 34.3$\pm$6.1 & possible background AGN in SDSS~J020938.10-100819.4\\
17 & 2$^{\text{h}}$09${\text{m}}$37.962$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 40.53\mbox{\arcm\arcm} & 1.72,1.20 & 147.4 & 58.8$\pm$9.6 & possibly thermal, CXO J020937.9-100840 \\
\hline
\multicolumn{7}{l}{\textit{NGC~833 / NGC~835 complex}} \\
10 & 2$^{\text{h}}$09${\text{m}}$22.900$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 24.93\mbox{\arcm\arcm} & 1.45,1.45 & 174.6 & 116.6$\pm$11.1 & ULX in tidal zone, CXO J020922.8-100824 \\
12 & 2$^{\text{h}}$09${\text{m}}$24.613$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 09.50\mbox{\arcm\arcm} & 1.61,1.70 & 82.3 & 2361.3$\pm$50.7 & NGC~835 nucleus\\
13 & 2$^{\text{h}}$09${\text{m}}$21.099$^{\text{s}}$ & -10$^\circ$08\hbox{$^\prime$} 03.15\mbox{\arcm\arcm} & 1.42,1.13 & 154.8 & 33.4$\pm$6.9 & ULX\\
14 & 2$^{\text{h}}$09${\text{m}}$20.861$^{\text{s}}$ & -10$^\circ$07\hbox{$^\prime$} 59.41\mbox{\arcm\arcm} & 1.53,1.34 & 86.7 & 718.4$\pm$27.7 & NGC~833 nucleus\\
15 & 2$^{\text{h}}$09${\text{m}}$20.427$^{\text{s}}$ & -10$^\circ$07\hbox{$^\prime$} 48.08\mbox{\arcm\arcm} & 1.64,1.20 & 37.0 & 17.9$\pm$4.5 & \\
16 & 2$^{\text{h}}$09${\text{m}}$27.578$^{\text{s}}$ & -10$^\circ$07\hbox{$^\prime$} 46.72\mbox{\arcm\arcm} & 1.39,1.51 & 166.1 & 421.1$\pm$20.8 & possible background AGN, CXO J020927.6-100746 \\
& & & & & & in SDSS~J020927.57-100746.2 \\
18 & 2$^{\text{h}}$09${\text{m}}$20.773$^{\text{s}}$ & -10$^\circ$07\hbox{$^\prime$} 46.18\mbox{\arcm\arcm} & 2.44,1.00 & 178.1 & 11.5$\pm$3.9 &
\enddata
\tablecomments{
$^a$ Radii and position angles of elliptical regions used to extract spectra and numbers of counts. $^b$ Counts in 0.5-7~keV band, summed over all observations with source in field of view.
}
\end{deluxetable*}
\begin{deluxetable*}{lccccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:variable}Point source fluxes in each observation and the stacked 2013 observations}
\tablehead{
\colhead{Source} & \multicolumn{6}{c}{L$^{cts}_{0.5-7}$ (10$^{38}$\ensuremath{\erg \ps} )} & \colhead{Variable?} \\
\colhead{} & \colhead{923} & \colhead{10394} & \colhead{15181} & \colhead{15666} & \colhead{15667} & \colhead{15XXX$^a$} & \colhead{} \\
}
\startdata
1 & 236.13$\pm$25.20 & - & 316.83$\pm$23.22 & 286.23$\pm$29.56 & 303.88$\pm$18.86 & 311.05$\pm$13.40 & N \\
2 & $<$33.12 & - & $<$8.10 & $<$17.42 & 13.52$\pm$3.55 & 8.28$\pm$1.76 & - \\
3 & $<$75.05 & 134.64$\pm$20.81 & 79.76$\pm$11.97 & 48.64$\pm$12.93 & 53.11$\pm$9.67 & 61.57$\pm$6.57 & $>$3$\sigma$\\
4 & $<$44.12 & $<$24.71 & 23.94$\pm$5.55 & $<$20.89 & 34.78$\pm$5.85 & 27.14$\pm$3.47 & N \\
5 & $<$58.25 & $<$70.62 & $<$33.76 & $<$4.43 & $<$29.87 & 10.91$\pm$6.18 & - \\
6 & $<$73.49 & $<$47.13 & $<$23.03 & $<$28.68 & $<$20.85 & 22.48$\pm$4.92 & - \\
7 & $<$48.51 & $<$24.75 & $<$10.12 & $<$15.39 & $<$9.01 & 5.41$\pm$1.65 & - \\
8 & $<$54.16 & $<$48.78 & $<$9.78 & $<$14.90 & $<$8.25 & 5.69$\pm$1.54 & - \\
9 & $<$78.22 & $<$22.61 & $<$10.47 & $<$17.42 & $<$9.74 & 4.70$\pm$1.47 & - \\
10 & $<$69.97 & - & 27.11$\pm$4.53 & 28.15$\pm$5.91 & 30.48$\pm$4.43 & 28.77$\pm$2.80 & N \\
11 & $<$48.01 & $<$23.76 & $<$7.98 & $<$16.50 & $<$6.76 & 7.71$\pm$1.60 & - \\
12 & 139.48$\pm$37.98 & 95.77$\pm$24.07 & 628.75$\pm$22.00 & 601.34$\pm$27.64 & 567.72$\pm$19.27 & 597.10$\pm$12.85 & $>$16$\sigma$\\
13 & $<$139.53 & - & $<$10.12 & 16.84$\pm$5.29 & 9.34$\pm$2.87 & 9.72$\pm$1.95 & N \\
14 & 302.29$\pm$69.55 & - & 158.36$\pm$11.11 & 203.09$\pm$16.13 & 196.44$\pm$11.36 & 184.16$\pm$7.16 & N \\
15 & $<$99.99 & - & $<$7.75 & $<$12.91 & $<$6.61 & 4.81$\pm$1.23 & - \\
16 & 101.76$\pm$30.23 & 54.63$\pm$13.57 & 113.51$\pm$9.09 & 58.91$\pm$8.40 & 108.12$\pm$8.07 & 99.29$\pm$5.06 & $>$3$\sigma$\\
17 & $<$42.13 & $<$29.22 & 17.56$\pm$4.50 & $<$19.56 & 11.78$\pm$3.64 & 14.28$\pm$2.58 & N \\
18 & $<$90.41 & - & $<$8.67 & $<$15.13 & $<$7.10 & 0.54$\pm$0.26 & -
\enddata
\tablecomments{Upper limits are shown for sources detected at $<$3$\sigma$ significance in a particular observation. $^a$: Column 7 shows the mean luminosity in the 2013 observations 15181, 15666 and 15667 co-added.}
\end{deluxetable*}
\begin{figure*}
\includegraphics[width=\textwidth,bb=30 550 570 760]{variability_plots_v2.pdf}
\caption{\label{fig:var}Variability plots for those sources whose 0.5-7~keV fluxes appear to change over time, with observations in chronological order from ObsID 923 (bin 1) to 15667 (bin 5). 1$\sigma$ error bounds on the counts-based fluxes from the five observations, or 3$\sigma$ upper limits are shown for each source, with the mean flux in the 2013 observations marked by a dashed line. Each plot is labelled with the source number.}
\end{figure*}
\subsection{Notes on individual point sources}
\label{sec:PSfits}
For sources 3, 6, 10 and 16 we are able to fit simple absorbed thermal plasma or powerlaw models. Spectra from datasets 923 and 10394 were not included in cases where they contain only a handful of counts for each source. The results of these fits are shown in Table~\ref{tab:bright}.
\begin{deluxetable*}{lccccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:bright}Spectral fits to the four point sources with highest fluxes not already known to be AGN}
\tablehead{
\colhead{Source} & \multicolumn{4}{c}{Thermal model} & \multicolumn{3}{c}{Powerlaw model} \\
\colhead{} & \colhead{kT} & \colhead{Abund.} & \colhead{$L_{0.5-7}$} & \colhead{red. $\chi^2$/d.o.f$^a$} & \colhead{$\Gamma$} & \colhead{$L_{0.5-7}$} & \colhead{red. $\chi^2$/d.o.f$^a$} \\
\colhead{} & \colhead{(keV)} & \colhead{(\ensuremath{\mathrm{~Z_{\odot}}})} & \colhead{(10$^{38}$\ensuremath{\erg \ps})} & \colhead{} & \colhead{} & \colhead{(10$^{38}$\ensuremath{\erg \ps})} & \colhead{}
}
\startdata
3 & $>$30.98 & 0.3$^b$ & 90.37$^{+6.61}_{-11.29}$ & 1.859/17 & 0.86$\pm$0.23 & 105.88$^{+14.63}_{-13.71}$ & 1.688/17 \\[+0.5mm]
6 & 1.44$^{+0.89}_{-0.20}$ & 0.3$^b$ & 15.39$^{+5.92}_{-4.32}$ & 1.63/4 & 2.85$^{+1.41}_{-1.54}$ & 20.66$^{+8.59}_{-5.81}$ & 2.05/4 \\[+0.5mm]
10 & 3.48$^{+2.73}_{-1.16}$ & 0.3$^b$ & 31.93$^{+4.24}_{4.16}$ & 0.532/6 & 2.00$\pm$0.30 & 34.60$^{+4.66}_{-4.58}$ & 0.60/6 \\[+0.5mm]
16 & 4.27$^{+1.30}_{-0.84}$ & 0.3$^b$ & 97.55$^{+7.83}_{-7.49}$ & 0.595/13 & 1.80$^{+0.12}_{-0.11}$ & 104.31$^{+7.03}_{-6.99}$ & 0.581/13
\enddata
\tablecomments{$^a$The reduced $\chi^2$ of the fit over its degrees of freedom. $^b$ Parameter fixed during fitting.}
\end{deluxetable*}
Sources 3 and 6 are located close together in the core of NGC~838. Source 6 is best fitted by a thermal plasma model with temperature $\sim$1.4~keV, somewhat hotter than the surrounding diffuse thermal emission, but not unphysically so. It has a relatively steep spectral index when fitted with a powerlaw. Given that NGC~838 hosts an extended starburst wind with a good deal of clumpiness in its X--ray structure, it seems likely that this source is in fact a cloud of hot gas.
By contrast, in source 3 the thermal model has an unconstrained high temperature while the powerlaw model has an index $\Gamma$=0.86$\pm$0.23. ULXs are typically observed to have spectral indices of $\Gamma\sim1.9$ \citep{Swartzetal04}, with only a very small fraction having indices as flat as source 3. The source is variable at $>$3$\sigma$ significance, and is located $<$1\mbox{\arcm\arcm}\ from the optical centroid of the galaxy. The source luminosity is $L_{0.5-7}$=[9.04$^{+0.66}_{-1.13}$]$\times$10$^{39}$\ensuremath{\erg \ps}, and it therefore seems likely that this is the previously undetected active nucleus of NGC~838.
Source 16 is located in the tidal arm extending east from NGC~835. However, the X--ray centroid matches the position of SDSS~J020927.57-100746.2, a faint ($g$=21.8) galaxy with no measured redshift but which is probably unassociated with the group. The source was detected in previous \emph{Chandra}\ observations \citep[CXO~J020927.6-100746,][]{Evansetal10}. The spectrum is noisy, and cannot distinguish between a $\sim$4~keV thermal plasma or a $\Gamma\sim$1.8 powerlaw. We conclude that this source is probably a background AGN unrelated to HCG~16.
Source 10 is located in the tidally disturbed structure on the southern edge of NGC~835 and NGC~833 and is bright and stable enough to have been previously detected as source CXO J020922.8-100824 \citep{Evansetal10}. The spectra are not sufficient to distinguish between a thermal or powerlaw origin for the emission, but it seems plausible that the source is either a ULX in the tidal zone, or a background AGN.
Of the fainter sources, three (4, 5 and 13) are bright enough to be ULXs if they are truly point sources located in HCG~16. Source 5 corresponds to another peak in diffuse emission east of the core of NGC~838, and thus may be a gas clump like source 6. Source 4 is at the western edge of the diffuse emission of NGC~838 and is clearly visible in the 2-7~keV band, suggesting that it is a spectrally hard source. Source 13 is located just southwest of the core of NGC~833, close to the position of some unclassified optical sources visible in HST WFPC2 $V$-band imaging. Source 11 is located $<$0.5\mbox{\arcm\arcm}\ from SDSS J020938.10-100819.4 a disk galaxy north of NGC~838 with no measured spectroscopic redshift, and a photometric redshift of $\sim$0.2-0.25.
Based on the cumulative luminosity function of background sources measured by the \emph{Chandra}\ Multiwavelength Project \citep[ChaMP,][]{Kimetal07a} we are able to estimate the number of background AGN we would expect to find within the \ensuremath{D_{25}}\ ellipses and tidal interaction regions of the four major galaxies. The \ensuremath{D_{25}}\ ellipses enclose an area of $\sim$3.05 square arcminutes, with the tidal regions adding $\sim$10\% to this value. Based on the sensitivity limit of the stacked 2013 observations, we would expect to see 2.4 background AGN within the area of the galaxies in the 0.5-8~keV band. Our true sensitivity is probably somewhat better than this, since we include the two earlier observations when detecting sources, but the number of expected sources will also be somewhat lower since we use a narrower energy band (0.5-7~keV) than the ChaMP team. However, since both effects will be relatively small, we still expect to find 2-3 background sources in our regions of interest. This agrees well with our findings that sources 7 and 11 are probably associated with background galaxies.
\section{Galaxy and AGN emission}
\label{sec:AGNdiff}
To examine the radio emission from each galaxy, we extracted total flux densities in each band from regions defined based on the 3$\sigma$ detection contours. These 610~MHz and 1.4~GHz flux densities are listed in Table~\ref{tab:Rflux}. From these fluxes we calculate a spectral index $\alpha$ for each galaxy, defined as $S_\nu\propto\nu^{-\alpha}$, $S_\nu$ being the flux density at frequency $\nu$. We also calculated the radio power at each frequency, where the power $P_\nu$ is defined as $P_\nu$=4$\pi$D$^2$(1+$z$)$^{\alpha-1}S_\nu$, where D is the distance and $z$ the redshift of the group. The 610~MHz data have sufficient spatial resolution to allow us to separate core and diffuse components of the radio emission for NGC~835 and NGC~838, and to suggest that some of the extended emission in NGC~848 may be associated with a background galaxy. We therefore report separate 610~MHz flux densities for the different components of these galaxies.
\begin{deluxetable*}{llccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:Rflux}Radio fluxes, spectral indices and powers for the five major galaxies}
\tablehead{
\colhead{Galaxy} & \colhead{} & \colhead{$S_{610}$} & \colhead{$S_{1400}$} & \colhead{$\alpha$} & \colhead{$P_{610}$} & \colhead{$P_{1400}$} \\
\colhead{} & \colhead{} & \colhead{(mJy)} & \colhead{(mJy)} & \colhead{} & \colhead{(10$^{21}$~W~Hz$^{-1}$)} & \colhead{(10$^{21}$~W~Hz$^{-1}$)}\\
}
\startdata
NGC~833 & total & 6.0$\pm$0.3 & 4.1$\pm$0.2 & 0.5$\pm$0.1 & 2.3$\pm$0.1 & 1.6$\pm$0.1 \\
NGC~835 & total & 90.9$\pm$4.6 & 47.6$\pm$2.4 & 0.8$\pm$0.1 & 34.6$\pm$1.8 & 18.1$\pm$0.9 \\
& core & 75.1$\pm$3.8 & - & - & - & - \\
& extended & 15.8$\pm$0.8 & - & - & - & - \\
NGC~838 & total & 175.2$\pm$8.8 & 93.1$\pm$4.7 & 0.8$\pm$0.1 & 66.7$\pm$3.4 & 35.5$\pm$1.8 \\
& core & 151.7$\pm$7.6 & - & - & - & - \\
& extended & 23.5$\pm$1.2 & - & - & - & - \\
NGC~839 & total & 56.3$\pm$2.8 & 37.5$\pm$1.9 & 0.5$\pm$0.1 & 21.4$\pm$1.1 & 14.2$\pm$0.7 \\
NGC~848 & total & 18.6$\pm$0.9 & 10.9$\pm$0.5 & 0.6$\pm$0.1 & 7.1$\pm$0.3 & 4.1$\pm$0.2 \\
& no extended & 12.0$\pm$0.6 & - & - & - & -
\enddata
\end{deluxetable*}
Moving on to the X-ray properties of the major galaxies, we initially extracted spectra from relatively large regions around each of them. Point sources within these regions, except those corresponding to the nuclear sources in NGC~833, NGC~835 and NGC~839, were not excluded, so as to allow an accurate measurement of the total hard X-ray flux from the stellar populations of the galaxies. With the exception of NGC~848, where the shallow off-axis observation contains only a small number of counts, we determined the size of the extraction regions by examining the cumulative growth in the number of counts with radius from the centre of each galaxy, and selected regions which contain 95\% of the background subtracted flux. In the case of NGC~833 and 835 there is a clear region of diffuse emission between the two galaxies, and in the eastern tidal arm extending from NGC~835. We therefore used 15\mbox{\arcm\arcm}\ radius circles to extract spectra of the AGN and galaxy cores, and a larger polygonal region to examine the diffuse component.
The galaxy spectra were fitted with models consisting of APEC thermal
plasma and powerlaw components, folded through an absorber whose column was
fixed at the Galactic value. In the three galaxies with bright nuclear
sources, a model of the galaxy emission with the AGN region (as defined in
Table~\ref{tab:sources}) removed was initially fitted. The AGN region was
then reintroduced, and the galaxy model allowed to vary only in overall
normalization, to account for the increased emission volume. Additional
components were then added to model any AGN. Previous studies suggest that
all three galaxies host partially absorbed AGN, so these were modelled
using a powerlaw observed through a partial-covering absorber at the
redshift of the galaxy. Although the partial covering fraction was allowed
to fit in each case, we found that the best fit value was always 1.0. This
suggests that all three AGN are in fact fully covered by their absorbers,
and any unabsorbed powerlaw component probably arises from the X--ray
binary population in the host galaxy. We outline the results of the fits to
each galaxy below and best-fitting model parameters are shown in
Table~\ref{tab:AGN}.
\begin{deluxetable*}{llcccccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:AGN}Best-fitting spectral model parameters for the five major galaxies}
\tablehead{
\colhead{Component} & \colhead{Parameter} & \colhead{NGC~833} & \colhead{NGC~835} & \multicolumn{2}{c}{NGC~838} & \multicolumn{2}{c}{NGC~839} & \colhead{NGC~848}
}
\startdata
Model & & AP+PL+abs.PL & AP+PL+abs.PL & AP+PL & AP+AP & AP+PL+abs.PL & AP+abs.PL & PL\\
\hline
\multicolumn{10}{l}{\textit{Galaxy emission}}\\
Soft & kT$^a$ & 0.47$^{+0.14}_{-0.10}$ & 0.61$\pm$0.03 & 0.80$\pm$0.02 & 0.78$\pm$0.02 & 0.88$\pm$0.05 & 0.79$^{+0.03}_{_-0.04}$ & - \\[+.5mm]
Thermal & Z$^b$ & 0.47$^{+0.47}_{-0.37}$ & 0.22$^{+0.15}_{-0.07}$ & 0.16$^{+0.08}_{-0.04}$ & 0.16$^{+0.05}_{-0.03}$ & 0.3$^f$ & 0.3$^f$ & - \\[+.5mm]
& L$^c_{0.5-7}$ & 41.25$^{+8.02}_{-6.88}$ & 200.91$^{+17.95}_{-17.57}$ & 415.56$^{+51.56}_{-64.17}$& 420.15$^{+45.45}_{-37.43}$& 92.81$\pm$11.84 & 106.56$^{+12.60}_{-13.75}$ & - \\[+.5mm]
Hard & \NH$^d$ & 0.0$^f$ & 0.0$^f$ & 0.0$^f$ & 0.0$^f$ & 0.0$^f$ & 0.47$^{+0.24}_{-0.18}$ & 0.0$^f$\\[+0.5mm]
& $\Gamma$ & 1.65$^f$ & 1.65$^f$ & 1.84$^{+0.18}_{-0.19}$ & kT$^a$=4.27$^{+2.28}_{-0.85}$ & 1.46$^{+0.15}_{-0.16}$ & 1.67$^{+0.17}_{-0.15}$ & 1.65$^{+0.28}_{-0.25}$ \\[+.5mm]
& L$^c_{0.5-7}$ & 79.83$^{+10.31}_{-11.84}$ & 76.01$^{+30.56}_{-32.08}$ & 409.83$^{+59.58}_{-56.53}$& 376.99$^{+36.67}_{-34.38}$ & 207.78$\pm$17.57 & 477.44$^{+48.89}_{-35.14}$ & 212.36$^{+47.36}_{-42.78}$ \\[+.5mm]
& red. $\chi^2$/d.o.f$^e$ & 1.54/23 & 1.10/48 & 1.12/233 & 1.10/233 & 0.98/73 & 0.99/109 & 0.59/5 \\
\hline
\multicolumn{7}{l}{\textit{Nuclear emission}}\\
AGN & \NH$^d$ & 22.42$^{+7.84}_{-6.30}$& 23.36$^{+3.74}_{-3.23}$& - & - & 0.92$^{+0.99}_{-0.55}$ & - & - \\[+.5mm]
& $\Gamma$ & 0.46$^{+0.96}_{-0.82}$ & 0.67$^{+0.48}_{-0.42}$ & - & - & 1.85$^{+0.55}_{-0.37}$ & - & - \\[+.5mm]
& L$^c_{0.5-7}$ & 1519.4$^{+1551.5}_{-467.5}$& 6924.8$^{+2979.2}_{-41.3}$& - & - & 274.62$^{+141.70}_{-46.98}$ & - & - \\[+.5mm]
Fe K$\alpha$ & Energy$^a$ & - & 6.41$\pm$0.05 & - & - & - & - & - \\[+.5mm]
& width $\sigma$ & - & 0.10$\pm$0.05 & - & - & - & - & - \\[+.5mm]
& L$^c_{0.5-7}$ & - & 116.11$^{+45.45}_{-40.87}$ & - & - & - & - & - \\[+.5mm]
& red. $\chi^2$/d.o.f$^e$ & 0.85/56 & 1.09/174 & - & - & 1.04/110 & - & -
\enddata
\tablecomments{Model indicates major model component, APEC thermal plasma (AP), powerlaw (PL) and powerlaw with intrinsic absorption (abs.PL). $^a$ Temperatures and energies are in units of keV. $^b$ Abundance relative to the solar value. $^c$ 0.5-7~keV luminositieses are given in units of 10$^{38}$\ensuremath{\erg \ps}. $^d$ Intrinsic absorption in units of 10$^{22}$\ensuremath{\cm^{-2}}. $^e$ The reduced $\chi^2$ of the fit over its degrees of freedom. $^f$ Parameters fixed during fitting.}
\end{deluxetable*}
\subsection{NGC~833 / HCG~16B}
NGC~833 is a disturbed Sb galaxy, with an asymmetric velocity distribution and misalignment between the gas and stellar rotation axes \citep{MendesdeOliveiraetal98}, indicating that it is interacting with NGC~835. The galaxy has a LINER nucleus \citep{VeronCettyVeron06} but there is no optical evidence of ongoing star formation, and the galaxy is relatively poor in ionized and molecular gas compared to its companions \citep{MendesdeOliveiraetal98}. It is also the least luminous of the four major galaxies in the infrared and ultraviolet bands. Using the \emph{XMM}\ first light observation, \citet{Turneretal01special} find that the galaxy spectrum is best modelled using a combination of a soft (0.47$\pm$0.12~keV) thermal model and two powerlaw components with $\Gamma$=1.8$\pm$0.5, one of which is strongly absorbed below $\sim$3~keV.
\begin{figure}
\includegraphics[width=\columnwidth,bb=36 126 576 666]{HCG16AB_ps_regions_SDSS.pdf}
\caption{\label{fig:sourcesAB} Images of NGC~835 (left) and NGC~833
(right). The upper panel shows an SDSS $i$-band image with contours of
GMRT 610~MHz flux density overlaid (starting at 3$\times$rms and
increasing in steps of factor 2). The lower panel shows a \emph{Chandra}\ 0.5-2
keV image with the same alignment and scale, smoothed with a 2\mbox{\arcm\arcm}\
Gaussian. Small cyan ellipses indicate spectral extraction regions used
to examine point sources. Dashed regions are those used to extract
spectra of the galaxy cores and associated diffuse emission. }
\end{figure}
The GMRT 610~MHz image shows only an unresolved source at the position of
NGC~833, presumably associated with the AGN. We model the X-ray spectrum of
the central 15\mbox{\arcm\arcm}-radius region of the galaxy with the thermal plus
powerlaw plus absorbed powerlaw model described above. The spectral
extraction region, and those of point sources in the galaxy, is shown in
Figure~\ref{fig:sourcesAB}. Our best fit to spectra from ObsIDs 15181,
15666 and 15667 is shown in Figure~\ref{fig:HCG16b_spec} and its parameters
are listed in Table~\ref{tab:AGN}. The temperature of the thermal component
agrees with that found by TRP01, and we find a roughly half solar
abundance. Our absorbed AGN model has a hydrogen column identical within
errors with that found by TRP01, and a flatter powerlaw index consistent
with TRP01 at the 2$\sigma$ level.
\begin{figure}
\includegraphics[width=\columnwidth,bb=10 30 710 585,clip=]{N833_bestfit_iplot.pdf}
\caption{\label{fig:HCG16b_spec} Spectra and best fitting model for the
central 15\mbox{\arcm\arcm}-radius region of HCG~16B / NGC~833. Green, black and red
points represent ObsIDs 15181, 15666 and 15667 respectively, solid lines
show the best fitting model and dashed lines contributions from the
various model components. The lower panel shows the significance of the
residuals to the fit.}
\end{figure}
NGC~833 is also in the field of view of ObsID~923. The observation is too short to constrain our chosen model. However, comparing our best fitting model to the spectrum we find that the soft emission is reasonably well described by the thermal component, but that the hard emission is stronger than the model predicts. To test whether this arises from a change in the AGN intrinsic luminosity or in the absorbing column, we allow either the powerlaw normalization or the column density to vary while holding all other parameters fixed at their best fit values. An increase in normalization provides a significantly better fit than an increase in absorbing column (reduced $\chi^2$=1.457 compared to 2.772 for 10 degrees of freedom), leading us to conclude that in 2000 November the AGN luminosity was a factor $\sim$3 greater than in 2013 July.
\subsection{NGC~835 / HCG~16A}
The other galaxy in the interacting pair, NGC~835, is a Seyfert~2 \citep{VeronCettyVeron06} with an apparent tidal tail and clumpy ring of UV-bright knots at the edge of its disk, suggestive of star formation (TRP01). As with NGC~833, TRP01 modelled the X-ray emission from the galaxy with a combination of thermal and absorbed powerlaw components, finding that the temperature of the thermal component fell from $\sim$0.5~keV in the galaxy core to $\sim$0.3~keV in the star forming ring. The GMRT 610~MHz data reveal a central point source surrounded by diffuse emission from the galaxy disk, probably arising from star formation.
We extract spectra from a 15\mbox{\arcm\arcm}-radius region (inside the ring) and fit the same model we used for NGC~833. The spectral extraction region is shown in Figure~\ref{fig:sourcesAB}. Fitting the 2013 observations simultaneously, we find a satisfactory solution, with an apparent 6.4~keV Fe K$\alpha$ line visible in the longest exposures, ObsIDs 15181 and 15667. We therefore add a redshifted Gaussian component to the model and allow its energy, flux and line width to fit. The best fitting parameters are shown in Table~\ref{tab:AGN} and the best fitting model in Figure~\ref{fig:HCG16a_spec}.
\begin{figure}
\includegraphics[width=\columnwidth,bb=10 30 710 585,clip=]{N835_bestfit_iplot.pdf}
\caption{\label{fig:HCG16a_spec} Spectra and best fitting model for the central 15\mbox{\arcm\arcm}-radius region of HCG~16A / NGC~835, using ObsIDs 15181, 15666 and 15667. Colors and lines are as described in Figure~\ref{fig:HCG16b_spec}.}
\end{figure}
Our best fitting thermal component has a temperature higher than, but consistent at the 2$\sigma$ level with, that found by TRP01. We find an an abundance 0.25\ensuremath{\mathrm{~Z_{\odot}}}\ lower than that in NGC~833, but consistent within the uncertainties. The AGN component has a significantly lower absorbing column than that found from \emph{XMM}\ by TRP01 (23.4$^{+3.7}_{-3.2}$$\times$10$^{22}$\ensuremath{\cm^{-2}}\ compared to 46$\pm$15$\times$10$^{22}$\ensuremath{\cm^{-2}}) and a flatter powerlaw index (0.67$^{+0.48}_{-0.42}$ compared to 2.25$\pm$0.23). However, we note that our model differs from that of TRP01; their model contains no component to account for the X-ray binary population, but does include an unabsorbed AGN component whose powerlaw index is partially constrained by soft emission.
As with NGC~833, we find that our best fitting thermal component provides a reasonable description of the spectra extracted from ObsIDs 923 and 10394 below $\sim$2~keV, but that the hard 2-7~keV band flux in these earlier observations is significantly less than expected. We again allow either the AGN powerlaw normalization or absorbing column to fit, to test the likely cause of the reduced hard fluxes. In both cases a reduced normalization produces the better fit; reduced $\chi^2$=1.374 compared to 1.559 for 13 degrees of freedom in ObsID 923 and reduced $\chi^2$=2.098 compared to 2.559 for 12 degrees of freedom in ObsID 10394. The best fitting normalisations are $\sim$21\% and $\sim$18\% of the value found for the 2013 observations, respectively. However, the poor quality of these fits indicates that these simple adjustments to the model are insufficient, and the overall shape of the spectrum may have changed. Unfortunately these earlier observations lack the depth to constrain the model further. We conclude that the Seyfert nucleus of NGC~835 is variable in the X-ray band on timescales of months to years, probably in large part owing to changes in intrinsic luminosity.
\subsection{Diffuse emission between NGC~833 and NGC~835}
We extract spectra of the diffuse emission around and between NGC~833 and NGC~835 using a polygon region chosen to follow the stellar envelope, excluding point sources and the 15\mbox{\arcm\arcm}\ radius galaxy regions described above (see Figure~\ref{fig:sourcesAB}. The region includes the tidal arm and star--forming ring of NGC~835, and the outer tidally-disturbed parts of NGC~833. All five observations cover this region, though only partially in the case of ObsID~10394. The spectra are adequately modelled by a thermal plasma model with Galactic absorption (reduced $\chi^2$=1.06 for 108 degrees of freedom). The thermal component has temperature kT=0.48$^{+0.05}_{-0.04}$~keV, abundance 0.09$^{+0.04}_{-0.03}$\ensuremath{\mathrm{~Z_{\odot}}}\ and a 0.5-7~keV luminosity of (1.08$\pm$0.05)$\times$10$^{40}$\ensuremath{\erg \ps}. Neither a two-temperature thermal model or a thermal plus powerlaw model provide a better fit, and both have unconstrained parameters, suggesting that the diffuse emission is at least primarily thermal and relatively soft.
\subsection{NGC~838 / HCG~16C}
NGC~838 is a Luminous Infrared Galaxy (LIRG) having an infrared luminosity
$>$10$^{11}$\ensuremath{\mathrm{~L_{\odot}}}, and is the most IR and UV luminous of the four major
galaxies. \citet{MendesdeOliveiraetal98} note a number of optical features
indicating disturbance and possibly a recent merger, including kinematic
warping and multiple kinematic components in the ionized gas which are
misaligned with the stellar major axis. \citet{Vogtetal13} argue that the
galaxy underwent a period of global star formation $\sim$500~Myr ago, with
star formation continuing in the galaxy core. This activity is driving an
asymmetric bipolar wind, visible in ionized gas, which is probably only a
few Myr old in its current phase \citep{Vogtetal13}. TRP01, using \emph{XMM},
identified diffuse X-ray emission in the galaxy, and found that it could be
modelled as two-temperature thermal emission with kT=0.59$\pm$0.04~keV and
3.2$\pm$0.8~keV.
The \emph{Chandra}\ image (Figure~\ref{fig:sourcesC}) shows that the diffuse
emission extends along a roughly north-south axis to $\sim$15\mbox{\arcm\arcm}\ north
of the galaxy centroid and $\sim$25\mbox{\arcm\arcm}\ southward. The northern emission
appears brighter. \citet{Vogtetal13} suggest that the northern side of the
galaxy disk is facing us with the central starburst obscured by a dust
lane, and that while the outflowing starburst wind has inflated bubbles on
both sides of the disk, the northern bubble is better confined by the
surrounding neutral gas and therefore more compact. The \emph{Chandra}\ X-ray
images are consistent with this picture, and we also see 610~MHz emission
extending $\sim$30\mbox{\arcm\arcm}\ north and south of the galactic disk. In the south
this is clearly correlated with the X-ray (and H$\alpha$) emission, but on
the north side of the galaxy the radio source extends furthest to the
northeast and is less clearly correlated with the X-ray emission. The
connection between the starburst winds and the surrounding IGM is discussed
in paper II.
\begin{figure}
\includegraphics[width=\columnwidth,bb=36 243 576 548]{HCG16C_ps_regions_SDSS.pdf}
\caption{\label{fig:sourcesC}Images of NGC~838. The left panel shows an SDSS $i$-band image with contours of GMRT 610~MHz flux density overlaid (starting at 3$\times$rms and increasing in steps of factor 2). The right panel shows a \emph{Chandra}\ 0.5-2 keV image with the same alignment and scale, smoothed with a 2\mbox{\arcm\arcm}\ Gaussian. Small cyan ellipses indicate spectral extraction regions used to examine point sources. Dashed regions are those used to extract spectra of the galaxy as a whole and its superwind.
}
\end{figure}
Extracting spectra from a $\sim$25\mbox{\arcm\arcm}\ radius circular region, we find that all five ObsIDs can be adequately modelled by either a two-temperature thermal plasma, or a low-temperature thermal model plus a powerlaw. As expected, there is no indication of significant luminosity variation with time. Best-fitting parameters for both models are listed in Table~\ref{tab:AGN} and APEC plus powerlaw model fit to all five spectra is shown in Figure~\ref{fig:HCG16c_spec}. We find a temperature for the soft component $\sim$0.2~keV hotter than the TRP01 estimate ($\sim$4.5$\sigma$ significant), and a low abundance, Z$\sim$0.16\ensuremath{\mathrm{~Z_{\odot}}}. While we might expect a high metallicity in a galactic wind driven by supernovae, it will be diluted by entrainment and mixing with less enriched cold gas. Our measurements may be biased low by the multiphase nature the wind, since a single-temperature model will typically underestimate the abundance of multi-temperature gas around 1~keV \citep[the ``Fe bias'',][]{Buotefabian98,Buote00b}. As suggested by TRP01, the spectrally hard component probably arises from young high-mass X-ray binaries formed in the star formation episodes, while the soft component is primarily the hot gas of the starburst wind.
\begin{figure}
\includegraphics[width=\columnwidth,bb=10 30 710 585,clip=]{N838_bestfit_iplot.pdf}
\caption{\label{fig:HCG16c_spec} Spectra and best fitting APEC+powerlaw model for the central 25\mbox{\arcm\arcm}-radius region of HCG~16C / NGC~838, showing all five \emph{Chandra}\ observations. Colors and lines are as described in Figure~\ref{fig:HCG16b_spec} with ObsID~923 marked in blue and ObsID~10394 in cyan.}
\end{figure}
To examine the temperature structure of the wind we divided the emission
into strips with widths of 5, 10, and 15\mbox{\arcm\arcm}\ (chosen to include a few
hundred counts) and extracted spectra from all five observations. A 5\mbox{\arcm\arcm}\
width strip was placed at the optical centroid of the galaxy, and includes
the majority of the hard (2-7~keV) emission. Figure~\ref{fig:sourcesC}
shows these regions. We fitted each region with an APEC or (in the central
region) APEC+powerlaw model, either allowing abundance to fit freely or
fixing it at 0.3\ensuremath{\mathrm{~Z_{\odot}}}. The results of these fits for the thermal component
are shown in Table~\ref{tab:wind}. Although the models are
under-constrained in some cases (reduced $\chi^2<1$), the fits suggest that
the wind has a fairly consistent temperature of $\sim$0.7-0.8~keV, falling
to 0.3~keV (a 4.4$\sigma$ significant decline) in the outer part of the more
extended southern outflow.
\begin{deluxetable}{lccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:wind}Best-fitting model parameters for the NGC~838 wind}
\tablehead{
\colhead{Region} & \multicolumn{3}{c}{Abundance free} & \multicolumn{2}{c}{Abundance=0.3\ensuremath{\mathrm{~Z_{\odot}}}} \\
\colhead{} & \colhead{kT} & \colhead{Abund.} & \colhead{red. $\chi^2$/dof} & \colhead{kT} & \colhead{red. $\chi^2$/d.o.f.} \\
\colhead{} & \colhead{(keV)} & \colhead{(\ensuremath{\mathrm{~Z_{\odot}}})} & \colhead{} & \colhead{(keV)} & \colhead{} \\
}
\startdata
Central & 0.92$^{+0.05}_{-0.04}$ & $>$0.14 & 1.28/101 & 0.92$^{+0.05}_{-0.04}$ & 1.26/102 \\
North 1 & 0.79$\pm$0.02 & 0.08$\pm$0.01 & 1.34/92 & 0.81$\pm$0.02 & 2.41/93 \\
South 1 & 0.74$^{+0.06}_{-0.07}$ & 0.06$\pm$0.03 & 0.67/23 & 0.81$\pm$0.04 & 1.47/24 \\
South 2 & 0.30$^{+0.07}_{-0.03}$ & $<$0.03 & 0.57/9 & 0.75$^{+0.08}_{-0.35}$ & 1.36/10
\enddata
\end{deluxetable}
The APEC thermal plasma model assumes collisional ionisation equilibrium, which may not hold in a complex, rapidly expanding galaxy wind \citep[e.g.,][]{Breitschwerdt03}. We therefore test the impact of using a non-equilibrium ionisation (NEI) model instead of APEC in our fits. In general the NEI model produces similar temperatures, abundances and fit statistics. Where significant differences are found (e.g., in the North 1 bin where the best fit NEI temperature is 0.70$\pm$0.02~keV) the NEI model produces a poorer fit to the data (reduced $\chi^2$=1.479). We therefore conclude that, while the wind gas is likely to be out of collisional equilibrium in some regions, our results are unlikely to be significantly effected. Deeper observations would be required to probe the equilibrium state of the wind.
\subsection{NGC~839 / HCG~16D}
NGC~839 is the second LIRG in HCG~16, and has also been classified as a
LINER-2 \citep{DeCarvalhoCoziol99}, though deep optical integral field
spectroscopy suggests that the LINER emission arises from shock excitation
in an outflowing starburst wind \citep{Richetal10}. This wind is visible as
a biconical polar outflow in H$\alpha$ emission, and stellar population
modelling shows that the galaxy contains a sizable population of A stars,
indicating a starburst age of $\sim$400~Myr. TRP01 show that the
\emph{XMM}\ spectrum is best fit by a model including a fairly heavily absorbed
powerlaw, and conclude that an active nucleus is present. Our GMRT 610~MHz
observation detects only an unresolved source coincident with the galaxy
centroid.
The galaxy is outside the field of view of ObsID~10394, but we extract
spectra from the other four observations, using the regions shown in
Figure~\ref{fig:sourcesD}. We initially extract a spectrum excluding the
central point source, and find that this is well fitted by an APEC plus
powerlaw model, representing the thermal emission from the galactic wind,
and the X-ray binary population. Applying this model to the spectra for the
whole galaxy, we find that the emission from the central point source is
best modelled by a second, intrinsically absorbed powerlaw.
However, we find that the emission from the galaxy as a whole (including
the central point source) can also be well modelled using an APEC component
and a single intrinsically-absorbed powerlaw. This model fit is shown in
Figure~\ref{fig:HCG16d_spec}. The fact that the hard component of the
spectra can be modelled by a single absorbed powerlaw component suggests
the possibility that it could be dominated by emission from an AGN, or that
the X-ray binary population dominates and the AGN emission is negligible.
In either case, the model requires a significant absorption column within
the galaxy. Parameters for both fits are listed in Table~\ref{tab:AGN}. The
abundance of the thermal component is poorly constrained and is therefore
fixed at 0.3\ensuremath{\mathrm{~Z_{\odot}}}\ in both fits.
\begin{figure}
\includegraphics[width=\columnwidth,bb=36 274 576 517]{HCG16D_ps_regions_DSS.pdf}
\caption{\label{fig:sourcesD}Images of NGC~839. The left panel shows an Digitized Sky Survey $R$-band image with contours of GMRT 610~MHz flux density overlaid (starting at 3$\times$rms and increasing in steps of factor 2). The right panel shows a \emph{Chandra}\ 0.5-2 keV image with the same alignment and scale, smoothed with a 2\mbox{\arcm\arcm}\ Gaussian. Small cyan ellipses indicate spectral extraction regions used to examine point sources. Dashed regions are those used to extract spectra of the galaxy as a whole and its superwind.}
\end{figure}
Our best fitting temperature is somewhat higher than that found by TRP01 (0.88$\pm$0.05 or 0.79$^{+0.03}_{-0.04}$ compared to 0.63$\pm$0.10). The powerlaw index is consistent with the value found by TRP01 (1.85$^{+0.55}_{-0.37}$ or 1.67$^{+0.17}_{-0.15}$ compared to 2.1$\pm$0.8) but we find a much lower absorbing column, [0.92$^{+0.99}_{-0.55}$] or [0.47$^{+0.24}_{-0.18}$]$\times$10$^{22}$\ensuremath{\cm^{-2}}\ compared to their best fitting value of [45$\pm$20]$\times$10$^{22}$\ensuremath{\cm^{-2}}. We will discuss the origin of the powerlaw emission, and the question of whether NGC~839 hosts an AGN, in Section~\ref{sec:disc839}
\begin{figure}
\includegraphics[width=\columnwidth,bb=10 30 710 585,clip=]{N839_bestfit_iplot.pdf}
\caption{\label{fig:HCG16d_spec} Spectra and best fitting APEC+absorbed powerlaw model for the central 26\mbox{\arcm\arcm}-radius region of HCG~16D / NGC~839, showing data from ObsIDs 923, 15181, 15666 and 15667. Colors and lines are as described in Figure~\ref{fig:HCG16b_spec} with ObsID 923 marked in blue.}
\end{figure}
As with NGC~838, there is enough diffuse emission from the wind regions north and south of the galaxy core to allow (crude) spectral fitting. We divide the galaxy into three regions (8 and 16\mbox{\arcm\arcm}\ wide strips), fitting the core with the model described above, and the wind regions with an absorbed APEC thermal plasma model. The wind emission is too faint to constrain abundance and we therefore fix it at 0.3\ensuremath{\mathrm{~Z_{\odot}}}. The parameters of the thermal components of the fits are listed in Table~\ref{tab:hcg16dwind}.
\begin{deluxetable}{lcc}
\tablewidth{0pt}
\tablecaption{\label{tab:hcg16dwind}Best-fitting model parameters for the NGC~839 wind}
\tablehead{
\colhead{Region} & \colhead{kT} & \colhead{red. $\chi^2$/d.o.f.} \\
\colhead{} & \colhead{(keV)} & \colhead{} \\
}
\startdata
Central & 0.92$^{+0.05}_{-0.10}$ & 0.922/74 \\[+0.5mm]
North & 1.24$^{+0.11}_{-0.12}$ & 1.652/8 \\[+0.5mm]
South & 0.94$^{+0.09}_{-0.12}$ & 0.701/7 \\[+0.5mm]
North+South & 0.99$\pm$0.07 & 1.183/17
\enddata
\tablecomments{Abundance fixed at 0.3\ensuremath{\mathrm{~Z_{\odot}}}\ in all fits.}
\end{deluxetable}
\subsection{NGC~848}
NGC~848 lies $\sim$15\hbox{$^\prime$}\ from the group core, $\sim$11.5\hbox{$^\prime$}\ southeast
of NGC~839, but is linked to the four main galaxies by the \ensuremath{\mathrm{H}\textsc{i}}\ filament \citep{VerdesMontenegroetal01}. It is a peculiar barred spiral galaxy hosting a starburst \citep{Continietal98}, and is only visible on the ACIS-S1 CCD of observation 923. There is a clear detection of emission from the galaxy, apparently extended along the galactic bar (see Figure~\ref{im:N848}) though the large off-axis angle means that individual sources cannot be resolved.
\begin{figure}
\includegraphics[width=\columnwidth,bb=36 254 576 538]{N848_opt_x_regions.pdf}
\caption{\label{im:N848} Images of NGC~848. The left panel shows a \textit{Digitized Sky Survey} $R$-band image with contours of GMRT 610~MHz flux density overlaid (starting at 3$\times$rms and increasing in steps of factor 2). The right panel shows a \emph{Chandra}\ 0.5-2 keV image with the same alignment and scale, binned to 2\mbox{\arcm\arcm}\ pixels and smoothed with a 6\mbox{\arcm\arcm}\ Gaussian. The dashed region indicates the \ensuremath{D_{25}}\ ellipse used to extract a spectrum of the galaxy emission. The two images have the same scale and orientation.}
\end{figure}
We extract a spectrum from the S1 data using an elliptical source region
corresponding to the \ensuremath{D_{25}}\ optical contour of the galaxy, and a background
spectrum from an annulus centred on the galaxy with radii 1.75\hbox{$^\prime$}\ and
3.5\hbox{$^\prime$}. The effective exposure is short enough that even a simple
absorbed powerlaw model is under-constrained, with reduced $\chi^2$=0.586
for 5 degrees of freedom. The best fitting parameters of this model are
listed in Table~\ref{tab:AGN}. The powerlaw index,
$\Gamma$=1.65$^{+0.28}_{-0.25}$, is consistent with emission from X-ray
binaries, but is also consistent with emission from a LINER or Seyfert
nucleus \citep[e.g.,][]{GuCao09}. However, comparison with the optical and
radio images shows that the galaxy core is located between the two most
luminous clumps of X-ray emission, indicating that any X-ray emission from
an AGN is weak compared to the emission associated with star formation and
the X-ray binary population. The spectral model gives luminosity
$L_{0.5-7}$=2.12$^{+0.47}_{-0.43}$$\times$10$^{40}$\ensuremath{\erg \ps}\ at our adopted
distance for the group. An APEC thermal model provides a fit of similar
quality, but the temperature is unphysical (kT=5.1$^{+8.4}_{-2.4}$~keV),
suggesting there is relatively little hot gas in the galaxy. Fixing the
temperature to kT=0.5~keV and the abundance to Z=0.3\ensuremath{\mathrm{~Z_{\odot}}}, we can place a
3$\sigma$ upper limit on the flux from any thermal component of
$F_{0.5-7}<$1.38$\times$10$^{-14}$\ensuremath{\erg \ps \pcmsq}\
($L_{0.5-7}<$5.27$\times$10$^{39}$\ensuremath{\erg \ps}), fainter than the thermal
emission from any of the other major group members except NGC~833. The
GMRT 610~MHz image shows a marginally extended central source, with an
extension to the southeast that appears to correspond to an unidentified
optical source. This may be a foreground star or background AGN.
\section{Stellar Population Modelling}
\label{sec:SSP}
The Sloan Digital Sky Survey partially covers HCG~16, and optical spectra
of the cores of NGC~833 and NGC~838 are available as part of Data Release
10 \citep[SDSS-DR10,][]{Ahnetal14}. We analysed these spectra and
estimated the ages, metallicities, and mass of the stellar populations
using the spectral fitting code STARLIGHT \citep{CidFernandesetal05}.
STARLIGHT fits the observed spectrum with a combination of simple stellar
population (SSP) models covering a range of ages and metallicities. The
code returns the contribution, as a fraction of total stellar mass, from
each {\it basis} SSP. Before running the code, the observed spectra are
corrected for foreground extinction and de-redshifted, and the models are
degraded to match the wavelength-dependent resolution of the spectrum of
each galaxy, as described in \citet{LaBarberaetal10}.
In order to check the model-dependency of the stellar population properties we used two sets of SSP models. One of them is based
on the Medium resolution INT Library of Empirical Spectra \citep[MILES,
][]{SanchezBlazquezetal06}, using the updated version 9.1
\citep{FalconBarrosoetal11} of the code presented in
\citet{Vazdekisetal10}. We selected models computed with \citet{Kroupa01}
universal initial mass function (IMF) and isochrones by
\citet{Girardietal00}. We also used \citet[][BC03]{BruzualCharlot03}
models, calculated with Padova 1994 evolutionary tracks
\citep{Girardietal96} and with \citet{Chabrier03} IMF. The basis grids
cover ages in the range $0.07 - 14.2$~Gyr for MILES models and $0.02
-14.2$~Gyr for BC03 models, with constant $\log({\rm Age})$ steps of 0.2.
We selected SSPs with metallicities [M/H]~=$\{-1.71, -0.71, -0.38, 0.00,
+0.20\}$.
For NGC~833, we adopted the \citet{Cardellietal89} extinction law ($R_{\rm
V} = 3.1$). NGC~838 contains several dust lanes, and we adopted
the \citet{Calzettietal00} law ($R_{\rm V} = 4.05$), which is more suitable for
starburst galaxies. We also allowed stellar populations younger than
$0.032$~Gyr to have an extra extinction in relation to the older
populations. The stellar masses -- computed within the fiber aperture --
are corrected to the full extent of the galaxy by computing the difference
between fiber and model magnitudes in the $z$ band. However, we note that
this correction is only approximate, since the stellar population in the fiber
aperture is probably not representative of the galaxy as a whole.
In NGC~833, we find that the galaxy is dominated by an old stellar
population (mass-weighted age $\sim$10~Gyr for MILES and $\sim$13.5~Gyr for
BC03), with a young component making up $\sim$0.01\% of the stellar mass.
The age of the young component is dependent on the models used, with star
formation beginning $\sim$300~Myr ago for MILES and $\sim$50~Myr ago for
BC03. However, in both cases the star formation rates are small
(1-3\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}) and continue to the lower age limit of the basis grid. The
fits suggest a total stellar mass for the galaxy of
1-4$\times$10$^{11}$\ensuremath{\mathrm{~M_{\odot}}}\ (in rough agreement with luminosity based
estimates, see Section~\ref{sec:gascont}), and metallicity
[M/Fe]=0.02-0.22.
For NGC~838, the properties of the stellar population are much more
sensitive to the choices made in the spectral fitting. Depending on the
models and extinction law adopted, the mass-weighted age varies from
$\sim$5 to $\sim$11~Gyr, metallicity varies across the range [M/Fe]=-1.2 to
[M/Fe=+0.3, and stellar masses derived with BC03 models are a factor
$\sim$10 greater than those derived from MILES models. The MILES fits find
that 30-50\% of the stars in the fiber aperture were formed in a burst
starting $\sim$300~Myr ago, with SFR $\sim$15-50\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}. The BC03 fits
suggest a much smaller young component ($\sim$2\% by mass) formed in a
burst starting 50~Myr ago and continuing to the present, with SFR
$\sim$50-80\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}, and an extended star-formation history for the old
stellar component, continuing as late as 2~Gyr ago.
Figure~\ref{fig:SSP838} shows the results of two fits to the NGC~838
spectrum using the MILES and BC03 models, compared with fits to the same
spectrum from the Versatile Spectral Analysis database
\citep[VESPA,][]{Tojeiroetal09}. VESPA fits use either the BC03 models or
the models of \citet{Maraston05}, and for the young population these
produce similar results to our BC03 fits. However, they differ in finding
little evidence of an extended period of formation for the old population.
\begin{figure}
\includegraphics[width=\columnwidth,bb=0 0 576 576]{SFH_NGC838d3.pdf}
\caption{\label{fig:SSP838} Star formation history of NGC~838 comparing our fits using the MILES and BC03 models, to results from the VESPA database, using BC03 or \protect\citet{Maraston05} models. Histograms show the results of the model fits, while errorbars and upper limits indicate uncertainties derived from monte-carlo simulations of the best-fitting STARLIGHT model.}
\end{figure}
To test the reliability of the STARLIGHT fits, we simulated spectra based on the BC03 and MILES modelling results. For each of the the two best-fitting model, 100 spectra were simulated and noise was added to acheive a signal-to-noise ratio of 60, matching that of the observed spectrum. Each simulated spectrum was then fitted using the STARLIGHT code, and the results of the fits used to estimate the 1$\sigma$ uncertainty on the SFR in each age bin, shown as errorbars in Figure~\ref{fig:SSP838}. Where the 1$\sigma$ uncertainty was consistent with zero, show the 1$\sigma$ upper limit on SFR. These uncertainties indicate that the general picture of the star formation history presented by each set of fits is fairly robust, and that even quite low SFRs ($\sim$0.1\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}) are unlikely in the 300~Myr-1~Gyr age range.
We must also consider previous observations of NGC~838, which show that the
stellar population in the galaxy core is probably not representative of the
galaxy as a whole. While \citet{Vogtetal13} did not fit SSP models to their
integral field spectra of NGC~838, they did examine those spectra for
simple age indicators. Examining spectra from the core and disk they found
that both regions have similar absorption features typical of a young,
A-type stellar population with an age $\sim$500~Myr post-starburst, but
that the core has a much stronger blue continuum, indicative of an O/B-type
stellar population consistent with very recent/ongoing star formation. They
conclude that the galaxy underwent a galaxy-wide star formation episode
starting $\sim$500~Myr ago, but that star formation rates have declined
with time and activity is now restricted to the core. Our BC03 fits (and
the VESPA results) agree with Vogt et al. in finding ongoing star
formation in the core, and suggest that it began $\sim$50~Myr ago. Our MILES
fit suggests that star formation began in the core as early as
$\sim$300~Myr ago, but does not find it to be ongoing. However, since the
SDSS spectrum only samples the core, we cannot compare our age estimates
with the overall burst age estimated by Vogt et al. based on spectra from
the disk of the galaxy. We therefore adopt their burst age of
$\sim$500~Myr, with the caveat that star formation probably began more
recently in the galaxy core. High-resolution optical spectra from elsewhere
in NGC~838 would be required to more accurately measure the burst age.
\section{Discussion}
\subsection{Point Sources}
\citet{Tzanavarisetal14} have used ObsID 923 to examine the point source population in and around the four original group members. Eight of their sources are located within our adopted \ensuremath{D_{25}}\ regions, including the cores of NGC~833, NGC~835 and NGC~839, and several sources in NGC~838. In general our results agree well with theirs for the brighter sources, though our ability to model the spectra of the absorbed AGN naturally produces more accurate powerlaw spectral indices. We find somewhat lower luminosities for the faintest sources, and fail to detect their source 46, probably owing to the different flux extraction methods and background regions used.
\subsection{Gas content of the galaxies}
\label{sec:gascont}
Table~\ref{tab:mgas} shows our estimates of the density and mass of the gas
in the four main galaxies. We estimate the electron number density of the
hot gas ($n_e$) from the spectral fits described in section~\ref{sec:AGNdiff}.
The core region of NGC~833 is reasonably well described by a spherical
$\beta$-model with $r_{core}$=4.1~kpc and $\beta$=0.51, and we estimate the gas
mass from this model and the normalisation of the APEC model. The core of
NGC~835 is not well described by a single $\beta$-model (or even two), so
to reduce the bias arising from assuming a constant mean density in a
volume where density actually varies, we break it into two spectral
regions, a central cylinder of length $\sim$3.1~kpc and radius 0.97~kpc
aligned north-south, and the remaining 15\mbox{\arcm\arcm}\ (4.1~kpc) radius sphere.
For the diffuse emission in the disks of the two galaxies we approximate
the volume as a rectangular slab with projected area equal to that of the
polygon region used for spectral extraction, $\sim$13$\times$44~kpc and
depth 10~kpc. The third row of Table~\ref{tab:mgas} shows results summed
over all three regions in the NGC~833/835 complex. The gas density outside
the galaxy cores is 1.58$^{+0.85}_{-0.78}$$\times$10$^{-3}$\ensuremath{\cm^{-3}}.
\begin{deluxetable*}{lcccccccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:mgas}Gas and stellar mass estimates}
\tablehead{
\colhead{Galaxy} & \colhead{log $L_K$} & \colhead{log $L_{3.6\mu m}$} & \colhead{$M_{\rm H\textsc{i}}$}& \colhead{$n_e$} & \colhead{$M_{\rm gas}^{\rm hot}$} & \colhead{$M_{*}^K$} & \colhead{$M_{*}^{3.6\mu m}$} & \colhead{$M_{\rm gas}^{\rm hot}$/$M_*^K$} & \colhead{$M_{\rm gas}$/$M_*^K$} \\
\colhead{} & \colhead{(\ensuremath{\mathrm{~L_{\odot}}})} & \colhead{(\ensuremath{\mathrm{~L_{\odot}}})} & (10$^9$\ensuremath{\mathrm{~M_{\odot}}}) & \colhead{(10$^{-2}$\ensuremath{\cm^{-3}})} & \colhead{(10$^8$\ensuremath{\mathrm{~M_{\odot}}})} & \colhead{(10$^{10}$\ensuremath{\mathrm{~M_{\odot}}})} & \colhead{(10$^{10}$\ensuremath{\mathrm{~M_{\odot}}})} & \colhead{($\times$10$^{-3}$)} & \colhead{($\times$10$^{-3}$)} \\
}
\startdata
NGC 833 & 11.06 & 10.99 & 0.79 & 0.80$^{+1.37}_{-0.76}$ & 0.61$^{+0.61}_{-0.18}$ & 7.46 & 4.83 & 0.62 & 11.21 \\[+.5mm]
NGC 835 & 11.27 & 11.28 & 1.17 & 2.15$\pm$1.29 & 0.64$\pm$0.36 & 12.10 & 9.62 & 1.03 & 10.70 \\[+.5mm]
NGC 833+NGC 835 & 11.48 & 11.45 & 1.96 & - & 3.57$^{+1.49}_{-1.27}$ & 19.56 & 14.45 & 1.83 & 13.19 \\[+.5mm]
NGC 838 & 10.95 & 11.17 & 3.02 & 2.27$^{+1.17}_{-1.31}$ & 2.57$^{+1.33}_{-1.48}$ & 5.79 & 7.40 & 4.43 & 56.60 \\[+.5mm]
NGC 839 & 10.92 & 11.03 & $>$4.47 & 1.86$^{+0.66}_{-0.69}$ & 0.41$\pm$0.15
& 5.41 & 5.36 & 0.76 & $>$83.38 \\[+.5mm]
NGC 848 & 10.67 & - & 0.77 & - & - & 3.04 & - & - & 25.33
\enddata
\tablecomments{We adopt $M_{*}$/$L_K$=0.65 and $L_{*}$/$L_{3.6\mu m}$=0.5
\citep{Eskewetal12,McGaughSchombert14}. \ensuremath{\mathrm{H}\textsc{i}}\ masses from
\citet{VerdesMontenegroetal01}. No 3.6$\mu$m data are available for
NGC~848, and we detect no hot gas in the galaxy.}
\end{deluxetable*}
In NGC~838 the diffuse emission is primarily located in the inner part of the galaxy or in the southern wind region. We approximate these two volumes as an oblate ellipsoid with radii 18.7\mbox{\arcm\arcm}\ and 10.1\mbox{\arcm\arcm}\ (5.1 and 2.76~kpc) and a conical section of a spherical shell with opening angle 43\hbox{$^\circ$}\ and radii 25.5\mbox{\arcm\arcm}\ and 47\mbox{\arcm\arcm}\ (7.0 and 12.8~kpc). While the conical geometry of the wind in NGC~839 is obvious in H$\alpha$ \citep{Richetal10} in the X-ray it is less clear, and most of the wind and galaxy disk emission can be enclosed in a 15$\times$30\mbox{\arcm\arcm}\ box, suggesting a cylindrical geometry.
Estimating the stellar mass of the major galaxies from their $K$-band and
3.6$\mu$m luminosities \citep[adopting stellar mass-to-light ratios
from][]{Eskewetal12,McGaughSchombert14} we find that, within the immediate boundaries
of the galaxies, the typical hot-gas-to-stellar mass ratio is
$\sim$10$^{-3}$, with the starburst superwind galaxy NGC~838 having the
highest ratio. However NGC~835 contains the largest quantity of hot gas
(2.57$^{+1.33}_{-1.48}$$\times$10$^8$\ensuremath{\mathrm{~M_{\odot}}}), perhaps unsurprising since it
has the highest stellar mass of the four. There is an order of magnitude
more cold \ensuremath{\mathrm{H}\textsc{i}}\ than hot gas in the galaxies, giving total gas-to-stellar
mass ratios $\gtrsim$1\% for all five galaxies.
We can estimate approximate isobaric cooling times for the hot gas in the
galaxies, and find values of $\sim$1.4~Gyr and $\sim$1.8~Gyr for NGC~839
and NGC~838 respectively. This suggests that little of the gas in these
galaxy winds is likely to have cooled out of the X-ray regime. However, if
cooling via mixing with colder entrained gas dominates over radiative
cooling, the true cooling times will be shorter.
We estimate outflow rates for the two superwind galaxies based on the
geometries adopted above and an assumed outflow velocity. The only direct
measurement of outflow velocity comes from line splitting in the optical
spectra of NGC~839 which, accounting for the inclination of the galaxy and
opening angle of the conical H$\alpha$ wind structure, suggests a wind
velocity $\sim$250\ensuremath{\km \ps}\ \citep{Richetal10}. This suggests an outflow of
$\sim$2.5\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}\ of hot gas in NGC~839. \citet{Vogtetal13} suggest, based
on the size of the wind-blown bubbles in NGC~838, that these structures are
5-50~Myr old with an outflow velocity of $\sim$130-1300\ensuremath{\km \ps}. If we
assume a wind velocity similar to that in NGC~839, the bubbles would be
$\sim$27~Myr old. The outflow rate of hot gas is then $\sim$17\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}.
As star formation seems to have peaked in the two galaxies 400-500~Myr ago,
and declined since then, these rates are likely underestimates of the
average outflow over that period. However, if we assume outflows at these
rates over this timescale, we find that the two galaxies may have ejected
9.5$\times$10$^9$\ensuremath{\mathrm{~M_{\odot}}}\ of hot gas. This value exceeds the \ensuremath{\mathrm{H}\textsc{i}}\ mass of the
galaxies (and the hot gas mass in the surrounding region, see paper II),
suggesting either that the outflow rates are overestimated, or that gas
infall is replenishing the galaxies. The outflow velocity measured in
NGC~839 \citep[$\sim$250\ensuremath{\km \ps},][]{Richetal10} is greater than the escape
velocity of the galaxy \citep[$\sim$200\ensuremath{\km \ps}, based on the stellar and
gaseous rotation velocities,][]{MendesdeOliveiraetal98} but only by
$\sim$20\%, suggesting that not all of the wind material will escape. The
two galaxies are also embedded in the \ensuremath{\mathrm{H}\textsc{i}}\ filament, which is likely to
impede outflows and may be responsible for containing the bubbles seen in
NGC~838. This may increase the fraction of outflowing gas which falls back
into the galaxy and is cooled and recycled. However, as we argue in paper
II, the morphology of the winds and of the surrounding diffuse emission
suggests that at least part of their outflows do escape and contribute to
the intra-group medium.
\subsection{Star Formation Rates}
\label{sec:starburst}
The star formation rates (SFRs) in the group member galaxies can be
estimated both from our X-ray and radio data, and from measurements in the
infra-red (IR) and ultra-violet (UV). The hard component of X-ray emission
in star forming galaxies arises primarily from high mass X-ray binaries,
formed from massive stars with short lifespans. It is therefore closely
linked to recent star formation, but not the current rate, since some time
is required for one member of each binary to evolve into a compact object.
\citet{Fragosetal13} suggest that the HMXB population is related to star
formation over the previous 2-100~Myr, but there is significant scatter in
the relationship. UV and IR measurements probe star formation via the most
massive, shortest-lived stars, and so may give a more accurate estimate of
the current star formation rate.
We use the $L_X$:SFR relation of \citet{Mineoetal12} to estimate star
formation rates from the 0.5-8~keV luminosity of the X-ray binary
population of each galaxy, as determined from the powerlaw component of the
galaxy emission from the best fitting models described in
section~\ref{sec:AGNdiff}, and including contributions from individual
point sources described in section~\ref{sec:PS}. We exclude the probable
background sources (numbers 11 and 16), and the nuclear sources (3, 12 and
14). Since individual ULXs may contribute a significant fraction of the
hard X-ray luminosity of an entire galaxy we calculate two star formation
rates for those galaxies containing ULX candidates, one including the ULX
contribution to luminosity, one with ULXs excluded. These estimates are
shown in Table~\ref{tab:SFR}. In NGC~838 and NGC~839 the estimates are
complicated by the question of whether the galaxies host AGN. The evidence
suggests that source 3 in NGC~838 is probably an AGN, but if it were a ULX
it would increase the total SFR estimate for the galaxy to
16.68$^{+2.42}_{-2.30}$\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}. As we will argue in
Section~\ref{sec:disc839}, NGC~839 probably does not host an AGN, and we
base our SFR estimate on the AP+abs.PL model. If we instead assume an AGN
is present and use the luminosity of the unabsorbed powerlaw in the
AP+PL+abs.PL model to estimate SFR, we obtain a value of
8.74$\pm$0.74\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}.
To estimate star formation rates from the radio luminosity, we use the
$L_{1.4~GHz}$:SFR relation of \citet{Bell03}. The GMRT 610~MHz data has
sufficient resolution for us to see that the AGN dominates the radio
emission in NGC~833 and for us to separate the central point source from
the extended star-formation emission in NGC~835. In NGC~838, the 610~MHz
image shows diffuse emission coincident with the wind bubbles north and
south of the galaxy, and we subtract this flux, which seems likely to be
associated with past rather than recent star formation. Where necessary we
convert 610~MHz to 1.4~GHz fluxes using either the observed spectral index
(0.5-0.8 in NGC~838, NGC~839 and NGC~848) or a canonical value of 0.5 (for
NGC~835). The resulting SFR estimates are shown in Table~\ref{tab:SFR}. As
expected, NGC~838 and NGC~839 have the highest SFRs
(12.05$^{+2.50}_{-2.37}$ and 20.09$^{+2.06}_{-1.48}$\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}\
respectively), NGC~848 has a moderately enhanced rate of
8.76$^{+1.94}_{-1.78}$\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}, while NGC~833 and NGC~835 have rates of
$\sim$3\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}} (all including ULXs).
\begin{deluxetable*}{lccccc}
\tablewidth{0pt}
\tablecaption{\label{tab:SFR}Star formation rate estimates}
\tablehead{
\colhead{Estimator} & \multicolumn{5}{c}{SFR (\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}})} \\
\colhead{} & \colhead{NGC~833} & \colhead{NGC~835} & \colhead{NGC~838} & \colhead{NGC~839} & \colhead{NGC~848}\\
}
\startdata
X-ray (incl. ULXs) & 3.30$^{+0.43}_{-0.45}$ & 3.14$^{+1.26}_{-1.33}$ & 12.05$^{+2.50}_{-2.37}$ & 20.09$^{+2.06}_{-1.48}$ & 8.76$^{+1.94}_{-1.78}$ \\[+0.5mm]
X-ray (excl. ULXs) & 2.90$^{+0.43}_{-0.50}$ & 1.71$^{+1.28}_{-1.34}$ & 10.51$^{+2.51}_{-2.53}$ & - & - \\[+0.5mm]
Radio & - & 2.49$\pm$0.30 & 16.87$\pm$1.03 & 7.90$\pm$0.40 & 2.04$\pm$0.21 \\
SSP & 1-3 & - & 15-80 & - & - \\
24$\mu$m$^a$ & 0.07 & 1.74 & 7.24 & 9.55 & - \\
UV/IR$^b$ & 5.37$\pm$0.62 & 0.33$\pm$0.03 & 14.38$\pm$1.83 & 17.06$\pm$2.31 & -
\enddata
\tablecomments{$^a$: from Brassington et al. (2014, submitted).\\
$^b$: from \citet{Tzanavarisetal10}.}
\end{deluxetable*}
For comparison, we draw on two studies in the IR and UV:
\citet{Tzanavarisetal10}, who estimate SFRs using a combination of
\textit{Spitzer} 24$\mu$m and \textit{Swift} $uvw2$-band luminosities, and
Brassington et al. (2014, submitted), who use only \textit{Spitzer} data,
but use an updated $L_{24\mu m}$:SFR relation which is more accurate at the
high luminosities seen in starburst galaxies. Neither study includes
NGC~848. We also include estimates of the recent star formation rates from
the SSP models described in Section~\ref{sec:SSP}.
Although it is clear that the different estimators identify NGC~838 and
NGC~839 as the most actively star forming systems in HCG~16, there is a
significant degree of divergence among the actual SFR values. The SSP
estimate for NGC~833 is in general agreement with the other estimators,
though somewhat higher than the Brassington et al. value. The lower end of
the range of SFRs estimated from the SSP modelling of NGC~838 is in
agreement with the other estimates (though again higher than the
Brassington et al. result) but the upper limit greatly exceeds any of the
other estimates. Neither our BC03 or MILES model fits agree with the other
SFR estimates in detail; BC03 predicts current SFRs a factor $\sim$4.5
higher than the other estimators, while the MILES models suggest that star
formation ended $\sim$100~Myr ago. The best agreement is seen between the
Tzanavaris et al. estimates for NGC~838 and NGC~839, and our X-ray
estimates. However, the Tzanavaris rates are a factor $\sim$2 higher than
those found by Brassington et al., which should be more accurate. Removing
the contribution of ULXs brings our X-ray estimates into line with the
Brassington et al. estimates for NGC~835 and NGC~838. In NGC~839 our radio
estimate is closest to the Brassington et al. estimate, while in NGC~838
it is closer to the Tzanavaris estimate. Brassington et al. find that both
NGC~833 and NGC~835 have quite low star formation rates, 0.07 and
1.74\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}\ respectively. The star forming ring in NGC~835 is probably
the main cause of the difference between the two galaxies. If the
Brassington et al. estimates are correct, our X-ray estimates (and except
in the case of NGC~838, radio estimates) generally overpredict the current
SFR (particularly when ULXs are included) suggesting that all four original
compact group member galaxies were significantly more active in forming
stars over the past 10$^8$~yr than they are now. This would agree with the
results of optical studies of NGC~838 and NGC~839
\citep{Vogtetal13,Richetal10}.
\subsection{Comparison with other star forming galaxies}
There is an extensive literature on star forming and starburst galaxies and
their winds, and comparison of NGC~838 and NGC~839 with population studies
offers the possibility of better understanding their physical state and
relationship to their environment. The sample of \citet{LiWang13} provides
a good basis for comparison. It consists of 53 nearby, highly inclined disk galaxies
selected to be star formation dominated with little or no AGN contribution,
all of which were observed with \emph{Chandra}.
We follow Li \& Wang in estimating three further parameters: 1) the total
supernova mechanical energy injection rate, $\dot E_{SN}$. This includes type
Ia supernovae, with a rate estimated based on the stellar mass \citep{Mannuccietal05} as well as core collapse supernovae with a rate based on the SFR \citep{Heckmanetal90}, and assumes 10$^{51}$~erg per
supernova; 2) the X-ray radiation efficiency, $\eta$, defined as the
0.5-2~keV gas luminosity divided by $\dot E_{SN}$; 3) the surface rate of core
collapse (CC) supernovae, $F_{SN(CC)}$, defined as the number of supernovae
divided by the area of the galactic disk, determined from its \ensuremath{D_{25}}. We find
values of $\dot E_{SN}$=1.7$\times$10$^{42}$\ensuremath{\erg \ps}\ and $\eta$=2.4\% for
NGC~838, and $\dot E_{SN}$=2.4$\times$10$^{42}$\ensuremath{\erg \ps}\ and $\eta$=0.4\% for
NGC~839. The difference in X-ray radiation efficiency is interesting, and
is clearly caused by the larger mass of hot gas retained by NGC~838. Li \&
Wang show that NGC~838 has a high $\eta$, at the upper end of the range
seen in star-forming or starburst systems, and is among the most X-ray
luminous systems with comparable supernova energy injection rates. NGC~839
is apparently a more typical starburst.
The surface rate of CC supernovae provides information on the ability of
star formation to launch a wind out of the galactic disk.
\citet{Stricklandetal04b} showed that there is a critical value of
$F_{SN(CC)}$=25~SN~Myr$^{-1}$~kpc$^{-2}$ above which superbubble blowout
becomes possible. As expected, both our galaxies fall well above this
limit, with $F_{SN(CC)}\simeq$150~SN~Myr$^{-1}$~kpc$^{-2}$. However,
comparing these values with the $F_{SN(CC)}$:$\eta$ relation established by
Li \& Wang, we again find that NGC~838 is an outlier, with an efficiency
$\eta$ well above that expected for its surface rate of CC supernovae.
NGC~839 falls on the best-fitting relation. Both NGC~838 and NGC~839 have
relatively high gas temperatures for their X-ray luminosity and star
formation rate, at the outer edge of the scatter in these properties across
the galaxy population.
