Graph Homomorphisms for Quantum Players
Extracted source text ยท arXiv 1212.1724
\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage[small, labelfont={bf}, margin=1cm]{caption}
\usepackage[
colorlinks,
linkcolor = blue,
citecolor = blue,
urlcolor = blue]{hyperref}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{thm-restate,colonequals,mathtools}
\usepackage{tikz}
\mathtoolsset{centercolon}
\usepackage{amssymb,amsmath,amsthm,enumerate}
% amssymb - Let's me use things like \textbf
% amsmath - Let's me use \text in math mode
% amsthm - Let's me use theorem environments
% Sets up my theorem environments
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{cor}[theorem]{Corollary}
% Sets up my definition environment
\theoremstyle{definition}
\newtheorem*{definition}{Definition}
\newtheorem{defn}[theorem]{Definition}
\theoremstyle{conjecture}
%\newtheorem*{definition}{Definition}
\newtheorem{conjecture}[theorem]{Conjecture}
% Sets up a few commands to make life easier
\DeclareMathOperator{\pg}{PG}
\newcommand{\Z}{\mathbb{Z}}
\newcommand{\qbinom}[2]{\left[{#1 \atop #2}\right]}
\newcommand{\xa}{x_{A}}
\newcommand{\xb}{x_{B}}
\newcommand{\ya}{y_{A}}
\newcommand{\yb}{y_{B}}
\newcommand{\ket}[1]{\left | #1 \right \rangle}
\newcommand{\bra}[1]{\left \langle #1 \right |}
\newcommand{\marrow}{\xrightarrow{\mathcal{M}}}
\newcommand{\qarrow}{\xrightarrow{q}}
\newcommand{\bartheta}{\bar{\vartheta}}
\DeclareMathOperator{\tr}{Tr}
\DeclareMathOperator{\supp}{supp}
\DeclareMathOperator{\vect}{vec}
\DeclareMathOperator{\rk}{rk}
\DeclareMathOperator{\og}{og}
\begin{document}
\title{Graph Homomorphisms for Quantum Players}
\author{David E. Roberson$^1$, Laura Man\v{c}inska$^{1,2}$ \\ \normalsize
$^1$\textit{Department of Combinatorics \& Optimization and} \\ \normalsize
\textit{$^2$Institute for Quantum Computing} \\ \normalsize
\textit{University of Waterloo}}
\date{}
\maketitle
\vspace{-.19in}
\abstract{A homomorphism from a graph $X$ to a graph $Y$ is an adjacency preserving
mapping $f:V(X) \rightarrow V(Y)$. We consider a nonlocal game in which Alice and
Bob are trying to convince a verifier with certainty that a graph $X$
admits a homomorphism to $Y$. This is a generalization of the
well-studied graph coloring game. Via systematic study of quantum
homomorphisms we prove new results for graph coloring. Most
importantly, we show that the Lov\'{a}sz theta number of the complement lower bounds the
quantum chromatic number, which itself is not known to be computable.
We also show that other quantum graph parameters, such as quantum
independence number, can differ from their classical counterparts.
Finally, we show that quantum homomorphisms closely relate to
zero-error channel capacity. In particular, we use quantum
homomorphisms to construct graphs for which entanglement-assistance
increases their one-shot zero-error capacity.}
%\vspace{-.2in}
\tableofcontents
%\begin{itemize}
%\item Intro
%\item Homomorphism game
% \begin{itemize}
% \item Classical strategy
% \item Quantum strategy
% \item Reformulation
% \item If $Y$ is vertex transitive, then $X \qarrow Y \Leftrightarrow X \ltimes Y$ has a projective representation of value $| V(Y)|$.
% \item Define $\xi_f$
% \item No-signalling
% \end{itemize}
%\item Basic Properties
% \begin{itemize}
% \item Transitivity
% \item $K_m \qarrow K_n$ if and only if $m \le n$
% \item $X$ connected, $Y$ has components $Y_i$ for $i \in [n]$, and $X \qarrow Y$ $\Rightarrow X \qarrow Y_i$ for some $i \in [n]$.
% \end{itemize}
%\item Quantum Parameters
% \begin{itemize}
% \item Definitions
% \item Quantum Chromatic Number
% \begin{itemize}
% \item Hadamard graphs have quantum chromatic EQUAL to n
% \end{itemize}
% \item Quantum Independence and Clique Numbers
% \begin{itemize}
% \item $X \qarrow Y \Leftrightarrow \alpha_q(X \ltimes Y) = |V(X)|$
% \item Separations
% \item Quantum Clique-Coclique Bound
% \end{itemize}
% \item Quantum Odd Girth
% \end{itemize}
%\item Relation to Zero-Error Capacity
% \begin{itemize}
% \item $\alpha_q(X) \le c_0^*(X)$ with equality if and only if $c_0^*(X)$ can be achieved with projective measurements on maximally entangled state.
% \item $X \qarrow Y \Rightarrow c_0^*(\overline{X}) \le c_0^*(\overline{Y})$
% \item $X \qarrow Y \Rightarrow \Theta^*(\overline{X}) \le \Theta^*(\overline{Y})$
% \end{itemize}
%\item Relation to Lov\'{a}sz $\vartheta$
%\end{itemize}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Introduction}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%Section labels: homogame, theta, parameters, capacity, rank
The $c$-coloring game on a graph $X$ consists of two players, Alice and Bob, attempting to convince a referee that they have a $c$-coloring of $X$ \cite{Galliard02,Cleve04}. The game is played as follows: the referee sends each of the players a vertex of $X$, and each player responds with a color from $[c] = \{1,2,\ldots,c\}$. To win, the players must respond with the same color when they are sent the same vertex, and with different colors when they are sent adjacent vertices. The players are not allowed to communicate during the game but may agree on a strategy beforehand. It can be easily seen that deterministic players with shared randomness (classical players) can win the $c$-coloring game on $X$ with certainty if and only if there exists a $c$-coloring of $X$. On the other hand, players allowed to make quantum measurements on some shared entangled state can sometimes win the $c$-coloring game on $X$ with certainty even when $X$ admits no $c$-coloring. Thus the quantum chromatic number, $\chi_q(X)$, is defined to be the smallest $c \in \mathbb{N}$ such that quantum players can win the $c$-coloring game on $X$ \cite{Avis}.
Quantum strategies for the coloring game and the quantum chromatic number have been well-studied \cite{Avis,qchrom,qchrom-aqis,SS12,MSS}. However, many questions remain unanswered. For example, it is not known whether $\chi_q(X)$ is computable, or whether there exists a family of graphs $X_n$ such that $\lim_{n \rightarrow \infty} \chi_q(X_n) < \infty$ but $\lim_{n \rightarrow \infty} \chi(X_n) = \infty$. Furthermore, there are few lower bounds known for quantum chromatic number. The authors of $\cite{qchrom}$ have shown that the orthogonal rank of a graph is a lower bound on a restricted version of quantum chromatic number, but the only general lower bound known is the clique number of a graph.
A \emph{homomorphism} from a graph $X$ to a graph $Y$ is a function, $\phi$, from the vertex set of $X$, denoted $V(X)$, to the vertex set of $Y$, denoted $V(Y)$, which preserves adjacency. More formally, $\phi: V(X) \rightarrow V(Y)$ is a homomorphism from $X$ to $Y$ if $\phi(x) \sim \phi(x')$ whenever $x \sim x'$, where ``$\sim$" denotes adjacency. We will write $X \rightarrow Y$ if there exists a homomorphism from $X$ to $Y$, and $X \not\rightarrow Y$ if not. It is straightforward to see that a homomorphism from a graph $X$ to the complete graph on $c$ vertices, denoted $K_c$, is equivalent to a $c$-coloring of $X$. Thus homomorphisms are a natural generalization of colorings. There is a well-developed and beautiful theory around graph homomorphisms \cite{Nesetril, Tardif}, and the study of them has given rise to original results on graph coloring.
Echoing the way in which homomorphisms generalize colorings, we have defined a homomorphism game which generalizes the coloring game. To play the $(X,Y)$-homomorphism game for graphs $X$ and $Y$, each player is sent a vertex of $X$, and must respond with a vertex of $Y$. In order to win, the players must respond with the same vertex of $Y$ when they are given the same vertex of $X$, and they must respond with adjacent vertices when given adjacent vertices. Similarly to the coloring game, classical players can win the $(X,Y)$-homomorphism game if and only if $X \rightarrow Y$. If quantum players can win the $(X,Y)$-homomorphism game, then we say that there is a \emph{quantum homomorphism} from $X$ to $Y$ and write $X \qarrow Y$.
Generalizing quantum colorings to quantum homomorphisms helps to place the idea into a broader and more natural mathematical context. Furthermore, a systematic study of quantum homomorphisms can potentially yield a better understanding of and new results concerning quantum colorings. Indeed, our Theorem~\ref{thetapreserved} can be applied to quantum colorings to obtain $\vartheta(\overline{X}) \le \chi_q(X)$, a previously unknown and efficiently computable lower bound. Here, $\overline{X}$ refers to the complement of $X$. Defining quantum homomorphisms also allows us to define quantum versions of many other graph parameters, including quantum independence number, which turns out to be intimately related to entanglement-assisted zero-error capacity. Lastly, quantum homomorphisms afford us many new examples of non-local games which potentially have perfect quantum strategies but no such classical strategy. We have already found an infinite family of such examples (see Section~\ref{alpha_q}).
%
%\vspace{1in}
%
%In this paper we will refer to orthogonal projectors (matrices which are both idempotent and Hermitian) as simply projectors, and we will say that two projectors $A$ and $B$ are orthogonal if $AB = 0$. Furthermore, all of our graphs are undirected, and unless explicitly defined otherwise they will be finite. We will denote the vertex set of a graph $X$ with $V(X)$, and we will write $x \sim x'$ to denote that vertices $x$ and $x'$ are adjacent. A function $\phi : V(X) \rightarrow V(Y)$ is a homomorphism from $X$ to $Y$ if $f(x) \sim f(x')$ whenever $x \sim x'$. We write $X \rightarrow Y$ whenever there exists a homomorphism from $X$ to $Y$. As an example, a coloring of the graph $X$ with $c$ colors is a homomorphism from $X$ to $K_c$.
%
%Here we investigate a game in which two players, Alice and Bob, try to convince a referee that they have a homomorphism from a graph $X$ to a graph $Y$. We refer to this as the $(X,Y)$-homomorphism game. Since colorings are a special case of homomorphisms, the homomorphism game is a generalization of the graph coloring game \cite{Cleve04,Galliard02,qchrom,Avis,qchrom-aqis,SS12,MSS} in which players try to convince a referee that they have a coloring of a graph $X$ with $c$ colors.
%
%If players are not allowed to communicate during the game, but allowed to decide on a classical strategy beforehand, then they can win the game with probability 1 if and only if $X \rightarrow Y$. However, if Alice and Bob are allowed to make quantum measurements on some shared entangled state, then they are sometimes able to win even when $X \not\rightarrow Y$. In the case that there exists a quantum strategy for winning the $(X,Y)$-homomorphism game, we say that there exists a quantum homomorphism from $X$ to $Y$ and write $X \qarrow Y$.
%----------------------------------------
\subsection{Preliminaries and Outline}
%----------------------------------------
In this paper graphs will be undirected and, unless explicitly defined otherwise, finite. A graph is said to be vertex/edge transitive if for any two vertices/edges, there exists an automorphism of the graph that maps one to the other. We will refer to orthogonal projectors (matrices which are both idempotent and Hermitian) simply as projectors, and we will say that two projectors $A$ and $B$ are orthogonal if $AB = 0$. We use $\mathbb{M}_{\mathbb{F}}(m,n)$ to denote the space of $m\times n$ matrices over the field $\mathbb{F}$. We say that a quantum state $\psi \in \mathbb{C}^d \otimes \mathbb{C}^d$ is \emph{maximally entangled} if $\psi = \frac{1}{\sqrt{d}}\sum_{i = 1}^d u_i \otimes v_i$, where $\{u_1,\ldots,u_d\}$ and $\{v_1,\ldots,v_d\}$ are some orthonormal bases for $\mathbb{C}^d$. The \emph{canonical maximally entangled state} in $\mathbb{C}^d \otimes \mathbb{C}^d$ is $\Phi = \frac{1}{\sqrt{d}}\sum_{i=1}^d e_i \otimes e_i$, where $e_i$ is the $i^\text{th}$ standard basis vector. A positive-operator valued measure (POVM) is a set $\{E_1,\ldots,E_n\}$ of positive semidefinite operators such that $\sum_{i \in [n]} E_i = I$. A POVM, or measurement, is projective if each $E_i$ is a projector. The rest of the paper is organized as follows:
In Section~\ref{homogame} we introduce in detail the homomorphism game and the general forms of classical and quantum strategies. We also reformulate the question of the existence of a winning quantum strategy for a particular homomorphism in terms of the existence of a particular set of projectors satisfying certain orthogonality constraints. This is achieved via an argument previously used in \cite{qchrom} to obtain a similar reformulation for the coloring game.
%This is done using the same argument given in \cite{qchrom} for the coloring game.
We then use the reformulation to prove some basic properties of quantum homomorphisms and to rephrase them in terms of homomorphisms.
In Section~\ref{theta} we show that the order of the Lov\'{a}sz theta function, $\vartheta$, is respected by quantum homomorphisms, i.e.~if $X \qarrow Y$, then $\vartheta(\overline{X}) \le \vartheta(\overline{Y})$. This is followed by section~\ref{parameters} in which we define and investigate quantum analogs of some graph parameters that can be defined using homomorphisms. We start with quantum chromatic number, which has been previously defined. We use the main result of Section~\ref{theta} to show that the quantum chromatic number of $X$ is lower bounded by $\vartheta(\overline{X})$. Using this bound we are able to show that the quantum colorings given in \cite{Avis} are optimal. Moving on to quantum independence number, we prove a quantum analog of a classical result relating the existence of a homomorphism from $X$ to $Y$ to the existence of an independent set of size $|V(X)|$ in a graph depending on $X$ and $Y$. We then use this to construct graphs which have large separation between classical and quantum independence number. We again use the result from Section~\ref{theta} to show that quantum independence number is upper bounded by Lov\'{a}sz theta number.
In Section~\ref{capacity} we show that the entanglement-assisted one-shot zero-error capacity of a graph is lower bounded by the quantum independence number. Furthermore, we show that if the entanglement-assisted one-shot zero-error capacity of a graph can be achieved using only projective measurements on a maximally entangled state, then it is equal to the quantum independence number. We also show that the order of entanglement-assisted one-shot zero-error capacity is respected by quantum homomorphisms, i.e.,~that $c_0^*(\overline{X}) \le c_0^*(\overline{Y})$ whenever $X \qarrow Y$.
In Section~\ref{rank} we introduce a new graph parameter which we refer to as \emph{projective rank} and denote $\xi_f$, that is a generalization of both orthogonal rank and fractional chromatic number. We show that this parameter lies between Lov\'{a}sz theta of the compliment and quantum chromatic number. We also show that the order of projective rank is respected by quantum homomorphisms, i.e.,~if $X \qarrow Y$, then $\xi_f(X) \le \xi_f(Y)$.
%
%In this article we present the notion of quantum homomorphisms as a generalization of quantum colorings. Both of these concepts are defined through games in which two separated players try to convince a verifier that they know an appropriate homomorphism or coloring.
%
%A graph homomorphism from a graph $X$ to a graph $Y$ is a map $\varphi: V(X) \rightarrow V(Y)$ such that $\varphi(x)$ and $\varphi(x')$ are adjacent in $Y$ whenever $x$ and $x'$ are adjacent in $X$. When such a map exists, we say that \emph{$X$ has a homomorphism to $Y$} and write $X \rightarrow Y$.
%
%A $c$-coloring of a graph $X$ is an assignment of ``colors" from a set of size $c$ to the vertices of $X$ such that adjacent vertices receive distinct colors. This can be viewed as a homomorphism from $X$ to $K_c$, the complete graph on $c$ vertices.
%
%Many graph theoretic problems can be formulated as questions about homomorphisms, and many graph parameters can be defined in terms of homomorphisms. As examples, we list a few such parameters below:
%\[
%\begin{array}{ll}
%\text{Chromatic number: } & \chi(X) = \min\{n \in \mathbb{N} : X \rightarrow K_n\} \\
%\text{Clique number: } & \omega(X) = \max\{n \in \mathbb{N} : K_n \rightarrow X\} \\
%\text{Independence number: } & \alpha(X) = \omega(\overline{X})
%%\og_q(X) &= \min\{n \in \mathbb{N}, \ n \text{ odd} : C_n \qarrow X\}
%\end{array}
%\]
%Here $\overline{X}$ refers to the \emph{complement of $X$}, which is the graph with the same vertex set as $X$ and two distinct vertices are adjacent in $\overline{X}$ if and only if they are not adjacent in $X$.
%
%We now take a moment to introduce some notation that will be used throughout the rest of the article.
%\begin{itemize}
%\item
%$\sim, \not\sim, \not\simeq$ : For vertices $x$ and $x'$ in some graph we write $x \sim x'$ to denote that $x$ and $x'$ are adjacent, $x \not\sim x'$ to denote that they are not adjacent (including the case where $x = x'$), and $x \not\simeq x'$ to denote that they are neither equal nor adjacent.
%\item
%We say that a graph $X$ is \emph{vertex/edge transitive} if its automorphism group acts transitively on its vertices/edges.
%\item
%An \emph{independent set} in a graph is a set of pairwise nonadjacent vertices.
%\item
%A clique in a graph is a set of pairwise adjacent vertices.
%\item
%An orthogonal projector is a Hermitian matrix $P$ such that $P^2 = P$. For brevity, we will simply refer to these as projectors.
%\item
%For a graph $X$, the graph $mX$ is the disjoint union of $m$ compies of $X$.
%\item
%The graph $K_m^n$ is the complete $m$-partite graph in which each part has size $n$, i.e. $V(K_m^n)$ can be partitioned into sets $V_i$ for $i \in [m]$ such that $x \sim x'$ if and only if they belong to different $V_i$ and each $V_i$ has size $n$.
%\item
%The Cartesian product, $X \square Y$, of $X$ and $Y$ is the graph with vertex set $V(X) \times V(Y)$ in which $(x,y) \sim (x',y')$ if ($x = x'$ and $y \sim y'$) OR ($x \sim x'$ and $y = y'$).
%\item
%The Strong product, $X \boxtimes Y$, of $X$ and $Y$ is the graph with vertex set $V(X) \times V(Y)$ in which $(x,y) \sim (x',y')$ if ($x = x'$ and $y \sim y'$) OR ($x \sim x'$ and $y = y'$) OR ($x \sim x'$ and $y \sim y'$).
%\end{itemize}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Homomorphism Game}\label{homogame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
For a graph $X$ and positive integer $c$, the $(X,c)$-coloring game is played as follows: a verifier sends Alice and Bob vertices $\xa$ and $\xb$ of $X$ respectively, and they respond with colors \mbox{$\ya,\yb \in [c]$} accordingly. The players may decide on a strategy beforehand, but are not allowed to communicate after receiving input from the verifier. To win, the following conditions need to be satisfied:
\begin{align}
&\text{if } \xa = \xb, \text{ then } \ya = \yb; \\
&\text{if } \xa \sim \xb, \text{ then } \ya \ne \yb.
\end{align}
The first condition is usually referred to as the consistency condition, while the second condition corresponds to the requirements of a proper coloring of a graph. The game is only played for one round, i.e.~one pair of inputs and one pair of outputs are involved. It is of particular interest when quantum players can win the $c$-coloring game for some graph $X$ for which classical players cannot. As mentioned above, the quantum chromatic number is the smallest $c \in \mathbb{N}$ such that quantum players can win the $(X,c)$-coloring game.