The morphology of the two galactic winds may provide an indication of the
cause of the difference in radiative efficiency between the two. In
NGC~838, the wind appears to be largely confined by the surrounding IGM,
forming two large (possibly leaking) bubbles. Conversely, the wind in
NGC~839 is conical and apparently unconfined. It seems plausible that
because of this confinement, NGC~838 has retained more gas and maintained a
higher X-ray luminosity. NGC~839 has a higher star formation rate, but a
lower hot gas content, suggesting that its wind has been able to escape the
galaxy, mixing with the surrounding IGM and cooling. This difference has
implications for the future development of the two galaxies, with NGC~838
potentially building up an enhanced metallicity and stellar fraction
compared to NGC~839. It is unclear why the NGC~838 wind is confined when
NGC~839 is not. Their specific star formation rates only differ by a factor
1.4-1.8 (depending on the band used to estimate stellar mass). Both
galaxies lie in the densest part of the \ensuremath{\mathrm{H}\textsc{i}}\ filament, and of the hot IGM
(see paper II), and observationally this seems the most likely factor in
determining the difference in their properties.
\subsection{Does NGC~839 / HCG~16D host an active nucleus?}
\label{sec:disc839}
While early optical spectroscopic studies of NGC~839 found evidence of a
LINER or Seyfert nucleus \citep{DeCarvalhoCoziol99}, more recent
integral-field spectroscopy shows that the LINER-like line ratios are found
in an extended region and are more likely to arise from shock-excitation of
gas in the outflowing galactic superwind \citep{Richetal10}. However,
TRP01 classify the galaxy as hosting an absorbed AGN, based
on a spectral fit to the first light \emph{XMM}\ EPIC-MOS spectra which is
dominated by an absorbed powerlaw component at energies above 2~keV. Our
deep \emph{Chandra}\ observations offer an opportunity to resolve this apparent
conflict.
We find that the \emph{Chandra}\ spectra for the galaxy as a whole can be
described either by a thermal model with a single intrinsically absorbed
powerlaw, or by a combination of a thermal component, an unabsorbed
powerlaw representing the X-ray binary population, and an intrinsically
absorbed powerlaw representing an AGN. The two component model is a
marginally better fit than the three component model, but both are good
fits to the data (reduced $\chi^2$=0.991 or 1.039). The powerlaw indices in
both models are consistent with either AGN or X-ray binaries. Since we are
viewing HCG~839 almost edge-on ($i$=67\hbox{$^\circ$}), emission from the central
regions of the galaxy is likely to be absorbed by dust and gas in the disk,
so the intrinsic absorption seen in the two component model could be
consistent with either an AGN or X-ray binaries. As discussed in
Section~\ref{sec:starburst}, we can estimate the star formation rate in the
galaxy over the past $\sim$100~Myr based on the luminosity of the HMXB
population. The two component model implies
SFR=20.09$^{+2.06}_{-1.48}$\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}, while the three component model
suggests SFR=8.74$\pm$0.74\ensuremath{\mathrm{~M_{\odot}~yr^{-1}}}. The lower rate is more consistent with
the estimated rate of Brassington et al., based on the \textit{Spitzer}
24$\mu$m flux. The higher rate is closer to the estimated rate of
Tzanavaris et al., and since NGC~839 appears to have hosted more extensive
star formation in the past few 10$^8$~yr, with star formation rates
declining to their current level, a high X-ray star formation rate estimate
might be expected, owing to a larger population of high-mass X-ray
binaries.
Imaging the galaxy in the 2-7~keV band reveals a $\sim$5$\times$2.5\mbox{\arcm\arcm}\
ellipsoidal distribution aligned with the galaxy disk, with no clear
central point source. To model this hard band emission, we made a 2-7~keV
image from the combined 2013 observations, covering a $\sim$1\hbox{$^\prime$}\ region
centred on the galaxy. We convolved all models with an appropriate exposure
map and PSF. The image is well fitted using a constant background and an
elliptical $\beta$ model to represent the source emission. Best fitting
parameters were $r_{\rm core}$=2.34\mbox{\arcm\arcm}$^{+0.63}_{-0.40}$,
$\beta$=1.09$^{+0.25}_{-0.14}$ and ellipticity=0.47$\pm$0.04. If we add a
point source to represent an AGN and allow its position to vary freely, the
fit rejects the extra component, moving it out of the image. If the
position is fixed at the centre of the $\beta$ model, it provides
12.2$^{+22.4}_{-7.6}$ per cent of the 2-7~keV flux; even with a fixed
position, the point source contribution is consistent with zero at the
2$\sigma$ level.
In summary, the X-ray data alone do not conclusively rule out an AGN in
NGC~839, but suggest that the AGN contribution is probably weak if it is
present. Combined with the probability that the optical LINER emission
arises from shocks in the galactic superwind, this suggests that the
majority of spectrally hard X-ray emission in NGC~839 arises from its X-ray
binary population.
\subsection{Galaxy interactions as the trigger for nuclear activity and star formation}
The presence of the diffuse \ensuremath{\mathrm{H}\textsc{i}}\ filament, combined with the \ensuremath{\mathrm{H}\textsc{i}}\ deficiency estimates, demonstrates that tidal interactions have transported gas out of the four original group members. Based on the morphology, it seems plausible that a series of encounters produced a cloud of \ensuremath{\mathrm{H}\textsc{i}}\ surrounding those four galaxies, which was then drawn out and elongated into a filament by the passage of NGC~848 through the group core. The motion of NGC~833/NGC~835 may also have helped extend the filament to the northwest.
NGC~848 must have passed close to the other galaxies to have produced the
observed \ensuremath{\mathrm{H}\textsc{i}}\ filament, and this raises the question of whether tidal
forces associated with its passage might be responsible for some of the
activity we observe. The estimated ages of the starbursts in NGC~838
\citep[500~Myr]{Vogtetal13} and NGC~839 \citep[400~Myr]{Richetal10} suggest
the possibility that these two starbursts were triggered in sequence, as
NGC~848 passed by from northwest to southeast. In such a scenario, the star
formation in NGC~848 could also have been triggered by these encounters.
It is less clear whether NGC~848 could have triggered star formation in
NGC~833 and NGC~835. The ongoing interaction between NGC~833 and NGC~835 is
probably sufficient to explain their nuclear activity. If we estimate a
velocity in the plane of the sky for NGC~848 based on the star formation
timescales, we find that we would expect it to have passed NGC~833
$\sim$600~Myr ago. A starburst of the magnitude seen in NGC~838 or NGC~839,
but 100~Myr older, should be clearly detected in the SDSS spectrum of
NGC~833, and ought to be obvious from optical colours in NGC~835. However,
NGC~833 and NGC~835 are both $\sim$80\% \ensuremath{\mathrm{H}\textsc{i}}\ deficient and are located in a
lower density region of the \ensuremath{\mathrm{H}\textsc{i}}\ filament than NGC~838 and NGC~839. If both
galaxies lost most of their cold gas before the interaction, it is possible
that any triggered star formation was weak, short-lived, and largely
limited to NGC~835. Alternatively, NGC~848 may not have passed close enough
to them to trigger a starburst.
\section{Conclusions}
\label{sec:conc}
HCG~16 forms an excellent natural laboratory in which to study the effects
of star formation and nuclear activity driven by the tidal interactions
which are common in compact groups. Combining deep, high spatial-resolution
\emph{Chandra}\ X-ray observations with VLA and GMRT radio data, we have examined
the five major galaxy members with the goal of determining the nature of
their present activity and its connection to the group environment. Our
results can be summarised as follows:
\begin{enumerate}
\item We identify eighteen point-like X-ray sources in and around the four original group member galaxies. Three of these are previously known sources associated with galactic nuclei, and one is probably a clump of gas in the starburst superwind of NGC~838. Two sources are located along the line of sight to background galaxies, consistent with our expectation of finding 2-3 background sources in the area of interest. The remaining twelve sources all appear to be associated with the group member galaxies, and five have luminosities $L_{0.5-8}\geq$10$^{39}$\ensuremath{\erg \ps}, making them candidate ultra-luminous X-ray sources (ULXs). One of the brightest point sources is found to be variable at $>$3$\sigma$ significance, is best described by a powerlaw spectral model with flat photon index ($\Gamma$=0.86$\pm$0.23), and is located $<$1\mbox{\arcm\arcm}\ from the optical centroid of NGC~838 (HCG~16C). We therefore suggest that this source may be a previously unidentified active nucleus.
\item We examine the X-ray spectra of NGC~833 and NGC~835 (HCG~16B and A)
and confirm the previous finding that they host obscured AGN. SSP
modelling of the SDSS spectrum of NGC~833 shows that little or no recent
star formation has occurred in the galaxy. We use the excellent spatial
resolution of \emph{Chandra}\ to separate nuclear and galactic emission, and
find that powerlaw emission previously thought to be scattered from the
nuclei is more likely to arise from the X-ray binary populations of the
two galaxies. The AGN components of both galaxies are found to be
variable on timescales of months to years (between \emph{Chandra}\
observations), and we find evidence that this variability is best
modelled as a change in the intrinsic luminosity of the AGN, and
therefore probably of the accretion rate, rather than as a change in the
optical depth of the absorber. We identify an Fe-K$\alpha$ emission line
in NGC~835 with energy 6.41$\pm$0.05~keV and width 100$\pm$50~eV.
\item We find that NGC~838 and NGC~839 (HCG~16C and D) are probably both
starburst-dominated systems, with minimal nuclear activity. Whereas
previous X-ray studies have found evidence of an obscured AGN in NGC~839,
in conflict with optical spectroscopy suggesting that LINER emission is
produced by shocks in the galactic superwind, we find that the spectrally
hard X-ray emission component is likely produced by the X-ray binary
population of the galaxy and absorbed by gas and dust in the galactic
disk. Both galaxies have mean temperatures $\sim$0.8~keV, and in
NGC~838 we are able to constrain the abundance to
0.16$^{+0.08}_{-0.04}$\ensuremath{\mathrm{~Z_{\odot}}}. However, this is likely biased low owing to
mass loading of the wind with cold gas, and the wide range of
temperatures in the hot phase of the wind. We examine the X-ray
temperature structures of the winds and estimate the mass of hot gas
involved and the rates of outflow. We also estimate the star formation
rates of the galaxies from both X-ray and radio luminosities, and compare
these to infra-red and ultra-violet estimates, finding that the X-ray
rates probably better represent the SFRs in the recent past. The NGC~838
wind appears to be confined by the surrounding \ensuremath{\mathrm{H}\textsc{i}}\ filament and/or IGM,
retaining more of its gas than has NGC~839. This has the potential to
affect the development of the galaxy, with NGC~838 likely to have a
higher metallicity and stellar fraction than NGC~839 in future.
\item We examine the X-ray and radio properties of NGC~848, the fifth largest galaxy in the group, using data from \emph{Chandra}\ cycle~1, in which the galaxy falls $\sim$15\hbox{$^\prime$}\ off-axis. We find that the X-ray emission follows the optical bar and is best modelled by a simple powerlaw, suggesting that this starburst galaxy is dominated by emission from its X-ray binary population. The radio emission is more centrally concentrated, and may arise either from star formation or an AGN.
\end{enumerate}
\acknowledgements The authors thank L. Verdes-Montenegro for providing her
VLA \ensuremath{\mathrm{H}\textsc{i}}\ map of the group, as well as C. Jones and the anonymous referee
for their helpful comments on the manuscript. EO'S thanks R. Barnard for
useful discussions on data analysis. MT acknowledges the support of the Funda\c{c}\~{a}o de Amparo \`{a} Pesquisa do Estado de S\~{a}o Paulo (FAPESP),
process no. 2012/05142-5. Support for this work was provided by the
National Aeronautics and Space Administration (NASA) through Chandra Award
Number G03-14143X issued by the Chandra X-ray Observatory Center (CXC),
which is operated by the Smithsonian Astrophysical Observatory (SAO) for
and on behalf of NASA under contract NAS8-03060. SG acknowledges the
support of NASA through the Einstein Postdoctoral Fellowship PF0-110071
awarded by the CXC, and this research has made use of data obtained from
the Chandra Data Archive and software provided by the CXC in the
application packages CIAO, ChIPS, and Sherpa, as well as SAOImage DS9,
developed by SAO. We thank the staff of the GMRT for their help during
observations. GMRT is run by the National Centre for Radio Astrophysics of
the Tata Institute for Fundamental Research. We acknowledge the usage of
the HyperLeda database (http://leda.univ-lyon1.fr). Funding for SDSS-III
has been provided by the Alfred P. Sloan Foundation, the Participating
Institutions, the National Science Foundation, and the U.S. Department of
Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
\textit{Facilities:} \facility{CXO} \facility{VLA} \facility{GMRT}
\bibliographystyle{apj}
|
\section{Introduction}
The past few years have witnessed unprecedented expansion of commercial computing operations as the idea of cloud computing has become more mainstream and widely adopted by forward thinking technical organizational leadership. This rate of adoption promises to increase in the near future as well. With this expansion has come opportunity as well as risk, embodied by recent major service outages at leading cloud providers like Amazon. These issues promise to become more difficult to control as managed infrastructure expands. This expansion will simply not be possible without large amounts of automation in all aspects of cloud computing systems.
The current state of the art in cloud systems is poorly differentiated and not as customer-focused as it could be. Current providers place the responsibility of monitoring performance and proving outages on the consumer \cite{ctrl:amazon-cloud-watch}. Furthermore, providers as a whole usually provide one type of service level agreement (SLA) in a loosely-defined one-size-fits-all type of arrangement. Current SLAs are also very difficult to evaluate and manage \cite{Hilley-2009}. This provides strong differentiating opportunities for smaller, second generation cloud system providers who have established the technology required to scalably manage multiple, competing SLAs on the same infrastructure in tandem with clear customer system visibility.
These second generation providers will rely on automated infrastructure management in order to scale. One of the first steps toward automating these systems is automating SLA management and compliance. Likewise, future cloud users are likely to build systems spanning multiple cloud providers, providing more difficult system management scenarios from their perspective as well \cite{ctrl:lamb-MCCCS}.
Herein, we elaborate the idea of applying usage management to a single system governed by multiple different types of SLAs. We will define the problem, formally describe SLAs, analyze the implications of that formality with an eye toward efficient computability and the implications thereof.
In Section \ref{sec:cloud-models}, this paper begins by describing the different types of cloud computing models that generally exist today and how they manage services. Immediately thereafter, we propose a possible future model in which users can have unique SLAs that more closely fit their needs rather than shoehorning their computing needs into a previously configured contract. Then, in Section \ref{sec:SLA-defined}, we more formally define an SLA, and show how to convert one to an evaluatable expression. In the following section we analyze the new SLA model and extract specific theoretical limits on computability and discuss implications thereof, showing solutions to SLAs in general to be NP-Complete.
\subsection{Previous Work}
As cloud computing is emerging as the future of utility systems hosting for consumer-facing applications. In these kinds of systems, components, applications, and hardware are provided as utilities over the Internet with associated pricing schemes pegged by system demand. Users accept specific Quality-of-Service (QoS) guidelines that providers use to provision and eventually allocate resources. These guidelines become the basis over which providers charge for services.
Over the past few years multiple service-based paradigms such as web services, cluster computing and grid computing have contributed to the development of what we now call cloud computing \cite{Bu:09}. Cloud computing distinctly differentiates itself from other service-based computing paradigms via a collective set of distinguishing characteristics: market orientation, virtualization, dynamic provisioning of resources, and service composition via multiple service providers \cite{BuYeVeBrBr:09}. This implies that in cloud computing, a cloud-service consumer's data and applications reside inside that cloud provider's infrastructure for a finite amount of time. Partitions of this data can in fact be handled by multiple cloud services, and these partitions may be stored, processed and routed through geographically distributed cloud infrastructures. These activities occur within a cloud, giving the cloud consumer an impression of a single virtual system. These operational characteristics of cloud computing can raise concerns regarding the manner in which cloud consumer's data and applications are managed within a given cloud. Unlike other computing paradigms with a specific computing task focus, cloud systems enable cloud consumers to host entire applications on the cloud (i.e. software as a service) or to compose services from different providers to build a single system. As consumers aggressively start exploiting these advantages to transition IT services to external utility computing systems, the manner in which data and applications are handled within those systems by various cloud services will become a matter of serious concern \cite{Jamkhedkar:2010:IUM:1866870.1866885}.
\section{Cloud System Models}\label{sec:cloud-models}
\noindent Current cloud systems do not ignore SLA restrictions; rather, they are designed from the ground up to support a single type of SLA. That SLA generally encompasses total system uptime and some kind of response time metric \cite{ctrl:amazon-sla,ctrl:rackspace-sla}. If for some reason the cloud provider can no longer adhere to the terms outlined, some kind of compensation strategy generally applies to affected customers. Future cloud providers can very well use the ability to support multiple SLAs as a way to differentiate available products from competitors.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{cloud-current}
\caption{Single SLA with Control Elements}
\label{fig:current-cloud-model}
\end{figure}
\subsection{Current Model}
Current systems like Amazon's EC2 or Rackspace products are designed around high availability, and this is reflected in the focus of their supplied SLAs. This common design focus is also evident in the artifacts generated by other vendors \cite{ctrl:google-arch}. Furthermore, Amazon offers clear guidance on how to develop systems that take advantage of their robust architecture as well as services that provide some measure of automatic scaling \cite{ctrl:amazon-best-practice,ctrl:amazon-fault-tolerant}. This combination of market leading position and products and the extensive supplied guidance make Amazon a clear choice to examine when reflecting on the current state-of-the-art.
Amazon's Cloud Watch products used in tandem with Auto Scaling provide the ability to control the number of deployed instances in response to specific system loads, as shown in Figure \ref{fig:current-cloud-model} \cite{ctrl:amazon-cloud-watch,ctrl:amazon-auto-scale}. Cloud Watch gives customers the ability to monitor various system performance metrics for their virtual machines, including but not limited to latency, processor use, and request counts. Furthermore, users can set resource levels at which additional EC2 instances are created or destroyed. This provides some level of personalized management and control over deployed systems within Amazon's cloud infrastructure.
\subsection{Future Reference Model}
While current cloud service providers focus on a single QoS metric, future providers may very well begin to provide multiple metrics over which they will define service levels, as shown in Figure \ref{fig:future-cloud-model}. This is not without precedent --- just as airlines provide the same essential product at different service points, cloud providers could supply system hosting via disparate service levels, including divergent service metric definitions. For example, current architectures support uptime and availability as the primary managed metric from an SLA perspective. Future architectures could support uptime and availability, as well as specific latency, bandwidth, and geo-location sensitive hosting parameters. These kinds of SLAs would also continue to outline penalties when any of the conditions of that SLA were violated. Unlike current SLAs however, these could also differentiate based on the magnitude of the imposed penalty, with different classifications of service mapping to increasingly large penalties on service failure.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{cloud-future}
\caption{Multiple SLA Architectural Integration}
\label{fig:future-cloud-model}
\end{figure}
While industry does seem to certainly be trending in this direction, as indicated by the development of tools supporting user-centric infrastructure monitoring and management, this kind of control is not yet embedded into contracts of any kind, much less agreements that are machine-readable. Furthermore, this kind of management is still manual and cannot scale to the levels needed to manage Internet-scale systems.
\section{Service Level Agreements Defined}\label{sec:SLA-defined}
As we have seen, SLAs generally consist of a set of conditions of use under which the SLA is binding, a set of obligations that the provider will adhere to if the customer adheres to the set conditions, and two sets of penalties, one penalizing the provider when breaching obligations, and another penalizing the customer when breaching conditions of use. Conditions are generally loosely defined, while provider obligations are much more rigorously constructed. Generally however, conditions and obligations in this context can be viewed as defined by {\it objectives} which are measured by {\it indicators}. In the case of provider-centric obligations, these are commonly defined as service level indicators (SLIs) and objectives (SLOs).
With this general understanding of SLAs and related SLIs and SLOs, we can create a non-specific definition of an SLA as a list of tuples related via boolean operators including $($, $)$, $\lor$, $\land$, and $\lnot$. Each tuple is a clause in the SLA:
\begin{eqnarray}
CLAUSE = \lbrace (I,O,E,P) \rbrace, \\
OPERATOR = \lbrace ( , ) , \lor, \land, \lnot \rbrace \\
SLA = \lbrace ( o_{0}, c_{0}, o_{1}, c_{1}, ..., c_{n}, o_{n + 1} ) \rbrace , \\ c \in CLAUSE, o \in OPERATOR \notag
\end{eqnarray}
Where $ I $ is an indicator function, $ \forall i \in I, i : () \rightarrow \tau $, which retrieve indicator values, generally related to some kind of SLI or customer condition indicator. Parameter $ O $ is a set of values derived from SLOs or customer condition objectives such that $ \forall o \in O, o : P(\tau) $. Parameter $ E $ is a set of predicates that indicate whether a specific indicator complies with its objectives, where $ \forall e \in E, e : ( () \rightarrow \tau ) \times P(\tau) \rightarrow bool $. Parameter $ P $ is a set of penalty functions, $ \forall p \in P, p : Z \rightarrow Z $, where the first argument is generally an elapsed time value.
For example, say we are a customer of Nimbus Cloud Corporation, and we have an SLA in which Nimbus provides guaranteed 100\% uptime and packet latency between 300 and 650 milliseconds. Nimbus does not have any customer conditions specified within its SLAs. This then gives us a machine evaluateable SLA:
\begin{align}
SLA_{nimbus} = ( ( uptime\_monitor() : bool, \notag\\
\lbrace true \rbrace, \notag \\
uptime\_evaluator( monitor : () \rightarrow bool, \lbrace true \rbrace) : bool, \notag \\
uptime\_penalty\_evaluator( T : Z ) : Z ), \notag \\
\land, \notag \\
( latency\_monitor() : Z, \notag \\
\lbrace 300, 650 \rbrace, \notag \\
latency\_evaluator( monitor : () \rightarrow Z, \lbrace 300, 650 \rbrace) : bool, \notag \\
latency\_penalty\_evaluator( T : Z ) : Z) ) \notag
\end{align}
This more rigorous SLA allows users to monitor obligations and determine penalties when triggered. Furthermore, when structuring SLAs in this way, we are able to combine any single SLA with other SLAs but extracting the clauses and creating a new aggregate SLA, which is just the conjunction of the combined SLAs, such that for $ A, B, C \in SLA $, $ ( A \land B \land C ) \in SLA $. We will further define SLA-SAT as the problem of establishing whether a given SLA or group of SLAs have a solution.
Also note that evaluators, as they compare the output of a monitor to an input value, can run in constant time. Monitors, as they retrieve specific contextual values from a running system, execute over some assumed constant time not influenced by the SLA.
\section{Evaluating and Verifying Service Level Agreements}\label{sec:SLA-analysis}
Now that we have rigorously defined our SLAs, notice that the SLA evaluation functions are predicates, and can be curried for later execution if needed. This allows us to begin a more fundamental analysis of SLAs and their capabilities.
\subsection{Computational and Space Complexity}
In Section \ref{sec:SLA-defined}, we defined an SLA to essentially be a sequence of evaluatable predicates. These evaluatable predicates are related in some way; currently, an SLA is the conjunction of these predicates. As these predicates can be created prior to evaluation, and at evaluation time require no specific arguments once appropriately curried, we can define these predicates as boolean {\it terms}. Ergo, once we have created a group of predicates and transformed them into terms, we are evaluating an arbitrary boolean equation - in other words, we are verifying an instance of the Boolean Satisfiability Problem, or SAT.\\
\noindent {\bf Claim 1:} SLA-SAT is in {\bf NP}.\\
\noindent {\bf Proof:} The following verifier runs in polynomial time in the length of an SLA-SAT.\\
\noindent {\sl Input}: $ \sigma \in $ SLA-SAT, a verifier $v_{SAT}$ for SAT. \\
\noindent {\sl Output}: TRUE if the SLA is satisfied FALSE otherwise.
\begin{tabbing}
~~~~\=~~~~\=~~~~\=~~~~\= \\
\noindent Verifier $V_{SLA-SAT} \langle \sigma, v_{SAT} \rangle$ : \\
\> $ l \leftarrow $ list \\
\> \underline{for} each element $ e \in \sigma $ :\\
\>\> \underline{if} e is a clause : \\
\>\>\> extract monitor $ m \in M $ from $ c $\\
\>\>\> extract objective $ o \in O $ from $ c $\\
\>\>\> extract evaluator $ v \in E $ from $ c $\\
\>\>\> $ r \leftarrow v ( m, o ) $\\
\>\>\> add $ r $ to $ l $ \\
\>\> \underline{else} : \\
\>\>\> add $ e $ to $ l $ \\
\>\> \underline{endif} \\
\> \underline{endfor} \\
\> \underline{return} $ v_{SAT}(l) $ \\
\end{tabbing}
\noindent We can furthermore establish that SLA-SAT is NP-Hard, establishing NP-Completeness.\\
\noindent {\bf Claim 2:} SAT $ \leq_{p} $ SLA-SAT.\\
\noindent {\bf Proof:} Define a function $ f : ( \beta ) \rightarrow \sigma $, where $ \beta $ is a generalized boolean formula consisting of boolean clauses related by the logical operators $($, $)$, $\land$, $\lor$, and $\lnot$ and $ \sigma \in $ SLA-SAT. $ f $ will parse through $ \beta $, skipping operators, and transforming boolean elements $ \zeta $ into tuples $ t \in CLAUSE $ such that $ t = ( null, null, f : ( () \rightarrow \tau \times P(\tau) \rightarrow bool ) \rightarrow \zeta, null ) $. $ f $ then adds the operator or transfurmed boolean element to a list. When $ f $ reaches the end of the boolean formula, it returns the list. \\
\noindent We can also show the inverse.\\
\noindent {\bf Claim 3:} SLA-SAT $ \leq_{p} $ SAT.\\
\noindent {\bf Proof:} Define a function $ g : ( \sigma ) \rightarrow \beta $, where $ \beta $ is a generalized boolean formula consisting of boolean clauses related by the logical operators $($, $)$, $\land$, $\lor$, and $\lnot$ and $ \sigma \in $ SLA-SAT. Evaluate each element within $\sigma$, copying each operator into a boolean expression $\beta$, and transforming each clause by evaluating $e \in E$ with contained indicators and objectives $i \in I$ and $o \in O$. Copy the resulting boolean value into the expression $\beta$. When finished traversing $\sigma$, return $\beta$. \\
Therefore, as SAT is NP-Complete, and provably difficult to solve, SLA-SAT is NP-Complete as well \cite{comptheory:sipser:intro-comp-theory}. However, 3SAT, a subset of SAT, is equally difficult, while 2SAT is not. 2SAT is firmly in the computational class P; in fact, 2SAT is NL-Complete as well, so we know it is solvable in an amount of space logarithmic in the number of boolean terms \cite{comptheory:papadimitriou:computational-complexity}; it is widely believed that both SAT and 3SAT cannot be solved in logarithmic or less space. Finally, as 2SAT is NL-Complete, we also know it is contained within NC$^{2}$, and as such is highly parallelizeable \cite{comptheory:papadimitriou:computational-complexity}. This implies that SLAs should be implemented or aggregated in a form no more complex than 2SAT to facilitate efficient processing.
\subsection{Verification v. Solution}
General SLA use focuses on verification rather than solution. That is to say, with respect to a given SLA, both the user and provider is more concerned with whether the system is currently compliant with all SLA terms. In future use, this may very well no longer be the case. For example, imagine a cloud system from the user's perspective that spans multiple cloud providers; a single general purpose provider for general computing, data storage, and queuing, a specific-use provider for an unique set of algorithms of some kind (say, market modeling algorithms), and finally a Content Delivery Network (CDN) provider. Each of these providers have a unique SLA with multiple conditions. In this particular case, the user may need to know if the given system can work together at all under the terms of the SLAs, and if so, under what conditions. As the user has combined all the SLAs composing the system and is attempting to find some combination of terms that satisfies the resulting boolean formula, the user is in essence attempting to find a solution for this instance of $ SAT $, a known NP-Complete problem. Likewise, a cloud provider may need to solve similar problems where the SLAs at issue are the whole of SLAs issued to the entire provider's customer base.
Nontrivial generalized SLAs may be too difficult to solve effectively without using some kind of approximate heuristic or $ SAT $ Solver \cite{Hochbaum:1996:AAN:241938,ctrl:satcompetition}. If these SLAs are formulated in at most a $ 2SAT $ style problem however, they are suddenly much more tractable, easier to work with, and amenable to efficient solution. Keep in mind, even generalized SLAs can be efficiently verified.
\section{Conclusions and Future Works}
\noindent Herein, we started by going over the current generalized state of most cloud systems from an SLA perspective, differentiating between current architectures that incorporate SLA ideas into the design itself with possible futures architectures that incorporate pluggable SLAs with varying indicators and objectives. We then generalized SLAs into sets of quadruples containing a monitoring function, a set of values defining acceptable ranges returned from the monitoring functions, an evaluation function, and a penalty evaluating function, and demonstrated how this formulation could be used with a specific example. With this in place, we then demonstrated that the generalized SLA problem is equivalent to SAT, and therefore is NP-Complete. We finally covered the implications and theoretical limits implied by this NP-Completeness, validating the applicability of this work by designing a realistic control model using these ideas.
This is preliminary work into establishing the theoretical bounds surrounding effective automated control of cloud systems within Internet-scale systems. Furthermore, the SLA modeled was fairly simple, and only took into account externally-evaluatable metrics in a black-box arrangement. SLAs can very well outline other operational parameters, like specific data routing, fine-grained usage management, or encryption requirements. Likewise, experimental evidence supporting these control ideas will be vital to promoting acceptance and action around these concepts within both cloud service provider systems and user configured applications.
\bibliographystyle{apalike}
{\small
|
\section{Introduction}
It is common practice in economics and statistics to conduct inference that is robust to within-cluster dependence. Examples of such clusters are households, classrooms, firms, cities, or counties. We have to expect that units within these clusters influence one another or are influenced by the same sociological, technical, political, or environmental shocks. To account for the presence of data clusters, the literature frequently recommends inference using cluster-robust versions of the bootstrap; see, among many others, \citet{bertrandetal2004} and \citet{cameronetal2008} for an overview in the context of linear regression models estimated by least squares. In this paper I develop a bootstrap method for cluster-robust inference in linear quantile regression (QR) models. The method, which I refer to as \emph{wild gradient bootstrap}, is an extension of a wild bootstrap procedure proposed by \citet{chenetal2003}. Despite the fact that it involves resampling the QR gradient process, the wild gradient bootstrap is fast and easy to implement because it does not involve finding zeros of the gradient during each bootstrap iteration. I show that the wild gradient bootstrap allows for the construction of asymptotically valid bootstrap standard errors, hypothesis tests both at individual quantiles or over ranges of quantiles, and confidence bands for the QR coefficient function.
Quantile regression, introduced by \citet{koenkerbassett1978}, has become an important empirical tool because it enables the researcher to quantify the effect of a set of covariates on the entire conditional distribution of the outcome of interest. This is in sharp contrast to conventional mean regression methods, where only the conditional mean can be considered. A disadvantage of QR in comparison to least squares methods is that the asymptotic variance of the QR coefficient function is notoriously difficult to estimate due to its dependence on the unknown conditional density of the response variable. An analytical estimate of the asymptotic variance therefore requires a user-chosen kernel and bandwidth. Hence, two researchers working with the same data can arrive at different conclusions simply because they used different kernels or bandwidths. Furthermore, a common concern in applied work is that analytical estimates of asymptotic variances perform poorly in the cluster context when the number of clusters is small or the within-cluster correlation is high; see, e.g., \citet{bertrandetal2004} and \citet{webb2013} for extensive Monte Carlo evidence in the least squares case. Their overall finding is that true null hypotheses are rejected far too often. Simulations in \citet{mackinnonwebb2014} suggest that similar problems also occur when the cluster sizes differ substantially.
I show that the wild gradient bootstrap is robust to each of these concerns: It performs well even when the number of clusters is small, the within-cluster dependence is high, and the cluster sizes are heterogenous. The wild gradient bootstrap consistently estimates the asymptotic distribution and covariance functions of the QR coefficients without relying on kernels, bandwidths, or other user-chosen parameters. As such, this paper complements recent work by \citet{parentesantossilva2013}, who provide analytical, kernel-based covariance matrix estimates for QR with cluster data. Their estimates have the advantage that they are simpler to compute than the bootstrap procedures presented here. However, as the Monte Carlo study in this paper shows, a drawback is that tests based on their covariance matrix estimates can suffer from severe size distortions in the same situations as those described for the mean regression case above. In addition, \citeauthor{parentesantossilva2013}'s method does not allow for uniform inference across quantiles because the limiting QR process generally has an analytically intractable distribution. In contrast, the bootstrap approximations of the distribution and covariance functions developed here can be combined to perform uniform Wald-type inference about the QR coefficient function.
The wild bootstrap procedure discussed in this paper was first introduced by \citet{chenetal2003} as a way to construct confidence intervals for QR coefficients at a single quantile. However, they only provide heuristic arguments for the consistency of the bootstrap approximation and note that ``as far as [they] know, there is no analytical proof that the bootstrap method is valid for the general quantile regression model with correlated observations.'' I considerably extend the scope of their method under explicit regularity conditions to allow for inference on the entire QR process and uniformly consistent covariance matrix estimates of that process. Some parts of the proofs below rely on a recent result by \citet{kato2011} regarding the convergence of bootstrap moments. In turn, his results build on a technique developed in \citet{alexander1985} and tail bounds for the empirical process given in \citet{vandervaartwellner1996}. I also utilize empirical process results of \citet{pollard1991b} and \citet{kosorok2003} to address some nonstandard issues arising from the fact that I allow clusters to be heterogeneous both in terms of their size and their distributions.
\changes{Other types of wild bootstrap procedures for QR are given in \citet{fengetal2010} and \citet{davidson2012}. They do not deal with cluster data but their methods are likely to generalize in this direction. Although typically only convergence of the bootstrap distribution is shown, these and other bootstrap methods have been suggested in the literature as ways to construct bootstrap standard errors or, more generally, bootstrap covariance matrix estimates. \citet{hahn1995} and \citet{goncalveswhite2005} explicitly caution against such conclusions because convergence in distribution does not imply convergence of moments without uniform integrability conditions. This paper establishes these conditions for QR estimates in the cluster context. As I show in my Monte Carlo study, the availability of a bootstrap covariance matrix estimate is crucial for bootstrap tests to have good size and power in many empirically relevant situations.}
Cluster-robust inference in linear regression has a long history in economics; see \citet{cameronmiller2014} for a recent survey. \citet{kloek1981} is an early reference. \citet{moulton1990} highlights the importance of correcting inference for within-cluster correlation when covariates do not vary within clusters. However, apart from \citet{chenetal2003} and \citet{parentesantossilva2013}, cluster inference in QR models has not received much attention. Notable exceptions are \citet{wanghe2007} and \citet{wang2009}, who develop methods for cluster-robust quantile rank score inference, \citet{chetverikovetal2013}, who introduce a method for instrumental variables estimation in a QR model with cluster-level treatment, and \citet{yoongalvao2013}, who discuss QR in a panel model where clusters arise from correlation of individual units over time.
The paper is organized as follows: Section \ref{s:qreg} states and discusses several assumptions that are then used to establish the large sample distribution of the QR estimator with cluster data. Section \ref{s:boot} introduces the wild gradient bootstrap procedure and shows how it can be applied to conduct bootstrap inference on the QR process. Section \ref{s:mc} illustrates the finite-sample behavior of the wild gradient bootstrap in three Monte Carlo experiments. Section \ref{s:star} contains a brief application of the proposed bootstrap procedure to Project STAR data. Section \ref{s:conc} concludes. The appendix contains auxiliary results and proofs.
I use the following notation throughout the paper: $|\cdot|$ is Euclidean norm and $1\{\cdot\}$ is the indicator function. Limits are as $n\to\infty$ unless otherwise noted and convergence in distribution is indicated by $\leadsto$.
\section{Quantile Regression with Clustered Data}\label{s:qreg}
This section discusses linear QR with clustered data and outlines the basic assumptions used to justify asymptotic inference. Then I establish weak convergence of the QR estimator in the cluster context.
Regression quantiles in a framework with cluster data express the conditional quantiles of a response variable $Y_{ik}$ in terms of an observed covariate vector $X_{ik}$. Here $i$ indexes clusters and $k$ indexes observations within that cluster. There are $n$ clusters and cluster $i$ has $c_i$ observations. The cluster sizes $c_i$ need not be identical across $i$, but will be taken to be small relative to $n$ for the asymptotic theory. Because there is typically no natural ordering of observations in the same cluster (unless $k$ indexes time), I allow for arbitrary within-cluster dependence of the data.
\begin{assumption}\label{as:data}
For all $i,j \geq 1$ with $i\neq j$ and all $1\leq k\leq c_i$, $1\leq l\leq c_j$, the random vectors $(Y_{ik}, X_{ik}^\top)^\top$ and $(Y_{jl}, X_{jl}^\top)^\top$ are independent. The cluster sizes $c_i$ are bounded by some $c_{\max} <\infty$ uniformly in $i\geq 1$.
\end{assumption}
The $\tau$th quantile function of $Y_{ik}$ conditional on $X_{ik}=x$ is given by $Q_{{ik}}(\tau\mid x) := \inf\{y: \prob(Y_{ik}\leq y\mid X_{ik} = x)\geq \tau\}$, where $\tau\in(0,1)$. I assume that the linear QR framework is an appropriate model for the data.
\begin{assumption}\label{as:model}
For $\{Y_{ik} : i\geq 1, 1\leq k\leq c_i\}$, the $\tau$th quantile function satisfies $$Q_{{ik}}(\tau\mid x) = x^\top \beta(\tau),\qquad x\in\mathcal{X}_{ik}\subset \mathbb{R}^d,~\tau\in\mathit{T},$$ where $\mathcal{X}_{ik}$ is the support of $X_{ik}$ and $\mathit{T}$ is a closed subset of $(0,1)$. For all $\tau\in\mathit{T}$, $\beta(\tau)$ is contained in the interior of a compact and convex set $\mathit{B}\subset\mathbb{R}^d$.
\end{assumption}
\begin{comments}
\begin{inparaenum}[(i)]
\item Because $\tau\mapsto\beta(\tau)$ does not depend on $i$, this assumption implicitly rules out cluster-level ``fixed effects'' as they would lead to incidental parameter problems; see \citet{koenker2004}. It does \emph{not} rule out covariates that vary at the cluster level and, more importantly, fixed effects for levels above the cluster level. For example, the application in Section \ref{s:star} has classroom-level clusters and school-level fixed effects. There are several ways to address the incidental parameters problem when more is known about the dependence structure in the data; see \citet{yoongalvao2013} and the references therein.
\item The assumption of compactness of $\mathit{B}$ has no impact on the estimation of the QR model in practice because $\mathit{B}$ can always be viewed as large. Compactness is, however, essential for the validity of bootstrap moment estimates in the QR context; see the discussion below Theorem \ref{th:bootse} in the next section.
\end{inparaenum}
\end{comments}
Estimates of the unknown QR coefficient function $\tau\mapsto \beta(\tau)$ can be computed with the help of the \citet{koenkerbassett1978} check function $\rho_\tau(z) = (\tau - 1\{z < 0\})z$. For clustered data, the QR problem minimizes
\begin{equation*}
\mathbb{M}_{n}(\beta,\tau) := \frac{1}{n}\sum_{i=1}^n\sum_{k=1}^{c_i}\rho_{\tau}(Y_{ik} - X_{ik}^\top \beta
\end{equation*}
with respect to $\beta$ for a given $\tau$ so that $\tau \mapsto \beta(\tau)$ is estimated by
\begin{equation*}
\tau\mapsto \hat{\beta}_n(\tau) := \argmin_{\beta\in\mathit{B}}\mathbb{M}_{n}(\beta,\tau), \qquad \tau\in\mathit{T}.
\end{equation*}
The main goal of this paper is to provide a method for cluster-robust bootstrap inference about the QR coefficient function that is valid uniformly on the entire set $\mathit{T}$ and leads to cluster-robust bootstrap covariance matrix estimates for $\tau\mapsto \hat{\beta}_n(\tau)$. The validity of this method relies on the asymptotic normal approximation to the distribution of $\sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau))$, which I turn to next.
Asymptotic normality of the QR estimates in the cluster context requires straightforward extensions of smoothness and moment conditions familiar from the iid case. The following assumption allows for arbitrary heterogeneity of the clusters as long as the observations within these clusters satisfy mild restrictions on the smoothness of their conditional distributions and the tail behavior of the covariates. This assumption is needed to ensure identification of the QR coefficient function and to justify an approximation argument following immediately below.
\begin{assumption}\label{as:smooth}
\begin{compactenum}[\upshape(i)]
\item\label{as:smooth:mom} $\ev |X_{ik}|^q< \infty$ uniformly in $i\geq 1$ and $1\leq k\leq c_i$ for some $q > 2$,
\item $n^{-1} \sum_{i=1}^n \sum_{k=1}^{c_i} \ev X_{ik}X_{ik}^\top$ is positive definite, uniformly in $n$,
\item\label{as:smooth:den} the conditional density $f_{{ik}}(y\mid X_{ik}=x)$ of $Y_{ik}$ and its derivative in $y$ are bounded above uniformly in $y$ and $x\in\mathcal{X}_{ik}$, uniformly in $i\geq 1$ and $1\leq k\leq c_i$, and
\item\label{as:smooth:der} $f_{{ik}}(x^\top\beta \mid X_{ik}=x)$ is bounded away from zero uniformly in $\beta\in\mathit{B}$ and $x\in\mathcal{X}_{ik}$, uniformly in $i\geq 1$ and $1\leq k\leq c_i$
\end{compactenum}
\end{assumption}
To establish distributional convergence of the QR estimator, I consider the recentered population objective function $\beta\mapsto M_n(\beta,\tau) := \ev (\mathbb{M}_{n}(\beta,\tau) - \mathbb{M}_{n}(\beta(\tau),\tau))$. The recentering ensures that $M_n$ is well defined without moment conditions on the response variable. Provided Assumptions \ref{as:model} and \ref{as:smooth} hold, the map $\beta\mapsto M_n(\beta,\tau)$ is differentiable with derivative $M_n'(\beta, \tau) := \partial M_n(\beta,\tau) / \partial \beta^\top$ and achieves its minimum at $\beta(\tau)$ by convexity. I show in the appendix that under Assumptions \ref{as:data}-\ref{as:smooth} we can write
\begin{equation}\label{eq:taylorex}
0 = \sqrt{n} M_n'(\beta(\tau), \tau) = \sqrt{n} M_n'(\hat{\beta}_n(\tau), \tau) - J_n(\tau)\sqrt{n}\bigl(\hat{\beta}_n(\tau) - \beta(\tau)\bigr) + o_{\prob} (1)
\end{equation}
uniformly in $\tau\in\mathit{T}$ by a mean value expansion about $\hat{\beta}_n(\tau)$. Here
$J_n(\tau)$ is the Jacobian matrix of the expansion evaluated at $\beta(\tau)$,
$$ J_n(\tau) = \frac{1}{n} \sum_{i=1}^n \sum_{k=1}^{c_i} \ev f_{ik}\bigl( X_{ik}^\top\beta(\tau) \mid X_{ik}\bigr) X_{ik}X_{ik}^\top.$$
After rearranging \eqref{eq:taylorex}, a stochastic equicontinuity argument (see the appendix for details) can be used to show that $\sqrt{n} M_n'(\hat{\beta}_n(\tau), \tau)$ is, uniformly in $\tau\in\mathit{T}$, within $o_{\prob} (1)$ of the first term on the right of
\begin{equation}
J_n(\tau) \sqrt{n}\bigl(\hat{\beta}_n(\tau) - \beta(\tau)\bigr) = \frac{1}{\sqrt{n}} \sum_{i=1}^n \sum_{k=1}^{c_i} \psi_\tau\bigl(Y_{ik}-X_{ik}^\top\beta(\tau)\bigr) X_{ik} + o_{\prob}(1),\label{eq:bahadur}
\end{equation}
where $\psi_\tau(z) = \tau - 1\{z < 0\}$. The outer sum on the right-hand side can be viewed as an empirical process evaluated at functions of the form $\sum_{k=1}^{c_i} \psi_\tau(Y_{ik}-X_{ik}^\top\beta(\tau)) X_{ik}$ indexed by $\tau\in\mathit{T}$.\footnote{In view of Assumption \ref{as:data}, we can always take $(Y_{ik},X_{ik}^\top)^\top=0$ for $c_i<k\leq c_{\max}$ whenever $c_i < c_{\max}$ to make this a well-defined class of functions from $\mathbb{R}^{c_{\max}}\times\mathbb{R}^{d\times c_{\max}}$ to $\mathbb{R}^d$.} This empirical process has covariances
$$\Sigma_n(\tau,\tau') := \frac{1}{n} \sum_{i=1}^n \sum_{k=1}^{c_i}\sum_{l=1}^{c_i} \ev \psi_\tau\bigl(Y_{ik}-X_{ik}^\top\beta(\tau)\bigr) \psi_{\tau'}\bigl(Y_{il}-X_{il}^\top\beta(\tau')\bigr)X_{ik}X_{il}^\top, \qquad \tau,\tau'\in\mathit{T}.$$
\phantomsection%
\label{rev:liangzeger}%
\changes{As a referee points out, similar covariances arise in the generalized estimating equations framework of \citet{liangzeger1986}.}
In the absence of clusters (i.e., $c_i\equiv 1$), $\Sigma_n(\tau,\tau')$ reduces to the familiar form of $n^{-1}\sum_{i=1}^n \ev X_{i1}X_{i1}^\top$ times the covariance function $(\min\{\tau,\tau'\}-\tau\tau')I_d$ of the standard $d$-dimensional Brownian bridge, where $I_d$ is the identity matrix of size $d$. Because of the within-cluster dependence, the structure of $\Sigma_n(\tau,\tau')$ is now significantly more involved. This does not change in the limit as $n\to\infty$, which is assumed to exist along with the limit of the Jacobian.
\begin{assumption}\label{as:gstat}
$J(\tau) = \lim_{n\to\infty} J_n(\tau)$, $\Sigma(\tau,\tau') = \lim_{n\to\infty} \Sigma_n(\tau,\tau')$ exist for $\tau,\tau'\in\mathit{T}$.
\end{assumption}
\begin{comment}
\noindent I show in the appendix that under Assumptions \ref{as:data}-\ref{as:smooth} the pointwise convergence in Assumption \ref{as:gstat} already implies uniform convergence of $J_n$ and $\Sigma_n$.
\end{comment}
The matrix limit $J(\tau)$ is positive definite by Assumption \ref{as:smooth}. Hence, the asymptotic distribution of $\sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau))$ can be determined via the continuous mapping theorem and an application of a central limit theorem to the right-hand side of \eqref{eq:bahadur}. The following theorem confirms that this even remains valid when $\tau \mapsto \sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau))$ is viewed as a random process indexed by $\mathit{T}$. The distributional convergence then occurs relative to $\ell^\infty(T)^d$, the class of uniformly bounded functions on $\mathit{T}$ with values in $\mathbb{R}^d$.
\begin{theorem}\label{th:clt}
Suppose Assumptions \ref{as:data}-\ref{as:gstat} hold. Then $\{\sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau)): \tau\in\mathit{T}\}$ converges in distribution to a mean-zero Gaussian process $\{\mathbb{Z}(\tau):\tau\in\mathit{T}\}$ with covariance function $\ev\mathbb{Z}(\tau)\mathbb{Z}(\tau')^\top= J^{-1}(\tau)\Sigma(\tau,\tau')J^{-1}(\tau')$, $\tau,\tau'\in\mathit{T}$.
\end{theorem}
\begin{comment}
\begin{inparaenum}[(i)]
\item The proof of the theorem proceeds via empirical process arguments similar to those used in \citet{angristetal2006}. Their results do not carry over to the present case because heterogeneous cluster data is not covered by their iid assumptions. This prevents the use of standard Donsker theorems and leads to measurability issues typically not encountered in such proofs. Both problems are taken care of through results of \citet{pollard1991b} and \citet{kosorok2003}.
\item The theorem implies joint asymptotic normality of $\sqrt{n}(\hat{\beta}_n(\tau_j) - \beta(\tau_j))$ for every finite set of quantile indices $\tau_j\in\mathit{T}$, $j=1, 2, \dots$; see, e.g., Theorem 18.14 of \citet{vandervaart1998}. The asymptotic covariance at $\tau_{j}$ and $\tau_{j'}$ is $J^{-1}(\tau_j)\Sigma(\tau_j,\tau_{j'})J^{-1}(\tau_{j'})$. If only this finite dimensional convergence is needed, then Assumptions \ref{as:smooth}\eqref{as:smooth:den} and \eqref{as:smooth:der} can be relaxed considerably using the approach of \citet{knight1998}.
\end{inparaenum}
\end{comment}
The non-iid structure of the data causes the asymptotic covariance function of $\sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau))$ to take on the sandwich form $J^{-1}(\tau)\Sigma(\tau,\tau')J^{-1}(\tau')$. Estimates of these covariances are needed for Wald-type inference. However, in addition to the usual problem of having to control the nuisance quantities $f_{ik}( X_{ik}^\top\beta(\tau) \mid X_{ik})$ contained in the Jacobian $J(\tau)$, the matrix $\Sigma(\tau,\tau')$ now also contains products of quantile crossing indicators $\psi_\tau(Y_{ik}-X_{ik}^\top\beta(\tau)) \psi_{\tau'}(Y_{il}-X_{il}^\top\beta(\tau'))$. For standard plug-in inference, the crossing indicators can be estimated by replacing $\tau\mapsto\beta(\tau)$ with $\tau\mapsto\hat{\beta}_n(\tau)$. The Jacobian is not directly affected by the within-cluster dependence and can be estimated using the bandwidth-driven methods of \citet{hendrickskoenker1992} and \citet{powell1986}.
\citet{parentesantossilva2013} propose such a plug-in estimator based on \citeauthor{powell1986}'s method and show that it leads to asymptotically valid covariance matrix estimates at individual quantiles in a setting with iid (necessarily equal-sized) clusters.\footnote{Their method is likely to generalize to allow for pointwise inference in the presence of clusters with unequal sizes.} However, both the \citetalias{hendrickskoenker1992} and \citeauthor{powell1986} estimators are sensitive to the choice of bandwidth. \citet[pp.\ 5-6]{parentesantossilva2013} give practical suggestions on how to select this bandwidth in the cluster context, but also note that some standard bandwidth rules derived for iid data seem to not perform well in some contexts. To my knowledge, bandwidth rules for QR that explicitly deal with cluster data are currently not available. Moreover, \citeauthor{parentesantossilva2013}'s method does not extend to uniform inference over ranges of quantiles because the limiting Gaussian process $\{\mathbb{Z}(\tau):\tau\in\mathit{T}\}$ from Theorem \ref{th:clt} is nuisance parameter dependent and cannot be normalized to be free of these parameters. Critical values for inference based on $\mathbb{Z}$ therefore cannot be tabulated.
In the next section I present a bootstrap method that is able to approximate the distribution of the limiting process, consistently estimates the covariance function of that process, and avoids the issue of choosing a bandwidth (and kernel) altogether.
\section{Bootstrap Algorithms for Cluster-Robust Inference}\label{s:boot}
In this section I describe and establish the validity of procedures for cluster-robust bootstrap inference (Algorithm \ref{al:binf} below) and cluster-robust confidence bands (Algorithm \ref{al:bcbands}) in QR models. Recall from the discussion above equation \eqref{eq:taylorex} that the population first-order condition of the QR objective function can be written as
\begin{equation}\label{eq:foc}
\sqrt{n}M'_{n}(\beta(\tau), \tau) =- \frac{1}{\sqrt{n}}\sum_{i=1}^n\sum_{k=1}^{c_i}\ev\psi_{\tau}\bigl(Y_{ik} - X_{ik}^\top\beta(\tau)\bigr)X_{ik} = 0
\end{equation}
where $\psi_\tau(z) = \tau - 1\{z < 0\}$ as before. The sample analogue of this condition,
\begin{equation}\label{eq:grad}
\frac{1}{\sqrt{n}}\sum_{i=1}^n\sum_{k=1}^{c_i}\psi_{\tau}(Y_{ik} - X_{ik}^\top\beta)X_{ik} = 0
\end{equation}
can be thought of as being nearly solved by the QR estimate $\beta=\hat{\beta}_n(\tau)$.
The idea is now to bootstrap by repeatedly computing solutions to perturbations of \eqref{eq:grad}. To account for the possible heterogeneity in the data, I use \citeauthor{chenetal2003}'s (\citeyear{chenetal2003}) modification of the wild bootstrap \citep{wu1986,liu1988,mammen1992} for QR with correlated data. \citeauthor{chenetal2003}\ use their method to obtain confidence intervals for QR estimators at a single quantile. Here, I considerably extend the scope of their method to allow for inference on the entire QR process and uniformly consistent covariance matrix estimates of that process via the bootstrap; confidence intervals at individual quantiles $\tau_j\in\mathit{T}$, $j=1, 2, \dots$, then follow as a special case. Because \citeauthor{chenetal2003}\ do not give explicit regularity conditions for the validity of their method, this paper also serves as a theoretical justification for their pointwise confidence intervals.
To ensure that the bootstrap versions of the QR estimate accurately reflect the within-cluster dependence, the resampling scheme perturbs the gradient condition \eqref{eq:grad} at the cluster level. Let $W_1,\dots,W_n$ be iid copies of a random variable $W$ with $\ev W = 0$, $\var W=1$, and $\ev|W|^q < \infty$, where $q >2$ as in Assumption \ref{as:smooth}\eqref{as:smooth:mom}. Here $W$ is independent of the data. Define the bootstrap gradient process as
\begin{equation*}
\mathbb{W}_n(\beta,\tau) = \frac{1}{\sqrt{n}}\sum_{i=1}^n W_i \sum_{k=1}^{c_i}\psi_{\tau}(Y_{ik} - X_{ik}^\top\beta)X_{ik}.
\end{equation*}
An obvious strategy for bootstrap resampling would now be to repeatedly solve $\mathbb{W}_n(\beta,\tau) =0$ for $\beta$ with different draws of
$W_1,\dots,W_n$. However, this type of resampling is impractical because zeros of $\beta\mapsto\mathbb{W}_n(\beta,\tau)$ are difficult to compute
due to the fact that $\mathbb{W}_n(\beta,\tau) =0$ is not a first-order condition of a convex optimization problem.
Instead, I use the bootstrap gradient process $\mathbb{W}_n(\tau) := \mathbb{W}_n(\hat{\beta}_n(\tau) ,\tau)$ evaluated at the original QR estimate to construct the new objective function
\begin{equation}\label{eq:bootmin}
\beta\mapsto\mathbb{M}^*_{n}(\beta,\tau) = \mathbb{M}_{n}(\beta,\tau) + \mathbb{W}_n(\tau)^\top \beta/\sqrt{n}
\end{equation}
and define the process $\tau\mapsto \hat{\beta}^*_n(\tau)$ as any solution to $\min_{\beta\in\mathit{B}}\mathbb{M}^*_{n}(\beta,\tau)$. Then
$\hat{\beta}^*_n(\tau)$ can be interpreted as the $\beta$ that nearly solves the corresponding ``first-order condition''
\begin{equation*}
\frac{1}{\sqrt{n}}\sum_{i=1}^n\sum_{k=1}^{c_i}\psi_{\tau}(Y_{ik} - X_{ik}^\top\beta)X_{ik} = \mathbb{W}_n(\tau)
\end{equation*}
This bootstrap, which I refer to as \emph{wild gradient bootstrap}, essentially perturbs the right-hand side of \eqref{eq:grad} instead of the left. Because $\mathbb{W}_n(\tau)$ mimics the original gradient process $n^{-1/2}\sum_{i=1}^n\sum_{k=1}^{c_i}\psi_{\tau}(Y_{ik} - X_{ik}^\top\hat{\beta}_n(\tau))X_{ik}$ just like the original gradient process mimics the population first-order condition \eqref{eq:foc}, choosing $\hat{\beta}^*_n(\tau)$ in such a way induces the left-hand side of the preceding display to match the behavior of $\mathbb{W}_n(\tau)$. The distributions of $\sqrt{n}(\hat{\beta}^*_n(\tau)-\hat{\beta}_n(\tau))$ and $\sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau))$ can then be expected to be similar. Theorem \ref{th:bootclt} ahead confirms that this is indeed the case, uniformly in $\tau\in\mathit{T}$.
The distributional convergence occurs both in the standard sense and with probability approaching one, conditional on the sample data $D_n := \{ (Y_{ik}, X_{ik}^\top)^\top : 1\leq k\leq c_i, 1\leq i\leq n\}$. The latter concept is the standard measure of consistency for bootstrap distributions; see, e.g., \citet[p.\ 332]{vandervaart1998}. Let $\mathrm{BL}_1(\ell^\infty(\mathit{T})^d)$ be the set of functions on $\ell^\infty(\mathit{T})^d$ with values in $[-1,1]$ that are uniformly Lipschitz and define $\ev^*(\cdot) := \ev(\cdot \mid D_n)$.
\begin{theorem}\label{th:bootclt}
If Assumptions \ref{as:data}-\ref{as:gstat} hold, then $\{\sqrt{n}(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau)):\tau\in\mathit{T}\}$ converges in distribution to the Gaussian process $\{\mathbb{Z}(\tau): \tau\in\mathit{T}\}$ described in Theorem \ref{th:clt}. The convergence also holds conditional on the data in the sense that
$$\sup_{h\in\mathrm{BL}_1(\ell^\infty(\mathit{T})^d)} \bigl| \ev^* h\bigl(\sqrt{n}\bigl(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau)\bigr)\bigr) - \ev h(\mathbb{Z}(\tau))\bigr|
\pto^{\prob}
0.$$
\end{theorem}
Minimizing the bootstrap objective function \eqref{eq:bootmin} is
a standard convex optimization problem. In fact, as the following algorithm shows, the problem can be implemented in statistical software as a linear QR with one additional observation. The idea is to pick a large enough $Y^*$ to ensure $Y^* > X^{*\top}\hat{\beta}^*_n(\tau)$ for all $\tau\in\mathit{T}$, where $X^* = -\sqrt{n}\mathbb{W}_{n}(\tau)/\tau$. Then $\sqrt{n}\mathbb{W}_n(\tau)^\top \beta = \rho_{\tau}(Y^* - X^{*\top} \beta) - \tau Y^*$ and $- \tau Y^*$ can be ignored because $\hat{\beta}^*_n(\tau)$ not only minimizes \eqref{eq:bootmin}, but also $\beta \mapsto n\mathbb{M}^*_{n}(\beta,\tau)- \tau Y^*$.
\begin{algorithm}[Wild gradient bootstrap]\label{al:bse}
\begin{compactenum}
\item\label{al:bse:1} Run a QR of $Y_{ik}$ on $X_{ik}$ and save $\tau\mapsto \hat{\beta}_n(\tau)$. Compute $Y^* = n \max_{1\leq i \leq n} c_i \max_{1\leq i \leq n, 1\leq k\leq c_i} |Y_{ik}|$.
\item\label{al:bse:2} Draw iid copies $W_1,\dots,W_n$ of $W$ and compute $\mathbb{W}_n(\tau) := \mathbb{W}_n(\hat{\beta}_n(\tau) ,\tau)$ for that draw. Generate $X^* = -\sqrt{n}\mathbb{W}_{n}(\tau)/\tau$ and rerun the QR from Step \ref{al:bse:1} with the additional observation $(Y^*, X^{*\top})^\top$ to obtain
$\tau\mapsto \hat{\beta}^*_{n}(\tau) = \argmin_{\beta}\sum_{i=1}^n\sum_{k=1}^{c_i}\rho_{\tau}(Y_{ik} - X_{ik}^\top \beta) + \rho_{\tau}(Y^* - X^{*\top} \beta).$
\item\label{al:bse:3} Repeat Step \ref{al:bse:2} $m$ times, each with a new realization of $W_1,\dots,W_n$.
\item Approximate the distribution of $\{\sqrt{n}(\hat{\beta}_n(\tau) - \beta(\tau)) : \tau\in\mathit{T} \}$ by the empirical distribution of the $m$ observations of $\{\sqrt{n}(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau)) : \tau\in\mathit{T}\}$.
\end{compactenum}
\end{algorithm}
\begin{comments}
\begin{inparaenum}[(i)]
\item The idea of representing a perturbed QR problem as linear QR with one additional observation is due to \citet{parzenetal1994}. The value of $Y^*$ given in the first step of the algorithm is similar to the one suggested by \citet[Algorithm A.4]{bellonietal2011}
\phantomsection%
\label{rev:mammenrec}%
\changes{\item The Monte Carlo experiments in the next section suggest that in practice $W$ should be drawn from the \citet{mammen1992} 2-point distribution that takes on the value $-(\sqrt{5}-1)/2$ with probability $(\sqrt{5}+1)/(2\sqrt{5})$ and the value $(\sqrt{5}+1)/2$ with probability $(\sqrt{5}-1)/(2\sqrt{5})$. Other distributions such as the Rademacher or \citet{webb2013} distributions can be used, but there is no evidence that this would lead to better inference.}
\end{inparaenum}
\end{comments}
By choosing the number of bootstrap simulations $m$ in Algorithm \ref{al:bse} large enough,\footnote{\citet{andrewsbuchinsky2000} and \citet{davidsonmackinnon2000} provide methods for determining an appropriate number of bootstrap simulations $m$ in practice.} the distribution of $\sqrt{n}(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau))$ or functionals thereof can be approximated with arbitrary precision. I therefore let $m\to\infty$ in the following and define the bootstrap estimate of the asymptotic covariance function $V(\tau,\tau') := J^{-1}(\tau)\Sigma(\tau,\tau')J^{-1}(\tau')$ directly as
\begin{equation*}
\hat{V}_n^*(\tau,\tau') = \ev^*n\bigl(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau)\bigr)\bigl(\hat{\beta}_n^*(\tau') - \hat{\beta}_n(\tau')\bigr)^\top, \qquad \tau,\tau'\in\mathit{T}.
\end{equation*}
In practice one simply computes $\hat{V}_n^*(\tau,\tau')$ as the sample covariance of the $m$ bootstrap observations of $\sqrt{n}(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau))$ and $\sqrt{n}(\hat{\beta}_n^*(\tau') - \hat{\beta}_n(\tau'))$. Cluster-robust standard errors of $\hat{\beta}_n(\tau)$ are the square-roots of the diagonal elements of $\hat{V}_n^*(\tau,\tau)/n$.
\pub{\addlines[2]}
Availability of a consistent estimate of the covariance function of the limiting process is not strictly required for valid bootstrap inference on the QR process. Algorithm \ref{al:binf} ahead shows how this is done. However, especially in the presence of data clusters, applied researchers frequently emphasize the importance of bootstrap covariance matrix estimates for Wald-type inference in mean regression models; see, among others, \citet{bertrandetal2004} and \citet{cameronetal2008}. As the Monte Carlo results in the next section show, reweighting by the bootstrap covariance matrix is equally important for cluster-robust inference in the QR context. Still, because convergence in distribution does not imply convergence in moments, consistency of $\hat{V}_n^*(\tau,\tau')$ does not immediately follow from Theorem \ref{th:bootclt}.
Fortunately, the wild gradient bootstrap is able to consistently approximate the asymptotic variance of $\sqrt{n}(\hat{\beta}_n(\tau) - \beta_n(\tau))$. If the covariates have moments of high enough order, then the approximation of the asymptotic covariance function $V(\tau,\tau')$ through its bootstrap counterpart $\hat{V}_n^*(\tau,\tau')$ is in fact uniform in $\tau,\tau'\in\mathit{T}$.
\begin{theorem}\label{th:bootse}
Suppose Assumptions \ref{as:data}-\ref{as:gstat} hold. Then,
\begin{compactenum}[\upshape(i)]
\item\label{th:bootse_pw} for all $\tau,\tau'\in\mathit{T}$, $\hat{V}_n^*(\tau,\tau') \pto^{\prob} V(\tau,\tau')$, and
\item\label{th:bootse_uni} if $q>4$ in Assumption \ref{as:smooth}, then $\sup_{\tau,\tau'\in\mathit{T}} | \hat{V}_n^*(\tau,\tau') - V(\tau,\tau') | \pto^{\prob} 0$.
\end{compactenum}
\end{theorem}
\begin{comments}
(i)~For the proof of this theorem I extend ideas developed by \citet{kato2011}, who in turn relies to some extent on the strategy used in the proof of Theorem 3.2.5 of \citet{vandervaartwellner1996} and \citeauthor{alexander1985}'s (\citeyear{alexander1985}) ``peeling device.'' \citeauthor{kato2011}'s results do not apply to the present case because he works with a single quantile, iid data, and a different bootstrap method. For the proof I develop new tail bounds on the QR gradient process and differences of such processes. They yield $\ev\sup_{\tau\in\mathit{T}}|\sqrt{n}( \hat{\beta}^*_n(\tau) - \hat{\beta}_n(\tau))|^{p}< \infty$ and $\ev\sup_{\tau\in\mathit{T}}|\sqrt{n}( \hat{\beta}_n(\tau) - \beta(\tau))|^{p}< \infty$ uniformly in $n$ for $p < q$. The first part of the theorem then follows from Theorem \ref{th:bootclt} and a bootstrap version of a standard uniform integrability result. The proof of the second part is considerably more involved, but relies on the same tail bounds.
(ii)~A byproduct of the proof of the theorem is the result that the wild gradient bootstrap correctly approximates other (possibly fractional) order moments of $\mathbb{Z}(\tau)$ if the covariates have moments of slightly higher order: As long as $p<q$, the results in the appendix immediately give $\ev^* |\sqrt{n}(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau)) |^p \pto^{\prob} \ev |\mathbb{Z}(\tau)|^p$.
(iii)~\citet{ghoshetal1984} show that the bootstrap variance estimate of an unconditional quantile can be inconsistent if the bootstrap observations are too likely to take on extreme values. This problem is generic and does not depend on the specific type of bootstrap. The boundedness of the parameter space imposed in Assumption \ref{as:model} prevents such behavior in the bootstrap estimates obtained from the perturbed QR problem \eqref{eq:bootmin}. As \citet{kato2011} points out, a possible (although not particularly desirable) alternative would be to restrict the moments on the response variable.
(iv)~Similar but somewhat simpler arguments can be used to prove analogues of Theorems \ref{th:clt}, \ref{th:bootclt}, and \ref{th:bootse} for the bootstrap method presented in \citet{parzenetal1994} for QR with independent data. For iid data, such analogues of Theorems \ref{th:clt} and \ref{th:bootclt} are essentially contained in the results of \citet{bellonietal2011} as special cases.
\end{comments}
I now turn to inference with the wild gradient bootstrap. Let $\tau \mapsto R(\tau)$ be a continuous, $(h\times d)$-matrix-valued function with $h\leq d$ and let $r\colon\mathit{T}\to\mathbb{R}^d$. Suppose $R(\tau)$ has full rank for every $\tau\in\mathit{T}$. I consider testing general pairs of hypotheses of the form
\begin{align*}
\mathrm{H_0}\colon R(\tau)\beta(\tau)=r(\tau)~\text{for all $\tau\in\mathit{T}$}, \qquad \mathrm{H_1}\colon R(\tau)\beta(\tau)\neq r(\tau)~\text{for some $\tau\in\mathit{T}$}.
\end{align*}
Many empirically relevant hypotheses can be tested with this framework. For example, a standard hypothesis in practice is that a single QR coefficient is zero for all $\tau\in\mathit{T}$. If the coefficient of interest is the first entry of $\beta(\tau)$, then $R(\tau)\equiv (1,0,\dots,0)$ and $r(\tau)\equiv 0$.
For inference I use generalized Kolmogorov-Smirnov statistics. Cram\'er-von-Mises versions of these statistics can be used as well, but are not discussed here to conserve space. For a positive definite weight matrix function $\tau\mapsto\Omega(\tau)$ with positive square root $\Omega^{1/2}(\tau)$, define the test statistic
\begin{equation}\label{eq:ksstat}
K_n(\Omega,T) = \sup_{\tau\in\mathit{T}} \bigl\vert\Omega^{-1/2}(\tau)\sqrt{n}\bigl(R(\tau)\hat{\beta}_n(\tau)-r(\tau)\bigr)\bigr\vert.
\end{equation}
I focus on two versions of the statistic: (i)~an unweighted version with $\Omega(\tau)\equiv I_d$ and (ii)~a Wald-type statistic with $\Omega(\tau)$ equal to $$ \hat{\Omega}_n^*(\tau) := R(\tau)\hat{V}_n^*(\tau,\tau)R(\tau)^\top. $$ Other choices are clearly possible. For example, $\hat{V}_n^*(\tau,\tau)$ can be replaced by any other uniformly consistent estimate of $V(\tau,\tau)$. However, the Monte Carlo study in the next section suggests that option (ii) leads to tests with better finite-sample size and power than tests based on (i) or analytical estimates of $V(\tau,\tau)$.
\pub{\addlines}
In the absence of within-cluster correlation, the process inside the Euclidean norm in \eqref{eq:ksstat} with $\Omega = \hat{\Omega}_n^*$ would converge weakly to a standard vector Brownian bridge. Consequently, $\mathit{K}_n(\hat{\Omega}_n^*,T)$ would converge in distribution to the supremum of a standardized, tied-down Bessel process whose critical values can be simulated or computed exactly; see \citet{koenkermachado1999} for details. In the presence of data clusters, the limiting Gaussian process of the quantity inside the Euclidean norm is no longer a Brownian bridge for any choice of weight matrix. Both $\mathit{K}_n(\hat{\Omega}_n^*,T)$ and $K_n(I_d,\mathit{T})$ are then, in general, asymptotically non-pivotal statistics.
Bootstrap tests based on $\mathit{K}_n(\hat{\Omega}_n^*,T)$ therefore do not necessarily outperform tests based on $\mathit{K}_n(I_d,\mathit{T})$ because of asymptotic refinements; see, e.g., \citet{hall1992}.
However, as I will show below, $\mathit{K}_n(\hat{\Omega}_n^*,T)$ still has the advantage that its square converges to a chi-square distribution if $\mathit{T}$ consists of only a single quantile.
The following algorithm describes how to conduct inference and how to test restrictions on the QR process uniformly over the entire set $\mathit{T}$. This includes, for example, individual quantiles, finite sets of quantiles, closed intervals, and disjoint unions of closed intervals.
\begin{algorithm}[Wild gradient bootstrap inference]\label{al:binf}
\begin{compactenum}
\item\label{al:binf:1} Do Steps \ref{al:bse:1}-\ref{al:bse:3} of Algorithm \ref{al:bse}.
\item\label{al:binf:2} If $\Omega(\tau) = \hat{\Omega}_n^*(\tau)$, compute $\hat{V}_n^*(\tau,\tau)$ as the sample variance of the $m$ bootstrap observations of $\sqrt{n}(\hat{\beta}_n^*(\tau) - \hat{\beta}_n(\tau))$ from Step \ref{al:binf:1}.
\item For each of the $m$ bootstrap observations from Step \ref{al:binf:1}, calculate
\begin{equation}\label{eq:ksstarstat}
K^*_n(\Omega,T) := \sup_{\tau\in\mathit{T}} \bigl\vert \Omega^{-1/2}(\tau)\sqrt{n}R(\tau)\bigl(\hat{\beta}^*_n(\tau)-\hat{\beta}_n(\tau)\bigr)\bigr\vert.
\end{equation}
\item Reject $\mathrm{H}_0$ in favor of $\mathrm{H}_1$ if $\mathit{K}_n(\Omega,\mathit{T})$ is larger than $q_{n,1-\alpha}(\Omega,\mathit{T})$, the $1-\alpha$ empirical quantile of the $m$ bootstrap statistics $\mathit{K}^*_n(\Omega,\mathit{T})$.
\end{compactenum}
\end{algorithm}
As before, I take the number of bootstrap simulations $m$ as large and view the bootstrap quantile $q=q_{n,1-\alpha}(\Omega,\mathit{T})$ directly as the minimizer of $$\ev^* \Bigl( \rho_{1-\alpha}\bigr(K^*_n(\Omega,T) - q\bigr) - \rho_{1-\alpha}\bigl(K^*_n(\Omega,T)\bigr)\Bigr).$$ Subtracting the second term here again ensures that this expression is necessarily finite without further conditions on the underlying variables.
To prove consistency of Algorithm \ref{al:binf} for the Wald-type weight $\hat{\Omega}_n^*$, we also need to guarantee that $\hat{\Omega}_n^*$ is non-singular with probability approaching one as $n\to\infty$. This requires the eigenvalues of $\Sigma(\tau,\tau)$ in $V(\tau,\tau) = J^{-1}(\tau)\Sigma(\tau,\tau)J^{-1}(\tau)$ to be bounded away from zero, uniformly in $\tau\in\mathit{T}$. In the absence of clusters, such a property would automatically follow from non-singularity of $n^{-1}\sum_{i=1}^n \ev X_{i1}X_{i1}^\top$. (Recall the discussion above Assumption \ref{as:gstat}.) In the cluster context, it is a separate restriction that rules out some scenarios where several clusters have similar forms of extreme within-cluster dependence.
\begin{assumption}\label{as:posdef}
For all non-zero $a\in\mathbb{R}^d$, $\inf_{\tau\in\mathit{T}}a^\top\Sigma(\tau,\tau)a > 0$.
\end{assumption}
The next result shows that Algorithm \ref{al:binf} is indeed a consistent test of the null hypothesis $R(\tau)\beta(\tau)=r(\tau)$ for all $\tau\in\mathit{T}$ against the alternative that $R(\tau)\beta(\tau)\neq r(\tau)$ for some $\tau\in\mathit{T}$
\begin{theorem}\label{c:bootse}
Suppose Assumptions \ref{as:data}-\ref{as:gstat} and \ref{as:posdef} hold. For $\alpha\in(0,1)$, we have
\begin{compactenum}[\upshape(i)]
\item under $\mathrm{H_0}$, $\prob(\mathit{K}_n(\hat{\Omega}_n^*,\mathit{T}) > q_{n,1-\alpha}(\hat{\Omega}_n^*,\mathit{T}))\to\alpha$ and
\item under $\mathrm{H_1}$, $\prob(\mathit{K}_n(\hat{\Omega}_n^*,\mathit{T}) > q_{n,1-\alpha}(\hat{\Omega}_n^*,\mathit{T}))\to 1$.
\end{compactenum}
Both results also hold without Assumption \ref{as:posdef} if $I_d$ is used instead of $\hat{\Omega}_n^*$ in all instances.
\end{theorem}
\begin{comment}
The theorem in fact remains valid if the Euclidean norm in the definition of $\mathit{K}_n(\Omega,\mathit{T})$ in \eqref{eq:ksstat} is replaced by any other norm on $\mathbb{R}^d$ as long as the same norm is also employed in the bootstrap statistic $\mathit{K}^*_n(\Omega,\mathit{T})$ in \eqref{eq:ksstarstat}. A natural choice other than the Euclidean norm is the maximum norm $|x|_{\max} = \max\{|x_1|,\dots,|x_d|\}$, i.e., the maximum absolute entry of a vector $x = (x_1,\dots,x_d)$. I will use this norm below to construct bootstrap confidence bands for the QR coefficient functions.
\end{comment}
I now discuss three useful corollaries of Theorems \ref{th:bootse} and \ref{c:bootse} regarding (i)~chi-square inference with the bootstrap covariance matrix, (ii)~bootstrap confidence bands, and (iii)~computation of the supremum in the Kolmogorov-Smirnov statistics. First, if $\mathit{T}$ consists of only a single quantile $\tau_0$, then the square of $K_n(\hat{\Omega}_n^*,\tau_0)$ is simply the ordinary Wald statistic
$$n\bigl(R(\tau_0)\hat{\beta}_n(\tau_0)-r(\tau_0)\bigr)^\top \hat{\Omega}_n^{*-1}(\tau_0)\bigl(R(\tau_0)\hat{\beta}_n(\tau_0)-r(\tau_0)\bigr).$$
Because $\sqrt{n}(R(\tau_0)\hat{\beta}_n(\tau_0)-r(\tau_0))$ is asymptotically multivariate normal under the null hypothesis and $\hat{\Omega}_n^*(\tau_0)$ is consistent for the variance of that multivariate normal distribution, the statistic in the preceding display has an asymptotic chi-square distribution. Hence, chi-square critical values can be used instead of bootstrap critical values for the test decision. The following corollary makes this precise.
\begin{corollary}\label{c:bootpw}
Suppose we are in the situation of Theorem \ref{c:bootse} with $\mathit{T} = \{\tau_0\}$ for some $\tau_0\in(0,1)$. Then
\begin{compactenum}[\upshape(i)]
\item under $\mathrm{H_0}$, $K_n(\hat{\Omega}_n^*,\tau_0)^2 \leadsto \chi^2_{\rank R(\tau_0)}$ and
\item under $\mathrm{H_1}$, $\prob(\mathit{K}_n(\hat{\Omega}_n^*,\tau_0)^2 > q)\to 1$ for every $q\in\mathbb{R}$.
\end{compactenum}
\end{corollary}
\begin{comments}
\begin{inparaenum}[(i)]
\item From this result it also follows immediately that a single QR coefficient at a single quantile can be studentized with its bootstrap standard error and compared to a standard normal critical value.
\phantomsection%
\label{rev:smallsample}%
\changes{\item The Monte Carlo study below suggests that asymptotic inference using the bootstrap covariance matrix generally performs well and is only slightly worse in terms of finite-sample size than bootstrapping both the covariance matrix and the critical values. Still, when there are very few clusters, asymptotic inference with bootstrap standard errors tends to over-reject while simultaneously having significantly lower power than the test with bootstrap critical values. The over-rejection could, in principle, be avoided by replacing standard normal and chi-square critical values with larger critical values from the Student $t_{n-1}$ and similarly scaled $F$ distributions \citep{donaldlang2007, besteretal2011}. However, such small-sample adjustments would decrease the power of the test even further. It is therefore recommended to bootstrap the critical values when only few clusters are available.}
\end{inparaenum}
\end{comments}
Next, the results in Theorems \ref{th:bootse} and \ref{c:bootse} allow for the construction of bootstrap confidence bands (uniform in $\tau\in\mathit{T}$) for the QR coefficient function. These bands can be computed jointly for the entire $d$-dimensional function or only a subset $\Delta\subset\{1,\dots,d\}$ of coefficients. As before, a positive definite weight matrix function, denoted here by $\tau\mapsto\Lambda(\tau)$, can be specified to improve the finite-sample performance. An obvious choice is $\Lambda(\tau) = \hat{V}_n^*(\tau,\tau)$. In the following algorithm and in the corollary immediately below, I write $a_j$ for the $j$th entry of a $d$-vector $a$ and $A_{jj}$ for the $j$th diagonal element of a $d\times d$ square matrix $A$.
\begin{algorithm}[Wild gradient bootstrap confidence bands]\label{al:bcbands}
\begin{compactenum}
\item Do Steps \ref{al:bse:1}-\ref{al:bse:3} of Algorithm \ref{al:bse} and, if $\Lambda(\tau) = \hat{V}_n^*(\tau,\tau)$, compute $\hat{V}_n^*(\tau,\tau)$ as in Step \ref{al:binf:2} of Algorithm \ref{al:binf}.
\item For each of the $m$ bootstrap observations, calculate
\begin{equation*}
K^*_n(\Lambda,\mathit{T},\Delta) := \sup_{\tau\in\mathit{T}} \max_{j\in\Delta} \biggl\vert\frac{\hat{\beta}^*_n(\tau)_j-\hat{\beta}_n(\tau)_j}{\sqrt{\Lambda(\tau)_{jj}/n}}\biggr\vert
\end{equation*}
and $q_{n,1-\alpha}(\Lambda,\mathit{T},\Delta)$, the $1-\alpha$ empirical quantile of $K^*_n(\Lambda,\mathit{T},\Delta)$.
\item For each $\tau\in\mathit{T}$ and $j\in\Delta$, compute the interval $$\Bigl[\hat{\beta}_n(\tau)_j \pm q_{n,1-\alpha}(\Lambda,\mathit{T},\Delta)\sqrt{\smash{\Lambda(\tau)_{jj}}/n} \Bigr].$$
\end{compactenum}
\end{algorithm}
The confidence band given in the last step of the algorithm has asymptotic coverage probability $1-\alpha$. The proof of this result is based on the fact that, as long as the maximum norm is used in \eqref{eq:ksstarstat} instead of the Euclidean norm, $K^*_n(\Lambda,\mathit{T},\Delta)$ is nothing but the bootstrap statistic $K^*_n(\Omega,\mathit{T})$ with a diagonal weight matrix and a matrix of restrictions $R(\tau)\equiv R$ that selects the coefficients given in $\Delta$.
\begin{corollary}\label{c:bootband}
Suppose we are in the situation of Theorem \ref{c:bootse}. For every $\Delta\subset\{1,\dots,d\}$
$$ \prob\biggl(\beta(\tau)_j \in \biggl[\hat{\beta}_n(\tau)_j \pm q_{n,1-\alpha}(\hat{V}_n^*,\mathit{T},\Delta)\sqrt{\hat{V}_n^*(\tau,\tau)_{jj}/n} \biggr]~\text{for all $\tau\in\mathit{T}$, all $j\in\Delta$}\biggr) $$ converges to $1-\alpha$ as $n\to\infty$. This continues to hold without Assumption \ref{as:posdef} if all instances of $\hat{V}_n^*$ are replaced by $I_d$.
\end{corollary}
Finally, if $\mathit{T}$ is not a finite set, computing $K_n(\Omega,\mathit{T})$ and the confidence bands is generally infeasible in practice due to the presence of a supremum in their definitions. This can be circumvented by replacing the supremum with a maximum over a finite grid $\mathit{T}_n\subset\mathit{T}$ that becomes finer as the sample size increases. For example, if $\mathit{T}$ is a closed interval, we can take $\mathit{T}_n = \{ j/n : j=0,1,\dots,n \}\cap \mathit{T}$. For any $\tau$ in the interior of $\mathit{T}$ and $n$ large enough, we can then find $\tau_n,\tau_n'\in\mathit{T}_n$ that differ by $1/n$ and satisfy $\tau_n\leq \tau < \tau_n'$. This gives $0\leq \tau - \tau_n < 1/n$. Furthermore, the endpoints of $\mathit{T}_n$ are less than $1/n$ away from the respective endpoints of $\mathit{T}$. Hence, every $\tau\in\mathit{T}$ is the limit of a sequence $\tau_n\in\mathit{T}_n$. This turns out to be the property needed to ensure that the approximation of $\mathit{T}$ by a finite set has no influence on the asymptotic behavior of the bootstrap test.
\begin{corollary}\label{c:bootgrid}
Suppose we are in the situation of Theorem \ref{c:bootse} and there exist sets $\mathit{T}_n\subset\mathit{T}$ such that for every $\tau\in\mathit{T}$ there is a sequence $\tau_n\in\mathit{T}_n$ such that $\tau_n\to\tau$ as $n\to\infty$. Then Theorem \ref{c:bootse} and Corollary \ref{c:bootband} continue to hold when $\mathit{T}_n$ is used instead of $\mathit{T}$.
\end{corollary}
The next section illustrates the finite-sample behavior of the wild gradient bootstrap in a brief Monte Carlo exercise. Section \ref{s:star} then provides an application of the wild gradient bootstrap to Project STAR data.
\section{Monte Carlo Experiments}\label{s:mc}
This section presents several Monte Carlo experiments to investigate the small-sample properties of the wild gradient bootstrap in comparison to other methods of inference. I discuss significance tests at a single quantile (Experiment \ref{ex:signi}), inference about the QR coefficient function (Experiment \ref{ex:KS}), and confidence bands (Experiment \ref{ex:bands}).
The data generating process (DGP) for the following experiments is
\begin{align*}
Y_{ik} = 0.1 U_{ik} + X_{ik} + X_{ik}^2 U_{ik},
\end{align*}
where $X_{ik} = \sqrt{\varrho}Z_{i} + \sqrt{1-\varrho}\varepsilon_{ik}$ with $\varrho \in [0, 1)$; $Z_i$ and $\varepsilon_{ik}$ are standard normal, independent of each other, and independent across their indices. This guarantees that the $X_{ik}$ are standard normal and, within each cluster, any two observations $X_{ik}$ and $X_{il}$ have a correlation coefficient of $\varrho$. The $U_{ik}$ are distributed as $\mathrm{N}(0,1/3)$ and drawn independently of $X_{ik}$ to ensure that the $X_{ik}^2 U_{ik}$ have mean zero and variance one. The correlation structure of $U_{ik}$ is chosen such that the within-cluster correlation coefficient of $X_{ik}^2 U_{ik}$ is also approximately $\varrho$.\footnote{By construction, the correlation coefficient of $X_{ik}^2 U_{ik}$ and $X_{il}^2 U_{il}$ is $\corr(U_{ik}, U_{il})(2\varrho^2+1)/3$. I generate data such that $\corr(U_{ik}, U_{il}) = \min\{1, 3\varrho/(2\varrho^2 + 1)\}$. The within-cluster correlation coefficient of $X_{ik}^2 U_{ik}$ is then exactly $\varrho$ for $\varrho \in [0, 0.5]$ and has a value slightly below $\varrho$ for $\varrho\in(0.5,1)$. This choice for $\corr(U_{ik}, U_{il})$ ensures that the other restrictions on the DGP hold for all values of $\varrho$ used in the experiments.} Both $X_{ik}$ and $U_{ik}$ are independent across clusters.
The DGP in the preceding display corresponds to the quadratic QR model
\begin{align}\label{eq:mc}
Q_{ik}(\tau\mid X_{ik}) = \beta_0(\tau) + \beta_1(\tau) X_{ik} + \beta_2(\tau)X_{ik}^2
\end{align}
with $\beta_0(\tau) = \Phi^{-1}(\tau)/\sqrt{300}$, $\beta_1(\tau) \equiv 1$, and $\beta_2(\tau) = \Phi^{-1}(\tau)/\sqrt{3}$, where $\Phi$ is the standard normal distribution function. I denote the QR estimates of the two slope parameters $\beta_1(\tau)$ and $\beta_2(\tau)$ by $\hat{\beta}_{1,n}(\tau)$ and $\hat{\beta}_{2,n}(\tau)$. Their bootstrap versions are $\hat{\beta}^*_{1,n}(\tau)$ and $\hat{\beta}^*_{2,n}(\tau)$. As before, I refer to the square roots of the diagonal elements of $\smash{\hat{V}_n^*}(\tau, \tau)/n$ as bootstrap standard errors and, for simplicity, now denote the bootstrap standard error of $\hat{\beta}_{2,n}(\tau)$ by $\se^*(\hat{\beta}^*_{2,n}(\tau))$.
In the following experiments, I consider inference about $\tau\mapsto \beta_1(\tau)$ and $\tau\mapsto \beta_2(\tau)$ for different values of the number of clusters $n$, the within-cluster correlation $\varrho$, and the variance of the cluster size $\var(c_i)$. In all experiments below, the smallest possible cluster size is $5$ and $c_i$ is distributed uniformly on $\{5, 6, \dots, c_{\max}\}$. Unless otherwise noted, the bootstrap weights are drawn from the Mammen distribution as defined in the remarks below Algorithm \ref{al:bse}.
\begin{experiment}[Significance tests at a single quantile]\label{ex:signi}
This Monte Carlo experiment illustrates the small-sample size and power of different methods for testing whether a single QR coefficient equals zero at a given quantile. To test the correct null hypothesis $\beta_2(.5) = 0$ in \eqref{eq:mc} against the alternative $\beta_2(.5) \neq 0$, I consider (i) wild gradient bootstrap inference as in Algorithm \ref{al:binf}, (ii) standard inference with bootstrap standard errors as in Corollary \ref{c:bootpw}, (iii) cluster-robust inference based on analytically estimating the standard errors, (iv) standard inference without cluster correction, (v) cluster-robust Rao score inference, and (vi) wild bootstrap inference without cluster correction.
For (i), note that $R\equiv (0, 0, 1)$ and $r\equiv 0$. Hence, Algorithm \ref{al:binf} is equivalent to testing whether $|\hat{\beta}_{2,n}(.5)|$ exceeds the empirical $1-\alpha$ quantile of the $m$ observations of $|\hat{\beta}^*_{2,n}(.5)-\smash{\hat{\beta}_{2,n}}(.5)|$ conditional on $\smash{\hat{\beta}_{2,n}}(.5)$. No weight matrix is needed because the test decision is independent of $\Omega(\tau)$ whenever $R(\tau)V(\tau,\tau)R(\tau)^\top$ is a scalar. Similarly, for (ii), the test decision in Corollary \ref{c:bootpw} is equivalent to simply comparing $|\hat{\beta}_{2,n}(.5)|/\se^*(\hat{\beta}^*_{2,n}(.5))$ to $\Phi^{-1}(1-\alpha/2)$. For (iii), I obtain standard errors by estimating $V(\tau,\tau) = J^{-1}(\tau)\Sigma(\tau,\tau)J^{-1}(\tau)$ analytically as suggested by \citet{parentesantossilva2013}. They propose the plug-in estimate
$
\frac{1}{n} \sum_{i=1}^n \sum_{k=1}^{c_i}\sum_{l=1}^{c_i}\psi_\tau\bigl(Y_{ik}-X_{ik}^\top\hat{\beta}_n(\tau)\bigr) \psi_{\tau}\bigl(Y_{il}-X_{il}^\top\hat{\beta}_n(\tau)\bigr)X_{ik}X_{il}^\top$$
for $\Sigma(\tau,\tau)$ and replace $J(\tau)$ by a \citet{powell1986} kernel estimate. The kernel estimate requires a bandwidth choice. The results here are based on the standard implementation in the \texttt{quantreg} package in \texttt{R} with the Hall-Sheather rule; see \citet[pp.\ 80-81]{koenker2005} and \citet{koenker2013}.\footnote{This bandwidth choice required a robust estimate of scale. \citet{koenker2013} uses the minimum of the standard deviation of the QR residuals and their normalized interquartile range. \citet{parentesantossilva2013} suggest the median absolute deviation of the QR residuals with a scaling constant of $1$. I chose \citeauthor{koenker2013}'s implementation because it yielded better results in nearly all cases.} For (iv), I use the regular version of the \citeauthor{powell1986} sandwich estimator described in \citet{koenker2005}. It employs the same kernel estimate of $J(\tau)$ as in (iii), but replaces $\Sigma(\tau,\tau)$ by $n^{-1} \tau(1-\tau) \sum_{i=1}^n \sum_{k=1}^{c_i}X_{ik}X_{ik}^\top$ and is therefore not designed to account for within-cluster correlation. For (v), I apply the $\mathrm{QRS}_0$ test of \citet{wanghe2007}, a cluster-robust version of the QR rank score test \citep{gutenbrunneretal1993}. \citeauthor{wanghe2007} derive their test statistic under homoscedasticity assumptions; the DGP considered here is highly heteroscedastic. For (vi), I compute critical values from the \texttt{quantreg} implementation of the \citeauthor{fengetal2010} (\citeyear{fengetal2010}, FHH hereafter) wild bootstrap for QR models. Their method perturbs the QR residuals via a carefully chosen weight distribution but presumes independent observations. An alternative wild bootstrap procedure due to \citet{davidson2012} had size properties similar to those of the FHH method but had lower power in nearly all of my experiments; results for this bootstrap are therefore omitted.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\textwidth]{fig_ex11}
\caption{Empirical rejection frequencies of a correct hypothesis $\mathrm{H}_0\colon \beta_2(.5) = 0$ (panels (a)-(c)) and the incorrect hypothesis $\mathrm{H}_0\colon \beta_2(.75) = 0$ (panels (d)-(f)) using wild gradient bootstrap critical values (solid black lines), bootstrap standard errors (solid grey), analytical cluster-robust standard errors (dotted black), regular standard errors without cluster correction (dotted grey), cluster-robust rank score inference (long-dashed black), and FHH wild bootstrap without cluster correction (long-dashed grey) at the 5\% level (short-dashed) as a function of the (a) number of clusters, (b) within-cluster correlation, and (c) maximal cluster size.} \label{fig:mc_ex11}
\end{figure}
Panels (a)-(c) in Figure \ref{fig:mc_ex11} show empirical rejection frequencies of a correct hypothesis $\mathrm{H}_0\colon \beta_2(.5) = 0$ for methods (i)-(vi) at the 5\% level (short-dashed line) as a function of (a) the number of clusters $n$, (b) the within-cluster correlation $\varrho$, and (c) the variance of the cluster size $\var(c_i)$. Each horizontal coordinate was computed from 2,000 simulations and all six methods were faced with the same data. The three bootstrap tests used $m=299$ bootstrap repetitions. The wild gradient bootstrap had Mammen weights. Results for other weight distributions are discussed below.
For panel (a), I set $\varrho = .5$, $\var(c_i) = 10$ (i.e., $c_{\max} = 15$), and considered $n\in\{10, 20, \dots, 100\}$. As can be seen, the wild gradient bootstrap critical values (solid black lines) and bootstrap standard errors (solid grey) provided tests that were essentially at the nominal level with as few as 20 clusters, with the bootstrap critical values performing slightly better. Tests based on analytical cluster-robust standard errors (dotted black) and cluster-robust rank scores (long-dashed black) over-rejected significantly, although this property became less pronounced for larger numbers of clusters. Regular standard errors without cluster correction (dotted grey) and wild bootstrap without cluster correction (long-dashed grey) led to severe over-rejection in all cases. For (b), I chose $n = 50$, $\var(c_i) = 10$, and varied $\varrho\in\{0, .1, \dots, .9\}$. At $\varrho = 0$, all tests except the rank score test apply and had almost correct size. For larger values of the within-cluster correlation, the three analytical tests and the FHH bootstrap test over-rejected considerably, although the rank score test improved for very high correlations. The test based on $\se^*(\hat{\beta}^*_{2,n}(\tau))$ over-rejected mildly. The bootstrap test given in Algorithm \ref{al:binf} was nearly at the nominal level in most cases. For (c), I fixed $\varrho = .5$ and changed $c_{\max}\in\{9, 11, \dots, 29\}$ so that $\var(c_i)$ increased from $2$ to $52$ over this range. I simultaneously decreased $n$ in order to keep the average total number of observations constant at approximately 250; this resulted in numbers of clusters between 36 and 15. The test based on the bootstrap standard error again over-rejected slightly but far less than the ones based on the analytical cluster-robust standard error and the cluster-robust rank score. Wild gradient bootstrap critical values again provided a test with almost proper size, while regular standard errors and the wild bootstrap for independent observations were not useful at any value of $\var(c_i)$.
Panels (d)-(f) show empirical rejection frequencies of the incorrect hypothesis $\mathrm{H}_0\colon \beta_2(.75) = 0$ for the same data. Rejection frequencies of the three analytical methods and the FHH wild bootstrap are only reported for completeness and, because of their size distortion, should not be interpreted as estimates of their power. The wild gradient bootstrap critical values tended to lead to a more powerful test than inference with bootstrap standard errors. This was, in particular, the case in small samples, at high within-cluster correlations, and for large variances of the cluster size. The rejection frequencies of all tests were increasing in the number of clusters, decreasing in the within-cluster correlations, and decreasing in the variance of the cluster size.
\phantomsection%
\label{rev:differentrho}%
Following \citet{mackinnonwebb2014}, I also experimented (not shown) with cases where I varied the within-cluster correlation of $X$ and $U$ in the Monte Carlo DGP independently. For the wild gradient bootstrap, I found that for any within-cluster correlation of $X$, the degree of correlation in $U$ had little impact, whereas increases in the within-cluster correlation in $X$ led to mild size distortions similar to the ones found in Figure \ref{fig:mc_ex11}. In contrast, increases in the within-cluster correlation of $U$ led to severe over-rejection in tests based on analytical cluster-robust standard errors; higher correlation in $X$ also induced over-rejection, but the impact was considerably less pronounced.
In light of the findings so far it should be noted that the small-sample results for the analytically estimated standard errors reported here do not contradict the ones reported by \citet{parentesantossilva2013}, who find a much better performance of their method in terms of finite-sample size. In comparison to their experiments, I consider data with smaller numbers of clusters, different correlation structures, and much stronger cluster heterogeneity in terms of cluster sizes. Computing the standard errors analytically worked well when the number of clusters was large, the within-cluster correlation was low, and the clusters were small.
\phantomsection%
\label{rev:wanghedging}%
Similarly, the rank score test of \citet{wanghe2007} is designed for homoscedastic models and performed much better in such settings. For heteroscedastic models, \citet{wang2009} shows that reweighting their test statistic can significantly improve inference when more is known about the specific form of heteroscedasticity; her reweighting schemes do not apply to the DGP in the present example and are therefore not discussed.
\begin{table}[thp]
\caption{Empirical size and power as in Figure \ref{fig:mc_ex11} for different bootstrap weights}\label{tab:weights}
\centering
{\small
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cp{0cm}ccp{0cm}ccp{0cm}ccp{0cm}ccp{0cm}ccp{0cm}cc}%
\hline
&
& \multicolumn{8}{c}{$\mathrm{H}_0\colon \beta_2(.5) = 0$ (size)}
&
& \multicolumn{8}{c}{$\mathrm{H}_0\colon \beta_2(.75) = 0$ (power)}
\\
\cline{3-10}\cline{12-19}
&
& \multicolumn{2}{c}{Mammen}
&
& \multicolumn{2}{c}{Rademacher}
&
& \multicolumn{2}{c}{Webb}
&
& \multicolumn{2}{c}{Mammen}
&
& \multicolumn{2}{c}{Rademacher}
&
& \multicolumn{2}{c}{Webb}
\\
\cline{3-4}\cline{6-7}\cline{9-10}\cline{12-13}\cline{15-16}\cline{18-19}
$n$
&
& cv & se
&
& cv & se
&
& cv & se
&
& cv & se
&
& cv & se
&
& cv & se \\
\hline
10 & & .098 & .114 & & .131 & .166 & & .128 & .146 & & .155 & .019 & & .104 & .016 & & .094 & .011 \\
20 & & .068 & .088 & & .076 & .102 & & .071 & .098 & & .328 & .086 & & .302 & .036 & & .270 & .026 \\
100 & & .054 & .068 & & .059 & .070 & & .054 & .070 & & .876 & .896 & & .864 & .886 & & .868 & .890 \\
\rule{0pt}{3ex}$\varrho$ \\
\hline
.1 & & .061 & .069 & & .063 & .071 & & .062 & .068 & & .998 & 1 & & .998 & 1 & & .999 & 1 \\
.5 & & .055 & .071 & & .059 & .074 & & .061 & .076 & & .602 & .613 & & .590 & .544 & & .589 & .502 \\
.9 & & .057 & .078 & & .067 & .088 & & .065 & .091 & & .378 & .308 & & .376 & .183 & & .371 & .156 \\
\rule{0pt}{3ex}$\var(c_i)$ \\
\hline
2 & & .056 & .070 & & .057 & .072 & & .056 & .072 & & .820 & .840 & & .808 & .830 & & .803 & .831 \\
24 & & .054 & .076 & & .062 & .078 & & .060 & .074 & & .580 & .578 & & .578 & .479 & & .570 & .446 \\
52 & & .056 & .082 & & .062 & .085 & & .066 & .086 & & .456 & .364 & & .459 & .235 & & .446 & .183 \\
\hline
\end{tabular}%
}}
\end{table}
Bootstrap weight distributions other than the Mammen distribution are often found to work well in regression settings.
\phantomsection%
\label{rev:bootweights}%
These include the standard normal distribution, the recentered $\mathrm{Exponential}(1)$ distribution, the Rademacher distribution, which takes on the values $-1$ and $1$ with equal probability, and the \citet{webb2013} 6-point distribution, which takes on $-\sqrt{1.5}, -1, -\sqrt{0.5}, \sqrt{0.5}, 1,$ and $\sqrt{1.5}$ with equal probability. In my experiments, the standard normal had size properties very similar to those of the Rademacher and Webb distributions, but lower power. I therefore do not present detailed results for this distribution. The same holds for the recentered $\mathrm{Exponential}(1)$, which behaved almost like the Mammen distribution in terms of size, but also had lower power. Comparisons of the other distributions are shown in Table \ref{tab:weights}. The experimental setup and data were the same as in Figure \ref{fig:mc_ex11}. The left-hand side of the table measures finite-sample size for different numbers of clusters, within-cluster correlations, and variances of the cluster size as in panels (a), (b), and (c) of Figure \ref{fig:mc_ex11}; the right-hand side corresponds to the power estimates in panels (d)-(f). As can be seen, the Mammen distribution had slightly better size and power, in particular when the number of clusters was small, the within-cluster correlation was high, and the variance of the cluster size was large.
\begin{figure}[t]
\centering
\includegraphics[width=.55\textwidth]{fig_ex12}
\caption{Rejection frequencies of $\mathrm{H}_0\colon \beta_2(\tau) = 0$ for different values of $\tau$ using the same methods as in Table \ref{tab:weights}. $\mathrm{H}_0$ is only true at $\tau = .5$.}
\label{fig:mc_ex12}
\end{figure}
To further investigate finite-sample power of the three bootstrap weight distributions, I plot in Figure~\ref{fig:mc_ex12} their rejection frequencies of $\mathrm{H}_0\colon\beta_2(\tau) = 0$ at 17 separate quantile indices $\tau\in\{.1, .15, \dots, .9\}$ for $n=75$. I again chose $\varrho=.5$, $\var(c_i) = 10$, $m=299$, and 2,000 Monte Carlo simulations. For this experiment all 17 possible null hypotheses were tested with the same data. Only $\mathrm{H}_0\colon\beta_2(.5)=0$ is true. As the plot shows, wild gradient bootstrap critical values led to tests with good size and power at all quantiles and for all weight distribution. Size and power of the tests based on bootstrap standard errors were similar for $\tau\in[.2,.8]$. However, for quantile indices outside this interval the tests with bootstrap standard errors from the Rademacher (dotted grey) and Webb (long-dashed grey) distributions showed a sharp decline in power; the Mammen distribution (solid grey) did not have this issue. I experimented with the parameters of the DGP and found that the power of the Rademacher and Webb distributions for small and large $\tau$ increased quickly when I increased the number of clusters or decreased the within-cluster correlation. For example, for the Rademacher distribution the rejection frequency at $\tau=.1$ and $.9$ was about 80\% when I either set $\varrho$ to $.3$ or $n$ to $100$.
\phantomsection%
\label{rev:skewdiscussion}%
The reason for the large differences in finite-sample power between the distributions appears to be the extreme skewness of the distribution of the summands in the gradient process for small and large $\tau$. The asymmetry in the Mammen distribution seems to mimic this property particularly well. The standard errors also improved when I used a recentered $\mathrm{Exponential}(1)$ or other asymmetric distributions, but the Mammen distribution provided the best results. \hfill$\square$
\end{experiment}
\begin{experiment}[Uniform inference on the QR process]\label{ex:KS}
This experiment illustrates the finite-sample performance of Algorithm \ref{al:binf} for inference on the entire QR process. I tested the true hypothesis $\mathrm{H}_0\colon \beta_1(\tau) = 1$ for all $\tau\in\mathit{T}$ and the false hypothesis $\mathrm{H}_0\colon \beta_1(\tau)= 0$ for all $\tau\in\mathit{T}$ at the 5\% level. I chose $\var(c_i)=10$, $m=199$ bootstrap simulations with the Mammen distribution, and, in view of Corollary \ref{c:bootgrid}, I approximated $\mathit{T} = [.1, .9]$ by $\{.1, .2, \dots, .9\}$. The test statistics $K_n(\Omega, \mathit{T})$ were either (i)~weighted by the bootstrap estimate $\hat{\Omega}_n^*$, (ii)~weighted by the analytical estimate of $\tau\mapsto R(\tau)V(\tau,\tau)R(\tau)^\top$ described in the preceding Monte Carlo exercise, or (iii)~unweighted ($\Omega = I$). All three methods were faced with the same data.
\begin{table}[thp]
\caption{Empirical size of Algorithm \ref{al:binf} at the 5\% level}\label{tab:ks}
\centering
{\small
\begin{tabular}{cp{0cm}cccp{0cm}cccp{0cm}ccc}%
\hline
&
& \multicolumn{3}{c}{$n=20$}
&
& \multicolumn{3}{c}{$n=30$}
&
& \multicolumn{3}{c}{$n=50$}
\\
\cline{3-5}\cline{7-9}\cline{11-13}
$\varrho$
&
& boot. & ana. & unw.
&
& boot. & ana. & unw.
&
& boot. & ana. & unw.
\\
\hline
0 & & .029 & .026 & .012 & & .034 & .014 & .011 & & .038 & .026 & .012 \\
.1 & & .030 & .022 & .004 & & .038 & .029 & .013 & & .029 & .029 & .011 \\
.2 & & .020 & .016 & .006 & & .025 & .026 & .010 & & .035 & .034 & .015 \\
.3 & & .020 & .017 & .002 & & .032 & .028 & .009 & & .042 & .039 & .017 \\
.4 & & .010 & .006 & .001 & & .030 & .026 & .009 & & .048 & .039 & .011 \\
.5 & & .008 & .001 & .000 & & .016 & .011 & .001 & & .036 & .025 & .006 \\
.6 & & .005 & .000 & .000 & & .015 & .006 & .000 & & .041 & .029 & .004 \\
.7 & & .011 & .001 & .000 & & .015 & .010 & .000 & & .039 & .030 & .001 \\
.8 & & .012 & .000 & .000 & & .008 & .005 & .000 & & .037 & .027 & .003 \\
.9 & & .012 & .000 & .000 & & .006 & .002 & .000 & & .032 & .020 & .002 \\
\hline
\end{tabular}%
}
\end{table}
Table \ref{tab:ks} reports the empirical rejection frequencies of the true null hypothesis for methods (i)-(iii) from an experiment with 1,000 Monte Carlo simulations for each $\varrho\in\{0, .1, \dots, .9\}$ and each $n\in\{20, 30, 50\}$. At $n = 20$, the test based on the bootstrapped Wald weight (``boot.'')\ was quite conservative for all degrees of within-cluster correlation but not overly so for $\varrho$ smaller than $.4$. The performance of the bootstrap test with analytical weights (``ana.'')\ was slightly worse, especially for higher within-cluster correlations. However, both of these tests under-rejected considerably less for larger $n$ so that at $n=50$ the size of bootstrap-weighted test was above .3 for all but one $\varrho$. In contrast, the unweighted version was very conservative for all within-cluster correlations and all numbers of clusters.
\begin{table}[thp]
\caption{Empirical power of Algorithm \ref{al:binf} at the 5\% level}\label{tab:kspow}
\centering
{\small
\begin{tabular}{cp{0cm}cccp{0cm}cccp{0cm}ccc}%
\hline
&
& \multicolumn{3}{c}{$n=10$}
&
& \multicolumn{3}{c}{$n=15$}
&
& \multicolumn{3}{c}{$n=20$}
\\
\cline{3-5}\cline{7-9}\cline{11-13}
$\varrho$
&
& boot. & ana. & unw.
&
& boot. & ana. & unw.
&
& boot. & ana. & unw.
\\
\hline
0 & & 1 & .735 & .692 & & 1 & .986 & .983 & & 1 & 1 & 1 \\
.1 & & .994 & .431 & .373 & & 1 & .924 & .904 & & 1 & .994 & .991 \\
.2 & & .944 & .173 & .133 & & .993 & .704 & .665 & & 1 & .962 & .945 \\
.3 & & .828 & .043 & .027 & & .964 & .352 & .320 & & .999 & .787 & .747 \\
.4 & & .665 & .013 & .001 & & .849 & .085 & .062 & & .963 & .417 & .362 \\
.5 & & .438 & .007 & .000 & & .594 & .005 & .001 & & .790 & .063 & .041 \\
.6 & & .412 & .003 & .000 & & .531 & .004 & .000 & & .703 & .034 & .016 \\
.7 & & .409 & .006 & .000 & & .474 & .001 & .000 & & .624 & .024 & .008 \\
.8 & & .388 & .006 & .000 & & .423 & .003 & .001 & & .527 & .014 & .002 \\
.9 & & .380 & .006 & .001 & & .338 & .006 & .001 & & .428 & .014 & .001 \\
\hline
\end{tabular}%
}
\end{table}
Table \ref{tab:kspow} shows empirical rejection frequencies of the false null hypothesis $\mathrm{H}_0\colon \beta_1(\tau)= 0$ for all $\tau\in\mathit{T}$ at $n\in\{10, 15, 20\}$ in the same experimental setup as above. For $n=10$, the Wald test with bootstrap weights had substantial power even for high within-cluster correlations. In sharp contrast, the unweighted and analytically weighted tests rejected considerably fewer false null hypotheses and exhibited a total loss of power starting from about $\varrho = .5$. Increases in the number of clusters translated into significant gains in the power of all tests, but the test based on the bootstrap weight matrix far outperformed the other two tests at all sample sizes.
\hfill $\square$
\end{experiment}
\begin{experiment}[Confidence bands]\label{ex:bands}
In this experiment I investigate the finite-sample properties of Algorithm \ref{al:bcbands}. The setup is as in the preceding experiment. The empirical coverage of $\tau\mapsto \beta_1(\tau)$ with a 95\% wild gradient bootstrap confidence band is, by construction, identical to $1$ minus the empirical size of the bootstrap test in Table \ref{tab:ks} and therefore not shown here. I instead consider a more complex scenario where I report the empirical coverage of a joint 95\% confidence band for the two slope functions $\tau\mapsto \beta_1(\tau)$ and $\tau\mapsto \beta_2(\tau)$ for $n \in \{ 10, 15 , 20\}$ and $\varrho\in\{0, .1, \dots, .9\}$.
Table \ref{tab:cb} contains the results.
\begin{table}[htp]
\caption{Empirical coverage of $\tau\mapsto (\beta_1, \beta_2)(\tau)$ by 95\% confidence band} \label{tab:cb}
\centering
{\small
\begin{tabular}{cp{0cm}cccp{0cm}cccp{0cm}ccc}%
\hline
&
& \multicolumn{3}{c}{$n=10$}
&
& \multicolumn{3}{c}{$n=15$}
&
& \multicolumn{3}{c}{$n=20$}
\\
\cline{3-5}\cline{7-9}\cline{11-13}
$\varrho$
&
& boot. & ana. & unw.
&
& boot. & ana. & unw.
&
& boot. & ana. & unw.
\\
\hline
0 & & .939 & .953 & .993 & & .938 & .949 & .993 & & .945 & .958 & .995 \\
.1 & & .949 & .977 & .997 & & .923 & .938 & .992 & & .938 & .942 & .992 \\
.2 & & .955 & .981 & .991 & & .920 & .946 & .988 & & .930 & .934 & .990 \\
.3 & & .960 & .994 & .997 & & .945 & .975 & .993 & & .934 & .947 & .994 \\
.4 & & .954 & .998 & 1 & & .966 & .993 & 1 & & .948 & .972 & .997 \\
.5 & & .962 & 1 & 1 & & .981 & .999 & 1 & & .974 & .998 & 1 \\
.6 & & .950 & .999 & 1 & & .958 & 1 & 1 & & .972 & .998 & 1 \\
.7 & & .937 & .999 & 1 & & .960 & .998 & 1 & & .969 & 1 & 1 \\
.8 & & .919 & .999 & 1 & & .949 & .997 & .999 & & .965 & .998 & 1 \\
.9 & & .900 & .998 & 1 & & .941 & .997 & .999 & & .960 & .996 & 1 \\
\hline
\end{tabular}%
}
\end{table}
As before, the procedure based on the bootstrapped Wald weight showed the most balanced performance with confidence bands that were close to 95\% in most cases. The only exceptions occurred at $n=10$ for very high within-cluster correlations, where the confidence bands were too thin. The unweighted confidence bands were consistently too wide. For analytical weights, the empirical coverage was near 95\% for small $\varrho$. However, at values of $\varrho$ larger than $.4$ the coverage was essentially 100\% even for $n=20$. Further increases in $n$ (not shown) yielded improvements for all versions of the confidence band but even the bootstrap-weighted confidence band needed a large number of clusters for the coverage to be fully balanced across $\varrho$. \hfill $\square$
\end{experiment}
In summary, the wild gradient bootstrap performs well even in fairly extreme (but empirically relevant) situations where the number of clusters is small, the within-cluster correlation is high, and the clusters are very heterogeneous. Here, reweighting the test statistic by the bootstrap covariance matrix is crucial for tests to have good size and power in finite samples. Analytical weights or no weights can be used when the number of clusters is large; otherwise they tend to lead to tests that are less reliable than those based on the bootstrapped Wald weight. For inference at a single quantile, testing with bootstrap standard errors and normal/chi-square critical values provides a simpler alternative to testing with bootstrap critical values that is, with some exceptions, nearly as good. These findings are also confirmed by an additional experiment in the next section, where I implement placebo interventions in the Project STAR data.
\section{Application: Project STAR}\label{s:star}
This section applies the wild gradient bootstrap to investigate the effects of a class size reduction experiment on the conditional quantiles of student performance on a standardized test. The data come from the first year of the Tennessee \emph{Student/Teacher Achievement Ratio} experiment, known as Project STAR.
I start by briefly describing Project STAR; the discussion closely follows \citetalias{wordetal1990} and \citet{graham2008}, where more details can be found. At the beginning of the 1985-1986 school year, incoming kindergarten students who enrolled in one of the 79 project schools in Tennessee were randomly assigned to one of three class types within their school: a small class (13-17 students), a regular-size class (22-25 students), or a regular-size class (22-25 students) with a full-time teacher's aide. Teachers were then randomly assigned to one of these class types. Each of the project schools was required to have at least one of each kindergarten class type. During the 1985-1986 school year, a total of 6,325 students in 325 different classrooms across Tennessee participated in the project. Classroom identifiers are not available, but \citeauthor{graham2008}'s (\citeyear{graham2008}) matching algorithm is able to uniquely identify 317 of these classrooms in the data. 5,727 students in these classrooms have the complete set of characteristics available that I use in the QR model below. I restrict the analysis to only these kindergarten students.
The outcome of interest is student performance on the \emph{Stanford Achievement Test} in mathematics and reading administered at the end of the 1985-1986 school year. I standardized the raw test scores as in \citet{krueger1999}: First, I computed the empirical distribution functions of the math and reading scores for the pooled sample of regular (with and without teacher's aide) students. Next, I transformed the math and reading scores for students in all three class types into percentiles using the math and reading empirical distribution functions, respectively, obtained in the first step. Finally, to summarize overall performance, I computed the average of the two percentiles for each student. I use this percentile score as the dependent variable in the following analysis. The idea behind \citeauthor{krueger1999}'s normalization is that in the absence of a class size effect, the transformed subject scores for both small and regular class types would have an approximately uniform distribution.
The two main covariates of interest are the treatment dummy $\mathit{small}$ indicating whether the student was assigned to a small class and the treatment dummy $\mathit{regaide}$ indicating whether the student was assigned to a regular class with an aide. I consider the following model for the conditional quantiles of the transformed scores:
\begin{align}\label{eq:star1}
Q_{ik}(\tau\mid X_{ik}) = \beta_0(\tau) + \beta_1(\tau) \mathit{small}_{ik} + \beta_2(\tau) \mathit{regaide}_{ik} + \beta_3(\tau)^\top Z_{ik}.
\end{align}
This specification is similar to the mean regression given in \citeauthor{krueger1999}'s (\citeyear{krueger1999}) Table V.4. The covariate vector $Z_{ik}$ contains a dummy indicating if the student identifies as $\mathit{black}$,\footnote{The sample also contains a large number of students who identify as white and a very small number of students who identify as Hispanic, Asian, American Indian, or other.} a student gender dummy, a dummy indicating whether the student is $\mathit{poor}$ as measured by their access to free school lunch, a dummy indicating if the teacher identifies as black ($\mathit{tblack}$, the other teachers in the sample identify as white), the teacher's years of teaching experience ($\mathit{texp}$), a dummy indicating whether the teacher has at least a master's degree ($\mathit{tmasters}$), and additive school ``fixed effects.'' Because of possible peer effects and unobserved teacher characteristics, I cluster at the classroom level.
\begin{figure}[htp]
\begin{center}
\includegraphics[width=0.9\textwidth]{fig_star1}
\caption{QR coefficient estimates $\tau\mapsto\hat{\beta}_n(\tau)$ (solid black lines) of model \eqref{eq:star1}. The regression includes additive school ``fixed effects'' (not shown). The grey areas are pointwise 95\% wild gradient bootstrap confidence intervals based on bootstrap quantiles clustered at the classroom level.} \label{fig:star1}
\end{center}
\end{figure}
The results are shown in Figure \ref{fig:star1}. The solid black lines in each panel plot a coefficient estimate corresponding to a coefficient in \eqref{eq:star1} as a function of $\tau$. The vertical scale is the average percentile score. The grey bands are pointwise 95\% wild gradient bootstrap confidence intervals based on bootstrap quantiles computed from $m=999$ bootstrap simulations with Mammen weights.
Students assigned to small classes mostly perform better than students assigned to regular classes (with or without aide), although the effect varies across the distribution. For scores above the .2 quantile of the score distribution, the difference is about five percentage points. This is in accordance with \citeauthor{krueger1999}'s (\citeyear{krueger1999}) findings. However, the benefits for students below the .2 quantile are much smaller and become insignificant at the .1 quantile. The impact of a smaller class on students at the very bottom of the score distribution is essentially zero.
In addition, as in \citeauthor{krueger1999}'s mean regression analysis, the effect of being assigned a full-time aide is insignificant.
I now briefly discuss the other covariates. Black students perform worse than non-black students with otherwise identical characteristics; this is particularly pronounced between the first and third quartiles of the conditional score distribution, where black students' scores are about 10 percentage points lower. Girls generally score higher than boys, although the gap is quite small near the tails of the conditional score distribution. Poor students score up to 15 percentage points lower than otherwise identical students; however, this difference is much smaller near the top of the conditional distribution. As \citet{krueger1999} and earlier studies have found, teacher characteristics seem to matter little: their race and education (measured by whether they have a master's degree) have no significant impact. Another year of teaching experience has a small, positive effect for all but the very best students.
As a referee points out, an issue with Monte Carlo studies such as those in the preceding section is that the data sets used in simulations are likely to be quite different from real data sets. I therefore also evaluate the performance of the wild gradient bootstrap and the alternative methods introduced in Experiment \ref{ex:signi} above through placebo interventions in the Project STAR data.
\begin{experiment}[Placebo interventions]\label{ex:placebo}
For this experiment, I removed all small classes from the sample so that only 194 regular-size classes (with and without teacher's aide) in the 79 project schools remained. Of these schools, 16 had two regular-size classes without aide and 2 had three such classes.
\begin{table}[htp]
\caption{Rejection frequencies of $\mathrm{H}_0\colon \beta_1(.5) = 0$ in placebo interventions for different values of $\beta_1(.5)$}\label{tab:placebo}
\centering
{\small
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lp{0cm}ccp{0cm}ccp{0cm}ccp{0cm}cccccccc}%
\hline
&
& \multicolumn{2}{c}{Mammen}
&
& \multicolumn{2}{c}{Rademacher}
&
& \multicolumn{2}{c}{Webb}
&
& \multicolumn{2}{c}{Ana.}
& \multicolumn{2}{c}{Reg.}
& \multicolumn{2}{c}{Rank}
& \multicolumn{2}{c}{FHH}
\\
\cline{3-4}\cline{6-7}\cline{9-10}
&
& cv & se
&
& cv & se
&
& cv & se
&
& \multicolumn{2}{c}{se}
& \multicolumn{2}{c}{se}
& \multicolumn{2}{c}{score}
& \multicolumn{2}{c}{cv} \\
\hline
$\beta_1(.5) = 0$ (size) & & .072 & .084 & & .077 & .093 & & .085 & .095 & &
\multicolumn{2}{c}{.277} & \multicolumn{2}{c}{.284} & \multicolumn{2}{c}{.098} &\multicolumn{2}{c}{.311} \\
$\beta_1(.5) = 5$ (power) & & .413 & .427 & & .449 & .446 & & .450 & .450 & &
\multicolumn{2}{c}{.701} & \multicolumn{2}{c}{.714} & \multicolumn{2}{c}{.516} &\multicolumn{2}{c}{.727} \\
\hline
\end{tabular}%
}}
\end{table}
In each of these 18 schools, I then randomly assigned one of the regular-size classes without aide the treatment indicator $\mathit{small}=1$. This mimics the random assignment of class sizes within schools in the original sample, even though in this case no student actually attended a small class. Next, I reran the QR in \eqref{eq:star1} and tested, at the 5\% level, the correct null hypothesis that the coefficient on $\mathit{small}$ is zero at the median, $\mathrm{H}_0\colon \beta_1(.5) = 0$, using the same methods as in Experiment \ref{ex:signi}. The rejection frequencies in the first line of results in Table \ref{tab:placebo} show the outcome of repeating this process 1,000 times. The bootstraps were again based on $m=299$ simulations. As can be seen, the wild gradient bootstrap test from Algorithm \ref{al:binf} with the Mammen distribution outperformed all other methods of inference, some by a very large margin. Still, the test over-rejected slightly. This can be attributed to the fact that the treatment effect is now identified off of comparisons within only 18 instead of 79 schools, which makes the estimation problem much more challenging than in the actual data. The size of the tests in the placebo experiment can, in that sense, be viewed as an upper bound for the size of the tests in the original sample.
I also investigated power by increasing the percentile scores of all students in the randomly drawn small classes of the placebo experiment by 5. This increase is of the same order of magnitude as the estimated treatment effect at the median in the actual sample. Then I repeatedly tested the incorrect hypothesis $\mathrm{H}_0\colon \beta_1(.5) = 0$ (the correct value is $\beta_1(.5) = 5$) with the same experimental setup as before. The results are shown in the second line of Table \ref{tab:placebo}. Despite the now much smaller sample, the wild gradient bootstrap was able to reject the null in a large number of cases. The other methods rejected more often, but this was likely driven by their size distortion. Notable here is the high power of the \citet{wanghe2007} rank score test despite its relatively mild over-rejection under the null.
\hfill $\square$
\end{experiment}
\phantomsection%
\label{rev:intraclass}%
The large differences in the finite-sample size of the methods of inference considered in the preceding experiment can be attributed to the within-cluster dependence in the data. This is also supported by a back-of-the-envelope comparison of the results here to the Monte Carlo experiments in Section \ref{s:mc}. For the Monte Carlo DGP \eqref{eq:mc}, the within-cluster correlation coefficient of the outcome variable can be shown to be approximately $\varrho$. For the Project STAR data, the \citet{karlinetal1981} intraclass correlation coefficient $$ \hat{\varrho}_n := \frac{\sum_{i=1}^n\sum_{k=1}^{c_i}\sum_{l\neq k} (Y_{ik} - \bar{Y}_n)(Y_{il} - \bar{Y}_n)/(c_i -1)}{\sum_{i=1}^n\sum_{k=1}^{c_i} (Y_{ik} - \bar{Y}_n)^2},\quad\text{where}\quad\bar{Y}_n = \frac{\sum_{i=1}^n\sum_{k=1}^{c_i}Y_{ik}}{\sum_{i=1}^n c_i},$$ of the percentile score is $.319$. This is a consistent estimate of the within-cluster correlation coefficient of the percentile score as long as both its mean and within-cluster covariance structure are identical across clusters. (Neither of these conditions is needed for any of the theoretical results in this paper.) At $\varrho = \hat{\varrho}_n$, the results of Experiments \ref{ex:signi} and \ref{ex:placebo} are quite similar, with the exception that the rank score test performed much better in Experiment \ref{ex:placebo} than the test based on analytical cluster-robust standard errors.
Finally, before concluding this section, Figure \ref{fig:star2} illustrates the difference between a 95\% pointwise confidence interval based on a Powell sandwich estimator (as described in Experiment \ref{ex:signi}) that does not control for within-cluster correlation (dotted lines), the wild gradient bootstrap confidence interval shown in Figure \ref{fig:star1} (grey), and a 95\% wild bootstrap confidence band for the entire coefficient function of $\mathit{small}$ weighted by the bootstrap covariance matrix (dashed). As can be seen from the size of the grey area, not accounting for the possibility of peer effects and unobserved teacher characteristics via cluster-robust inference appears to give a false sense of precision at most quantiles. However, as the confidence band shows, we can conclude that the effect of the small class size is significantly positive over a large part of the support of the score distribution.
\begin{figure}[thp]
\begin{center}
\includegraphics[width=0.4\textwidth]{fig_star2}
\caption{QR coefficient estimate and pointwise confidence interval for $\mathit{small}$ from Figure \ref{fig:star1}, a 95\% pointwise confidence interval not robust to within-cluster correlation (dotted lines), and a 95\% wild bootstrap confidence band for the entire coefficient function (dashed).} \label{fig:star2}
\end{center}
\end{figure}
\section{Conclusion}\label{s:conc}
In this paper I develop a wild bootstrap procedure for cluster-robust inference in linear QR models. I show that the bootstrap leads to asymptotically valid inference on the entire QR process in a setting with a large number of small, heterogeneous clusters and provides consistent estimates of the asymptotic covariance function of that process. The proposed bootstrap procedure is easy to implement and performs well even when the number of clusters is much smaller than the sample size. A brief application to Project STAR data is provided. It is still an open question how cluster-level fixed effects that correspond to the intuitive notion of identifying parameters from within-cluster variation can fit into the present framework; this is currently under investigation by the author. Another question is if the jackknife can improve on the bootstrap in the current context; recent results by \citet{portnoy2014} for censored regression quantiles suggest this possibility.
\pre{
\section*{Supplementary Materials}
\begin{description}
\item[\rmfamily Supplementary appendix:] Technical appendix containing auxiliary results and proofs.
\end{description}
\phantomsection
\addcontentsline{toc}{section}{References}
\putbib[qspec]
\end{bibunit}
\newpage\setcounter{page}{1}
}
|
\section{Introduction}
The motion of particulate matter in fluid flows has long been a central problem in both fundamental and applied fluid mechanics.
The dynamics of spherical particles was a natural starting point for Stokes, who considered this problem in the 19th century~\cite{Stokes1851}. A long list of researchers continued work to describe the motion of spherical particles at low Reynolds number, leading to the Maxey-Riley-Gatignol equations~\cite{Maxey-Riley1983,Gatignol1983}.
The development of numerical simulations~\cite{Yeung1989, Squires1990} and experimental tools~\cite{Dracos1996, LaPorta2001, Mordant2001} capable of revealing the motion of spherical particles in complex flows has led to a recent resurgence of work on particle dynamics~\cite{Toschi2009}.
Extending the problem to non-spherical particles in general is important for many applications ranging from icy clouds~\cite{Pinsky1998}, to bio-locomotion ~\cite{Guasto2012}, to suspension flows in industrial settings~\cite{Lundell2011}.
In 1922, Jeffery analyzed the motion of axisymmetric ellipsoidal particles in Stokes flow~\cite{Jeffery1922}. Their tumbling rate can be expressed as:
\begin{equation}
\dot{p}_i = \Omega_{ij} p_j + \frac{\alpha^2-1}{\alpha^2+1} [S_{ij}p_j - p_i p_j S_{jk} p_k],
\label{eqn:jeffery}
\end{equation}
where $\mathbf{p}$ is a unit vector along the symmetry axis of the ellipsoid, $\alpha$ is the aspect ratio, $\mathbf{\Omega}$ is the anti-symmetric part of the velocity gradient tensor (vorticity), and $\mathbf{S}$ is the symmetric part of the velocity gradient tensor (strain rate).
A wide range of studies has explored deviations from Jeffery dynamics due to particle inertia, particle shape, and Reynolds number, among other factors~\cite{Leal1980}.
Analytic work~\cite{Szeri1992,Gustavsson2014} and numerical simulations~\cite{Shin2005,Marchioli2010,Pumir2011,Chevillard2013} have made significant progress extending studies of non-spherical particles in complex flows, but experimental measurements of the dynamics of anisotropic particles have lagged far behind due to the difficulty of measuring their time-dependent orientation in three dimensions.
Recently, methods have been developed for measuring time-resolved orientation and position of thin rods in 3D turbulence with stereoscopic optical imaging~\cite{Parsa2012}.
Another technique uses large transparent anisotropic particles with tracer particles inside and measures their rotations with particle image velocimetry~\cite{Bellani2012}.
In this paper, we introduce a new way to measure the orientations and rotational dynamics of anisotropic particles that behave like ellipsoids.
Bretherton has shown that many anisotropic particles have equivalent ellipsoids so that their tumbling rate follows \autoref{eqn:jeffery} with an effective aspect ratio~\cite{Bretherton1962a}.
We have identified particles of this class whose orientations can also be directly measured from stereoscopic imaging.
Three-dimensional printing allows us to create anisotropic particles made of mutually perpendicular thin rods.
Two perpendicular thin rods form a cross, while three perpendicular rods form a jack.
Arguments in Bretherton~\cite{Bretherton1962a} and resistive force theory calculations in Appendix A show that a cross rotates like a disk, an ellipsoid with aspect ratio $\alpha \ll 1$.
Similarly, a jack rotates like a sphere, which is an ellipsoid with $\alpha=1$.
We have used stereoscopic video imaging to directly measure the orientation of these objects as a function of time.
The methods developed in this paper thus allow us to obtain Lagrangian measurements of the full solid-body rotation rate.
Since a jack rotates like a sphere, its solid-body rotation rate couples only to the vorticity.
Thus, measurements of the rotation of neutrally buoyant Kolmogorov-scale jacks are direct Lagrangian vorticity measurements---a long sought goal of fundamental fluid mechanics.
Several other methods for measuring Lagrangian vorticity have also been developed.
For example, the vorticity optical probe measures the reflections from planar mirrors embedded in spherical particles~\cite{Frish1981}.
Stereoscopic imaging has been used to measure the rotation of large spheres by applying patterns to the sphere surface~\cite{Zimmermann2011a, Zimmermann2011b} or embedding small fluorescent tracers in transparent particles~\cite{Klein2013}.
Our technique of measuring the solid-body rotation of small jacks allows us to accurately measure the fluid vorticity using straightforward imaging methods.
\section{Experiment}
\label{sec:experiment}
\subsection{Printing 3D Particles}
\label{sec:particles}
We use 3D printing to fabricate anisotropic particles in the shapes of crosses and jacks in order to measure the dynamics of axisymmetric ellipsoids across the range of aspect ratios.
To print at both high resolution and in high quantity, we used a Connex 500 to make 10,000 of each particle shape.
Arm lengths were 3 mm, which is 6 times the Kolmogorov length scale in our flow.
The diameter of any given arm on a rod, cross, or jack is 300$\mu$m, the smallest we could achieve while maintaining the structural integrity of the particles.
To print particles with such small cylindrical arms, it was important to ensure that none of the arms were along the build axes.
Arms that lay along the vertical build axis had defects and would often break off, while arms lying in the horizontal plane tended to flatten.
Another important difficulty in printing $\mathcal{O}(10^4)$ particles is removing the support material.
Connex printers use a different material for the support structure and the printed particles. The support material can be partially dissolved using a strong base solution (e.g., NaOH) without affecting the particles themselves.
We found that using an ultrasonic bath made the removal process much more efficient, and the particles could be filtered out of the solution with almost no loss.
In order to have neutrally buoyant particles, the density of our fluid is matched to the particle density.
The density is adjusted by adding calcium chloride to water~\cite{Parsa2012}.
We used the print material VeroClear, whose bulk density was quoted at 1.17~g/cm$^3$.
However, we found that the manufacturer's quote differed significantly from the density of the fluid in which the particles were neutrally buoyant.
After the particles were immersed in the fluid for several hours, we found that different particles were neutrally buoyant at slightly different fluid densities ranging from 1.21 to 1.23~g/cm$^3$.
We chose the density of the fluid for the experiment as 1.22~g/cm$^3$ based on the population average.
More work is needed to understand the mass density distribution inside 3D printed objects, but our particles are sufficiently density matched that their rotations should accurately represent the neutrally buoyant case~\cite{Voth2002}.
To make the particles fluorescent, they were placed in a high concentration Rhodamine solution at elevated temperature (60-80$^\circ C$) for several hours.
\subsection{Turbulent flow between oscillating grids}
\label{sec:flow}
The turbulence is generated in a $1\times 1 \times 1.5$ m$^3$ octagonal prism using two parallel 8\,cm-mesh grids oscillating in phase (see \autoref{fig:tank}).
\begin{figure}
\centering
\includegraphics{Figure1.pdf}
\caption{Experimental setup (figure from \cite{Wijesinghe2012}). In the octagonal flow between oscillating grids, a central viewing volume in the focus of the four stereoscopically arranged video cameras is illuminated by a green Nd:YAG laser.}
\label{fig:tank}
\end{figure}
The measurements presented here were performed at a grid frequency of 1 Hz, which produced a flow with a root mean squared (rms) velocity $U \equiv \sqrt{\langle u_i u_i \rangle} /3= 20 $ mm/s and energy injection length scale, $L=U^3/\varepsilon=60$ mm so that $R_\lambda \equiv (15 U L/\nu)^{1/2} = 91$.
The energy dissipation rate, $\varepsilon$ = 133 mm$^2$s$^{-3}$, was measured from the mean square tumbling rate of jacks as described in \autoref{sec:results}.
The kinematic viscosity of the CaCl$_2$ solution that is approximately density matched to the particles is $\nu = 2.17$ mm$^2$/s.
The Kolmogorov length scale is $\eta = (\nu^3 / \varepsilon)^{1/4} = 526~\mu$m, and the Kolmogorov time scale is $\tau_\eta = (\nu / \varepsilon)^{1/2} = 128$~ms.
We chose this low Reynolds number to ensure that our particles were only 6$\eta$, where the mean square rotation shows only a small deviation from the tracer limit ~\cite{Shin2005,Parsa2014}.
We use four stereoscopic cameras (1280$\times$1024 at 450 Hz) with a custom real-time image compression system~\cite{Chan2007, Wijesinghe2012} to allow continuous imaging.
A green Nd:YAG laser with 50 W average power illuminates a detection volume of roughly $3 \times 3 \times 3$~cm$^3$ at the center of the tank, where the flow is quite homogeneous~\cite{Blum2010a}.
We used two perpendicular expanded beams in the horizontal plane, each reflected back upon itself, in order to provide 4 directions of illumination and minimize shadowing of some arms of a particle by other arms.
With an average particle number density of only $5\times 10^{-3}$~cm$^{-3}$, there was a particle in view less than 20\% of the time, which made the image compression system essential for acquiring enough trajectories to converge statistics~\cite{Chan2007}.
\subsection{Image Analysis.}
\label{sec:analysis}
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{Figure2.pdf}
\caption{A sample image of a jack from each of the four cameras. The pairs of red and green asterisks denote the two ends of an arm determined by the orientation-finding algorithm.}
\label{fig:projex}
\end{figure}
\emph{Orientation.}---An example of a jack imaged by all four cameras is shown in \autoref{fig:projex}.
When a particle is visible on all four cameras, we find the three dimensional position of the particle using stereomatching methods~\cite{Ouellette2006}.
We developed a nonlinear fitting algorithm to determine the orientation of a particle from a set of stereoscopic images.
Any orientation is specified by a rotation matrix, $\op{O}$, which can be parameterized by three Euler angles $(\phi, \theta, \psi)$~\cite{Goldstein}.
From a measurement of the particle's center and an initial guess of its orientation, we construct a model of the particle and project it onto each of the four image planes using the calibrated camera parameters~\cite{Tsai1987}.
The total difference in intensity between a model image and an experimental image provides the residual that is minimized by a nonlinear search in Euler angle space.
To minimize computational cost, we project the endpoints of each arm onto the image plane of each camera, and then model the intensity distributions in two dimensions.
The model image of an arm is formed by a Gaussian intensity distribution along the width of each arm (in 2D) and a Fermi intensity distribution across the length of the arm.
For jacks and crosses, each arm in the model has an identical intensity distribution.
As can be seen from \autoref{fig:projex}, the arms in the experimental images are not of uniform intensity.
The observed intensity has a non-trivial dependence on the angles between the arms, the illumination, and the viewing direction~\cite{Parsa2014NPr}, which has not yet been included in our model. However, we found the simple model adequate enough for our purposes.
The orientation-finding algorithm must be seeded with an initial guess for the orientation.
Except for the first frame of a track, we use the previous frame as the initial guess, since the rotation between frames is less than 0.01 radians.
The initial guess for the beginning of a particle track is reliably generated using an optical tomographic reconstruction algorithm~\cite{Elsinga2006}.
We also compared tomographic reconstruction as a method for finding particle orientation, but we found it to be both less accurate and more computationally expensive than our own algorithm.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{Figure3.png}
\caption{\textbf{(a)} A reconstructed trajectory of a cross in three-dimensional turbulence.
The two different color sheets trace out the path of the particle through space and time.
The length of the particle track is 336 frames, or $5.7 \,\tau_\eta$. A cross is shown every 15 frames.
\textbf{(b)} A reconstructed trajectory of a jack in three dimensional turbulence.
The three different colors distinguish the arms of the jack and trace out their path as the particle rotates.
The dark green line denotes the trajectory of the jack's center.
The length of the particle track is 1025 frames, or $17.5 \,\tau_\eta$.
A jack is shown every 50 frames. (\emph{Note:} neither the crosses nor the jacks shown above are drawn to scale.)}
\label{fig:cross+jack_track}
\end{figure}
The Euler angles found for a jack give 1 of the 24 orientations related by symmetry.
To see this, consider 1 of the 6 arms of a jack and define it by the vector ${\vec{p}}=\hat{\vec{z}}$; there are 4 symmetrically identical orientations obtained by rotations of $\pi/2$ about the $z$ axis.
There are 4 such orientations for each of the 6 arms, for a total for 24.
A cross has 8 symmetrically identical orientations, and a rod has 2.
We ensure that we have a consistent series of orientations along a single trajectory by comparing the orientation found in each frame with that from the previous frame, and choose the orientation of the particle that produces the smallest total rotation between frames.
\autoref{fig:cross+jack_track}(a) shows a representative track for a cross, which is $5.7\,\tau_\eta$ long.
It demonstrates the effectiveness of our algorithm at determining the full range of orientations.
At various points along the track, the rotation of the cross about a vector nearly coplanar with its two arms is clearly visible.
In \autoref{fig:cross+jack_track}(b) we also show an example jack trajectory that is $17.5\,\tau_\eta$ long.
\emph{Solid-body rotation rate.}---Once we have measurements of orientation and position at each time step, the natural quantity to consider is the rate of change of the unit vector defining the particle's orientation, which we call the tumbling rate, $\dot{\vec{p}}$.
For crosses and jacks, we can also measure the full solid-body rotation rate vector, $\vec{\omega}_s$, from a series of orientation measurements smoothed along the particle's trajectory.
One method for doing this has been described in~\cite{Zimmermann2011b}.
We take a different approach using the tools we already developed for least squares optimization in Euler angle space.
The problem can be framed as finding the initial orientation matrix, $\op{O}(t_i)$ and the rotation matrix over a single time step, $\mathbb{R}$, that together give the particle orientation matrix as a function of time,
\begin{align}
\op{O}(t) = \mathbb{R}^{\frac{t-t_i}{\tau_f}}\op{O}(t_i),
\end{align}
where $\tau_f$ is the period between images.
A non-linear least squares fit is used to find the Euler angles for the matrices $\op{O}(t_i)$ and $\mathbb{R}$ that best match the measured orientation matrices.
Then, the rotation matrix $\mathbb{R}$ can be decomposed as a rotation by an angle $\Phi$ about the solid body rotation axis, $\hat{\vec{\omega}}_s$, in accordance with Euler's theorem~\cite{Goldstein}, from which we obtain the magnitude of the solid body rotation rate, $\omega_s = \Phi / \tau_f$.
The solid-body rotation rate, $\vec{\omega}_s$, is related to the tumbling rate by $\dot{\vec{p}}=\vec{\omega}_s \times \vec{p}$.
The difference between the two quantities is that $\dot{\vec{p}}$ does not depend on the vector component of $\vec{\omega}_s$ lying along $\vec{p}$.
We use this relationship to determine the tumbling rate from measurements of particle orientation and their solid-body rotation rate.
In order to correct for the contributions from orientation measurement errors to a measurement of the mean square tumbling rate, one needs to measure $\dot{\vec{p}}$ over a a range of fit-lengths, $\tau$~$=t-t_i$.
\autoref{fig:ppfl} shows our measurement of $\langle \dot{p}_i \dot{p}_i \rangle$ as a function of $\tau$. The solid lines are fits of the function,
\begin{align}
f(\tau) = A \tau^B + C \exp \left( D \tau + E \tau^2 \right),
\end{align}
where the first term models the random error that dominates at small $\tau$ and the second term models the effect of filtering experimental measurements of orientation~\cite{Voth2002}.
The fit parameter $C$ gives an estimate of the function value at $\tau=0$ if there were no random errors.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{Figure4.pdf}
\caption{Measurements of the mean square tumbling rate, $\langle \dot{p}_i \dot{p}_i \rangle $, as a function of the fit length for both jacks (blue) and crosses (red). To determine the true value, we extrapolate to zero fit-length by fitting the data to a stretched exponential~\cite{Voth2002}. }
\label{fig:ppfl}
\end{figure}
\section{Results}
\label{sec:results}
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{Figure5.pdf}
\caption{The PDF of the mean square tumbling rate for our experimental measurements of crosses (red squares) and jacks (blue circles) as well as direct numerical simulations of spheres (solid line).}
\label{fig:PDF_pd2}
\end{figure}
From the measured solid-body rotation rates of many jacks and crosses, we can obtain the probability density function (PDF) of the squared tumbling rate, which is shown in \autoref{fig:PDF_pd2}.
Also shown is the PDF obtained from direct numerical simulations of spheres~\cite{Parsa2012}. All three PDFs agree within experimental uncertainties.
Numerical work has shown that there should be slightly more probability density in the tail of the PDF for rods and disks than there is for spheres ~\cite{Parsa2012}. For large tumbling rates, our data shows the opposite, with jacks having slightly higher probability than crosses, although the sources contributing to this small discrepancy are known.
The error bars shown in \autoref{fig:PDF_pd2} account for random error as well as the systematic error that results from the fit-length dependence of the tumbling rate measurements.
An additional source of error that is not included in the error bars is the self-shadowing of particles as they pass through certain orientations dependent on the camera configuration.
The most dramatic cases are when an entire arm is missing from each of the four cameras, such that a jack, for example, will look like a cross along part of a given trajectory.
The reduction in accuracy in determining these particular orientations occasionally leads to erroneously high measurements of the solid-body rotation rate, which pushes additional probability density towards the tail of the PDF.
This effect is stronger for jacks than for crosses, which is why the jack PDF is slightly higher in our measurements at large tumbling rates.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{Figure6.pdf}
\caption{Mean square tumbling rate as a function of aspect ratio. Solid squares are our data for crosses and jacks. The other data is from \protect{\cite{Parsa2012}}. Plus symbols are numerical simulations. The triangle and circle are experimental measurements of rods at $R_\lambda=214$ and 160. The solid line is an analytic result for randomly oriented ellipsoids. }
\label{fig:pd2_alpha}
\end{figure}
Previous work by Parsa \emph{et al.} showed that the mean square tumbling rate for axisymmetric ellipsoids, $\left< p_i p_i \right>$, is strongly affected by alignment of anisotropic particles by turbulence~\cite{Parsa2012}.
They found agreement between simulations and experimental measurements of rods, but particles with a wider range of aspect ratios could not be measured experimentally with the tools available at that time.
We have now measured the mean square tumbling rate of crosses and jacks using the techniques described in \autoref{sec:experiment}.
In \autoref{fig:pd2_alpha}, we show the mean square tumbling rate normalized by the Kolmogorov timescale as a function of aspect ratio.
Our measurements show crosses tumbling at a considerably higher rate than jacks or rods, but still more slowly than the randomly-oriented prediction. There is good agreement with simulations across the full range of aspect ratios.
In turbulence experiments, it is always a challenge to measure the energy dissipation rate, which appears in the normalization of the vertical axis in \autoref{fig:pd2_alpha}.
We attempted to make independent measurements using non-fluorescent tracer particles, but did not succeed because of reduced light scattering from tracers in the density-matched CaCl$_2$ solution.
However, the measurements of jack rotations provide a new way to measure the energy dissipation rate.
Because they rotate like spheres, they give a direct measurement of the vorticity, and in isotropic turbulence their vorticity is directly related to the energy dissipation through $\langle \Omega_{ij} \Omega_{ij} \rangle$=$\langle S_{ij}S_{ij} \rangle=\varepsilon/\nu$. This implies that for spheres $\langle \dot{p}_i \dot{p}_i \rangle = \frac{1}{6} \varepsilon/\nu$~\cite{Parsa2012}.
We use this to determine our energy dissipation rate.
This makes our jack data at $\alpha=1$ match the simulations by definition, and the agreement of the cross data at aspect ratio $\alpha=0.1$ with numerical simulations is an independent result of the measurement.
The solid green curve in \autoref{fig:pd2_alpha} gives the mean square tumbling rate as a function of aspect ratio for randomly oriented ellipsoids.
While the tumbling rate of jacks is unmodified from the randomly oriented case, both rods and crosses show smaller tumbling rates due to the effects of alignment by turbulence.
Shin and Koch were the first to notice that the tumbling rate of rods that have correlated with a turbulent flow is reduced in comparison to that of randomly oriented rods~\cite{Shin2005}.
More recent numerical work showed that the effects of alignment persist across the full range of aspect ratios~\cite{Parsa2012}.
The leading-order effects are contained in the Lagrangian three-point correlations of the velocity gradients, which imply higher tumbling rates for disks than for rods~\cite{Gustavsson2014}.
A number of studies on the dynamics of rods show that they align with the vorticity, and this has been used to explain the slower tumbling rate of rods since the component of the vorticity along the rod axis does not contribute to its tumbling rate~\cite{Pumir2011, Chevillard2013, Wilkinson2012, Ni2014}.