%We say that $X$ has a ``quantum $c$-coloring" if quantum players can win the $c$-coloring game for $X$, and the \emph{quantum chromatic number of $X$}, denoted $\chi_q(X)$, is the smallest number $c$ for which quantum players can win the $(X,c)$-coloring game. As we will see below, there do exists graphs $X$ such that $\chi_q(X) < \chi(X)$.
For an ordered pair of graphs $(X,Y)$, the $(X,Y)$-homomorphism game, consists of a verifier sending Alice and Bob vertices $\xa$ and $\xb$ of $X$ respectively, and Alice and Bob responding with vertices $\ya$ and $\yb$ of $Y$ accordingly. To win the $(X,Y)$-homomorphism game, the following conditions need to be satisfied:
\begin{align}
&\text{if } \xa = \xb, \text{ then } \ya = \yb; \\
&\text{if } \xa \sim \xb, \text{ then } \ya \sim \yb.
\end{align}
Like above, the first condition is the consistency condition and the second corresponds to the adjacency-preserving property of homomorphisms. When $Y = K_c$, this game reduces to the $(X,c)$-coloring game since inequality is adjacency in the complete graph. When we say that two players can ``win" a game, we mean that they can win with probability 1.
%------------------------------------
\subsection{The Classical Strategy}
%------------------------------------
%A general deterministic strategy for Alice and Bob playing the $(X,Y)$-homomorphism game is as follows: upon receiving a vertex $\xa \in V(X)$, Alice responds with a vertex $\ya \in V(Y)$ which is completely determined by her input $\xa$. Bob acts symmetrically, receiving $\xb \in V(X)$ and responding with vertex $\yb \in V(Y)$. We can think of Alice and Bob each having a table with all possible inputs in one column, and the outputs they correspond to in the column next to them. More mathematically, we can define functions $f_A, f_B : V(X) \rightarrow V(Y)$ such that Alice responds with $f_A(x)$ whenever she receives $x$ as an input, and $f_B$ is defined similarly.
Any deterministic strategy for the $(X,Y)$-homomorphism game can be specified by some functions $f_A, f_B : V(X) \rightarrow V(Y)$, where Alice responds with $f_A(x)$ whenever she receives $x$ as an input, and $f_B$ is defined similarly. Due to the consistency condition, we must have that $f_A = f_B$ for a winning strategy. Furthermore, it must be the case that $f_A$ is in fact a homomorphism from $X$ to $Y$. Conversely, if $\phi : X \rightarrow Y$ is a homomorphism, then each player responding with $\phi(x)$ upon receiving $x$ is a winning deterministic strategy for the $(X,Y)$-homomorphism game.
Classical players are in general not restricted to deterministic strategies, they are allowed probabilistic strategies which depend on shared randomness. However, in this case their strategy can be viewed as a probability distribution over deterministic strategies. Since they must win with certainty, each deterministic strategy which occurs with nonzero probability must be a winning deterministic strategy. Therefore, by the above there must still exist a homomorphism from $X$ to $Y$ in this case. Thus classical players can win the $(X,Y)$-homomorphism game if and only if $X \rightarrow Y$.
%-----------------------------------
\subsection{The Quantum Strategy}
%-----------------------------------
For quantum Alice and Bob, the most general strategy for playing the $(X,Y)$-homomorphism game is as follows: upon receiving input $x \in V(X)$, Alice performs POVM $M_x = \{E_{xy}\}_{y \in V(Y)}$ on her part of a shared state $\psi \in \mathbb{C}^{d_A} \otimes \mathbb{C}^{d_B}$ and obtains some outcome $y \in V(Y)$, which she sends to the verifier as her answer. Bob acts similarly, except that he uses POVMs $N_x = \{F_{xy}\}_{y \in V(Y)}$ for $x \in V(X)$. The probability that Alice outputs $y \in V(Y)$ and Bob outputs $y' \in V(Y)$ upon receiving inputs $x,x' \in V(X)$ respectively is given by $\psi^*(E_{xy} \otimes F_{x'y'})\psi$.
%They will win with certainty if and only if they answer incorrectly with probability zero. Therefore, in order to win the $(X,Y)$-homomomorphism game, Alice and Bob's measurements and shared state must satisfy
Therefore, to win we need that
\begin{align*}
\psi^* \left(E_{xy} \otimes F_{x'y'}\right) \psi &= 0 \text{ for } y \ne y'; \\
\psi^* \left(E_{xy} \otimes F_{x'y'}\right) \psi &= 0 \text{ for } x \sim x' \text{ and } y \not\sim y',
\end{align*}
where $y \not\sim y'$ includes the case where $y = y'$. Since quantum players can always perform at least as well as classical ones, we have that $X \rightarrow Y \Rightarrow X \qarrow Y$.
%More specifically, if $\phi$ is a homomorphism from $X$ to $Y$, then letting
%\[E_{xy} = F_{xy} = \left\{
% \begin{array}{ll}
% I & \text{if } y = \phi(x) \\
% 0 & \text{o.w.}
% \end{array}
% \right.
%\]
%gives a winning quantum strategy for the $(X,Y)$-homomorphism game.
The following theorem was proven by Cameron et. al. in \cite{qchrom} for the coloring game, and the same proof works more generally for the homomorphism game.
\begin{restatable}{theorem}{NormalForm}\label{thm:normal}
If the $(X,Y)$-homomorphism game can be won by a quantum strategy, then it can be won by a quantum strategy such that
\begin{enumerate}[\indent a.]
\item the $E_{xy}$ and $F_{xy}$ are $d \times d$ projectors for some $d \in \mathbb{N}$;
\item $E_{xy} = \overline{F}_{xy}$ for all $x \in V(X)$, $y \in V(Y)$;
\item $\psi = \frac{1}{\sqrt{d}} \sum_{i = 1}^d e_i \otimes e_i$.
\end{enumerate}
Furthermore, if $Y$ is vertex transitive, then we may assume that:
\begin{enumerate}[\indent a.]
\item[d.] The projectors $E_{xy}$ all have the same rank.
\end{enumerate}
\end{restatable}
%The idea of the last part of this proof is that Alice and Bob can take a strategy for winning the $(X,Y)$-homomorphism game and make many strategies by simply relabeling the vertices of $Y$ according to some automorphisms. They can then pick among all of these strategies according to some correlated randomness that they set up using entanglement.
%The following corollary was shown to be a direct consequence of the above theorem in \cite{qchrom} for the case of quantum coloring. The same reasoning shows that it holds for quantum homomorphisms in general.
Further following the argument of \cite{qchrom} we obtain:
\begin{cor}\label{cor:reformulation}
There exists a quantum homomorphism from $X$ to $Y$ if and only if there exist projectors $E_{xy}$ for $x \in V(X), y \in V(Y)$ such that
\begin{align}
\label{cond:ident} \sum_{y \in V(Y)} E_{xy} &= I \ \forall x \in V(X);\\
\label{cond:orthog} E_{xy}E_{x'y'} &= 0 \text{ if } (x = x' \ \& \ y \ne y') \text{ or } (x \sim x' \ \& \ y \not\sim y').
\end{align}
\end{cor}
We say that ``$E_{xy}$ are projectors that give a quantum homomorphism from $X$ to $Y$" if the $E_{xy}$ satisfy the conditions of Corrolary~\ref{cor:reformulation}.
The following lemma shows that we can further restrict to using only real projectors. We will need this fact later when proving results about Lov\'{a}sz theta.
\begin{lemma}\label{real}
There exists a quantum homomorphism from $X$ to $Y$ if and only if there exist real projectors $E_{xy}$ for all $x \in V(X), y \in V(Y)$ satisfying (\ref{cond:ident}) and (\ref{cond:orthog}) from Corollary \ref{cor:reformulation}.
\end{lemma}
\proof
The `if' direction is trivial by Corollary \ref{cor:reformulation}. To prove the other direction, suppose that $X \qarrow Y$. By Corollary \ref{cor:reformulation} there exist (possibly complex) projectors $E_{xy} \in \mathbb{M}_\mathbb{C}(d,d)$ for $x \in V(X), y \in V(Y)$ satisfying (\ref{cond:ident}) and (\ref{cond:orthog}) above. Let $R : \mathbb{M}_\mathbb{C}(d,d) \rightarrow \mathbb{M}_\mathbb{R}(2d,2d)$ be the map defined by
\[R(A) = \left(\begin{array}{cc}
\Re(A) & \Im(A) \\
-\Im(A) & \Re(A)
\end{array}\right)\]
where $\Re(A)$ and $\Im(A)$ are the real and imaginary parts of $A$ respectively. It is routine to check that $R(A + B) = R(A) + R(B)$, $R(AB) = R(A)R(B)$, and that $R$ takes Hermitian matrices to symmetric matrices. Therefore the $R(E_{xy})$ for $x \in V(X), y \in V(Y)$ are real projectors that satsify (\ref{cond:ident}) and (\ref{cond:orthog}).\qed
%For each $x \in V(X), y \in V(Y)$, let $R_{xy}, C_{xy}$ be real matrices such that $P_{xy} = R_{xy} + iC_{xy}$. Note that $R_{xy}$ is symmetric and $C_{xy}$ is skew-symmetric. For each $x \in V(X), y \in V(Y)$ define $Q_{xy}$ as follows:
%\[Q_{xy} =
%\left(\begin{array}{cc}
%R_{xy} & C_{xy} \\
%-C_{xy} & R_{xy}
%\end{array}\right).\]
%Note that $Q_{xy}$ is real and symmetric. To see that it is also a projector, consider
%\begin{align*}
%Q_{xy}^2 &= \left(\begin{array}{cc}
%R_{xy} & C_{xy} \\
%-C_{xy} & R_{xy}
%\end{array}\right)
%\left(\begin{array}{cc}
%R_{xy} & C_{xy} \\
%-C_{xy} & R_{xy}
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%R_{xy}^2 - C_{xy}^2 & R_{xy}C_{xy} + C_{xy}R_{xy} \\
%- (R_{xy}C_{xy} + C_{xy}R_{xy}) & R_{xy}^2 -C_{xy}^2
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%\Re(P_{xy}^2) & \Im(P_{xy}^2) \\
%-\Im(P_{xy}^2) & \Re(P_{xy}^2)
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%\Re(P_{xy}) & \Im(P_{xy}) \\
%-\Im(P_{xy}) & \Re(P_{xy})
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%R_{xy} & C_{xy} \\
%-C_{xy} & R_{xy}
%\end{array}\right) \\
%&= Q_{xy}
%\end{align*}
%Now suppose that $x,x' \in V(X), y,y' \in V(Y)$ are such that $P_{xy}P_{x'y'} = 0$. Then
%\begin{align*}
%Q_{xy}Q_{x'y'} &= \left(\begin{array}{cc}
%R_{xy} & C_{xy} \\
%-C_{xy} & R_{xy}
%\end{array}\right)
%\left(\begin{array}{cc}
%R_{x'y'} & C_{x'y'} \\
%-C_{x'y'} & R_{x'y'}
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%R_{xy}R_{x'y'} - C_{xy}C_{x'y'} & R_{xy}C_{x'y'} + C_{xy}R_{x'y'} \\
%- (R_{xy}C_{x'y'} + C_{xy}R_{x'y'}) & R_{xy}R_{x'y'} - C_{xy}C_{x'y'}
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%\Re(P_{xy}P_{x'y'}) & \Im(P_{xy}P_{x'y'}) \\
%-\Im(P_{xy}P_{x'y'}) & \Re(P_{xy}P_{x'y'})
%\end{array}\right) \\
%&= 0
%\end{align*}
%This shows that (\ref{cond:orthog}) from Corollary \ref{cor:reformulation} holds for the $Q_{xy}$. To see that (\ref{cond:ident}) holds as well, consider
%\begin{align*}
%\sum_{y \in V(Y)} Q_{xy} &= \sum_{y \in V(Y)} \left(\begin{array}{cc}
%R_{xy} & C_{xy} \\
%-C_{xy} & R_{xy}
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%\sum_{y \in V(Y)} R_{xy} & \sum_{y \in V(Y)} C_{xy} \\
%-\sum_{y \in V(Y)}C_{xy} & \sum_{y \in V(Y)}R_{xy}
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%\Re\left(\sum_{y \in V(Y)} P_{xy}\right) & \Im\left(\sum_{y \in V(Y)} P_{xy}\right) \\
%-\Im\left(\sum_{y \in V(Y)}P_{xy}\right) & \Re\left(\sum_{y \in V(Y)}P_{xy}\right)
%\end{array}\right) \\
%&= \left(\begin{array}{cc}
%I & 0 \\
%0 & I
%\end{array}\right) \\
%&= I
%\end{align*}
%Therefore the $Q_{xy}$ satisfy (\ref{cond:ident}) and (\ref{cond:orthog}) from Corollary \ref{cor:reformulation}.\qed
Recall from Theorem \ref{thm:normal} that if $Y$ is vertex transitive we can assume that the projectors used in a winning quantum strategy all have the same rank. This gives us the following corollary.
\begin{cor}\label{reformvtxtrans}
If $Y$ is vertex transitive, then $X \qarrow Y$ if and only if there exist rank $r$ projectors $E_{xy} \in \mathbb{M}_\mathbb{C}(d,d)$ for $x \in V(X), y \in V(Y)$ for some $r,d \in \mathbb{N}$ such that
\begin{align}
&d = r|V(Y)|; \\
&E_{xy}E_{x'y'} = 0 \text{ if } (x = x' \ \& \ y \ne y') \text{ or } (x \sim x' \ \& \ y \not\sim y').
\end{align}
\end{cor}
\proof
Suppose that such projectors exist. Since the sum of mutually orthogonal projectors is equal to identity if and only if the sum of their ranks is equal to the dimension in which they live, we have that
\[\sum_{y \in V(Y)} E_{xy} = I \ \forall x \in V(X).\]
Therefore by Corollary \ref{cor:reformulation} we have that $X \qarrow Y$.
If $X \qarrow Y$, then by Corollary \ref{cor:reformulation} we have that there exist projectors $E_{xy} \in \mathbb{M}_\mathbb{C}(d,d)$ for $x \in V(X), y \in V(Y)$ for some $d \in \mathbb{N}$ that satisfy (\ref{cond:ident}) and (\ref{cond:orthog}). Furthermore, since $Y$ is vertex transitive, we can assume that the $E_{xy}$ all have the same rank, $r$, for some $r \in \mathbb{N}$. Condition (\ref{cond:ident}) then implies that $d = r|V(Y)|$.\qed
%-------------------------------------------------------
\subsection{Basic Properties of Quantum Homomorphisms}
%-------------------------------------------------------
In this section we discuss some basic properties of quantum homomorphisms which we need later. First, we show that, like classical homomorphisms, quantum homomorphisms are transitive.
\begin{lemma}\label{lem:trans}
The relation $\qarrow$ is transitive: if $X \qarrow Y$ and $Y \qarrow Z$, then $X \qarrow Z$.
\end{lemma}
\proof
Assume that $X \qarrow Y$ and $Y \qarrow Z$. Consider the $(X,Z)$-homomorphism game; we now describe a strategy to win this game which relies only on the fact that Alice and Bob can win both the $(X,Y)$- and $(Y,Z)$-homomorphism games. Alice is given $\xa \in V(X)$ and Bob is given $\xb \in V(X)$. Alice and Bob first act as if they are playing the $(X,Y)$-homomorphism game to obtain outputs $\ya,\yb \in V(Y)$ respectively. They then act as if they are playing the $(Y,Z)$-homomorphism game with inputs $\ya$ and $\yb$ to obtain outputs $z_A,z_B \in V(Z)$ respectively. If $\xa = \xb$, then $\ya = \yb$ and then $z_A = z_B$ since Alice and Bob win the $(X,Y)$-homomorphism game and $(Y,Z)$-homomorphism game. Similarly, if $\xa \sim \xb$, then $\ya \sim \yb$ and so $z_A \sim z_B$. Therefore, Alice and Bob can win the $(X,Z)$-homomorphism game, i.e. $X \qarrow Z$.\qed
The above proof does not use that fact that the players are using a particular type of strategy, therefore this also proves that classical homomorphisms are transitive, though that is well-known and easy to see. In the quantum case, the method of ``composition" given in the proof corresponds to the following: if $\{E_{xy}: x \in V(X), y \in V(Y)\}$ are the projectors which give a quantum homomorphism from $X$ to $Y$, and $\{E'_{yz}: y \in V(Y), z \in V(Z)\}$ are the projectors which give a quantum homomorphism from $Y$ to $Z$, then the following projectors give a quantum homomorphism from $X$ to $Z$:
\[E''_{xz} = \sum_{y \in V(Y)} E_{xy}\otimes E'_{yz} \text{ for } x \in V(X), z \in V(Z).\]
The next result was given in different terms in \cite{qchrom}. It shows that when restricted to complete graphs, quantum homomorphisms behave the same as homomorphisms.
\begin{lemma}\label{complete}
$K_m \qarrow K_n$ if and only if $m \le n$.
\end{lemma}
\proof
If $m \le n$, then $K_m \rightarrow K_n$ and thus $K_m \qarrow K_n$.
Now suppose that $K_m \qarrow K_n$. Let $E_{xy}$ for $x \in V(K_m), y \in V(K_n)$ be the projectors satisfying the conditions of Corollary \ref{reformvtxtrans}. The $E_{xy}$ have rank some $r \in \mathbb{N}$ and live in $\mathbb{M}_\mathbb{C}(rn,rn)$. For fixed $y \in V(K_n)$ the projectors $\{E_{xy}\}_{x \in V(K_m)}$ are pairwise orthogonal. Therefore,
\[\rk\left(\sum_{x \in V(K_m)} E_{xy}\right) = \sum_{x \in V(K_m)} \rk\left(E_{xy}\right) = rm\]
and thus $rm \le rn$ which of course implies that $m \le n$.\qed
\begin{lemma}
Suppose that $X$ is a connected graph and $Y$ is a graph with components $Y_1,\ldots,Y_n$. If $X \qarrow Y$, then there exists $k \in [n]$ such that $X \qarrow Y_k$.
\end{lemma}
\proof
If $X$ is a single vertex then we are done, otherwise $X$ has an edge. Suppose that $X \qarrow Y$ and that $E_{xy}$ for $x \in V(X), y \in V(Y)$ are projectors satisfying the conditions of Corollary \ref{cor:reformulation}. For each $i \in [n]$ and $x \in V(X)$ define $E_{xi}$ as follows:
\[E_{xi} = \sum_{y \in V(Y_i)} E_{xy}.\]
Since $E_{xy}E_{xy'} = 0$ for $y \ne y'$, we have that the $E_{xi}$ are projectors. Furthermore, since the $V(Y_i)$ partition $V(Y)$, we have that $E_{xi}E_{xj} = 0$ for $i \ne j$. Now suppose that $x',x'' \in V(X)$ are adjacent. Since $y \not\sim y'$ for $y \in V(Y_i), y' \in V(Y_j), i \ne j$ we have that
\[E_{x'i}E_{x''j} = 0 \text{ for all } i \ne j.\]
This implies that $E_{x'i}\left(I - E_{x''i}\right) = 0$ and $\left(I - E_{x'i}\right)E_{x''i} = 0$ for all $i \in [n]$, as $\sum_{i\in[n]} E_{x'i}= \sum_{j\in[n]} E_{x''j}=I$. Therefore, $E_{x'i} = E_{x'i}E_{x''i} = E_{x''i}$. Since $X$ is connected, we further obtain that for each $i\in[n]$ there exists a projector $E_i$ such that $E_{xi} = E_i$ for all $x\in V(X)$.
Since $\sum_{i \in [n]} E_{i} = I$, we can fix a $k \in [n]$ such that $E_{k} \ne 0$. Then for all $x\in V(X), y\in V(Y_k)$ we have that
\begin{align*}
&\sum_{y \in V(Y_k)} E_{xy} = E_k \\
&E_{xy}E_{x'y'} = 0 \text{ if } (x = x' \ \& \ y \ne y') \text{ or } (x \sim x' \ \& \ y \not\sim y').