We find that alignment with vorticity is also responsible for the reduction of the tumbling rates for crosses.
We can directly measure the preferential alignment of a particle by measuring the angle between the orientation of the particle and its solid body rotation rate vector.
In \autoref{fig:cospw_all}, we plot the PDF of the magnitude of the cosine of this angle, $\lvert \vec{p} \cdot \hat{\vec{\omega}}_s \rvert$, for both crosses and jacks, and compare them with numerical simulations.
The peak near $\lvert \vec{p} \cdot \hat{\vec{\omega}}_s \rvert=0$ for crosses in \autoref{fig:cospw_all}(a) shows that disks preferentially align with $\vec{p}$ perpendicular to $\vec{\omega}_s$.
\autoref{fig:cospw_all}(b) confirms the same story as \autoref{fig:cospw_all}(a).
It shows that an arm of a cross is preferentially aligned with the solid body rotation rate vector.
The height of the peak in \autoref{fig:cospw_all}(b) is lower than that in \autoref{fig:cospw_all}(a) because any vector in the plane of the disk is equally likely to align with $\vec{\omega}_s$.
In a turbulent flow, disks rotate like a spun coin rather than a tossed frisbee.
For jacks in \autoref{fig:cospw_all}(c), there is no preferential alignment because they rotate like spheres. Numerics and experiment are in quite good agreement.
The deviations near $\lvert \vec{p} \cdot \hat{\vec{\omega}}_s \rvert=0$ for crosses are likely the result of measurement error in the solid body rotation rate vector.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{Figure7.pdf}
\caption{The PDF of the alignment between a particle's orientation and its solid body rotation rate, $|\hat{\vec{p}} \cdot\hat{\vec{\omega}}_s|$, for (a) crosses and (c) jacks. Symbols are experimental measurements and solid lines are numerical simulations. (b) shows the alignment of one of the arms of a cross, $\vec{p}^\prime$, with its solid body rotation rate vector. }
\label{fig:cospw_all}
\end{figure}
The observed alignment in \autoref{fig:cospw_all} seems natural if it is thought of as a result of the Lagrangian stretching of the fluid. Stretching will align both the vorticity and a long axis of a particle with the stretching direction~\cite{Ni2014}. For rods this means that $\vec{p}$ aligns with the vorticity, which decreases the tumbling rate.
Because disks have their long axis perpendicular to $\vec{p}$, they preferentially have $\vec{p}$ perpendicular to the vorticity, which makes $\vec{p}$ also preferentially perpendicular to the solid-body rotation rate. This creates a larger tumbling rate for disks since $\vec{\dot{p}}=\vec{\omega}_s \times \vec{p}$, consistent with the data in \autoref{fig:pd2_alpha}.
Although disks tumble faster than rods, they still tumble slightly slower than if they were randomly oriented, as seen in~\autoref{fig:pd2_alpha}.
Since disks have a much smaller difference from the randomly oriented case, one might conclude that disks are less strongly aligned than rods by the turbulence.
However, this is largely a result of the way $\dot{\vec{p}}$ is defined.
We have measured the normalized mean square tumbling rate of a unit vector along one of the arms of a cross as $0.12 \pm 0.02$, which is smaller than the tumbling of either crosses or jacks and the same as measurements for rods within experimental error.
This explicitly indicates that crosses (disks) are also strongly aligned by turbulence, and it is only the definition of $\vec{p}$ as perpendicular to the two arms that increases the mean square tumbling rate so much in comparison to rods.
It is not easy to directly compare the degree of alignment of rods and disks because disks have an entire plane that can align with the stretching, while rods have only a director.
The picture that emerges from our data and previous work is that anisotropic particles are aligning with the Lagrangian stretching direction, and this suppresses the tumbling rate for all particles except nearly spherical oblate ellipsoids.
The amount of suppression is strongly dependent on the axis whose tumbling rate is considered.
\section{Conclusions}
We have developed a method for measuring the time-resolved Lagrangian orientation and solid body rotation rate of anisotropic particles in a turbulent flow.
By measuring the rotation of 3D printed jacks and crosses we are able to extend previous measurements of rods to cover the full range of aspect ratios of axisymmetric ellipsoids.
Moreover, we have provided a way to directly probe Lagrangian vorticity with a single particle measurement, which has potential for application in a wide range of flows and Reynolds numbers.
We find that the mean square tumbling rate, $\langle \dot{p}_i \dot{p}_i \rangle$, agrees with DNS data at points spanning the full range of aspect ratios.
Our measurements show that crosses are preferentially aligned in turbulence with their orientation axis perpendicular to their solid body rotation rate vector.
To the best of our knowledge, this is the first direct experimental measurement of the preferential alignment of particles with their rotational motion in a turbulent flow.
Our results support a natural picture of alignment in turbulence where particles have their long axes aligned with the Lagrangian stretching direction of the fluid flow.
|
\section{INTRODUCTION}
\label{sec:introduction}
Terrestrial regions of planetary systems are not well studied. The majority of known exoplanets were discovered using the radial velocity and transit techniques, which are biased to massive and very short-period planets. While some rocky planets have now been discovered in the terrestrial zone (e.g. Kepler-186f; \citealp{quintana2014}) their frequency and statistical properties can only be estimated by extrapolation from planets nearer their stars (e.g. \citealp{petigura2013}). Direct imaging of exoplanets, on the other hand, is currently limited to massive planets on wide orbits outside of the terrestrial zone.
Observations of debris disks provide an alternative method to study planetary systems. Debris disks are the results of the collisional processing of the solid material left over from planet formation, and their locations may be gravitationally influenced by unseen planets (for a recent review of debris disks, see \citealp{matthews2014}). The disks are often partitioned into concentric components, and it is useful to categorize these components by equilibrium temperature \citep{su2014}. ``Cold" components ($\lesssim$100 K), located in the outer parts of planetary systems, are analogous to the Solar System's Kuiper Belt and may trace the radial limits of planet formation or migration \citep{ballering2013}. ``Warm" components ($\sim$200 K) are analogous to the asteroid belt in the Solar System, and their locations may be set by the water ice line \citep{morales2011}. ``Very hot" components ($\sim$1000 K) trace small refractory dust grains located very near to the star (e.g. \citealp{absil2013}). Between the very hot and warm components lies the ``hot" component ($\sim$300 K), also referred to as the terrestrial zone. Dust in this zone may be analogous to the zodiacal dust in the Solar System, likely brought inward from Jupiter Family comets and the main asteroid belt \citep{nesvorny2010}. Studying the terrestrial regions of planetary systems via observations of exozodiacal dust is the subject of this paper.
Debris in terrestrial zones has been largely inaccessible to observation. The {\it Herschel} and {\it Spitzer} space telescopes have imaged disks at mid- to far-IR wavelengths (e.g., \citealp{su2005}; \citealp{su2008}; \citealp{booth2013}; \citealp{morales2013}), but the resolution of these telescopes is not sufficient to resolve the terrestrial regions of these systems. The Atacama Large Millimeter/sub-millimeter Array (ALMA) has high resolution, but works at sub-mm and radio wavelengths so is primarily sensitive to cold components. Debris disks have also been imaged via the starlight they scatter in the visible and near-infrared by the {\it Hubble Space Telescope} (e.g. \citealp{kalas2005}; \citealp{schneider2009}; \citealp{soummer2014}) and large ground-based telescopes equipped with advanced adaptive optics systems (e.g. \citealp{buenzli2010}; \citealp{currie2012}; \citealp{rodigas2014}), but current scattered light observations are limited to regions outside the terrestrial zone by the high dust/star contrast and small inner working angles required. Interferometric observations in the near-infrared have detected very hot dust around some stars (see the current list of detections in \citealp{vanlieshout2014}), but this material resides nearer to the star than the terrestrial region. A promising method to spatially resolve dust in terrestrial regions is interferometry at $\lambda \sim$ 10 $\micron$, such as the detections of exozodiacal dust around $\eta$ Crv \citep{millangabet2011}. A large program to expand such results is being undertaken with the Large Binocular Telescope Interferometer (LBTI), although it will be limited to relatively nearby stars \citep{hinz2009}. The study of exozodiacal dust presented in this paper is complementary to these interferometric observations.
Debris disk characteristics around many stars have been inferred through their spectral energy distributions (SEDs) that show infrared flux in excess of that from the star's photosphere. SED studies, using data primarily from {\it Spitzer} and {\it Herschel}, have succeeded in discovering and characterizing hundreds of debris disks \citep{morales2011,ballering2013,chen2014}. These studies classify disks by temperature and find that they consist of cold components, warm components, or both. However, these components are usually colder than 300 K, which is the characteristic temperature expected for dust in the terrestrial zone.\footnote{The recent analysis by \citet{chen2014} shows some components reaching higher temperatures.}
Conversion of the apparent temperature of a debris disk to its orbital location is uncertain, as there is a degeneracy between the distance from the star and the optical and physical properties of the dust (minimum grain size, grain size distribution, and grain composition). The degeneracy arises from small grains that are superheated above their equilibrium temperatures because they absorb visible starlight more efficiently than they cool by emitting longer wavelength radiation. \citet{booth2013} resolve several cold debris belts with {\it Herschel} and find that they can be up to 2.5 times further from the star than predicted by blackbody fits to their SEDs. \citet{rodriguez2012} find that the disk orbital radii can be up to 5 times their blackbody radii. The degree of difference between the true location of a debris disk and that derived from the SED-measured dust temperature is likely not uniform for all disks, as the size of the smallest grains in a system is determined by the radiation forces exerted on them by the central star and by the collisional and dynamical processes occurring in the system. These uncertainties make it difficult to determine whether or not warm dust is located in the terrestrial zone. Furthermore, this degeneracy operates such that debris disk components seem to be nearer to their stars than they actually are because grains tend to be warmer than their equilibrium blackbody temperatures. Because most warm components have measured temperatures less than 300 K, this degeneracy makes it unlikely that these components are probing the terrestrial zones.
While the emitted flux density of most debris disks can be well modelled with one or two blackbody functions, a minority of disks show solid-state emission features in their spectra, most prominently at $\sim$10 $\micron$ and $\sim$18 $\micron$ from Si-O stretching and O-Si-O bending vibrations in the silicate material, respectively. These features only arise from warm, small (sub-$\micron$ to few-$\micron$) grains. By using the extra information present in the emission features, the degeneracies in modelling a debris disk SED are broken, and the location of the disk and its grain properties can be more accurately determined.
At least sixteen warm debris disks with prominent spectral features have been studied. These include: HR3927 \citep{chen2006}, $\eta$ Crv \citep{chen2006,lisse2012}, HD113766 \citep{chen2006,lisse2008,olofsson2012,smith2012,olofsson2013}, HD172555 \citep{chen2006,lisse2009,smith2012,johnson2012}, $\eta$ Tel \citep{chen2006,smith2009}, HD69830 \citep{beichman2005,lisse2007,beichman2011,olofsson2012}, BD+20 307 \citep{song2005,weinberger2011,olofsson2012}, HD15407A \citep{melis2010,fujiwara2012,olofsson2012}, HD169666 \citep{moor2009,olofsson2012}, [GBR2007] ID8 \citep{meng2012,olofsson2012}, EF Cha \citep{rhee2007b,currie2011}, HD145263 \citep{honda2004}, HD165014 \citep{fujiwara2010}, HD23514 \citep{meng2012,rhee2008}, HD72905 \citep{beichman2006}, and the very well-studied debris disk around $\beta$ Pic \citep{telesco1991,knacke1993,okamoto2004,chen2007,li2012}. In general, it has been concluded that these are exceptional systems, in many cases possibly the sites of elevated dynamical activity that has temporarily boosted the amount of very small dust in their debris systems.
Here we present $22$ additional warm debris disks with evidence for silicate emission features in their {\it Spitzer} Infrared Spectrograph (IRS; \citealp{houck2004}) data. The presence of a warm debris disk component was previously known to exist around these stars, but no analysis of their spectral features has been published\footnote{One exception is HIP86305. \citet{morales2013} created a model for this system that does reproduce the IRS emission feature, as we discuss in $\S$\ref{sec:specifictargets}}. The large number of new detections of features implies that such behavior is not exceptional. We fit to each spectrum physically-motivated SED models and determine the disk location and grain properties. We find that the locations of these disks can be well constrained, and that they are probing dust in the terrestrial regions of these systems. That is, analysis of subtle silicate features can be used to probe the terrestrial zones in many warm debris disks.
\section{METHODS}
\label{sec:methods}
\subsection{Target Selection}
\label{sec:targetselection}
For stars identified in \citet{ballering2013} to host warm debris disk components, we inspected the IRS data for signs of spectral features. We limited our search to targets with data available from all four IRS low-resolution spectral orders (LL1, LL2, SL1, and SL2). We found $22$ targets with signs of features out of the 106 warm components that had all four spectral orders; \citet{ballering2013} identified 125 total targets with warm components. The stars in our sample range in spectral type from B9 to F7, and their properties are given in Table \ref{table:targetlist}.
We inferred $T_\star$ and $M_\star$ from the known spectral types according to the tabulated values (or interpolations between those values) from \citet{carrollandostlie2006}. $L_\star$ values were computed from the bolometric magnitudes, as $L_\star = 10^{-0.4(M_\text{bol}-4.74)}L_\sun$, where $M_\text{bol} = V - 5\log_{10}(D) + 5 - A_V + BC$ and $A_V = 1.15(V-K-(V-K)_0)$. $BC$ is the bolometric correction inferred from the spectral type according to \citet{carrollandostlie2006}, and $(V-K)_0$ is the intrinsic color inferred from the spectral type according to \citet{cox2000}. We computed $R_\star$ from $T_\star$ and $L_\star$ using the Stefan-Boltzmann Law. MIPS 24 $\micron$ flux density values were taken from \citet{ballering2013}.
We used stellar ages from \citet{ballering2013} when available. These were estimated by combining chromospheric activity, x-ray emission, placement on the HR diagram, surface gravity, membership in clusters and associations, and gyrochronology. The references for these measurements are given in Table \ref{table:targetlist}. We also provide a quality flag for the age accuracy, giving the number of independent age measurements with good agreement. When ages were not available from \citet{ballering2013}, we found age references in the literature from studies that used reliable HR diagram fitting \citep{nielsen2013,zorec2012,chen2014}. For these targets, the age uncertainty can be large, sometimes $\gtrsim 50 \%$. \citet{zorec2012} provided ages in terms of the fraction of the main sequence lifetime; we obtained total main sequence lifetimes for these targets from Table 45 of \citet{schaller1992}, the $M_\star$ versus main sequence lifetime relation for the stellar evolution models employed by \citet{zorec2012}. We used $M_\star$ values for our targets from \citet{zorec2012} when using this table.
\subsection{IRS Data Reduction}
\label{sec:irsdatareduction}
The IRS Astronomical Observation Requests (AORs) for our targets are listed in Table \ref{table:targetlist}. The basic reduction was performed using the Spectroscopic Modeling Analysis and Reduction Tool (SMART) software package \citep{higdon2004}, as detailed in \citet{ballering2013}. In summary, bad pixels were removed using IRSCLEAN, multiple Data Collection Events (DCEs) for each nod position were combined, the background was removed from each 2D spectrum by subtraction of the opposite nod, the 2D spectra were converted into 1D spectra using optimal 2 nod extraction \citep{lebouteiller2010}, and the 1D spectra from both nods were combined. The result was a wavelength, flux, and uncertainty vector of each spectral order for each AOR. The ``bonus" third order data were not used. We compared our results with the reduction provided by the Cornell AtlaS of Spitzer/IRS Sources\footnote{The Cornell Atlas of Spitzer/IRS Sources (CASSIS) is a product of the Infrared Science Center at Cornell University, supported by NASA and JPL. http://cassis.astro.cornell.edu} (CASSIS; \citealp{lebouteiller2011}) to check that any structures in our spectra -- potential spectral features -- were not unique to our reduction procedure. We found no serious discrepancies between the two reductions, although there was typically a systematic difference in the absolute flux level.
We used the MATLAB software package for subsequent data reduction and analysis. We examined each spectral order and trimmed data from the ends, where the data are less reliable. The exact location of the trimming was determined individually for each target by eye, although we ensured some degree of overlap remained between adjacent orders. Next, we cut outlying data points from each order. To do this, we fit each spectral order with a third-degree polynomial, calculated the standard deviation of the residuals around this fit, and then discarded points lying more than three standard deviations from the fit. In practice, this process removed very few data points (primarily large outliers). We avoided a more aggressive cutting procedure as we did not want to erase any signs of emission features from our data. We corrected flux offsets between the orders by applying a multiplicative correction factor to the LL1, SL1, and SL2 flux values to bring them in line with the LL2 data and with each other. These corrections were determined by eye, and were typically less than 5\% and almost always less than 10\%. We combined the data from the four orders by interleaving the data at the overlapping regions and then smoothing the entire spectrum by binning to a wavelength resolution of 0.03 $\micron$. The silicate features we found are in the middle of the SL1 and LL2 wavelength coverages so they are not significantly affected by the order matching procedure. We then normalized the entire IRS spectrum to agree with the measured Multiband Imaging Photometer for {\it Spitzer} (MIPS; \citealp{rieke2004}) flux at 24 $\micron$, as the absolute calibration of MIPS is known to be more accurate than that of IRS.
With reduced IRS data in hand, we used a variety of empirical fitting approaches that indicated the presence of features roughly at the positions expected for silicate emission. The features were generally too weak to detect when simply viewing the data by eye -- they only became evident after subtracting the contribution from the stellar photosphere. We set out to confirm the signs of features by fitting the spectra with physically-motivated disk models capable of reproducing the emission features. We present the details of our model fitting procedure in the following section.
\subsection{Model Fitting}
\label{sec:modelsedfitting}
To explore the location of the dust grains producing the emission features, we carried out fits to the debris disk spectra, as discussed below. In summary, we found that fitting with a single dust belt almost always resulted in a ``belt" so broad that a better physical explanation would be two belts, and for 9 of the sources single belt models could not even produce acceptable fits. We therefore fit all the systems with two debris belts. These fits indicate that the features arise from fairly narrow rings, which, as we show in $\S$\ref{sec:terrestrialzones}, are largely confined to the terrestrial zones around these stars.
The observed flux from a single dust grain is given by
\begin{equation}
\label{eq:grainflux}
F_\nu(\lambda,a,T_\text{d}) = \left(\frac{a}{D}\right)^2 Q_\text{abs}(\lambda,a) \pi B_\nu(\lambda,T_\text{d}),
\end{equation}
where $D$ is the distance to the system, $a$ is the grain size, $T_\text{d}$ is the temperature of the dust grain, and $B_\nu$ is the blackbody function. $Q_\text{abs}(\lambda,a)$ is the efficiency at which a dust grain absorbs and emits light, which depends on the dust composition. We assumed that all dust grains were composed of amorphous olivine (MgFeSiO$_4$), and we obtained the optical constants $(n,k)$ for this material as a function of wavelength from 0.2 to 500 $\micron$ from \citet{dorschner1995}. We then used the Mie Theory code \texttt{miex} \citep{wolf2004} with these optical constants to compute $Q_\text{abs}(\lambda,a)$ for a range of grain sizes.
The dust temperature is set by the energy balance of absorbed and emitted stellar radiation,
\begin{align}
\label{eq:grainbalance}
& \int_0^\infty \left(\frac{R_\star}{r}\right)^2 \pi a^2 Q_\text{abs}(\lambda,a) \pi B_\lambda(\lambda,T_\star) \,\mathrm{d}\lambda \nonumber\\ &= \int_0^\infty 4\pi a^2 Q_\text{abs}(\lambda,a) \pi B_\lambda(\lambda,T_\text{d}) \,\mathrm{d}\lambda,
\end{align}
where $r$ is the distance between the dust grain and the star. This equation cannot be solved explicitly for $T_\text{d}$, however it can be solved for $r$:
\begin{equation}
\label{eq:graindistance}
r = \frac{R_\star}{2}\sqrt{\frac{\int_0^\infty Q_\text{abs}(\lambda,a) B_\lambda(\lambda,T_\star) \,\mathrm{d}\lambda}{\int_0^\infty Q_\text{abs}(\lambda,a) B_\lambda(\lambda,T_\text{d}) \,\mathrm{d}\lambda}}.
\end{equation}
For each target in our sample, we computed $r$ over a grid of input $a$ and $T_\text{d}$ values. We then inverted the tabulated results to find $T_\text{d}(a,r)$. The integrals in Equation \ref{eq:graindistance} were carried out using MATLAB's \texttt{trapz} function with 200 wavelength values logarithmically spaced between 0.2 and 500 $\micron$. When generating models we checked that no dust grains reached temperatures above 1550 K, olivine's sublimation temperature.
We assumed the dust grains were distributed in a ring between $r_\text{in}$ and $r_\text{out}$ with surface density $\Sigma(r) \propto r^{-q}$. We also assumed that the grain size distribution was $n(a) \propto a^{-p}$ from $a_\text{min}$ to $a_\text{max}$, and was identical at all $r$. The total flux from a debris belt is then
\begin{align}
\label{eq:beltflux}
F_{\nu,\text{belt}}(\lambda) = &A \int_{a_\text{min}}^{a_\text{max}} \int_{r_\text{in}}^{r_\text{out}} \left(\frac{a}{D}\right)^2 Q_\text{abs}(\lambda,a) \nonumber\\ &\times \pi B_\nu(\lambda,T_\text{d}) (2 \pi r) r^{-q} a^{-p} \,\mathrm{d}r \, \mathrm{d}a.
\end{align}
When generating models, we evaluated Equation \ref{eq:beltflux} by summing the integrand in 500 $\times$ 500 bins in $r$ and $a$, distributed logarithmically between the maximum and minimum values. The normalization $A$ was set so that the total mass of dust represented by the model was $10^{-10} M_\sun$. We tested our procedure by comparing our models with those generated by the Debris Disk Radiative Transfer Simulator\footnote{http://www1.astrophysik.uni-kiel.de/dds/index.html} (DDS; \citealp{wolf2005}). We found a very good agreement between the resulting theoretical spectra.
Because we had SL IRS data at wavelengths as short as $\sim$5 $\micron$, and the infrared excess generally arose at somewhat longer wavelengths, our spectra provided sufficient information to determine the brightness of the stellar photosphere without relying on photometry from other instruments at shorter wavelengths. Thus, we modelled the photosphere and the excess together, and the normalization of the photosphere was included as a free parameter in our fits. The stellar photosphere was assumed to emit as a blackbody of temperature $T_\star$. A blackbody is appropriate because the stars in our sample, mostly A and F types, have virtually no spectral features in the mid-IR.
First, we attempted to fit our data with the stellar photosphere plus a single belt model. We limited the number of free parameters by fixing the grain size distribution exponent to $p=3.65$, as suggested by \citet{gaspar2012}. We also fixed the maximum grain size to $a_\text{max}=1000$ $\micron$, as the largest grains in a disk generally contribute little to the total flux at these wavelengths, making this parameter difficult to constrain. We limited $q$ to vary from 0 to 2. Thus, the form of our model was
\begin{align}
\label{eq:onemodel}
F_{\nu,\text{model}}(\lambda) &= C_pB_\nu(\lambda,T_\star) \nonumber\\ &+ C_0F_{\nu,\text{belt}}(\lambda,r_\text{in},r_\text{out},q,a_\text{min}),
\end{align}
with six free parameters $r_\text{in}$, $r_\text{out}$, $q$, $a_\text{min}$, $C_p$, and $C_0$.
Our fitting procedure entailed first defining a broad 4D grid of belt parameters and generating the model spectra for all points in this grid. We then found the best fit to the data (the optimal $C_p$ and $C_0$) for each model using MATLAB's \texttt{lsqcurvefit} algorithm by minimizing the standard $\chi^2$ metric (calculated in linear space). The parameter set corresponding to the lowest overall $\chi^2$ was deemed the best model. We examined how the minimum $\chi^2$ varied with each disk parameter to determine if we had located a global minimum in parameter space. We then created a revised parameter grid with greater precision centered on the previous minimum and repeated the fitting procedure, iterating this process until we were confident that the overall best fit was found. We found acceptable one-belt fits to $13$ of the $22$ targets. The parameters of the best fits for these targets are given in Table \ref{table:1beltresults}, and the the fits are shown in Figure \ref{fig:onebeltfits}.
The IRS SL1 order can show spurious excess signal at 13.5 to 15 $\micron$ due to the ``SL 14 micron teardrop" effect (IRS Instrument Handbook). This artifact is thought to be caused by an internal reflection in the instrument, and is evident as a teardrop shape overlapping and slightly to the left of the spectral trace on the detector. We examined the 2D spectra of several sources and noticed signs of this effect. To avoid mistaking the teardrop signal for an emission feature, we excluded the 13.5 to 15 $\micron$ data for all of our targets when fitting models. However, in general the models fit this spectral range reasonably well (see Figures \ref{fig:onebeltfits} and \ref{fig:twobeltfits}) and including it in our $\chi^2$ minimization would not have modified the fits significantly.
\begin{figure*}
\centering
\includegraphics[scale=0.8]{f1_1.eps}
\caption{Model spectra for the $13$ targets that could be fit well by one belt. Targets marked with an asterisk have marginally detected features. IRS data are shown in black and the models are in solid green. The left panels show the total flux, while the right panels show the excess flux above the photosphere. Uncertainty in the data is shown in gray shading on the left panels, but is omitted from the right panels for clarity. Data between 13.5 and 15 $\micron$ were not included in the fitting procedure. The blackbody fits from \citet{ballering2013} are shown in dashed green (warm component) and dashed blue (cold component). Cases where the blackbody fits do not match the data well are due to the differences between this work and \citet{ballering2013} in how the stellar photosphere component was removed.}
\label{fig:onebeltfits}
\end{figure*}
\begin{figure*}
\centering
\figurenum{1}
\includegraphics[scale=0.8]{f1_2.eps}
\caption{(Continued)}
\end{figure*}
\begin{figure*}
\centering
\figurenum{1}
\includegraphics[scale=0.8]{f1_3.eps}
\caption{(Continued)}
\end{figure*}
For the targets that could not be fit by one-belt models, the difficulty arose because the structure of the features at $\sim$10 $\micron$ and the levels of continuum excess at longer wavelengths could not both be reproduced with a single belt. Furthermore, the best fitting single-belt models were generally very large in radial extent, whereas many spatially resolved images of debris disks reveal them to be comprised of multiple, narrower belts. In fact, although all of the fits have inner radii within the terrestrial zones of the stars, 10 of the 13 fits have outer radii beyond 30 AU. Realistically, all of these fits could just as well be described as two-belt fits, since it is not plausible that there is a single component that is so broad. Therefore, we next fit all of our targets with two-belt models. We again fixed $p=3.65$ and $a_\text{max}=1000$ $\micron$ for both belts, and the form of the model was
\begin{align}
\label{eq:twomodel}
F_{\nu,\text{model}}(\lambda) &= C_pB_\nu(\lambda,T_\star) \nonumber\\ &+ C_1F_{\nu,\text{belt}}(\lambda,r_\text{in1},r_\text{out1},q_1,a_\text{min1}) \nonumber\\ &+ C_2F_{\nu,\text{belt}}(\lambda,r_\text{in2},r_\text{out2},q_2,a_\text{min2}),
\end{align}
with eleven free parameters $r_\text{in1}$, $r_\text{out1}$, $q_1$, $a_\text{min1}$, $r_\text{in2}$, $r_\text{out2}$, $q_2$, $a_\text{min2}$, $C_p$, $C_1$, and $C_2$. The fitting again entailed defining a 4D grid of parameters for the inner belt and for the outer belt and generating the single-belt model spectra for all points in these grids. We found the best fit to the data (the optimal $C_p$, $C_1$, and $C_2$) for each possible pairing of inner and outer models\footnote{As a point of clarity, when using these models (specified in terms of dust location), we refer to ``inner" and ``outer" belts. This is in contrast to blackbody models (specified in terms of dust temperature), for which we refer to ``warm" and ``cold" components.} such that $r_\text{out1} < r_\text{in2}$. The parameters of these best fits are presented in Table \ref{table:2beltresults}, and the fits are plotted in Figure \ref{fig:twobeltfits}. The fits were unable to constrain $q$; we generally found that we could fit the data nearly equally well while varying this parameter from 0 to 2 (although for the one-belt fits $q$ tended to be closer to 0).
For most targets, the 10 $\micron$ feature was fit almost entirely by the flux from the inner belt model; hence, the parameters of the inner belts were more constrained by our fitting than those of the outer belts, and we only report the parameters of the inner belts in Table \ref{table:2beltresults}. Virtually all the inner belts lie entirely within the terrestrial zones.
In Tables \ref{table:1beltresults} and \ref{table:2beltresults} we give the total mass of dust (in grains from $a_\text{min}$ to 1000 $\micron$) for the belts, computed from $C_0 \times 10^{-10} M_\sun$ and $C_1 \times 10^{-10} M_\sun$. We also give the fractional luminosity of each belt, $L_\text{belt}/L_\star$, where the belt's emitting luminosity was calculated from
\begin{equation}
\label{eq:Lbelt}
L_\text{belt} = 4 \pi D^2 \int_{1\,\micron}^{2000\,\micron} \left(\frac{c}{\lambda^2}\right) F_{\nu,\text{belt}}(\lambda) \, \mathrm{d}\lambda.
\end{equation}
The uncertainties in our model fits were likely dominated by systematic errors, rather than by the statistical errors in the IRS flux density measurements. Calibration errors in the data were one source of systematic error, although we mitigated this by allowing the normalization of stellar photosphere flux density to be a free parameter in the fitting. Any errors in the stellar properties ($L_\star$, $T_\star$, D, etc.) influenced the models, adding systematic error to the best fit parameters. By using Mie theory we implicitly assumed the dust grains are spherical, but real grains are not perfect spheres, which added uncertainty to our models via our computed $Q_\text{abs}$ values. The robustness of our fits was also limited by using only one dust composition, fixing the form of $\Sigma(r)$ and $n(a)$, fixing $a_\text{max}$ and $p$, and using a maximum of two belts in our models. Varying the grain composition can result in changes in the radial scale of the belt to fit the observations, but variations by more than a factor of two are unlikely based on the range of optical constants available via the DDS website. Some silicate compositions fail to reproduce the shapes of the observed features entirely. The validity of our assumptions depends in part on the source of feature-producing dust, which we discuss in $\S$\ref{sec:discussion}. However, our best fit models generally reproduce the data well with physically reasonable parameters, so further increasing the model complexity and number of free parameters likely would have simply added degeneracies among the parameters. Furthermore, few if any of these additional free parameters could significantly undermine the detection of silicate features in the spectra.
\begin{figure*}
\centering
\includegraphics[scale=0.8]{f2_1.eps}
\caption{Two-belt model fits for each target. Targets marked with an asterisk have marginally detected features. IRS data are shown in black, the inner belt models are in solid green, the outer belt models are in solid blue, and the total models are in solid red. The left panels show the total flux, the center panels show the excess flux above the photosphere, and the right panels show the remaining excess with the outer belt models removed. Uncertainty in the data is shown in gray shading on the left panels, but is omitted from the other panels for clarity. Data between 13.5 and 15 $\micron$ were not included in the fitting procedure. The blackbody fits from \citet{ballering2013} are shown in the center panels in dashed green (warm component) and dashed blue (cold component). Cases where the blackbody fits do not match the data well are due to the differences between this work and \citet{ballering2013} in how the stellar photosphere component was removed.}
\label{fig:twobeltfits}
\end{figure*}
\begin{figure*}
\centering
\figurenum{2}
\includegraphics[scale=0.8]{f2_2.eps}
\caption{(Continued)}
\end{figure*}
\begin{figure*}
\centering
\figurenum{2}
\includegraphics[scale=0.8]{f2_3.eps}
\caption{(Continued)}
\end{figure*}
\begin{figure*}
\centering
\figurenum{2}
\includegraphics[scale=0.8]{f2_4.eps}
\caption{(Continued)}
\end{figure*}
\begin{figure*}
\centering
\figurenum{2}
\includegraphics[scale=0.8]{f2_5.eps}
\caption{(Continued)}
\end{figure*}
\begin{figure*}
\centering
\figurenum{2}
\includegraphics[scale=0.8]{f2_6.eps}
\caption{(Continued)}
\end{figure*}
After inspecting the model fits, we segregated our targets into those with clear features and those with only marginally detected features. This designation is listed in Tables \ref{table:1beltresults} and \ref{table:2beltresults}, and targets with marginal features are marked with asterisks in Figures \ref{fig:onebeltfits} and \ref{fig:twobeltfits}. Some spectra with marginal features had only a very weak excess at 10 $\micron$, although the shape of the excess resembled a silicate emission feature. Others had a strong excess at 10 $\micron$, but its shape only differed slightly from a featureless blackbody. The majority of targets presented here do have clearly detected features, and these features vary in strength and the degree to which their signal is potentially confused by flux from the continuum. This suggests a natural variation in these features, predicting that there should be some low-level features that could be only marginally detected. Thus, it is likely that at least some of marginally-detected features here are true detections, and we include them in our target list for completeness.
\section{RESULTS}
\label{section:results}
\subsection{A Window to Terrestrial Zones}
\label{sec:terrestrialzones}
Are the inner belts of these systems located in the terrestrial zones? The terrestrial zone is most easily defined in terms of an equilibrium temperature of $\sim$300 K. In Figure \ref{fig:newvsoldtdust} we plot the equilibrium temperatures of the dust in our best-fit models versus the temperatures of the warm components found by blackbody fitting from \citet{ballering2013} for these systems. We calculated the equilibrium temperatures using $T_\text{eq} = (\text{278.7 K})\left(\frac{L_\star}{L_\sun}\right)^{1/4} \left(\frac{r}{\text{1 AU}}\right)^{-1/2}$. Note that $T_\text{eq}$ is not the temperature of all dust grains in the belt at $r$, as the temperature varies significantly with grain size, and the smallest grains are significantly hotter than $T_\text{eq}$.
From the discussion in $\S$\ref{sec:modelsedfitting}, the single belt models are not plausible, either because they do not produce satisfactory fits (9 cases) or because they require belts that are so wide that they are in fact indicating the need for two belts. We computed $T_\text{eq}$ at the midpoint of the inner belt, $r = (r_\text{in1} + r_\text{out1})/2$. $T_\text{eq}$ values are listed in Table \ref{table:2beltresults}. We found that the inner belts are typically nearer to their stars than predicted by blackbody fitting. Within the errors, 18/22 of the inner belts lie within the terrestrial zones of their stars; there are two cases where the belts are too hot (HIP43121, HIP79797) and two where they are too cold (HIP61049, HIP86305).
\begin{figure}
\centering
\includegraphics[scale=0.44,angle=0]{f3.eps}
\caption{The equilibrium temperatures of our best fitting models versus the temperatures of the warm component blackbody fits to these targets from \citet{ballering2013}. The equilibrium temperatures are calculated at the midpoint of the inner belt. Targets with clear features are stars and those with marginal features are squares. Green points are those that \citet{ballering2013} fit with two blackbodies and cyan points are those that \citep{ballering2013} fit with only one blackbody. This figure illustrates that fitting models to emission features can detect exozodiacal dust in the terrestrial zones of these systems (or even hotter zones), while the blackbody fits would find only asteroid belt zone dust.}
\label{fig:newvsoldtdust}
\end{figure}
It might seem surprising that disk models using realistic grain properties would predict dust belts to be nearer their stars than derived from blackbody fitting. As discussed in the Introduction, blackbody models tend to place debris disks closer to their stars than they actually are due to the presence of small, superheated grains. However, blackbody models often miss the emission features entirely (see the dashed lines in the right panels of Figure \ref{fig:onebeltfits} and the center panels of Figure \ref{fig:twobeltfits}). The signal of an emission feature (governed by $Q_\text{abs}$) is modulated by a blackbody function at the dust temperature (see Equation \ref{eq:grainflux}). Thus, disk spectra that show features are more likely to host an underlying population of dust at a temperature such that its blackbody peaks around 10 $\micron$. The spectral shape of the emission from this population of dust differs enough from a blackbody that a blackbody fitting routine is likely to ignore, rather than attempt to reproduce, the feature.
From our two-belt fits, we see that the flux from the feature-producing inner belt falls off quickly towards longer wavelengths, requiring an additional outer belt to fit the data. Because the flux from the inner belt is concentrated around the wavelength of the feature, the outer belt must account for more of the remaining flux than a cold component typically will when fitting with blackbodies. Indeed, the best fitting outer belt models for our targets are typically radially very broad. There may actually be three components in these systems, with dust in the terrestrial, asteroid belt, and Kuiper belt zones. Our inner belt models fit to the terrestrial zone dust, while our outer belt models accounted for all of the asteroid belt and the Kuiper belt zone dust that contributed to the IRS data. Our one-belt fits lead to a similar result: dust must be located at a wide range of radial locations, with some dust at least as close to the star as the terrestrial zone. The exact number of individual belts and the radial width of each belt cannot be determined from these data. However, the models require dust in the terrestrial zones to reproduce the emission features. Thus, the presence of these features is a useful tracer of exozodiacal dust in terrestrial zones.
\subsection{Notes on Specific Targets}
\label{sec:specifictargets}
Most of the debris disks in our sample are not well studied in the literature. The presence of a warm component was reported for most of these targets, but only from IRAS, MIPS, or WISE photometry. The IRS data for many of these targets were not published prior to \citet{ballering2013}. Some targets, however, have been well-studied in the past. We discuss these here, and how our discovery of emission features in their spectra fits with what was known about these systems. We also discuss notable aspects of our model fits for certain targets.
\textit{HIP2578 (HD3003, HR136) --} An infrared excess has been known to exist around HIP2578 for some time. The excess was detected with IRAS at 25 $\micron$ \citep{oudmaijer1992}, with MIPS at 24 and 70 $\micron$ \citep{smith2006}, and with AKARI at 18 $\micron$ but not at 9 $\micron$ \citep{fujiwara2013}. \citet{smith2010} detected the excess at 18 $\micron$ with ground based observations, but could not confirm an excess at 10 or 12 $\micron$. \citet{schutz2009} also detected no excess in an 8-13 $\micron$ spectrum obtained from the ground. At $\sim$10 $\micron$, these ground based measurements had uncertainties roughly as large or larger than the excess flux level we found from the IRS data; thus, the ground based observations are consistent with our results. The IRS spectrum was previously published by \citet{zuckerman2011} and by \citet{donaldson2012}. \citet{donaldson2012} fit these data and {\it Herschel} PACS photometry for this target with a single broad belt extending from 7.8 to 120 AU and a minimum grain size of 3.5 $\micron$. This fit also required an unusually steep grain size distribution with $p$=4.4. They noted no signs of emission features in the IRS spectrum. \citet{ballering2013} fit the IRS and MIPS data with a single blackbody of temperature 194 K, while \citet{chen2014} fit these data with two blackbodies at 472 and 173 K. Our inner belt location of 1.5-3 AU is consistent with the $<6.5$ AU constraint from (unresolved) images by \citet{smith2010}, although both our fits require an outer population of cold dust as well. HIP2578 may be a binary system of two A stars separated by 0\farcs1 (4.6 AU), measured in 1925 and 1964 \citep{dommanget1994,mason2001}. \citet{eggleton2008} report that HIP2578 is part of a 6 component system consisting of a wide hierarchical triple of three close binaries. The binarity of this system could impact both the heating and orbital stability of the debris disk. A detailed consideration of these factors is beyond the scope of this paper.
\textit{HIP26966 (HD38206, HR1975) --} The mid-IR excess around HIP26966 was detected by both IRAS \citep{mannings1998} and MIPS \citep{rieke2005,su2006}. The IRS data have been published several times \citep{morales2009,morales2011,zuckerman2011,ballering2013,chen2014}, but the presence of a 10 $\micron$ emission feature has not been previously discussed. \citet{smith2010} did not detect an excess at 10 $\micron$ with ground based observations. \citet{moerchen2010} also performed ground-based observations, detecting excess at 18.3 $\micron$ but not at 10.4 $\micron$. Their results are consistent with our measurements, but their uncertainties at 10.4 $\micron$ were too large to claim a significant excess detection. The 18.3 $\micron$ images of \citet{moerchen2010} were not spatially resolved, constraining the radius of the disk to $<$10.8 AU, which is consistent with our inner belt extending to 4 AU (nearly all of the excess flux at 18.3 $\micron$ is emanating from the inner belt in our model). No circumstellar dust was detected around this target in scattered light with HST \citep{krist2010}.
\textit{HIP27288 ($\zeta$ Lep, HD38678, HR1998) --} HIP27288 has long been known to host a warm debris disk. The mid-IR excess was detected by IRAS \citep{cote1987,aumann1991,mannings1998}, ISO \citep{habing2001}, MIPS \citep{rieke2005,su2006}, and a number of ground-based instruments \citep{fajardoacosta1998,chen2001,jayawardhana2001,schutz2005}. \citet{wyatt2007} noted that this excess is unusually bright for a system of its age, suggesting that this dust may be transient. No evidence for a very hot dust component was seen from near-IR interferometric observations \citep{absil2013}. \citet{chen2006} fit the IRS excess with a single 190 K blackbody, and noted no signs of emission features. This disk was also spatially resolved by ground-based observations at 18.3 $\micron$, locating the disk at 3 AU with some emission extending out to 8 AU \citep{moerchen2007}. Our inner belt (5-6 AU) is consistent with these results. Our models do differ significantly from the literature in that we include outer dust as well in the form of an outer belt in our two-belt model, or out to $\sim$40 AU in our one-belt model. There is clearly some structure in the IRS data at $\sim$10 $\micron$, but we classify the detection of a feature in this spectrum as marginal because this structure has a less peaked shape than features seen in other targets. If there is only one belt, then this structure must be the result of an emission feature. On the other hand, the structure could be the result of overlapping emission from multiple belts. A single blackbody model, however, is not a good fit to these data.
\textit{HIP43121 (50 Cnc, HD74873, HR3481) --} The debris disk around HIP43121 has not been particularly well studied in the literature. However, the IRS data were published by \citet{morales2009} and \citet{morales2011}. Both of these studies fit the excess with a single blackbody function at 190 K, but a footnote in \citet{morales2009} remarked that there were hints of a 10 $\micron$ feature in the data.
\textit{HIP58220 (HD103703) --} While fitting two-belt models to this target, we discovered a degeneracy in terms of which belt produced the majority of the 10 $\micron$ feature. Both local minima in parameter space produced fits with nearly identical $\chi^2$. The model we present in Figure \ref{fig:twobeltfits} relied on the inner belt to reproduce the feature while the outer belt resembled a nearly-smooth continuum, as was the case for most of our targets. In the alternative model, a broad outer belt with a small minimum grain size contributed significantly to the 10 $\micron$ feature while the inner belt with only larger grains provided primarily continuum. The parameters for this second case were: $a_\text{min1}$ = 6 $\micron$, $r_\text{in1}$ = 0.8 AU, $r_\text{out1}$ = 1 AU, $M_\text{dust1} = 0.218 \times 10^{-5} M_\earth$, $L_\text{belt1}/L_\star = 18.7 \times 10^{-5}$, $a_\text{min2}$ = 1 $\micron$, $r_\text{in2}$ = 1 AU, $r_\text{out2}$ = 100 AU, $M_\text{dust2}=73.7 \times 10^{-5} M_\earth$, and $L_\text{belt2}/L_\star=15 \times 10^{-5}$.
\textit{HIP61049 (HD108857) --} Like HIP58220, this target also exhibited a degeneracy in terms of which belt produced the emission feature. In Figure \ref{fig:twobeltfits} we show the fit where the inner belt could reproduce the entirety of the data with no need for an outer belt. In the other fit, the outer belt contributed significantly to the spectrum, including to the emission feature, due to having a smaller minimum grain size than the inner belt. The parameters of this fit were: $a_\text{min1}$ = 5 $\micron$, $r_\text{in1}$ = 0.7 AU, $r_\text{out1}$ = 0.8 AU, $M_\text{dust1}=0.17 \times 10^{-5} M_\earth$, $L_\text{belt1}/L_\star = 23.3 \times 10^{-5}, a_\text{min2}$ = 1.7 $\micron$, $r_\text{in2}$ = 4 AU, $r_\text{out2}$ = 7 AU, $M_\text{dust2} = 4.98 \times 10^{-5}$, and $L_\text{belt2}/L_\star = 26.1 \times 10^{-5}$.
\textit{HIP79797 (HD145689, HR6037) --} \citet{zuckerman2011} fit the MIPS and IRS excess with a 220 K blackbody, but did not mention any emission feature. The noteworthy aspect of this system is that the primary star is orbited at a projected separation of 350 AU by a binary system of brown dwarfs, separated from each other by 3 AU \citep{huelamo2010,nielsen2013}.
\begin{figure}
\centering
\includegraphics[scale=0.44,angle=0]{f4.eps}
\caption{This figure shows that $a_\text{min}$ is consistent with $a_\text{BOS}$ for our targets. Stars are targets with clear features and squares are those with marginal features. The bounding trend lines show the effect of varying $a_\text{BOS}$ by a factor of two (e.g., due to more complex grain structure). Low-level silicate emission features can arise in debris disk spectra, even when grains are not below the blowout size.}
\label{fig:aminvsabos}
\end{figure}
\textit{HIP86305 ($\pi$ Ara, HD159492, HR6549) --} Excess emission was first detected around HIP86305 with IRAS \citep{cheng1992,mannings1998}. No circumstellar dust was seen in scattered light with HST \citep{doering2007}. \citet{morales2009} and \citet{morales2011} fit blackbodies to the IRS excess and found a warm component temperature of 160 K, well outside of the terrestrial zone, although \citet{morales2009} noted that there may be signs of a faint spectral feature in the data. \citet{morales2013} resolved the outer edge of the outer belt with \textit{Herschel} (116 AU) and reanalyzed the IRS data, incorporating the constraint from \textit{Herschel}, and using physically-motivated belt models that could reproduce the structure in the IRS spectrum. They found an inner belt location of 9.1 or 9.8 AU, depending on the grain composition, which falls between the inner and outer edges of our best fit inner belt model. The analysis of \citet{morales2013} required grains three or four times smaller than the blowout size (requiring $a_\text{min}$ $\sim$1 $\micron$), whereas we found $a_\text{min1}$ consistent with the blowout size for this system (we found a larger minimum grain size and a slightly lower blowout size). \citet{morales2013} used a lower value for the stellar luminosity ($\sim$10 $L_\sun$) and different grain compositions, which may have contributed to the discrepancy with our results. Furthermore, neither we nor \citet{morales2013} could reproduce all of the structure seen in the IRS data, which may arise from crystalline grains.
\begin{figure}
\centering
\includegraphics[scale=0.44]{f5.eps}
\caption{The stellar type and age for targets with clear features (green stars), marginal features (green squares), and no features (gray circles). The properties of the targets with no features are from \citet{ballering2013}. This illustrates that features can be used to probe the terrestrial regions of young planetary systems over a range of stellar types, and also for older planetary systems of early type stars. Note that there are actually three 30 Myr-old A0 targets with clear features and two 20 Myr-old A0 targets with marginal features in our sample.}
\label{fig:typevsage}
\end{figure}
\section{DISCUSSION}
\label{sec:discussion}
Our analysis shows that a significant number of debris disks with excesses in the mid-IR exhibit low-level silicate emission features, indicating the presence of exozodiacal dust in their terrestrial zones. In this section we compare properties of these systems with others that host featureless warm debris disks, and we discuss potential sources for this exozodiacal dust.
Dust produced by a steady state collisional cascade of planetesimals is expected to have a minimum grain size set by the blowout size of the system. Dust smaller than this size will be removed from the system by the star's radiation pressure. The presence of grains smaller than the blowout size could indicate that a large amount of dust was produced recently, such as in a massive collision. If the smallest grains were found to be significantly larger than the blowout size, then there might be other forces acting to remove small grains such as interactions with the ISM. We calculated the blowout size for each of our targets using Equation 5 of \citet{donaldson2012}:
\begin{align}
\label{eq:abos}
a_\text{BOS} = &(1.15 ~\micron)\left(\frac{L_\star}{L_\sun}\right) \left(\frac{M_\star}{M_\sun}\right)^{-1} \nonumber\\ &\times \left(\frac{\rho}{\text{1 g cm}^{-3}}\right)^{-1},
\end{align}
where $\rho$ is the density of the grain material (we assumed $\rho=3.71 \, \text{g cm}^{-3}$). The results for our targets are given in Tables \ref{table:1beltresults} and \ref{table:2beltresults}. In Figure \ref{fig:aminvsabos} we plot the minimum grain size of the inner belt of the two-belt models versus the blowout size for the targets. We see that the minimum grain size does track the blowout size, suggesting that these systems are not being influenced by rare or extreme circumstances. Equation \ref{eq:abos} assumes grains are solid spherical particles. Real grains may have more complex structures, so $a_\text{BOS}$ is likely only accurate to within a factor of $\sim$2.
We next investigated the age and stellar type of targets that exhibited features, as shown in Figure \ref{fig:typevsage}. The gray points are systems with warm disks from the sample of \citet{ballering2013} that did not show any features. For targets with no age determination in \citet{ballering2013}, we found age values from \citet{nielsen2013}, \citet{zorec2012}, and \citet{chen2014}, as we did for the targets with features (see $\S$\ref{sec:targetselection}). In Figures \ref{fig:fracfeaturesvstype} and \ref{fig:fracfeaturesvsage} we show the fraction of warm debris disks that exhibit features, binned by stellar type and by age.
We find that many of the disks with features are young ($\sim$10-30 Myr), but that there is a significant number of older disks (hundreds of Myr) that also have features. Young disks with features have stellar types spanning the range of the parent sample (late B through F), while the only older disks that show features are the early to mid A types. While the presence of features does not appear uniform across all stellar types and ages, it is clear that analyzing features in debris disks can provide a means to study the terrestrial zones of planetary systems with a large range of stellar types and ages. Disks with clear features and with marginal features are distributed in approximately the same way by stellar type and age, lending further evidence to the notion that there is natural variation in feature strengths for systems with dust in their terrestrial zones.
\begin{figure}
\centering
\includegraphics[scale=0.44]{f6.eps}
\caption{The fraction of warm debris disks showing features in bins of stellar type. The fraction with clear features is in green, the fraction with marginal features is in magenta, and the total (the sum of green and magenta) is in black. Note that the fractions with clear and with marginal features are consistent with each other in their variation with stellar type.}
\label{fig:fracfeaturesvstype}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.44]{f7.eps}
\caption{The fraction of warm debris disks showing features in bins of system age. The fraction with clear features is in green, the fraction with marginal features is in magenta, and the total (the sum of green and magenta) is in black. Features are found around both young and older systems. Note that the fractions with clear and with marginal features are consistent with each other in their variation with age. The sharp spike at 25-60 Myr and minimum at 60-130 Myr may be a result of small number statistics, as there are few total warm disks in this age range.}
\label{fig:fracfeaturesvsage}
\end{figure}
Are features preferentially found in bright debris disks? In Figure \ref{fig:fraclumvsage} we plot the fractional luminosity of the mid-IR excess of these systems versus age. The gray points are again warm disks from \citet{ballering2013} that do not show signs of features. The fractional luminosity values in this plot are not the same values as given in Tables \ref{table:1beltresults} and \ref{table:2beltresults}. As discussed in $\S$\ref{sec:terrestrialzones} and illustrated in Figures \ref{fig:onebeltfits} and \ref{fig:twobeltfits}, the models we used in this paper fit the IRS data differently than did the one or two blackbody functions used by \citet{ballering2013}. We did not have acceptable one-belt fits for all targets in this paper, nor is it appropriate to compare the brightness of just our inner (outer) belt with the warm (cold) component from blackbody fitting, since the two fits in the two papers are constrained by different wavelength ranges. Comparing the total fractional luminosities of all components would also not be a valid comparison, as \citet{ballering2013} included MIPS 70 $\micron$ data in the fitting and so were sensitive to a significant amount of cold dust that was not measured in this study. To properly compare these two samples, we calculated the luminosity of the total model (inner + outer belts using our two-belt models; one or two blackbodies) using Equation \ref{eq:Lbelt}, but only over the wavelength range from 1 to 30 $\micron$.
Figure \ref{fig:fraclumvsage} shows that targets with features (either clear or marginal) are not extraordinary in terms of the brightness of their mid-IR excess. This suggests that the detection of features is not a selection effect limited to very bright disks. Some planetary systems have a detectable population of dust in their terrestrial zones while others have significantly less dust in this region, but otherwise the systems with features and without features are quite similar.
\begin{figure}
\centering
\includegraphics[scale=0.44]{f8.eps}
\caption{The fractional luminosity (of the infrared excess from 1 to 30 $\micron$) and age of targets with clear features (green stars), marginal features (green squares), and no features (gray circles). The properties of the targets with no features are from \citet{ballering2013}. The overall decrease in disk brightness is well-known from previous debris disk studies. We see that the fractional luminosities of disks with features (clear or marginal) are consistent with those of featureless disks.}
\label{fig:fraclumvsage}
\end{figure}
What is the source of exozodiacal dust? There could be a belt of parent body planetesimals in the terrestrial zone undergoing a collisional cascade and producing the dust. This model can explain most standard debris disks with dust in the asteroid belt or Kuiper belt zones, although debris disks on smaller orbits are expected to grind down and dissipate on much shorter timescales. \citet{kennedy2013} investigated the possibility of in situ dust production to explain the detection of excesses at 12 $\micron$ with WISE photometry, which they interpreted as emission from exozodiacal dust. As in this study, they found that most systems with exozodiacal dust were young, but that some older systems had 12 $\micron$ excess as well. They could reproduce this finding by assuming that all systems start with a population of parent bodies in this zone that steadily collide and decay with time (this explains the young systems), but additionally there are occasional random collisions between the remaining parent bodies, as required to explain the exozodiacal dust in the older systems.
Another possibility is that the dust was produced farther out and migrated inwards via Poynting-Robertson (PR) drag or other drag forces. Theoretical studies find that the interaction of PR drag and dust sublimation can create a population of dust extending inwards from the parent body belt where the dust is created to the sublimation radius \citep{kobayashi2008,vanlieshout2014}.
Exozodiacal debris could also be delivered to terrestrial zones by planetesimals scattered inwards from an outer belt. \citet{bonsor2012} modelled this scenario and found that a chain of closely packed planets is required to move material inwards effectively. Furthermore, for older systems the outer belt must be located at a large orbital radius such that it can be massive enough to deliver sufficient material inwards without quickly grinding itself down. The inward scattering of planetesimals can occur at a higher rate and be sustained for a longer time period if the outermost planet is actively migrating outwards into the planetesimal belt \citep{bonsor2014}. The efficiency of this mechanism depends sensitively on the number, locations, and masses of the planets in the chain, and on the properties of the planetesimal belt. If this mechanism is the dominant source of the exozodiacal dust, the detection of silicate emission features also implies that these targets host rich planetary systems.
A dynamical instability in a planetary system can also deliver planetesimals to the terrestrial region. In this case, planets scatter each other, destabilizing their orbits and scattering many planetesimals. \citet{bonsor2013} simulated such events and found that their signals are short-lived (0.7-2.8 Myr). Detecting a significant number of systems that recently underwent a dynamical instability is unlikely, and this mechanism cannot be the dominant source of exozodiacal dust.
Exozodiacal dust may be the byproduct of collisions associated with the final, chaotic stage of terrestrial planet formation. \citet{jackson2012} modeled the dust production from such a collision (analogous to our Moon-forming impact) and found that, after an initial spike in infrared excess from vapor condensates, a low but detectable amount of debris can persist in the terrestrial regions for at least 1 Myr and often longer than 10 Myr. Each terrestrial planet undergoes multiple giant impacts during the chaotic phase of its formation. Terrestrial planet formation is expected to have finished once the system reaches an age of 100 Myr, so this explanation is plausible for the younger targets in our sample, but not for the older systems.
Our targets potentially represent a different source of exozodiacal dust than the debris disks with previously studied features listed in $\S$\ref{sec:introduction}. The excesses in those systems tend to be anomalously bright, and the emission features are often very strong. Those systems may have very recently experienced massive collisions such that transient populations of dust (e.g. vapor condensates) are still present. These massive collisions may be part of the terrestrial planet formation process (for the younger systems) or from periods of dynamical instability.
\section{CONCLUSIONS}
\label{sec:conclusions}
In summary, we found low-level silicate emission features in the IRS spectra of $22$ warm debris disks that were previously studied by \citet{ballering2013}. $13$ of these had clearly detected features, while $9$ were only marginally detected. We fit these data with physically-motivated models, allowing us to constrain the radial locations and minimum grain sizes of the dust belts more precisely than was possible when fitting with blackbodies. Our fits place dust in the terrestrial zones of these targets, which was missed when these data were fit with only blackbody functions. An outer population of dust was also required to fit the data. The minimum grain sizes of the terrestrial zone dust were consistent with the blowout sizes of these systems, and the mid-IR fractional luminosities of debris disks with features were comparable to those of warm debris disks without features, implying that disks with features are ``normal". The properties of systems with marginally detected features were distributed almost identically as those of systems with clear features, suggesting that many of the marginal cases are likely true emission features.
We found systems with features at a range of stellar types and ages (although no features were found around older, later type stars). The analysis of emission features in the spectra of unresolved debris disks provides a powerful method to probe the terrestrial zones of planetary systems at various stages of their evolution. These results will complement mid-IR interferometric studies of exozodiacal dust, allowing for the robust characterization of regions of planetary systems that have, until now, remained largely out of reach.
\acknowledgments
We would like to thank Kate Su for many helpful comments and suggestions. We also thank the referee for providing valuable feedback. We made use of the NASA/IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with NASA. This work is based on observations made with the {\it Spitzer} Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA.
\bibliographystyle{apj}
\input{DebrisDisks.bbl}
\clearpage
\begin{landscape}
\begin{deluxetable}{ccccccccccccccc}
\setlength{\tabcolsep}{0.001in}
\tabletypesize{\scriptsize}
\tablewidth{0pt}
\tablecolumns{15}
\tablecaption{Target Properties \label{table:targetlist}}
\tablehead{\colhead{HIP} & \colhead{Other} & \colhead{Spectral} & \colhead{D} & \colhead{Age} & \colhead{Age\tablenotemark{a}} & \colhead{Age} & \colhead{$T_\star$} & \colhead{$L_\star$} & \colhead{$R_\star$} & \colhead{$M_\star$} & \colhead{V} & \colhead{K} & \colhead{$F_{\nu}$(24 $\micron$)} & \colhead{IRS} \\ \colhead{Identifier} & \colhead{Identifiers} & \colhead{Type} & \colhead{(pc)} & \colhead{(Myr)} & \colhead{Quality} & \colhead{References} & \colhead{(K)} & \colhead{($L_\sun$)} & \colhead{($R_\sun$)} & \colhead{$M_\sun$} & \colhead{(mag)} & \colhead{(mag)} & \colhead{(mJy)} & \colhead{AOR}}
\startdata
HIP2578 & HD3003, HR136 & A0V & 45.6 & 30 & 2 & 4,7 & 9800 & 22.09 & 1.64 & 2.90 & 5.07 & 4.99 & 232 $\pm$ 2.31 & 21789184 \\
HIP18437 & HD24966 & A0V & 105.8 & 10 & 1 & 2 & 9800 & 21.01 & 1.59 & 2.90 & 6.89 & 6.86 & 68.43 $\pm$ 0.67 & 21792256 \\
HIP26395 & HD37306, HR1919 & A2V & 62.9 & 10 & 2 & 4 & 9020 & 13.42 & 1.50 & 2.43 & 6.09 & 5.97 & 65.46 $\pm$ 0.64 & 21794560 \\
HIP26966 & HD38206, HR1975 & A0V & 75.1 & 30 & 1 & 2,3 & 9800 & 28.22 & 1.85 & 2.90 & 5.73 & 5.78 & 112 $\pm$ 1.1 & 12713472 \\
HIP27288 & $\zeta$ Lep, HD38678, HR1998 & A2IV-Vn: & 21.6 & 230 & 1 & 1 & 9020 & 19.04 & 1.79 & 2.43 & 3.54 & 3.29 & 878 $\pm$ 8.78 & 4932864 \\
HIP41081 & HD71043, HR3300 & A0V & 70 & 30 & 2 & 3 & 9800 & 22.86 & 1.66 & 2.90 & 5.89 & 5.87 & 59.79 $\pm$ 0.63 & 16177408 \\
HIP43121 & 50 Cnc, HD74873, HR3481 & A1V & 54 & 147 & 1 & 8 & 9400 & 16.53 & 1.54 & 2.64 & 5.89 & 5.55 & 66.88 $\pm$ 0.68 & 14140928 \\
HIP57971 & HD103266, HR4553 & A2V & 74.1 & 760 & 1 & 10 & 9020 & 18.53 & 1.77 & 2.43 & 6.17 & 5.99 & 41.49 $\pm$ 0.44 & 21800704 \\
HIP58220 & HD103703 & F3V & 98.9 & 20 & 2 & 5 & 6919 & 4.27 & 1.44 & 1.48 & 8.5 & 7.39 & 25.87 $\pm$ 0.3 & 16170496 \\
HIP58528 & HD104231 & F5V & 110.5 & 20 & 2 & 5 & 6650 & 4.27 & 1.56 & 1.40 & 8.54 & 7.42 & 18.6 $\pm$ 0.33 & 21800960 \\
HIP59394 & 3 Crv, HD105850, HR4635 & A1V & 58.8 & 369 & 1 & 9 & 9400 & 23.61 & 1.84 & 2.64 & 5.47 & 5.32 & 80.24 $\pm$ 0.83 & 21801216 \\
HIP60561 & HD107947 & A0V & 91.1 & 20 & 2 & 5 & 9800 & 19.34 & 1.53 & 2.90 & 6.6 & 6.62 & 43.46 $\pm$ 0.44 & 22800896 \\
HIP61049 & HD108857 & F7V & 97 & 20 & 2 & 5 & 6388 & 3.86 & 1.61 & 1.26 & 8.6 & 7.07 & 40.07 $\pm$ 0.44 & 22802432 \\
HIP61558 & f Vir, HD109704, HR4799 & A3V & 69 & 412 & 1 & 9 & 8710 & 18.81 & 1.91 & 2.26 & 5.88 & 5.7 & 62.69 $\pm$ 0.65 & 21801984 \\
HIP63439 & HD112810 & F4IV/V & 143.3 & 10 & 2 & 5 & 6784 & 4.49 & 1.54 & 1.45 & 9.14 & 8.04 & 10.33 $\pm$ 0.17 & 21802752 \\
HIP65965 & HD117484 & B9V & 147.3 & 20 & 2 & 5 & 10500 & 30.78 & 1.68 & 3.29 & 7.53 & 7.52 & 39.31 $\pm$ 0.41 & 22803200 \\
HIP66068 & HD117665 & A1/A2V & 147.9 & 20 & 2 & 5 & 9400 & 29.57 & 2.06 & 2.64 & 7.21 & 7.08 & 43.71 $\pm$ 0.46 & 22801920 \\
HIP71271 & HD127750 & A0V & 175.7 & 20 & 2 & 6 & 9800 & 30.98 & 1.94 & 2.90 & 7.6 & 7.54 & 19.51 $\pm$ 0.2 & 26312704 \\
HIP78641 & HD143675 & A5IV/V & 113.4 & 20 & 2 & 5 & 8190 & 7.38 & 1.35 & 2.00 & 8.04 & 7.62 & 51.64 $\pm$ 0.51 & 22806528 \\
HIP79797 & HD145689, HR6037 & A4V & 52.2 & 203 & 1 & 8 & 8434 & 10.30 & 1.51 & 2.12 & 5.95 & 5.66 & 52.24 $\pm$ 0.53 & 21809152 \\
HIP86305 & $\pi$ Ara, HD159492, HR6549 & A5IV-V & 44.6 & 562 & 1 & 10 & 8190 & 15.68 & 1.97 & 2.00 & 5.25 & 4.78 & 198.3 $\pm$ 1.97 & 14143232 \\
HIP99742 & $\rho$ Aql, HD192425, HR7724 & A2V & 46 & 433 & 1 & 9 & 9020 & 21.91 & 1.92 & 2.43 & 4.95 & 4.77 & 132.6 $\pm$ 1.32 & 14143744 \\
\enddata
\tablenotetext{a}{1 means there was a single age determination, 2 means there were two independent and consistent age determinations.}
\tablerefs{
(1) \citet{vican2012} -- isochrone ages;
(2) \citet{rhee2007};
(3) \citet{su2006};
(4) \citet{tetzlaff2010};
(5) \citet{rizzuto2011};
(6) \citet{hoogerwerf2000};
(7) \citet{zuckerman2004};
(8) \citet{nielsen2013};
(9) \citet{zorec2012};
(10) \citet{chen2014}.
}
\end{deluxetable}
\clearpage
\end{landscape}
\clearpage
\begin{deluxetable}{ccccccccc}
\tabletypesize{\scriptsize}
\tablewidth{0pt}
\tablecolumns{9}
\tablecaption{One Belt Fitting Results \label{table:1beltresults}}
\tablehead{\colhead{HIP} & \colhead{Other} & \colhead{$a_\text{BOS}$} & \colhead{$a_\text{min}$} & \colhead{$r_\text{in}$} & \colhead{$r_\text{out}$} & \colhead{$M_\text{dust}$} & \colhead{$L_\text{belt}/L_\star$} & \colhead{Feature} \\ \colhead{Identifier} & \colhead{Identifiers} & \colhead{($\micron$)} & \colhead{($\micron$)} & \colhead{(AU)} & \colhead{(AU)} & \colhead{($\times 10^{-5} M_\earth$)} & \colhead{($\times 10^{-5}$)} & \colhead{Detection}}
\startdata
HIP2578 & HD3003, HR136 & 2.4 & 2.2 & 0.6 & 100 & 83.5 & 10.5 & clear \\
HIP27288 & $\zeta$ Lep, HD38678, HR1998 & 2.4 & 3 & 1.5 & 40 & 19 & 7.94 & marginal \\
HIP43121 & 50 Cnc, HD74873, HR3481 & 1.9 & 2.2 & 0.3 & 180 & 72.9 & 3.58 & clear \\
HIP57971 & HD103266, HR4553 & 2.4 & 2.6 & 0.2 & 80 & 16.7 & 3.52 & marginal \\
HIP58220 & HD103703 & 0.9 & 2.7 & 0.2 & 10 & 3.78 & 33.9 & clear \\
HIP58528 & HD104231 & 0.9 & 1.8 & 1.1 & 15 & 4.16 & 14.4 & clear \\
HIP59394 & 3 Crv, HD105850, HR4635 & 2.8 & 2.6 & 0.2 & 200 & 102 & 3.93 & clear \\
HIP61049 & HD108857 & 1.0 & 2 & 0.8 & 10 & 5.99 & 42.4 & clear \\
HIP66068 & HD117665 & 3.5 & 2.8 & 1.4 & 40 & 36.5 & 16.1 & clear \\
HIP78641 & HD143675 & 1.1 & 1.8 & 1.6 & 200 & 1420 & 50.2 & clear \\
HIP79797 & HD145689, HR6037 & 1.5 & 1.2 & 0.4 & 80 & 7.49 & 2.38 & clear \\
HIP86305 & $\pi$ Ara, HD159492, HR6549 & 2.4 & 2 & 3.4 & 80 & 54.4 & 7.3 & clear \\
HIP99742 & $\rho$ Aql, HD192425, HR7724 & 2.8 & 3.2 & 0.3 & 100 & 34.8 & 3.97 & marginal \\
\enddata
\tablecomments{$a_\text{BOS}$ is the blowout size for grains in the system, calculated from Equation \ref{eq:abos}. $a_\text{min}$ is the minimum grain size of our best-fit model. $r_\text{in}$ and $r_\text{out}$ are the inner and outer orbital radii of our best-fit model, respectively. $M_\text{dust}$ is the total mass of dust (in grains from $a_\text{min}$ to 1000 $\micron$) of our best-fit model. $L_\text{belt}/L_\star$ is the fractional luminosity of our best fit model, with $L_\text{belt}$ calculated from Equation \ref{eq:Lbelt}. The final column notes whether the detection of features was clear or marginal, as discussed in $\S$\ref{sec:modelsedfitting}.}
\end{deluxetable}
\begin{deluxetable}{cccccccccc}
\tabletypesize{\scriptsize}
\tablewidth{0pt}
\tablecolumns{10}
\tablecaption{Two Belt Fitting Results: Inner Belt Properties \label{table:2beltresults}}
\tablehead{\colhead{HIP} & \colhead{Other} & \colhead{$a_\text{BOS}$} & \colhead{$a_\text{min1}$} & \colhead{$r_\text{in1}$} & \colhead{$r_\text{out1}$} & \colhead{$T_\text{eq}$} & \colhead{$M_\text{dust1}$} & \colhead{$L_\text{belt1}/L_\star$} & \colhead{Feature} \\ \colhead{Identifier} & \colhead{Identifiers} & \colhead{($\micron$)} & \colhead{($\micron$)} & \colhead{(AU)} & \colhead{(AU)} & \colhead{(K)} & \colhead{($\times 10^{-5} M_\earth$)} & \colhead{($\times 10^{-5}$)} & \colhead{Detection}}
\startdata
HIP2578 & HD3003, HR136 & 2.4 & 3 & 1.5 & 3 & 402 & 0.277 & 6.12 & clear \\
HIP18437 & HD24966 & 2.2 & 2 & 2 & 3 & 376 & 0.285 & 5.98 & marginal \\
HIP26395 & HD37306, HR1919 & 1.7 & 1.5 & 1.6 & 3 & 351 & 0.0812 & 2.91 & clear \\
HIP26966 & HD38206, HR1975 & 3.0 & 2.6 & 2.3 & 4 & 361 & 0.653 & 7.34 & clear \\
HIP27288 & $\zeta$ Lep, HD38678, HR1998 & 2.4 & 3 & 5 & 6 & 248 & 1.58 & 5.36 & marginal \\
HIP41081 & HD71043, HR3300 & 2.4 & 2 & 2.5 & 4 & 337 & 0.252 & 3.26 & clear \\
HIP43121 & 50 Cnc, HD74873, HR3481 & 1.9 & 1.7 & 0.5 & 1 & 647 & 0.00538 & 1.72 & clear \\
HIP57971 & HD103266, HR4553 & 2.4 & 3 & 2 & 3 & 365 & 0.104 & 1.81 & marginal \\
HIP58220 & HD103703 & 0.9 & 2 & 0.4 & 3 & 306 & 0.341 & 21.6 & clear \\
HIP58528 & HD104231 & 0.9 & 2 & 1.5 & 2 & 302 & 0.133 & 6.25 & clear \\
HIP59394 & 3 Crv, HD105850, HR4635 & 2.8 & 2 & 1.5 & 3 & 409 & 0.0638 & 1.72 & clear \\
HIP60561 & HD107947 & 2.1 & 3 & 2.5 & 4 & 323 & 0.565 & 5.59 & marginal \\
HIP61049 & HD108857 & 1.0 & 2 & 0.7 & 10 & 168 & 5.83 & 43.4 & clear \\
HIP61558 & f Vir, HD109704, HR4799 & 2.6 & 3 & 1.5 & 2 & 438 & 0.0772 & 2.6 & marginal \\
HIP63439 & HD112810 & 1.0 & 1.7 & 1.5 & 1.8 & 315 & 0.0914 & 5.2 & marginal \\
HIP65965 & HD117484 & 2.9 & 2.5 & 4.5 & 7 & 273 & 1.86 & 6.22 & marginal \\
HIP66068 & HD117665 & 3.5 & 2.5 & 4.5 & 5 & 297 & 2.06 & 10.3 & clear \\
HIP71271 & HD127750 & 3.3 & 2 & 3 & 5 & 328 & 0.23 & 2.07 & marginal \\
HIP78641 & HD143675 & 1.1 & 1 & 3 & 4 & 245 & 0.632 & 11.5 & clear \\
HIP79797 & HD145689, HR6037 & 1.5 & 0.6 & 0.6 & 1 & 557 & 0.00323 & 1.73 & clear \\
HIP86305 & $\pi$ Ara, HD159492, HR6549 & 2.4 & 2 & 4 & 50 & 106 & 24.3 & 6.71 & clear \\
HIP99742 & $\rho$ Aql, HD192425, HR7724 & 2.8 & 2.5 & 1 & 4 & 380 & 0.0779 & 1.65 & marginal \\
\enddata
\tablecomments{$a_\text{BOS}$ is the blowout size for grains in the system, calculated from Equation \ref{eq:abos}. $a_\text{min1}$ is the minimum grain size of our best-fit model's inner belt. $r_\text{in1}$ and $r_\text{out1}$ are the inner and outer orbital radii of our best-fit model's inner belt, respectively. $T_\text{eq}$ is the equilibrium temperature at the midpoint of our best-fit model's inner belt. $M_\text{dust1}$ is the total mass of dust (in grains from $a_\text{min}$ to 1000 $\micron$) of our best-fit model's inner belt. $L_\text{belt1}/L_\star$ is the fractional luminosity of our best fit model's inner belt, with $L_\text{belt}$ calculated from Equation \ref{eq:Lbelt}. The final column notes whether the detection of features was clear or marginal, as discussed in $\S$\ref{sec:modelsedfitting}.}
\end{deluxetable}
\end{document} |
\section{\label{s1}Optimization}
Even though Eq. (3) can be applied for any proof masses, including those
chosen in \cite{c2,c3}, we present here results of former calculations
performed for specific case shown in Fig \ \ref{f1}
\begin{figure}[!t]
\includegraphics[width=11cm]{f1.eps}
\caption{The proof mass as a whole is
parallelepiped $2L_{h}\times 2L_{h}\times 2L_{z}$ with narrow $2L_{n}\times
2L_{n}\times 2L_{z}$ hole for Raman fields and atom trajectories. Atoms are
launched vertically from the points $z_{1}$ and $z_{2}$ with velocities
v_{1z}$ and $v_{2z}.$ Proof mass consists from 2 halves. (a) Top view.