\end{align*}
Hence, when restricted to the image of $E_k$, the projectors $E_{xy}$ for $x\in V(X),y\in V(Y)$ satisfy the conditions of Corollary \ref{cor:reformulation} and $X \qarrow Y_k$ as desired.\qed
%-----------------------------------------------------------------------------
\subsection{Reformulating Quantum Homomorphisms in Terms of Homomorphisms}
%-----------------------------------------------------------------------------
Using Corollary~\ref{cor:reformulation}, we can rephrase the question of the existence of a quantum homomorphism from $X$ to $Y$ as a question of the existence of a homomorphism from $X$ to a graph, $\mathfrak{M}(Y)$, which depends only on $Y$. In theory, this may offer some assistance in the study of quantum homomorphisms, since it allows one to apply results concerning homomorphisms, which have been extensively studied. In practice it may not be of much use because the graphs $\mathfrak{M}(Y)$ seem to be difficult to analyse, and in particular contain an uncountable number of vertices.
\begin{definition}
For a finite graph $Y$ and $d \in \mathbb{N}$, define $M(Y,d)$ to be the graph whose vertices are functions $p:V(Y) \rightarrow \mathbb{M}_\mathbb{C}(d,d)$ such that $p(y)$ is a projector for all $y \in V(Y)$, and $\sum_{y \in V(Y)} p(y) = I$. Two such functions, $p$ and $p'$, are adjacent if $p(y)p'(y') = 0$ for all $y \not\sim y'$. Further, define $\mathfrak{M}(Y)$ to be the disjoint union of $M(Y,d)$ over all $d \in \mathbb{N}$.
\end{definition}
Note that the vertices of $M(Y,d)$ are exactly projective measurements in dimension $d$ whose outcomes are indexed by the vertices of $Y$. We have the following:
\begin{theorem}\label{qhom2homd}
If $X$ and $Y$ are graphs, then $X \qarrow Y$ if and only if $X \rightarrow M(Y,d)$ for some $d \in \mathbb{N}$.
\end{theorem}
\proof
Suppose that $\phi: X \rightarrow M(Y,d)$ for some $d \in \mathbb{N}$. For all $x \in V(X)$ and $y \in V(Y)$, let $E_{xy} = \phi(x)(y)$. It is straightforward to check that the $E_{xy}$ satisfy the conditions of Corollary~\ref{cor:reformulation} and thus $X \qarrow Y$.
Conversely, if $X \qarrow Y$, then there exist projectors $E_{xy}$ satisfying the conditions of Corollary~\ref{cor:reformulation}. For each $x \in V(X)$, define $p_x : V(Y) \rightarrow \mathbb{M}_\mathbb{C}(d,d)$ as $p_x(y) = E_{xy}$. The function $\phi : V(X) \rightarrow V(M(Y,d))$ defined by $\phi(x) = p_x$ is clearly a homomorphism from $X$ to $M(Y,d)$.
\qed
If we really want to only consider a single target graph instead of a family, then we can use $\mathfrak{M}(Y)$:
\begin{theorem}\label{qhom2hom}
If $X$ and $Y$ are graphs, then $X \qarrow Y$ if and only if $X \rightarrow \mathfrak{M}(Y)$.
\end{theorem}
\proof
If $X \qarrow Y$ then by Theorem~\ref{qhom2homd} there exists $d \in \mathbb{N}$ such that $X \rightarrow M(Y,d)$. Since $M(Y,d)$ is a subgraph of $\mathfrak{M}(Y)$, we have that $X \rightarrow \mathfrak{M}(Y)$.
Now suppose that $\phi: X \rightarrow \mathfrak{M}(Y)$. If $X'$ is a component of $X$, then the image of $X'$ under $\phi$ is connected and therefore must be contained in $M(Y,d)$ for some $d \in \mathbb{N}$. Therefore $X' \qarrow Y$ by Theorem~\ref{qhom2homd}, and similarly all components of $X$ have a quantum homomorphism from $X$ to $Y$. Now by considering the $(X,Y)$-homomorphism game, it is easy to see that $X$ must have a quantum homomorphism to $Y$: each player plays according to the strategy required for the component of $X$ from which their input vertex comes.\qed
%--------------------------------------------------------
\subsection{Non-signalling Strategies}\label{nonsignal}
%--------------------------------------------------------
We here consider the case in which Alice and Bob have access to more general nonlocal correlations than those produced by shared entanglement. Mathematically, a correlation can be described as a joint conditional probability distribution $P(Y_A, Y_B | X_A, X_B)$, where $X_A$ and $X_B$ are random variables representing Alice and Bob's respective inputs to the correlation, and $Y_A$ and $Y_B$ their respective outputs. The correlation is \emph{non-signalling} (with respect to the partition $\{(X_A,Y_A),(X_B,Y_B)\}$) if
\begin{align*}
P(Y_A = y | X_A = x, X_B = x') &= P(Y_A = y | X_A = x, X_B = x'') \ \ \ \text{for all } x,x',x'',y \\
P(Y_B = y | X_A = x', X_B = x) &= P(Y_B = y | X_A = x'', X_B = x) \ \ \ \text{for all } x,x',x'',y .
\end{align*}
The non-signalling restriction corresponds to the fact that the correlation cannot be used by Alice and Bob to communicate information to one another.
It turns out that players with access to arbitrary non-signalling correlations can play the homomorphism game just as well as if they were able to communicate during the game. To see this note that for any graph $X$, the $(X,K_2)$-homomorphism game can be won using the following non-signalling correlation:
\begin{displaymath}
P(Y_A = \ya, Y_B = \yb| X_A = \xa, X_B = \xb) = \left\{
\begin{array}{ll}
\frac{1}{2} & \text{if } \delta(\xa,\xb) = \delta(\ya,\yb) \\
0 & \text{o.w.}
\end{array}
\right.
\end{displaymath}
Here $X_A, X_B$ take values from $V(X)$ and $Y_A, Y_B$ take values from $V(K_2)$, and $\delta$ is the Kronecker delta function. Note that there is no way to win the $(X,K_1)$-homomorphism game if $X$ is not empty, since if the players are given adjacent vertices they will always lose.
%In general, Alice and Bob's strategy may not be limited to the use of quantum mechanics. In the \emph{non-signalling} model for nonlocal games, the only restriction is that the players' strategy may not transmit information from one party to another. For a two player game, if we let $X_A$ and $X_B$ be the random variable representing Alice and Bob's respective inputs, and $Y_A$ and $Y_B$ the random variables representing their respective outputs, then the mathematical description of the non-signalling restriction is given by
%\begin{align*}
%P(Y_A = y | X_A = x, X_B = x') &= P(Y_A = y | X_A = x, X_B = x'') \ \ \ \forall x,x' \in V(X), \ y \in V(Y) \\
%P(Y_B = y | X_A = x', X_B = x) &= P(Y_B = y | X_A = x'', X_B = x) \ \ \ \forall x,x' \in V(X), \ y \in V(Y).
%\end{align*}
%In other words, Alice cannot gain any information regarding Bob's input and vice versa. If a joint conditional probability distribution satisfies the above two conditions then we say that it is \emph{non-signalling} (with respect to the partition $\{(X_A,Y_A),(X_B,Y_B)\}$). Restricting to non-signalling strategies turns out to be no restriction at all, as it allows players to win the $(X,K_2)$-homomorphism game for any graph $X$. To see this consider the following joint conditional probability distribution:
%\begin{displaymath}
% P(Y_A = \ya, Y_B = \yb| X_A = \xa, X_B = \xb) = \left\{
% \begin{array}{ll}
% \frac{1}{2} & \text{if } \delta(\xa,\xb) = \delta(\ya,\yb) \\
% 0 & \text{o.w.}
% \end{array}
% \right.
%\end{displaymath}
%Here $\delta$ is the Kronecker delta function. Note that there is no way to win the $(X,K_1)$-homomorphism game if $X$ is not empty, since if the players are given adjacent vertices they will always lose. This means that players able to use arbtitrary non-signalling strategies can play the homomorphism game just as well as if they were able to communicate during the game.
%
%Perhaps there is some deeper theory that allows them to win even more instances of the homomorphism game. This motivates the consideration of strategies whose only restriction is that Alice and Bob are not able to communicate any information to each other. One can imagine that Alice and Bob each have some box which accepts a vertex of $X$ as an input and gives a vertex of $Y$ as an output (in a reasonable amount of time). If Alice and Bob are sufficiently far apart during the game, then as long as nothing can exceed the speed of light (a speed limit obeyed by seemingly all non-neutrino based entities), then it is not possible for Alice, or her box, to communicate her input to Bob, or his box, in any way. Note that the box need not be deterministic, as is the case when it represents performing quantum measurements on a shared state. If we let $X_A$ and $X_B$ be the random variable representing Alice and Bob's respective inputs, and $Y_A$ and $Y_B$ the random variables representing their respective outputs, then the mathematical description of not allowing any information to pass between the two parties is given below:
%\begin{align*}
%P(Y_A = y | X_A = x, X_B = x') &= P(Y_A = y | X_A = x, X_B = x'') \ \ \ \forall x,x' \in V(X), \ y \in V(Y) \\
%P(Y_B = y | X_A = x', X_B = x) &= P(Y_B = y | X_A = x'', X_B = x) \ \ \ \forall x,x' \in V(X), \ y \in V(Y).
%\end{align*}
%If a joint conditional probability distribution satisfies these two conditions then we say that it is \emph{non-signalling} (with respect to the partition $\{(X_A,Y_A),(X_B,Y_B)\}$). Restricting to non-signalling strategies turns out to be no restriction at all as it allows players to win the $(X,K_2)$-homomorphism game for any graph $X$. To see this consider the following joint conditional probability distribution:
%\begin{displaymath}
% P(Y_A = \ya, Y_B = \yb| X_A = \xa, X_B = \xb) = \left\{
% \begin{array}{ll}
% \frac{1}{2} & \text{if } \delta(\xa,\xb) = \delta(\ya,\yb) \\
% 0 & \text{o.w.}
% \end{array}
% \right.
%\end{displaymath}
%Here $\delta$ is the Kronecker delta function. Note that there is no way to win the $(X,K_1)$-homomorphism game if $X$ is not empty, since if the players are given adjacent vertices they will always lose. This means that players able to use arbtitrary non-signalling strategies can play the homomorphism game just as well as if they were able to communicate during the game.
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Lov\'{a}sz $\vartheta$}\label{theta}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
The Lov\'{a}sz theta number of a graph $X$, denoted $\vartheta(X)$, was introduced in \cite{lovasz} as an upper bound on the Shannon capacity $\Theta(X)$. It was later shown to upper bound the entanglement-assisted Shannon capacity $\Theta^*(X)$ as well \cite{Winter,Beigi}. Since $\vartheta$ admits a semidefinite program formulation, it can be computed efficiently and is thus of great practical use. In this paper we will be mostly interested in Lov\'{a}sz theta of the complement of a graph, denoted $\bartheta(X)$, as this is preserved by homomorphisms. There are various definitions for $\vartheta$, and hence $\bartheta$, but the one below suits us since it directly defines $\bartheta$ in terms of homomorphisms.
\begin{defn}\label{LVhomo}
For $m \in \mathbb{N}$ and $\alpha < 0$, let $S(m,\alpha)$ be the infinite graph whose vertices are unit vectors in $\mathbb{R}^m$ and two vectors $u$ and $v$ are adjacent if $u^T v = \alpha$. Then
\[\bartheta(X) = \min \left\{1 - \frac{1}{\alpha} : X \rightarrow S(m,\alpha), \ \alpha < 0 \right\}.\]
Note that $m$ is not restrained, but we can assume that $m = |V(X)|$ if we like, since any set of $|V(X)|$ vectors lives in a space of dimension at most $|V(X)|$.
\end{defn}
From this definition it is clear that if $X \rightarrow Y$, then $\bartheta(X) \le \bartheta(Y)$. Surprisingly, the same is true for quantum homomorphisms.
\begin{theorem}\label{thetapreserved}
If $X \qarrow Y$, then $\bartheta(X) \le \bartheta(Y)$.
\end{theorem}
\proof
Suppose that $X \qarrow Y$ and $Y \rightarrow S(m,\alpha)$ for some $m \in \mathbb{N}$ and $\alpha < 0$. To prove the theorem it suffices to show that $X \rightarrow S(m',\alpha)$ for some $m' \in \mathbb{N}$. Since $Y \rightarrow S(m,\alpha)$, there exist unit vectors $v_y \in \mathbb{R}^m$ for $y \in V(Y)$ such that $v_y^T v_{y'} = \alpha$ whenever $y \sim y'$. Since $X \qarrow Y$, Lemma~\ref{real} implies that there exist \emph{real} projectors $E_{xy}$ in some dimension $d$ that give a quantum homomorphism from $X$ to $Y$. Define vectors $u_x$ for $x \in V(X)$ by
\[u_x = \frac{1}{\sqrt{d}}\sum_{y \in V(Y)} v_y \otimes \vect(E_{xy})\]
where $\vect(M)$ is the vector obtained by stacking the rows of $M$ on top of each other. Then we have
\begin{align*}
u_x^T u_{x'} &= \frac{1}{d}\left(\sum_{y \in V(Y)} v_y \otimes \vect(E_{xy})\right)^T \left(\sum_{y' \in V(Y)} v_{y'} \otimes \vect(E_{x'y'})\right) \\
&= \frac{1}{d}\sum_{y,y' \in V(Y)} v_y^T v_{y'} \tr(E_{xy}E_{x'y'}).
\end{align*}
Since $\tr(E_{xy}E_{xy'}) = 0$ for all $y \ne y'$, we further get
\begin{align*}
u_x^T u_x &= \frac{1}{d}\sum_{y \in V(Y)} v_y^T v_y \tr(E_{xy}E_{xy}) \\
&= \frac{1}{d}\tr\left(\sum_{y \in V(Y)} E_{xy}\right) \\
&= \frac{1}{d}\tr(I) = 1.
\end{align*}
So the $u_x$ are unit vectors and now we just need to check that $u_x^T u_{x'} = \alpha$ whenever $x \sim x'$. In this case, we have that $\tr(E_{xy}E_{x'y'}) = 0$ whenever $y \not\sim y'$ and so
\begin{equation}\label{eq:theta}
\begin{split}
u_x^T u_{x'} &= \frac{1}{d} \sum_{y \sim y'} v_y^T v_{y'} \tr(E_{xy}E_{x'y'}) = \frac{\alpha}{d} \sum_{y \sim y'} \tr(E_{xy}E_{x'y'}) \\
&= \frac{\alpha}{d} \sum_{y,y' \in V(Y)} \tr(E_{xy}E_{x'y'}) = \frac{\alpha}{d} \tr\left(\left(\sum_{y \in V(Y)}E_{xy}\right)\left(\sum_{y' \in V(Y)}E_{x'y'}\right)\right) \\
&= \alpha.
\end{split}
\end{equation}
Hence $X \rightarrow S(md^2,\alpha)$ and we are done.
\qed
As we will see in the next section, this theorem can be used to bound quantum parameters defined in terms of homomorphisms, such as quantum chromatic number. In general, no algorithm is known for deciding whether $X \qarrow Y$. However, if $\bartheta(X) > \bartheta(Y)$, the above theorem allows us to conclude that $X$ does not admit a quantum homomorphism to $Y$.
There is another graph parameter, $\chi_\text{vec}$, known as the \emph{vector chromatic number} \cite{chivec}, which is closely related to $\bartheta$. The only difference is that in the definition of $\chi_\text{vec}$, the vectors assigned to adjacent vertices must have inner product at most some $\alpha < 0$, as opposed to requiring equality. The analog of the theorem above for $\chi_\text{vec}$ holds, and the same proof works with the exception of changing the second '=' in Equation (\ref{eq:theta}) to '$\le$', and noting that $\tr(E_{xy}E_{x'y'}) \ge 0$ since both matrices are positive semidefinite.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%$$$$$$$$$
\section{Quantum Parameters}\label{parameters}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%$$$$$$$$$
%The quantum chromatic number, $\chi_q(X)$, is defined as the smallest $c$ for which quantum players can win the $(X,c)$-coloring game with certainty \cite{Galliard02,Cleve04}. In terms of quantum homomorphisms, this is equivalent to the smallest $c$ such that $X \qarrow K_c$. This parameter has been relatively well-studied \cite{Avis,qchrom,qchrom-aqis,SS12,MSS}. In particular, it is known that for the family of graphs $\Omega_{4n}$ there is an exponential separation between $\chi(\Omega_{4n})$ and $\chi_q(\Omega_{4n})$. Here, $\Omega_n$ is the graph with vertex set $\{\pm 1\}^n$ and $v \sim w$ if $v^T w = 0$. Also, a complete characterization of graphs with $\chi_q(X)<\chi(X)$ has been given \cite{MSS}. However, many questions remain open. For example, it is not known whether $\chi_q(X)$ is computable, or whether there exists a family of graphs $X_n$ such that $\lim_{n\rightarrow\infty}\chi(X_n)=\infty$ but $\lim_{n\rightarrow\infty}\chi_q(X_n)<\infty$. A systematic study of quantum homomorphisms could aid in answering these and other questions.
%%We hope that systematic study of quantum homomorphisms will aid in answering these and other questions. (See Section~\ref{sec:Lovasz} for new results concerning $\chi_q$ obtained via our study of quantum homomorphisms.)
%
%Using the framework of quantum homomorphisms, we can introduce a quantum analogue for any graph parameter defined in terms of graph homomorphisms (e.g., clique number, independence number, odd girth, etc.). Here we only consider the following:
%%Using quantum homomorphisms, we can define quantum analogues of other graph parameters. Here, we only consider the following:
%\begin{itemize}
% \item quantum chromatic number, $\chi_q(X) = \min\{n : X \qarrow K_n\}$;
% \item quantum clique number, $\omega_q(X) = \max\{n : K_n \qarrow X\}$;
% \item quantum independence number, $\alpha_q(X) = \omega_q(\overline{X})$.
%% \item quantum odd girth, $\og_q(X) = \min\{n : C_n \qarrow X, \ n \text{ odd}\}$, where $C_n$ is the cycle on $n$ vertices.
%\end{itemize}
%
%We are about to see that quantum clique and independence number can be different from their classical counterparts. Moreover, we show how to construct a graph with such a separation using any two graphs $X$ and $Y$ such that $X \qarrow Y$ but $X \not\rightarrow Y$.
There are many graph parameters that can be defined in terms of homomorphisms. Having defined quantum homomorphisms, it is natural to define quantum analogs of such graph parameters by simply replacing ``homomorphism'' with ``quantum homomorphism" in the definition. Here we consider the following:
\[
\begin{array}{ll}
\text{quantum chromatic number: } & \chi_q(X) := \min\{n \in \mathbb{N} : X \qarrow K_n\}; \\
\text{quantum clique number: } & \omega_q(X) := \max\{n \in \mathbb{N} : K_n \qarrow X\}; \\
\text{quantum independence number: } & \alpha_q(X) := \omega_q(\overline{X}); \\
\text{quantum odd girth: } & \og_q(X) := \min\{n \in \mathbb{N}, \ n \text{ odd} : C_n \qarrow X\}.
\end{array}
\]
The above definitions of quantum clique and independence numbers are different from those given in \cite{Beigi} and \cite{Winter}. The quantum clique number of \cite{Beigi} and the various independence numbers of \cite{Winter} are defined in terms of the amount of quantum or classical information one can send over a quantum channel, and are therefore more analogous to capacities than our notion of quantum independence number.
It is known that there exist graphs $X$ such that $\chi_q(X) < \chi(X)$ \cite{Avis,qchrom,qchrom-aqis,SS12,MSS}; we will see that there can also be (large) separations between $\alpha(X)$ and $\alpha_q(X)$ (and thus between $\omega(X)$ and $\omega_q(X)$). We are not currently aware of any graphs with a separation between $\og(X)$ and $\og_q(X)$. Theorem~\ref{thetapreserved} along with the known values of $\bartheta(C_n)$ can be used to show that for odd $m$ and $n$, $C_n \qarrow C_m$ if and only if $C_n \rightarrow C_m$ (i.e. $m \le n$). However, this does not preclude the existence of separations between odd girth and quantum odd girth.
%Since $\qarrow$ is transitive, we have that $X \qarrow Y$ implies $\chi_q(X) \le \chi_q(Y), \omega_q(X) \le \omega_q(Y), \og_q(X) \ge \og_q(Y),$ etc. We can use these statements to derive some simple facts about quantum parameters. In particular, it is easy to see that $\chi_q(X) = 1$ if and only if $X$ is empty if and only if $\chi(X) = 1$, and $\chi_q(X) = 2$ if and only if $X$ is non-empty and bipartite if and only if $\chi(X) = 2$ (these facts are not new, the latter was proven in \cite{qchrom}). This trend does not continue, indeed in \cite{qchrom-aqis} they exhibit a graph which has quantum chromatic number 3, but chromatic number 4. In the next section we will see that the separation can be much larger.
%---------------------------------------
\subsection{Quantum Chromatic Number}
%---------------------------------------
The quantum chromatic number has been relatively well-studied \cite{Avis,qchrom,qchrom-aqis,SS12,MSS}. In particular a family of graphs exhibiting an exponential separation \cite{BCT99,BCW98,Avis}. In \cite{qchrom}, the authors show that when restricted to using rank one projectors, larger than than exponential separations cannot be achieved. However, in \cite{qchrom-aqis} it is shown that rank one projectors are not always sufficient to attain $\chi_q(X)$. Despite these advances, some fundamental questions regarding quantum chromatic number remain unsolved. For example, it is not known whether $\chi_q(X)$ is computable or whether there exists a sequence of graphs $X_n$ for which $\lim_{n \rightarrow \infty} \chi(X_n) = \infty$ but $\lim_{n \rightarrow \infty} \chi_q(X_n) < \infty$. %We hope that a systematic study of quantum homomorphisms would aid in answering these and other questions.
Our main result concerning quantum chromatic number is the following lower bound which is a direct consequence of Theorem~\ref{thetapreserved}:
\begin{cor}\label{chiqlowerbound}
For any graph $X$,
\[\bartheta(X) \le \chi_q(X).\]
\end{cor}
\proof
Let $c = \chi_q(X)$. Then $X \qarrow K_c$ and by Theorem~\ref{thetapreserved}, $\bartheta(X) \le \bartheta(K_c) = c$.\qed
This lower bound on $\chi_q(X)$ is new\footnote{Through private communication we have learned that $\vartheta(\overline{X}) \le \chi_q(X)$ has been independently proven (but not published) by M.~Laurent, G.~Scarpa, and A.~Varvitsiotis.} and, as $\omega(X) \le \bartheta(X)$, it improves upon the best previously known lower bound, $\omega(X)$. Furthermore, since $\omega(X)$ is NP-hard to compute, the above is also an improvement in terms of efficiency. This is particularly useful since it is not known if $\chi_q(X)$ is even computable.
since $\chi_q$ is not known to even be computable, it is useful to be able to efficiently bound it.
We now use this lower bound to compute the exact value of the quantum chromatic number for a well-known class of graphs. For $n \in \mathbb{N}$, let $\Omega_n$ be the graph whose vertices are the $\pm 1$ vectors of length $n$ with orthogonal vectors being adjacent. We only consider the case when $4 | n$, since otherwise $\Omega_n$ is either empty or bipartite. In such a case, a result of Frankl and R\"{o}dl \cite{frankrod} implies that $\chi(\Omega_n)$ is exponential in $n$. On the other hand, it is known that $\chi_q(\Omega_n) \le n$ for all $n \in \mathbb{N}$ \cite{Avis}. However, it it remained unknown whether this inequality is tight for $4 | n$. To show this, we first compute $\bartheta(\Omega_n)$ for $4 | n$.
\begin{lemma}\label{thetahadamard}
If $n$ is divisible by 4, then $\bartheta(\Omega_n) = n$.
\end{lemma}
\proof\hspace{-.07in}\footnote{This proof is a minor modification of a proof from unpublished notes of Simone Severini.}
Let $n = 4m$. In $\cite{lovasz}$, it was shown that if $X$ is regular (all vertices have equal degree) and edge transitive, then
\[\vartheta(X) = |V(X)|\frac{\lambda_{\min}}{\lambda_{\min} - \lambda_{\max}}\]
where $\lambda_{\max}$ and $\lambda_{\min}$ are the largest and smallest eigenvalues of the adjacency matrix of $X$, respectively. Since $\Omega_n$ is regular and edge transitive, we can use this equality to compute $\vartheta(\Omega_n)$. The maximum eigenvalue of a regular graph is just the degree of the graph, and so for $\Omega_n$ this is equal to $\binom{4m}{2m}$. The minimum eigenvalue of $\Omega_n$ for $4|n$ was computed in \cite{interestinggraphs} and is equal to $-\binom{4m}{2m}/(4m-1)$. Therefore,
\[\vartheta(\Omega_n) = 2^n\frac{-\frac{\binom{4m}{2m}}{4m-1}}{-\frac{\binom{4m}{2m}}{4m-1}-\binom{4m}{2m}} = 2^n\frac{1}{1+(4m-1)} = \frac{2^n}{n}.\]
Another result from \cite{lovasz} states that for a vertex transitive graph $X$, \[\vartheta(X)\bartheta(X) = |V(X)|.\]
Since $\Omega_n$ is vertex transitive as well, we obtain that
\[\bartheta(\Omega_n) = n\]
as desired.\qed
Combining the above with Corollary~\ref{chiqlowerbound} yields the following:
\begin{cor}
If $n$ is divisible by 4, then $\chi_q(\Omega_n) = n$.\qed
\end{cor}
%An interesting situation arises from this result: the graphs $\Omega_n$ for $4 | n$ are an infinite family of graphs for which the quantum chromatic number is known exactly, but the chromatic number is not (for $n > 8$). %To the best of the authors' knowledge, there is not even a conjectured value for $\chi(\Omega_n)$.
Curiously, $\{\Omega_{4n}\}_{n \in \mathbb{N}}$ is an infinite family of graphs for which the quantum chromatic number is known exactly, while the chromatic number remains unknown (for $n > 2$).
%--------------------------------------------------------------------
\subsection{Quantum Independence and Clique Numbers}\label{alpha_q}
%--------------------------------------------------------------------
In this section we show that any graphs $X$ and $Y$ such that $X \qarrow Y$ but $X \not\rightarrow Y$ can be used to construct a graph $Z$ for which $\alpha_q(Z) > \alpha(Z)$, and thus $\omega_q(\overline{Z}) > \omega(\overline{Z})$. Such constructions show that quantum advantage can also occur for homomorphism games with non-complete target graphs. We begin with a definition from \cite{Nesetril}:
\begin{definition}
For graphs $X$ and $Y$, define their \emph{homomorphic product}, denoted $X \ltimes Y$, to be the graph with vertex set $V(X) \times V(Y)$ with distinct vertices $(x,y)$ and $(x',y')$ being adjacent if either $x = x'$, or $x \sim x'$ and $y \not\sim y'$.
\end{definition}
It is useful to consider this product because the edges in $X \ltimes Y$ correspond exactly to the required orthogonalities for projectors giving a quantum homomorphism from $X$ to $Y$. We will often use the following two properties:
\begin{align}
\label{prop:omega} &K_{|V(Y)|} \rightarrow X \ltimes Y \\
\label{prop:chi} &\overline{X \ltimes Y} \rightarrow K_{|V(X)|}.
\end{align}
Both of these properties follow from the fact that the sets $V_x = \{(x,y): y \in V(Y)\}$ for $x \in V(X)$ induce cliques of size $|V(Y)|$ and partition $V(X \ltimes Y)$.
The following lemma is the original motivation for defining the homomorphic product, and it is stated as Exercise 7 in Chapter 2 of \cite{Nesetril}.
\begin{lemma}\label{lem:hom2indy}
For graphs $X$ and $Y$, we have that $X \rightarrow Y$ if and only if $\alpha(X \ltimes Y) = |V(X)|$.
\end{lemma}
\proof
Let $m = |V(X)|$ and $n = |V(Y)|$ and note that
\[\alpha(X \ltimes Y) = \omega(\overline{X \ltimes Y}) \le \chi(\overline{X \ltimes Y}) \le m\]
by Property~(\ref{prop:chi}). So $\alpha(X \ltimes Y) = m$ if and only if $X \ltimes Y$ has an independent set of size $m$.
Suppose that $\varphi : X \rightarrow Y$ is a homomorphism. Then the set $\{(x,\varphi(x)): x \in V(X)\}$ is an independent set of size $m$ in $X \ltimes Y$.
Conversely, let $S$ be an independent set of size $m$ in $X \ltimes Y$. Since the set $\{(x,y) : y \in V(Y)\}$ induces a clique for all $x \in V(X)$, and there are $m$ such sets partitioning $V(X\ltimes Y)$, there must be exactly one vertex of $S$ whose first coordinate is $x$ for every $x \in V(X)$. Define $\phi: V(X) \rightarrow V(Y)$ as follows:
\[\phi(x) = y \text{ for the unique } y \text{ such that } (x,y) \in S.\]
It is straightforward to check that $\phi$ is a homomorphism.\qed
Somewhat surprisingly, a quantum analog of this lemma also holds.
\begin{lemma}\label{lem:qhom2qindy}
For graphs $X$ and $Y$, we have that $X \qarrow Y$ if and only if $\alpha_q(X \ltimes Y) = |V(X)|$.
\end{lemma}
\proof
Let $m = |V(X)|$ and $n = |V(Y)|$. Note that
\[\alpha_q(X \ltimes Y) = \omega_q(\overline{X \ltimes Y}) \le \chi_q(\overline{X \ltimes Y}) \le \chi(\overline{X \ltimes Y}) \le m\]
by Property~(\ref{prop:chi}). So $\alpha_q(X \ltimes Y) = m$ if and only if $K_m \qarrow \overline{X \ltimes Y}$.
Suppose that $X \qarrow Y$ and consider Alice and Bob playing the $(K_m,\overline{X \ltimes Y})$-homomorphism game. We will describe a winning strategy for this game, thus proving one direction of the lemma. Since $m = |V(X)|$, we can assume that $K_m$ and $X$ have the same vertex set. When Alice and Bob are given inputs $\xa,\xb \in V(K_m) = V(X)$, they act as if they are playing the $(X,Y)$-homomorphism game with $\xa$ and $\xb$ as their inputs in order to obtain $\ya, \yb \in V(Y)$. They then output $(\xa,\ya)$ and $(\xb,\yb)$ respectively. If $\xa = \xb$, then $\ya = \yb$ and thus $(\xa,\ya) = (\xb,\yb)$. If $\xa \ne \xb$, then either $\xa \sim \xb$ in $X$ or $\xa \not\simeq \xb$ in $X$ (note that $\not\simeq$ means ``neither equal nor adjacent to" and is equivalent to ``adjacent in the complement"). In the former case, we will have that $\ya \sim \yb$ and thus $(\xa,\ya) \not\simeq (\xb,\yb)$ in $X \ltimes Y$ and therefore $(\xa,\ya) \sim (\xb,\yb)$ in $\overline{X \ltimes Y}$. In the latter case, we have that $(\xa,\ya) \not\simeq (\xb,\yb)$ in $X \ltimes Y$ regardless of $\ya$ and $\yb$, and therefore $(\xa,\ya) \sim (\xb,\yb)$ in $\overline{X \ltimes Y}$. So Alice and Bob can win the $(K_m,\overline{X \ltimes Y})$-homomorphism game and therefore $\alpha_q(X \ltimes Y) = |V(X)|$.
Conversely, suppose that $\alpha_q(X \ltimes Y) = m$, and that $P_{i(x,y)} \in \mathbb{M}_\mathbb{C}(d,d)$ for $i \in [m], \ x \in V(X), \ y \in V(Y)$ are the projectors corresponding to a winning strategy for the $(K_m,\overline{X \ltimes Y})$-homomorphism game. Then for all $i \in [n]$
\[\sum_{x \in V(X), y \in V(Y)} P_{i(x,y)} = I\]
and
\[P_{i(x,y)}P_{j(x',y')} = 0\]
whenever $\left(i = j \ \& \ (x,y) \ne (x',y')\right)$ OR $\left(i \ne j \ \& \ (x,y) \not\sim (x',y') \text{ in } \overline{X \ltimes Y}\right)$. Considering how adjacency is defined in $X \ltimes Y$, we can rewrite the orthogonality conditions above as $P_{i(x,y)}P_{j(x',y')} = 0$ if one of the following hold:
\begin{enumerate}[a.]
\item
$i = j$ and $(x,y) \ne (x',y')$, or
\item
$i \ne j$ and $x = x'$, or
\item
$i \ne j$ and $x \sim x'$ and $y \not\sim y'$.
\end{enumerate}
For each $x \in V(X), y \in V(Y)$ define
\[Q_{xy} = \sum_{i \in [m]} P_{i(x,y)}.\]
We claim that the $Q_{xy}$ are projectors which satisfy the conditions of Corollary \ref{cor:reformulation} for the $(X,Y)$-homomorphism game. They are indeed projectors since by (b) they are the sum of mutually orthogonal projectors.
%First we must check that they are projectors:
%\begin{align*}
%Q_{xy}^2 &= \left(\sum_{i \in [m]} P_{i(x,y)}\right)\left(\sum_{j \in [m]} P_{j(x,y)}\right) = \sum_{i,j \in [m]} P_{i(x,y)}P_{j(x,y)} \\
%&= \sum_{i \in [m]} P^2_{i(x,y)} = \sum_{i \in [m]} P_{i(x,y)} = Q_{xy}
%\end{align*}
%where the third equality is due to 2 above.
Now we must check that the $Q_{xy}$ satisfy the orthogonality conditions (\ref{cond:orthog}):
\begin{align*}
Q_{xy}Q_{x'y'} &= \left(\sum_{i \in [m]} P_{i(x,y)}\right)\left(\sum_{j \in [m]} P_{j(x',y')}\right) = \sum_{i,j \in [m]} P_{i(x,y)}P_{j(x',y')} \\
&= \left\{
\begin{array}{ll}
0 & \text{if } x = x' \ \& \ y \ne y' \text{ by (a) and (b)} \\
0 & \text{if } x \sim x' \ \& \ y \not\sim y' \text{ by (a) and (c)} \\
\text{something} & \text{o.w.}
\end{array}
\right.
\end{align*}
These are exactly the needed orthogonalities and so it is only left to show that for all $x \in V(X)$
\begin{equation}\label{eq:Q}
\sum_{y \in V(Y)} Q_{xy} = I.
\end{equation}
Since for a fixed $x \in V(X)$, the $Q_{xy}$ are mutually orthogonal, we have that $\sum_{y \in V(Y)} \rk(Q_{xy}) \le d$ with equality if and only if (\ref{eq:Q}) holds. By (a) we have that
\[\sum_{x \in V(X), y \in V(Y)} \rk\left(P_{i(x,y)}\right) = \rk\left(\sum_{x \in V(X), y \in V(Y)} P_{i(x,y)}\right) = \rk(I) = d\]
for all $i \in [m]$. Therefore,
\begin{align*}
md &= \sum_{i \in [m]}\left(\sum_{x \in V(X), y \in V(Y)} \rk\left(P_{i(x,y)}\right)\right) = \sum_{x \in V(X), y \in V(Y)} \rk\left(\sum_{i \in [m]}P_{i(x,y)}\right) \\
&= \sum_{x \in V(X)} \left(\sum_{y \in V(Y)} \rk\left(Q_{xy}\right)\right) \le |V(X)|d = md
\end{align*}
where the second equality follows from (b). Of course this implies that $\sum_{y \in V(Y)} \rk(Q_{xy}) = d$ for all $x \in V(X)$ and thus (\ref{eq:Q}) holds and we are done.\qed
%Let $d \in \mathbb{N}$ be such that $P_{i(x,y)} \in M_{\mathbb{C}}(d,d)$. Then by (a) we have that
%\[\sum_{i \in [m], x \in V(X), y \in V(Y)} \rk\left(P_{i(x,y)}\right) = \sum_{i \in [m]} \rk\left(\sum_{x \in V(X), y \in V(Y)} P_{i(x,y)}\right) = \sum_{i \in [m]} \rk(I) = md.\]
%%since
%%\[\sum_{x \in V(X), y \in V(Y)} P_{i(x,y)} = I \text{ for all } i \in [m],\]
%%we have that
%%\[\sum_{i \in [m], x \in V(X), y \in V(Y)} \rk\left(P_{i(x,y)}\right) = md.\]
%However, by the above, $Q_{xy}Q_{xy'} = 0$ for $y \ne y'$ and $P_{i(x,y)}P_{j(x,y)} = 0$ for $i \ne j$, therefore
%\[\sum_{y \in V(Y)} \sum_{i \in[m]} \rk\left(P_{i(x,y)}\right) = \sum_{y \in V(Y)} \rk\left(Q_{xy}\right) = \rk\left(\sum_{y \in V(Y)} Q_{xy}\right) \le d \text{ for all } x \in V(X).\]
%But this implies that
%\[md = \sum_{x \in V(X)}\left(\sum_{y \in V(Y),i \in [m]} \rk\left(P_{i(x,y)}\right)\right) \le |V(X)|d = md.\]
%Therefore we must have that
%\[\rk\left(\sum_{y \in V(Y)} Q_{xy}\right) = d \text{ for all } x \in V(X),\]
%and thus
%\[\sum_{y \in V(Y)} Q_{xy} = I \text{ for all } x \in V(X)\]
%as desired.\qed
Interestingly, one direction of the above lemma can be proved without appealing to the actual strategy being used, while the other direction cannot. Indeed, if $X$ is the path on three vertices, and $Y = K_1$, then there is no non-signalling correlation which wins the $(X,Y)$-homomorphism game (if Alice and Bob receive adjacent vertices they have no way of winning). However, $\overline{X \ltimes Y}$ is a single edge plus an isolated vertex, and so any graph (including arbitrarily large cliques) can be mapped to $\overline{X \ltimes Y}$ using a non-signalling correlation as described in Section~\ref{nonsignal}. Therefore the non-signalling version of the ``if" direction of this lemma does not hold, and thus any proof of it has to appeal to the particular type of strategy being used.
We now show how to construct graphs exhibiting separations between independence and quantum independence numbers. This construction is inspired by, and a generalization of, Theorem 16 of \cite{MSS}.
\begin{cor}
If $X \not\rightarrow Y$ and $X \qarrow Y$, then $\alpha(X \ltimes Y) < \alpha_q(X \ltimes Y)$.
\proof
If $X \not\rightarrow Y$, then $\alpha(X \ltimes Y) < |V(X)|$. If $X \qarrow Y$, then $\alpha_q(X \ltimes Y) = |V(X)|$.\qed
\end{cor}
By the definition of quantum independence number, the above corollary states that if $X \not\rightarrow Y$ and $X \qarrow Y$, then $K_n \not\rightarrow \overline{X\ltimes Y}$ and $K_n \qarrow \overline{X\ltimes Y}$, where $n = |V(Y)|$. Therefore the $(K_n,\overline{X \ltimes Y})$-homomorphism game exhibits quantum advantage, and by Lemma~\ref{complete} the graph $\overline{X\ltimes Y}$ is not (homomorphically equivalent to) a complete graph.
Unfortunately, the above result does not guarantee a large separation. However, we will see how to remedy this in some cases.