Joined halves. (b) Top view. Halves separated on the distance $2L_{d}$ along
$x$ access. (c) Side view, cross-section $x=0.$}
\label{f1}
\end{figure}
Scale of the parameters chosen for calculations are pieced together in the
Table \ref{t1}. The chosen value of density corresponds to pure Tungsten
\cite{c7}.
\begin{table}[h]
\caption{Order of magnitude of the atom interferometer and proof mass
parameters
\begin{tabular}{ll}
Atom & $^{133}Cs$ \\
Effective wave vector & $\vec{k}=\left\{ 0,0,k\right\} ,~k=1.47\ast 10^{7}$m
^{-1}$ \\
Time between launch and first Raman pulse & $t_{1}\sim 40$ms \\
Time between Raman pulses & $T\sim 250$ms \\
Relative acuracy of atom interferometer phase measurement & $err=10^{-4}$ \\
Earth gravity field & $\vec{g}=\left\{ 0,0,-9.8\text{m/s}^{\text{2}}\right\}
$ \\
Proof mass & $W=100$kg \\
Proof mass density & $19250$kg/m$^{3}$ \\
The hole size & $L_{n}=0.02$
\end{tabular
\label{t1}
\end{table}
For the given proof mass difference between phases of the interferometers
\left\{ z_{1},v_{1z}\right\} $ and $\left\{ z_{2},v_{2z}\right\} $ is
maximal when $\left\{ z_{1},v_{1z}\right\} $ is an absolute maximum of the
phase and $\left\{ z_{2},v_{2z}\right\} $ is an absolute minimum of the
phase. To find out these extrema we used an iterative process, which was
continued until the new value of the extremum differs relatively from the
previous value less than measurement accuracy
\end{subequations}
\begin{equation}
err=10^{-4}. \label{4p}
\end{equation
Our choice of the proof mass shape is convenient because for the gravity
potential of the parallelepiped having homogeneous density $\rho $ and sizes
$2a_{x}\times 2a_{y}\times 2a_{z}$ one has analytic expression \cite{c9,c6}
\begin{eqnarray}
\Phi &=&-G\rho
\sum_{j_{x}=-1}^{1}\sum_{j_{y}=-1}^{1}\sum_{j_{z}=-1}^{1}j_{x}j_{y}j_{z}
\left( x+j_{x}a_{x},y+j_{y}a_{y},z+j_{z}a_{z}\right) , \notag \\
f\left( u,v,w\right) &=&-\dfrac{1}{2}w^{2}\arctan \left( \dfrac{uv}{wr
\right) -\dfrac{1}{2}v^{2}\arctan \left( \dfrac{uw}{vr}\right) -\dfrac{1}{2
u^{2}\arctan \left( \dfrac{vw}{ur}\right) \notag \\
&&+vw\ln \left( u+r\right) +uw\ln \left( v+r\right) +uv\ln \left( w+r\right)
, \notag \\
r &=&\sqrt{u^{2}+v^{2}+w^{2}}. \label{5p}
\end{eqnarray
We performed calculations for Newtonian gravitational constant
G=6.67428\ast 10^{-11}$m$^{3}$s$^{-2}$kg$^{-1}$. The proof mass shown in
Fig. \ref{f1} consists of the parallelepipeds 1,2 and 3 for one half and 4,5
and 6 for another half.
Dependences of the maximal phase difference, position and velocity of
maximum and minimum on the half-size of the proof mass $2L_{z}$ are shown in
the Figs \ref{f2}-\ref{f4}.
\begin{figure}[!t]
\includegraphics[width=11cm]{f2.eps}
\caption{Dependence of the maximum of
phase difference on parallelepiped half-size}
\label{f2}
\end{figure}
\begin{figure}[!t]
\includegraphics[width=11cm]{f3.eps}
\caption{Optimal position of the 1st and
2nd atom cloud}
\label{f3}
\end{figure}
\begin{figure}[!t]
\includegraphics[width=11cm]{f4.eps}
\caption{Optimal launching velocities of
the 1st and 2nd atom clouds}
\label{f4}
\end{figure}
From the Fig. \ref{f2} one sees that the optimal phase difference has its
own maximum. The value of this maximum and values of parameters we recommend
to choose to observe it are presented in Table \ref{t2}.
\begin{table}[h]
\caption{Optimal proof mass sizes and atom clouds positions and velocities
to maximize the phase difference
\begin{tabular}{ll}
phase difference & $\Delta \phi =0.55271113$ rad \\
vertical half-size & $L_{z}=0.09$m \\
horisontal half-size & $L_{h}=0.08726$m \\
1st cloud position & $z_{1}=-0.4904$m \\
1st cloud launching velocity & $v_{1z}=2.849$m/s \\
2nd cloud position & $z_{2}=-0.2823$m \\
2nd cloud launching velocity & $v_{2z}=2.846$m/
\end{tabular
\label{t2}
\end{table}
\section{\label{s2}Error model}
To achieve high precision of the interferometers' phase measurements one has
to prepare with great accuracy both the atomic and proof mass system. In
this section we determine requirements for preparation to achieve phase
measurements with accuracy (\ref{4p}).
The most challenging here is precise positioning of the atom clouds \cite{c3
. The preferable here are, evidently, extrema of the clouds position. That
is why found above extrema in $\left\{ z,v\right\} $ space allow one not
only maximize the response, but also make less severe requirements for atom
clouds position, velocity, temperature and size because the response becomes
quadratic on variations of positions and velocities nea extrema.
Lets allow now small variations of the atom clouds initial positions,
velocities and effective wave vector (atomic variables) and small
displacement and rotation of proof mass halves (see Fig. \ref{f5}). We
expect that main contribution to response arises from joined proof mass
halves, while for the separated halves contributions to the error decrease
when the distance between halves increases. We determine below, in Sec. \re
{s2.1.2.1}, the minimal distance $L_{d},$ starting from which the variations
of contribution to the response from separated halves becomes smaller than
ultimate phase error (\ref{4p}).
\begin{figure}[!t]
\includegraphics[width=11cm]{f5.eps}
\caption{Top view. Small variations of
the atomic and proof mass variables. Variations of proof mass halves
orientation and effective wave vector rotation are not shown.}
\label{f5}
\end{figure}
\subsection{\label{s2.1}Atomic variables}
\subsubsection{\label{s2.1.1}Joined proof mass halves.}
For the given atom cloud (1 or 2 in Table \ref{t2}) lets denote as $\left\{
\vec{a}_{0},\vec{v}_{0}\right\} $ extremal position and velocity and $\vec{k
_{0}=\left\{ 0,0,k\right\} ^{T}$ the vertical Raman field effective wave
vector. Including variations one has to substitute in the Eq. (\ref{3p})
\begin{subequations}
\label{2}
\begin{align}
\vec{a}& =\vec{a}_{0}+\delta \vec{a}, \label{2a} \\
\vec{v}& =\vec{v}_{0}+\delta \vec{v}, \label{2b} \\
\vec{k}& =R_{\vec{k}}\vec{k}_{0}, \label{2c}
\end{align
where $\delta \vec{a}_{i}=\left\{ \delta x_{i},\delta y_{i},\delta
z_{i}\right\} ^{T}$ for interferometer $i.$ We assumed in (\ref{2c}) that
Raman field consists only from counterpropagating wave vectors, but laser
axis could be slightly rotated from direction of $\vec{k}_{0}.$ For the
rotation matrix of this rotation we use Rodriguez rotation formula \cite{c8}
\end{subequations}
\begin{equation}
R_{\vec{k}ij}=\cos \left( \psi \right) \delta _{ij}+\dfrac{1-\cos \left(
\psi \right) }{\psi ^{2}}\psi _{i}\psi _{j}+\dfrac{\sin \left( \psi \right)
}{\psi }\varepsilon _{ijm}\psi _{m}, \label{3}
\end{equation
where $\vec{\psi}$ is an angle of rotation, $\delta _{ij}$ is Kronecker
symbol, $\varepsilon _{ijm}$ is absolutely antisymmetric tensor. For $\psi
\ll 1,
\begin{equation}
R_{\vec{k}ij}\approx \delta _{ij}+\varepsilon _{ijm}\psi _{m}-\dfrac{1}{2
\left( \psi ^{2}\delta _{ij}-\psi _{i}\psi _{j}\right) . \label{4}
\end{equation}
Using this expression and expanding in Eq. (\ref{3p}) up to the 2nd order in
respect to $\delta \vec{a},\delta \vec{v}$ and $\vec{\psi}$ one arrives to
the following approximate expression for the phase
\begin{subequations}
\label{5}
\begin{align}
\phi & \approx \vec{k}_{0}\cdot \left( \tau _{3}\vec{u}_{30}-t_{1}\vec{u
_{20}+\vec{u}_{21}-\vec{u}_{31}\right) \notag \\
& -\left( \vec{\psi}\times \vec{k}_{0}\right) \cdot \left( \tau _{3}\vec{u
_{30}-t_{1}\vec{u}_{20}+\vec{u}_{21}-\vec{u}_{31}\right) \notag \\
& +\delta \vec{a}_{p}\vec{k}_{0i}\left( \tau
_{3}b_{pi30}-t_{1}b_{pi20}+b_{pi21}-b_{pi31}\right) \notag \\
& +\delta v_{p}\vec{k}_{0i}\left( \tau
_{3}b_{pi31}-t_{1}b_{pi21}+b_{pi22}-b_{pi32}\right) \notag \\
& +\dfrac{1}{2}\delta \vec{a}_{p}\delta \vec{a}_{q}\vec{k}_{0i}\left( \tau
_{3}d_{pqi30}-t_{1}d_{pqi20}+d_{pqi21}-d_{pqi31}\right) \notag \\
& +\dfrac{1}{2}\delta v_{p}\delta v_{q}\vec{k}_{0i}\left( \tau
_{3}d_{pqi32}-t_{1}d_{pqi22}+d_{pqi23}-d_{pqi33}\right) \notag \\
& +\dfrac{1}{2}\left[ \vec{\psi}\times \left( \vec{\psi}\times \vec{k
_{0}\right) \right] \cdot \left( \tau _{3}\vec{u}_{30}-t_{1}\vec{u}_{20}
\vec{u}_{21}-\vec{u}_{31}\right) \notag \\
& +\delta \vec{a}_{p}\delta \vec{v}_{q}\vec{k}_{0i}\left( \tau
_{3}d_{pqi31}-t_{1}d_{pqi21}+d_{pqi22}-d_{pqi32}\right) \notag \\
& -\delta \vec{a}_{p}\left( \vec{\psi}\times \vec{k}_{0}\right) _{i}\left(
\tau _{3}b_{pi30}-t_{1}b_{pi20}+b_{pi21}-b_{pi31}\right) \notag \\
& -\delta v_{p}\left( \vec{\psi}\times \vec{k}_{0}\right) _{i}\left( \tau
_{3}b_{pi31}-t_{1}b_{pi21}+b_{pi22}-b_{pi32}\right) , \label{5a} \\
\vec{u}_{\alpha \beta }& =\int_{\tau _{\alpha -1}}^{\tau _{\alpha
}}dtt^{\beta }\delta \vec{g}\left( \vec{a}_{0}+\vec{v}_{0}t+\vec{g}\dfrac
t^{2}}{2}\right) , \label{5b} \\
b_{pi\alpha \beta }& \equiv \int_{\tau _{\alpha -1}}^{\tau _{\alpha
}}dtt^{\beta }\partial _{p}\delta \vec{g}_{i}\left( \vec{a}_{0}+\vec{v}_{0}t
\vec{g}\dfrac{t^{2}}{2}\right) , \label{5c} \\
d_{pqi\alpha \beta }& =\int_{\tau _{\alpha -1}}^{\tau _{\alpha }}dtt^{\beta
}\partial _{p}\partial _{q}\delta \vec{g}_{i}\left( \vec{a}_{0}+\vec{v}_{0}t
\vec{g}\dfrac{t^{2}}{2}\right) . \label{5d}
\end{align
A summation convention is implicit in Eq. (\ref{5a}) that will be used in
all subsequent equations, in which repeated indices and symbols are to be
summed over.
We calculated numerically coefficients in the expansion (\ref{5a}) for the
optimal conditions found in Sec. \ref{s1}. Different terms in the Eq. (\re
{5a}) are presented in Table \ref{t100}. We changed sign of the terms
associated with interferometer 2.
\begin{table}[h]
\caption{Error model for 100 kg proof mass
\begin{tabular}{|c|c|}
\hline
Term & relative weight \\ \hline
Linear in position &
\begin{tabular}{c}
$-.03159\delta z_{1}$ \\
$-0.05332\delta z_{2}
\end{tabular}
\\ \hline
Linear in velocity &
\begin{tabular}{c}
$0.0005170\delta v_{1z}$ \\
$-0.0001622\delta v_{2z}
\end{tabular}
\\ \hline
nonlinear in position &
\begin{tabular}{c}
$46.32\left( \delta x_{1}^{2}+\delta y_{1}^{2}\right) $ \\
$-92.64\delta z_{1}^{2}$ \\
$17.89\left( \delta x_{2}^{2}+\delta y_{2}^{2}\right) $ \\
$-35.79\delta z_{2}^{2}
\end{tabular}
\\ \hline
nonlinear in velocity &
\begin{tabular}{c}
$3.981\left( \delta v_{1x}^{2}+\delta v_{1y}^{2}\right) $ \\
$-7.962\delta v_{1z}^{2}$ \\
$1.587\left( \delta v_{2x}^{2}+\delta v_{2y}^{2}\right) $ \\
$-3.173\delta v_{2z}^{2}
\end{tabular}
\\ \hline
nonlinear in rotation &
\begin{tabular}{c}
$-0.2782\left( \psi_{1x}^{2}+\psi_{1y}^{2}\right) $ \\
$-0.2218\left( \psi_{2x}^{2}+\psi_{2y}^{2}\right)
\end{tabular}
\\ \hline
position-velocity cross term &
\begin{tabular}{c}
$27.00\left( \delta v_{1x}\delta x_{1}+\delta v_{1y}\delta y_{1}\right) $ \\
$-53.99\delta v_{1z}\delta z_{1}$ \\
$10.42\left( \delta v_{2x}\delta x_{2}+\delta v_{2y}\delta y_{2}\right) $ \\
$-20.83\delta v_{2z}\delta z_{2}
\end{tabular}
\\ \hline
position-rotation cross term &
\begin{tabular}{c}
$-0.01580\left( \delta x_{1}\psi_{1y}-\delta y_{1}\psi_{1x}\right) $ \\
$-0.02666\left( \delta x_{2}\psi_{2y}-\delta y_{2}\psi_{2x}\right)
\end{tabular}
\\ \hline
velocity-rotation cross term &
\begin{tabular}{c}
$0.0002585\left( \delta v_{1x}\psi_{1y}-\delta v_{1y}\psi_{1x}\right) $ \\
$-0.00008111\left( \delta v_{2x}\psi_{2y}-\delta v_{2y}\psi_{2x}\right)
\end{tabular}
\\ \hline
\end{tabular
\label{t100}
\end{table}
One sees that in spite of the using extremum points $\left\{
z_{i},v_{i}\right\} $ linear terms are not equal $0.$ It is because extrema
\left\{ z_{i},v_{i}\right\} $ have been found in Sec. \ref{s1}
approximately. One can find that coefficients in the linear dependences so
small that for allowed variations of position and velocity (see below Table
\ref{te100}) linear contributions are negligible.
One can use nonlinear terms to estimate atom clouds' radii and temperatures.
Consider for example relative contribution
\end{subequations}
\begin{equation}
\delta \varphi _{z}=\alpha \delta z_{i}^{2}. \label{8}
\end{equation
If Raman fields are sufficiently flat to neglect ac-Stark shift variation
across the atom cloud and if Raman pulses are sufficiently short to neglect
the Doppler broadening of the Raman transition, then one needs just to
average (\ref{8}) over atoms' spatial distribution. For Gaussian
distribution, $\dfrac{\exp \left[ -\delta z_{i}^{2}/\delta z_{i\max }^{2
\right] }{\sqrt{\pi }\delta z_{i\max }}$, after averaging one get
\begin{equation}
\left\langle \delta \varphi _{z}\right\rangle =\dfrac{\alpha }{2}\delta
z_{i\max }^{2} \label{9}
\end{equation
Requiring it to be equal expected relative error of phase measurement, $err,$
one finds for atom cloud radiu
\begin{equation}
\delta z_{i\max }=\sqrt{\dfrac{2\ast err}{\alpha }}. \label{10}
\end{equation
In the same manner we determine atom cloud velocities' variations,
temperatures and angle of the wave vector rotation. These quantities are
pieced together in the Table \ref{te100} for relative error value \ref{4p}
\begin{table}[h]
\caption{Parameters of the atom interferometers one has to hold for proof
mass 100 kg and relative error $10^{-4}.$
\begin{tabular}{cc}
1st cloud vertical radius $\delta z_{1\max }$ [m] & $0.001469$ \\
1st cloud vertical velocity $\delta v_{1z\max }$ [m/s] & $0.005012$ \\
1st cloud vertical temperature [K] & $2.070\ast 10^{-7}$ \\
1st cloud horizontal radius $\delta x_{1\max }$ [m] & $0.002078$ \\
1st cloud horizontal velocity $\delta v_{1x\max }$[m/s] & $0.007088$ \\
1st cloud horizontal temperature [K] & $4.139\ast 10^{-7}$ \\
1st interferometer wave vector rotation angle $\psi _{1\max }$ [rad] &
0.02681$ \\
2nd cloud vertical radius $\delta z_{2\max }$ [m] & $0.002364$ \\
2nd cloud vertical velocity $\delta v_{2z\max }$ [m/s] & $0.007939$ \\
2nd cloud vertical temperature [K] & $5.193\ast 10^{-7}$ \\
2nd cloud horizontal radius $\delta x_{2\max }$ [m] & $0.003343$ \\
2nd cloud horizontal velocity $\delta v_{2x\max }$ [m/s] & $0.01123$ \\
2nd cloud horizontal temperature [K] & $1.039\ast 10^{-6}$ \\
2nd interferometer wave vector rotation angle $\psi _{2\max }$ [rad] &
0.03003
\end{tabular
\label{te100}
\end{table}
\subsubsection{\label{s2.1.2}Separated proof mass halves.}
Contribution to the response from different terms in Eq. (\ref{5a}) arising
for separated proof mass halves are pieced together in the Table \ref{td100}.
\begin{table}[h]
\caption{Contribution to response from separated proof mass halves. Phase
decrease is a ratio of response to the phase difference for joined proof
mass halves. We changed sign for terms related to interferometer 1. Three
values in the curls correspond to the half-distance between proof masses
L_{d}=0.15$m, $0.3$m, and $1$m respectively. Values of $\protect\delta
x_{i\max },\ \protect\delta z_{i\max },\ \protect\delta v_{ix\max },\
\protect\delta v_{iz\max },\ \protect\psi _{i\max }$ are taken from table
\protect\ref{te100}.
\begin{tabular}{|c|c|c|}
\hline
Term & relative weight &
\begin{array}{c}
\delta x_{i}=\delta y_{i}=\delta x_{i\max },\delta z_{i}=\delta z_{i\max },
\\
\delta v_{ix}=\delta v_{iy}=\delta v_{ix\max },\delta v_{iz}=\delta
v_{iz\max },\psi _{i}=\psi _{i\max
\end{array
$ \\ \hline
Phase decrease & $\left\{ 0.84669512,0.95775894,0.9979968\right\} $ & \\
\hline
\begin{tabular}{c}
Linear \\
in positio
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.3435,0.1461,0.009211\right\} \delta z_{1}$ \\
$\left\{ -0.5453,-0.1844,-0.009591\right\} \delta z_{2}
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.0005046762,0.00021462856,0.000013534148\right\} $ \\
$\left\{ -0.0012890354,-0.00043602961,-0.000022672662\right\}
\end{tabular}
\\ \hline
\begin{tabular}{c}
Linear \\
in velocit
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.09949,0.04268,0.002719\right\} \delta v_{1z}$ \\
$\left\{ -0.1619,-0.05464,-0.002839\right\} \delta v_{2z}
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.00049866914,0.00021392743,0.000013627868\right\} $ \\
$\left\{ -0.001285252,-0.00043379244,-0.000022539355\right\}
\end{tabular}
\\ \hline
\begin{tabular}{c}
nonlinear \\
in positio
\end{tabular}
&
\begin{tabular}{c}
$\left\{ -5.532,-0.8019,-0.006320\right\} \left( \delta x_{1}^{2}+\delta
y_{1}^{2}\right) $ \\
$\left\{ 3.751,0.5657,0.004688\right\} \delta z_{1}^{2}$ \\
$\left\{ -4.643,-0.6147,-0.004004\right\} \left( \delta x_{2}^{2}+\delta
y_{2}^{2}\right) $ \\
$\left\{ 3.241,0.4444,0.002988\right\} $ $\delta z_{2}^{2}
\end{tabular}
&
\begin{tabular}{c}
$\left\{ -4.777\ast 10^{-5},-6.924\ast 10^{-6},-5.458\ast 10^{-8}\right\} $
\\
$\left\{ 8.098\ast 10\symbol{94}-6,1.221\ast 10\symbol{94}-6,1.012\ast 1
\symbol{94}-8\right\} $ \\
$\left\{ -0.0001038,-0.00001374,-8.951\ast 10\symbol{94}-8\right\} $ \\
$\left\{ 0.00001811,2.484\ast 10\symbol{94}-6,1.670\ast 10\symbol{94
-8\right\}
\end{tabular}
\\ \hline
\begin{tabular}{c}
nonlinear \\
in velocit
\end{tabular}
&
\begin{tabular}{c}
$\left\{ -0.5135,-0.07803,-0.0006637\right\} \left( \delta v_{1x}^{2}+\delta
v_{1y}^{2}\right) $ \\
$\left\{ 0.3424,0.05444,0.0004912\right\} \delta v_{1z}^{2}$ \\
$\left\{ -0.3943,-0.05179,-0.0003348\right\} \left( \delta v_{2x}^{2}+\delta
v_{2y}^{2}\right) $ \\
$\left\{ 0.2758,0.03748,0.0002499\right\} $ $\delta v_{2z}^{2}
\end{tabular}
&
\begin{tabular}{c}
$\left\{ -0.00005160,-7.840\ast 10\symbol{94}-6,-6.669\ast 10\symbol{94
-8\right\} $ \\
$\left\{ 8.602\ast 10\symbol{94}-6,1.367\ast 10\symbol{94}-6,1.234\ast 1
\symbol{94}-8\right\} $ \\
$\left\{ -0.00009940,-0.00001306,-8.442\ast 10\symbol{94}-8\right\} $ \\
$\left\{ 0.00003477,4.725\ast 10\symbol{94}-6,3.151\ast 10\symbol{94
-8\right\}
\end{tabular}
\\ \hline
\begin{tabular}{c}
nonlinear \\
in rotatio
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.04398,0.01255,0.0006208\right\} \left( \psi _{1x}^{2}+\psi
_{1y}^{2}\right) $ \\
$\left\{ 0.03268,0.008566,0.0003808\right\} \left( \psi _{2x}^{2}+\psi
_{2y}^{2}\right)
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.00006323,0.00001805,8.926\ast 10\symbol{94}-7\right\} $ \\
$\left\{ 0.00005893,0.00001544,6.867\ast 10\symbol{94}-7\right\}
\end{tabular}
\\ \hline
\begin{tabular}{c}
position- \\
velocity \\
cross ter
\end{tabular}
&
\begin{tabular}{c}
$\left\{ -3.245,-0.4732,-0.003779\right\} \left( \delta v_{1x}\delta
x_{1}+\delta v_{1y}\delta y_{1}\right) $ \\
$\left\{ 2.196,0.3332,0.002802\right\} \delta v_{1z}\delta z_{1}$ \\
$\left\{ -2.686,-0.3551,-0.002309\right\} \left( \delta v_{2x}\delta
x_{2}+\delta v_{2y}\delta y_{2}\right) $ \\
$\left\{ 1.876,0.2568,0.001724\right\} \delta v_{2z}\delta z_{2}
\end{tabular}
&
\begin{tabular}{c}
$\left\{ -0.00009559,-0.00001394,-1.113\ast 10\symbol{94}-7\right\} $ \\
$\left\{ 0.00001617,2.454\ast 10\symbol{94}-6,2.063\ast 10\symbol{94
-8\right\} $ \\
$\left\{ -0.0002017,-0.00002666,-1.734\ast 10\symbol{94}-7\right\} $ \\
$\left\{ 0.00003521,4.819\ast 10\symbol{94}-6,3.235\ast 10\symbol{94
-8\right\}
\end{tabular}
\\ \hline
\begin{tabular}{c}
position- \\
rotation \\
cross ter
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 1.114,0.3560,0.01893\right\} \left( \delta x_{1}\psi _{1y}-\delta
y_{1}\psi _{1x}\right) $ \\
$\left\{ -1.434,-0.4119,-0.01944\right\} \left( \delta x_{2}\psi
_{2y}-\delta y_{2}\psi _{2x}\right)
\end{tabular}
&
\begin{tabular}{c}
$0$ \\
$0
\end{tabular}
\\ \hline
\begin{tabular}{c}
velocity- \\
rotation \\
cross ter
\end{tabular}
&
\begin{tabular}{c}
$\left\{ 0.3259,0.1045,0.005593\right\} \left( \delta v_{1x}\psi
_{1y}-\delta v_{1y}\psi _{1x}\right) $ \\
$\left\{ -0.4251,-0.1220,-0.005755\right\} \left( \delta v_{2x}\psi
_{2y}-\delta v_{2y}\psi _{2x}\right)
\end{tabular}
&
\begin{tabular}{c}
$0$ \\
$0
\end{tabular}
\\ \hline
\end{tabular
\label{td100}
\end{table}
The point here is that even if the contribution to the response from
separated halves is small this case could be dangerous because launching
positions and velocities found above become no more extrema of the phase,
and therefore major contribution to the phase arises from the linear terms
in Table \ref{td100}. The only way to decrease these linear error is to
increase distance between proof masses $L_{d}$. Indeed for $L_{d}=0.15$m
linear in velocity errors can be 13 times large than ultimate relative
accuracy (\ref{4p}). For $L_{d}=0.3$m they are still 4 times larger. But for
$L_{d}=1$m all errors linear and nonlinear are well below than parameter
err.$
\paragraph{\label{s2.1.2.1}Minimal distance}
Since moving proof mass halves on the distances $\pm 1m$ could be a
technological challenge, we determine here minimal half-distance $L_{d}$ of
proof mass halves separation. For quantitative consideration we accept here
that the minimal $L_{d}$ is a distance at which all relative errors in the
3rd columns of the Table \ref{td100} are smaller than parameter $err.$ For
example, for $100$kg proof mass largest error in table \ref{td100} is linear
in position of the second interferometer cloud $\delta z_{2}.$ When
effective wave vector is vertical, $\vec{k}=\left\{ 0,0,k\right\} ,$ from
Eq. (\ref{5a}), one finds for this ter
\begin{equation}
\varphi _{d}\equiv \left\vert k\left( \tau
_{3}b_{3330}-t_{1}b_{3320}+b_{3321}-b_{3331}\right) \right\vert \delta
z_{2\max }, \label{11.1}
\end{equation
where tensor $b$ is defined in Eq. (\ref{5c}) and maximal variation of the
atom cloud vertical position, $\delta z_{2\max },$ one finds in the table
\ref{te100}. Fig. \ref{f6} shows dependence of the term (\ref{11.1}) on the
half distance $L_{d}$.
\begin{figure}[!t]
\includegraphics[width=11cm]{f6.eps}
\caption{Dependence of the error
\protect\ref{11.1}) on the half-distance between proof mass halves $L_{d}.$}
\label{f6}
\end{figure}
One sees that $\varphi _{d}$ becomes
smaller than $err$ at $L_{d}=0.58m.$ From the error model for this
half-distance , presented at the table \ref{td100ld} , one sees that all
other errors are also smaller than $err.
\begin{table}[tbph]
\caption{The same as in the Table \protect\ref{td100}, but for the distance
between proof mass halves $L_{d}=0.58$m.
\begin{tabular}{|c|c|c|}
\hline
Term & relative weight &
\begin{array}{c}
\delta x_{i}=\delta y_{i}=\delta x_{i\max },\delta z_{i}=\delta z_{i\max },
\\
\delta v_{xi}=\delta v_{yi}=\delta v_{xi\max },\delta v_{zi}=\delta
v_{zi\max },\psi _{i}=\psi _{i\max
\end{array
$ \\ \hline
Phase decrease & $0.99128250$ & \\ \hline
\begin{tabular}{c}
Linear \\
in positio
\end{tabular}
&
\begin{tabular}{c}
$0.03715\delta z_{1}$ \\
$-0.04099\delta z_{2}
\end{tabular}
&
\begin{tabular}{c}
$0.00005459$ \\
$-0.00009691
\end{tabular}
\\ \hline
\begin{tabular}{c}
Linear \\
in velocit
\end{tabular}
&
\begin{tabular}{c}
$0.01093\delta v_{1z}$ \\
$-0.01214\delta v_{2z}
\end{tabular}
&
\begin{tabular}{c}
$0.00005477$ \\
$-0.00009635
\end{tabular}
\\ \hline
\begin{tabular}{c}
nonlinear \\
in positio
\end{tabular}
&
\begin{tabular}{c}
$-0.06795\left( \delta x_{1}^{2}+\delta y_{1}^{2}\right) $ \\
$0.04963\delta z_{1}^{2}$ \\
$-0.04620\left( \delta x_{2}^{2}+\delta y_{2}^{2}\right) $ \\
$0.03420$ $\delta z_{2}^{2}
\end{tabular}
&
\begin{tabular}{c}
$-0.5868\ast 10^{-6}$ \\
$0.1071\ast 10^{-6}$ \\
$-0.1033\ast 10^{-5}$ \\
$0.1911\ast 10^{-6}
\end{tabular}
\\ \hline
\begin{tabular}{c}
nonlinear \\
in velocit
\end{tabular}
&
\begin{tabular}{c}
$-0.006947\left( \delta v_{1x}^{2}+\delta v_{1y}^{2}\right) $ \\
$0.005046\delta v_{1z}^{2}$ \\
$-0.003872\left( \delta v_{2x}^{2}+\delta v_{2y}^{2}\right) $ \\
$0.002867$ $\delta v_{2z}^{2}
\end{tabular}
&
\begin{tabular}{c}
$-0.6980\ast 10^{-6}$ \\
$0.1268\ast 10^{-6}$ \\
$-0.9761\ast 10^{-6}$ \\
$0.1807\ast 10^{-6}
\end{tabular}
\\ \hline
\begin{tabular}{c}
nonlinear \\
in rotatio
\end{tabular}
&
\begin{tabular}{c}
$0.002667\left( \psi _{1x}^{2}+\psi _{1y}^{2}\right) $ \\
$0.001692\left( \psi _{2x}^{2}+\psi _{2y}^{2}\right)
\end{tabular}
&
\begin{tabular}{c}
$0.3835\ast 10^{-5}$ \\
$0.3051\ast 10^{-5}
\end{tabular}
\\ \hline
\begin{tabular}{c}
position- \\
velocity \\
cross ter
\end{tabular}
&
\begin{tabular}{c}
$-0.04042\left( \delta v_{1x}\delta x_{1}+\delta v_{1y}\delta y_{1}\right) $
\\
$0.02949\delta v_{1z}\delta z_{1}$ \\
$-0.02666\left( \delta v_{2x}\delta x_{2}+\delta v_{2y}\delta y_{2}\right) $
\\
$0.01974\delta v_{2z}\delta z_{2}
\end{tabular}
&
\begin{tabular}{c}
$-0.1191\ast 10^{-5}$ \\
$0.2172\ast 10^{-6}$ \\
$-0.2001\ast 10^{-5}$ \\
$0.3704\ast 10^{-6}
\end{tabular}
\\ \hline
\begin{tabular}{c}
position- \\
rotation \\
cross ter
\end{tabular}
&
\begin{tabular}{c}
$0.07976\left( \delta x_{1}\psi _{1y}-\delta y_{1}\psi _{1x}\right) $ \\
$-0.08506\left( \delta x_{2}\psi _{2y}-\delta y_{2}\psi _{2x}\right)
\end{tabular}
&
\begin{tabular}{c}
$0$ \\
$0
\end{tabular}
\\ \hline
\begin{tabular}{c}
velocity- \\
rotation \\
cross ter
\end{tabular}
&
\begin{tabular}{c}
$0.02351\left( \delta v_{1x}\psi _{1y}-\delta v_{1y}\psi _{1x}\right) $ \\
$-0.02518\left( \delta v_{2x}\psi _{2y}-\delta v_{2y}\psi _{2x}\right)
\end{tabular}
&
\begin{tabular}{c}
$0$ \\
$0
\end{tabular}
\\ \hline
\end{tabular
\label{td100ld}
\end{table}
\subsection{\label{s2.2}Proof mass variables}
In this section we consider errors arising from variations of the joined
proof mass halves position and orientation. When the proof mass frame
shifted on $\delta \vec{c}$ and rotated on angle $\vec{\psi}$ in respect to
the lab. frame Eqs. (\ref{3p}) have to be rewritten as
\begin{subequations}
\begin{align}
\phi & =\vec{k}^{\prime }\cdot \left( \tau _{3}\vec{u}_{30}-t_{1}\vec{u
_{20}+\vec{u}_{21}-\vec{u}_{31}\right) , \label{12a} \\
\vec{u}_{\alpha \beta }& =\int_{\tau _{\alpha -1}}^{\tau _{\alpha
}}dtt^{\beta }\delta \vec{g}\left( \vec{a}^{\prime }+\vec{v}^{\prime }t+\vec
g}_{E}^{\prime }\dfrac{t^{2}}{2}\right) , \label{12b}
\end{align
where
\end{subequations}
\begin{subequations}
\label{13}
\begin{align}
\vec{k}^{\prime }& =R\vec{k}_{0}, \label{13a} \\
\vec{a}^{\prime }& =R\left( \vec{a}_{0}-\delta \vec{c}\right) , \label{13b}
\\
\vec{v}^{\prime }& =R\vec{v}_{0}, \label{13c} \\
\vec{g}_{E}^{\prime }& =R\vec{g}_{E} \label{13d}
\end{align
are, respectively, wave vector, atoms' launching position, atoms launching
velocity, and Earth gravity field in the proof mass frame, $R$ is rotation
matrix. Configurations considered above could not be optimal for both halves
of the proof mass and we allow the variations of these halves to be
independent then linear in $\delta \vec{c}$ and $\vec{\psi}$ terms should
dominate. So in this section we consider only linear corrections to the
phase, when
\end{subequations}
\begin{equation}
R_{ij}\approx \delta _{ij}+\varepsilon _{ijm}\psi _{m}. \label{14}
\end{equation
Expanding in Eq. (\ref{12b}) to the linear terms brings one to the following
expression for the phase$\
\begin{align}
\phi & \approx \vec{k}_{0}\left( \tau _{3}\vec{u}_{30}-t_{1}\vec{u}_{20}
\vec{u}_{21}-\vec{u}_{31}\right) \notag \\
& -\left( \vec{\psi}\times \vec{k}_{0}\right) \left( \tau _{3}\vec{u
_{30}-t_{1}\vec{u}_{20}+\vec{u}_{21}-\vec{u}_{31}\right) \notag \\
& -\left( \delta \vec{c}+\vec{\psi}\times \vec{a}_{0}\right) _{p}\vec{k
_{i}\left( \tau _{3}b_{pi30}-t_{1}b_{pi20}+b_{pi21}-b_{pi31}\right) \notag
\\
& -\left( \vec{\psi}\times \vec{v}_{0}\right) _{p}\vec{k}_{i}\left( \tau
_{3}b_{pi31}-t_{1}b_{pi21}+b_{pi22}-b_{pi32}\right) \notag \\
& -\dfrac{1}{2}\left( \vec{\psi}\times \vec{g}_{E}\right) _{p}\vec{k
_{i}\left( \tau _{3}b_{pi32}-t_{1}b_{pi22}+b_{pi23}-b_{pi33}\right) ,
\label{15}
\end{align
where tensor $b_{pi\alpha \beta }$ is defined in Eq. (\ref{5c}). For the
chosen proof mas halves' geometry, location and orientation and unperturbed
atomic variables, numeric integration brings one to the following linear
dependence of the phase differenc
\begin{equation}
\Delta \phi \approx 0.5527\left[ 1+6.782\left( \text{$\delta $}c_{1x}-\text{
\delta $}c_{2x}\right) +0.04246\left( \text{$\delta $}c_{1z}+\text{$\delta $
c_{2z}\right) -0.09261\left( \psi _{1y}-\psi _{2y}\right) \right] \label{16}
\end{equation
where variation of the left (right) half-proof mass position and angle of
rotation are $\delta \vec{c}_{1}$ $\left( \delta \vec{c}_{2}\right) $ and
\vec{\psi}_{1}\ \left( \vec{\psi}_{2}\right) .$ One sees that the phase is
most sensitive to displacement along $x-$axis (see Fig. \ref{f1}a). From the
symmetric shapes there are no linear sensitivity to the displacement along
y-$axis, rotations in respect to the $z-$ and $x-$axes. When one synchronize
displacement along $x-$axis and rotation of both proof mass halves
corresponding linear dependences disappear. Since in the absence of
rotation\ synchronized displacement of the proof mass halves is equivalent
to the synchronized displacement of both interferometers in the opposite
directions, the slopes of the linear dependences on $\delta z$$_{i}$ equal
to the average slopes in the linear dependences on the interferometers'
displacement taken with opposite sign [compare corresponding coefficients in
Eq. (\ref{16}) and first 2 rows in the Tables \ref{t100}]. Since for optimal
configuration $z-$coordinates of the atom clouds launching points are closed
to the extrema, the slopes of the dependence on $\delta c_{iz}$ in Eq. (\re
{16}) are 2 orders of magnitude smaller than slopes of the dependence on
\delta c_{ix}.$
From the Eq. (\ref{16}) one concludes that ultimate accuracy (\ref{4p}) can
be achieved for proof mass halves positioning with accuracy
\begin{subequations}
\begin{eqnarray}
\left\vert \text{$\delta $}c_{ix}\right\vert &<&14.74\mu , \label{17a} \\
\left\vert \psi _{iy}\right\vert &<&1.08\text{mrad.} \label{17b}
\end{eqnarray}
\section{Conclusion}
We showed that $100$kg Tungsten proof mass can produce change in the atom
interferometers phase double difference
\end{subequations}
\begin{equation}
\delta \Delta \phi =0.54789287\text{rad} \label{18}
\end{equation
for the $^{133}Cs$ atom interferometers, with parameters listed in Table \re
{t1}, extremal values of atom interferometers launching position and
velocities and proof mass sizes listed in Table \ref{t2}. The response (\re
{18}) is comparable with that observed in \cite{c3}, but the choice of
phase's extrema allowed us to make requirements for atoms' positioning 2
orders of magnitude less severe than requirements for proof mass halves
positioning [compare (\ref{17a}) and data in Table \ref{te100}].
\acknowledgments
Author is appreciated to Drs. M. Kasevich, B. Young, S. Libby, M. Matthews,
T. Loftus, M. Shverdin, V. Sonnad and A. Zorn for fruitful discussion and
collaboration. Special gratefulness to S. Libby and V. Sonnad who showed me
their unpublished results, and to Dr. A. Zorn, who brought to my attention
article \cite{c9}.
The support of this work by Lawrence Livermore National Laboratory, LDRD 12-
LW-009: \textquotedblleft High-Precision Test of the Gravitational
Inverse-Square Law with an Atom Interferometer,\textquotedblright is gratefully acknowledged.
This work was partially performed under the auspices of the U.S. Department
of Energy by Lawrence Livermore National Laboratory under Contract
DE-AC52-07NA27344.
|
\section{Introduction}
The volume weighted average price (VWAP) occurs frequently in finance. It is an average price which gives more weight to periods of high trading than to periods of low trading in its calculation. A broker's daily performance is frequently measured against the VWAP and it is becoming increasingly popular for institutional investors to place buy and sell orders at the VWAP. The VWAP also appears in Australian taxation law as part of
determining the prices of share buy-backs in publicly listed companies (Woellner {\it et al.} (2009)).
Most of the existing literature on VWAP focuses on strategies and algorithms
to execute orders as close as possible to the VWAP price (see e.g. Konishi (2002), Bialkowski {\it et al.} (2008), Fuh {\it et al.} (2010), Frei \& Westray (2013)).
On the other hand, surprisingly few results on actual pricing methodologies related to VWAP options have been published (Stace (2007), Novikov {\it et al.} (2014)).
This can be explained by pointing out that the VWAP pricing problem is set in an incomplete market since there is no underlying with which to hedge the volume risk, and hence there is no uniquely defined price. Any price obtained will include a market price of volume risk which must be determined from the corresponding volume statistics.
In this paper, we propose a new model to price VWAP options in which the volume data is modelled by a gamma process. Exact closed-form expressions are derived for the first two moments of the VWAP, which may be used price VWAP options via well-known moment matching techniques. We then compare our results against the technique suggested by Stace (2007)
as well as with Monte Carlo modelling results.
The rest of this paper is organised as follows. Section 2 briefly describes some of the previously suggested models for volume data.
Section 3 justifies our choice of the gamma process as the preferred volume model by presenting goodness-of-fit results and other analyses of volume data. Section 4 formally introduces our model for both stock price (lognormal) and stock volume (gamma).
Section 5 presents the main results of the paper, which include closed-form expressions for the VWAP moments and option prices (based on a moment matching technique) as well as a comparison to Monte Carlo results.
Section 6 contains some concluding remarks.
Detailed derivations of VWAP moments, for both discrete and continuous-time cases, can be found in Appendix A.
\section{Previously suggested models for volume process.}
It is common to model the underlying stock price $S_t$ using the standard geometric Brownian motion. On the other hand, the choice of model for the volume traded $V_t$ is far less obvious. A few versions of the volume process $V_t$ have been suggested in the literature. We shall briefly outline some of these existing approaches before selecting our own.
For example, Stace (2007)
has considered the following mean reverting volume process:
\begin{gather} \label{Stace_V_evolution}
{d V_t = \lambda(V_{\rm{mean}} - V_t)dt + \beta V_t^f dW,}
\end{gather}
where $V_{t=0}$ is given, $\lambda$ is the speed of mean reversion, $V_{\rm{mean}}$ is the long term average of the
volume process, $\beta$ is the volatility of the volume process, $W$ is a standard Brownian motion (which may be partially correlated to the stock price process Brownian motion), and $f$ is either $1$ or $0.5$.
In the more recent work of Novikov {\it et al.} (2014)
a different class of volume processes has been suggested:
\begin{gather} \label{Novikov_V_evolution}
V_t = X_t^2 + \delta,\qquad
d X_t = \lambda(X_{\rm{mean}} - X_t) dt + \beta dW,
\end{gather}
where $X$ is a standard Ornstein-Uhlenbeck (OU) process with $\lambda$, $X_{\rm{mean}}$ and $\beta$ being its speed of mean-reversion, level of mean-reversion and OU volatility, respectively.
Yet another $V_t$ model, also based on presence of a second Brownian motion in underlying dynamics was suggested in Fuh {\it et al.} (2010).
However, all these and many other VWAP-related publications are predominantly concentrated on the description of their pricing or trading algorithms, with little attention given to any justification of the corresponding volume process model choice or any comprehensive empirical analysis of volume statistics. A notable exception is Frei \& Westray (2013),
in which a considerable amount of attention is paid to such a justification, as well as to outlining of useful approaches for empirical checks. Importantly, Frei \& Westray (2013)
argues that there is substantial empirical evidence to suggest that $V_t$ can be modelled by i.i.d.\! gamma random noises. This choice may look unusual for a financial model, but as Brody {\it et al.} (2008)
demonstrates, gamma processes actually have a broad range of applications in many areas of insurance and finance, where cumulative processes are involved. These include the modelling of aggregative claims, credit portfolio losses, defined benefit pension schemes and so on.
\section{Empirical analysis of volume data.}
In this section we present our empirical analysis justifying the use of gamma variables to model underlying dynamics of traded volume. This, in turn, allows us to progress with the development of a VWAP option pricing model in the next section.
In particular, we analyse the price and the volume data series for a few ASX stocks.
Data for our analysis is obtained from Bloomberg and covers the period from 15/02/2013 until 28/08/2013. Each original data point represents a traded volume and the corresponding VWAP price for a 10 minute interval. Typically 38 traded volume data points $V_i$ per day are available (6 per hour for 6 hours, plus one pre-market trading [10 am] point and one post-market [4.10 pm] point). Thus, on average, we have about 5130 data points per stock. This number varies slightly from stock to stock, because occasionally some data points are missing (e.g., due to the lack of trading within certain 10 min intervals).
For each equity volume data set we construct 4 secondary sets $V_i^{(L)}$ by combining volumes of $L$ consecutive points (and thus amalgamating the original $L$-point groups to single points in the newly derived data sets), where $L = 5, 10, 20$ and $40$, corresponding to approximately $1/8$, $1/4$, $1/2$ and $1$ day incremental volumes, respectively. Then gamma distributions are fitted to these newly constructed amalgamated volume data sets. We recall that the standard gamma distribution $\Gamma(\alpha,\theta)$ has a mean $\alpha \theta$ and variance $\alpha \theta^2$ (see e.g., Brody {\it et al.} (2008)
for a detailed discussion of gamma distribution properties and additional relevant references).
Due to properties of the gamma distribution, if the original (non-amalgamated) volume data follows a true gamma distribution, then its $\theta$ parameter would stay constant with respect to $L$, its $\alpha$ parameter would scale so that $\alpha(L)/L$ is constant, and all time series and autocorrelation coefficients would satisfy $C_{\rm auto}(L) \ll 1$. In addition to the analysis of $\theta(L)$, $\alpha(L)$ and $C_{\rm auto}(L)$, we perform two goodness-of-fit tests: Anderson-Darling (A-D) and Kolmogorov-Smirnov (K-S), based on the hypothesis of $V_i^{(L)}$ having gamma distribution. The A-D and K-S goodness-of-fit analysis was conducted using the Mathematica 9.0 software package and below we only report the corresponding $P$-values (with higher values meaning higher probability of the hypothesis being correct).
\begin{table}[ht]
\caption{Statistical analysis of CBA, WDC and FMG stock volume data (data source: Bloomberg) and its comparison with analysis of synthetic (computer-generated) data sample (with gamma distribution parameters $\alpha = 1.0$ and $\theta = 0.2 \times 10^6$). For stock data each original data point represents a traded volume and the corresponding VWAP price for a 10 minute interval (approx. 5130 points are available for each stock; for synthetic data we also generate 5130 points). Then 4 secondary sets $V_i^{(L)}$ are constructed by combining volumes of $L$ consecutive points ($L = 5, 10, 20, 40$). Then gamma distributions are fitted to these amalgamated volume data sets (i.e., we obtain fitted gamma distribution parameter values $\theta(L)$ and $\alpha(L)$ for each set) and autocorrelation coefficients $C_{\rm auto}(L)$ are also calculated. In addition, we perform two goodness-of-fit tests: Anderson-Darling (A-D) and Kolmogorov-Smirnov (K-S), based on the hypothesis of $V_i^{(L)}$ having a gamma distribution and report the corresponding $P$-values (with higher values meaning higher probability of the hypothesis being correct).}
{\begin{tabular}{|c|c|c|c|c|c|c|c|} \hline
& & & & & & & \\
${\rm stock/}$ & $L$ & $\theta(L)/10^6$ & $\alpha(L)$ & $\alpha/L$ & $C_{\rm auto}(L)$ & $P^{\rm(A-D)}$ & $P^{\rm(K-S)}$ \\
${\rm synthetic}$ & & & & & & &\\
& & & & & & &\\
\hline
& & & & & & &\\
${\rm CBA}$ & 5 & 0.33 & 3.23 & 0.65 & 39\% & 0.01\% & 0.11\% \\
${\rm CBA}$ & 10 & 0.54 & 3.95 & 0.39 & 21\% & 5.0\% & 4.6\% \\
${\rm CBA}$ & 20 & 0.73 & 5.84 & 0.29 & 8\% & 59.1\% & 49.4\% \\
${\rm CBA}$ & 40 & 0.77 & 11.07 & 0.28 & 45\% & 90.8\% & 74.8\% \\
& & & & & & & \\
\hline
& & & & & & &\\
${\rm WDC}$ & 5 & 0.95 & 2.19 & 0.44 & 21\% & 0.00\% & 0.00\% \\
${\rm WDC}$ & 10 & 1.46 & 2.87 & 0.29 & 2\% & 2.8\% & 1.9\% \\
${\rm WDC}$ & 20 & 1.77 & 4.73 & 0.24 & -14\%& 54.6\% & 31.6\% \\
${\rm WDC}$ & 40 & 1.55 & 10.82 & 0.27 & 48\%& 99.7\% & 98.7\% \\
& & & & & & & \\
\hline
& & & & & & &\\
${\rm FMG}$ & 5 & 2.58 & 2.82 & 0.56 & 40\% & 1.7\% & 42.3\% \\
${\rm FMG}$ & 10 & 3.88 & 3.75 & 0.38 & 25\% & 37.4\% & 32.3\% \\
${\rm FMG}$ & 20 & 5.14 & 5.66 & 0.28 & 10\% & 73.1\% & 58.4\% \\
${\rm FMG}$ & 40 & 5.29 & 10.99 & 0.27 & 33\%& 93.9\% & 95.9\% \\
& & & & & & & \\
\hline
& & & & & & &\\
${\rm synthetic}$ & 5 & 0.20 & 4.99 & 1.00 & 2\% & 93.2\% & 90.4\% \\
${\rm synthetic}$ & 10 & 0.21 & 9.50 & 0.95 & 0\% & 99.9\% & 94.0\% \\
${\rm synthetic}$ & 20 & 0.22 & 17.65 & 0.88 & 2\% & 90.7\% & 89.6\% \\
${\rm synthetic}$ & 40 & 0.20 & 38.18 & 0.98 & 4\% & 84.2\% & 76.2\% \\
& & & & & & & \\
\hline
\end{tabular}}
\label{CBA}
\end{table}
Our results for CBA, WDC and FMG stocks are presented in Table 1 (all these stocks are part of the ASX 200 index). These results typically display significant deviations from true gamma distribution behaviours for smaller bucket size $L$, but strongly support the hypothesis of $V_i^{(L)}$ having gamma distributions for $L \ge 20$ (i.e., with half a day or longer averaging).
For the sake of giving the reader a better feel of how close our stock volume statistics hypothesis is to reality, we use the algorithm of Press {\it et al.} (2007)
to generate the same number of sample points of ``true'' gamma-distributed noise (5130 points) and repeat our analysis. The corresponding results are also included in Table 1.
\begin{figure}[h]
\begin{center}
\resizebox*{8cm}{!}{\includegraphics{Fig1v3.eps}
\caption{Correlation between cumulative volume $V^{(c)}_i = \sum_{j = 1}^i V_j$ and the corresponding relative trading volume $V_i$ for CBA, WDC and FMG. Each correlation point is calculated for data sets obtained for the same time period $[t_i,t_i+\Delta t]$ of available trading days.
\label{Fig1}
\end{center}
\end{figure}
Finally we note that some other tests suggested by Frei \& Westray (2013)
are also performed. The most important one being a test demonstrating a low correlation between cumulative ($\sum_{j = 1}^i V_j$) and incremental volumes ($V_i$) within the same time period of a trading day $[t_i,t_i+\Delta t]$ (i.e. day-to-day independence of $\sum_{j = 1}^{i} V_j$ and $V_i$; see Frei \& Westray (2013) for details).
Figure 1 shows the corresponding intra-day correlations for CBA, WDC and FMG stocks, which are indeed reasonably low.
All results reported in this section strongly support the hypothesis that at least for the averaging interval of half a day or longer the stock volume dynamics can indeed be described as the standard gamma process: if the volume traded within the time period $[t_i,t_i+\Delta t]$ is given by $V_{i}$, then we assume that it has the gamma distribution $\Gamma(\alpha,\theta)$, with mean $\alpha \theta$ and variance $\alpha \theta^2$, where $\alpha$ depends linearly on the averaging period $\Delta t$. Due to the independent increment property of the gamma process, volumes traded in disjoint time intervals of equal length are i.i.d.\! gamma variables.
\section{Model for VWAP option pricing.}
We now formally define a new model for the VWAP option pricing using a gamma process for volume dynamics.
We work under the filtered probability space $(\Omega, \mathcal{F}, \mathbb{F}, \mathbb{Q})$. The filtration $\mathbb{F}=\{\mathcal{F}_t\}$ represents the flow of information available to market participants. In particular, it is the augmented filtration generated by a standard Brownian motion $W_t$ and a sequence of i.i.d.\! gamma variables $V_1,\ldots, V_N \sim \Gamma(\alpha,\theta)$. For every $i=1,\ldots,N$, the gamma variable $V_i$ is $\mathcal{F}_{t_i}$-measurable where $t_i:=i\Delta t$ for some fixed time increment $\Delta t$.
Note that the process $W_t$ and the random variables $\{V_i\}$ are assumed to be independent. For convenience we choose $W_t$ to be the Brownian motion under the risk-neutral pricing measure. So the measure $\mathbb{Q}$ is the product of the risk-neutral pricing measure and the real-world measure associated with $\{V_i\}$.
To summarise: the \emph{stock price} process $S_t$ is given by the standard geometric Brownian motion
\begin{gather} \label{S_evolution}
d S_t = r S_t dt + \sigma S_t dW_t,
\end{gather}
where $r$ is the risk-free interest rate and $\sigma$ is the volatility. Note that the discounted stock price $e^{-rt}S_t$ is a martingale under $\mathbb{Q}$.
The \emph{volumes traded} $V_{t_i}$ traded during the time periods $[t_{i-1},t_{i}]$ (where $t_{i} = i\Delta t$) are directly modelled by the i.i.d.\! gamma variables,
\begin{gather} \label{Z_evolution}
V_{t_i} := V_i \sim \Gamma(\alpha,\theta),
\end{gather}
with mean $\alpha \theta$ and variance $\alpha \theta^2$.
Now we can define the VWAP on a time interval $[0, T = N \Delta t]$ as
\begin{align} \label{VWAPd}
S^{\rm (VWAP)} = \frac{\sum^N_{i=1} S_i V_i}{\sum^N_{i=1} V_i}.
\end{align}
This paper will focus on the discrete-time formulation of VWAP \eqref{VWAPd} which is a weighted average of $N$ incremental trade volumes as opposed to a weighted integral over time. For completeness, the results for the continuous-time limit can be found in Appendix A.
Note that we will consider the standard call and put VWAP options:
\begin{align} \label{VWAP_payouts}
C &= \max(S^{\rm (VWAP)} - K,0),\\
P &= \max(K - S^{\rm (VWAP)}, 0).
\end{align}
\section{Closed-form pricing formulas for VWAP options}
It is natural to start further analysis with a brief description of currently known approaches (Stace (2007), Novikov \& Kordzakhia (2013), Novikov {\it et al.} (2014)).
for VWAP option pricing. Note that two of these works (Stace (2007), Novikov {\it et al.} (2014))
utilise a well-known moment-matching technique, which works reasonably well for problems such as pricing Asian options (see e.g., Hull (2006)).
In this technique, we match the first two moments of the VWAP to the corresponding moments of the standard lognormal process (for which all pricing formulas are well-known) which have the following form (see, e.g., Glasserman (2003)):
\begin{align} \label{Lognormal_moments}
M_1^{(LN)}&={\E}[S_t] = S_0 \exp{(r t}),\\
M_2^{(LN)}&={\Var}[S_t] = S_0^2 [\exp{(\sigma^2 t)} - 1] \exp{(2 r t)},
\end{align}
where all notations are the same as (\ref{S_evolution}). After matching calculated VWAP moments to the lognormal moments (i.e. taking $M_1^{(LN)}=M_1^{(VWAP)}$ and $M_2^{(LN)}=M_2^{(VWAP)}$), one can use the standard Black formulas for options pricing with the forward $F_0 = M_1$ and the volatility $\sigma^2 =\ln{(M_2/M_1^2+1)}/T$ (see, e.g., Hull (2006)).
The approach is known to work very well for low volatilities ($\sigma \le 0.2$) and reasonably well for higher volatility levels ($0.2 < \sigma \le 0.4$).
In the work of Stace (2007)
the standard moment matching approach (described above) is supplemented by further approximations during the calculation of the VWAP moments, utilising the following expressions:
\begin{align} \label{Ratios}
{\rm \E}\left(\frac{Y}{Z}\right) &\approx \frac{{\rm \E}\left(Y \right)}{{\rm \E}\left(Z \right)}
- \frac{{\rm Cov}\left(Y, Z \right)}{\left[{\rm \E}\left(Z \right)\right]^2} + \frac{{\rm \E}\left(Y \right)}{\left[{\rm \E}\left(Z \right) \right]^3} {\rm Var}(Z),\\
{\rm Var}\left(\frac{Y}{Z}\right) &\approx \left(\frac{{\rm \E}\left(Y \right)}{{\rm \E}\left(Z \right)}\right)^2 \left( \frac{{\rm Var}\left(Y \right)}{\left[{\rm \E}\left(Y \right)\right]^2}
+\frac{{\rm Var}\left(Z \right)}{\left[{\rm \E}\left(Z \right)\right]^2} - 2 \frac{{\rm Cov}\left(Y,Z \right)}{{\rm \E}\left(Y \right){\rm \E}\left(Z \right)} \right),\label{Ratios2}
\end{align}
where all terms may be computed explicitly after lengthy, but rather straightforward calculations. Note that these approximations are rather well-known (see e.g., Mood {\it et al.} (1974), p.\! 181).
Another possible approach for VWAP option pricing was recently outlined in Novikov \& Kordzakhia (2013)
and is essentially based on an upper/lower bounds approach, which is also well-known for Asian options (see, e.g., Curran (1994), Rogers \& Shi (1995), Thompson (2000), Lord (2006) to name a few).
Although more accurate than the moment-matching approach for higher volatility values ($\sigma > 0.2$), this approach typically does not allow closed-form analytical representations (some numerical search for minimum/maximum values over one or more parameters is present).
At this point we choose to follow the simpler and more tractable moment matching method of pricing, aiming to get a simple closed-form analytical result at the end. Thus we need to calculate the first and the second moments for $S^{\rm (VWAP)}$ which is given by Eq. (\ref{VWAPd}), where evolution of $S_t$ and $V_t$ are defined by Eq. (\ref{S_evolution}) and Eq. (\ref{Z_evolution}), respectively.
It is important to note that, in contrast to lognormal volume dynamics models, the system (\ref{S_evolution})--(\ref{Z_evolution}) allows us to obtain a closed-form moment matching-based pricing solution, both with and without using an additional approximation (\ref{Ratios})--(\ref{Ratios2}).
Detailed derivation of the solution to the system (\ref{S_evolution})--(\ref{Z_evolution}) can be found in Appendix A. Here we only present some final expressions for the volatility correction factor $R^{\rm (VWAP)} \equiv \sigma_D^{\rm (VWAP)}/\sigma_D^{\rm (AA)}$, (where $\sigma_D^{\rm (AA)}$ represents the standard moment matching-based implied volatility result for arithmetic averaging Asian options) noting that compact closed-form asymptotic expansion formulas are available in the limit $T/N \ll 1$:
\begin{align} \label{Exact}
R^{\rm (VWAP)}_{\rm Exact} &= \sqrt{\frac{N (3 + \alpha + 2 \alpha N)}{(1+2 N) (1 + \alpha N)}+ O(T/N)},\\
\label{Stace_like}
R^{\rm (VWAP)}_{\rm Stace} &= \sqrt{\frac{(3 + \alpha + 2 \alpha N)}{(\alpha + 2 \alpha N)}+ O(T/N)}.
\end{align}
Also note that for our gamma-process model $F^{\rm(VWAP)} = F^{\rm (AA)}$, i.e., no VWAP forward correction is required in comparison to the corresponding arithmetic averaging Asian result.
A few observations can be made from the analysis of (\ref{Exact})--(\ref{Stace_like}):\\
(i) the parameter $\theta$ is not present in the result (due to its absence in the relative variance formula for the gamma distribution: $\Var(\Gamma)/\E(\Gamma)^2 = 1/\alpha$),\\
(ii) in case of daily ($T/N = 1/252$) or more frequent averaging, all the extra terms $O(T/N)$ are negligible in comparison to the leading order terms for any value of $T$,\\
(iii) for up to $T/N \sim 0.5$, the dependence of $R^{\rm VWAP}$ on $T$ is weak (in comparison with its dependence on $\alpha$ or $N$) and may be ignored, at least as a first approximation.
We may now check the quality of the results (\ref{Exact})--(\ref{Stace_like}) by comparing them to direct Monte-Carlo (MC) simulations. The results are presented in Table 2.
\begin{table}[h]
\caption{Comparison of the gamma model effective volatility results with the corresponding MC results for the following parameter values: $\sigma = 0.2$, $r = 0.05$, $\theta = 0.00067$, MC number of paths $= 10000000$, $T = 2/52$ (with $N=10$ averaging points). We note that the $R^{\rm (VWAP)}_{\rm Exact}$ column gives a ratio $\sigma^{\rm (VWAP)}_{\rm ExactD}/\sigma_D^{\rm (AA)}$ result based on the discrete-time exact formulas of Appendix A; the $R^{\rm (VWAP)}_{\rm Stace}$ column gives $\sigma^{\rm (VWAP)}_{\rm StaceD}/\sigma_D^{\rm (AA)}$ result based on the discrete-time ``Stace-like'' formulas of Appendix A; $R^{\rm (VWAP)}_{\rm MC}$ column gives the corresponding result of direct MC modelling $\sigma_{\rm MC}^{\rm (VWAP)}/\sigma_{\rm MC}^{\rm (AA)}$; and the last column provides an error estimate for the ``Stace-like'' ratio result $(R^{\rm (VWAP)}_{\rm Stace}-1)/(R^{\rm (VWAP)}_{\rm MC}-1)$.}
{\begin{tabular}{|c|c|c|c|c|} \hline
& & & & \\
$1/\alpha$ & $R^{\rm (VWAP)}_{\rm Exact}$ & $R^{\rm (VWAP)}_{\rm Stace}$ & $R^{\rm (VWAP)}_{\rm MC}$ & $\frac{(R^{\rm (VWAP)}_{\rm Stace}-1)}{(R^{\rm (VWAP)}_{\rm MC}-1)}$ \\
& & & & \\
\hline
& & & & \\
0.00 & 1.0000 & 1.0000 & 1.0000 $\pm$ 0.0002 & - \\
0.02 & 1.0004 & 1.0014 & 1.0004 $\pm$ 0.0002 & 3.40 \\
0.20 & 1.0042 & 1.0142 & 1.0042 $\pm$ 0.0002 & 3.36 \\
0.50 & 1.0102 & 1.0351 & 1.0102 $\pm$ 0.0002 & 3.43 \\
0.75 & 1.0148 & 1.0351 & 1.0150 $\pm$ 0.0002 & 3.49 \\
1.00 & 1.0193 & 1.0522 & 1.0194 $\pm$ 0.0002 & 3.56 \\
1.20 & 1.0227 & 1.0823 & 1.0228 $\pm$ 0.0002 & 3.62 \\
1.50 & 1.0276 & 1.1019 & 1.0276 $\pm$ 0.0002 & 3.69 \\
1.80 & 1.0322 & 1.1212 & 1.0322 $\pm$ 0.0002 & 3.76 \\
2.00 & 1.0351 & 1.1339 & 1.0352 $\pm$ 0.0002 & 3.80 \\
& & & & \\
\hline
\end{tabular}}
\label{Gamma}
\end{table}
The parameter values in Table 2 are as follows: $\sigma = 0.2$, $r = 0.05$, $\theta = 0.00067$, MC number of paths $= 10000000$, $T = 2/52$ (with $N=10$ averaging points, i.e. a two-week term with daily averaging). More specifically:
the $R^{\rm (VWAP)}_{\rm Exact}$ column gives a ratio $\sigma^{\rm (VWAP)}_{\rm ExactD}/\sigma_D^{\rm (AA)}$ result based on the discrete-time exact formulas of Appendix A; the $R^{\rm (VWAP)}_{\rm Stace}$ column gives $\sigma^{\rm (VWAP)}_{\rm StaceD}/\sigma_D^{\rm (AA)}$ result based on the discrete-time ``Stace-like'' formulas of Appendix A; $R^{\rm (VWAP)}_{\rm MC}$ column gives the corresponding result of direct MC modelling $\sigma_{\rm MC}^{\rm (VWAP)}/\sigma_{\rm MC}^{\rm (AA)}$; and the last column provides an error estimate for the ``Stace-like'' ratio result $(R^{\rm (VWAP)}_{\rm Stace}-1)/(R^{\rm (VWAP)}_{\rm MC}-1)$.
Note that we omit a similar comparison for $R^{\rm (VWAP)}_{\rm Exact}$ ratio as it perfectly matches the MC result (i.e. is always within MC error margin).
We see a substantial disagreement of Stace approximation (\ref{Stace_like}) with numerics, whereas the exact solution (\ref{Exact}) matches MC results perfectly (within MC error bounds).
Table 3 provides the comparison of VWAP implied volatilities (based on the gamma model considered in this Section) and the corresponding vanilla option prices with their arithmetic Asian option counterparts. For all examples, the parameters are $\sigma = 0.2$,
$r = 0.05$ and $S = K = \$ 100.00$.
\begin{table}[h]
\caption{Comparison of VWAP implied volatilities $\sigma^{\rm (VWAP)}_D$ and VWAP option prices $P^{\rm (VWAP)}_D$ with the corresponding arithmetic Asian volatilities $\sigma^{\rm (AA)}_D$ and option prices $P^{\rm (AA)}_D$ for some typical values of gamma parameter $\alpha$, option tenor $T$ and number of averaging points $N$. For all examples, the parameters are $\sigma = 0.2$,
$r = 0.05$ and $S = K = \$ 100.00$. }
{\begin{tabular}{|c|c|c|c|c|c|c|c|c|} \hline
& & & & & & & & \\
Type & $T$ & $\alpha$ & $N$ & $\sigma^{\rm (AA)}_D$ & $\sigma^{\rm (VWAP)}_D$
& $P^{\rm (AA)}_D$ & $P^{\rm (VWAP)}_D$ & $\frac{P^{\rm (VWAP)}_D}{ P^{\rm (AA)}_D} - 1$ \\
& [years] & & & [$\%$] & [$\%$] & [$\$ $] & [$\$ $] & [$\%$] \\
& & & & & & & & \\
\hline
& & & & & & & & \\
put & 0.317 & 10 & 80 & 11.68 & 11.69 & 2.217 & 2.217 & 0.04 \\
call & 0.317 & 10 & 80 & 11.68 & 11.69 & 3.012 & 3.012 & 0.03 \\
put & 0.079 & 10 & 20 & 11.99 & 12.00 & 1.242 & 1.242 & 0.12 \\
call & 0.079 & 10 & 20 & 11.99 & 12.00 & 1.450 & 1.450 & 0.11 \\
put & 0.020 & 10 & 5 & 13.27 & 13.32 & 0.716 & 0.718 & 0.37 \\
call & 0.020 & 10 & 5 & 13.27 & 13.32 & 0.775 & 0.778 & 0.34 \\
put & 0.020 & 5 & 5 & 13.27 & 13.36 & 0.716 & 0.721 & 0.73 \\
& & & & & & & & \\
\hline
\end{tabular}}
\label{Examples}
\end{table}
The last three columns of Table 3 report arithmetic Asian and VWAP option prices and their relative difference for a typical for ASX stocks ($\alpha = 10.0$) and lowest observed ($\alpha = 5.0$) values of the parameter $\alpha$. Clearly this relative difference is quite noticeable for some examples.
\section{Conclusions and discussion}
In conclusion, we suggest a VWAP option pricing model based on modelling the underlying volume as the standard gamma process - an assumption which is backed by our empirical analysis of volume statistics of a few ASX-traded stocks. Our model allows us to obtain simple closed-form formulas for implied volatility adjustments (with no forward adjustments needed) for Black-Scholes-style pricing formulas. These formulas are excellent approximations to the VWAP gamma model exact results, obtained by Monte Carlo simulations. In addition we have demonstrated that some rather conventional approximations (see Eqs. (\ref{Ratios})--(\ref{Ratios2})), which our model can avoid, should be used with a great deal of care since they may lead to substantial pricing errors.
It is also reasonably clear how one could further improve on our results. One should aim to derive a more sophisticated model for the volume process which should have a richer parameter space to better fit the empirical stock volume data, than the simple gamma distribution utilised in this work and/or take into account the partial correlation between volume and stock processes. Another possible direction is to obtain upper/lower bound results based on the gamma distributed volume model.
\section*{Appendix A} \label{AppendixA}
We model the volumes $V_i$ traded within $[t_{i-1},t_{i}]$ time interval (where $t_i=i\Delta t$) by i.i.d.\! gamma variables $V_i\sim\Gamma(\alpha,\theta)$, independent of the underlying stock process $S_t$. Note that the parameter $\alpha$ depends linearly on the choice of $\Delta t$.
The vector of scaled increments $X_i=V_i/\sum_{j=1}^N V_j$ are given by
\[
\left(X_1,\ldots,X_N\right)=\left(\frac{V_1}{\sum_{j=1}^N V_j},\ldots,\frac{V_N}{\sum_{j=1}^N V_j}\right)\sim D(\alpha, \ldots, \alpha),
\]
where $D$ is a Dirichlet distribution. The following formulas for various expectations are well-known:
\begin{gather*}
\E(X_i)=\frac{1}{N}=\frac{\Delta t}{T},\quad \E(X_i^2)=\frac{\alpha + 1}{N(\alpha N+1)},\quad \Var(X_i)=\frac{N-1}{N^2(\alpha N+1)},\\
\E(X_i X_j)=\frac{\alpha}{N(\alpha N+1)},\quad \Cov(X_i,X_j)=\frac{-1}{N^2(\alpha N+1)},\quad i\neq j.
\end{gather*}
\subsection*{Discrete-time case}
In the discrete-time case, the VWAP is defined to be
\[
S := S^{\rm(VWAP)} =\frac{\sum_{i=1}^N S_iV_i}{\sum_{i=1}^N V_i}= \sum_{i=1}^N S_i X_i
\]
where $S_i=S_{i\Delta t}$.
The first two moments can then be computed using Fubini's Theorem, Fubini (1958):
\begin{align*}
M_1&=\E S= \sum_{i=1}^N \E S_i \E X_i
=\E\left(\frac{1}{N}\sum_{i=1}^N S_i\right),\\
\E S^2 &=\sum_{i=1}^N \E S_i^2 \E X_i^2 + 2\sum_{i< j} \E (S_i S_j) \E (X_i X_j) \\
&= \frac{\alpha + 1}{N(\alpha N+1)} \sum_{i=1}^N \E S_i^2 + \frac{2\alpha}{N(\alpha N+1)} \sum_{i< j} \E (S_iS_j) \\
&= \E\left( \frac{1}{N}\sum_{i=1}^N S_i\right)^2 + \frac{1}{\alpha N+1}\left( \E\left(\frac{1}{N}\sum_{i=1}^N S_i^2\right)-\E\left( \frac{1}{N}\sum_{i=1}^N S_i\right)^2\right),\\
M_2&=\E S^2-M_1^2\\
&= \Var \left( \frac{1}{N}\sum_{i=1}^N S_i\right) + \frac{1}{\alpha N +1}\left( \E\left(\frac{1}{N}\sum_{i=1}^N S_i^2\right)-\E\left( \frac{1}{N}\sum_{i=1}^N S_i\right)^2\right).
\end{align*}
Note that instead of solving gamma process VWAP problem exactly as above, we can adopt the ``Stace-like'' approach as well, by utilising the approximations (\ref{Ratios})--(\ref{Ratios2}). In this case $M_1^{\rm (Stace)} = M_1$, but the final expression for $M_2$ changes into:
\begin{align*}
M_2^{\rm (Stace)} &= \Var \left( \frac{1}{N}\sum_{i=1}^N S_i\right) + \frac{1}{\alpha N}\left( \E\left(\frac{1}{N}\sum_{i=1}^N S_i^2\right)-\left[\E\left( \frac{1}{N}\sum_{i=1}^N S_i\right)\right]^2 \right).