For graphs $X$ and $Y$, their Cartesian product, denoted $X\square Y$, is the graph with vertex set $V(X) \times V(Y)$ with two vertices being adjacent if they are equal in one coordinate and adjacent in the other. Note that $X \ltimes K_n = X \square K_n$ for all graphs $X$. A classic theorem of Vizing's states:
\[\alpha(X \square Y) \le \min\{\alpha(X)|V(Y)|, \alpha(Y)|V(X)|\}.\]
Suppose that $X$ is a graph such that $\chi_q(X) = c \ll |V(X)|/\alpha(X)$ (note that this necessarily implies that $\chi_q(X) \ll \chi(X)$). We have that $X \qarrow K_c$ and therefore $\alpha_q(X \square K_c) = |V(X)|$. However, applying the result of Vizing, we obtain $\alpha(X \square K_c) \le c\alpha(X) \ll |V(X)|$.
Applying this approach to the graphs $\Omega_n$, we get the following:
\begin{cor}\label{qindyexample}
There exists an $\epsilon > 0$, such that for $n \in \mathbb{N}$ divisible by 4
\[\frac{\alpha_q(\Omega_n \square K_n)}{\alpha(\Omega_n \square K_n)} \ge \frac{1}{n}\left(\frac{2}{2-\epsilon}\right)^n.\]
\end{cor}
\proof
We have mentioned that $\Omega_n \qarrow K_n$, so by Lemma \ref{lem:qhom2qindy}, $\alpha_q(\Omega_n \square K_n) = |V(\Omega_n)| = 2^n$. A result by Frankl and R\"{o}dl \cite{frankrod} shows that there exists an $\epsilon > 0$ such that for $n \in \mathbb{N}$ divisible by 4, $\alpha(\Omega_n) \le (2 - \epsilon)^n$. Applying the Vizing result gives that $\alpha(\Omega_n \square K_n) \le n(2-\epsilon)^n$.
\qed
The idea and proof for the above corollary is identical to Theorem 17 in \cite{MSS}, though their result is about one-shot zero-error capacities rather than independence numbers.\\
As with quantum chromatic number, Theorem~\ref{thetapreserved} lets us bound the quantum clique number of a graph.
\begin{cor}\label{omegaqupperbound}
For any graph $X$,
\[\omega_q(X) \le \bartheta(X) \ \ \& \ \ \alpha_q(X) \le \vartheta(X).\]
\end{cor}
\proof
Let $n = \omega_q(X)$. Then $K_n \qarrow X$, and by Theorem~\ref{thetapreserved} we have $n = \bartheta(K_n) \le \bartheta(X)$.\qed
Applying the above to Lemma~\ref{lem:qhom2qindy}, we further obtain:
\begin{cor}\label{thetaproduct}
If $X \qarrow Y$, then
\[\vartheta(X \ltimes Y) = |V(X)|.\]
\end{cor}
\proof
By Property~(\ref{prop:chi}) of the homomorphic product, we get
\[\vartheta(X \ltimes Y) = \bartheta(\overline{X \ltimes Y}) \le \chi(\overline{X \ltimes Y}) \le |V(X)|.\]
Suppose that $X \qarrow Y$, then by Corollary~\ref{omegaqupperbound} and Lemma~\ref{lem:qhom2qindy} we have
\[\vartheta(X \ltimes Y) \ge \alpha_q(X \ltimes Y) = |V(X)|.\]\qed
Though this corollary is weaker than Lemma~\ref{lem:qhom2qindy}, it may be of practical use since $\vartheta$ can be computed efficiently.
Another nice consequence of Theorem~\ref{thetapreserved}, and more specifically Corollary~\ref{omegaqupperbound}, is a quantum version of the well-known clique-coclique bound for vertex transitive graphs:
\begin{cor}\label{cliquecoclique}
If $X$ is a vertex transitive graph, then
\[\alpha_q(X) \omega_q(X) \le |V(X)|.\]
\end{cor}
\proof
In \cite{lovasz}, it is shown that $\vartheta(X)\bartheta(X) = |V(X)|$ for any vertex transitive graph $X$. Applying Corollary~\ref{omegaqupperbound} we see that
\[\alpha_q(X)\omega_q(X) \le \vartheta(X)\bartheta(X) = |V(X)|.\]\qed
%Those who study graph homomorphisms might note that the clique-coclique bound is a special case of the well known no homomorpism lemma which states that if $Y$ is vertex transitive and $X \rightarrow Y$, then
%\[\frac{|V(X)|}{\alpha(X)} \le \frac{|V(Y)|}{\alpha(Y)}.\]
%Given that a quantum version of the clique-coclique bound holds, one might wonder if a quantum no homomorphism lemma holds more generally. This would, in particular, imply that
%\[\chi_q(X) \ge \frac{|V(X)|}{\alpha_q(X)}\]
%for all graphs $X$. However, in considering $\Omega_n$, we recall that
%\[\chi_q(\Omega_n) = n \ \ \& \ \ \alpha_q(\Omega_n) \le \vartheta(\Omega_n) = \frac{2^n}{n} \ \ \& \ \ |V(\Omega_n)| = 2^n.\]
%This means that
%\[\frac{|V(X)|}{\alpha_q(X)} \ge \frac{2^n}{\lfloor \frac{2^n}{n} \rfloor} > n = \chi_q(\Omega_n)\]
%for $n$ a multiple of 4 but not a power of 2. So at least this quantum version of the no homomorphism lemma does not hold.
The clique-colique bound is a special case of the no-homomorphism lemma \cite{nohomo}, which states that if $X \rightarrow Y$, and $Y$ is vertex transitive, then
\[\frac{|V(X)|}{\alpha(X)} \le \frac{|V(Y)|}{\alpha(Y)}.\]
So it is natural to ask if a quantum version of the no-homomorphism lemma holds true. In particular this would imply that for all graphs $X$,
\[\frac{|V(X)|}{\alpha_q(X)} \le \chi_q(X).\]
It turns out that this does not hold. Recalling the graph $\Omega_n$, for $4 | n$ we have the following:
\[\chi_q(\Omega_n) = n \ \ \text{and} \ \ \alpha_q(\Omega_n) \le \vartheta(\Omega_n) = \frac{2^n}{n} \ \ \text{and} \ \ |V(\Omega_n)| = 2^n.\]
Since, $\alpha_q(\Omega_n)$ is integer, we have that $\alpha_q(\Omega_n) \le \lfloor \frac{2^n}{n} \rfloor$, and thus $\alpha_q(\Omega_n) < \frac{2^n}{n}$ whenever $n$ is not a power of 2. Therefore, we have that
\[\frac{|V(\Omega_n)|}{\alpha_q(\Omega_n)} > \frac{2^n}{\frac{2^n}{n}} = n = \chi_q(\Omega_n)\]
when $n$ is a multiple of 4 but not a power of 2.
%----------------------------------------------------
\subsection{Quantum Homomorphisms to Kneser Graphs}
%----------------------------------------------------
%The chromatic number of a graph $X$ can be written as the solution to an integer program, and the solution to its linear program relaxation is known as the fractional chromatic number $\chi^*(X)$. Though this is the usual definition, it is a nontrivial result that the following definition is equivalent:
%
%\begin{definition}
%The Kneser graph, $K_{n:r}$, has the $r$-subsets of $[n]$ as its vertices with two being adjacent if they are disjoint. For a graph $X$,
%\[\chi^*(X) = \min\left\{\frac{n}{r} : X \rightarrow K_{n:r}\right\}.\]
%\end{definition}
%
%Using this we can define quantum fractional chromatic number as follows:
%
%\begin{definition}
%For any graph $X$, the quantum fractional chromatic number of $X$ is denoted by $\chi^*_q(X)$ and is defined as
%\[\chi^*_q(X) = \inf\left\{\frac{n}{r} : X \qarrow K_{n:r}\right\}.\]
%
%In this case it is necessary
%\end{definition}
%The Kneser graph, $K_{n:r}$, has the $r$-subsets of $[n]$ as its vertices, with two being adjacent if they are disjoint.
%\begin{theorem}
%If $n \ge 2r$, then
%\[\alpha(K_{n:r}) = \binom{n-1}{r-1}.\]
%\end{theorem}
%
%\begin{lemma}
%If $X \rightarrow Y$ and $Y$ is vertex transitive, then
%\[\frac{|V(X)|}{\alpha(X)} \le \frac{|V(Y)|}{\alpha(Y)}.\]
%\end{lemma}
Here we prove a lemma which relates the existence of quantum homomorphism from $X$ to $Y$ to the existence of a quantum homomorphism from $X \ltimes Y$ to a particular Kneser graph. The Kneser graph, $K_{n:r}$, is the graph whose vertices are the $r$-subsets of the set $[n]$ wtih disjoint sets being adjacent.
\begin{lemma}\label{kneser2hom}
Let $X$ and $Y$ be graphs. If $n,r \in \mathbb{N}$ are such that $\frac{n}{r} = |V(Y)|$, then
\[X \ltimes Y \rightarrow K_{n:r} \Rightarrow X \rightarrow Y\]
and
\[X \ltimes Y \qarrow K_{n:r} \Rightarrow X \qarrow Y\]
\end{lemma}
\proof
Suppose that $X \ltimes Y \rightarrow K_{n:r}$. The $r$-subsets of $[n]$ containing the element 1 give an independent set of $K_{n:r}$ of size $\binom{n-1}{r-1}$, and thus $\alpha(K_{n:r}) \ge \binom{n-1}{r-1}$ for all values of $n$ and $r$. %If $n < 2r$, then $n = r$ since $n/r = |V(Y)|$, and therefore $K_{n:r} \cong K_1$. In this case, $X$ must be empty and thus $X \rightarrow Y$. If $n \ge 2r$, then by the famous Erd\H{o}s-Ko-Rado theorem \cite{EKR}, $\alpha(K_{n:r}) = \binom{n-1}{r-1}$.
Since $K_{n:r}$ is vertex transitive, we can apply the no-homomorphism lemma \cite{nohomo} to obtain that
\[\frac{|V(X\ltimes Y)|}{\alpha(X\ltimes Y)} \le \frac{|V(K_{n:r})|}{\alpha(K_{n:r})} \le \frac{n}{r} = |V(Y)|.\]
Therefore, $\alpha(X \ltimes Y) \ge |V(X)|$ and by Lemma~\ref{lem:hom2indy} $X \rightarrow Y$.
The proof of the quantum statement uses results and discussion from Section~\ref{rank}. Suppose that $X \ltimes Y \qarrow K_{n:r}$. Since $K_{n:r}$ has a projective representation of value $\frac{n}{r} = |V(Y)|$, by Theorem~\ref{thm:xipreserved} so does $X \ltimes Y$, and therefore by Corollary~\ref{hom2rep}, $X \qarrow Y$.\qed
\begin{lemma}\label{hom2kneser}
If $X$ and $Y$ are graphs and $G$ is a subgroup of $\text{Aut}(Y)$ that acts transitively on $V(G)$, then
\[X \rightarrow Y \Leftrightarrow X \ltimes Y \rightarrow K_{|G|:\frac{|G|}{|V(Y)|}}\]
and
\[X \qarrow Y \Leftrightarrow X \ltimes Y \qarrow K_{|G|:\frac{|G|}{|V(Y)|}}\]
\end{lemma}
\proof
Let $G$ be as stated and let $n = |G|$ and $r = |G|/|V(Y)|$. Note that we can view the vertices of $K_{n:r}$ as being $r$-subsets of $G$. We will prove both forward implications simultaneously. Suppose that Alice and Bob can win the $(X,Y)$-homomorphism game. For all $y,y' \in V(Y)$ let
\[G(y,y') = \{\phi \in G : \phi(y) = y'\}.\]
Note that $G(y,y')$ is a coset of the stabilizer of $y$. Since $G$ acts transitively on $V(Y)$, the orbit of any vertex is $V(Y)$, and therefore
\[|G(y,y')| = |G|/|V(Y)| = r\]
for all $y,y' \in V(Y)$. Consider the following strategy for the $(X \ltimes Y,K_{n:r})$-homomorphism game. Upon receiving $(x_A,y_A), (x_B,y_B) \in V(X\ltimes Y)$, Alice and Bob act as if they are playing the $(X,Y)$-homomorphism game with $x_A, x_B \in V(X)$ to obtain $y'_A, y'_B \in V(Y)$. They then respond with $G(y_A,y'_A)$ and $G(y_B,y'_B)$ respectively. To show that this is a valid strategy for the $(X \ltimes Y,K_{n:r})$-homomorphism game, we have two cases:
\begin{itemize}
\item If $(x_A,y_A) = (x_B,y_B)$, then $x_A = x_B$ and therefore $y'_A = y'_B$. Furthermore, $y_A = y_B$ and so $G(y_A,y'_A) = G(y_B,y'_B)$.
\item If $(x_A,y_A) \sim (x_B,y_B)$, then we have two subcases:
\begin{itemize}
\item If $x_A = x_B$ and $y_A \ne y_B$, then $y'_A = y'_B$. Since no automorphism of $Y$ can map both $y_A$ and $y_B$ to the same vertex, $G(y_A,y'_A)$ and $G(y_B,y'_B)$ are disjoint.
\item If $x_A \sim x_B$ and $y_A \not\sim y_B$, then $y'_A \sim y'_B$. Since no automorphism of $Y$ can map two (possibly equal) nonadjacent vertices to two adjacent vertices, $G(y_A,y'_A)$ and $G(y_B,y'_B)$ are disjoint.
\end{itemize}
\end{itemize}
The other direction of each statement is proven by Lemma~\ref{kneser2hom} and the fact that
\[\frac{|G|}{|G|/|V(Y)|} = |V(Y)|.\]\qed
Applying Theorem~\ref{thetapreserved} to the above, we obtain a corollary similar to Corollary~\ref{thetaproduct}:
\begin{cor}\label{thetaproducttrans}
If $Y$ is vertex transitive and $X \qarrow Y$, then
\[\bartheta(X \ltimes Y) = |V(Y)|.\]
\end{cor}
\proof
Recalling property~\ref{prop:omega}, we see that
\[\bartheta(X \ltimes Y) \ge \omega(X \ltimes Y) \ge |V(Y)|.\]
By Lemma~\ref{hom2kneser}, $X \ltimes Y \qarrow K_{n:r}$ for $n = |\text{Aut}(Y)|$, and $r = |\text{Aut}(Y)|/|V(Y)|$. Applying Theorem~\ref{thetapreserved} we have that
\[\bartheta(X \ltimes Y) \le \bartheta(K_{n:r}) = \frac{n}{r} = |V(Y)|\]
where the value of $\bartheta(K_{n:r})$ comes from \cite{lovasz}.\qed
As with Corollary~\ref{thetaproduct}, the above result is strictly weaker than Lemma~\ref{hom2kneser}, but since $\bartheta$ is efficient to compute, it may be of more practical use. Note that in the case of $Y$ being vertex transitive, the above is strictly stronger than Corollary~\ref{thetaproduct}. To see this, note that $\vartheta(Z)\bartheta(Z) \ge |V(Z)|$ for any graph $Z$ was proven in \cite{lovasz}. Therefore, if $X\qarrow Y$, then by the above corollary we have that
\[\vartheta(X\ltimes Y) \ge \frac{|V(X)||V(Y)|}{\bartheta(X\ltimes Y)} = |V(X)|.\]
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Zero-Error Capacity}\label{capacity}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%Here we show that the quantum independence number is always at most the one-shot entanglement-assisted zero-error capacity, which we denote by $c_0^*$.
%
%First, we will introduce some background about entanglement assisted capacity. The set up is as follows: Alice has access to some noisy channel $\mathcal{N}$ that she can use to send messages to Bob. The set of inputs to $\mathcal{N}$ is $V$ and the set of outputs is $E$. The channel is equipped with a probability distribution and we denote by $\mathcal{N}(y|x)$ the probability of the channel outputting $y$ given that it received $x$ as input. We can then say that two inputs $x,x' \in V$ are confusable if there exists $y \in E$ such that $\mathcal{N}(y|x) > 0$ and $\mathcal{N}(y|x') > 0$. Note that this means that $x$ is confusable with itself.
%
%The \emph{confusability graph} of $\mathcal{N}$, denoted $X(\mathcal{N})$, has $V$ as its vertex set, and two distinct vertices $x,x' \in V$ are adjacent if they are confusable (we don't have loops even though a vertex is confusable with itself).
%
%It is straightforward to see that the most messages that can be sent using $\mathcal{N}$ classically is simply equal to the size of the largest set $S$ of inputs such that no two distinct elements of $S$ are confusable. This is simply the independence number of $X(\mathcal{N})$, and so $c_0(\mathcal{N}) = \alpha(X(\mathcal{N}))$. This means that the classical capacity of a channel depends only on its confusability graph. It is a nontrivial result that the same is true for entanglement-assisted capacity.
%
%Given a channel $\mathcal{N}$, the general quantum strategy for sending $m$ messages with $\mathcal{N}$ is as follows: Alice and Bob have a shared state $\rho$ (here $\rho$ is a density matrix, not a vector), and for each message $i \in [m]$, Alice has a POVM $\mathcal{E}^i = \{E_{ix}\}_{x \in V}$. To send message $i$, Alice measures her part of $\rho$ with POVM $\mathcal{E}^i$ and gets outcome $E_{ix}$ for some $x \in V$, and then she inputs $x$ into the channel. The residual state that Bob is left with after Alice gets outcome $E_{ix}$, is
%\[\rho^i_x = \tr_A((E_{ix} \otimes I)\rho)/p^i_x\]
%where $p^i_x$ is the probability that Alice inputs $x$ given she wants to send message $i$. If we let
%\[\beta^i_x = p^i_x \rho^i_x\]
%then
%\[\sum_{x \in V} \beta^i_x = \tr_A(\rho) =: \rho_B.\]
%Bob will be able to perfectly determine what message Alice wanted to send if and only if
%\[\tr(\beta^i_x \beta^j_{x'}) = 0 \text{ for all } i \ne j, \ x \text{ and } x' \text{ confusable.}\]
%
%In \cite{entcap} they show that it suffices to only consider picking residual states which satisfy the above, thus doing away with Alice's measurements.
%
%\begin{theorem}[Cubitt, Leung, Matthews, Winter, 2010]\label{betas}
%There exists an entanglement-assisted strategy for sending $m$ messages over a channel $\mathcal{N}$ if and only if there exists a density matrix $\rho_B$ and positive semidefinite operators $\beta^i_x$ for all $i \in [m]$ and $x \in V$ such that
%\[\sum_{x \in V} \beta^i_x = \rho_B \text{ for all } i \in [m]\]
%and
%\[\tr(\beta^i_x \beta^j_{x'}) = 0 \text{ for all } i \ne j, \ x \text{ and } x' \text{ confusable.}\]
%In particular, the entanglement-assisted capacity of a channel only depends on its confusability graph.
%\end{theorem}
%
%Since the entanglement-assisted capacity of a channel only depends on the confusability graph, we often speak of the capacity of a graph $X$. If we want to construct a channel whose confusability graph is some given $X$, then we just let the vertices of $X$ be the inputs and let $\mathcal{N}(y|x) = \frac{1}{\text{deg}(x)}$ for all edges $y$ incident to $x$ for all $x \in V(X)$. This means that we can think of the edges as the outputs of the channel. Note that if there are isolated vertices, then either we can add loops, or simply allow for isolated vertices to be outputs of our channel.
In this section we investigate the relationship of quantum
homomorphisms to zero-error channel capacity. We start by explaining
the setting of zero-error communication.
Suppose Alice wants to use a noisy classical channel $\mathcal{N}$ to
communicate messages to Bob with zero probability of decoding error.
The channel $\mathcal{N}$ is completely specified by the set of inputs
$Z$, the set of outputs $W$ and a conditional probability distribution
$\mathcal{N}(w|z)$, the probability to output $w$ upon inputting
$z$. Inputs $z$ and $z'$ are said to be \emph{confusable} if there
exists $w\in W$ such that $\mathcal{N}(w|z) > 0$ and
$\mathcal{N}(w|z') > 0$. To represent a channel $\mathcal{N}$, we use
its \emph{confusability graph} $X_{\mathcal{N}}$. This is a graph with
vertex set $Z$ in which distinct vertices $z,z'$ are adjacent whenever
inputs $z$ and $z'$ are confusable. Note that $X_{\mathcal{N}}$ does
not have loops even though input $z$ is confusable with itself.
The one-shot zero-error capacity, $c_0(\mathcal{N})$, is the largest
number number of different messages that Alice can send to Bob with
zero probability of error by one use of $\mathcal{N}$. Since
$c_0(\mathcal{N})$ is the largest set of inputs among which no two are
confusable, we have that $c_0(\mathcal{N})=\alpha(X_{\mathcal{N}})$.
It turns out that one-shot zero-error capacity can increase if Alice
and Bob are allowed to use shared entanglement \cite{entcap}. The capacity in this case is referred to as the entanglement-assisted one-shot zero-error capacity, and denoted $c_0^*(\mathcal{N})$. As in the unassisted case, it can be shown that
$c_0^*(\mathcal{N})$ only depends on the confusability graph
$X_{\mathcal{N}}$ \cite{entcap}. Thus, from now on we can consider one-shot capacities of graphs and write $c_0(X)$ or $c_0^*(X)$. For
each graph $X$, we choose a canonical channel $\mathcal{N}$ that has
$X$ as its confusability graph. This is the channel with input set
$V(X)$, output set $E(X)$, and $\mathcal{N}(e|x)=\frac{1}{deg(x)}$.
For isolated vertices we can either add loops or allow for isolated
vertices to be outputs of our channel.
We now describe the general form of a quantum protocol that Alice
and Bob can use to transmit one of $m$ different messages over
$\mathcal{N}$ using a shared entangled state $\psi$. To send message
$i \in [m]$, Alice measures her part of $\psi$ using a POVM
$\mathcal{E}^i = \{E_{iz}\}_{z \in Z}$. She then uses the measurement
outcome $z$ as the input to the channel. Bob receives some output $w$
such that $\mathcal{N}(w|z)>0$. To recover the original message $i$,
he performs a measurement on his residual state $\beta^i_z :=
\tr_A((E_{iz} \otimes I)\psi\psi^*)$, where $\tr_A : \mathbb{C}^{d_A} \otimes \mathbb{C}^{d_B} \rightarrow \mathbb{C}^{d_B}$ is the map defined by $\tr_A(M\otimes N) = \tr(M)N$ and extended linearly. A successful recovery can be
ensured if and only if
\begin{equation}
\tr(\beta_z^i \beta_{z'}^j) = 0 \text{ for all distinct $i,j\in[m]$ and
all confusable $z,z'\in Z$.}
\label{eq:betas}
\end{equation}
Note that $\tr(\beta_z^i\beta_{z'}^j) = 0$ if and only if $\beta_z^i\beta_{z'}^j = 0$ since $\beta_z^i$ and $\beta_{z'}^j$ are positive semidefinite. The above is necessary since confusable inputs can result in Bob
receiving the same output $w$. It is sufficient since, given $w$, Bob
can recover $i$ with certainty by discriminating between the mutually
orthogonal operators
\[\beta^i := \sum_{z : \mathcal{N}(w|z)>0} \beta_z^i.\]
%------------------------------------------------------
\subsection{Relation to Quantum Independence Number}
%------------------------------------------------------
Recalling that $c_0(X) = \alpha(X)$ for any graph $X$, it is natural to ask if there is any relation between the entanglement-assisted one-shot capacity and the quantum independence number. The following theorem shows that these two parameters are indeed related.
\begin{theorem}\label{thm:qindyqcap}
For any graph $X$,
\[\alpha_q(X) \le c_0^*(X)\]
with equality if and only if $c_0^*(X)$ can be achieved using a strategy in which all of Alice's measurements are projective and the shared state is $\Phi = \frac{1}{\sqrt{d}}\sum_{i = 1}^d e_i \otimes e_i \in \mathbb{C}^d \otimes \mathbb{C}^d$ for some $d \in \mathbb{N}$.
\end{theorem}
\proof
We will make use of the following identity, which can be proved easily using the techniques from \cite{calculus}:
\[\tr(\tr_A((E \otimes I)\Phi \Phi^*)\tr_A((F \otimes I) \Phi \Phi^*)) = \frac{1}{d^2}\tr(EF).\]
Let $E_{ix} \in M_{\mathbb{C}}(d,d)$ be projectors that give a quantum homomorphism from $K_m$ to $\overline{X}$, where $m = \alpha_q(X)$. We will show that the $E_{ix}$ along with the shared state $\Phi$ give a valid strategy for sending $m$ messages with $X$. Bob's residual states for this strategy are
\[\beta^i_x = \tr_A((E_{ix} \otimes I)\Phi \Phi^*).\]
It remains to show that the above operators satisfy Condition~(\ref{eq:betas}). If $i \ne j$ and ($x \sim x'$ or $x = x'$), then by the winning conditions of the $(K_m,\overline{X})$-homomorphism game, we have that $\tr(E_{ix}E_{jx'}) = 0$, and therefore
\[\tr(\beta^i_x \beta^j_{x'}) = 0.\]
That is, Condition~(\ref{eq:betas}) is satisfied. Therefore $\alpha_q(X) \le c_0^*(X)$.
%For a fixed $i \in [m]$, we have that
%\[\sum_{x \in V(X)} \beta ^i_x = \sum_{x \in V(X)} \tr_A((E_{ix} \otimes I)\Phi \Phi^*) = \tr_A(\Phi \Phi^*) = \frac{1}{d}I_d.\]
%Since $\frac{1}{d}I_d$ is a density matrix, this shows that Alice and Bob can send $m$ messages with $X$ and thus $c_0^*(X) \ge \alpha_q(X)$.
Now suppose that Alice can use projective measurements $E_{ix}$ and the shared state, $\Phi$, to send $m$ messages with $X$. From the discussion above, we have that
\begin{align*}
0 = \tr(\tr_A((E_{ix} \otimes I)\Phi \Phi^*)\tr_A((E_{jx'} \otimes I) \Phi \Phi^*)) = \frac{1}{d^2}\tr(E_{ix}E_{jx'})
%&= \tr(\tr_A(\frac{1}{d}\vect(E_{ix})\vect(I)) \tr_A(\frac{1}{d}\vect(E_{jx'})\vect(I)))
\end{align*}
for all $i \ne j$, and confusable $x,x'$. Note that $x$ is confusable with $x'$ if and only if $x \sim x'$ or $x = x'$. Since the $E_{ix}$'s are projectors, we also have that $\tr(E_{ix}E_{ix'}) = 0$ for all $i \in [m]$ and $x \ne x'$. These are all the orthogonalities required to win the $(K_m,\overline{X})$-homomorphism game.
\qed
If in fact $c_0^*(X)$ can always be achieved using projective measurements on the maximally entangled state, then $\alpha_q(X) = c_0^*(X)$. In light of this we conjecture the following:
%If it could be shown that projective measurements and the state $\frac{1}{\sqrt{d}}\sum_{i = 1}^d e_i \otimes e_i \in \mathbb{C}^d \otimes \mathbb{C}^d$ always suffice to achieve $c_0^*(X)$, then we would have that $\alpha_q(X) = c_0^*(X)$ for all graphs $X$. In light of this we give the following:
\begin{conjecture}\label{conj:qindyqcap}
For any graph $X$,
\[\alpha_q(X) = c_0^*(X).\]
\end{conjecture}
%It is significant that we used the actual quantum strategies in the proof above, instead of just considering the games themselves as in some other proofs. This is because the above theorem is not true for arbitrary correlations. In fact the above theorem does not really make sense for arbitrary correlations because, unlike in the classical and quantum cases, the capacity of a channel is not completely determined by its confusability graph.
As a corollary to the above theorem, we obtain a method for constructing a graph $Z$ with $c_0(Z) < c_0^*(Z)$ given graphs $X$ and $Y$ such that $X \qarrow Y$ but $X \not\rightarrow Y$. This was already proven for $Y = K_n$ in \cite{MSS}.
\begin{cor}
If $X \qarrow Y$ but $X \not\rightarrow Y$, then
\[c_0(X \ltimes Y) < c_0^*(X \ltimes Y) = |V(X)|.\]
\end{cor}
\proof
First, note that
\[c_0^*(X\ltimes Y) \le \vartheta(X \ltimes Y) \le \chi(\overline{X\ltimes Y}) \le |V(X)|.\]
Now since $X \not\rightarrow Y$ and $X \qarrow Y$, we have that
\[c_0(X \ltimes Y) = \alpha(X \ltimes Y) < |V(X)| = \alpha_q(X\ltimes Y) \le c_0^*(X \ltimes Y) \le |V(X)|.\]\qed
%----------------------------------------------------
\subsection{Relation to Quantum Homomorphisms}
%----------------------------------------------------
Here we show that quantum homomorphisms respect the order of entanglement-assisted one-shot zero-error capacity.
\begin{lemma}
If $X \qarrow Y$, then $c_0^*(\overline{X}) \le c_0^*(\overline{Y})$.
\end{lemma}
\proof
Suppose that $\overline{X} \qarrow \overline{Y}$ and that $P_{xy} \in\mathbb{M}_\mathbb{C}(d,d)$, for $x \in V(X), y \in V(Y)$ are projectors which give such a quantum homomorphism. Furthermore, let $E_{ix}$, $i \in [m]$, $x \in V(X)$, be POVM elements and $\psi$ the shared state giving a strategy for sending $m$ messages with $X$. Then Bob's residual states for this strategy
\[\beta_x^i = \tr_A((E_{ix} \otimes I)\psi\psi^*)\]
satisfy the orthogonality condition~(\ref{eq:betas}).
Let $\Phi = \frac{1}{\sqrt{d}}\sum_{i = 1}^d e_i \otimes e_i$, and let
\[F_{iy} = \sum_{x \in V(X)} P_{xy} \otimes E_{ix}\]
for all $i \in [m]$, $y \in V(Y)$. We claim that the POVMs $\{F_{iy}\}_{y \in V(Y)}$ and the state $\Phi \otimes \psi$ give a valid strategy for sending $m$ messages with $Y$. First, note that
\[\sum_{y \in V(Y)} F_{iy} = \sum_{x \in V(X), y \in V(Y)} P_{xy} \otimes E_{ix} = \sum_{x \in V(X)} I \otimes E_{ix} = I\]
Second, each $F_{iy}$ is positive semidefinite since it is the sum of tensor products of positive semidefinite matrices, therefore the $\{F_{iy}\}_{y \in V(Y)}$ are indeed POVMs. If we use $A'$ and $A$ to denote the spaces in Alice's posession, then the the residual states are given by
\begin{align*}
\alpha_y^i &= \tr_{A'A}\left(\sum_{x \in V(X)}\left((P_{xy} \otimes I)\Phi\Phi^*\right) \otimes \left((E_{ix} \otimes I)\psi\psi^*\right)\right) \\
&= \frac{1}{d}\sum_{x \in V(X)} P_{xy} \otimes \tr_A((E_{ix} \otimes I)\psi\psi^*) = \frac{1}{d}\sum_{x \in V(X)} P_{xy} \otimes \beta_x^i.
\end{align*}
Suppose that $i \ne j$ and $y,y'$ are confusable, i.e.~$y$ is not adjacent to $y'$ in $\overline{Y}$. Then
\[
\alpha_y^i \alpha_{y'}^j = \frac{1}{d^2}\sum_{x,x' \in V(X)} P_{xy}P_{x'y'} \otimes \beta_x^i \beta_{x'}^j = \frac{1}{d^2}\sum_{x \not\simeq x' \hspace{.015in} \text{in} \hspace{.015in} X} P_{xy}P_{x'y'} \otimes \beta_x^i \beta_{x'}^j = 0
\]
where the last equality holds because of the orthogonality constraints on the $P_{xy}$. This shows that Alice and Bob can send $m$ messages using $Y$.\qed
% As $\tr_A((E_{ix} \otimes I)\psi\psi^*)$ are the residual states from the strategy for sending messages with $X$, we have that
%\[\tr_A((E_{ix} \otimes I)\psi\psi^*)\tr_A((E_{jx'} \otimes I)\psi\psi^*) = 0\]
%whenever $i \ne j$ and ($x = x'$ or $x \sim x'$). We also have that $P_{xy}P_{x'y'} = 0$ whenever ($x = x'$ and $y \ne y'$) or ($x \not\simeq x'$ and ($y = y'$ or $y \sim y'$)) from the conditions of the $(\overline{X},\overline{Y})$-homomorphism game. Therefore, if $i \ne j$ and $y$ and $y'$ are confusable,
%\[\beta_y^i \beta_{y'}^j = \sum_{x,x' \in V(X)} P_{xy}P_{x'y'} \otimes \]
%
% Furthermore, let $\beta_x^i$, $i \in [m]$, $x \in V(X)$ and $\rho_B$ satisfy the conditions of Theorem \ref{betas} for $X$. Define
%\[\alpha_y^i = \frac{1}{d} \sum_{x \in V(X)} P_{xy} \otimes \beta_x^i\]
%for all $i \in [m]$, $y \in V(Y)$. We claim that the $\alpha_y^i$ satisfy the conditions of Theorem \ref{betas} for $Y$.
%
%Suppose that $i \ne j$ and $y,y'$ are confusable, i.e.~$y = y'$ or $y \sim y'$. Then
%\[
%\alpha_y^i \alpha_{y'}^j = \frac{1}{d^2}\sum_{x,x' \in V(X)} P_{xy}P_{x'y'} \otimes \beta_x^i \beta_{x'}^j = \frac{1}{d^2}\sum_{x \not\simeq x' \in V(X)} P_{xy}P_{x'y'} \otimes \beta_x^i \beta_{x'}^j = 0
%\]
%where the last equality holds because of the orthogonality constraints on the $P_{xy}$. Checking the appropriate sums we see that
%\begin{align*}
%\sum_{y\in V(Y)} \alpha_y^i &= \frac{1}{d}\sum_{x \in V(X), y \in V(Y)} P_{xy} \otimes \beta_x^i \\
%&= \frac{1}{d}\sum_{x\in V(X)}\left( \sum_{y\in V(Y)}P_{xy} \otimes \beta_x^i \right) \\
%&= \frac{1}{d}\sum_{x\in V(X)} I \otimes \beta_x^i \\
%&= \frac{1}{d}(I \otimes \rho_B).
%\end{align*}
%Since $\frac{1}{d}(I \otimes \rho_B)$ is a density matrix, we have shown that $c_0^*(Y) \ge c_0^*(X)$.\qed
The entanglement-assisted Shannon capacity, $\Theta^*$, is the asymptotic version of $c_0^*$, and is defined as
\[\Theta^*(X) = \lim_{n \rightarrow \infty} \sqrt[n]{c_0^*(X^{\boxtimes n})}.\]
Here, $X \boxtimes Y$ is the graph with vertex set $V(X) \times V(Y)$ with two distinct vertices being adjacent if each coordinate is either equal or adjacent, and $X^{\boxtimes n}$ is the product of $X$ with itself $n$ times.
The value of $\Theta^*$ corresponds to the per use capacity of a channel with entanglement-assistance in the limit of many uses. We now prove an asymptotic version of the above lemma. To avoid messy notation, we define the following:
\[X^{n} := \overline{\overline{X}^{\boxtimes n}}.\]
\begin{lemma}
If $X \qarrow Y$, then $\Theta^*(\overline{X}) \le \Theta^*(\overline{Y})$.
\end{lemma}
\proof
Suppose that $X \qarrow Y$. It suffices to show that $c_0^*\left(\overline{X^n}\right) \le c_0^*\left(\overline{Y^n}\right)$ for all $n \in \mathbb{N}$. By the previous lemma this holds if $X^n \qarrow Y^n$ for all $n \in \mathbb{N}$. It is straightforward to check that two vertices $(x_1,\ldots,x_n)$ and $(x'_1,\ldots,x'_n)$ are adjacent in $X^n$ if and only if $x_i \sim x'_i$ for some $i \in [n]$.
To win the $\left(X^n, Y^n\right)$-homomorphism game Alice and Bob may play as follows: When they receive $(x_1,\ldots,x_n)$ and $(x'_1,\ldots,x'_n)$ respectively, they act as if they are playing the $(X,Y)$-homomorphism game with each coordinate to obtain $(y_1,\ldots,y_n)$ and $(y'_1,\ldots,y'_n)$, which they respond with accordingly. If the vertices they received were equal, then each coordinate was equal and therefore they responded with equal vertices. If the vertices they received were adjacent, then there exists $i \in [n]$ such that $x_i \sim x'_i$ and therefore $y_i \sim y'_i$ and thus they responded with adjacent vertices.
Therefore, $X^n \qarrow Y^n$ and so $c_0^*\left(\overline{X^n}\right) \le c_0^*\left(\overline{Y^n}\right)$ for all $n \in \mathbb{N}$.\qed
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Projective Rank}\label{rank}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
As we have seen, the existence of a quantum homomorphism from a graph $X$ to a graph $Y$ is related to the existence of a set of projectors satisfying certain orthogonality conditions. Furthermore, when $Y$ is vertex transitive, it can be assumed that the projectors all have the same rank. Noting that the orthogonality conditions of Corollary~\ref{reformvtxtrans} exactly correspond to the adjacencies of $X \ltimes Y$, we see that a set of projectors satisfying the conditions in Corollary~\ref{reformvtxtrans} is exactly an assignment of rank $r$ projectors from $\mathbb{M}_\mathbb{C}(d,d)$ to the vertices of $X \ltimes Y$ such that adjacent vertices receive orthogonal projectors and $d = r|V(Y)|$. This motivates the following definition:
\begin{definition}
A \emph{$d/r$-projective representation} (or simply a $d/r$-representation) of a graph $X$ is an assignment of rank $r$ projectors in $\mathbb{M}_\mathbb{C}(d,d)$ to the vertices of $X$ such that adjacent vertices are assigned projectors which are orthogonal to each other. We say that the \emph{value} of a $d/r$-representation is the rational number $\frac{d}{r}$, i.e. a $3/1$-representation is not a $6/2$-representation, but they have the same value, namely 3.
\end{definition}
If one prefers a more geometric definition, they may think of a $d/r$-representation as assigning $r$-dimensional subspaces of $\mathbb{C}^d$ such that adjacent vertices receive orthogonal subspaces. Since two projectors are orthogonal if and only if their images are orthogonal, this is clearly equivalent to the above definition.
Using the same trick as in the proof of Lemma~\ref{real}, we see that if a graph $X$ has a $d/r$-representation, then it has a $2d/2r$-representation using real projectors. Therefore, if $X$ has a projective representation of value $\alpha$, it has some projective representation of value $\alpha$ that uses only real projectors.%To see this, let $R : \mathbb{C}^{d\times d} \rightarrow \mathbb{R}^{2d\times 2d}$ be defined as in Lemma \ref{real}, and note that if $P \in \mathbb{C}^{d\times d}$ is a projector, then $R(P)$ is a projector, $\tr(P) = \tr(\Re(P))$ since $P$ is Hermitian and so
%\[\rk(R(P)) = \tr(R(P)) = 2\tr(\Re(P)) = 2\tr(P) = 2\rk(P).\]
The following corollary is the original motivation for defining projective representations, and simply amounts to a rephrasing of Corollary~\ref{reformvtxtrans}.
\begin{cor}\label{hom2rep}
If $Y$ is vertex transitive, then $X \qarrow Y$ if and only if $X \ltimes Y$ has a projective representation of value $|V(Y)|$.
\end{cor}
\proof
The orthogonality conditions in Corollary~\ref{reformvtxtrans} exactly correspond to adjacencies in $X \ltimes Y$, and a projective representation of $X \ltimes Y$ having value $|V(Y)|$ corresponds to the condition $d = r|V(Y)|$.\qed
Since $X \ltimes Y$ contains a clique of size $|V(Y)|$, it has a projective representation of value $|V(Y)|$ if and only if this is the minimum value of any projective representation of $X \ltimes Y$. This motivates the following definition.
\begin{definition}
The \emph{projective rank} of a graph $X$, denoted $\xi_f(X)$, is given by
\[\xi_f(X) = \inf\left\{\frac{d}{r} : X \text{ has a } d/r\text{-representation}\right\}.\]
\end{definition}
If we let $\Omega(r,d)$ be the graph whose vertices are the rank $r$ projectors in $\mathbb{M}_\mathbb{C}(d,d)$, and in which adjacency is orthogonality, then it is clear that a $d/r$-representation is simply a homomorphism to $\Omega(r,d)$. Therefore,
\[\xi_f(X) = \inf\left\{\frac{d}{r} : X \rightarrow \Omega(r,d)\right\}.\]
From this it is trivial to see that $\xi_f$ is preserved by homomorphisms.