\end{align*}
Now let us recall that $S_t$ is a geometric Brownian motion with drift $r$ and volatility $\sigma$, all terms in the expressions for $M_1$, $M_2$ and $M_2^{\rm (Stace)}$ can all be explicitly calculated:
\begin{align*}
\E\left(\frac{1}{N}\sum_{i=1}^N S_i\right)&=\frac{S_0}{N}\sum_{i=1}^N e^{r i\Delta t},\\
\E\left(\frac{1}{N}\sum_{i=1}^N S_i^2\right)&=\frac{S_0^2}{N}\sum_{i=1}^N e^{(2r+\sigma^2) i\Delta t},\\
\E\left(\frac{1}{N}\sum_{i=1}^N S_i\right)^2&=\frac{S_0^2}{N^2}\left(\sum_{i=1}^N e^{(2r+\sigma^2) i\Delta t}+2\sum_{j=1}^N\sum_{i=1}^{j-1}e^{r(i+j)\Delta t+\sigma^2 i\Delta t}\right).
\end{align*}
This, in turn, allows the calculation of the corresponding implied volatilities by moment matching. For example, for equidistant spacing of averaging points, the following expressions can be obtained:
\begin{align*}
\sigma^{\rm (VWAP)}_{\rm ExactD} = \sigma \sqrt{\frac{(1+N) (3 + \alpha + 2 \alpha N)}{6 N (1 + \alpha N)}+ O(T/N)},
\end{align*}
or (if taken relative to $\sigma^{(AA)}_D$):
\begin{align*}
\frac{\sigma^{\rm (VWAP)}_{\rm ExactD}}{\sigma^{(AA)}_D} = \sqrt{\frac{N (3 + \alpha + 2 \alpha N)}{(1+2 N) (1 + \alpha N)}+ O(T/N)} ,
\end{align*}
and
\begin{align*}
\sigma^{\rm (VWAP)}_{\rm StaceD} = \sigma \sqrt{\frac{(1+N) (3 + \alpha + 2 \alpha N)}{6 \alpha N^2}+ O(T/N)},
\end{align*}
or (if taken relative to $\sigma^{(AA)}_D$):
\begin{align*}
\frac{\sigma^{\rm (VWAP)}_{\rm StaceD}}{\sigma^{(AA)}_D} = \sqrt{\frac{(3 + \alpha + 2 \alpha N)}{(\alpha + 2 \alpha N)}+ O(T/N)},
\end{align*}
\subsection*{Continuous-time case}
The continuous-time VWAP is the limit of the discrete-time case as $N\rightarrow \infty$ or $\Delta t\rightarrow 0$:
\[
\tilde{S}:=S^{\rm(VWAP)}=\frac{\int_0^T S_t\, dZ_t}{Z_T}=\lim_{N\rightarrow \infty} \sum_{i=1}^N S_{(i-1)\Delta t} X_i.
\]
where \emph{the aggregate volume traded} $Z_t$ is a gamma process $\Gamma(t;\tilde \alpha, \theta)$.
Here we use the parameter $\tilde \alpha$ which relates to the discrete-time case parameter via $\tilde \alpha = \alpha / \Delta t$ or $\tilde \alpha T= \alpha N$. Using arguments similar to the discrete-time case, as well as the dominated convergence theorem, the first two exact moments are:
\begin{align*}
M_1&=\E\left(\frac{1}{T}\int_0^T S_{t}\, dt\right),\\
\E \tilde{S}^2 &= \E\left(\frac{1}{T}\int_0^T S_t\, dt\right)^2 + \frac{1}{\tilde{\alpha} T+1}\left(\E\left(\frac{1}{T}\int_0^T S_t^2\, dt\right)- \E\left(\frac{1}{T}\int_0^T S_t\, dt\right)^2\right),\\
M_2 &=\Var\left(\frac{1}{T}\int_0^T S_t\, dt\right) + \frac{1}{\tilde{\alpha} T+1}\left(\E\left(\frac{1}{T}\int_0^T S_t^2\, dt\right)- \E\left(\frac{1}{T}\int_0^T S_t\, dt\right)^2\right).
\end{align*}
And again we can use \eqref{Ratios}--\eqref{Ratios2} to obtain a Stace-like approximation,
\[
M_2^{\rm (Stace)}=\Var\left(\frac{1}{T}\int_0^T S_t\, dt\right) + \frac{1}{\tilde{\alpha} T}\left(\E\left(\frac{1}{T}\int_0^T S_t^2\, dt\right)- \left[\E\left(\frac{1}{T}\int_0^T S_t\, dt\right)\right]^2\right).
\]
The terms involving the stock process $S_t$ are given by:
\begin{align*}
\E\left(\frac{1}{T}\int_0^T S_t\, dt\right)&=\frac{S_0}{T}\left(\frac{e^{r T}-1}{r}\right),\\
\E\left(\frac{1}{T}\int_0^T S_t\, dt\right)^2&=\frac{2S_0^2}{(r+\sigma^2)T^2}\left(\frac{e^{(2r+\sigma^2)T}-1}{2r+\sigma^2}-\frac{e^{r T}-1}{r}\right),\\
\E\left(\frac{1}{T}\int_0^T S_t^2\, dt\right) &= \frac{S_0^2}{T}\left(\frac{e^{(2r+\sigma^2)T}-1}{2r+\sigma^2}\right).
\end{align*}
Finally the corresponding implied volatilities for the continuous limit can also be calculated:
\begin{align*}
\frac{\sigma^{\rm (VWAP)}_{\rm Exact}}{\sigma^{(AA)}} = \sqrt{\frac{(3 + 2 \tilde{\alpha})}{(2 + 2 \tilde{\alpha})}+ \frac{(6 (1 + \tilde{\alpha}) r^2 + 3 (1 + \tilde{\alpha}) r \sigma^2 +
\tilde{\alpha} \sigma^4) T}{24 (1 + \tilde{\alpha})^2 \sigma^2} + O(T^2)} ,
\end{align*}
and
\begin{align*}
\frac{\sigma^{\rm (VWAP)}_{\rm Stace}}{\sigma^{(AA)}} = \sqrt{\frac{(3 + 2 \tilde{\alpha})}{2 \tilde{\alpha}}+ \frac{(2 \tilde{\alpha} r^2 + \tilde{\alpha} r \sigma^2 - 3 \sigma^4 -
\tilde{\alpha} \sigma^4) T}{8 \tilde{\alpha}^2 \sigma^2} + O(T^2)}.
\end{align*}
\subsection*{Acknowledgment}
The authors are grateful to A. Barber, A. Brace, T. Glass, V. Frishling, D. Maher, T. Ling, A. Novikov and W. Wright for fruitful discussions and useful suggestions.
|
\section{Introduction}
\begin{table}[t]
\caption{State-of-the-art datasets for action recognition based on depth or skeletal features, sorted from more quoted to less quoted according to Google Scholar.}
\centering
\begin{tabu} to \linewidth{|X[3.4,l]|X[0.9,c]|X[0.9,c]|X[0.9,c]|X[1,c]|X[1,c]|X[0.8,c]|}
\hline
\textbf{\scriptsize Name} & \textbf{\scriptsize Actions} & \textbf{\scriptsize Actors} & \textbf{\scriptsize Times} & \textbf{\scriptsize Samples} & \textbf{\scriptsize Citations} & \textbf{\scriptsize Year} \\
\hline
MSR Action3D~\cite{Li2010} & 20 & 10 & 2 or 3 & 567 & 176 & 2010 \\
\hline
MSR DailyActivity3D~\cite{Wang2012} & 16 & 10 & 2 & 320 & 138 & 2012 \\
\hline
RGBD-HuDaAct~\cite{Ni2011} & 12 & 30 & 2 or 4 & 1189 & 86 & 2011 \\
\hline
CAD-60~\cite{Jaeyong2012} & 12 & 2+2 & - & 60 & 80 & 2012 \\
\hline
UTKinect Action~\cite{Xia2012} & 10 & 10 & 2 & - & 73 & 2012 \\
\hline
MSRC-12 KinectGesture~\cite{Fothergill2012} & 12 & 30 & - & 594 & 39 & 2012 \\
\hline
CAD-120~\cite{Koppula2013} & 10 & 2+2 & - & 120 & 33 & 2013 \\
\hline
MSR ActionPairs~\cite{Oreifej2013} & 6 & 10 & 3 & 180 & 29 & 2013 \\
\hline
MSR Gesture3D~\cite{Kurakin2012} & 12 & 10 & 2 or 3 & 336 & 25 & 2012 \\
\hline
LIRIS Human Activities~\cite{Wolf2012} & 10 & 21 & - & - & 24 & 2012 \\
\hline
Berkeley MHAD~\cite{Ofli2013} & 11 & 7+5 & 5 & $\sim660$ & 18 & 2013 \\
\hline
G3D~\cite{Bloom2013} & 20 & 10 & 3 & - & 11 & 2012 \\
\hline
ACT4 Dataset~\cite{Cheng2012} & 14 & 24 & \textgreater1 & 6844 & 9 & 2012 \\
\hline
UPCV Action~\cite{Theodorakopoulos2014} & 10 & 20 & - & - & 6 & 2014 \\
\hline
WorkoutSu-10 Gesture~\cite{Negin2013} & 10 & 15 & 10 & 1500 & 6 & 2013 \\
\hline
IAS-Lab Action~\cite{Munaro2013} & 15 & 12 & 3 & 540 & 3 & 2013 \\
\hline
Florence 3D Action~\cite{Seidenari2013} & 9 & 10 & 2 or 3 & 215 & 2 & 2012 \\
\hline
\end{tabu}
\label{tab:datasets}
\end{table}
In recent years, interest has grown on affordable devices (\eg \emph{Microsoft Kinect} or \emph{ASUS Xtion Pro}) that capture depth quite reliably. Such devices provide a depth image (D), along with an RGB image (thus RGB-D). A depth image can be further processed to obtain marker-less body pose estimation by means of a skeleton model consisting of a series of joints. Due to their low cost, high sample rate and capability to combine visual and depth information, these devices have become widespread in both research and commercial applications. Furthermore, their use has not been restricted to games, for which they were initially designed, but other applications where natural human-computer interaction is required.
These devices are widely used in the field of human action recognition (HAR), particularly in indoor scenarios for the recognition of activities of daily living. For research purposes, a variety of datasets for human action (or gesture) recognition have been recorded using RGB-D devices (see Table~\ref{tab:datasets}). The MSR Action3D dataset~\cite{Li2010} from Microsoft Research stands out as one of the most used in the literature, as many developed methods for action recognition have been validated with this dataset. Hence, it should be easy to determine the best human action recognition method in a straightforward way by comparing their success and processing rates. However, to the best of our knowledge, this is not possible at the moment as we found that almost all the works compare results obtained with different validation methods.
Therefore, this work aims to fill the existing gap in order to enable a fair comparison of the state of the art. We have reviewed 176 papers that make reference to the MSR Action3D dataset. Out of these 176 papers, 62 papers have been considered as they use the MSR Action3D dataset for the validation of the human action (or gesture) recognition methods proposed. They are classified according to the validation method and ranked based on their success rate.
The remainder of this paper is organised as follows: Section~\ref{sec:msr-action3d} describes the MSR Action3D dataset employed by the reviewed works. In section~\ref{sec:instances}, an explanation of the inconsistencies found in the number of the used samples is given. Section~\ref{sec:validation-methods} presents the validation methods used in the reviewed papers and provides a classification of each work according to this. Finally, section~\ref{sec:conclusion} presents some conclusions and recommendations for the future.
\section{MSR Action3D dataset}
\label{sec:msr-action3d}
The MSR Action3D dataset \cite{Li2010} contains 20 different actions, performed by 10 different subjects with up to 3 different repetitions. This makes a total of 567 sequences and each one includes depth and skeleton joints. However 10 sequences are not valid in this dataset because the skeletons were either missing or wrong, as explained by the authors\footnote{MSR Action Recognition Datasets and Codes, http://research.microsoft.com/en-us/um/people/zliu/actionrecorsrc/default.htm (last access: 06/26/2014)}. The authors divided the dataset in three subsets of 8 gestures each, as shown in Table~\ref{tab:subsets}. Most of the papers working with this dataset have also used them. This was due to the high computational cost of dealing with the overall dataset. The AS1 and AS2 subsets were intended to group actions with similar movement, while AS3 was intended to group complex actions together.
\begin{table}[t]
\caption{Actions in each of the MSR Action3D subsets.}
\centering
\begin{tabular}{|c | l | c | l | c | l |}
\hline
\multicolumn{2}{|c|}{\textbf{AS1}} & \multicolumn{2}{c}{\textbf{AS2}} & \multicolumn{2}{|c|}{\textbf{AS3}}\\
\hline
Label & Action name & Label & Action name & Label & Action name\\% inserts table
\hline
a02 & Horizontal arm wave & a01 & High arm wave & a06 & High throw \\
a03 & Hammer & a04 & Hand catch & a14 & Forward kick \\
a05 & Forward punch & a07 & Draw cross & a15 & Side-kick \\
a06 & High throw & a08 & Draw tick & a16 & Jogging \\
a10 & Hand clap & a09 & Draw circle & a17 & Tennis swing \\
a13 & Bend & a11 & Two-hand wave & a18 & Tennis serve \\
a18 & Tennis serve & a14 & Forward kick & a19 & Golf swing \\
a20 & Pick-up and throw & a12 & Side-boxing & a20 & Pick-up and throw \\
\hline
\end{tabular}
\label{tab:subsets}
\end{table}
\section{How many samples are used for testing?}
\label{sec:instances}
Despite of the fact that the MSR Action3D dataset is made up of 567 sequences, the number of instances used in some works is unclear~\cite{Azary2013,Gao2014,Shen2014}. There is a lot of confusion concerning this topic.
As far as we know, the authors of the dataset firstly described it as made up of twenty actions, where each one was performed by seven subjects for three times~\cite{Li2010}. However, actions are performed by ten subjects with up to three repetitions as described in the previous section. Many works have compared their results with Li et al. and most of them used ten subjects~\cite{Xia2012,Wang2013,Yang2014}. In other words, they may have used a higher number of instances than the work they aim to compare to. Wang et al.~\cite{Wang2012} described the dataset as made up of 402 sequences. For the sake of clarity, this mistake is advertised at the dataset web page\footnote{A list of the used sequences is also provided in the website}. The authors explain that 10 sequences out of the 567 are not used because a number of skeletons are either missing or too erroneous. So, the dataset is eventually composed of 557 sequences. However, it is curious to see how recent works~\cite{Gao2014,Shen2014,Yuan2014} still mention that the dataset is composed of 402 sequences and directly compare their results with the state-of-the-art papers that use other number of instances. Furthermore, other authors have intentionally used a subset of the whole dataset, \eg 17 actions, 8 subjects and 3 repetitions (408 samples). Due to this, the AS1, AS2 and AS3 subsets are composed of different actions too, thereby they compare their results with works that use a different number of instances.
As a consequence, it is very difficult to confirm whether these works use 402, 557 or 567 samples as we are not sure whether the authors are aware of these key aspects concerning the dataset, or if those are only naive text mistakes. Moreover, the missing information concerning the number of instances prevents to make a fair comparison between different methods.
\section{Which is the validation method used?}
\label{sec:validation-methods}
Regarding the experimentation method used by many authors working with the MSR Action3D dataset, it is worth to mention that there is a lack of agreement. In the paper by Li et al.~\cite{Li2010} where the dataset was firstly presented, three tests are performed: 1/3, 2/3 and cross-subject test. In the first two tests, 1/3 and 2/3 of the instances are respectively used as training samples and the rest as testing samples. In the third test, half of the subjects are used for training and the remainder for testing. However, it is not described which instances or subjects are actually used in each partition of the dataset.
Given that information is missing, we could assume that the 1/3 means to split the dataset using the first repetition of each action performed by each subject as training, and to use the remainder for testing. The same could be assumed for the 2/3. However, if we only consider instances as a whole, we can split the dataset in a different way. For instance, the dataset can be split using 1/3 (or 2/3) of all the instances for training. The same is true for the cross-subject test. It is not stated which instances are used. Any half of all the subjects can be used for training, \eg 1, 2, 3, 9 and 10; and the remainder for testing, \latinphrase{i.e.}\xspace 4, 5, 6, 7 and 8. Given that it is not clear which instances are used, each researcher is free to interpret anything, thereby comparing different methods where a distinct methodology has been used for the experimentation. However, this is not desirable to compare and decide which method performs better.
In the cross-subject test employed by Li et al.~\cite{Li2010} the actual samples of subjects 1, 3, 5, 7 and 9 are used for training, whereas actors 2, 4, 6, 8 and 10 are used for validation. This test is followed by many authors as shown in Table~\ref{tab:li-cross-subject-test}. While some authors use the mentioned settings for their training and validation sets, other authors use subjects 1-5 for training and 6-10 for validation (see Table~\ref{tab:1-5-train-6-10-test}). Regardless of the used setup, most of the works state that they follow the same settings as Li et al. but do not provide a description of such a setup. Due to this, we assume that they follow the same validation than Li et al., so Table~\ref{tab:li-cross-subject-test} and Table~\ref{tab:1-5-train-6-10-test} can even have classification mistakes. Anyway, a fair comparison cannot be performed. Indeed, when it is sure that the same setup has been used, sometimes results only show an accuracy score and the authors do not give an explanation of what it represents, \latinphrase{i.e.}\xspace the average of the AS1, AS2 and AS3 tests, or the overall accuracy of using the whole dataset (20 actions).
\tabulinesep = 0.9mm
\begin{longtabu} to \linewidth{|>{\scriptsize}X[5.7,m]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|}
\caption{Li et al.'s cross-subject test (1-3-5-7-9 training, 2-4-6-8-10 test). The first 12 methods explicitly describe both the training and validation sets. Results are ordered by the average result for the AS1, AS2, and AS3 subsets; and then by the results for the whole dataset.}
\label{tab:li-cross-subject-test} \\
\hline
\textbf{Method} & \textbf{Year} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
\endfirsthead
\multicolumn{7}{c}%
{\tablename\ \thetable\ -- \textit{Continued from previous page}} \\
\hline
\textbf{Method} & \textbf{Year} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
\endhead
\hline
\multicolumn{7}{c}{\textit{Continued on next page}} \\
\endfoot
\hline
\endlastfoot
Deep Convolutional Neural Networks for Action Recognition Using Depth Map Sequences, Wang et al. \cite{Wang2015} & 2015 & - & - & - & - & ¿100? \\
\hline
Efficient Pose-Based Action Recognition, Eweiwi et al. \cite{Eweiwi2015} & 2015 & - & - & - & - & 92.3? \\
\hline
Action Recognition from Depth Sequences Using Depth Motion Maps-based Local Binary Patterns,
Chen et al. \cite{Chen2015} & 2015 & 98.1 & 92 & 94.6 & 94.9 & 91.94 \\
\hline
HOPC: Histogram of Oriented Principal Components of 3D Pointclouds for Action Recognition, Rahmani et al. \cite{Rahmani2014eccv} & 2014 & - & - & - & - & 91.64? \\
\hline
Fusion of Skeletal and Silhouette-Based Features for Human Action Recognition with RGB-D Devices, Chaaraoui et al. \cite{Chaaraoui2013} & 2013 & 92.38 & 86.61 & 96.4 & 91.8 & - \\
\hline
Real-time human action recognition based on depth motion maps, Chen et al. \cite{Chen2013} & 2013 & 96.2 & 83.2 & 92 & 90.47 & - \\
\hline
Skeletal Quads: Human Action Recognition Using Joint Quadruples, Evangelidis et al. \cite{Evangelidis2014} & 2014
& 88.39 & 86.61 & 94.59 & 89.86 & - \\
\hline
Random Occupancy Patterns, Wang et al. \cite{Wang2014}
& 2014 & - & - & - & 86.50? & - \\
\hline
Action recognition based on a bag of 3d points, Li et al. \cite{Li2010}
& 2010 & 72.9 & 71.9 & 79.2 & 74.67 & - \\
\hline
Learning Maximum Margin Temporal Warping for Action Recognition, Wang and Wu \cite{Wang2013}
& 2013 & - & - & - & - & 92.7? \\
\hline
Learning Actionlet Ensemble for 3D Human Action Recognition, Yuan et al. \cite{Yuan2014}
& 2014 & - & - & - & - & 88.2? \\
\hline
Mining actionlet ensemble for action recognition with depth cameras, Wang et al. \cite{Wang2012}
& 2012 & - & - & - & - & 88.2? \\
\hline
\hline
\hline
Group Sparsity and Geometry Constrained Dictionary Learning for Action Recognition from Depth Maps, Luo et al. \cite{Luo2013}
& 2013 & 97.2 & 95.5 & 99.1 & 97.26 & 96.7 \\
\hline
Fusing Spatiotemporal Features and Joints for 3D Action Recognition, Zhu et al. \cite{Zhu2013}
& 2013 & - & - & - & 94.3 & - \\
\hline
Human Action Recognition by Mining Discriminative Segment with Novel Skeleton Joint Feature, Zou et al. \cite{Zou2013}
& 2013 & - & - & - & 94.0? & \\
\hline
Pose-based human action recognition via sparse representation in dissimilarity space , Theodorakopoulos et al. \cite{Theodorakopoulos2014}
& 2014 & 91.23 & 90.09 & 99.5 & 93.61 & - \\
\hline
Action recognition on motion capture data using a dynemes and forward differences representation, Kapsouras and Nikolaidis \cite{Kapsouras2014} & 2014
& - & - & - & 93.6 & 91.4 \\
\hline
Super Normal Vector for Activity Recognition Using Depth Sequences, Yang and Tian \cite{Yang2014b} & 2014
& - & - & - & 93.09? & - \\
\hline
Action Recognition Using Ensemble Weighted Multi-Instance Learning, Chen et al. \cite{Chen2014} & 2014
& - & - & - & 92? & - \\
\hline
Histogram of Oriented Displacements (HOD): Describing Trajectories of Human Joints for Action Recognition, Gowayyed et al. \cite{Gowayyed2013} & 2013 & 92.39 & 90.18 & 91.43 & 91.26 & - \\
\hline
DMM-Pyramid Based Deep Architectures for Action Recognition with Depth Cameras, Yand and Yang \cite{YangYang2015} & 2015 & - & - & - & 91.21? & - \\
\hline
Human Action Recognition Using a Temporal Hierarchy of Covariance Descriptors on 3D Joint Locations, Hussein et al. \cite{Hussein2013} & 2013
& 88.04 & 89.29 & 94.29 & 90.53 & - \\
\hline
Body Surface Context: A New Robust Feature for Action Recognition From Depth Videos, Song et al. \cite{Song2014} & 2014
& - & - & - & 90.36? & - \\
\hline
An Approach to Pose-Based Action Recognition, Wang et al. \cite{Wang2013b} & 2013
& - & - & - & 90.22 & - \\
\hline
Human Action Recognition Via Multi-modality Information, Gao et al. \cite{Gao2014} & 2014
& 92 & 85 & 93 & 90 & - \\
\hline
Human Behavior Recognition Based on Axonometric Projections and PHOG Feature, Shen et al. \cite{Shen2014} & 2014 & 90.6 & 81.4 & 94.6 & 88.87 & - \\
\hline
On the improvement of human action recognition from depth map sequences using Space–Time Occupancy Patterns, Vieira et al. \cite{Vieira2014} & 2014 & 91.7 & 72.2 & 98.6 & 87.5 & 81.55 \\
\hline
STOP: Space-Time Occupancy Patterns for 3D Action Recognition from Depth Map Sequences, Vieira et al. \cite{Vieira2012} & 2012 & 84.7 & 81.3 & 88.4 & 84.8 & - \\
\hline
Effective 3D action recognition using EigenJoints, Yang And Tian \cite{Yang2014} & 2014
& - & - & - & 83.3? & - \\
\hline
Home Monitoring Musculo-skeletal Disorders with a Single 3D Sensor, Wang et al. \cite{Wang2013c} & 2013 & - & - & - & 81.9? & - \\
\hline
Online Human Gesture Recognition from Motion Data Streams, Zhao et al. \cite{Zhao2013} & 2013 & - & - & - & 81.7? & - \\
\hline
Effective approaches in human action recognition, Li et al. \cite{Li2013} & 2013
& - & - & - & 81.5 or 91.5? & - \\
\hline
Gesture recognition from depth images using motion and shape features, Qin et al. \cite{Qin2013} & 2013 & 81 & 79 & 82 & 80.66 & - \\
\hline
Human activity recognition using multi-features and multiple kernel learning, Althloothi et al. \cite{Althloothi2014} & 2014 & 74.3 & 76.8 & 86.7 & 79.27 & - \\
\hline
View invariant human action recognition using histograms of 3D joints, Xia et al. \cite{Xia2012} & 2012 & 87.98 & 85.48 & 63.46 & 78.97 & - \\
\hline
Three Dimensional Motion Trail Model for Gesture Recognition, Liang and Zheng \cite{Liang2013} & 2013 & 73.7 & 81.5 & 81.6 & 78.93 & -\\
\hline
Attractor-Shape for Dynamical Analysis of Human Movement: Applications in Stroke Rehabilitation and Action Recognition, Venkataraman et al. \cite{Venkataraman2013} & 2013 & 77.5 & 63.1 & 87 & 75.87 & - \\
\hline
Exploring the Trade-off Between Accuracy and Observational Latency in Action Recognition, Ellis et al. \cite{Ellis2013} & 2013 & - & - & - & 65.7? & - \\
\hline
Robust 3D Action Recognition with Random Occupancy Patterns, Wang et al. \cite{Wang2012b} & 2012 & - & - & - & - & 86.5? \\
\hline
\end{longtabu}
\normalsize
\begin{longtabu} to \linewidth{|>{\scriptsize}X[5.7,m]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|}
\caption{Cross-subject test (1-5 training, 6-10 test). The first seven methods explicitly describe both the training and validation sets.}
\label{tab:1-5-train-6-10-test} \\
\hline
\textbf{Method} & \textbf{Year} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
\endfirsthead
\multicolumn{7}{c}%
{\tablename\ \thetable\ -- \textit{Continued from previous page}} \\
\hline
\textbf{Method} & \textbf{Year} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
\endhead
\hline
\multicolumn{7}{c}{\textit{Continued on next page}} \\
\endfoot
\hline
\endlastfoot
Sparse spatio-temporal representation of joint shape-motion cues for human action recognition in depth sequences, Tran and Ly \cite{Tran2013b} & 2013
& - & -& - & 91.92? & - \\
\hline
The Moving Pose: An Efficient 3D Kinematics Descriptor for Low-Latency Action Recognition and Detection, Zanfir et al. \cite{Zanfir2013} & 2013
& - & - & - & 91.7? & - \\
\hline
An effective fusion scheme of spatio-temporal features for human action recognition in RGB-D video, Tran and Ly \cite{Tran2013} & 2013
& - & - & - & 88.89? & - \\
\hline
Real Time Action Recognition Using Histograms of Depth Gradients and Random Decision Forests, Rahmani et al. \cite{Rahmani2014wacv} & 2014
& - & - & - & 88.8? & - \\
\hline
Iterative temporal learning and prediction with the sparse online echo state gaussian process, Soh and Demiris \cite{Soh2012} & 2012
& 80.6 & 74.9 & 87.1 & 80.87 & - \\
\hline
Joint Angles Similarities and HOG2 for Action Recognition, Ohn-Bar and Trivedi \cite{Ohn-Bar2013} & 2013
& - & - & - & - & 94.84 \\
\hline
HON4D: Histogram of Oriented 4D Normals for Activity Recognition from Depth Sequences, Oreifej and Liu \cite{Oreifej2013} & 2013
& - & - & - & - & 88.89?\\
\hline
\hline
\hline
Spatio-temporal feature extraction and representation for RGB-D human action recognition, Luo et al. \cite{Luo2014} & 2014 & 96.1 & 90.8 & 98.33 & 95.08 & 93.83?\\
\hline
Action Classification with Locality-constrained Linear Coding, Rahmani et al. \cite{Rahmani2014icpr} & 2014 & - & - & - & 90.9? & - \\
\hline
Spatio-temporal Depth Cuboid Similarity Feature for Activity Recognition Using Depth Camera, Xia and Aggarwal \cite{Xia2013} & 2013 & - & - & - & 89.3? & - \\
\hline
Optimal Joint Selection for Skeletal Data from RGB-D Devices Using a Genetic Algorithm, Climent et al. \cite{Climent2013} & 2013 & - & - & - & - & 71.1 \\
\hline
\end{longtabu}
Due to all this confusion about how to split the dataset in two sets for training and validation, some authors randomly choose half of the subjects for the training set, and select the rest of the subjects for the validation set. As in 2-fold cross validation, they repeat the test using the previous validation set as the training set and vice versa. In this case, the final result is the average of both tests (see Table~\ref{tab:2-fold-cross-validation}). In other works, instead of performing a 2-fold cross validation, some authors randomly select the two sets and repeat the experiment several times. For example, Miranda et al.~\cite{Miranda2012} perform a random selection of half of the actors as training set and the other half as validation set. This is repeated 10 times and the final result is the average of the results of each run. Other authors repeat the test 100 times instead of 10~\cite{Çeliktutan2013,Çeliktutan2014}, and even 200 times~\cite{Chen2013} (see Table~\ref{tab:miranda-test}). However, although the tests are repeated many times, all the possible splits are not considered, \latinphrase{i.e.}\xspace all the possible combinations (252 tests) of using 5 subjects for training and the remaining ones for testing. Only a few works perform this test. In Table~\ref{tab:miranda-test} these works have been included with the 252 number in the third column. This indicates that they performed the test with all the possible combinations.
\tabulinesep = 0.6mm
\begin{table}
\caption{2-fold cross-validation test}
\begin{tabu} to \linewidth{|>{\scriptsize}X[5.7,m]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|}
\hline
\textbf{Method} & \textbf{Year} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
Evolutionary joint selection to improve human action recognition with RGB-D devices, Chaaraoui et al. \cite{Chaaraoui2014} & 2014 & 91.59 & 90.83 & 97.28 & 93.23 & - \\
\hline
Combining unsupervised learning and discrimination for 3D action recognition, Chen et al. \cite{ChenG2015} & 2015 & - & - & - & 91,4 & - \\
\hline
3D Action Classification Using Sparse Spatio-temporal Feature Representations, Azary and Savakis \cite{Azary2012} & 2012 & 77.66 & 73.17 & 91.58 & 80.8 & 63.23 \\
\hline
\end{tabu}
\label{tab:2-fold-cross-validation}
\end{table}
\begin{table}
\caption{Miranda et al.'s test: Random selection of training and test sets repeated a number of times.}
\begin{tabu} to \linewidth{|>{\scriptsize}X[5.7,m]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|}
\hline
\textbf{Method} & \textbf{Year} & \textbf{Tests} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
Efficient Pose-Based Action Recognition, Eweiwi et al. \cite{Eweiwi2015} & 2015 & 252 & - & - & - & - & 88.38? \\
\hline
HOPC: Histogram of Oriented Principal Components of 3D Pointclouds for Action Recognition, Rahmani et al. \cite{Rahmani2014eccv} & 2014 & 252& - & - & - & - & 86.49? \\
\hline
Sparse spatio-temporal representation of joint shape-motion cues for human action recognition in depth sequences, Tran and Ly \cite{Tran2013b} & 2013 & 252
& - & - & - & 84.54? & -\\
\hline
Real Time Action Recognition Using Histograms of Depth Gradients and Random Decision Forests, Rahmani et al. \cite{Rahmani2014wacv} & 2014 & 252
& - & - & - & - & 82.7? \\
\hline
HON4D: Histogram of Oriented 4D Normals for Activity Recognition from Depth Sequences, Oreifej and Liu \cite{Oreifej2013} & 2013 & 252
& - & - & - & - & 82.15? \\
\hline
Real-time human action recognition based on depth motion maps, Chen et al. \cite{Chen2013} & 2013 & 200 & 90.1 & 90.6 & 97.6 & 92.77 & - \\
\hline
Fast Exact Hyper-graph Matching with Dynamic Programming for Spatio-temporal Data, \c{C}eliktutan et al. \cite{Çeliktutan2014} & 2014 & 100 & 84.5 & 85 & 72.2 & 80.57 & - \\
\hline
Graph-based Analysis of Physical Exercise Actions, Çeliktutan et al. \cite{Çeliktutan2013} & 2013 & 100 & 84.5 & 85 & 72.2 & 80.5 & - \\
\hline
Informative joints based human action recognition using skeleton contexts, Jiang et al. \cite{Jiang2015} & 2015 & 10 & 88.7 & 87.7 & 88.5 & 88.3 & - \\
\hline
Online gesture recognition from pose kernel learning and decision forests, Miranda et al. \cite{Miranda2014} & 2014 & 10 & 96 & 57.1 & 97.3 & 83.5 & - \\
\hline
Real-Time Gesture Recognition from Depth Data through Key Poses Learning and Decision Forests, Miranda et al. \cite{Miranda2012} & 2012 & 10 & 93.5 & 52 & 95.4 & 80.3 & - \\
\hline
Space-Time Pose Representation for 3D Human Action Recognition, Devanne et al. \cite{Devanne2013} & 2013 & 10 & 84.8 & 67.8 & 87.1 & 79.9 & - \\
\hline
\end{tabu}
\label{tab:miranda-test}
\end{table}
Another approach used by some authors is to perform a leave-one-actor-out cross-validation test. In this case, actor invariance is specifically tested by training with all but one actor, and testing the method with the unseen one. This is repeated for all the actors, averaging the returned success rates (see Table~\ref{tab:leave-one-actor-out-test}).
\begin{table}
\caption{Leave-one-actor-out cross-validation test}
\begin{tabu} to \linewidth{|>{\scriptsize}X[5.7,m]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|>{\scriptsize}X[1,c]|}
\hline
\textbf{Method} & \textbf{Year} & \textbf{AS1} & \textbf{AS2} & \textbf{AS3} & \textbf{Avg.} & \textbf{All} \\
\hline
Evolutionary joint selection to improve human action recognition with RGB-D devices, Chaaraoui et al. \cite{Chaaraoui2014} & 2014 & 91.46 & 91.78 & 97.13 & 93.46 & - \\
\hline
Fusion of Skeletal and Silhouette-Based Features for Human Action Recognition with RGB-D Devices, Chaaraoui et al. \cite{Chaaraoui2013} & 2013 & 90.65 & 85.15 & 95.93 & 90.58 & - \\
\hline
3D Action Classification Using Sparse Spatio-temporal Feature Representations, Azary and Savakis \cite{Azary2012} & 2012 & 80.73 & 77.11 & 93.89 & 83.91 & 72.11 \\
\hline
Grassmannian Sparse Representations and Motion Depth Surfaces for 3D Action Recognition, Azary and Savakis \cite{Azary2013} & 2013 & - & - & - & - & 78.48?\\
\hline
Fast Exact Hyper-graph Matching with Dynamic Programming for Spatio-temporal Data, \c{C}eliktutan et al. \cite{Çeliktutan2014} & 2014 & - & - & - & - & 72.9?\\
\hline
\end{tabu}
\label{tab:leave-one-actor-out-test}
\end{table}
Finally, in addition to the described validation methods that are frequently used in the literature, there are other authors that have not been included in any table because either the validation method is unclear~\cite{Barnachon2014,Lillo2014} or the employed settings are not used by more than one author~\cite{Ofli2012,Cottone2013,Chaudhry2013,Sabinas2013}. For instance, Ofli et al.~\cite{Ofli2012} use a subset of 17 actions and 8 subjects. They train with 5 subjects and validate with 3 subjects in order to obtain the success rate (41.18\% for the whole dataset). These results are improved in~\cite{Chaudhry2013} with the same setup (83.89\% for the whole dataset). Cottone et al.~\cite{Cottone2013} perform a leave-one-sequence-out cross validation, training with all the sequences in the dataset but one that is used for testing. Then, they perform 10 of these tests obtaining the average success rate (90.47\% for the average of AS1, AS2 and AS3). Sabinas et al.\cite{Sabinas2013} are focused on early detection of gestures, \latinphrase{i.e.}\xspace without seeing all the information, and their experimentation is based on one-shot learning. Therefore, their results are not directly comparable (47\% for the average of AS1, AS2 and AS3).
\section{Conclusions}
\label{sec:conclusion}
In this work, we have aimed to give an answer to the question of which is the best action recognition method based on features extracted from depth and skeletal data. Based on the review the present work has performed, it can be observed that we cannot answer this question with confidence. In other words, we cannot know so far. Hence, we have presented the most important divergences in the comparison of action recognition methods that use the MSR Action3D dataset. Among these, we can highlight the mismatch in the number of samples used by most of the works and the different validation methods that have been used. As we have seen, the validation performed by Li et al. is one of the most used. However, the missing information about how to split the dataset into training and validation sets has led to a lot of confusion. Furthermore, most of the authors do not describe how this division is performed in their works. Therefore, experiments cannot be reproduced and fair comparisons cannot be made. Thus, in this work we have tried to clear up the existing confusion. This may enable to improve future comparisons and increase the awareness of the need of clarifying experimental settings.
Among all the validation methods reviewed in this work, we consider that the cross validation considering all the possible splits of the dataset, \latinphrase{i.e.}\xspace all the possible combinations (252 tests) of using 5 subjects for training and the remaining ones for testing, is the most robust validation method. However, if testing your method with the 5-5 splits cross validation is very demanding concerning computational cost, then the leave-one-actor-out cross validation is the one we recommend under these conditions.
\section*{Notes to authors}
As it has been difficult in some cases to understand the validation method of the papers, we encourage authors of the reviewed works to contact us in case their works had been misclassified in the previous tables. This way, we will be able to update the document and correct it. Similarly, authors of new works are also encouraged to contact us in order to incorporate their works if so desired.
\bibliographystyle{splncs}
|
\section{Introduction}
In attempting to construct a theory beyond the standard model and a theory of inflation \cite{Guth, Linde1},
it is desirable for us to be guided by some fundamental principles and symmetries such as
the equivalence principle and the general coordinate invariance in general relativity.
As one of such fundamental symmetries, we often encounter a scale or conformal symmetry
in particle physics and cosmology.
For instance, on the microscopic scale of particle physics, it is well known that the standard model
has a global scale symmetry if the (negative) Higgs mass term is neglected. This fact might suggest that
the underlying symmetry behind the standard model would be a global or local scale symmetry and the Higgs
mass term could emerge as the result of the spontaneous symmetry breakdown of the scale symmetry \cite{Bardeen}.
In addition, even in cosmology, the scale symmetry plays a critical role. For instance, the cosmic mocrowave background
radiation has almost scale-invariant fluctuations \cite{Planck} so it
seems to be natural at least for the present authors to conjecture that the universe was
controlled by the scale symmetry around the beginning of its creation. Together with these two
examples from particle physics and cosmology, it seems to be of interest to pursue an idea such that a theory
beyond the standard model is not only scale-invariant but also couples to a scale-invariant gravity
in order to account for inflation scenario.
The inflaton is a unknown scalar field which triggers the inflationary expansion and must be very weakly
interacting in order to explain the observed isotropy of the microwave background radiation. A natural question
to be asked frequently is whether the inflaton could be identified with recently discovered Higgs particle \cite{ATLAS, CMS}.
The recent revival of interest to this idea, dubbed Higgs inflation, is based on the observation that
when the simple chaotic inflation with the potential $\lambda \phi^4$ is combined with the non-minimal
gravitational coupling $\xi \phi^2 R$ in the Jordan frame, the inflation model can explain the observational data
if $\lambda = {\cal O}(1)$ and $\xi \approx 10^5$ \cite{Bezrukov}.
The non-minimal gravitational coupling $\xi \phi^2 R$ is invariant under a global scale transformation \cite{Oda1, Oda2, Oda3},
so there might exist a possibility of generalizing the Higgs inflation to a locally scale-invariant gravitational
theory. Indeed, in recent years, a locally scale-invariant class of inflation models has been established and
discussed from various viewpoints in \cite{Kallosh1}-\cite{Costa}, but a detailed analysis has not yet done for
the standard model embeded in a locally scale-invariant gravity.
In this article, we construct a theory of the standard model coupled to the Weyl-invariant gravity. In particular,
we discuss the Higgs mechanism in this theory and explore the question of whether this theory could explain the recent remarkable
results of the BICEP2 \cite{BICEP2}.
The structure of this article is the following: In Section 2, we present a simple model which accomodates
a local scale symmetry and the abelian gauge symmetry and explain the formalism in the Einstein
gauge and the Jordan gauge. In Section 3, we comment on an $SO(1,1)$ accidental symmetry and a shift symmetry.
In Section 4, we discuss the Higgs mechanism and derive the value of mass of the gauge field.
In Section 5, the abelian gauge model constructed in Section 2 is generalized to a model with the non-abelian
gauge field.
In Section 6, we argue inflation derived from the abelian gauge model and show that it is difficult to account
for the BICEP2 results.
We conclude in Section 7.
\section{An abelian gauge model}
Let us start with an abelian model where a $U(1)$ gauge field $A_\mu$ couples to a Weyl-invariant
gravity of two scalar fields, one of which, $\phi$, is real while the other, $H$ which is nothing but the
Higgs doublet field, is complex. In this article, we will ignore fermions completely since their existence
does not change our conclusion at all.
With a background curved metric $g_{\mu\nu}$, the Lagrangian takes the form\footnote{We follow
notation and conventions by Misner et al.'s textbook \cite{MTW}, for instance, the flat Minkowski metric
$\eta_{\mu\nu} = diag(-, +, +, +)$, the Riemann curvature tensor $R^\mu \ _{\nu\alpha\beta} =
\partial_\alpha \Gamma^\mu_{\nu\beta} - \partial_\beta \Gamma^\mu_{\nu\alpha} + \Gamma^\mu_{\sigma\alpha}
\Gamma^\sigma_{\nu\beta} - \Gamma^\mu_{\sigma\beta} \Gamma^\sigma_{\nu\alpha}$,
and the Ricci tensor $R_{\mu\nu} = R^\alpha \ _{\mu\alpha\nu}$.
The reduced Planck mass is defined as $M_p = \sqrt{\frac{c \hbar}{8 \pi G}} = 2.4 \times 10^{18} GeV$.
Through this article, we adopt the reduced Planck units where we set $c = \hbar = M_p = 1$.
In this units, all quantities become dimensionless.
Finally, note that in the reduced Planck units, the Einstein-Hilbert Lagrangian density takes the form
${\cal L}_{EH} = \frac{1}{2} \sqrt{-g} R$.}:
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} &=& \frac{1}{12} \left( \phi^2 - 2 |H|^2 \right) R + \frac{1}{2} g^{\mu\nu}
\partial_\mu \phi \partial_\nu \phi - g^{\mu\nu} (D_\mu H)^\dagger (D_\nu H) \nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho} F_{\nu\sigma} - V(\phi, H),
\label{Lagr 1}
\end{eqnarray}
where $\phi$ is a ghost, but this is not a problem because it can be removed by fixing the Weyl symmetry.
The covariant derivative and field strength in the abelian gauge group are respectively defined as
\begin{eqnarray}
D_\mu H = (\partial_\mu + i g A_\mu) H, \quad
(D_\mu H)^\dagger = H^\dagger (\overleftarrow{\partial}_\mu - i g A_\mu) , \quad
F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu,
\label{Def 1}
\end{eqnarray}
with $g$ being a $U(1)$ real coupling constant. (We use the same alphabet "$g$" to denote the gauge coupling
and the determinant of the metric tensor, but the difference would be obvious since the latter always appears
in the form of $\sqrt{-g}$.)
The Lagrangian (\ref{Lagr 1}) is invariant under a local scale transformation (or Weyl transformation). In fact,
with a local parameter $\Omega(x)$ the Weyl transformation is defined as
\begin{eqnarray}
g_{\mu\nu} &\rightarrow& \tilde g_{\mu\nu} = \Omega^2(x) g_{\mu\nu}, \quad
g^{\mu\nu} \rightarrow \tilde g^{\mu\nu} = \Omega^{-2}(x) g^{\mu\nu}, \quad
\nonumber\\
\phi &\rightarrow& \tilde \phi = \Omega^{-1}(x) \phi, \quad
H \rightarrow \tilde H = \Omega^{-1}(x) H, \quad
A_\mu \rightarrow \tilde A_\mu = A_\mu.
\label{Weyl transf}
\end{eqnarray}
Actually, it is straightforward to prove the Weyl invariance of ${\cal L}$
when we use the formulae $\sqrt{-g} = \Omega^{-4} \sqrt{- \tilde g}$ and
\begin{eqnarray}
R = \Omega^2 ( \tilde R + 6 \tilde \Box f - 6 \tilde g^{\mu\nu} \partial_\mu f \partial_\nu f ),
\label{Curvature}
\end{eqnarray}
with being defined as $f = \log \Omega$ and $\tilde \Box f = \frac{1}{\sqrt{- \tilde g}}
\partial_\mu (\sqrt{- \tilde g} \tilde g^{\mu\nu} \partial_\nu f) = \tilde g^{\mu\nu}
\tilde \nabla_\mu \tilde \nabla_\nu f$.
As a gauge condition for the Weyl symmetry, two gauge conditions, those are, the Einstein gauge (E-gauge) and
the Jordan gauge (J-gauge), are usually taken in \cite{Kallosh1}-\cite{Linde2}. First, we shall take the Einstein gauge and investigate
its implication.
With the unitary gauge $H(x) = \frac{1}{\sqrt{2}} e^{i \alpha \theta(x)} (0, h(x))^T$ where $\alpha$, $\theta(x)$ and
$h(x)$ are respectively a real number, the Nambu-Goldstone boson and the physical Higgs field,
the Einstein gauge (E-gauge) is of the form
\begin{eqnarray}
\phi^2 - 2 |H|^2 = \phi^2 - h^2 = 6.
\label{E-gauge}
\end{eqnarray}
This $SO(1,1)$ invariant gauge choice can be parametrized in terms of a canonically normalized
real scalar field $\varphi$ as
\begin{eqnarray}
\phi = \sqrt{6} \cosh \frac{\varphi}{\sqrt{6}}, \quad
h = \sqrt{6} \sinh \frac{\varphi}{\sqrt{6}}.
\label{Parametrization}
\end{eqnarray}
With the E-gauge (and the unitary gauge), the Lagrangian (\ref{Lagr 1}) can be rewritten as
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} &=& \frac{1}{2} R - \frac{1}{2} g^{\mu\nu}
\partial_\mu \varphi \partial_\nu \varphi - 3 g^2 g^{\mu\nu} B_\mu B_\nu \sinh^2 \frac{\varphi}{\sqrt{6}} \nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho} F_{\nu\sigma} - V(\varphi),
\label{Lagr-E}
\end{eqnarray}
where $V(\varphi)$ is the potential term $V(\phi, H)$ substituted by Eq. (\ref{Parametrization}).
And we have defined a new gauge field $B_\mu$ as $B_\mu = A_\mu + \frac{\alpha}{g} \partial_\mu \theta$
and consequently the gauge strength is now given by $F_{\mu\nu} \equiv \partial_\mu A_\nu - \partial_\nu A_\mu
= \partial_\mu B_\nu - \partial_\nu B_\mu$. Note that the third term in the RHS of Eq. (\ref{Lagr-E}) is written as
$\frac{1}{2} g^2 g^{\mu\nu} B_\mu B_\nu h^2$ which corresponds to the mass term of the gauge field in the
conventional formulation when the spontaneous symmetry breakdown (SSB) happens. In other words,
the gauge field $A_\mu$ eats the Nambu-Goldstone boson $\theta$, thereby becoming massive after the SSB.
However, it is not $h$ but $\varphi$ that one now regards as a fundamental field,
so the interpretation of the third term as the Higgs mass term is not suitable within the present framework.
Next, let us take the Jordan gauge (J-gauge) instead of the E-gauge. With the J-gauge $\phi = \sqrt{6}$,
the Lagrangian (\ref{Lagr 1}) is reduced to
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} = \frac{1}{2} \left( 1 - \frac{|H|^2}{3} \right) R - g^{\mu\nu} (D_\mu H)^\dagger (D_\nu H)
- \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho} F_{\nu\sigma} - V(H),
\label{Lagr-J}
\end{eqnarray}
where $V(H) \equiv V(\phi = \sqrt{6}, H)$. In order to show the equivalence of the Lagrangian in between the E-gauge and
the J-gauge, one can make use of a scale transformation in Eq. (\ref{Weyl transf}) by putting
\begin{eqnarray}
\Omega = 1 - \frac{|H|^2}{3} = 1 - \frac{h^2}{6},
\label{Omega}
\end{eqnarray}
where the unitary gauge is used in the second equality. After a straightforward calculation, we obtain the following
Lagrangian:
\begin{eqnarray}
\frac{1}{\sqrt{- g}}{\cal L} &=& \frac{1}{2} R - \frac{1}{2 \Omega^4} g^{\mu\nu} \partial_\mu h \partial_\nu h
- \frac{1}{2 \Omega^2} g^2 g^{\mu\nu} B_\mu B_\nu h^2 \nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho} F_{\nu\sigma} - \Omega^{-4} V(H),
\label{Lagr-J2}
\end{eqnarray}
where we have omitted to write tildes on fields denoting the transformed fields for simplicity of the
presentation.
Moreover, introducing $h' = \frac{1}{h}$ and then defining the canonically normalized scalar field $\varphi$
in terms of a differential equation
\begin{eqnarray}
\frac{d \varphi}{d h'} = \frac{1}{h'^2 - \frac{1}{6}},
\label{Diff-eq}
\end{eqnarray}
which is simply solved to
\begin{eqnarray}
h' \equiv \frac{1}{h} = \frac{1}{\sqrt{6} \tanh \frac{\varphi}{\sqrt{6}}},
\label{Diff-eq-sol}
\end{eqnarray}
the Lagrangian (\ref{Lagr-J2}) can be cast to
\begin{eqnarray}
\frac{1}{\sqrt{- g}}{\cal L} &=& \frac{1}{2} R - \frac{1}{2} g^{\mu\nu} \partial_\mu \varphi \partial_\nu \varphi
- 3 g^2 g^{\mu\nu} B_\mu B_\nu \sinh^2 \frac{\varphi}{\sqrt{6}} \nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho} F_{\nu\sigma} - \Omega^{-4} V(H).
\label{Lagr-J3}
\end{eqnarray}
Finally, to complete the proof of equivalence, we must specify the form of the potential $V(\phi, H)$. As the most general quadratic expression
which preserves the Weyl invariance and reduces to the Higgs potential in the low energy, let us consider the following potential:
\begin{eqnarray}
V(\phi, H) = \frac{\lambda}{36} F(z) \left[ 2 |H|^2 - G(z) \phi^2 \right]^2,
\label{Potential}
\end{eqnarray}
where $z$ is a gauge-invariant quantity defined as
\begin{eqnarray}
z \equiv \frac{\sqrt{2 |H|^2}}{\phi} = \frac{h}{\phi} = \tanh \frac{\varphi}{\sqrt{6}},
\label{z}
\end{eqnarray}
where at the last equality, the J-gauge and Eq. (\ref{Diff-eq-sol}) are used.
In the E-gauge, the potential term is written as
\begin{eqnarray}
V(\varphi) = \lambda F(z) \left[ \sinh^2 \frac{\varphi}{\sqrt{6}}
- G(z) \cosh^2 \frac{\varphi}{\sqrt{6}} \right]^2,
\label{Potential-E}
\end{eqnarray}
where $z = \tanh \frac{\varphi}{\sqrt{6}}$. On the other hand, the potential
in the J-gauge takes the form
\begin{eqnarray}
\Omega^{-4} V(H) &=& \cosh^4 \frac{\varphi}{\sqrt{6}} \cdot \frac{\lambda}{36}
F(z) \left[ 6 \tanh^2 \frac{\varphi}{\sqrt{6}}
- 6 G(z) \right]^2 \nonumber\\
&=& \lambda F(z) \left[ \sinh^2 \frac{\varphi}{\sqrt{6}}
- G(z) \cosh^2 \frac{\varphi}{\sqrt{6}} \right]^2,
\label{Potential-J}
\end{eqnarray}
which is equivalent to the potential (\ref{Potential-E}) in the E-gauge,
so we have completed the proof of the equivalence of the Lagrangian in both the gauge conditions.
\section{SO(1,1) accidental symmetry and shift symmetry}
For explanation of a relation between an $SO(1,1)$ global symmetry and a shift symmetry,
let us consider the simplest $SO(1,1)$ invariant potential
\begin{eqnarray}
V_0 (\phi, H) = \frac{\lambda}{36} \left( 2 |H|^2 - \phi^2 \right)^2
= \frac{\lambda}{36} \left( h^2 - \phi^2 \right)^2.
\label{Sim-Potential}
\end{eqnarray}
Moreover, we switch off the gauge field, $A_\mu = 0$. Then, in the unitary gauge,
the Lagrangian (\ref{Lagr 1}) is of form
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} = \frac{1}{12} \left( \phi^2 - h^2 \right) R
+ \frac{1}{2} g^{\mu\nu} \partial_\mu \phi \partial_\nu \phi
- \frac{1}{2} g^{\mu\nu} \partial_\mu h \partial_\nu h - V_0 (\phi, H).
\label{Lagr-SO}
\end{eqnarray}
It is obvious that there is a global $SO(1,1)$ symmetry in this Lagrangian.
Namely, the Lagrangian (\ref{Lagr-SO}) is invariant under the global $SO(1,1)$
transformation ($\varphi_0$ is a constant parameter)
\begin{eqnarray}
\left(
\begin{array}{c}
\phi' \\
h' \\
\end{array}
\right)
= \left(
\begin{array}{cc}
\cosh \frac{\varphi_0}{\sqrt{6}} & \sinh \frac{\varphi_0}{\sqrt{6}} \\
\sinh \frac{\varphi_0}{\sqrt{6}} & \cosh \frac{\varphi_0}{\sqrt{6}} \\
\end{array}
\right)
\left(
\begin{array}{c}
\phi \\
h \\
\end{array}
\right).
\label{SO(1,1)}
\end{eqnarray}
Using the parametrization (\ref{Parametrization}) in the E-gauge which respects the $SO(1,1)$ symmetry,
this $SO(1,1)$ transformation can be written as
\begin{eqnarray}
\left(
\begin{array}{c}
\phi' \\
h' \\
\end{array}
\right)
\equiv
\left(
\begin{array}{c}
\sqrt{6} \cosh \frac{\varphi'}{\sqrt{6}} \\
\sqrt{6} \sinh \frac{\varphi'}{\sqrt{6}} \\
\end{array}
\right)
= \left(
\begin{array}{c}
\sqrt{6} \cosh \frac{\varphi + \varphi_0}{\sqrt{6}} \\
\sqrt{6} \sinh \frac{\varphi + \varphi_0}{\sqrt{6}} \\
\end{array}
\right),
\label{SO(1,1)-E}
\end{eqnarray}
where we have used addition theorem of hyperbolic functions
\begin{eqnarray}
\sinh (\alpha \pm \beta) &=& \sinh \alpha \cosh \beta \pm \cosh \alpha \sinh \beta, \nonumber\\
\cosh (\alpha \pm \beta) &=& \cosh \alpha \cosh \beta \pm \sinh \alpha \sinh \beta.
\label{Addition}
\end{eqnarray}
Eq. (\ref{SO(1,1)-E}) clearly means that the $SO(1,1)$ transformation is nothing but a shift transformation
of $\varphi$ in the E-gauge, that is,
\begin{eqnarray}
\varphi \rightarrow \varphi' = \varphi + \varphi_0.
\label{Shift}
\end{eqnarray}
In inflation models, up to derivative terms, the potential term which is invariant under the shift symmetry
is known to be only a cosmological constant. Actually, the Lagrangian (\ref{Lagr-SO}) is written in the
E-gauge as
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} = \frac{1}{2} R
- \frac{1}{2} g^{\mu\nu} \partial_\mu \varphi \partial_\nu \varphi
- \lambda.
\label{Lagr-SO-E}
\end{eqnarray}
From this observation, it turns out that the existence of the non-trivial potential term for inflation
demands that one needs to have the potential which is not invariant under the $SO(1,1)$ transformation.
However, when we incorporate the gauge field in the Lagrangian as in the present formalism, the situation
completely changes, that is, the gauge field couples to only the Higgs field, thereby breaking the $SO(1,1)$
symmetry explicitly. In this sense, the $SO(1,1)$ symmetry is an "accidental" symmetry which exists only
in the gravitational sector. The accidental symmetry or custodial symmetry has some important consequences
in the standard model and a theory beyond standard model \cite{Logan}. In the present theory, however, it is
not clear how this accidental symmetry leads to important consequences.
\section{Higgs mechanism}
In this section, we wish to take account of the Higgs mechanism in the formalism at hand. For generality,
let us consider the most general form of the potential, Eq. (\ref{Potential}), in the Einstein gauge.
If we impose boundary conditions on the functions $F(z)$ and $G(z)$ in a small field limit
\begin{eqnarray}
\lim_{z \to 0} F(z) = 9, \quad \lim_{z \to 0} G(z) = \frac{v^2}{6},
\label{BC1}
\end{eqnarray}
the potential (\ref{Potential}) takes the following form in the limit $z \rightarrow 0$,
or equivalently $\varphi \rightarrow 0$:
\begin{eqnarray}
V \simeq \frac{\lambda}{4} ( \varphi^2 - v^2 )^2 \simeq \frac{\lambda}{4} ( h^2 - v^2 )^2,
\label{Potential-S}
\end{eqnarray}
where we have used $\varphi \simeq h$ for $\varphi \ll 1$. This is the conventional Higgs potential, so when
the Higgs field takes the vacuum expectation value $< h > = v$, the gauge field $B_\mu$
has the mass $M_B = g v$ as seen in Eq. (\ref{Lagr-E}). Note that $\phi \simeq \sqrt{6}$ for $\varphi \ll 1$,
which implies that the dynamical degree of freedom associated with $\phi$ naturally disappears and is fixed to
be a definite value of the J-gauge in a small field limit.
In a large field limit $\varphi \gg 1$, the term $3 g^2 B_\mu^2 \sinh^2 \frac{\varphi}{\sqrt{6}}$, giving
rise to the mass of the gauge field in a small field limit, loses a role as the mass term, and becomes an
interaction term between the gauge field $B_\mu$ and the inflaton $\varphi$.
In fact, since for $\varphi \gg 1$, Eq. (\ref{Parametrization}) gives us
\begin{eqnarray}
\phi \simeq h \simeq \sqrt{\frac{3}{2}} \ e^{\frac{\varphi}{\sqrt{6}}},
\label{Phi-h}
\end{eqnarray}
we obtain
\begin{eqnarray}
3 g^2 B_\mu^2 \sinh^2 \frac{\varphi}{\sqrt{6}} = \frac{1}{2} g^2 B_\mu^2 h^2
\simeq \frac{3}{4} g^2 B_\mu^2 e^{\sqrt{\frac{2}{3}} \varphi}.
\label{Non-renorma}
\end{eqnarray}
This term gives rise to non-renormalizable interactions between the gauge field and the inflaton,
and plays a role in the process of reheating \cite{Tomoyose}.
\section{The generalization to non-abelian gauge field}
For completeness, let us extend the abelian gauge field to the non-abelian gauge field
since the standard model is constructed on the basis of the gauge group $SU(3)_C \times SU(1)_L \times U(1)_Y$. For clarity,
we shall consider only the $SU(2)$ gauge group since the generalization to a general non-abelian
gauge field is straightforward.
Let us start with the $SU(2)$ generalization of the Lagrangian (\ref{Lagr 1})
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} &=& \frac{1}{12} \left( \phi^2 - 2 |H|^2 \right) R + \frac{1}{2} g^{\mu\nu}
\partial_\mu \phi \partial_\nu \phi - g^{\mu\nu} (D_\mu H)^\dagger (D_\nu H) \nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho}^a F_{\nu\sigma}^a - V(\phi, H),
\label{NA-Lagr 1}
\end{eqnarray}
where $a$ is an $SU(2)$ index running over $1, 2, 3$, and the covariant derivative and field strength
are respectively defined as
\begin{eqnarray}
D_\mu H &=& (\partial_\mu - i g \tau^a A_\mu^a) H, \quad
(D_\mu H)^\dagger =H^\dagger (\overleftarrow{\partial}_\mu + i g \tau^a A_\mu^a), \quad \nonumber\\
F_{\mu\nu}^a &=& \partial_\mu A_\nu^a - \partial_\nu A_\mu^a + g \varepsilon^{abc} A_\mu^b A_\nu^c.
\label{NA-Def 1}
\end{eqnarray}
Here $g$ is an $SU(2)$ coupling constant. Furthermore, the matrices $\tau^a$ are defined as half of the Pauli ones, i.e.,
$\tau^a = \frac{1}{2} \sigma^a$, so the following relations are satisfied:
\begin{eqnarray}
\{ \tau^a, \tau^b \} = \frac{1}{2} \delta^{ab}, \quad
[ \tau^a, \tau^b ] = i \varepsilon^{abc} \tau^c.
\label{tau-matrix}
\end{eqnarray}
In order to see the Higgs mechanism discussed in the previous section explicitly, it is
convenient to go to the unitary gauge. To do that, we first parametrize the Higgs doublet as
\begin{eqnarray}
H(x) = \frac{1}{\sqrt{2}} U^{-1}(x) \left(
\begin{array}{c}
0 \\
h(x) \\
\end{array}
\right),
\label{U-gauge}
\end{eqnarray}
where a unitary matrix $U(x)$ is defined as $U(x) = e^{ -i \alpha \tau^a \theta^a(x)}$ with $\alpha$ and
$\theta^a(x)$ being a real number and the Nambu-Goldstone fields, respectively.
Then, we will define new fields in the unitary gauge by
\begin{eqnarray}
H^u(x) &=& U(x) H(x) = \frac{1}{\sqrt{2}} \left(
\begin{array}{c}
0 \\
h(x) \\
\end{array}
\right), \nonumber\\
\tau^a B_\mu^a &=& U(x) \tau^a A_\mu^a U^{-1}(x) - \frac{i}{g} \partial_\mu U(x) U^{-1}(x).
\label{NA-new fields}
\end{eqnarray}
Using these new fields, after an easy calculation, we find the following relations
\begin{eqnarray}
D_\mu H = U^{-1}(x) D_\mu H^u, \quad
F_{\mu\nu}^a F^{a \mu\nu} = F_{\mu\nu}^a(B) F^{a \mu\nu}(B),
\label{NA-Rel 1}
\end{eqnarray}
where $D_\mu H^u$ and $F_{\mu\nu}^a(B)$ are respectively defined as
\begin{eqnarray}
D_\mu H^u = (\partial_\mu - i g \tau^a B_\mu^a) H^u, \quad
F_{\mu\nu}^a(B) = \partial_\mu B_\nu^a - \partial_\nu B_\mu^a + g \varepsilon^{abc} B_\mu^b B_\nu^c.
\label{NA-Def 2}
\end{eqnarray}
Thus, with the unitary gauge, the Lagrangian (\ref{NA-Lagr 1}) becomes
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} &=& \frac{1}{12} \left( \phi^2 - 2 |H^u|^2 \right) R + \frac{1}{2} g^{\mu\nu}
\partial_\mu \phi \partial_\nu \phi - g^{\mu\nu} (D_\mu H^u)^\dagger (D_\nu H^u) \nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho}^a(B) F_{\nu\sigma}^a(B) - V(\phi, U^{-1} H^u).
\label{NA-Lagr 1-U}
\end{eqnarray}
This Lagrangian is easily evaluated in the Einstein gauge to
\begin{eqnarray}
\frac{1}{\sqrt{-g}}{\cal L} &=& \frac{1}{2} R - \frac{1}{2} g^{\mu\nu}
\partial_\mu \varphi \partial_\nu \varphi - \frac{3 g^2}{4} g^{\mu\nu} B_\mu^a B_\nu^a \sinh^2 \frac{\varphi}{\sqrt{6}}
\nonumber\\
&-& \frac{1}{4} g^{\mu\nu} g^{\rho\sigma} F_{\mu\rho}^a(B) F_{\nu\sigma}^a(B) - V(\phi, U^{-1} H^u).
\label{NA-Lagr 1-E}
\end{eqnarray}
Here the potential is taken to be the most general quadratic form (\ref{Potential}) and is rewritten in the E-gauge as
\begin{eqnarray}
V(\phi, H) = V(\phi, U^{-1} H^u) = V(\phi, H^u)
= \lambda F(z) \left[ \sinh^2 \frac{\varphi}{\sqrt{6}} - G(z) \cosh^2 \frac{\varphi}{\sqrt{6}} \right]^2,
\label{NA-Potential}
\end{eqnarray}
where $z \equiv \tanh \frac{\varphi}{\sqrt{6}}$. Then, with the boundary conditions (\ref{BC1}) this potential takes
the same form as (\ref{Potential-S}) in small field values so that when the Higgs field takes the vacuum expectation value
$< h > = v$, the non-abelain gauge field $B_\mu$ has the mass $M_B = \frac{1}{2} g v$. Note that in the non-abelain
gauge field, the Higgs mass squared term depends on the inflaton field $\varphi$ like
$\frac{3 g^2}{4} g^{\mu\nu} B_\mu^a B_\nu^a \sinh^2 \frac{\varphi}{\sqrt{6}}$, so its physical property has a perfectly
similar aspect to that in the abelian gauge field.
\section{Inflation}
Based on our abelian gauge model, we will turn our attention to inflation. In particular,
we wish to ask ourselves if the abelian model accounts for the recent BICEP2 results
or not. It will turn out that the answer is not affirmative even if we consider a rather singular potential
in addition to a non-singular one since our model shares a common feature with the Starobinsky
inflation model \cite{Starobinsky}.
As a first step, let us select a boundary condition in a large field limit
\begin{eqnarray}
\lim_{z \to 1} F(z) = F(1) < \infty, \quad \lim_{z \to 1} G(z) = 1,
\label{BC2}
\end{eqnarray}
where $F(1)$ is a finite quantity. The second boundary condition $G(1) = 1$ is chosen in a such way that
our model does not lead to cyclic cosmology \cite{Bars1, Bars2}.
In the Einstein gauge, the most general quadratic potential (\ref{Potential}) is written in
a concise form as
\begin{eqnarray}
V(\phi, H) &=& \frac{\lambda}{36} F(z) \left[ 2 |H|^2 - G(z) \phi^2 \right]^2 \nonumber\\
&=& \lambda F(z) \left[ \frac{z^2 - G(z)}{z^2 - 1} \right]^2,
\label{Potential-Inf}
\end{eqnarray}
where $z \equiv \frac{\sqrt{2 |H|^2}}{\phi} = \frac{h}{\phi} = \tanh \frac{\varphi}{\sqrt{6}}$.
However, this concise expression turns out to be not convenient for analysing cosmological parameters
at large field values, so instead we will use the following equivalent form of the potential
\begin{eqnarray}
V(\varphi) = \lambda F(\tanh \frac{\varphi}{\sqrt{6}}) \cdot \sinh^4 \frac{\varphi}{\sqrt{6}}
\left[ 1- G(\tanh \frac{\varphi}{\sqrt{6}}) \tanh^{-2} \frac{\varphi}{\sqrt{6}} \right]^2.
\label{Potential-Inf2}
\end{eqnarray}
At large field values $\varphi \gg 1$, hyperbolic functions are approximated to
\begin{eqnarray}
\sinh \frac{\varphi}{\sqrt{6}} \simeq \frac{1}{2} e^{\frac{\varphi}{\sqrt{6}}}, \quad
\tanh \frac{\varphi}{\sqrt{6}} \simeq 1 - 2 e^{- \sqrt{\frac{2}{3}} \varphi}
\left( 1 - e^{- \sqrt{\frac{2}{3}} \varphi} \right).
\label{Hyperbolic}
\end{eqnarray}
Next let us assume that $F(x)$ and $G(x)$ are analytic around $x = 1$. Then, for $\varphi \gg 1$
we obtain
\begin{eqnarray}
F(\tanh \frac{\varphi}{\sqrt{6}}) &\simeq& F(1) - 2 F'(1) e^{- \sqrt{\frac{2}{3}} \varphi}
\left( 1 - e^{- \sqrt{\frac{2}{3}} \varphi} \right), \nonumber\\
G(\tanh \frac{\varphi}{\sqrt{6}}) &\simeq& 1 - 2 G'(1) e^{- \sqrt{\frac{2}{3}} \varphi}
\left( 1 - e^{- \sqrt{\frac{2}{3}} \varphi} \right),
\label{FG}
\end{eqnarray}
where the boundary condition $G(1) = 1$ is used and $F'(x) \equiv \frac{d F(x)}{d x}$.
Using these approximation formulae holding at large field values, the potential (\ref{Potential-Inf2})
takes the form
\begin{eqnarray}
V(\varphi) \simeq A \left( 1- B e^{- \sqrt{\frac{2}{3}} \varphi} \right),
\label{Potential-Inf3}
\end{eqnarray}
where we have defined
\begin{eqnarray}
A &=& \frac{\lambda}{4} F(1) \left( G'(1) - 2 \right)^2, \nonumber\\
B &=& \frac{F'(1)}{F(1)} - \frac{3 G'(1) + 2}{G'(1) - 2}.
\label{AB}
\end{eqnarray}
Here we assume that both $A$ and $B$ are positive and $G'(1) \neq 2$.
The potential (\ref{Potential-Inf3}) yields the slow-roll parameters
\begin{eqnarray}
\epsilon_V &\equiv& \frac{1}{2} \left( \frac{V_{, \varphi}}{V} \right)^2
\simeq \frac{B^2}{3} e^{- 2 \sqrt{\frac{2}{3}} \varphi}, \nonumber\\
\eta_V &\equiv& \frac{V_{, \varphi\varphi}}{V}
\simeq - \frac{2 B}{3} e^{- \sqrt{\frac{2}{3}} \varphi},
\label{SL parameters}
\end{eqnarray}
where $V_{, \varphi} \equiv \frac{d V}{d \varphi}$ and $V_{, \varphi\varphi}
\equiv \frac{d^2 V}{d \varphi^2}$. The number of e-folding can be evaluated to
\begin{eqnarray}
N \equiv \int_{\varphi_e}^\varphi d \varphi \frac{1}{\sqrt{2 \epsilon_V}}
\simeq \frac{3}{2B} \left( e^{\sqrt{\frac{2}{3}} \varphi} - e^{\sqrt{\frac{2}{3}}
\varphi_e} \right).
\label{N}
\end{eqnarray}
The inflation ends when the slow-roll parameters are approximately the unity,
$\epsilon_V(\varphi_e) = 1$, so we obtain $e^{\sqrt{\frac{2}{3}}
\varphi_e} = \frac{B}{\sqrt{3}}$. Then, we can derive the relation
holding $N \gg 1$
\begin{eqnarray}
e^{\sqrt{\frac{2}{3}} \varphi} = \frac{2 B N}{3} + \frac{B}{\sqrt{3}}
\simeq \frac{2 B N}{3}.
\label{e-varphi}
\end{eqnarray}
Using this relation (\ref{e-varphi}), the slow-roll parameters can be
expressed in terms of the number of e-folding like
\begin{eqnarray}
\epsilon_V = \frac{3}{4 N^2}, \quad
\eta_V = - \frac{1}{N}.
\label{SL parameters 2}
\end{eqnarray}
Then, the tensor-to-scalar ratio $r$ and the spectral index $n_s$ are
given by
\begin{eqnarray}
r \equiv 16 \epsilon_V = \frac{12}{N^2}, \quad
n_s \equiv 1 - 6 \epsilon_V + 2 \eta_V \simeq 1 - \frac{2}{N},
\label{Index}
\end{eqnarray}
which are in a perfect agreement with the values obtained in the Starobinsky model \cite{Starobinsky}.
Concretely, for $N = 60$, they take the values $r \simeq 0.003$ and $n_s \simeq 0.967$,
which do not coincide with the BICEP2 results $r \simeq 0.16$ and $n_s \simeq 0.96$.
So far, we have considered only analytic functions for $F$ and $G$. One possible
extension is to consider non-analytic functions and check if they lead to
the BICEP2 results or not. A general treatment of non-analytic functions is
not easy, but an instance turns out to be enough to derive a general feature of the
non-analytic functions.