Note that it is not clear whether or not the infimum in the definition of $\xi_f$ is always attained for every graph, or even if it is always rational. Recall that if a graph has a $d/r$-representation, then it has a $2d/2r$-representation using real projectors. Therefore restricting to real projectors does not change the value of $\xi_f(X)$.
%-------------------------------------------------------------
\subsection{Relation to Other Parameters}\label{rankvsother}
%-------------------------------------------------------------
An orthogonal representation of a graph is an assignment of nonzero vectors to its vertices such that adjacent vertices receive orthogonal vectors. The minimum dimension in which there exists an orthogonal representation of a graph $X$ is known as its \emph{orthogonal rank}, and is denoted by $\xi(X)$. This parameter has been investigated in various papers including \cite{qchrom}. Clearly, a projective representation using rank 1 projectors is equivalent to an orthogonal representation, and one can think of projective rank as a fractional version of orthogonal rank. Due to this, we have that
%projective representation using rank 1 projectors is equivalent to an orthogonal representation (an assignment of nonzero vectors such that adjacent vertices receive orthogonal vectors. Orthogonal representations
%If we restrict to representations using rank 1 projectors only, then we obtain the usual orthogonal rank, $\xi$, which has been investigated in various papers including \cite{qchrom}. One can think of projective rank as a fractional version of orthogonal rank (which is where the notation comes from), and so it is clear that
\[\xi_f(X) \le \xi(X)\]
for all graphs $X$.
One can also think of projective rank as a subspace version of fractional chromatic number: a homomorphism from $X$ to $K_{d:r}$ can be transformed into a homomorphism to $\Omega(r,d)$ by simply mapping elements of $[d]$ to standard basis vectors and an $r$-subset of $[d]$ to the projector whose image is the subspace spanned by the vectors corresponding to its elements. Since the fractional chromatic number of a graph $X$, denoted $\chi_f(X)$, is the minimum value of $d/r$ such that $X \rightarrow K_{d:r}$ \cite{AGT}, we see that
\[\xi_f(X) \le \chi_f(X)\]
for all graphs $X$.
Slightly less trivially, we have the following:
\begin{lemma}\label{xifchiq}
For any graph $X$,
\[\xi_f(X) \le \chi_q(X).\]
\end{lemma}
\proof
Suppose that $X$ has a quantum $c$-coloring, i.e.~a quantum homomorphism to $K_c$. This means that there exists projectors $E_{xi}$ for all $x \in V(X)$ and $i \in [c]$ of rank $r$ in dimension $cr$ such that $E_{xi}E_{x'j} = 0$ whenever ($x = x'$ \& $i \ne j$) or ($x \sim x'$ \& $i = j$). In particular, $E_{x1}E_{x'1} = 0$ whenever $x \sim x'$ and thus the projectors $E_{x1}$ give a projective representation of $X$ of value $\frac{cr}{r} = c$. Therefore $\xi_f(X) \le c$.
\qed
Note that the above lemma is also implied by the fact that quantum homomorphisms respect projective rank, Theorem \ref{thm:xipreserved} below, since the projective rank of $K_c$ can easily be seen to be $c$.
%
%In the proof of the above lemma, the projectors which give the quantum homomorphism from $X$ to $K_c$ give a $cr/r$-representation of $X \square K_c$ (not just $X$). More generally, we have the following:
%
%\begin{lemma}\label{qhom2xi}
%If $X \qarrow Y$, and $Y$ is vertex transitive, then the projectors which give the quantum homomorphism give a $r|V(Y)|/r$-representation of $X \ltimes Y$ for some $r \in \mathbb{N}$, and thus $\xi_f(X\ltimes Y) = |V(Y)|$.
%\end{lemma}
%\proof
%The first part follows trivially from Corollary \ref{hom2rep}, and since $X \ltimes Y$ has a clique of size $|V(Y)|$ we have that $\xi_f(X) = |V(Y)|$.\qed
In \cite{lovasz}, it was shown that $\bartheta(X) \le \xi(X)$ and $\bartheta(X) \le \chi_f(X)$ for any graph $X$. Here we show that the same is true for projective rank. The proof uses the following definition of $\vartheta$:
\begin{defn}\label{LThandle}
For any graph $X$ with vertices $x_1, \ldots, x_n$, we have that
\[\vartheta(X) = \min_{\phi,c}\max_{i \in [n]} \frac{1}{(c^T \phi(x_i))^2}\]
where $\phi$ is an assignment of real unit vectors to the vertices of $X$ such that nonadjacent vertices are assigned orthogonal vectors and $c$ is any unit vector.
\end{defn}
We are now prepared to prove the above mentioned result.
\begin{theorem}\label{thetaxif}
For any graph $X$,
\[\bartheta(X) \le \xi_f(X).\]
\end{theorem}
\proof
We will show that $\vartheta(X) \le \xi_f(\overline{X})$ which is equivalent. The proof is similar to the proof of $\vartheta(X) \le \xi(\overline{X})$ given in \cite{lovasz}. Suppose that $E_x$ for $x \in V(X)$ gives a $d/r$-representation of $\overline{X}$ using only real projectors. Let
\[v_x = \frac{1}{\sqrt{r}}\vect(E_x)\]
for all $x \in V(X)$, and let
\[c = \frac{1}{\sqrt{d}}\vect(I_d)\]
Note that
\begin{align*}
v_x^T v_x &= \frac{1}{r} \vect(E_x)^T \vect(E_x) = \frac{1}{r} \tr \left(E_x E_x\right) \\
&= \frac{1}{r} \tr (E_x) = \frac{1}{r} r = 1
\end{align*}
and that
\[c^T c = \frac{1}{d} \vect(I)^T \vect(I) = \frac{1}{d} d = 1.\]
So the $v_x$ and $c$ are unit vectors. Also, if $x \not\simeq x'$, then $\tr(E_x E_{x'}) = 0$ and therefore $v_x^T v_{x'} = 0$. Therefore the $v_x$ give an orthonormal representation of $\overline{X}$ as required in Definition \ref{LThandle}. Furthermore,
\begin{align*}
c^T v_x &= \frac{1}{\sqrt{rd}} \vect(I)^T \vect(E_x) = \frac{1}{\sqrt{rd}} \tr(IE_x) \\
&= \frac{1}{\sqrt{rd}} \tr(E_x) = \frac{1}{\sqrt{rd}} r = \sqrt{\frac{r}{d}}
\end{align*}
for all $x \in V(X)$. This means that
\[\vartheta(X) \le \frac{1}{(\sqrt{r/d})^2} = \frac{d}{r}.\]
This shows that $\vartheta(X) \le \xi_f(\overline{X})$, and taking complements yields the theorem.\qed\\
For a vertex transitive graph, one can show that its projective rank and quantum independence number are related.
\begin{lemma}
If $X$ is vertex transitive, then $\xi_f(X) \le \frac{|V(X)|}{\alpha_q(X)}$.
\end{lemma}
\proof
Let $X$ be a vertex transitive graph and let $n = |V(X)|$, $m = \alpha_q(X)$, and let $E_{ix} \in \mathbb{M}_\mathbb{C}(d,d)$, for $i \in [m]$, $x \in V(X)$ be projectors which give a qunatum homomorphism from $K_m$ to $\overline{X}$. Since $X$ is vertex transitive, we can assume that all of the $E_{ix}$ have the same rank $r$, and thus $d = rn$. For each $x \in V(X)$, let $F_x = \sum_{i \in [m]} E_{ix}$. We claim that the $F_x$ give a projective representation of $X$ with value $n/m$. Since $E_{ix}E_{jx} = 0$ when $i \ne j$, the $F_x$ are the sum of orthogonal projectors and are thus projectors. Furthermore, they have rank $mr$ since each $E_{ix}$ has rank $r$. If $x \sim x'$, then
\[F_x F_{x'} = \sum_{i,j \in [m]} E_{ix}E_{jx'} = 0\]
since $E_{ix}E_{jx'} = 0$ if ($i = j$ \& $x \ne x'$) OR ($i \ne j$ \& ($x \sim x'$ or $x = x'$)). Therefore, the $F_x$ give a projective representation of $X$ of value $d/mr = n/m = |V(X)|/\alpha_q(X)$.\qed
%
%As a corollary of this result, we can compute the quantum independence number of the Kneser graphs.
%
%\begin{cor}
%For $n \ge 2r$, $\alpha_q(K_{n:r}) = \binom{n-1}{r-1}$.
%\end{cor}
%
Note that for vertex transitive graphs, this lemma gives an upper bound on $\alpha_q(X)$ which is strictly stronger than the bound in Corollary~\ref{omegaqupperbound}. We have that
\[\alpha_q(X) \le \frac{|V(X)|}{\xi_f(X)} \le \frac{|V(X)|}{\bartheta(X)} = \vartheta(X).\]
%-----------------------------------------------------
\subsection{Relation to Quantum Homomorphisms}
%-----------------------------------------------------
Our main result on projective rank is that, like $\bartheta$, it is respected by quantum homomorphisms.
\begin{theorem}\label{thm:xipreserved}
If $X \qarrow Y$, and $Y$ has a projective representation of value $\gamma \in \mathbb{Q}$, then $X$ has a projective representation of value $\gamma$, and therefore $\xi_f(X) \le \xi_f(Y)$.
\end{theorem}
\proof
Suppose that $X \qarrow Y$, and that $E_{xy} \in \mathbb{M}_\mathbb{C}(d',d')$ for $x \in V(X)$, $y \in V(Y)$ are the projectors which give the quantum homomorphism. Furthermore, let $F_y$ for $y \in V(Y)$ give a $d/r$-representation of $Y$. Define $F_x$ for $x \in V(X)$ as follows:
\[F_x = \sum_{y \in V(Y)} E_{xy} \otimes F_y.\]
We will show that this is a projective representation of value $\frac{d}{r}$ for $X$. First we must check that the $F_x$ are projectors. They are clearly Hermitian since the $E_{xy}$ and $F_y$ are, and so we only need to check that they are idempotents. We have that
\begin{align*}
F_x F_x &= \left(\sum_{y \in V(Y)} E_{xy} \otimes F_y \right) \left(\sum_{y' \in V(Y)} E_{xy'} \otimes F_{y'}\right) \\
&= \sum_{y,y' \in V(Y)} E_{xy}E_{xy'} \otimes F_y F_{y'} = \sum_{y \in V(Y)} E_{xy} \otimes F_y = F_x
\end{align*}
where the nontrivial equality follows from the fact that $E_{xy}E_{xy'} = 0$ for $y \ne y'$ and that the $E_{xy}$ and $F_y$ are idempotents. Next we must check that $F_x F_{x'} = 0$ for $x \sim x'$. We have that
\begin{align*}
F_x F_{x'} &= \left(\sum_{y \in V(Y)} E_{xy} \otimes F_y \right) \left(\sum_{y' \in V(Y)} E_{x'y'} \otimes F_{y'}\right) = \sum_{y,y' \in V(Y)} E_{xy}E_{x'y'} \otimes F_y F_{y'}.
\end{align*}
However, for $x \sim x'$, whenever $y \not\sim y'$ we have that $E_{xy} E_{x'y'} = 0$, and whenever $y \sim y'$ we have that $F_y F_{y'} = 0$. Therefore, all of the terms in the sum above are 0 for $x \sim x'$ and thus $F_x F_{x'} = 0$ in this case.
Now all that is left to do is check that this representation has value $\frac{d}{r}$. The dimension of this representation is simply $dd'$ since the $E_{xy}$ are in dimension $d'$ and the $F_y$ are in dimension $d$. To compute the rank of $F_x$, we use the fact that $\rk(A + B) = \rk(A) + \rk(B)$ if $A$ and $B$ are projectors such that $AB = 0$, and the fact that $\sum_{y \in V(Y)} E_{xy} = I$. With these facts in mind, we have that
\begin{align*}
\rk(F_x) &= \rk\left(\sum_{y \in V(Y)} E_{xy} \otimes F_y \right) = \sum_{y \in V(Y)} \rk\left(E_{xy} \otimes F_y \right) \\
&= \sum_{y \in V(Y)} \rk\left(E_{xy}\right) \rk\left(F_y \right) = r \sum_{y \in V(Y)} \rk\left(E_{xy}\right) = rd'.
\end{align*}
Therefore, the representation has value $\frac{dd'}{rd'} = \frac{d}{r}$ and thus $\xi_f(X) \le \frac{d}{r}$.
\qed
The contrapositive of the above theorem gives us a way to show that there is no quantum homomorphism from some graph to another, however the projective rank may be very difficult to compute (it may not be computable at all) and so this may not be very useful in practice.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Concluding Remarks}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
We have seen that quantum homomorphisms offer a natural generalization of quantum colorings and that they respect the order of some known graph parameters such as Lov\'{a}sz theta and entanglement-assisted zero-error capacity (after taking complements). We have also introduced a new parameter, projective rank, whose order is respected by quantum homomorphisms. Furthermore, we have used the notion of quantum homomorphisms to introduce quantum versions of graph parameters other than quantum chromatic number. Also, we have established various relations between these and the above mentioned parameters. As a summary, in Figure~\ref{Hasse} we give a Hasse diagram comparing the investigated graph parameters. For two graph parameters $f$ and $g$, we say that $f \le g$ if $f(X) \le g(X)$ for all graphs $X$. We also use the notation $\bar{f}(X) := f(\overline{X})$. Below we explain the relations depicted in Figure~\ref{Hasse}.
\begin{figure}[h]
\begin{center}
\input{Diagram.tex}
\caption{Partial Order of Graph Parameters. A circle enclosing two parameters indicates that it is not known whether they are comparable. A dashed line connecting two parameters $f$ and $g$ indicates that is not known whether $f \ne g$.}\label{Hasse}
\end{center}
\end{figure}
We have that $\chi \ge \chi_f,\chi_q,\xi$, since an $n$-coloring of a graph can be used to obtain fractional and quantum $n$-colorings and an orthogonal representation in dimension $n$. Furthermore, we have seen examples for which $\chi$ is strictly greater than these three parameters, such as Kneser graphs for $\chi_f$ and $\Omega_n$ for $\chi_q$ and $\xi$. Therefore $\chi > \chi_f,\chi_q,\xi$.
Since $\chi_f(C_5) = 5/2 < 3 = \chi_q(C_5) = \xi(C_5)$ and $\chi_f(\Omega_n) = |V(\Omega_n)|/\alpha(\Omega_n) > n = \chi_q(\Omega_n) = \xi(\Omega_n)$ for $n$ a sufficiently large multiple of four, we have that $\chi_f$ and $\chi_q$ are incomparable, as well as $\chi_f$ and $\xi$. The fact that $\xi_f \le \xi,\chi_f$ was discussed in Section~\ref{rankvsother}, and $\xi_f \le \chi_q$ is Lemma~\ref{xifchiq} from that section. The incomparabilities among $\chi_f, \chi_q$, and $\xi$ imply that $\xi_f$ cannot be equal to any of these, therefore $\xi_f < \chi_f, \chi_q, \xi$.
Theorem~\ref{thetaxif} states that $\bartheta \le \xi_f$, and equality does not hold since $\bartheta(C_5) = \sqrt{5}$, and it can be shown that $\xi_f(C_5) = 5/2$. In \cite{Winter} and \cite{Beigi} it is shown that $\bar{\Theta}^* \le \bartheta$, however it is an open question as to whether these two parameters are equal.
Since $\bar{\Theta}^*$ is the entanglement-assisted version of $\bar{\Theta}$, and the asymptotic version of $\bar{c}_0^*$, clearly $\bar{\Theta}^* \ge \bar{\Theta},\bar{c}_0^*$. In \cite{LMMOR} and \cite{Jop}, examples of graphs for which $\bar{c}_0^*$, and thus $\bar{\Theta}^*$, is strictly larger than $\bar{\Theta}$ are given, thus $\bar{\Theta} < \bar{\Theta}^*$. Since $\bar{c}_0^*(C_5) \le \bartheta(C_5) = \sqrt{5}$ and $\bar{c}_0^*$ is an integer, we have that $\bar{c}_0^*(C_5) = 2 < \sqrt{5} = \bar{\Theta}^*(C_5)$, and thus $\bar{c}_0^* < \bar{\Theta}^*$.
Our Theorem~\ref{thm:qindyqcap} is equivalent to $\omega_q \le \bar{c}_0^*$. However, as mentioned in Conjecture~\ref{conj:qindyqcap}, we believe that $\omega_q$ and $\bar{c}_0^*$ are in fact equal. Since an $n$-clique is also a quantum $n$-clique, we have that $\omega \le \omega_q$, and our Corollary~\ref{qindyexample} provides examples of graphs with quantum clique number strictly greater than clique number (after taking complements).
Since $\bar{\Theta}$ is the asymptotic version of $\omega$, we have that $\omega \le \bar{\Theta}$. Also, $\omega(C_5) = 2 < \sqrt{5} = \bar{\Theta}(C_5)$, so these parameters are not equal.
To see that $\bar{\Theta}$ is not comparable with either $\bar{c}_0^*$ or $\omega_q$, note that $2 = \omega(C_5) \le \omega_q(C_5) \le \bar{c}_0^*(C_5) = 2 < \sqrt{5} = \bar{\Theta}(C_5)$. Combining this with the examples from \cite{LMMOR} and \cite{Jop} mentioned above, we see that $\bar{c}_0^*$ and $\bar{\Theta}$ are incomparable. Furthermore, to obtain the lower bound on $\bar{c}_0^*$ for those examples, the authors use projective measurements and a maximally entangled state. Therefore by our Theorem~\ref{thm:qindyqcap} we have that $\omega_q$ is not comparable to $\bar{\Theta}$.\\
%------------------------------
\subsection{Open Questions}
%------------------------------
Perhaps the most central question is whether it is decidable if $X \qarrow Y$ for arbitrary $X$ and $Y$. If we fix the dimension of the projectors Alice and Bob are allowed to use in their strategy for the $(X,Y)$-homomorphism game, then this question reduces to determining if a particular set of quadratic equations in a finite number of variables has a solution and is therefore decidable. So bounding the dimension needed for a quantum homomorphism would resolve this problem. This question remains open even for the specific case of quantum colorings.
We have seen separations between classical and quantum versions of chromatic number and independence/clique number, but what about other parameters? In particular, can we find a graph $X$ that has odd girth strictly greater than quantum odd girth? See the introduction to Section~\ref{parameters}.
%We briefly mentioned that quantum homomorphisms can only map longer odd cycles to shorter ones (this can be proved using the known values of $\bartheta$ for these graphs or by showing that $\xi_f(C_{2k+1}) = \frac{2k+1}{k}$ for all $k \in \mathbb{N}$). However, a similar statement holds for complete graphs, so this does not mean that no graph exhibits a separation between odd girth and quantum odd girth.
%We may also want to consider quantum fractional chromatic number, which can be defined using quantum homomorphisms to Kneser graphs. We actually have already seen that the graphs $\Omega_n$ from Lemma~\ref{thetahadamard} have quantum fractional chromatic number equal to $n$, but fractional chromatic number exponential in $n$. If we can find new examples of $X$ and $Y$ such that $X \not\rightarrow Y$ and $X \qarrow Y$ and $Y$ is vertex transitive, then applying Lemmas~\ref{kneser2hom} and \ref{hom2kneser} may produce new examples of graphs with separations between fractional chromatic and quantum fractional chromatic number. However, it is possible that the lemmas produce examples which are not fundamentally different than the graphs one started with.