Let us focus on an example of non-analytic functions such that
\begin{eqnarray}
F(z) = 10 - (1 - z^n)^\alpha, \quad
G(z) = \frac{v^2}{6} + \left(1 - \frac{v^2}{6} \right) z^m,
\label{Non-FG}
\end{eqnarray}
where $0 < \alpha < 1$. Note that these specific functions satisfy the boundary conditions
(\ref{BC1}) and (\ref{BC2}), but the derivatives of $F$ at $z=1$ diverge, $F'(1) = \infty,
F''(1) = \infty, \cdots $. Using a behavior of $F$ for $\varphi \gg 1$
\begin{eqnarray}
F(\varphi) \simeq 10 - (2n)^\alpha e^{- \alpha \sqrt{\frac{2}{3}} \varphi},
\label{Non-F}
\end{eqnarray}
the potential reads in a large field limit $\varphi \gg 1$
\begin{eqnarray}
V(\varphi) \simeq \frac{5 \lambda}{2} (m-2)^2 \left[ 1 - \frac{(2n)^\alpha}{10}
e^{- \alpha \sqrt{\frac{2}{3}} \varphi} \right],
\label{Non-V}
\end{eqnarray}
where a term involving the factor $e^{- \alpha \sqrt{\frac{2}{3}} \varphi}$ in the non-analytic
function $F(\varphi)$ becomes more dominant compared to terms including the factor
$e^{- \sqrt{\frac{2}{3}} \varphi}$ coming from hyperbolic functions $\sinh \frac{\varphi}{\sqrt{6}}$
and $\tanh \frac{\varphi}{\sqrt{6}}$.
Along the same line of argument as in the potential (\ref{Potential-Inf3}), it is straightforward
to derive the slow-roll parameters and cosmological indices whose result is summarized as
\begin{eqnarray}
\epsilon_V &=& \frac{3}{4 \alpha^2 N^2}, \quad
\eta_V = - \frac{1}{N}, \nonumber\\
r &=& \frac{12}{\alpha^2 N^2}, \quad
n_s = 1 - \frac{9}{2 \alpha^2 N^2} - \frac{2}{N}.
\label{Indices}
\end{eqnarray}
As a remark, in deriving this result, there appears one subtle point to be checked carefully.
Namely, we have approximated the factor $e^{- \alpha \sqrt{\frac{2}{3}} \varphi}$ by the number of
e-folding $N$ through the condition $\epsilon_V(\varphi_e) = 1$ as
\begin{eqnarray}
e^{\alpha \sqrt{\frac{2}{3}} \varphi} = \frac{(2n)^\alpha \alpha^2}{15} N
+ \frac{(2n)^\alpha \alpha}{10 \sqrt{3}}
\approx \frac{(2n)^\alpha \alpha^2}{15} N.
\label{Non-e-varphi}
\end{eqnarray}
The last approximation obviously depends on the size of $\alpha$ because of $0 < \alpha < 1$.
We will show the validity of this approximation shortly by taking numerical values concretely\footnote
{Of course, we can derive the similar result without using the approximation (\ref{Non-e-varphi}).
Then, the expressions in Eq. (\ref{Indices}) take more complicated forms
\begin{eqnarray}
\epsilon_V &=& \frac{3}{4 \left(\alpha N + \frac{\sqrt{3}}{2} \right)^2}, \quad
\eta_V = - \frac{\alpha}{\alpha N + \frac{\sqrt{3}}{2}}, \nonumber\\
r &=& \frac{12}{\left(\alpha N + \frac{\sqrt{3}}{2} \right)^2}, \quad
n_s = 1 - \frac{9}{2 \left(\alpha N + \frac{\sqrt{3}}{2} \right)^2} - \frac{2 \alpha}{\alpha N + \frac{\sqrt{3}}{2}}.
\label{Indices2}
\end{eqnarray}
}.
In order to show that the result (\ref{Indices}) does not match the BICEP2 experiment,
let us impose the condition $r = 0.16$ from the BICEP2 results. Then, we have $\epsilon_V = 0.01$
owing to the definition $r = 16 \epsilon_V$. Using the definition of the spectral index,
$n_s \equiv 1 - 6 \epsilon_V + 2 \eta_V$ and the result $\eta_V = - \frac{1}{N}$ in Eq. (\ref{Indices}),
for $N = 60$ the spectral index is given by $n_s = 0.91$, which does not coincide with the BICEP2
result $n_s = 0.96$. Incidentally, using the result $r = \frac{12}{\alpha^2 N^2}$ in Eq. (\ref{Indices}),
we can evaluate $\alpha = \frac{\sqrt{3}}{12}$, thereby making it possible to show
$\frac{(2n)^\alpha \alpha^2}{15} N \gg \frac{(2n)^\alpha \alpha}{10 \sqrt{3}}$ as promised in the
above.
Alternatively, we can present the same conclusion by fixing the value of the spectral index $n_s$ first, and
then calculate the tensor-to-scalar ratio $r$. With the values $n_s = 0.96$ from the BICEP2 and
$N = 60$, the tensor-to-scalar ratio $r$ can be evaluated via Eq. (\ref{Indices}) to
$r \simeq 0.027$, which is much smaller than the BICEP2 result $r = 0.16$.
At this stage, it is worthwhile to consider the role of the non-analytical functions in a general framework.
The point is that the non-analytical functions in general change the coefficient in the exponential
in the potential, that is, the exponential is changed from $e^{- \sqrt{\frac{2}{3}} \varphi}$ to
$e^{- a \varphi}$ where $0 < a < \sqrt{\frac{2}{3}}$. Thus, in order to show that general
non-analytic functions do not reproduce the BICEP2 results, it is sufficient to consider a general
potential in a large field limit $\varphi \gg 1$
\begin{eqnarray}
V(\varphi) = V_0 \left( 1- c e^{- a \varphi} \right),
\label{G-Potential}
\end{eqnarray}
where both $V_0$ and $c$ are positive. It is easy to see that this more general potential
produces the same result as (\ref{Indices}) in the more specific potential (\ref{Non-V}).
\section{Conclusion}
In this article, we have investigated a locally scale-invariant theory which describes the coupling of gravity
and the standard model in detail. After constructing an abelian gauge model, it is shown that this model
exhibits a peculiar feature of a coupling between the gauge field and the inflaton in large field
values while it describes the standard model coupled to general relativity in small field values.
Moreover, we have extended the abelian model to non-Abelian gauge groups. It is also straightforward to
generalize our model to the case of multi-component scalar fields.
We have also discussed a possibility that our model could explain the recent BICEP2 results, but
we have shown that it is very difficult to do so even if we take into consideration non-analytic types
of the potential. This fact is naturally understood since our model shares many of physical properties
with the Starobinsky $R^2$ inflation model \cite{Starobinsky}.
One interesting direction of future work is to spell out effects on the reheating coming from the coupling
between the gauge field and the inflaton \cite{Tomoyose}. Another future problem is to clarify quantum effects
in our model along our previous study \cite{Oda-Tomo}. The latter problem is very important since
we have to show that the running coupling constant $\lambda(\mu)$ could take the value $\lambda = {\cal O}(0.1)$ at the TeV
scale to make our model to be consistent with the standard model.
We wish to return these problems in near future.
\begin{flushleft}
{\bf Acknowledgements}
\end{flushleft}
This work is supported in part by the Grant-in-Aid for Scientific
Research (C) No. 25400262 from the Japan Ministry of Education, Culture,
Sports, Science and Technology.
|
\section*{Supplemental Material}
Here we present the detailed proofs of the results stated in the main text. We follow the same notation and all the numberings of equations and statements
refer to the main text.
\section{Bounds for Bell inequalities.}
We start by giving upper bounds on the maximum violation of a Bell inequality achieved by a quantum state in the single copy scenario.
In the main text we have proved Theorem 1 for positive coefficients Bell inequality, here we present the proof for general coefficients:
{\it Proof of Theorem 1}.
Theorem 1 for general coefficients follows directly by the relations:
\begin{align}
|\mathbf{S}(\rho) - \mathbf{S}(\sigma)| = & |\sum_{a,b,x,y} Tr s^{a,b}_{x,y} A_{a|x}\otimes B_{b|y} (\rho - \sigma)| \nonumber \\
=&|\sum_{a,b,x,y} Tr s^{a,b}_{x,y} A_{a|x}\otimes (B_{b|y})^T (\rho - \sigma)^{\Gamma}| \nonumber\\
=& | Tr\, \mathbf{S}^{\Gamma}(\rho^{\Gamma} -\sigma^{\Gamma}) | \label{eq:infty-norm}\\
\leq & Tr\, |\mathbf{S}^{\Gamma}(\rho^{\Gamma} -\sigma^{\Gamma}) | \nonumber \\
\leq &||\mathbf{S}^{\Gamma}||_{\infty}||\rho^{\Gamma} - \label{eq:positive} \sigma^{\Gamma}||.\nonumber
\end{align}
where in the fourth step we use the triangle inequality, and the last step follows from H\"{o}lder's inequalities for $p-$norms.
\qed
\begin{corone} For any bipartite Bell expression $\mathcal{S}$, and state $\rho$, it holds that:
\begin{equation}
Q_{\mathcal{S}}(\rho) \leq C(\mathcal{S}) + Q(\mathcal{S})\inf_{\sigma \in SEP} ||\rho^{\Gamma} - \sigma^{\Gamma}|| . \nonumber
\end{equation}
\end{corone}
{\it Proof of Corollary 1}.
First note that by substituting any separable state $\sigma$ in eq. (1), and using the fact that
$\mathbf{S}(\sigma_{AB}) \leq C(\mathcal{S}) \;\; \forall \; \sigma_{AB} \in SEP$
we have:
\begin{equation}
\mathbf{S}(\rho) \leq C(\mathcal{S}) + ||\mathbf{S}^{\Gamma}||_{\infty}\inf_{\sigma \in SEP} ||\rho^{\Gamma} - \sigma^{\Gamma}||.
\end{equation}
Now taking supremum over POVMS $\{A_{a|x}\otimes B_{b|y}\}$
on both sides we have the desired result.
\qed
Based on \cite{karol-PhD,BCHW-swapping}, we have by Corollary 1 an immediate observation that certain private states have a limited possibility of violating
a Bell inequality.
{\observation For any bipartite Bell inequality $\mathcal{S}$, if the states $\sqrt{XX^{\dagger}}$ and $\sqrt{X^{\dagger}X}$ are separable,
then a private state $\gamma_X$, described by $X$ according to eq. (7), satisfies:
\begin{equation}
Q(\gamma_X) \leq C(\mathcal{S}) + Q(\mathcal{S})||X^{\Gamma}||.
\end{equation}
While, as shown before, $||X^{\Gamma}||$ can be vanishing exponentially fast in number of qubits that composes $\gamma_X$.}
\begin{prop1}
There exist bipartite states (see eq. \eqref{eq:rec-state-presented} bellow) $\rho \in B({\cal C}^2\otimes{\cal C}^2 \otimes ({\cal C}^{d^k}\otimes{\cal C}^{d^k})^{\otimes m})$ with $d=m^2$, $k=m$ satisfying
$K_D(\rho^{\Gamma}\otimes\rho) \rightarrow 1$ with increasing $m$, such that:
\begin{equation}
Q_{\mathcal{S}}(\rho\otimes\rho^{\Gamma})\leq C(\mathcal{S}) +{ Q(\mathcal{S})\over 2^{m-1}}. \nonumber
\end{equation}
\end{prop1}
{\it Proof of Proposition 1}. Consider $\rho$ defined in eq. (149) of \cite{BCHW-swapping}.
It has the property that its distillable key is almost 1, and there is a separable state $\sigma_{\rho}$
such that $||\rho^{\Gamma}- \sigma_{\rho}^{\Gamma}||\leq p$, with
$p = {({1\over 2} - q)^m\over 2q^m+ 2({1\over 2} - q)^m}$ and $q ={1\over 3}$. Note that $p \leq {1\over 2^m}$ for natural $m$. Knowing this, we
bound the violation achieved by the above states in two steps.
We first apply Theorem 1 to state $\rho\otimes\rho^{\Gamma}$, with $\sigma = \sigma_{\rho}\otimes \rho^{\Gamma}$, in order to obtain:
\begin{equation}
\mathbf{S}(\rho\otimes\rho^{\Gamma}) \leq \mathbf{S}(\sigma_{\rho}\otimes\rho^{\Gamma}) + Q(\mathcal{S})||(\rho\otimes\rho^{\Gamma})^{\Gamma} - (\sigma_{\rho}\otimes\rho^{\Gamma})^{\Gamma}||,
\end{equation}
which in turn is bounded by
\begin{equation}
\mathbf{S}(\sigma_{\rho}\otimes\rho^{\Gamma}) + Q(\mathcal{S})||\rho^{\Gamma} - \sigma_{\rho}^{\Gamma}||.
\end{equation}
We now use the fact that $\sigma_{\rho}$ is separable, and that we can write $\rho^{\Gamma} = (1-r) \rho_{sep} + r\rho_{nsep}$,
with $\rho_{sep} \in SEP$ and $r \leq p$. By linearity of trace we obtain:
\begin{equation}
\mathbf{S}(\sigma_{\rho}\otimes\rho^{\Gamma}) = (1-r) \mathbf{S}(\sigma_{\rho}\otimes\rho_{sep}) + r \mathbf{S}(\sigma_{\rho}\otimes\rho_{nsep}).
\end{equation}
First term of RHS is bounded by $C(\mathcal{S})$, as the state $\sigma_{\rho}\otimes \rho_{sep}$ is separable. The second term is bounded
by $Q(\sigma_{\rho}\otimes\rho_{nsep})$, which is in fact equal to $Q(\rho_{nsep})$.
This lead us to the following bound:
\begin{eqnarray}
\mathbf{S}(\rho\otimes \rho^{\Gamma}) \leq C(\mathcal{S}) + pQ(\rho_{nsep})+ p Q(\mathcal{S}) \leq C(\mathcal{S})+ 2pQ(\mathcal{S}) \leq C(\mathcal{S}) + {Q(\mathcal{S})\over 2^{m-1}},
\end{eqnarray}
which proves the result.
\qed
An example of states
satisfying Proposition 1 is the family $\hat{\rho}_{p,d,k,m}$ on $B\left({\cal C}^2\otimes{\cal C}^2\otimes({\cal C}^{d^k}\otimes{\cal C}^{d^k})^{\otimes m}\right)$
\cite{pptkey} (see \cite{BCHW-swapping}).
Their matrix form is given below, up to the normalization factor $N_m = 2(p^m)+2({1\over 2} -p)^m$:
\begin{equation}
\left[\begin{array}{cccc}
[p({\tau_1+\tau_2\over 2})]^{\otimes m} &0&0&[p({\tau_1-\tau_2\over 2})]^{\otimes m} \\
0& [({1\over 2}-p)\tau_2]^{\otimes m}&0&0 \\
0&0&[({1\over 2}-p)\tau_2]^{\otimes m}& 0\\
{[p({\tau_1-\tau_2\over 2})]}^{\otimes m}&0&0&{[p({\tau_1+\tau_2\over 2})]}^{\otimes m}\\
\end{array}
\right].
\label{eq:rec-state-presented}
\end{equation}
$\tau_1 = ({\rho_a + \rho_s\over 2})^{\otimes k}$ and $\tau_2=(\rho_s)^{\otimes k}$, while $\rho_s$ and $\rho_a$ are the $d$-dimensional symmetric and antisymmetric Werner state, respectively.
\section{ Bound on asymptotic nonlocality}
To treat the asymptotic scenario we introduce the restricted regularized relative entropy of nonlocality.
This quantity is an entanglement measure, and is related to the relative entropy of nonlocality:
\begin{eqnarray}
T^{\infty}(\rho_{AB}) \equiv \varlimsup_{n\rightarrow \infty} {1\over n} \sup_{\Lambda\in LOCC} \sup_{\{M_{xy}\}} \sup_{p(x,y)} \inf_{\sigma \in SEP} \sum_{x,y} p(x,y) D(\{Tr M_{xy}\Lambda(\rho^{\otimes n})\}||\{Tr M_{xy}\Lambda(\sigma^{\otimes n})\}).
\end{eqnarray}
In the expression $D(\{Tr M_{xy}\Lambda(\rho)\}||\{Tr M_{xy}\Lambda(\sigma)\})$, we can treat $\{Tr M_{xy}\Lambda(\rho)\}$
as a diagonal matrix
with elements given by the probability distribution $P_{xy}(ab|xy)$, and then $D(\cdot||\cdot)$ is the quantum relative entropy.
A similar quantity has been introduced by Piani in \cite{Piani2009-relent}. This function is easier to deal with than $R(\rho)$.
Note that the definition of $T^{\infty}$ originates from $R$ by relaxing the optimization over local boxes to an optimization over separable states
and same local measurements, other relaxations can be defined similarly giving rise to upper bounds of independent interests.
Recalling the definitions introduced in the main text, we use the relative entropy as a measure of nonlocality
\begin{equation}
\mathcal{N}(\mathcal{P}) = \sup_{\{p(x,y)\}} \inf_{P_L \in L} \sum_{x,y} p(x,y) D(P_{xy}(ab|xy) || P_L(ab|x,y)),
\end{equation}
and for the asymptotic scenario we define the asymptotic relative entropy of nonlocality:
\begin{equation}
R(\rho_{AB}) \equiv \varlimsup_{n \rightarrow \infty}{1\over n} \sup_{\Lambda \in LOCC} \sup_{\DE{M_{xy}}} \mathcal{N}(\{Tr M_{xy}\Lambda(\rho_{AB}^{\otimes n})\}).
\end{equation}
In all the following proofs we consider optimization over the probability distribution of the inputs $\{p(x,y)\}$, but one can also restrict to the uniform case
\cite{vDamGrunwaldGill} and all the results follow in the same way.
We are now ready to prove Theorems 2.
\begin{thm2} For any bipartite state it holds that
\begin{equation}
R(\rho_{AB}) \leq T^{\infty}(\rho_{AB})\leq E_r(\rho_{AB}).
\label{eq:RT2}
\end{equation}
For $\rho_{AB}$ a PPT state, it holds that
\begin{equation}\label{T-Er}
T^{\infty}(\rho_{AB}) \leq E_r(\rho_{AB}^{\Gamma}).
\end{equation}
\end{thm2}
{\it Proof of Theorem 2}. We first prove that $R\leq T^{\infty}$.
In the first step let us note that $\Lambda(\sigma^{\otimes n})$ is a separable state, since $\Lambda$ is an LOCC operation. Hence,
if we place infimum over all separable states $\sigma$ instead of that of the form $\Lambda(\sigma^{\otimes n})$ in
definition of $T^{\infty}$, we may only decrease the quantity. Second, we observe that instead of obtaining the local quantum
box via the same POVMs, $\DE{M_{xy}}$, as for $\Lambda(\rho^{\otimes n})$, we can place also infimum over all $\tilde{M}_{xy}$ acting on $\sigma$,
which also can only lower the quantity. In the last step we observe that the set of such obtained quantum boxes is included
in the set of the local ones, hence we can place infumum over the latter instead, which proves the desired result.
Now we prove the relation $T^{\infty}(\rho) \leq E_r(\rho^{\Gamma})$ for PPT states. The proof of $T^{\infty}(\rho) \leq E_r(\rho)$ follows
in the same way, without use of partial transposition. To prove that $T^{\infty}(\rho) \leq E_r(\rho^{\Gamma})$ note that,
since $\Lambda \in LOCC$, it has a separable representation: $\Lambda(\rho) = \sum_{i,j} C_i\otimes D_j(\rho)C_i^{\dagger} \otimes D_j^{\dagger}$.
Using properties of trace and the separable representation we focus now on the term:
\begin{eqnarray}
D(\{Tr \sum_{ij}C_i^{\dagger}A_{a|x}C_i\otimes D_j^{\dagger}B_{b|y}D_j\rho^{\otimes n}\} ||\{Tr\sum_{ij}C_i^{\dagger}A_{a|x}C_i\otimes D_j^{\dagger}B_{b|y}D_j \sigma'^{\otimes n} \}).
\end{eqnarray}
Applying to both its components the identity $Tr XY = Tr X^{\Gamma}Y^{\Gamma}$, we have
\begin{eqnarray}
D(\{Tr \sum_{ij}C_i^{\dagger}A_{a|x}C_i\otimes (D_j^{*})^{\dagger}B_{b|y}^{T}D_j^{*}(\rho^{\Gamma})^{\otimes n})\}||\{Tr\sum_{ij}C_i^{\dagger}A_{a|x}C_i\otimes (D_j^{*})^{\dagger}B_{b|y}^{T}D_j^{*} (\sigma'^{\Gamma})^{\otimes n}\}) ,
\end{eqnarray}
which implies that $T^{\infty}$
can be written as:
\begin{eqnarray}
\lim_{n\rightarrow \infty} {1\over n} \sup_{\Lambda' \in LOCC} \sup_{\DE{M'_{xy}}} \sup_{\DE{p(x,y)}} \inf_{\sigma' \in SEP} \sum_{x,y} p(x,y) D(\{Tr M'_{xy}\Lambda'({\rho^{\Gamma}}^{\otimes n})\}||\{Tr M'_{xy}\Lambda'({\sigma'^{\Gamma}}^{\otimes n})\})
\label{eq:Tb1}
\end{eqnarray}
with $\Lambda'$ being a new separable operation with $D_j$ operators complex conjugated, and $M'$ being a new set of POVMs with $B_{b|y}$ transposed.
Now, since $\sigma'^{\Gamma}$ is also a separable state, and by the fact that the relative entropy is non-increasing under completely
positive trace-preserving maps, we have:
\begin{eqnarray}
\inf_{\sigma' \in SEP}\sum_{x,y}p(x,y) D(\{Tr M'_{xy}\Lambda'({\rho^{\Gamma}}^{\otimes n})\}||\{Tr M'_{xy}\Lambda'({\sigma'^{\Gamma}}^{\otimes n})\}) \leq \inf_{\sigma \in SEP} \sum_{x,y} p(x,y) D({\rho^{\Gamma}}^{\otimes n}||\sigma^{\otimes n}).
\label{eq:Tb2}
\end{eqnarray}
We finally use the identity $D({\rho^{\Gamma}}^{\otimes n}||\sigma^{\otimes n}) = n D(\rho^{\Gamma}||\sigma)$. Since the latter term does not depend on $\Lambda$ and $M$ as well as $p(x,y)$,
and the number of copies $n$ cancels with the regularization term ${1\over n}$, we obtain via (\ref{eq:Tb1}) and (\ref{eq:Tb2}) the bound:
\begin{equation}
T^{\infty}(\rho) \leq \inf_{\sigma \in SEP} D(\rho^{\Gamma}||\sigma) \equiv E_r(\rho^{\Gamma}).
\end{equation}
\qed
\section{Bound on post-selected nonlocality}
We can also treat the case where a filtering (non-trace-preserving) operation is performed before the Bell test. To quantify the nonlocality achieved in this case we
define the \textit{asymptotic relative entropy of hidden-nonlocality}:
\begin{align}
R_H(\rho_{AB}) \equiv & \varlimsup_{n \rightarrow \infty}{1\over n} \sup_{\Lambda \in LOCC} \sup_{\DE{M_{xy}}} \sup_{F_0} p^{F_0} \mathcal{N}(\{Tr M_{xy}\Lambda(F_0(\rho_{AB}^{\otimes n}))\}),
\end{align}
where a filtering process, $F_0$, takes state $\Lambda(\rho_{AB}^{\otimes n})$
to flag form, $F(\rho)=\sum_i \ketbra{i}{i}\otimes F_{i} \rho F_{i}^{\dagger}$, and later erasures all other results except the ``good''
one that leads to the highest violation of the Bell inequality. $p^{F_0}= Tr F_{0}\Lambda(\rho_{AB}^{\otimes n})F_0^{\dagger}$ is
the probability that the filter results in the desired outcome.
Analogously to Theorem 2, for the hidden-nonlocality scenario, we have the following result:
\begin{thm3}
For any bipartite state $\rho_{AB}$ it holds that
\begin{equation}
R_H(\rho_{AB}) \leq T^{\infty}(\rho_{AB})\leq E_r(\rho_{AB}).
\end{equation}
For a bipartite PPT state $\rho_{AB}$ it holds that
\begin{equation}\label{RH-Er}
R_H(\rho_{AB}) \leq E_r(\rho_{AB}^{\Gamma}).
\end{equation}\
\end{thm3}
{\it Proof of Theorem 3}. We just have to show that $R_H(\rho_{AB}) \leq T^{\infty}(\rho_{AB})$ and \eqref{RH-Er} follows from \eqref{T-Er}.
Let us consider:
\begin{align}
T_{n}(\rho_f) \equiv {1\over n} \sup_{\Lambda\in LOCC} \sup_{\{M_{xy}\}} \sup_{p(x,y)} \inf_{\sigma \in SEP}\sum_{x,y} p(x,y) D(\{Tr M_{xy}\Lambda(\rho^{\otimes n})\}||\{Tr M_{xy}\Lambda(\sigma^{\otimes n})\}).
\end{align}
Now let us restrict to a map $\Lambda$ of the form $\Lambda= F \circ \Lambda_0$, where $\Lambda_0$ is an arbitrary LOCC operation that acts on
$n$ copies of the system, and is followed by measurement $F$,
\textit{i.e. } $\Lambda(\rho^{\otimes n})=\sum_i p^{F_i} \ketbra{i}{i} \otimes F_i (\Lambda_0(\rho^{\otimes n})) F_i^{\dagger}$.
This restriction just decrease the RHS, so we have
\begin{align}
T_n(\rho^{\otimes n}) \geq {1\over n}& \sup_{\Lambda_0\in LOCC} \sup_{F} \sup_{\{M_{xy}\}} \sup_{p(x,y)} \inf_{\sigma \in SEP} \\
& p^{F_0} \sum_{x,y} p(x,y) D(\{Tr M_{xy}\sum_i p^{F_i} \ketbra{i}{i} \otimes F_i (\Lambda_0(\rho^{\otimes n})) F_i^{\dagger}||\{Tr M_{xy}\sum_i q^{F_i} \ketbra{i}{i} \otimes F_i (\Lambda_0(\sigma^{\otimes n})) F_i^{\dagger}),\nonumber
\end{align}
where $p^{F_i}= Tr F_i (\Lambda_0(\rho^{\otimes n})) F_i^{\dagger}$ and $q^{F_i}= Tr F_i (\Lambda_0(\sigma^{\otimes n})) F_i^{\dagger}$.
Using the following property of relative entropy \cite{Piani2009-relent}:
\begin{equation}
D\de{\sum_i p_i \rho_i \otimes \ketbra{i}{i} \middle| \middle|\sum_i q_i \sigma_i \otimes \ketbra{i}{i}}=\sum_i p_i D(\rho_i||\sigma_i)+D(p||q),
\end{equation}
we obtain
\begin{align}
T_n(\rho^{\otimes n}) \geq {1\over n} \sup_{\Lambda_0\in LOCC} \sup_{F_0} \sup_{\{M_{xy}\}} \sup_{p(x,y)} \inf_{\sigma \in SEP} p^{F_0} \sum_{x,y} p(x,y) D(\{Tr M_{xy}F_0(\Lambda_0(\rho^{\otimes n}))F_0^{\dagger}\}||\{Tr M_{xy}F_0(\Lambda_0(\sigma^{\otimes n}))F_0^{\dagger}\}).
\end{align}
where we have dropped the terms $D(p||q)\geq 0$, $\sum_{F_i\neq F_0} p^{F_i} D(\rho_i||\sigma_i) \geq 0$.
Now note that the RHS is an upper bound for $R_H$ and then we have the desired result.
\qed
\end{document}
|
\section{Introduction}
In the \ProblemName{Channel Assignment} problem, we are given a symmetric weight function $w:V^2\to\mathbb{N}$
(we assume that $0\in\mathbb{N}$).
The elements of $V$ will be called vertices (as $w$ induces a graph on the
vertex set $V$ with edges corresponding to positive values of $w$).
We say that $w$ is $\ell$-bounded when for every $x, y\in V$ we have
$w(x, y) \le \ell.$
An assignment $c:V\to\mathbb{Z}$ is called {\em proper} when for each pair of vertices
$x, y$ we have $|c(x) - c(y)|\ge w(x, y).$
The number $\left(\max_{v\in V} c(v) - \min_{v\in V} c(v) + 1\right)$ is called
the {\em span} of $c.$
The goal is to find a proper assignment of minimum span.
Note that the special case when $w$ is $1$-bounded corresponds to the classical
graph coloring problem.
It is therefore natural to associate the instance of the channel assignment
problem with an edge-weighted graph $G=(V, E)$ where
$E=\set{uv\ :\ w(u, v) > 0}$ with edge weights $w_E:E\to\mathbb{N}$ such that
$w_E(xy) = w(x, y)$ for every $xy\in E$ (in what follows we abuse the notation
slightly and use the same letter $w$ for both the function defined on $V^2$
and $E$).
The minimum span is called also the span of $(G, w)$ and denoted by
${\rm span}(G, w).$
It is interesting to realize the place \ProblemName{Channel Assignment} in a kind of hierarchy of constraint
satisfaction problems.
We have already seen that it is a generalization of the classical graph
coloring.
It is also a special case of the constraint satisfaction problem (CSP).
In CSP, we are given a vertex set $V,$ a constraint set $\mathcal{C}$ and a
number of colors $d.$
Each constraint is a set of pairs of the form $(v, t)$ where $v\in V$ and
$t\in\set{1, \ldots, d}.$
An assignment $c:V\to\set{1, \ldots, d}$ is {\em proper} if every constraint
$A\in\mathcal{C}$ is satisfied, i.e.\ there exists $(v, t)\in A$ such that
$c(v)\ne t.$
The goal is to determine whether there is a proper assignment.
Note that \ProblemName{Channel Assignment} corresponds to CSP where $d=s$ and every edge $uv$ of weight
$w(uv)$ in the instance of \ProblemName{Channel Assignment} corresponds to the set of constraints of the
form $\set{(u, t_1), (v, t_2)}$ where $|t_1 - t_2|<w(uv).$
In the general case th best known algorithm runs in $O^*(n!)$ time
(see McDiarmid~\cite{mcdiarmid}).
However, there has been some progress on the $\ell$-bounded variant.
McDiarmid~\cite{mcdiarmid} came up with an $O^*((2\ell+1)^n)$-time algorithm
which has been next improved by Kral~\cite{kral} to $O^*((\ell+2)^n),$
further to $O^*((\ell+1)^n)$ by Cygan and Kowalik~\cite{cygan}
and to $O^*((2\sqrt{\ell + 1})^n)$ by Kowalik and Socala~\cite{meet}.
These are all dynamic programming (and hence exponential space) algorithms.
The last but one applies the fast zeta transform to get a minor speed-up
and the last one uses the meet-in-the-middle approach.
Interestingly, all these works show also algorithms which {\em count} all
proper assignments of span at most $s$ within the same running time
(up to polynomial factors) as the decision algorithm.
Since graph coloring is solvable in time $O^*(2^n)$~\cite{bhk:coloring}
it is natural to ask whether \ProblemName{Channel Assignment} is solvable in time $O^*(c^n),$ for some
constant $c.$
It is a major open problem (see~\cite{kral,cygan,dagstuhl}) to find such a
$O(c^n)$-time algorithm for $c$ independent of $\ell$ or prove that it does not
exist under a reasonable complexity assumption.
A complexity assumption commonly used in such cases is the Exponential Time
Hypothesis (ETH),
introduced by Impagliazzo, Paturi and Zane~\cite{PZ01}.
It states that \ProblemName{3-CNF-SAT} cannot be computed in time $2^{o(n)},$ where $n$ is
the number of variables in the input formula.
The open problem mentioned above becomes even more interesting when we realize
that under ETH, CSP does not have a $O^*(c^n)$-time algorithm for a constant
$c$ independent of $d,$ as proved by Traxler~\cite{traxler}.
\heading{Our Results}
Our main result is a proof that \ProblemName{Channel Assignment} does not admit a $O(c^n)$-time for a
constant $c$ under the ETH assumption.
By applying a sequence of reductions (see Figure~\ref{fig:sequence})
starting in \ProblemName{3-CNF-SAT} and ending in \ProblemName{Channel Assignment}
we were able to solve this open problem and to show that there is no
$2^{o(n\log n)}$-time algorithm solving \ProblemName{Channel Assignment} unless the ETH fails.
Note that the currently best known algorithm works in time
$O^*(n!) = 2^{O(n\log n)}$ so our lower bound is tight.
\newpage
\heading{\ProblemName{Common Matching Weight} as a generic problem without $2^{o(n\log n)}$-time algorithm}\\
In order to prove that there is no $2^{o(n\log n)}$-time algorithm for some
problem we may want to use a reduction from some better studied problem,
say from \ProblemName{3-CNF-SAT} for which we know that there is no $2^{o(n)}$-time
algorithm unless the ETH fails.
Therefore
in this case we need to be able to transform an instance of \ProblemName{3-CNF-SAT} of
size $n$ into an instance of our target problem of size $O\of{\ndivlog{n}}.$
Then $2^{o(n\log n)}$-time algorithm for our target problem would imply
$2^{o(n)}$-time algorithm for \ProblemName{3-CNF-SAT} which contradicts the ETH.
However such reductions which compress the size of the instance from
$O(n)$ to e.g. $O\of{\frac{n}{\log n}}$ are very rare.
As shown in the Figure~\ref{fig:sequence} we do this for the problem \ProblemName{Common Matching Weight}
defined as follows:
\defproblem{\ProblemName{Common Matching Weight}}{
Two complete weighted bipartite graphs $G_1=(V_1\cup W_1, E, w_1)$
and $G_2=(V_2\cup W_2, E, w_2)$ such that $\pow{V_1} = \pow{W_1}$
and $\pow{V_2} = \pow{W_2}.$
The weight functions $w_1, w_2$ have nonnegative integer values.
}{
Are there two perfect matchings $M_1$ in $G_1$ and $M_2$ in $G_2$
such that $w_1\of{M_1} = w_2\of{M_2}?$
}
\newcommand{\stack}[2]{\begin{tabular}{c} #1 \\ #2 \\ \end{tabular}}
\begin{figure}[t]
\begin{center}
\begin{minipage}{0.14\textwidth}
\begin{center}
\ProblemName{3-CNF-SAT}\\
$O(n)$
\end{center}
\end{minipage}
$\rightarrow$
\begin{minipage}{0.16\textwidth}
\begin{center}
\ProblemName{Family Intersection}\\
$O(n)$
\end{center}
\end{minipage}
$\rightarrow$
\begin{minipage}{0.22\textwidth}
\begin{center}
\ProblemName{Common Matching Weight}\\
$O\of{\ndivlog{n}}$
\end{center}
\end{minipage}
$\rightarrow$
\begin{minipage}{0.25\textwidth}
\begin{center}
\ProblemName{Channel Assignment}\\
$O\of{\ndivlog{n}}$
\end{center}
\end{minipage}
\caption{
The sequence of the used reductions and the size of the instance.
The compression follows between \ProblemName{Family Intersection} and \ProblemName{Common Matching Weight}.
While the definition of \ProblemName{Family Intersection} is rather technical
the \ProblemName{Common Matching Weight} problem is quite natural and it can be used as a generic problem
without $2^{o(n\log n)}$-time algorithm.
}
\label{fig:sequence}
\end{center}
\vspace{-1cm}
\end{figure}
Note that in order to show that a new problem \ProblemName{P} does not
admit a $2^{o(n\log n)}$-time algorithm it suffices to give a linear reduction
from \ProblemName{Common Matching Weight} to \ProblemName{P}.
We shown such reduction for \ProblemName{Channel Assignment} and we hope that the same thing can be done
also for other problems.
\heading{Organization of the paper}
In Section~\ref{sec:CMW} we describe a sequence of reductions
starting in \ProblemName{3-CNF-SAT} and ending in \ProblemName{Common Matching Weight} and the conclusions from
the existence of these reductions leading to the theorem on the hardness
of \ProblemName{Common Matching Weight}.
In Section~\ref{sec:CA} we present a reduction from \ProblemName{Common Matching Weight} to \ProblemName{Channel Assignment} and prove
the hardness of \ProblemName{Channel Assignment}
\heading{Notation}
Throughout the paper $n$ denotes the number of the vertices of the graph under
consideration.
For an integer $k,$ by $\el{k}$ we denote the set $\set{1, 2, \ldots, k}.$
\section{Hardness of \ProblemName{Common Matching Weight}}
\label{sec:CMW}
In this section
we describe a sequence of reductions starting in \ProblemName{3-CNF-SAT} and ending in \ProblemName{Common Matching Weight}
and the consequences of these reductions on the complexity of \ProblemName{Common Matching Weight}.
In the second of these two reductions we compress the instance from the size
$O(n)$ to the size $O\of{\ndivlog{n}}$ which is an important part of our
result.
\subsection{From \ProblemName{3-CNF-SAT} to \ProblemName{Family Intersection}.}
\label{subsec:ThreeSAT-to-FInt}
The intuition is that for a given instance of \ProblemName{3-CNF-SAT} we consider a set
of the \emph{occurrences} of the variables in the formula
i.e. we treat any two different
occurrences of the same variable as they were two different variables.
Note that in a \ProblemName{3-CNF-SAT} instance with $n$ variables and $m$ clauses we have
$3m$ occurrences of the $n$ variables so there are $2^{3m}$ assignments of the
occurrences.
We would like to represent two useful subsets of the set of all $2^{3m}$
assignments of the occurrences.
The first is the set of the \emph{consistent} assignments i.e.
such assignments of the occurrences that
all the occurrences of the same variable have the same value.
The second is the set of the assignments of the occurrences such that every
clause is satisfied
(although they are allowed to have different values for different occurrences
of the same variable i.e. they do not need to be consistent).
Note that the instance of \ProblemName{3-CNF-SAT} is a YES-instance if and only if
the intersection of these two sets is nonempty.
To represent those two sets we would like to use the following concept.
For a function $f:\feln{a}{b}$ we define
$X_f = \set{\sum_{i=1}^{a} f\of{i, \sigma\of{i}} | \sigma: \fel{a}{b}}.$
We call this set an \emph{$f$-family}.
We will define a function $f$ such that the elements of the $f$-family $X_f$
correspond to the assignments of the occurrences
such that every two occurrences of the same variable have the same value.
Then we define another function $g$ such that the
$g$-family $X_g$ represents the assignments of the
occurrences such that every clause is satisfied.
Thus we reduce \ProblemName{3-CNF-SAT} into the following problem:
\defproblem{\ProblemName{Family Intersection}}{
A function $f:\feln{a}{b}$ and a function $g:\feln{c}{d}.$
}{
Does $X_f \cap X_g \neq \emptyset$?
}\\
\begin{example} \label{exmp:TwoSAT-to-FInt}
Let us illustrate our approach on a \ProblemName{2-CNF-SAT} formula
$\varphi = (\alpha \vee \beta) \wedge (\neg \alpha \vee \gamma).$
We can make a distinction between different occurrences of the same variables
$\varphi' = (\alpha_1 \vee \beta_1) \wedge (\neg \alpha_2 \vee \gamma_1).$
So we have four occurrences of the variables and $2^4=16$ assignments.
We represent those assignments as
numbers from the set $\set{0, 1, \ldots, 2^4 - 1}$
However, it will be convenient to refer to these numbers as bit vectors
of length $4$ where
the $i$-th bit represents the value of the $i$-th occurrence
(among all the occurrences of all the variables).
To represent the set of the consistent assignments of the occurrences we can
use a function $f:\feln{n}{2}$ such that the value of $f(i, 1)$ is a bit vector
representing all the occurrences of the $i$-th variable and $f(i, 2) = 0.$
So in our example we have
$f(1, 1) = 1010_2,$ $f(2, 1) = 0100_2, $ and $f(3, 1) = 0001_2.$
Therefore
$X_f = \set{0000_2, 0001_2, 0100_2, 0101_2, 1010_2, 1011_2, 1110_2, 1111_2}.$
To represent the set of the assignments which satisfies all the clauses we
can use a function $g:\feln{m}{3}$ such that $g(i, j)$ is a $j$-th assignment
(in some fixed order)
of the occurrences of the variables in the $i$-th clause which satisfies
this clause.
Note that every clause in \ProblemName{2-CNF-SAT} have $3$ assignments of the occurrences
which satisfies this clause.
So in our example we have $g(1, 1) = 1000_2,$ $g(1, 2) = 0100_2,$
$g(1, 3) = 1100_2,$ $g(2, 1) = 0000_2,$ $g(2, 2) = 0001_2$
and $g(2, 3) = 0011_2.$
Therefore
$X_g = \set{0100_2, 0101_2, 0111_2, 1000_2, 1001_2, 1011_2, 1100_2, 1101_2,
1111_2}.$
The set $X_f \cap X_g = \set{0100_2, 0101_2, 1011_2, 1111_2}$ is the set
of all the consistent assignments of the occurrences such that each clause
is satisfied.
\end{example}
We can formalize our observation as following.
\begin{lemma} \label{lem:ThreeSAT-to-FInt} ($\bigstar$)
There is a polynomial time reduction from a given instance of \ProblemName{3-CNF-SAT} with
$n$ variables and $m$ clauses
into an instance of \ProblemName{Family Intersection} with $f:\feln{n}{2}$ and $g:\feln{m}{7}$
such that $\max{X_f} < 2^{3m}$ and $\max{X_g} < 2^{3m}.$
\end{lemma}
The proof is straightforward and its idea should be illustrated
by the Example~\ref{exmp:TwoSAT-to-FInt}.
It is moved to the Appendix due to space limitations.
\subsection{From \ProblemName{Family Intersection} to \ProblemName{Common Matching Weight}.}
\label{subsec:FInt-to-CMW}
Consider an $f$-family $X_f$
and a $g$-family $X_g$ for some functions $f:\feln{n}{2}$ and $g:\feln{m}{7}.$
In this section we show how to encode $X_f$ in some weighted bipartite
graph $G_1$ so that the set of the weights of the perfect matchings in $G_1$
will be equal to $X_f.$
Similarly we will encode $X_g$ in some bipartite graph $G_2$
such that the set of the weights of the perfect matchings in $G_2$
will be equal to $X_g.$
So the set $X_f \cap X_g$ is nonempty if and only if
$G_1$ and $G_2$ contain perfect matchings with the same weight.
Moreover the number of the vertices of the graph $G_1$ will be
$O\of{\ndivlog{n}}$ and the number of the vertices of the graph $G_2$
will be $O\of{\ndivlog{m}}.$
This is a crucial step of our construction, because the instance size
decreases (by a logarithmic factor).
\newcommand{\h}[1]{\hat{#1}}
\newcommand{\h{\N}}{\h{\mathbb{N}}}
\newcommand{\hel}[1]{\h{\el{#1}}}
Before we describe the reduction we need the following technical lemma which
describe constructions of permutations of some specified
properties.
The permutations correspond naturally to perfect matchings in bipartite graphs.
Elements of $\el{k}^b$ will be treated as $b$-character words over alphabet
$\el{k},$ i.e. for $x\in\el{k}$ and $w\in[k]^b$ by $xw$ we mean the word of
length $b + 1$ obtained by concatenating $x$ and $w.$
For convenience we define a set $\h{\N} = \set{\h{0}, \h{1}, \h{2}, \ldots}$
as a copy of the natural numbers $\mathbb{N}$
and for every $n\in\mathbb{N}$ we define $\hel{n} = \set{\h{1}, \h{2}, \ldots, \h{n}}.$
Every set $\hel{k}^b$ is just a copy of $\el{k}^b$
so we refer to bijections between $\el{k}^b$ and $\hel{k}^b$
as to permutations.
\begin{lemma} \label{lem:bijection} ($\bigstar$)
Let $b\in \mathbb{N}$ and $\alpha:\hel{k}^b \times \el{b} \to \el{k} \cup \set{\bot}$
such that for every $\h{w}\in\hel{k}^b$ and for every $i\in\el{b}$
holds $\alpha\of{\h{w}, i} \neq \bot$ if and only if $\h{w}_i = \h{1}.$
There is a permutation $\phi:\hel{k}^b \to \el{k}^b$
such that for every $\h{w}\in\hel{k}^b$ and for every $i\in\el{b}$
if $\h{w}_i = \h{1}$ then $\phi\of{\h{w}}_i = \alpha\of{\h{w}, i}.$
\end{lemma}
Now we can describe the reduction.
\begin{lemma} \label{lem:family-to-matchings}
For a function $f:\feln{n}{k}$ there is a full bipartite graph
$G=\p{V_1 \cup V_2, E, w}$ such that
\begin{itemize}
\item for every $x\in X_f$ there exists a perfect matching $M$ of $G$
such that $w(M) = x,$
\item for every perfect matching $M$ of $G$ we have $w(M) \in X_f,$
\item $\pow{V_1} = \pow{V_2} = O\of{\frac{nk^2\log{k}}{\log{n} + \log{k}}}.$
\end{itemize}
\end{lemma}
\newcommand{\red}[1]{{\color{red}#1}}
\newcommand{\green}[1]{{\color{green}#1}}
\newcommand{\blue}[1]{{\color{blue}#1}}
\newcommand{\cyan}[1]{{\color{cyan}#1}}
\begin{figure}[t]
\begin{center}
\begin{tabular}{c|c|c|c|c|}
&
$\bt{\h{1}_{\red{(1)}}, \h{1}_{\blue{(2)}}}$ &
$\bt{\h{1}_{\green{(3)}}, \h{2}}$ &
$\bt{\h{2}, \h{1}_{\cyan{(4)}}}$ &
$\bt{\h{2}, \h{2}}$ \\
\hline
$\bt{1_{(i)}, 1_{(ii)}}$ &
$f\of{\red{1}, 1_{(i)}) + f(\blue{2}, 1_{(ii)}}$ &
$f\of{\green{3}, 1_{(i)}}$ &
$f\of{\cyan{4}, 1_{(ii)}}$ &
$0$ \\
\hline
$\bt{1_{(i)}, 2_{(ii)}}$ &
$f\of{\red{1}, 1_{(i)}) + f(\blue{2}, 2_{(ii)}}$ &
$f\of{\green{3}, 1_{(i)}}$ &
$f\of{\cyan{4}, 2_{(ii)}}$ &
$0$ \\
\hline
$\bt{2_{(i)}, 1_{(ii)}}$ &
$f\of{\red{1}, 2_{(i)}) + f(\blue{2}, 1_{(ii)}}$ &
$f\of{\green{3}, 2_{(i)}}$ &
$f\of{\cyan{4}, 1_{(ii)}}$ &
$0$ \\
\hline
$\bt{2_{(i)}, 2_{(ii)}}$ &
$f\of{\red{1}, 2_{(i)}) + f(\blue{2}, 2_{(ii)}}$ &
$f\of{\green{3}, 2_{(i)}}$ &
$f\of{\cyan{4}, 2_{(ii)}}$ &
$0$ \\
\hline
\end{tabular}
\caption{
The weights on the edges
of a graph encoding an $f$-family for $f:\feln{4}{2}.$
The lower indices $(1), (2), (3)$ and $(4)$
are added to indicate the correspondence between
the occurrences of $\h{1}$ and the elements of $\el{n}$
(the first argument of the function $f$).
The lower indices $(i)$ and $(ii)$ are added to indicate the correspondence
between the second argument of the function $f$ and the position in the
(two element) sequence $\langle \cdot, \cdot \rangle.$
}
\label{fig:k44-tabular}
\end{center}
\vspace{-1cm}
\end{figure}
\begin{proof}
Let us consider the smallest $b\in\mathbb{N}_+$ such that $c = b \cdot k^{b-1} \geq n.$
Later we will show that $\pow{V_1} = \pow{V_2} = k^b$ is sufficient.
For convenience we extend our chosen function $f:\feln{n}{k}$
to $f:\feln{c}{k}$ in such a way that
for every $i = n + 1, n + 2, \ldots, c$ and for every $j\in\el{k}$
we put $f\of{i, j} = 0.$
Note that the $f$-family $X_f$
does not change after this extension.
Let $V_1 = \hel{k}^b$ and $V_2 = \el{k}^b$ be the sets of words of length
$b$ over the alphabets respectively $\hel{k}$ and $\el{k}.$
Note that $\pow{V_1} = \pow{V_2} = k^b.$
Let $\beta:V_1\times\el{b}\to\el{c}\cup\set{\bot}$
be any function such that
if $\h{w}_j \neq \h{1}$ then $\beta\of{\h{w}, j} = \bot$
and every value from the set $\el{c}$ is used exactly once,
i.e.,
for every $x\in\el{c}$ there is exactly one argument
$(\h{w}, j) \in V_1 \times \el{b}$
such that
$\beta\of{\h{w}, j} = x.$
Note that such a function always exists because the total number of the
occurrences of $\h{1}$ in all the words in $V_1$
is exactly $c = b \cdot k^{b-1}.$
Now we define our weight function $w:V_1\times V_2\to\mathbb{N}$
as follows
\[
w\of{\h{t}, u} =
\sum_{\substack{i\in\el{b} \\ \beta\of{\h{t}, i}\neq\bot}}
f\of{\beta\of{\h{t}, i}, u_i}.
\]
An example of such weight function can be found in
Figure~\ref{fig:k44-tabular}
(or in the Appendix in Figure~\ref{fig:k44} as a picture of a bipartite
graph).
Note that because $\beta$ picks every value from the set $\el{c}$
exactly once then for every permutation $\phi:V_1\to V_2$ we have
\[
\sum_{\h{t}\in V_1} w\of{\h{t}, \phi\of{\h{t}}} =
\sum_{\h{t}\in V_1}
\sum_{\substack{i\in\el{b} \\ \beta\of{\h{t}, i}\neq\bot}}
f\of{\beta\of{\h{t}, i}, \phi\of{\h{t}}_i}
\in X_f.
\]
In other words the set of the weights of all perfect matchings in $G$
is a subset of $X_f$ as required.
We also need to show that for every $x \in X_f$ there exists some
permutation $\phi:V_1\to V_2$ such that
$\sum_{\h{t}\in V_1} w\of{\h{t}, \phi\of{\h{t}}} = x.$
This permutation gives us a corresponding perfect matching of weight $x$
in $G.$
Let us take a function $\sigma:\fel{c}{k}$ such that
$x = \sum_{i\in\el{c}} f\of{i, \sigma\of{i}},$
which exists by the definition of $X_f.$
Define $\alpha:\hel{k}^b\times\el{b}\to\el{k}\cup\set{\bot}$ as follows:
\[
\alpha\of{\h{u}, i} =
\begin{cases}
\sigma\of{\beta\of{\h{u}, i}} & \mbox{for } \h{u}_i = \h{1}\\
\bot & \mbox{for } \h{u}_i \neq \h{1}.\\
\end{cases}
\]
Now we can use Lemma~\ref{lem:bijection} with function $\alpha$
to obtain a permutation $\phi:\hel{k}^b \to \el{k}^b$ such that
for every $\h{u}\in\hel{k}^b$ and for every $i\in\el{b}$
if $\h{u}_i = \h{1}$ then $\phi\of{\h{u}}_i = \sigma\of{\beta\of{\h{u}, i}}.$
So we have that
\[
\sum_{\h{u}\in\hel{k}^b} w\of{\h{u}, \phi\of{\h{u}}} =
\sum_{\h{u}\in\hel{k}^b}
\sum_{\substack{i\in\el{b} \\ \beta\of{\h{u}, i}\neq\bot}}
f\of{\beta\of{\h{u}, i}, \sigma\of{\beta\of{\h{u}, i}}}
= \sum_{i\in\el{c}} f\of{i, \sigma\of{i}}
= x.
\]
Hence we have shown that
$X_f$ is the set of weights of all perfect matchings in graph $G.$
The last thing is to show that the number of the vertices
is sufficiently small.
We know that $\p{b - 1} \cdot k^{b - 2} < n$ so
$\p{b - 1} \cdot k^{b - 1} < nk$
and then
$b - 1 < \log_{k}{\rnlogkn{nk}{k}}.$
Therefore
$k^b
< k \cdot \rnlogkn{nk}{k}
= O\of{\frac{nk^2}{\log_k{nk}}}
= O\of{\frac{nk^2\log{k}}{\log{n} + \log{k}}}.$
So $|V_1| = |V_2| = k^b = O\of{\frac{nk^2\log{k}}{\log{n} + \log{k}}},$
as required.
\end{proof}
Lemma~\ref{lem:family-to-matchings} immediately implies the following result.
\begin{lemma} \label{lem:FInt-to-CMW}
There is a polynomial time reduction that
for an instance $I = (f, g)$ of \ProblemName{Family Intersection} with
$f:\feln{a}{b}$ and $g:\feln{c}{d}$ reduces it into
an instance of \ProblemName{Common Matching Weight} $J = (G_1, G_2)$ with
$\pow{V\of{G_1}} = O\of{\frac{a b^2 \log{b}}{\log{a} + \log{b}}}$
and $\pow{V\of{G_2}} = O\of{\frac{c d^2 \log{d}}{\log{c} + \log{d}}}$
vertices.
The sets of the weights of all perfect matchings in $G_1$ and in $G_2$
are equal respectively to $X_f$ and $X_g.$
\end{lemma}
Together with Lemma~\ref{lem:ThreeSAT-to-FInt} we obtain the following
theorem.
\begin{theorem} \label{thm:ThreeSAT-to-CMW}
There is a polynomial time reduction from a given instance of \ProblemName{3-CNF-SAT} with
$n$ variables and $m$ clauses
into an instance of \ProblemName{Common Matching Weight} with $|V(G_1)| = O\of{\ndivlog{n}},$
$|V(G_2)| = O\of{\ndivlog{m}}$
and the maximum matching weights bounded by $2^{3m}.$
\end{theorem}
A commonly know corollary of the Sparsification Lemma from \cite{PZ01} is:
\begin{corollary} \label{cor:ThreeSAT-hardness}
There is no algorithm solving \ProblemName{3-CNF-SAT} in $2^{o(n + m)}$-time
where $n$ is the number of variables
and $m$ is the number of clauses
unless ETH fails.
\end{corollary}
Using Corollary~\ref{cor:ThreeSAT-hardness}
we can prove the following lower bound.
\begin{corollary} \label{cor:CMW-hardness}
There is no algorithm solving \ProblemName{Common Matching Weight} in $2^{o(n\log n)}\poly{r}$-time
where $n$ is the total number of vertices,
and $r$ is the bit size of the input,
unless ETH fails.
\end{corollary}
\begin{proof}
For a given instance of \ProblemName{3-CNF-SAT} with $n$ variables and $m$ clauses we can use
the reduction from Theorem~\ref{thm:ThreeSAT-to-CMW} to obtain an instance of
\ProblemName{Channel Assignment}.
The total number of the vertices in the new instance is
$$|V(G_1)| + |V(G_2)| = O\of{\ndivlog{n} + \ndivlog{m}}
= O\of{\frac{n + m}{\log\of{n + m}} + \frac{n + m}{\log\of{n + m}}}
= O\of{\frac{n + m}{\log\of{n + m}}}$$
because the function $\ndivlog{n}$ is nondecreasing for the sufficiently big
values of $n.$
Weights of the matchings are bounded by $2^{3m}$ and the bit size
of the instance
$$r = O\of{\p{\frac{n + m}{\log\of{n + m}}}^2 \log 2^{3m}} = \poly{nm}.$$
Then let us assume that there is an algorithm solving \ProblemName{Common Matching Weight} in
$2^{o(n\log n)} \poly{r}$-time.
Then we could solve our instance in time
$$2^{o\of{\frac{n + m}{\log(n + m)}
\log\of{\frac{n + m}{\log(n + m)}}}} \poly{\poly{nm}}
= 2^{o\of{\frac{n + m}{\log(n + m)}
\log(n + m)}} \poly{nm}
= 2^{o(n + m)}
$$
which contradicts ETH by Corollary~\ref{cor:ThreeSAT-hardness}.
\end{proof}
\section{Hardness of \ProblemName{Channel Assignment}}
\label{sec:CA}
Consider two weighted
full bipartite graphs $G_1$ and $G_2.$ We would like to encode them
in a \ProblemName{Channel Assignment} instance in such a way that this \ProblemName{Channel Assignment} instance is a YES-instance
if and only if there are two perfect matchings,
one in $G_1$ and the other in $G_2,$ of the same weight.
Consider an instance $I=(V,d,s)$ of \ProblemName{Channel Assignment}.
We say that $c:V\to\mathbb{Z}$ is a YES-coloring if $c$ is a proper coloring and has
span at most $s.$
Note that an instance of \ProblemName{Channel Assignment} is a YES-instance
if and only if it has a YES-coloring.
Our approach is that we encode those graphs $G_1$ and $G_2$
separately in such a way
that we have a special vertex $v_M$ whose color in every YES-coloring
represents a weight of some perfect matching in $G_1$
and on the other hand in every YES-coloring
its color represents (in a similar way) a weight
of some perfect matching in $G_2.$
So a YES-coloring
coloring would be possible if and only if the graphs $G_1$ and $G_2$
have two perfect matchings, one in $G_1$ and the other in $G_2$, with equal
weights.
Before we present a way to encode a weighted full bipartite graph in a
\ProblemName{Channel Assignment} instance we would like to present the two lemmas to merge those two encoded
graphs into a one instance of \ProblemName{Channel Assignment}.
In order to do that we use the following concepts.
We say that instance $I$ is \emph{$(x, y)$-spanned}
for some vertices $x, y\in V$
if for every YES-coloring $c$ of $I$
we have $\abs{c\of{x} - c\of{y}} = s - 1.$
We say that an instance $I = (V, d, s)$ of \ProblemName{Channel Assignment} is \emph{$(X, Y)$-spanned}
for some nonempty subsets of the vertices $\emptyset\neq X, Y\subseteq V$
if it is $(x, y)$-spanned for every two vertices $x\in X$ and $y\in Y.$
\begin{lemma} \label{lem:ca-merge} ($\bigstar$)
For every $(u, v)$-spanned instance $I_1=(V_1, d_1, s)$ and
$(w, z)$-spanned instance $I_2=(V_2, d_2, s)$
of \ProblemName{Channel Assignment}
there is a $(\set{u,w}, \set{v,z})$-spanned instance $I=(V_1 \cup V_2, d, s)$
of \ProblemName{Channel Assignment}
such that
\begin{enumerate}[(i)]
\item for every YES-coloring $c$ of $I$
the coloring $c\mid_{V_1}$ is a YES-coloring of $I_1$
and the coloring $c\mid_{V_2}$ is a YES-coloring of $I_2,$
\item for every YES-coloring $c_1$ of $I_1$
and every YES-coloring $c_2$ of $I_2$
such that $c_1\of{u} = c_2\of{w},$ $c_1\of{v} = c_2\of{z}$ and
for every $x\in V_1 \cap V_2$ we have $c_1\of{x} = c_2\of{x}$
there exists a YES-coloring $c$ of $I$
such that $c\mid_{V_1} = c_1$ and $c\mid_{V_2} = c_2.$
\end{enumerate}
\end{lemma}
\begin{lemma} \label{lem:ca-extend} ($\bigstar$)
For every $(v_L, v_R)$-spanned instance $I=(V, d, s)$ of \ProblemName{Channel Assignment} and
for every numbers $l, r \in \mathbb{N}$ there exists
a $(w_L, w_R)$-spanned instance $I'=(V \cup \set{w_L, w_R}, d', l + s + r)$
such that
\begin{enumerate}[(i)]
\item for every YES-coloring $c$ of $I$
there is a YES-coloring $c'$ of $I'$ such that $c'\mid_V = c,$
\item for every YES-coloring $c'$ of $I'$ such that
$c'\of{w_L} \leq c'\of{w_R}$ we have that
\begin{itemize}
\item a coloring $c'\mid_V$ is a YES-coloring of $I,$
\item $c'\of{v_L} = c'\of{w_L} + l$ and
$c'\of{v_R} = c'\of{w_R} - r.$
\end{itemize}
\end{enumerate}
\end{lemma}
The proofs of these two lemmas are straightforward.
They are moved to the Appendix due to space limitations.
\begin{lemma} \label{lem:matchings-to-ca}
Let $G = (V_1 \cup V_2, E, w)$ be a weighted full bipartite graph
with nonnegative weights and such that $\pow{V_1} = \pow{V_2}.$
Let $n = \pow{V_1},$
$m = \max_{e\in E} w(e),$
$M = n \cdot m + 1,$
$l = \p{4n - 1} \cdot M$ and
$s = (8n - 1) \cdot M.$
There exists a $(v_L, v_R)$-spanned instance $I=(V, d, s)$ of \ProblemName{Channel Assignment}
with $\pow{V} = O\of{n}$ and
such that for some vertex $v_M \in V,$
\begin{enumerate}[(i)]
\item for every YES-coloring $c$ of $I$
such that $c\of{v_L} \leq c\of{v_R}$
there exists a perfect matching $M_G$ in $G$ such that
$c\of{v_M} = c\of{v_L} + l + w\of{M_G}$ and
\item for every perfect matching $M_G$ in $G$ there exists a YES-coloring
$c$ of $I$ such that $c\of{v_L} \leq \of{v_R}$ and
$c\of{v_M} = c\of{v_L} + l + w\of{M_G}.$
\end{enumerate}
\end{lemma}
\begin{figure}[t]
\begin{center}
\def5in{5in}
\input{matching-to-ca.pdf_tex}
\vspace{0.2cm}
\caption{
A weighted full bipartite graph $(G, w)$ with $|V_1| = |V_2| = 2$
encoded in a \ProblemName{Channel Assignment} form.
The color of the vertex $v_M = w_2$ corresponds to the weight of the
perfect matching in $G$ given by the permutation $\pi$
and is equal to $c\of{v_M} = c\of{v_L} + 7M + w\of{M_{\pi}}.$
The picture is simplified.
Some of the edges and corresponding to them minimum distances
are omitted in the picture.
}
\label{fig:matching-to-ca}
\end{center}
\vspace{-1cm}
\end{figure}
\newcommand{\vind}[3]{{#1}^{\p{#2}}_{#3}}
\newcommand{\vv}[2]{\vind{v}{#1}{#2}}
\newcommand{\ww}[2]{\vind{w}{#1}{#2}}
\newcommand{4n}{4n}
\newcommand{\seq{v}{\maxv}}{\seq{v}{4n}}
\newcommand{\el{\maxv}}{\el{4n}}
\newcommand{2n - 1}{2n - 1}
\newcommand{\seq{w}{\maxw}}{\seq{w}{2n - 1}}
\newcommand{\el{\maxw}}{\el{2n - 1}}
\newcommand{n}{n}
\newcommand{\seq{a}{\maxa}}{\seq{a}{n}}
\newcommand{\el{\maxa}}{\el{n}}
\newcommand{n}{n}
\newcommand{\seq{b}{\maxb}}{\seq{b}{n}}
\newcommand{\el{\maxa}}{\el{n}}
\begin{proof}
Let $V_1 = \set{\vv{1}{1}, \vv{1}{2}, \ldots, \vv{1}{n}}$
and $V_2 = \set{\vv{2}{1}, \vv{2}{2}, \ldots, \vv{2}{n}}.$
We will build our \ProblemName{Channel Assignment} instance step by step.
A simplified picture of
the instance can be found in Figure~\ref{fig:matching-to-ca}.
Let us introduce the vertices $v_L = \seq{v}{\maxv} = v_R$ to set $V.$
Because of the symmetry we can assume that for every coloring $c$ of our
instance we have $c\of{v_L} \leq c\of{v_R}.$
We set the minimum distance $d\of{v_L, v_R} = s - 1.$
Then for every YES-coloring $c$ we have that
$\abs{c\of{v_L} - c\of{v_R}} = s - 1$ so our instance is $(v_L, v_R)$-spanned.
For every $i, j \in \el{\maxv}$ such that $i \neq j$ and
$\set{i, j} \neq \set{1, 4n}$
we set the minimum distance
$d\of{v_i, v_j} = \abs{i - j} \cdot 2M.$
Then we can prove the following claim.
{\bf Claim 1}
For every YES-coloring $c$ and for every
$i < j$ we have that $c\of{v_i} < c\of{v_j}.$
{\em Proof of the claim:}
Indeed if the colors $c\of{v_1}, c\of{v_2}, \ldots, c\of{v_{4n}}$
are not strictly increasing or strictly decreasing then the distances
between the consecutive colors of the set $c\of{\set{\seq{v}{\maxv}}}$
are at least $2M$ and at least one of them is at least $4M.$
So it would need to use at least
$\p{4n - 2} \cdot 2M + 4M + 1 > (8n - 1) \cdot M = s$
colors.
In addition we have assumed that for all colorings
$c\of{v_L} \leq c\of{v_R}$ so the colors are strictly increasing.
This proves the claim.
Note that for every YES-coloring and for every $i \in \el{4n - 1}$
we have
\begin{equation} \label{eq:vdists}
2M \leq c\of{v_{i + 1}} - c\of{v_i} \leq 2M + n\cdot m < 3M,
\end{equation}
for otherwise $c(v_R) - c(v_L) \ge (4n - 2) \cdot 2M + 2M + n \cdot m + 1 = s,$
so $c$ has the ${\rm span}$ at least $s + 1,$ a contradiction.
Let us introduce new vertices $\seq{w}{\maxw}$ to set $V.$
For every $i\in\el{\maxw}$ and $j\in\el{\maxv}$
we set the minimum distances
$d\of{w_i, v_j} = \abs{4i + 1 - 2j} \cdot M.$
For every YES-coloring $c$ and for every $i\in\el{\maxw}$
we have
$c\of{v_{2i}} + M \leq c\of{w_i} \leq c\of{v_{2i + 1}} - M$
by \eqref{eq:vdists},
for otherwise we have that $c(v_j)\leq c(w_i) \leq c(v_{j + 1})$
for some $j \neq 2i$
(because $c(v_{4n}) - c(v_1) = s - 1$ so every YES-coloring uses only
colors from the interval $[c(v_1), c(v_{4n})]$)
and then
$c(v_{j + 1}) - c(v_j) \ge d(v_j, w_i) + d(w_i, v_{j + 1})$
and $\{v_j, v_{j + 1}\} \ne \{v_{2i}, v_{2i + 1}\}$
so at least one of these two distances is at least $3M$
and therefore $c(v_{j + 1}) - c(v_j) \ge 3M + M > 3M,$ a contradiction
with \eqref{eq:vdists}.
Thus infer the following claim.
{\bf Claim 2} For every YES-coloring $c$ the colors of the vertices in the
sequence
\begin{equation} \label{eq:vwseq}
v_1, v_2, w_1, v_3, v_4, w_2, v_5\ldots,
v_{4n - 2}, w_{2n - 1}, v_{4n - 1}, v_{4n}
\end{equation}
are increasing.
We introduce new vertices $\seq{a}{\maxa}$ and
for every $i\in\el{\maxa}$ and $j\in\el{\maxv}$ we set the minimum distances
$$d\of{a_i, v_j} = \begin{cases}
M + w\of{\vv{1}{i}, \vv{2}{j/2}} &
\mbox{when } j \leq 2n \mbox{ and } 2 \mid j\\
M & \mbox{when } j \leq 2n \mbox{ and } 2 \nmid j\\
\p{j - 2n} \cdot 2M + M & \mbox{when } j > 2n.\\
\end{cases}$$
Then for every YES-coloring $c$ and for every $i\in\el{\maxa}$ we have
$c\of{a_i} \leq c\of{v_{2n}}$
because in other case we have $c(v_j) \leq c(a_i) \leq c(v_{j + 1})$
for some $j \ge 2n$ and then
$c(v_{j + 1}) - c(v_j) \ge d(v_j, a_i) + d(a_i, v_{j + 1}) \ge M + 3M > 3M,$
a contradiction with \eqref{eq:vdists}.
Moreover for every $i\in\el{\maxa}$ and every $j\in\el{\maxw}$ we set the minimum distance
$d\of{a_i, w_j} = 2M.$
Therefore by \eqref{eq:vdists} and \eqref{eq:vwseq}
for every YES-coloring $c$ and every $i\in\el{\maxa}$ the vertex $a_i$
is colored with the color from one of the intervals
$\p{c\of{v_{2j - 1}}, c\of{v_{2j}}}$ for some $j \in \el{n}.$
Finally for every $i, j \in \el{\maxa}$ such that $i \neq j$
we set the minimum distance
$d\of{a_i, a_j} = 4M$ so by \eqref{eq:vdists} we know that
for every YES-coloring $c$ and every $i\in \el{n}$
exactly one one vertex $a_j$ of the vertices $\seq{a}{\maxa}$ is colored with the
color from the interval $\p{c\of{v_{2i - 1}}, c\of{v_{2i}}}.$
The assignment of vertices $\seq{a}{\maxa}$ to intervals
$(c(v_1), c(v_2)), (c(v_3), c(v_4)), \ldots, (c(v_{2n - 1}), c(v_{2n}))$
determines a permutation $\pi_c:[n]\to[n],$ i.e., $\pi_c(i) = j$
if $a_j$ gets a color from $(c(v_{2i - 1}), c(v_{2i})).$
Hence we get the following claim:
{\bf Claim 3} For every YES-coloring $c$ there is a permutation $\pi_c$ such
that the colors of the vertices of the sequence
$$v_1, a_{\pi_c\of{1}}, v_2, w_1, v_3, a_{\pi_c\of{2}}, v_4, w_2, v_5 \ldots,
v_{2n - 1}, a_{\pi_c\of{n}}, v_{2n}$$
are increasing.
Similarly we introduce new vertices $\seq{b}{\maxb}$ and
for every $i\in\el{\maxa}$ and $j\in\el{\maxv}$ we set the minimum distances
$$d\of{b_i, v_j} = \begin{cases}
\p{2n - j + 1} \cdot 2M + M & \mbox{when } j \leq 2n \\
M + m - w\of{\vv{1}{i}, \vv{2}{j/2 - n}} &
\mbox{when } j > 2n \mbox{ and } 2 \mid j \\
M & \mbox{when } j > 2n \mbox{ and } 2 \nmid j.\\
\end{cases}$$
Also for every $i\in\el{\maxa}$ and every $j\in\el{\maxw}$ we set the minimum distance
$d\of{b_i, w_j} = 2M$ and for every $i, j \in \el{\maxa}$ such that $i \neq j$
we set the minimum distance $d\of{b_i, b_j} = 4M.$
Hence similarly as before,
for every YES-coloring $c$ and every $i\in\el{n}$ exactly one
vertex $b_j$ of the vertices $\seq{b}{\maxb}$ is colored with the color from
the interval $\p{c\of{v_{2n + 2i - 1}}, c\of{v_{2n + 2i}}}.$
Analogously as before, the colors of the vertices $\seq{b}{\maxb}$
determine a permutation $\rho_c:[n]\to[n].$
Thus we have the following claim.
{\bf Claim 4} For every YES-coloring $c$ there is a permutation $\rho_c$
such that the colors of the vertices in the sequence
$$v_{2n + 1}, b_{\rho_c\of{1}}, v_{2n + 2}, w_{n + 1},
v_{2n + 3}, b_{\rho_c\of{2}}, v_{2n + 4}, w_{n + 2}, v_{2n + 5}\ldots
v_{4n - 1}, b_{\rho_c\of{n}}, v_{4n}$$
are increasing.
For every $i\in\el{n}$ we set the minimum distance
$d\of{a_i, b_i} = n \cdot 4M.$
Then we know that for every YES-coloring $c$ we have
$\pi_c^{-1}\of{i} \leq \rho_c^{-1}\of{i}$
for otherwise we can take
$j = 2\pi_c^{-1}(i) - 1$ and
$k = 2n + 2\rho_c^{-1}(i)$
and then
$(c(b_i) - c(a_i)) + 2M \le c(v_j) - c(v_k)$
and $k - j \le 2n$
so the sequence $v_1, v_2, \ldots, v_j, v_k, \ldots, v_{4n}$
has at least $4n - 2n + 1 = 2n + 1$ elements
so
$ c(v_{4n}) - c(v_1)
\ge (2n - 1) \cdot 2M + (c(v_k) - c(v_j))
\ge (2n - 1) \cdot 2M + (c(b_i) - c(a_i)) + 2M
\ge (2n - 1) \cdot 2M + n \cdot 4M + 2M
= n \cdot 8M
> (n - 1) \cdot 8M - 1
= s - 1,$
a contradiction.
Since $\pi_c$ and $\rho_c$ are permutations, we further infer that for every
YES-coloring $c$ we have $\pi_c = \rho_c.$
Hence we have the following claim.
{\bf Claim 5} For every YES-coloring $c$ there is a permutation $\pi_c$
such that the colors of the vertices in the sequence
$$v_1, a_{\pi_c\of{1}}, v_2, w_1, v_3, a_{\pi_c\of{2}}, v_4, w_2, v_5 \ldots,
v_{2n - 1}, a_{\pi_c\of{n}}, v_{2n}, w_n,$$ $$
v_{2n + 1}, b_{\pi_c\of{1}}, v_{2n + 2}, w_{n + 1},
v_{2n + 3}, b_{\pi_c\of{2}}, v_{2n + 4}, w_{n + 2}, v_{2n + 5}\ldots
v_{4n - 1}, b_{\pi_c\of{n}}, v_{4n}$$
are increasing.
This ends the description of the instance $I.$
Note that $I$ is $(v_L, v_R)$-spanned because $d(v_L, v_R) = s - 1.$
Let us put $v_M = w_n.$
We are going to show the following claim.
{\bf Claim 6} ($\bigstar$)
Let $\pi:\el{n}\to\el{n}$ be any permutation and
$M_{\pi} = \set{\vv{1}{i} \vv{2}{\pi\of{i}} : i\in\el{n}}$ be
the corresponding perfect matching in $G.$
There is exactly one YES-coloring $c$ such that
$\pi_c = \pi.$
Moreover $c(v_M) = c(v_L) + l + w(M_{\pi}).$
To proove the claim it is sufficient to check all the introduced
minimum allowed distances $d$ and the ${\rm span}$
for the coloring implied by the sequence as in Claim 5.
This is a simple manual check but due to its length
the proof of the claim is moved to the Appendix.
Thus there is a one-to-one correspondence between permutations and
YES-colorings.
Moreover we know that for every YES-coloring $c$ we have
$c\of{v_M} = c\of{v_L} + l + w\of{M_{\pi_c}}$
where $M_{\pi_c}$ is the perfect matching in $G$ corresponding
to permutation $\pi_c.$
Hence we have shown (i) and (ii) as required.
\end{proof}
\begin{lemma} \label{lem:CMW-to-CA} ($\bigstar$)
There is a polynomial time reduction such that
for a given instance $I=(G_1, G_2)$ of \ProblemName{Common Matching Weight}
with $n_1 = \pow{V\of{G_1}}, n_2 = \pow{V\of{G_2}}$
and such that the weight functions of $G_1$ and $G_2$
are bounded by respectively $m_1$ and $m_2$
reduces it into an instance of \ProblemName{Channel Assignment} with
$O\of{n_1 + n_2}$ vertices
and the maximum edge weight in $O\of{n_1^2 m_1 + n_1^2 m_2}.$
\end{lemma}
\begin{figure}[t]\begin{center}
\def5in{5in}
\input{matchings-to-ca.pdf_tex}
\caption{
Two weighted full bipartite graphs $(G_1, w_1)$ (with $n_1 = 2$)
and $(G_2, w_2)$ (with $n_2 = 3$)
encoded in a \ProblemName{Channel Assignment} form.
The color of the vertex $v_M = w_2$ corresponds to the weight of some
perfect matching in $G_1$
and to the weight of some perfect matching in $G_2.$
These two weights have to be equal.
The picture is simplified.
Some of the edges are omitted in the picture.
Note that the values $M$ and $m$ can be different for $(G_1, w_1)$ and
for $(G_2, w_2)$.
}
\label{fig:matchings-to-ca}
\end{center}
\vspace{-1cm}
\end{figure}
In the proof we use Lemma~\ref{lem:matchings-to-ca} to encode
$G_1$ and $G_2$ in two instances of \ProblemName{Channel Assignment},
then we extend them to the common length using
Lemma~\ref{lem:ca-extend}
and finally we merge them using Lemma~\ref{lem:ca-merge}.
The proof is straightforward and is moved to the Appendix
due to space limitations.
The simplified picture of the obtained \ProblemName{Channel Assignment} instance can
be found in Figure~\ref{fig:matchings-to-ca}.
Now we can use the results of Section~\ref{sec:CMW} to get following
two corollaries.
\begin{corollary}
There is no algorithm solving \ProblemName{Channel Assignment} in
$2^{o(n\log n)} \poly{r}$
where $n$ is the number of the vertices
and $r$ is the bit size of the instance
unless ETH fails.
\end{corollary}
\begin{proof}
For a given instance of \ProblemName{Common Matching Weight} with $n$ vertices and the weights bounded by $m$
we can transform it by
Lemma~\ref{lem:CMW-to-CA} into an instance of \ProblemName{Channel Assignment} with $n' = O(n)$ vertices
and the weights bounded by $\ell = O(n^2m).$
Note that for the bit size $r'$ of the new instance we have
$\poly{r'} = \poly{(n')^2\ell} = \poly{n, m} = \poly{r}.$
Let us assume that we can solve \ProblemName{Channel Assignment} in $2^{o(n\log n)}\poly{r}$-time.
Then we can solve our instance in time
$2^{o(n'\log n')}\poly{r'} = 2^{o(n\log n)} \poly{r}$
which contradicts ETH by Corollary~\ref{cor:CMW-hardness}.
\end{proof}
\begin{corollary} \label{cor:ca-nloglogl} ($\bigstar$)
There is no algorithm solving \ProblemName{Channel Assignment} in
$2^{n \cdot o(\log\log\ell)} \poly{r}$
where $n$ is the number of the vertices
and $r$ is the bit size of the instance
unless ETH fails.
\end{corollary}
\subsection*{Acknowledgments}
We thank {\L}ukasz Kowalik who was the originator of this studies for
a comprehensive support.
\bibliographystyle{abbrv}
|
\section{\label{sec:intro}Introduction}
The major goal in the studies of the dense baryonic matter at Nuclotron (BM@N project) \cite{bmn_CDR} is the measurement of strange and multi-strange baryons and mesons in
heavy ion collisions at the beam energies between 2 and 6 A$\cdot$GeV \cite{bmn_PoS}.
The physics program can be extended to the measurements of the in-medium effects for
strange particles decaying in hadronic modes \cite{brat1},
hard probes and correlations \cite{vasiliev_npps2011}, spin and polarization effects
\cite{lambda, bmn_dspin2013} etc.
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig1.eps}}
\caption{The vertical component $B_y$ of the magnetic field in the centre of
the SP41 magnet before the modification. }
\label{fig:fig1}
\end{figure}
For these purposes an experimental set-up will be installed
at the 6V beamline in the fixed-target hall of the Nuclotron.
The 6V beamline contains the quadrupole lenses doublet, two dipole magnets
allowing to
correct the beam position in the vertical and horizontal planes, and large aperture SP41 dipole magnet
for the momentum measurements \cite{bmn_CDR}.
The first results with the relativistic deuteron \cite{terekhin_PoS} and
carbon \cite{piyadin_C12} beams are demonstrated the feasibility of the
dense baryonic matter studies with light nuclei using 6V beamline infrastructure.
The modified SP41 dipole magnet will be used as an analyzing magnet \cite{bmn_CDR}.
Initially, the length of the magnet pole along the beam was 2.50~m, width in the horizontal
direction was 1.70~m and the height was about 0.75~m.
This magnet had also a hole for the photo-camera in the
upper pole since it was used previously for the experiments with streamer chamber.
As the result the
magnetic field components are demonstrated the non-uniform behaviour.
The $2D$ dependence of the magnetic field vertical component $B_y$ of the non-modified magnet SP41
in $XZ$ plane is shown in Fig.\ref{fig:fig1}.
However, the detection of multi-strange baryons requires the large aperture
silicon tracking system placement into the homogeneous magnetic field \cite{bmn_CDR,bmn_PoS,bmn_dspin2013}.
The SP41 magnet was modified to satisfy this requirement, namely,
the distance between the poles has been
enlarged up to 1.05 m and the hole in the upper pole and horizontal
beams has been filled by the steel-15.
In this paper the results of the magnetic field measurements are presented for the
modified dipole magnet SP41. These results are compared with the 3D magnetic field TOSCA calculations.
\section{\label{sec:TOSCA}Magnetic field calculations}
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig2.eps}}
\caption{SP41 dipole magnet model for the magnetic field TOSCA \cite{tosca}
calculations.}
\label{fig:fig2}
\end{figure}
\begin{figure}[hbtp]
\centering
\resizebox{8cm}{!}{\includegraphics{magnet_fig3.eps}}
\caption{The 3D TOSCA calculation results for the magnetic field vertical component
$B_y$ in the centre of the modified magnet SP41 along the beam direction.}
\label{fig:fig3}
\end{figure}
The magnetic field calculations for modified SP41 dipole magnet
has been performed using 3D TOSCA code \cite{tosca}.
Steel-10, steel-15 and copper were taken as the material for the yoke, magnet poles
and coils, respectively. The coordinate system was chosen as following:
$X$ axis is perpendicular to the beam direction in the horizontal plane, $Y$ is vertical and $Z$ is along the beam and parallel to the magnet poles.
The 3D model of the modified SP41 dipole magnet
for the magnetic field TOSCA calculations is presented in
Fig.~\ref{fig:fig2}.
The 3D TOSCA calculation results for the magnetic field vertical component $B_y$
in the centre of the
modified magnet SP41 along the beam direction are demonstrated in
Fig.~\ref{fig:fig3}. The current in the coils was taken as 1900~A.
The maximal value of the vertical component $B_y$ was found $\sim$0.9~T. The field integral is
$\sim$2.9~T$\cdot$m, which is
approximately 30~\% less than for the non-modified SP41 magnet.
\begin{figure}[htbt]
\begin{minipage}[t]{0.33\textwidth}
\centering
\resizebox{5cm}{!}{\includegraphics{magnet_fig4a.eps}}
\end{minipage}\hfill\begin{minipage}[t]{0.33\textwidth}
\centering
\resizebox{5cm}{!}{\includegraphics{magnet_fig4b.eps}}
\end{minipage}\hfill\begin{minipage}[t]{0.33\textwidth}
\centering
\resizebox{5cm}{!}{\includegraphics{magnet_fig4c.eps}}
\end{minipage}
\caption{
From left to right: $2D$ distributions in $XZ$ plane of the magnetic field
vertical component $B_y$ of modified SP41 magnet at $Y = 0$~cm (centre of the
magnet), $Y = 30$~cm, and $Y = 50$~cm.
}
\label{fig:fig4}
\end{figure}
The $2D$ distributions in $XZ$ plane of the magnetic field vertical
component $B_y$
of the modified SP41 magnet at $Y$=0 cm (centre of the magnet), $Y$=30 cm and
$Y$=50 cm are
presented in the left, middle, and right panels in Fig.~\ref{fig:fig4}.
The field demonstrates very smooth behaviour except the region around pole
of the magnet ($Y$=50 cm). The $3D$ results for $B_y$, $B_x$ and $B_z$
components of the magnetic field were incorporated
into the BM@N setup description for the simulation software \cite{bmn_dspin2013}.
\section{\label{sec:measur} Commissioning of the SP41 magnet}
During 2012--2013 the warm SP41 dipole magnet has been significantly modernized.
The magnet vertical gap has been enlarged by 30 cm up to 1.05 m. The upper pole and upper
horizontal beams have been filled by steel-15 in order to improve the magnetic field
homogeneity. The upper coil (with renovated pole and horizontal beams) and
lower coil alone have been rotated by 180$^\circ$
in the horizontal plane to provide optimal access to the magnet infrastructure
and detectors inside the BM@N experimental zone \cite{bmn_CDR}.
The magnet infrastructure, namely, pipes for cooling water,
the current leads and diagnostics have been also rotated by 180$^\circ$ and renovated.
The view of modernized SP41 dipole magnet is presented in Fig.\ref{fig:fig5}.
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig5.eps}}
\caption{The view of modernized SP41 dipole magnet.}
\label{fig:fig5}
\end{figure}
The commissioning of the SP41 magnet has been performed in two steps.
Firstly, the operation of the magnet with the 1650 A current
in the coils has been checked several times during several hours.
Second step includes the measurements of the magnetic field
in the centre of the magnet, the magnetic field integral and magnetic fringe field within
working range of the current in the coils (up to 1960 A).
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig6.eps}}
\caption{The
$B_y$ vertical component of the magnetic field
in the centre of modernized SP41 dipole magnet
as a function of the current in the coils \cite{rukoyatkin}. The dashed curve is the result of the quadratic function approximation.}
\label{fig:fig6}
\end{figure}
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig7.eps}}
\caption{The
magnetic field integral $\int Bdl$ of modernized SP41 dipole magnet
as a function of the current in the coils \cite{rukoyatkin}. The dashed curve is the result of the quadratic function approximation.}
\label{fig:fig7}
\end{figure}
The measurements of the $B_y$ vertical component of the magnetic field
in the centre of modernized SP41 dipole magnet
($X=0$, $Y=0$, $Z=0$) have been performed using planar Hall probe \cite{rukoyatkin}.
The
$B_y$ in the centre of modernized SP41 dipole magnet
as a function of the current in the coils is shown in Fig.\ref{fig:fig6}.
The dashed curve is the result of the quadratic function approximation.
The $B_y$ demonstrates the linear dependence on the current with negligible contribution
of the quadratic term. The $B_y$ at 1900 A is equal to 0.933$\pm$0.003 T
being in good
agreement with the 3D TOSCA calculation shown in Fig.\ref{fig:fig2}.
The measurements of the magnetic field integral $\int Bdl$ of modernized SP41 dipole magnet has been
performed using the method of the current-carrying filament \cite{rukoyatkin}.
The
magnetic field integral $\int Bdl$ of modernized SP41 dipole magnet
as a function of the current in the coils is shown in Fig.\ref{fig:fig7}.
The dashed curve is the result of the quadratic function approximation.
The magnetic field integral $\int Bdl$ is equal to 2.986$\pm$0.009 T$\cdot$m at 1900~A.
The non-linear contribution is about 5\%. The measured value is in good accordance
with the 3D TOSCA calculation results \cite{bmn_CDR}.
The field integral provides the bending angle of $\sim$70 mrad for
charged particles with $p/Z$=13 GeV/$c$.
On the other hand, since the saturation effect for the magnetic
field integral is not observed at 1900~A, the current value in the coils can be increased. This
will provide larger bending angle and better momentum resolution.
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig8.eps}}
\caption{The
magnetic field $|B|$ measured with the planar 3D Hall probe \cite{bmn_CDR}
at $X=0.2$~m, $Y=-0.22$~m, $Z=1.6$~m
as a function of the current in the coils.
The dashed curve is the result of the quadratic function approximation.}
\label{fig:fig8}
\end{figure}
\begin{figure}[hbtp]
\centering
\resizebox{10cm}{!}{\includegraphics{magnet_fig9.eps}}
\caption{The
magnetic field $|B|$ for the
current value in the coils of 1900~A as a function of $Z$ coordinate ($X=0.2$~m and $Y=-0.22$~m).
The open circles and triangles are the data obtained with the planar and coaxial 3D Hall probes \cite{bmn_CDR},
respectively.
The solid curve is the result of 3D TOSCA calculation scaled by factor 0.92.}
\label{fig:fig9}
\end{figure}
The measurements of the magnetic field $|B|$ at $X=0.2$~m, $Y=-0.22$~m, $Z=1.6$~m
have been performed using planar 3D Hall probe described in details in ref.\cite{bmn_CDR}.
The position of this point
is out of the magnet poles.
The
measured magnetic field $|B|$
as a function of the current value in the coils is shown in Fig.\ref{fig:fig8}.
The dashed curve is the result of the quadratic function approximation.
The $|B|$ demonstrates the non-negligible contribution
of the quadratic term at large values of the current in the coils.
For instance, the non-linear contribution is about 12\% at the current value of 1900~A.
Therefore, the fringe field demonstrates the saturation effect at large
currents in the coils, which is absent for the field in central region of the magnet.
The measurements of the fringe magnetic field as a function
of the distance from the magnet centre $Z$ at fixed values of
$X=0.2$~m and $Y=-0.22$~m coordinates have been performed using planar and coaxial 3D Hall probes \cite{bmn_CDR}.
The
magnetic field $|B|$ for the
current value in the coils of 1900~A as a function of $Z$ coordinate is shown in Fig.\ref{fig:fig9}.
The open circles and triangles represent the data obtained with the planar and coaxial 3D Hall probes \cite{bmn_CDR},
respectively. One can see that
the results obtained by these two probes are in good agreement.
The solid curve is the result of 3D TOSCA calculation scaled by factor 0.92.
The shape of the measured magnetic field $|B|$ reproduces the TOSCA results.
The observed difference can be recovered by the optimization of
the SP41 dipole magnet model for the magnetic field TOSCA \cite{tosca}
calculations.
\section{\label{sec:conclusions} Conclusions}
\begin{itemize}
\item
The SP41 dipole magnet has been modernized for the studies of dense baryonic matter
properties at Nuclotron. Namely,
the magnet vertical gap has been enlarged up to 1.05 m to increase the angular acceptance for
detection of hyperons \cite{bmn_PoS, lambda, bmn_dspin2013},
the magnetic field homogeneity improvement by the filling of the existed hole in the upper
pole and horizontal beams by steel-15 has been achieved,
the renovation of the magnet infrastructure has been made.
\item
The magnetic field in the centre of the magnet and out of the poles as well as field integral
have been measured as a function of current value in the coils. The
value of the magnetic field in the magnet centre and field integral at 1900~A
are in good agreement with the results of 3D TOSCA calculation,
while the measured fringe field at the same current is several percents below
the calculated one.
\item
Further steps are the measurements of the magnetic field components along the
optical axis of the SP41 magnet and 3D mapping of the magnetic field.
\end{itemize}
\vspace{0.5cm}
The authors thanks the technical services of LHEP participated in the modification
of the SP41 dipole magnet, especially, G.S.~Berezin, I.Ya.~Nefediev, A.V.~Shabunov, V.I.~Sharapov.
The authors are grateful to P.A.~Rukoyatkin for the magnetic measurements, I.A.~Bolshakova and S.~Timoshin for their assistance in
the use of 3D Hall probes. The work has been supported in part by RFBR
under grant $N^o$13-02-00101a.
|
\section{Introduction}\label{sec.intro}
Consider the periodic nonlinear Schr\"{o}dinger equation
\begin{align}\label{p-nls0}
i\partial_t{\psi} &= \Delta \psi + \lambda|\psi|^{2p} \psi
\end{align}
where $p \in \mathbb{N}$, $x\in\mathbb{T}^d$ and $\Delta$ is the standard Laplace-Beltrami operator. We wish to investigate the orbital stability of plane-wave solutions to \eqref{p-nls0}.
For $m \in \mathbb{Z}^d$, let $w_m(x,0):= \varrho e^{im\cdot x}$ be the initial datum concentrated at the $m$th mode. We will denote by $w_m(x,t)$ the plane-wave solution to equation \eqref{p-nls0} with initial datum $w_m(x,0)$. We will show that for any $K \in \mathbb{N}$, there exist $s_0$ and $\varepsilon_0$ so that any solution $\psi$ to \eqref{p-nls0} with initial datum that is $\varepsilon$-close to $w_m(x,0)$ in $H^s(\mathbb{T}^d)$, for $\varepsilon<\varepsilon_0$ and $s>s_0$, will meet the condition
\begin{align*}
\inf_{\varphi\in\mathbb{R}}\|e^{-i\varphi}e^{-im\cdot \bullet}w_m(\bullet,t)-e^{-im\cdot \bullet}\psi(\bullet,t)\|_{H^s(\mathbb{T}^d)} <
\varepsilon C(K,s_0,\varepsilon_0)
\end{align*}
for $t< \varepsilon^{-K}$.
Here $H^s(\mathbb{T}^d)$ is the Sobolev space.
Much has been written about this topic as outlined in the paper by Faou, Gauckler and Lubich \cite{FGL}. For instance, instability has been demonstrated in low regularity cases by Christ, Colliander, and Tao \cite{CCT031}, \cite{CCT032} ($d=1$, $s<0$), Carles, Dumas and Sparber \cite{CDS10} ($s<0$), and Hani \cite{H11} ($0<s<1$, $p=1$). On the other hand, for $d=1$ and $s=1$, there are stability results that can be found in Gallay and Haragus \cite{GH071}, \cite{GH072} and Zhidkov \cite{Z01}.
Results in the cubic case of our setting ($d\geq 1$, $s>1$) include results on the growth of the Sobolev norm of solutions and instability near $0$ by Bourgain \cite{B96}, Colliander, Keel, Staffilani, Takaoka and Tao \cite{CKSTT10} and Guardia and Kaloshin \cite{GK12}.
This paper, along with \cite{FGL}, uses the theory of Birkhoff normal forms in the same manner as Bambusi and Gr\'{e}bert \cite{B07, BG, G07} and Gauckler and Lubich \cite{GL10} in which the theory was applied to a modified cubic NLS, which requires high regularity, $s\gg 1$. As opposed to proving an instability result, we show long time orbital stability of plane wave solutions to \eqref{p-nls0} in the defocusing case.
We will emulate the argument presented in \cite{FGL} using the theory of Birkhoff normal forms presented in \cite{BG}. In \cite{FGL}, they prove
\begin{theorem}
Let $\rho_0>0$ be such that $1-2\lambda \rho_0^2>0$, and let $N>1$ be fixed arbitrarily. There exists $s_0>0$, $C \geq 1$ and a set of full measure $\mathcal{P}$ in the interval $(0,\rho_0]$ such that for every $s\geq s_0$ and every $\rho \in \mathcal{P}$, there exists $\varepsilon_0$ such that for every $m \in \mathbb{Z}^d$ the following holds: if the initial data $u(\bullet, 0)$ are such that
\begin{align*}
\|u(\bullet,0)\|_{L^2}=\rho \hspace{.5cm} \mbox{ and } \hspace{.5cm} \| e^{-im \cdot \bullet}u(\bullet,0)-u_m(0)\|_{H^s} = \varepsilon \leq \varepsilon_0
\end{align*}
then the solution of \eqref{p-nls0} (with $p=1$) with these initial data satisfies
\begin{align*}
\|e^{-im\cdot \bullet} u(\bullet,t)-u_m(t)\|_{H^s} \leq C \varepsilon \mbox{ for } t \leq \varepsilon^{-N}
\end{align*}
\end{theorem}
We prove the same result for any $p\in\mathbb{N}$, namely
\begin{theorem}\label{mainthm}
Let $L_0>0$ be such that $1-2p\lambda L_0^{p}>0$, and let $K>1$ be fixed arbitrarily. There exists $s_0>0$, $C \geq 1$ and a set of full measure $\mathcal{P}$ in the interval $(0,L_0]$ such that for every $s\geq s_0$ and every $L \in \mathcal{P}$, there exists $\varepsilon_0$ such that for every $m \in \mathbb{Z}^d$ the following holds: if the initial data $u(\bullet, 0)$ are such that
\begin{align*}
\|u(\bullet,0)\|^2_{L^2}=L \hspace{.5cm} \mbox{ and } \hspace{.5cm} \| e^{-im \cdot \bullet}u(\bullet,0)-u_m(0)\|_{H^s} = \varepsilon \leq \varepsilon_0
\end{align*}
then the solution of \eqref{p-nls0} with these initial data satisfies
\begin{align*}
\|e^{-im\cdot \bullet} u(\bullet,t)-u_m(t)\|_{H^s} \leq C \varepsilon \mbox{ for } t \leq \varepsilon^{-K}
\end{align*}
\end{theorem}
In essence, we show that the phenomena that allows for stability in the case $p=1$ are present for every $p\in \mathbb{N}$. We will demonstrate a more transparent and generalized argument than what has been shown before. One aspect that is made apparent is that the techniques used in \cite{FGL}, for the case $p=1$, can be applied to the case $p>1$ while examining the vector field of the normalized Hamiltonian as we do in the final two sections of this paper. This examination reaffirms that the stability is derived entirely from the type of linear combinations of the frequencies that are degenerate and the algebraic properties of the nonlinearity.
We should not readily expect this type of extension from the $p=1$ case. For the one-dimensional NLSE, the references mentioned above and \cite{GK14} show us that the $p=1$ case and the $p>1$ case exhibit very different phenomena. An important example is the fact that the one-dimensional NLSE with $p=1$ is completely integrable. Thus, it is not obvious that the same results should follow directly from the arguments made in \cite{FGL}.
In fact, after employing the normal form change of variables, the lower degree terms that remain can either be grouped into terms that preserve, as they are called in \cite{FGL}, ``super-actions":
\begin{align*}
\sum_{|n|^2=q} |z_n|^2
\end{align*}
or are small enough to be grouped with the high degree terms which determine the extent to which the solution remains close to the orbit of $w_m$.
\section{Functional Setting}
We first establish a setting in which to prove Theorem \ref{mainthm}. Similar definitions, as well as the proof of the lemmas in this section, appear in \cite{BG}.
\begin{definition}
For $x=\{x_n\}_{n \in \mathbb{Z}^d}$, define the standard Sobolev norm as
\begin{align*}
\|x\|_s := \sqrt{ \sum_{n\in \mathbb{Z}^d}|x_n|^2 \langle n \rangle^{2s} }
\end{align*}
Define $H^s$ as
\begin{align*}
H^s:= \left\{ x= \{x_n\}_{n\in \mathbb{Z}^d} \,\left| \, \|x\|_s<\infty \right. \right\}
\end{align*}
\end{definition}
Consider a vector-valued homogeneous polynomial, $X$, of degree $\ell$ be written as
\begin{align*}
X(z) = \sum_{\|\alpha\|_{\ell_1}=\ell} X_{\alpha} z^{\alpha}
\end{align*}
We will denote the majorant of $X$ by
\begin{align*}
\tilde{X}(z):=\sum_{\|\alpha\|_{\ell_1}=\ell} |X_{\alpha}| z^{\alpha}
\end{align*}
\begin{definition}[Tame Modulus] \label{tamemod}
Let $X$ be a vector-valued homogeneous polynomial of degree $\ell$. $X$ is said to have $s$-tame modulus if there exists $C>0$ such that
\begin{align*}
\left\| \tilde{X}(z^{(1)},...,z^{(\ell)}) \right\|_s \leq C \frac{1}{\ell} \sum_{k=1}^{\ell} \|z^{(1)}\|_{\frac{d+1}{2}}\cdots \|z^{(k-1)}\|_{\frac{d+1}{2}}\|z^{(k)}\|_s\|z^{(k+1)}\|_{\frac{d+1}{2}}\cdots \|z^{(\ell)}\|_{\frac{d+1}{2}}
\end{align*}
for all $z^{(1)},...,z^{(\ell)} \in H^s$. The infimum over all $C$ for which the inequality holds is called the tame $s$-norm and is denoted $|X|_s$.
\end{definition}
Definition \ref{tamemod} is an extension of Definition 2.2 in \cite{BG} with $d=1$. The inequality is related to the following property of Sobolev spaces: Consider two functions $u,v \in H^s(\mathbb{T}^d)$ for $s > \frac{d}{2}$, then by Leibnitz rules and Sobolev embedding we have
\begin{align*}
\|u\cdot v\|_{H^s(\mathbb{T}^d)} &\leq C_s\big( \|u\|_{H^s(\mathbb{T}^d)} \|v\|_{\infty}+\|u\|_{\infty} \|v\|_{H^s(\mathbb{T}^d)} \big) \\
&\leq C_{s,t}\big( \|u\|_{H^s(\mathbb{T}^d)} \|v\|_{H^t(\mathbb{T}^d)}+\|u\|_{H^t(\mathbb{T}^d)} \|v\|_{H^s(\mathbb{T}^d)} \big)
\end{align*}
for any $t>\frac{d}{2}$. On the Fourier side, the product $u\cdot v$ becomes a convolution of Fourier coefficients $\hat{u} *\hat{v}$. Therefore, we note that if $X$ is the function on sequences, $X(z^{(1)},\ldots,z^{(\ell)})=\tilde{X}(z^{(1)},\ldots,z^{(\ell)})=z^{(1)}*\ldots* z^{(\ell)}$,
by the same logic, there exists $C(s)$ such that
$$
\|X(z^{(1)},\ldots,z^{(\ell)})\|_s\leq C(s,d)
\sum_{k=1}^{\ell} \|z^{(1)}\|_{\frac{d+1}{2}}\cdots \|z^{(k-1)}\|_{\frac{d+1}{2}}\|z^{(k)}\|_s\|z^{(k+1)}\|_{\frac{d+1}{2}}\cdots \|z^{(\ell)}\|_{\frac{d+1}{2}}
$$
which is usually called ``tame property of the $H^s$ norm'' when $d=1$ (see, for instance, \cite{LM}). We choose $\frac{d+1}{2}$ in replace of $t$ for convenience and
in order to be consistent with \cite{BG} when $d=1$.
Of course, when $X(z^{(1)},\ldots,z^{(\ell)})$ is not a vector-valued homogeneous polynomial of degree $\ell$,
this property might not be satisfied. We now define two more norms on vector fields
\begin{definition}
Let $X$ be an vector-valued analytic function from $B_s(R)$ to $H^s$
where $B_s(R)=\{ x \in H^s \,|\, \|x\|_s \leq R\}$. We denote
\begin{align*}
\|X\|_{s,R} := \sup_{\|z\|_s \leq R} \|X(z)\|_s
\end{align*}
\end{definition}
\begin{definition}
Let $X$ be a nonhomogeneous vector-valued polynomial and consider its Taylor expansion
\begin{align*}
X= \sum X_{\ell}
\end{align*}
where $X_{\ell}$ is homogeneous of degree $\ell$ and assume $|X_{\ell}|_s<\infty$ for all $\ell$. For $R>0$, we define
\begin{align*}
\langle X\rangle_{s,R} := \sum_{\ell\geq 1} |X_{\ell}|_s R^{\ell}
\end{align*}
\end{definition}
The following lemmas provide context and a theoretical foundation for which to articulate and understand the reasoning behind the boundedness of the change of variables in Theorem \ref{normalform} and Lemma \ref{iterlemma} and the bounds on the norms of the resulting vector fields. Let
\begin{align*}
B_s(R):= \{ x\in H^s\,|\, \|x\|_s \leq R\}
\end{align*}
\begin{lemma} \label{comp.est}
Let $H$ be a Hamiltonian and $X_H:B_s(R) \rightarrow H^s$ the associated Hamiltonian vector field. Fix $0< r<R$, and assume that $\|X_H\|_{s,R} <\frac{r}{3}$ , and consider the time $t$ flow $\phi_t$ of $X_H$. Then, for $|t|\leq 1$,
\begin{align*}
\sup_{\|z\|_s\leq R-\frac{r}{3}} \|\phi_t(z) -z\|_s \leq \|X_H\|_{s,R}
\end{align*}
and for any analytic vector field $Y$ one has
\begin{align*}
\|Y \circ \phi_t\|_{s,R-r} \leq 2 \|Y\|_{s,R}
\end{align*}
\end{lemma}
The next lemma is especially important for establishing the negligibility of the terms in our transformed vector that are not normalized. We will not eliminate all nonresonant terms of small degree. Rather, we will eliminate terms so that the remaining low degree terms can be made to be as small as we want with an application of the following lemma:
\begin{lemma}
Fix $N$, and consider the decomposition $z=\bar{z}+\tilde{z}$. Where $\bar{z}:= \{ z_n\}_{|n|\leq N}$ and $\tilde{z}:=\{ z_n\}_{|n|\geq N}$. Let $X$ be a vector-valued polynomial with finite tame $s$-norm and assume that $X$ has a zero of order two in the variables $\tilde{z}$. then one has
\begin{align*}
\|X\|_{s,R} \lesssim \frac{\langle X \rangle_{s,R}}{N^{s-\frac{d+1}{2}}}
\end{align*}
\end{lemma}
For any two vector fields $X$ and $Y$, let $[X,Y]$ be the standard Lie bracket and define that adjoint function
\begin{align*}
\ad_Y(X):=[Y,X].
\end{align*}
Then we have the following two lemmas necessary for managing the effect of applying the $\ad$ function on the vector field infinitely many times in the definition of the normal form change of variables.
\begin{lemma}
For any $r<R$, one has
\begin{align*}
\langle [X,Y] \rangle_{s,R-r} \leq \frac{1}{r} \langle X\rangle_{s,R} \langle Y \rangle_{s,R}
\end{align*}
\end{lemma}
\begin{lemma}
For any $r<R$, one has
\begin{align*}
\langle \ad_Y^n(X) \rangle_{s,R-r} \leq \frac{e^n}{r^n} \langle X\rangle_{s,R} \left( \langle Y\rangle_{s,R} \right)^n
\end{align*}
\end{lemma}
\section{Symmetry Reduction and Diagonalization}
Let us consider equation \eqref{p-nls0} with the assumption $\lambda=-1$
\begin{align}\label{p-nls}
i\partial_t{\psi} &= \Delta \psi - |\psi|^{2p} \psi
\end{align}
where $p \in \mathbb{N}$, $x\in\mathbb{T}^d$ and $\Delta$ is the standard Laplace-Beltrami operator.
By the gauge invariance of \eqref{p-nls}, it suffices to continue assuming $m=0$. In Appendix \ref{appA}, we show it is sufficient to prove that
\begin{align*}
\|\psi(\cdot,t)-\psi_0(t)\|^2_{H^s(\mathbb{T}^d)} <\varepsilon C(N,s_0,\varepsilon_0)
\end{align*}
for $t< \varepsilon^{-N}$.
We denote $L:=\|\psi(0)\|^2_{L^2}$ and we assume that the $H^s$ norm of the initial datum is concentrated
at the zero mode for some $s>0$, i.e. $\|\psi(0)-{\psi}_0(0)\|_s=\varepsilon$. In order to eliminate the zero mode, we will construct a symplectic map on the Hamiltonian. Recall that the Hamiltonian corresponding to \eqref{p-nls} is
\begin{align} \label{ham}
H:= \sum_{k \in \mathbb{Z}^d} |k|^2 |u_k|^2+ \frac{1}{p+1}\sum_{\sum_{i=1}^{p+1} k_i= \sum_{i=1}^{p+1} h_i} u_{k_1} \dots u_{k_{p+1}}\bar{u}_{h_1}\dots \bar{u}_{h_{p+1}}.
\end{align}
Define the symplectic reduction of $u_0$:
\begin{align*}
&\{u_k, \bar{u}_k\}_{k \in \mathbb{Z}^d} \rightarrow (L,\nu_0, \{v_k,\bar{v}_k\}_{k \in \mathbb{Z}^d \setminus \{0\}}), \\
&u_0 = e^{i\nu_0}\sqrt{L-\sum_{k \in \mathbb{Z}^d} |v_k|^2}, \hspace{.3cm} u_k= v_ke^{i\nu_0}, \hspace{.3cm} \forall k \in \mathbb{Z}^d\setminus \{0\}.
\end{align*}
Inserting this change of variables into \eqref{ham} we obtain
\begin{align} \label{mainham}
&\sum_{k \in \mathbb{Z}^d \setminus\{0\}} |k|^2|v_k|^2+ \frac{1}{p+1}\big(L - \sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2 \big)^{p+1}+ \big(L - \sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2 \big)^{p}\Big( \sum_{k \in \mathbb{Z}^d\setminus \{0\}} (p+1)|v_k|^2 +\frac{p}{2}(v_kv_{-k}+\bar{v}_k\bar{v}_{-k})\Big)\\
&+ \big(L - \sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2 \big)^{p-\frac{1}{2}}\sum_{k_1,k_2 \in \mathbb{Z}^d\setminus \{0\} \atop k_1+k_2\neq 0} \Big(\frac{p(p-1)}{6} (v_{k_1}v_{k_2}v_{-k_1-k_2}+c.c) +\frac{(p+1)p}{2}(v_{k_1}v_{k_2}\bar{v}_{k_1+k_2}+c.c.)\Big) \nonumber\\
&+\big(L - \sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2 \big)^{p-1}\sum_{k_i \in \mathbb{Z}^d\setminus \{0\} \atop k_1+k_2\neq k_3+k_4} \Big(\frac{p^2(p+1)}{4} (v_{k_1}v_{k_2}\bar{v}_{k_3}\bar{v}_{k_4}+c.c) +\frac{(p+1)p(p-1)}{6}(v_{k_1}v_{k_2}v_{k_3}\bar{v}_{k_4}+c.c.)\Big) \nonumber\\
&+\big(L - \sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2 \big)^{p-1}\Big(\frac{p(p-1)(p-2)}{12}\sum_{k_i \in \mathbb{Z}^d\setminus \{0\} \atop k_1+k_2\neq k_3+k_4} (v_{k_1}v_{k_2}v_{k_3}v_{k_4}+c.c)\Big)+h.o.t. \nonumber
\end{align}
Expanding we have
\begin{align*}
&\frac{1}{p+1}L^{p+1}+ \sum_{k \in \mathbb{Z}^d \setminus\{0\}} (|k|^2+pL^p)|v_k|^2+ L^p\Big(\frac{p}{2}(v_kv_{-k}+\bar{v}_k\bar{v}_{-k})\Big) \\
&+ L^{p-\frac{1}{2}}\sum_{k_1,k_2 \in \mathbb{Z}^d\setminus \{0\} \atop k_1+k_2\neq 0} \Big(\frac{p(p-1)}{6} (v_{k_1}v_{k_2}v_{-k_1-k_2}+c.c) +\frac{(p+1)p}{2}(v_{k_1}v_{k_2}\bar{v}_{k_1+k_2}+c.c.)\Big)\\
&+ \big(- pL^{p-1}\sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2 \big)\Big( \sum_{k \in \mathbb{Z}^d\setminus \{0\}} (p+1)|v_k|^2 +\frac{p}{2}(v_kv_{-k}+\bar{v}_k\bar{v}_{-k})\Big)+\big(\frac{p}{2}L^{p-1}\big(\sum_{k \in \mathbb{Z}^d \setminus \{0\}} |v_k|^2\big)^2\big)\\
&+L^{p-1}\sum_{k_i \in \mathbb{Z}^d\setminus \{0\} \atop k_1+k_2\neq k_3+k_4} \Big(\frac{p^2(p+1)}{4} (v_{k_1}v_{k_2}\bar{v}_{k_3}\bar{v}_{k_4}+c.c) +\frac{(p+1)p(p-1)}{6}(v_{k_1}v_{k_2}v_{k_3}\bar{v}_{k_4}+c.c.)\Big)\\
&+L^{p-1}\Big(\frac{p(p-1)(p-2)}{12}\sum_{k_i \in \mathbb{Z}^d\setminus \{0\} \atop k_1+k_2\neq k_3+k_4} (v_{k_1}v_{k_2}v_{k_3}v_{k_4}+c.c)\Big)+h.o.t.
\end{align*}
We now diagonalize the quadratic part of the Hamiltonian:
\begin{align} \label{quada}
H_0= \sum_{k \in \mathbb{Z}^d\setminus \{0\}} (k^2+L^{p}p)|v_k|^2+L^p\frac{p}{2}(v_kv_{-k}+\bar{v}_k\bar{v}_{-k})
\end{align}
which amounts to diagonalizing the matrices:
\begin{align*}
J_k = k^2 \left(\begin{array}{cc} 1 &0 \\ 0 & -1\end{array} \right)+ L^pp \left(\begin{array}{cc} 1 &1 \\ -1&-1\end{array} \right)
\end{align*}
which group together $v_k, \bar{v}_{-k}$. We set
\begin{align*}
x_k = a_k v_k+ b_k \bar{v}_{-k}, ~ k\neq 0
\end{align*}
and in these variables
\begin{align} \label{quadb}
H_0= \sum_{k \in \mathbb{Z}^d} \frac{\Omega_k}{2} (|x_k|^2+|x_{-k}|^2)
\end{align}
with $\Omega_k= \sqrt{|k|^2(|k|^2+2pL^p)}$.
We note that $\Omega_n=\Omega_{m}$ whenever $|n|=|m|$. Therefore it might be convenient to
group together the modes having the same frequency i.e. to denote
\begin{equation}\label{omegas}
\omega_{q}:=\sqrt{q^2(q^2+2pL^{p})},\qquad q\geq 1.
\end{equation}
Before we continue, we note a crucial feature of our Hamiltonian and the vector field in the variables $x_k, \bar{x}_k$. Every monomial in $H(x,\bar{x})$,
\begin{align*}
x_{k_1}\cdots x_{k_p}\bar{x}_{n_1}\cdots \bar{x}_{n_q},
\end{align*}
obeys the law of {\em Conservation of Momentum}. Namely
\begin{align} \label{COM}
k_1+\cdots + k_p = n_1+\cdots + n_q
\end{align}
This property will be extremely important to the dynamics of the Hamiltonian.
\section{Normal Form}
We are now in the position to apply the theory of Birkhoff normal forms from Bambusi and Gr\'ebert \cite{BG}. We demonstrate, for completeness, the formal argument and introduce the nonresonance condition. After we demonstrate the normalization of our vector field, we can proceed to developing dynamical properties of the system.
Let us consider an auxiliary Hamiltonian $H(x)$, denote by $X_H$ the corresponding vector
field and by $\phi^t_H(x_0)$ the
time $t$ flow associated to $H$.
We note that
for any vector field $Y$, its transformed vector field under the time 1 flow generated by $X_H$ is
\begin{align}\label{lie}
e^{\ad_{X_H}}Y= \sum_{k=0}^{\infty} \frac{1}{k!} \ad_{X_H}^k Y
\end{align}
where $\ad_{X} Y:=[Y,X]$.
\subsection{Formal Argument}
Consider the equation
\begin{align} \label{form}
i\partial_t(y)_q= \omega_q (y)_q + \sum_{k\geq 2} \big(f_k(y)\big)_q
\end{align}
where for any sequence $y$ indexed by $\mathbb{Z}^d$ and $q \geq 1$
\begin{align*}
(y)_q:=
\left( \begin{array}{c} y_{n_1} \\ \cdots \\ y_{n_{k_q}} \end{array} \right)
\end{align*}
with $k_q :=\# \{n \in \mathbb{Z}^d \,|\, |n|=q\}$. Suppose that each $f_k$ is a vector-valued homogeneous polynomial of degree $k$.
We note that if we group together the components $x_n$ with $|n|=q$ and use the change of variables that takes \eqref{quada} to \eqref{quadb}, then we can rewrite the vector field for \eqref{mainham} in the form of \eqref{form}.
Our aim is to use an iterative argument which puts \eqref{form} into ``normal form'' up to some predetermined degree.
As usual, at each step we use a change of variables given by a time-1 flow map
associated with a suitable Hamiltonian vector field. We proceed by demonstrating this process of normalizing the vector field in \eqref{form} at degree $K_0\geq 2$.
Let $H$ be a Hamiltonian of degree $K_0$ and consider the change of variables
\begin{align*}
y=\Phi_H(x)
\end{align*}
where $\Phi_H(x)$ is the time-1 flow map associated with the Hamiltonian vector field $X_H$.
Using the identity \eqref{lie}, one obtains
\begin{align*}
i \partial_t (y)_q= \omega_q (y)_q+ \sum^{K_0-1}_{k= 2} \big(f_k(y)\big)_q
+([X_H,\omega y ](y))_q +
(f_{K_0}(y))_q
+ h.o.t.
\end{align*}
where $\omega y$ is the vector field with components
$$
\begin{pmatrix}
(\omega y)_n \cr \ \cr \overline{(\omega y)}_{-n}
\end{pmatrix}=
\left( \begin{array}{cc} \omega_q & 0\\ 0 & -\omega_q \end{array} \right) \left( \begin{array}{c} y_n\\ \bar{y}_{-n} \end{array} \right).
$$
The idea is to choose $H$ and another vector-valued homogeneous polynomial of degree $K_0$, $R_{K_0}$, in such a way that we can decompose $f_{K_0}$ as follows
\begin{align} \label{homo}
f_{K_0}(y)=R_{K_0}(y)-[X_H,\omega y ](y).
\end{align}
We can find $H$ so that $R_{K_0}$ is in the kernel of the following function from the space of polynomial vector fields into itself
\begin{align*}
\ad_{\omega} (X):=[X,\omega y].
\end{align*}
Any $Y \in \ker \ad_{\omega}$ is referred to as "normal" or "resonant".
In order to correctly choose $H$ and $R_{K_0}$, we will use the theory developed in \cite{BG}. First, let us characterize the normal terms with respect to the nonresonace condition of the frequencies $\{\Omega_n\}$. The monomials in $f_{K_0}$ that are normal are those terms $y^{\alpha}\bar{y}^{\beta}\partial_{y_m}$, where $\alpha,\beta\in\mathbb{N}^{\infty}$ with
$\|\alpha\|_1+\|\beta\|_1=K_0$ with components $\delta_{j,m}$ ($\delta_{j,m}$ being
the Kronecker symbol), such that
\begin{align*}
y^{\alpha}\bar{y}^{\beta}\partial_{y_m} \in \ker \ad_{\omega }.
\end{align*}
We note that
\begin{align*}
\ad_{\omega }(y^{\alpha}\bar{y}^{\beta}\partial_{y_m})= [(\alpha-\beta)\cdot \Omega-\Omega_m]y^{\alpha}\bar{y}^{\beta}\partial_{y_m}
\end{align*}
so that we can characterize $\ker \ad_{\omega }$ as
\begin{align*}
\ker \ad_{\omega }= \mbox{span} \left\{ y^{\alpha}\bar{y}^{\beta}\partial_{y_m} \,| \,(\alpha-\beta)\cdot \Omega-\Omega_m=0 \right\}
\end{align*}
where $(\alpha-\beta)\cdot \Omega=\sum_{i \in \mathbb{Z}^d} \alpha_i \Omega_i-\sum_{i \in \mathbb{Z}^d} \beta_i \Omega_i$.
Let $X_H$ be a homogeneous vector-valued polynomial of degree $K_0$. We Taylor expand $Y$, $X_H$, and $R_{K_0}$
\begin{align*}
Y(y,\bar{y})= \sum_{\alpha,\beta,m} Y_{\alpha,\beta,m}y^{\alpha}\bar{y}^{\beta} e_m\\
X_H(y,\bar{y}) = \sum_{\alpha,\beta,m} X_{\alpha,\beta,m}y^{\alpha}\bar{y}^{\beta} e_m\\
R_{K_0}(y,\bar{y}) = \sum_{\alpha,\beta,m} R_{\alpha,\beta,m}y^{\alpha}\bar{y}^{\beta} e_m
\end{align*}
The homological equation \eqref{homo} becomes
\begin{align*}
R_{\alpha,\beta,m}-(\Omega\cdot (\alpha-\beta)-\Omega_m)X_{\alpha,\beta,m} = Y_{\alpha,\beta,m}
\end{align*}
Now we define $X_H$ and $R_{K_0}$ as follows:
\begin{align*}
&\begin{array}{l} R_{\alpha,\beta,m} := Y_{\alpha,\beta,m} \\ X_{\alpha,\beta,m}:=0 \end{array} ~\mbox{ when }~ \Omega\cdot (\alpha-\beta)-\Omega_m=0\\
&X_{\alpha,\beta,m}:=\frac{-Y_{\alpha,\beta,m}}{(\Omega\cdot (\alpha-\beta)-\Omega_m)} ~\mbox{ when }~ \Omega\cdot (\alpha-\beta)-\Omega_m\neq 0
\end{align*}
We note that through this definition $H$ will be a Hamiltonian and that this change of variables preserves conservation of momentum, \eqref{COM}.
If we define $\lambda_q := \sum_{|i| =q} \alpha_i -\beta_i$, then the expression
$
(\alpha-\beta)\cdot \Omega-\Omega_m
$
becomes
\begin{align*}
\sum_{q \geq 1} \lambda_q \omega_q.
\end{align*}
\subsection{Nonresonance Condition}
Now that we have a formal characterization of resonant polynomials, we can state a normal form theorem and determine dynamical properties of our system. Given an $M\in \mathbb{N}$ dependent on the highest degree at which we will perform a normal form reduction, we have the following condition applicable to our parameter $L$ from the definition \eqref{omegas}:
\begin{definition}[Nonresonance Condition]
There exists $\gamma=\gamma_{M}>0$ and $\tau=\tau_M >0$ such that for any $N$ large enough, one has
\begin{equation}\label{nrcond}
\left| \sum_{q\geq1} \lambda_{q}\omega_{q} \right|\geq \frac{\gamma}{N^{\tau}}\hspace{1cm}\mbox{for } \|\lambda\|_1\leq M,
\ \ \sum_{q>N}|\lambda_q|\leq 2
\end{equation}
where $\lambda\in \mathbb{Z}^{\infty}\setminus\{0\}$.
\end{definition}
The following generalization of the ``non-resonance'' result in \cite{BG} holds.
\begin{theorem} \label{resthm}
For any $P>0$, there exists a set $J \subset (0,P)$ of full measure such that if $L^{p} \in J$ then for any $M>0$
the Nonresonance Condition holds.
\end{theorem}
\begin{proof}
The proof goes exactly as the one in Lemma 2.2 of \cite{FGL} with $L^{p}$ playing the role of $\rho^2$
and $p$ the one of $\lambda$ in their notations.
\end{proof}
If the Nonresonance condition is fulfilled, then we can conclude that for appropriate $\lambda$
\begin{align*}
\sum_{q \geq 1} \lambda_q \omega_q= 0
\end{align*}
implies $\lambda_q=0$ for all $q$ and
\begin{align*}
\sum_{q \geq 1} \lambda_q \omega_q \neq 0
\end{align*}
implies
\begin{align*}
\Big|\sum_{q \geq1} \lambda_q \omega_q \Big| \geq \frac{\gamma}{N^{\tau}}.
\end{align*}
\subsection{Normal Form Theorem}
Now we state the normal form theorem in \cite{BG} which we shall use in order
to prove our main result.
\begin{theorem} \label{normalform}
Consider the equation
\begin{align} \label{transe}
i\dot{x}= \omega x+ \sum_{k\geq 2} f_k(x).
\end{align}
and assume the nonresonance condition \eqref{nrcond}.
For any $\ell \in \mathbb{N}$, there exists ${s}_0={s}_0(\ell,\tau)$ such that for any $s\geq{s}_0$
there exists $r_{s}>0$ such that for $r<r_s$, there exists an analytic canonical change of variables
\begin{align*}
y=\Phi^{(\ell)}(x)\\
\Phi^{(\ell)}: B_s( r) \rightarrow B_s(3r)
\end{align*}
which puts \eqref{transe} into the normal form
\begin{align} \label{normal}
i\dot{y} = \omega y +\mathcal{R}^{(\ell)}(y)
+ \mathcal{X}^{(\ell)}(y).
\end{align}
Moreover
there exists a constant $C=C_s$ such that:
\begin{itemize}
\item \begin{align*} \sup_{x \in B_s(r)} \|x-\Phi^{(\ell)}(x)\|_s \leq C r^2 \end{align*}
\item $\mathcal{R}^{(\ell)}$ is at most of degree $\ell+2$, is resonant, and has tame modulus
\item the following bound holds
\begin{align*} \|\mathcal{X}^{(\ell)}\|_{s,r} \leq C r^{\ell+\frac{3}{2}} \end{align*}
\end{itemize}
\end{theorem}
\subsection{Normal terms}
Let's further characterize the resonant terms by starting with the case $d=1$.
\begin{lemma}[One-dimensional Case] \label{1dres}
Fix $K \in \mathbb{N}$. Consider $\ad_{\omega}$ as a function on homogeneous vector-valued polynomials of degree $K$. The degree $K$ resonant terms of equation \eqref{normal} are of the form
\begin{align*}
P_m(\{|y_n|^2\},\{y_n\bar{y}_{-n}\}) y_m \partial_{y_m}+ Q_m(\{|y_n|^2\},\{y_n\bar{y}_{-n}\})y_{-m}\partial_{y_m}
\end{align*}
where $P_m \in \mathbb{R}$ and $\bar{Q}_m=Q_{-m}$.
\end{lemma}
\begin{proof}
A degree $K$ resonant monomial is of the form $y^{\alpha}\bar{y}^{\beta}\partial_{y_m}$, where $\alpha, \beta$ and $m$ satisfy
\begin{align*}
(\alpha-\beta)\cdot \Omega -\Omega_m&=0
\end{align*}
which can be rewritten as
\begin{align*}
\sum_{i=1}^{M}\Omega_{n_i} - \sum_{i=1}^{M-1} \Omega_{k_i} -\Omega_m&=0
\end{align*}
for some $M>0$.
The nonresonance condition on the eigenvalues $\Omega_n$ implies that $(\alpha-\beta)\cdot \Omega -\Omega_m=0$ is (possibly up to reordering) equivalent to
\begin{align*}
\Omega_{n_i}=\Omega_{k_i}, \hspace{.5cm} \Omega_{m}=\Omega_{n_M}\\
\Leftrightarrow |n_i|=|k_i|, \hspace{.5cm} |m|=|n_M|.
\end{align*}
On the other hand, the conservation of momentum provides the following relation:
\begin{align*}
\sum_{i=1}^M n_i- \sum_{i=1}^{M-1} k_i -m=0
\end{align*}
In other words
the system of equations,
\begin{align*}
|n_i|=|k_i|, \hspace{.5cm} |m|=|n_M|\\
\sum^M_{i=1} n_i- \sum^{M-1}_{i=1} k_i -m=0
\end{align*}
characterizes the resonant terms. We will break up the characterization into cases. The first case is when $m=n_M$ and we have:
\begin{equation}\label{case1}
\begin{aligned}
|n_i|=|k_i|, \hspace{.5cm} m=n_M \\
\sum^{M-1}_{i=1} n_i- \sum^{M-1}_{i=1} k_i =0.
\end{aligned}
\end{equation}
For $n_{i},k_i$ satisfying \eqref{case1} above, there exists
$S=S(k_1,\ldots,k_{M-1})\geq0$ such that we can write \eqref{case1} as
\begin{equation}\label{case1a}
\begin{aligned}
n_{i_1}=-k_{i_1},\ldots,n_{i_S}=-k_{i_S}, \\
n_{i_{S+1}}=k_{i_{S+1}},\ldots,n_{i_{M-1}}=k_{i_{M-1}} \\ m=n_M \\
\sum^{M-1}_{i=1} n_i- \sum^{M-1}_{i=1} k_i =0.
\end{aligned}
\end{equation}
From the equation $n_{i_1}=-k_{i_1},\ldots,n_{i_S}=-k_{i_S}$ the resonant term contains a factor of the form
\begin{align*}
\prod_{1 \leq j \leq S } y_{n_{i_j}}\bar{y}_{-n_{i_j}}
\end{align*}
where we note that $\sum_{j=1}^S n_{i_j}=0$. From $n_{i_{S+1}}=k_{i_{S+1}},\ldots,n_{i_{M-1}}=k_{i_{M-1}}$ we obtain
\begin{align*}
\prod_{S< j \leq M-1} |y_{n_{i_j}}|^2.
\end{align*}
The full resonant term corresponding to \eqref{case1a} will be
\begin{align*}
y_m \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{1 \leq j \leq S } y_{n_{i_j}}\bar{y}_{-n_{i_j}} \partial_{y_m}.
\end{align*}
Therefore the resonant terms for the case $m=n_M$ will be the sum over all $\{n_i,k_i\}$ that satisfy \eqref{case1a}, namely:
\begin{align}\label{sum}
\left(\sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=0}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} y_{n_{i_j}}\bar{y}_{-n_{i_j}}\right) y_m \partial_{y_m}
\end{align}
For each $n_{i_1},...,n_{i_{M-1}} \in \mathbb{Z}$ and $S\in \{0,...,M-1\}$, we observe that the condition $\sum_{1\leq j \leq S} n_{i_j}=0$ implies that the terms
\begin{align*}
\prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{1 \leq j \leq S } y_{n_{i_j}}\bar{y}_{-n_{i_j}} \mbox{ and }\prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{1 \leq j \leq S }\bar{y}_{n_{i_j}}y_{-n_{i_j}}
\end{align*}
both appear in \eqref{sum} and thus
\begin{align*}
\left(\sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=0}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} y_{n_{i_j}}\bar{y}_{-n_{i_j}}\right) \in \mathbb{R}
\end{align*}
and we define
\begin{align*}
P_m(\{|y_n|^2\},\{y_n\bar{y}_{-n}\}) := \sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=0}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} y_{n_{i_j}}\bar{y}_{-n_{i_j}}
\end{align*}
We now consider the case $-m=n_M$. With this assumption we have
\begin{align*}
|n_i|=|k_i|, \hspace{.5cm} -m=n_M\\
\sum^{M-1}_{i=1} n_i- \sum^{M-1}_{i=1} k_i =2m
\end{align*}
and by the same argument the resonant terms from this case will be
\begin{align*}
\left(\sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=m}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} y_{n_{i_j}}\bar{y}_{-n_{i_j}}\right) y_{-m} \partial_{y_m}
\end{align*}
We now define
\begin{align*}
Q_m(\{|y_n|^2\},\{y_n\bar{y}_{-n}\}):= \sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=m}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} y_{n_{i_j}}\bar{y}_{-n_{i_j}}
\end{align*}
and note that
\begin{align*}
\bar{Q}_m&= \sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=m}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} \bar{y}_{n_{i_j}}y_{-n_{i_j}}\\
&= \sum_{0\leq S\leq M-1} \sum_{\substack{n_{i_1},\ldots,n_{i_{M-1}}\in\mathbb{Z}\\ \sum_{j=1}^S n_{i_j}=-m}} \prod_{S< j \leq M-1} |y_{n_{i_j}}|^2\prod_{\substack{1 \leq j \leq S \\ }} y_{n_{i_j}}\bar{y}_{-n_{i_j}}= Q_{-m}
\end{align*}
\end{proof}
The first step of analyzing the dynamical characteristics of the resonant terms is observing that the linear and resonant parts of the normalized Hamiltonian can be decoupled as a family of self-adjoint matrices.
\begin{corollary} \label{selfadjoint}
The truncation of \eqref{normal},
\begin{align*}
\dot{y}= \omega y + \mathcal{R}^{(\ell)}(y)
\end{align*}
can be decoupled in the following way:
\begin{align} \label{action}
i\partial_t \left( \begin{array}{c} y_n \\ y_{-n} \end{array} \right) = \mathcal{M}_n \left( \begin{array}{c} y_n \\ y_{-n} \end{array} \right)
\end{align}
where $\mathcal{M}_n=\mathcal{M}_n\left(\omega, \{|y_m|^2\}, \{y_m\bar{y}_{-m}\}\right)$ is a self-adjoint matrix for all $t$.
\end{corollary}
\begin{proof}
We Taylor expand $\mathcal{R}^{(\ell)}$:
\begin{align*}
\mathcal{R}^{(\ell)}=\sum_{i=2}^{\ell+2}c_i\mathcal{R}_i
\end{align*}
where each $c_i$ is a multiplicity constant and $\mathcal{R}_i$ is homogeneous and degree $i$. In fact, from Lemma \ref{1dres}
\begin{align*}
\mathcal{R}_i = \sum_{m\in \mathbb{Z}} \big(P^{(i)}_m(\{|y_n|^2\},\{y_n\bar{y}_{-n}\}) y_m + Q^{(i)}_m(\{|y_n|^2\},\{y_n\bar{y}_{-n}\})y_{-m}\big)\partial_{y_m}.
\end{align*}
Now we define $\mathcal{M}_n$ with the following components
\begin{align*}
(\mathcal{M}_n)_{11}&:= \omega_{|n|} + \sum_{i=2}^{\ell+2} c_i P^{(i)}_m &(\mathcal{M}_n)_{12}:= \sum_{i=2}^{\ell+2} c_i Q^{(i)}_m\\
(\mathcal{M}_n)_{21}&:= \sum_{i=2}^{\ell+2} c_i Q^{(i)}_{-m} &(\mathcal{M}_n)_{22}:= \omega_{|n|} + \sum_{i=2}^{\ell+2} c_i P^{(i)}_{-m}
\end{align*}
It follows immediately that $\mathcal{M}_n$ is self-adjoint.
\end{proof}
The one-dimensional case is included for instructive purposes. We can extend that arguments directly to the case $d>1$. It will be even more evident that the form of the resonant terms depends entirely on two properties of the Hamiltonian that have been mentioned previously:
\begin{itemize}
\item The Hamiltonian obeys the Conservation of Momentum law and
\item $\{ \omega_q\}_{ q<N}$ is a linearly independent set
\end{itemize}
\begin{lemma}[Higher-dimensional Case]
The resonant terms are of the form
\begin{align*}
\ker \ad_{\omega }=\mbox{span}\left(
\big\{Q_{m,i}y_{i}\partial_{y_m} \,\big|\, i , m \in \mathbb{Z}^d \big\} \right)
\end{align*}
where $\bar{Q}_{m,i}=Q_{i,m}$ and $Q_{i,m}=0$ when $|i|\neq |m|$.
In particular, $Q_{m,i}$ depends on $y_m$ only through terms of the form $y_{n}\bar{y}_k$ with $|n|=|k|$.
\end{lemma}
\begin{proof}
Just as in the one-dimensional case, the general resonant monomial of degree $k$ is of the form $y^{\alpha}\bar{y}^{\beta}\partial_{y_m}$, where $\alpha, \beta$ and $m$ satisfy
\begin{align*}
(\alpha-\beta)\cdot \Omega -\Omega_m&=0\\
\sum_{i=1}^{M}\Omega_{n_i} - \sum_{i=1}^{M-1} \Omega_{k_i} -\Omega_m&=0
\end{align*}
for some $M$ such that $2M-1=k$.
The nonresonance condition on the eigenvalues $\omega_q$ implies that, possibly up to reordering,
$(\alpha-\beta)\cdot \Omega -\Omega_m=0$ is equivalent to
\begin{align*}
\Omega_{n_i}=\Omega_{k_i}, \hspace{.5cm} \Omega_{m}=\Omega_{n_M} \\
\Leftrightarrow |n_i|=|k_i|, \hspace{.5cm} |m|=|n_M|
\end{align*}
Conservation of momentum provides the following relation:
\begin{align*}
\sum_{i=1}^{M} n_i- \sum_{i=1}^{M-1} k_i -m=0
\end{align*}
The system of equations,
\begin{align*}
|n_i|=|k_i|, \hspace{.5cm} |m|=|n_M|\\
\sum^M_{i=1} n_i- \sum^{M-1}_{i=1} k_i -m=0
\end{align*}
characterizes the resonant terms. We will again break up the characterization into cases. The first case is when $m=n_M$ and we have:
\begin{align}\label{highdsys}
|n_i|=|k_i|, \hspace{.5cm} m=n_M\\
\sum^{M-1}_{i=1} n_i- \sum^{M-1}_{i=1} k_i =0 \nonumber
\end{align}
and the resonant term corresponding to these equations will be of the form
\begin{align*}
y_m \prod_{i \in \{1,...,M-1\}}
y_{n_i}\bar{y}_{k_i} \partial_{y_m}
\end{align*}
At degree $2M-1$, the resonant terms for the case $m=n_M$ will be the sum over all $\{n_i,k_i\}$ that satisfy \eqref{highdsys}:
\begin{align}\label{hsum}
\left( \sum_{\sum n_i-\sum k_i=0} \prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i}\right) y_m \partial_{y_m}
\end{align}
We observe that the condition $\sum_{i \in \{1,...M-1\}} n_i-k_i=0$ implies that for each $\{n_i,k_i\}$ the terms
\begin{align*}
\prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i} \mbox{ and } \prod_{i \in \{1,...M-1\}} y_{k_i}\bar{y}_{n_i}
\end{align*}
both appear in \eqref{hsum} and thus
\begin{align*}
\left( \sum_{\sum n_i-\sum k_i=0} \prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i}\right) \in \mathbb{R}
\end{align*}
and we define
\begin{align*}
Q_{m,m} := \left( \sum_{\sum n_i-\sum k_i=0} \prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i}\right)
\end{align*}
For the case $n_M\neq m$. With this assumption we have
\begin{align*}
|n_i|=|k_i|, \hspace{.5cm} m^*:=n_M\\
\sum^{M-1}_{i=1} n_i- \sum^{M-1}_{i=1} k_i =m-m^*
\end{align*}
and by the same argument the resonant terms from this case will be
\begin{align*}
\left( \sum_{\sum n_i-\sum k_i=m-m^*} \prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i}\right) y_{m^*} \partial_{y_m}
\end{align*}
We now let
\begin{align*}
Q_{m,m*}:= \sum_{\sum n_i-\sum k_i=m-m^*} \prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i}
\end{align*}
and note that
\begin{align*}
\bar{Q}_{m,m^*}=\sum_{\sum n_i-\sum k_i=m-m^*} \prod_{i \in \{1,...M-1\}} \bar{y}_{n_i}y_{k_i}=\sum_{\sum n_i-\sum k_i=m^*-m} \prod_{i \in \{1,...M-1\}} y_{n_i}\bar{y}_{k_i}= Q_{m^*,m}
\end{align*}
\end{proof}
The analogue to Corollary \ref{selfadjoint} is as follows:
\begin{corollary} \label{highdselfadjoint}
The truncation of \eqref{normal},
\begin{align*}
i\dot{y}= \omega y + \mathcal{R}^{(\ell)}(y)
\end{align*}
can be decoupled in the following way:
\begin{align} \label{action}
i\partial_t \left( \begin{array}{c} y_{n_1} \\ \cdots \\ y_{n_k} \end{array} \right) = \mathcal{M}_q \left( \begin{array}{c} y_{n_1}\\ \cdots \\ y_{n_k} \end{array} \right)
\end{align}
where $q>0$, $\{n_1,\ldots,n_k\}:=\{n\in\mathbb{Z}^d\,:\,|n|=q\}$, $\mathcal{M}_q=\mathcal{M}_q\left(\omega, \{y_j\}\right)$ is a self-adjoint matrix for all $t$.
\end{corollary}
\begin{proof}
As in \ref{selfadjoint}, we expand $\mathcal{R}^{(\ell)}$,
\begin{align*}
\mathcal{R}^{(\ell)}= \sum_{i=2}^{\ell+2} c_i \mathcal{R}_i
\end{align*}
with
\begin{align*}
\mathcal{R}_i = \sum_{m\in \mathbb{Z}} \big(\sum_{j \in \mathbb{Z}^d\atop |j|=|m|}Q^{(i)}_{m,j}(\{y_n\})y_{j}\big)\partial_{y_m}.
\end{align*}
In conclusion, we define the components of $\mathcal{M}_q$:
\begin{align*}
\big(\mathcal{M}_q \big)_{mj}:= \delta_{mj}\omega_q +\sum_{i=2}^{\ell+2}c_i Q^{(i)}_{m,j}.
\end{align*}
\end{proof}
\subsection{Iterative Lemma}
We now present the inductive lemma that is used to produce Theorem \ref{normalform}. First consider a general Hamiltonian $H=H_0+P$. Expand $P$ in Taylor series up to order $\ell_0+2$:
\begin{align*}
P=P^{(1)}+ \mathscr{R}_*\\
P^{(1)}:= \sum_{i=1}^{\ell_0} P_i
\end{align*}
where $P_i$ is homogeneous of degree $i+2$ and $\mathscr{R}_*$ is the remainder of the Taylor expansion.
\begin{lemma}[Iterative Lemma, \cite{BG} Proposition 4.20, Corollary 4.21] \label{iterlemma}
Consider the Hamiltonian $H=H_0 + P^{(1)}+\mathscr{R}_*$, and fix $s\geq \frac{d+1}{2}$. For any $\ell\leq \ell_0$, there exists a positive $R_0\ll1$, and for any $N>1$, there exists an analytic canonical transformation
\begin{align*}
\mathscr{T}^{(\ell)}: B_s\Big( \frac{R_0(2\ell_0-\ell)}{2N^{\tau}\ell_0}\Big) \rightarrow H^s
\end{align*}
which transforms $H$ into
\begin{align*}
H^{(\ell)}:= H \circ \mathscr{T}^{(\ell)}= H_0 + \mathscr{L}^{(\ell)}+ f^{(\ell)} + \mathscr{R}^{(\ell)}_N + \mathscr{R}^{(\ell)}_T+ \mathscr{R}_* \circ \mathscr{T}^{(\ell)}.
\end{align*}
Let $L=\frac{2\ell_0-\ell}{2\ell_0}$. For any $R < R_0N^{-\tau}$, there exists a constant $C$ such that the following properties are fulfilled:
\begin{enumerate}
\item the transformation $\mathscr{T}^{(\ell)}$ satisfies
\begin{align*}
\sup_{z \in B_s(RL)} \|z-\mathscr{T}^{(\ell)}(z)\|_s \leq CN^{\tau}R^2;
\end{align*}
\item $\mathscr{L}^{(\ell)}$ is a polynomial of degree at most $\ell+2$ and has tame modulus; it is resonant and has a zero of order three at the origin; $f^{(r)}$ is a polynomial of degree at most $\ell_0+2$ and has zero of order $\ell+3$ at the origin; moreover, the following estimates hold:
\begin{align*}
\langle X_{\mathscr{L}^{(\ell)}} \rangle_{s, RL} &\leq C R^2 \\
\langle X_{f^{(\ell)}} \rangle_{s,RL} &\leq C R^2(RN^{\tau})^{\ell};
\end{align*}
\item the remainder terms, $\mathscr{R}^{(\ell)}_N$ and $\mathscr{R}^{(\ell)}_T$, have tame modulus and satisfy
\begin{align*}
\|X_{\mathscr{R}^{(\ell)}_T}\|_{s,RL} &\leq C (RN^{\tau})^{\ell_0+2},\\
\|X_{\mathscr{R}^{(\ell)}_N}\|_{s,RL} &\leq C R^2 N^{\frac{d+1}{2}-s},\\
\|X_{ \mathscr{R}_* \circ \mathscr{T}^{(\ell)}}\|_{s,RL} &\leq C (RN^{\tau})^{\ell_0+2}.
\end{align*}
\end{enumerate}
\end{lemma}
\section{Dynamics}
Finally, we can state the dynamical consequences of the normal form transformation given by Theorem \ref{normalform} by characterizing the normal terms. The characterization allows us to show, as in \cite{FGL}, that these terms preserve the super actions:
\begin{proposition} \label{cutoff}
Suppose $y \in H^s$ satisfies \eqref{action}, then
\begin{align*}
\partial_t \|y\|^2_s= \partial_t \sum_{q \geq 1} \left(\sum_{|n_i|=q}|y_{n_i}|^2 \right)\langle q \rangle^{2s}=0
\end{align*}
\end{proposition}
\begin{proof}
Fix $q$, let $\{n_1,\ldots,n_k\}:=\{n\in\mathbb{Z}^d\,:\,|n|=q\}$ and define
\begin{align*}
v:= \left( \begin{array}{c} y_{n_1} \\ \cdots \\ y_{n_k} \end{array} \right)
\end{align*}
Then by Corollary \ref{selfadjoint}
\begin{align*}
\partial_t \sum_{|n_i|=q}|y_{n_i}|^2 &= \partial_t v \cdot {\bar{v}}=\dot{v} \cdot {\bar{v}}+v \cdot \dot{{\bar{v}}}\\
&= \left(-i\mathcal{M}_q v\right) \cdot {\bar{v}}+v \cdot \left( i\overline{\mathcal{M}}_q {\bar{v}} \right)\\
&=\left(-i\mathcal{M}_q v\right) \cdot {\bar{v}}+\left(i\overline{\mathcal{M}}^T_q v \right) \cdot {\bar{v}} \\
&=0
\end{align*}
\end{proof}
Proposition \ref{cutoff} shows that for any $y(t)$ satisfying the truncated equation
\begin{align*}
i\partial_t y= \omega y + \mathcal{R}^{(\ell)}(y)
\end{align*}
there is no transference of mass between ``shells",
\begin{align*}
v_q:= \left( \begin{array}{c} y_{n_1} \\ \cdots \\ y_{n_k} \end{array} \right)
\end{align*}
although there may be transference between $y_n$ and $y_m$ when $|n|=|m|$. The following theorem completes our analysis on the dynamics of our system and allows us to prove the quantitative aspect of the stability in the main theorem, Theorem \ref{mainthm}.
\begin{theorem}\label{dynam}
Suppose $y \in B_s(r)$ satisfies \eqref{normal} with $r$ small enough. Then there exists a constant $C=C_s$
\begin{align*}
\partial_t \|y\|_s \leq C r^{\ell+\frac{5}{2}}
\end{align*}
\end{theorem}
\begin{proof}
\begin{align*}
\partial_t \|y\|_s&=\partial_t \langle y, y \rangle_s\leq |\langle\mathcal{X}^{(\ell)}(y) ,y \rangle_s|+
|\langle y, \mathcal{X}^{(\ell)}(y) \rangle_s | \\
&\leq C \|\mathcal{X}^{(\ell)} \|_{s,r}\|y\|_s
\end{align*}
The conclusion follows from Theorem \ref{normalform}.
\end{proof}
\section{Conclusion}
We conclude by assembling the proof of Theorem \ref{mainthm}. Theorem \ref{dynam}, shows that given a solution $y$ to equation \eqref{normal} with $\|y(0)\|_s <3r$, then for all $0<t <r^{-(\ell+\frac{3}{2})}$, $\|y\|_s< C_sr$. Assuming $r$ is small enough, Theorem \ref{normalform} implies that we then have the same bound for any $x$ that solves \eqref{transe} with $\|x(0)\|_s<r$. Finally, the transformation that takes a solution of equation \eqref{quada} to equation \eqref{quadb} is a change of coordinates on vectors of the form
\begin{align*}
v_q=
\left( \begin{array}{c} y_{n_1} \\ \cdots \\ y_{n_{k_q}} \end{array} \right)
\end{align*}
where $k_q =\# \{n \in \mathbb{Z}^d \,|\, |n|=q\}$. The transformation then preserves $\|v_q\|_{2}$ and therefore preserves the $H^s$ norm. Thus, the bound on $\|x\|_s$, for $x$ satisfying \eqref{transe}, can be applied to $\|v\|_s$, for $v$ satisfying \eqref{mainham}. Since we obtain $v$ by a gauge change to $(u_n)_{n \neq 0}$ fulfilling \eqref{ham}, we have
\begin{align*}
\|\psi(\cdot,t)-\psi_0(t)\|_{H^s(\mathbb{T}^d)}=\|u\|_s <C r
\end{align*}
for $t<r^{-(\ell+\frac{3}{2})}$.
The condition $1-2p\lambda L_0^{p}>0$ which implies $1-2p\lambda L^{p}>0$ when $\lambda>0$, is necessary so that $\Omega_n \in \mathbb{R}$ for all $n \in \mathbb{Z}^d$.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.