In light of Theorem~\ref{thm:qindyqcap}, it is natural to ask if $\alpha_q(X) = c_0^*(X)$ for all graphs $X$. We believe this to be the case. A positive answer would imply that the entanglement-assisted zero-error capacity of a channel can always be attained using projective measurements on a maximally entangled state, thus resolving an open problem from in quantum information theory.
The definition of projective rank uses an infimum, and it is of interest whether or not that infimum can be replaced by a minimum, i.e.~that the value of $\xi_f(X)$ is always obtained by some projective representation of $X$. Even if this is not the case, one may ask if $\xi_f(X)$ is always rational. One possible approach to the former question is to try to come up with a different proof that the fractional chromatic number of a graph is always attained by some homomorphism into a Kneser graph. The current proof relies on the fact that fractional chromatic number can be written as a linear program and therefore can always be attained.
So far the only examples of graphs $X$ and $Y$ such that $X \not\rightarrow Y$ and $X \qarrow Y$ are with one of $X$ and $Y$ being complete. Note that Lemma~\ref{hom2kneser} can be used to construct examples in which neither $X$ nor $Y$ is complete. However, for the examples we constructed it turned out that $X \qarrow K_n$, $X \not\rightarrow K_n$, and $K_n$ is a subgraph of $Y$ for some $n \in \mathbb{N}$. Therefore we are interested in whether there exist graphs $X$ and $Y$ such that $X \qarrow Y$, $X \not\rightarrow Y$, and there is no $n \in \mathbb{N}$ such that $X \qarrow K_n \qarrow Y$. Finding a graph $Y$ with $\og(Y) > \og_q(Y) \ge 5$ would give us such an example in addition to providing a separation for odd girth.
The relation $``\rightarrow"$ gives rise to a partial order on homomorphic equivalence classes of graphs. This partial order is known as the homomorphism order of graphs, and has been heavily studied and possesses many remarkable properties \cite{Nesetril,Tardif}. In the same way, one can consider the quantum homomorphism order of graphs. The quantum homomorphism order can easily be shown to be a lattice with the same meet and join operations as the homomorphism order. Since $X \rightarrow Y \Rightarrow X \qarrow Y$, the quantum homomorphism order is a (poset) homomorphic image of the homomorphism order. It is now natural to ask which properties these orders share? In particular, is the quantum homomorphism order dense?
%The relation $``\rightarrow"$ on graphs is transitive and reflexive and is therefore a preorder. This can be used to construct an equivalence relation in the usual way, $X$ is a equivalent to $Y$ if $X \rightarrow Y$ and $Y \rightarrow X$, and this relation is known as ``homomorphic equivalence". We can quotient out the original preorder by this equivalence relation to obtain a partial order on homomorphic equivalence classes of graphs known as the homomorphism order of graphs. This partial order has been heavily studied and possesses many remarkable properties \cite{Nesetril,Tardif}. In particular this partial order is a lattice whose meet and join operations correspond to categorical product and disjoint union of graphs. Since $``\qarrow"$ is transitive and reflexive, we can perform the same trick to construct the quantum homomorphism order on graphs. Though we do not do it here, it is straightforward to show that this is a lattice as well, and has the same meet and join operations as the homomorphism order. Since $X \rightarrow Y \Rightarrow X \qarrow Y$, the quantum homomorphism order is a (poset) homomorphic image of the homomorphism order. It is now natural to ask which properties of the homomorphism order does the quantum homomorphism order share? In particular one could ask if the quantum homomorphism order is dense, or if one could prove a quantum version of the sparse incomparability lemma. Related to this is whether one can find graphs with arbitrarily large quantum chromatic number and arbitrarily large quantum odd girth. A potential candidate class of graphs would be the Kneser graphs, which can have arbitrarily large chromatic number and odd girth.
Although we have focused on quantum strategies for the homomorphism game, some of the proofs do not rely on the type of strategy being used. What separates results proved in this manner from results which require explicit use of the type of strategy being considered?
%We have also seen that classical strategies correspond to the usual homomorphisms, and that non-signalling strategies render the game trivial. Are there other interesting classes of strategies which could be considered? If there are many such classes, then proofs which do not require use of the strategy become even more important, as results proved in this manner will hold true for all types of strategies.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section*{Acknowledgements}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
The authors would like to thank Giannicola Scarpa for stimulating discussions and Simone Severini for his helpful notes. DR would like to thank Chris Godsil for his support throughout this project. We would also like to thank the anonymous referees of QIP 2013 for their helpful comments on our technical report.
%
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%\appendix
%\section{Appendix}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%\NormalForm*
%\proof
%The proof is structured as follows: First we show that we can assume WLOG that the shared state has full Schmidt rank. Then we show that, assuming this, the first and second claims above must hold. Next we show that using a maximally entangled state still gives us a winning strategy. Finally, we show that given a strategy satisfying the first three claims, if $Y$ is vertex transitive we can construct a strategy in which all of the projectors have the same rank.
%
%Let $\psi' \in \mathbb{C}^{d_A} \otimes \mathbb{C}^{d_B}$ be the state that Alice and Bob share and let $E'_{xy}, F'_{xy}$ for $x \in V(X), y \in V(Y)$ be the measurement operators used by Alice and Bob respectively. We can write $\psi'$ in its Schmidt decomposition as
%\[\psi' = \sum_{i=1}^d \sqrt{\lambda_i} \alpha_i \otimes \beta_i\]
%for some orthonormal sets $\{\alpha_1,\ldots,\alpha_d\} \subseteq \mathbb{C}^{d_A}$, $\{\beta_1,\ldots,\beta_d\} \subseteq \mathbb{C}^{d_B}$ with $\lambda_i > 0$. Now let
%\[P_A = \sum_{i=1}^d e_i \alpha_i^* \text{ and } P_B = \sum_{i=1}^d e_i \beta_i^*\]
%and let
%\begin{align*}
%\psi &:= (P_A \otimes P_B) \psi' = \sum_{i=1}^d \sqrt{\lambda_i}e_i \otimes e_i\\
%E_{xy} &:= P_A E'_{xy} P_A^* \ \ \forall x \in V(X), y \in V(Y)\\
%F_{xy} &:= P_B F'_{xy} P_B^* \ \ \forall x \in V(X), y \in V(Y).
%\end{align*}
%Note that $\{E_{xy}\}_{y \in V(Y)}$ and $\{F_{xy}\}_{y \in V(Y)}$ are indeed measurements and $\psi$ is a valid quantum state (has unit norm). The following calculation shows that replacing the shared state and measurement operators with their unprimed versions does not change the conditional probabilities $P(y,y'|x,x')$:
%\begin{align*}
%\psi^* E_{xy} \otimes F_{x'y'}\psi &= \psi^* \left(P_A E'_{xy} P_A^* \right) \otimes \left(P_B F'_{x'y'} P_B^* \right) \psi \\
%&= \psi^* (P_A \otimes P_B) (E'_{xy} \otimes F'_{x'y'}) (P_A^* \otimes P_B^*)\psi \\
%&= \psi'^* E'_{xy} \otimes F'_{x'y'} \psi'
%\end{align*}
%since $(P_A^* \otimes P_B^*)\psi = \psi'$. Clearly $\psi$ has full Schmidt rank, so we have proven the desired claim.
%
%Now we will show that the $E_{xy}$'s and $F_{xy}$'s are orthogonal projectors. Let $D := \text{diag}(\sqrt{\lambda_i})$. This means that $\vect(D) = \psi$, where $\vect(D)$ is the vector obtained from $D$ by stacking its rows on top of each other with the first row on top. In the following, the support of a matrix $E$, denoted $\supp(E)$ refers to the column space of $E$. However, we will also sometimes use $\supp(E)$ to refer to the orthogonal projection onto the column space of $E$. It will be up to the reader to distinguish which we mean from the context.
%
%Using that $A \otimes B \vect(X) = \vect(AXB^T)$ and $\vect(A)^* \vect(B) = \tr(A^* B)$ for all matrices $A,B,X$ such that these products are defined, we get that
%\begin{align*}
%\psi^*E_{xy} \otimes F_{x'y'} \psi &= \vect(D)^* \left(E_{xy} \otimes F_{x'y'}\right) \vect(D) \\
%&= \tr(D^* E_{xy} D F^T_{x'y'}) \\
%&= \tr(E_{xy} D F^T_{x'y'}D)
%\end{align*}
%since $D^* = D$. From the consistency conditions we get that $\tr(E_{xy} D F^T_{xy'}D) = 0$ for $y \ne y'$. Using this and the fact that $\sum_{y' \in V(Y)} E_{xy'} = I$, we obtain
%\[
%\tr\left(D F^T_{xy}D\right) = \sum_{y' \in V(Y)}\tr\left(E_{xy'} D F^T_{xy}D\right) = \tr\left(E_{xy} D F^T_{xy}D\right).
%\]
%This implies that the 1-eigenspace of $E_{xy}$ contains the support of $DF^T_{xy}D$ (this fact can be proved using spectral decomposition). All 1-eigenspaces of the $E_{xy'}$ are mutually orthogonal since $\sum_{y' \in V(Y)} E_{xy'} = I$ and the $E_{xy'}$ are all positive semidefinite. This implies that $\supp(DF^T_{xy}D) \bot \supp(DF^T_{xy'}D)$ for $y \ne y'$. Also since $D^2$ has full rank and
%\[\sum_{y' \in V(Y)} DF^T_{xy'}D = D^2\]
%we get that
%\[\bigoplus_{y' \in V(Y)} \supp(DF^T_{xy'}D) = \mathbb{C}^d.\]
%Also we have that for $y \ne y'$, $\tr(E_{xy} D F^T_{xy'}D) = 0$ which is equivalent to $\supp(E_{xy}) \bot \supp(DF^T_{xy'}D)$. Together these two facts give
%\[
%\supp(E_{xy}) \subseteq \left(\bigoplus_{y' \ne y} \supp(DF^T_{xy'}D)\right)^\bot = \supp(DF^T_{xy}D).
%\]
%Combining this with the fact that $\supp(DF^T_{xy}D)$ is contained in the 1-eigenspace of $E_{xy}$, we get that $\supp(E_{xy})$ is equal to the 1-eigenspace of $E_{xy}$ and thus $E_{xy}$ is a projector. This of course implies that $E_{xy} = \supp(DF^T_{xy}D)$. Similarly, we can conclude that the $F_{xy}$ are projectors and that $F_{xy} = \supp(DE^T_{xy}D)$.
%
%Now we show that $D$ commutes with both $E_{xy}$ and $F_{xy}$, and use this to obtain that $E_{xy} = \overline{F}_{xy}$. Recall that for any two positive semidefinite matrices $A$ and $B$, $\tr(AB) = 0$ if and only if $\tr(A\supp(B)) = 0$ if and only if $AB = 0$. Using this and that $D$ has full rank, we get that
%\[0 = \tr(DE_{xy}DF^T_{xy'}) = \tr\left((DE_{xy}D)\supp(DE_{xy'}D)\right) \ \Leftrightarrow \ 0 = E_{xy}D^2E_{xy'}.\]
%Therefore,
%\[D^2 = \sum_{y,y' \in V(Y)} E_{xy} D^2 E_{xy'} = \sum_{y \in V(Y)} E_{xy} D^2 E_{xy}.\]
%Multiplying the first and last expressions above on either side by $E_{xy'}$ gives that
%\[D^2 E_{xy'} = E_{xy'} D^2 E_{xy'} = E_{xy'} D^2.\]
%Since the $E_{xy}$ commute with $D^2$, they also commute with $D$. Similarly the $F_{xy}$ commute with $D$. Using this we get that
%\[E_{xy} = \supp(DF^T_{xy}D) = \supp(F^T_{xy}D^2) = \supp(F^T_{xy}) = F^T_{xy} = \overline{F}_{xy}\]
%since $D$ is invertible and $F_{xy}$ is Hermitian. So we have proven the second claim.
%
%Now using the fact that $D$ commutes with $F^T_{xy}$, we see that
%\[0 = \psi^* E_{xy} \otimes F_{x'y'} \psi = \tr\left(E_{xy} (D F^T_{x'y'}D)\right)\]
%if and only if
%\[E_{xy} D F^T_{x'y'}D = 0\]
%if and only if
%\[E_{xy}F^T_{x'y'} = 0.\]
%
%This means that we can replace $\psi$ with any state which will give an invertible $D$ that commutes with the $E_{xy}$ and $F_{xy}$. If we use the state $\Psi = \frac{1}{\sqrt{d}}\sum_i e_i \otimes e_i$, then $D$ is a scalar multiple of the identity matrix and thus commutes with all other matrices. Therefore we can assume WLOG that we use the state $\Psi$ and so we have proven the third claim.
%
%Now suppose that $Y$ is vertex transitive and that $G = \{\varphi_1,\ldots,\varphi_m\}$ is a subgroup of $\text{Aut}(Y)$ that acts transitively on $V(Y)$. Let
%\begin{align*}
%\psi'' &= \frac{1}{\sqrt{md}} \sum_{i \in [d], j \in [m]} e_i \otimes e_j \otimes e_i \otimes e_j \\
%E''_{xy} &= \sum_{i=1}^m E_{x\varphi_j(y)} \otimes e_j e_j^* \\
%F''_{xy} &= \sum_{i=1}^m F_{x\varphi_j(y)} \otimes e_j e_j^*.
%\end{align*}
%A straightforward but somewhat tedious calculation shows that these states and operators still give a valid quantum strategy for the $(X,Y)$-homomorphism game, and the operators will all have the same rank since $G$ acts transitively on $V(Y)$. So we have proven the final claim.
%\qed
%
%
%
\bibliographystyle{alphaurl}
%\bibliography{Homos}
\input{Paper.bbl}
\end{document}
%\documentclass{article}
%\usepackage{tikz}
%\usetikzlibrary{arrows}
%\begin{document}
\begin{tikzpicture}
[vx/.style = {inner sep = 3pt, text height=1.5ex,text depth=.25ex}];
\def\hstep{-1};
\def\wstep{1.5};
\def\wstepn{1.1};
\foreach \i/\name in
{1/$\chi$,2/$\chi_q$,3/$\xi_f$,4/$\bar{\vartheta}$,
5/$\;\,\bar{\Theta}^*$, 8/$\omega$}{
\node[vx](\i) at (0,\i*\hstep) {\name};}
\node[vx](2l) at (-1*\wstep,2*\hstep) {$\chi_f$};
\node[vx](2r) at (\wstep,2*\hstep) {$\xi$};
\node[vx](6l) at (-1*\wstepn,6*\hstep) {$\bar{c}_0^*$};
\node[vx](7l) at (-1*\wstepn,7*\hstep) {$\omega_q$};
\node[vx](65r) at (\wstepn,6.5*\hstep) {$\bar{\Theta}$};
\draw (1) -- (2) -- (3) -- (4);
\draw (1) -- (2l) -- (3);
\draw (1) -- (2r) -- (3);
\draw[dashed] (4) -- (5);
\draw (5) -- (6l);
\draw[dashed] (6l) -- (7l);
\draw (7l) -- (8);
\draw (5) -- (65r) -- (8);
\node[rectangle,minimum height=15pt, minimum width=60pt,
rounded corners=7pt,draw=black!40](rect)
at (0.5*\wstep,2*\hstep) {};
\end{tikzpicture}
%\end{document}
\newcommand{\etalchar}[1]{$^{#1}$}
\begin{thebibliography}{CLMW10}
\bibitem[AC85]{nohomo}
Michael~O. Albertson and Karen~L. Collins.
\newblock Homomorphisms of {$3$}-chromatic graphs.
\newblock {\em Discrete Math.}, 54(2):127--132, 1985.
\newblock \href {http://dx.doi.org/10.1016/0012-365X(85)90073-1}
{\path{doi:10.1016/0012-365X(85)90073-1}}.
\bibitem[AHKS06]{Avis}
David Avis, Jun Hasegawa, Yosuke Kikuchi, and Yuuya Sasaki.
\newblock A quantum protocol to win the graph colouring game on all hadamard
graphs.
\newblock {\em IEICE Trans. Fundam. Electron. Commun. Comput. Sci.},
E89-A(5):1378--1381, 2006.
\newblock \href {http://arxiv.org/abs/quant-ph/0509047v4}
{\path{arXiv:quant-ph/0509047v4}}, \href
{http://dx.doi.org/10.1093/ietfec/e89-a.5.1378}
{\path{doi:10.1093/ietfec/e89-a.5.1378}}.
\bibitem[BBG12]{Jop}
Jop Bri\"{e}t, Harry Buhrman, and Dion Gijswijt.
\newblock Violating the {S}hannon capacity of metric graphs with entanglement.
\newblock 2012.
\newblock \href {http://arxiv.org/abs/1207.1779v1} {\path{arXiv:1207.1779v1}}.
\bibitem[BCT99]{BCT99}
Gilles Brassard, Richard Cleve, and Alain Tapp.
\newblock Cost of exactly simulating quantum entanglement with classical
communication.
\newblock {\em Phys. Rev. Lett.}, 83:1874--1877, 1999.
\newblock \href {http://arxiv.org/abs/quant-ph/9901035}
{\path{arXiv:quant-ph/9901035}}, \href
{http://dx.doi.org/10.1103/PhysRevLett.83.1874}
{\path{doi:10.1103/PhysRevLett.83.1874}}.
\bibitem[BCW98]{BCW98}
Harry Buhrman, Richard Cleve, and Avi Wigderson.
\newblock Quantum vs. classical communication and computation.
\newblock In {\em Proceedings of the thirtieth annual ACM symposium on Theory
of computing}, STOC '98, pages 63--68. ACM, 1998.
\newblock \href {http://arxiv.org/abs/quant-ph/9802040}
{\path{arXiv:quant-ph/9802040}}, \href
{http://dx.doi.org/10.1145/276698.276713} {\path{doi:10.1145/276698.276713}}.
\bibitem[Bei10]{Beigi}
Salman Beigi.
\newblock Entanglement-assisted zero-error capacity is upper-bounded by the
{L}ov\'{a}sz theta function.
\newblock {\em Phys. Rev. A}, 82:10303--10306, 2010.
\newblock \href {http://arxiv.org/abs/1002.2488} {\path{arXiv:1002.2488}},
\href {http://dx.doi.org/10.1103/PhysRevA.82.010303}
{\path{doi:10.1103/PhysRevA.82.010303}}.
\bibitem[CHTW04]{Cleve04}
Richard Cleve, Peter Hoyer, Ben Toner, and John Watrous.
\newblock Consequences and limits of nonlocal strategies.
\newblock In {\em 19th IEEE Annual Conference on Computational Complexity},
pages 236--249, 2004.
\newblock \href {http://arxiv.org/abs/quant-ph/0404076v2}
{\path{arXiv:quant-ph/0404076v2}}, \href
{http://dx.doi.org/10.1109/CCC.2004.1313847}
{\path{doi:10.1109/CCC.2004.1313847}}.
\bibitem[CLMW10]{entcap}
Toby~S. Cubitt, Debbie Leung, William Matthews, and Andreas Winter.
\newblock Improving zero-error classical communication with entanglement.
\newblock {\em Phys. Rev. Lett.}, 104:230503, Jun 2010.
\newblock \href {http://arxiv.org/abs/0911.5300} {\path{arXiv:0911.5300}},
\href {http://dx.doi.org/10.1103/PhysRevLett.104.230503}
{\path{doi:10.1103/PhysRevLett.104.230503}}.
\bibitem[CMN{\etalchar{+}}07]{qchrom}
Peter~J. Cameron, Ashley Montanaro, Michael~W. Newman, Simone Severini, and
Andreas Winter.
\newblock On the quantum chromatic number of a graph.
\newblock {\em Electr. J. Comb.}, 14(1), 2007.
\newblock \href {http://arxiv.org/abs/quant-ph/0608016}
{\path{arXiv:quant-ph/0608016}}.
\bibitem[DSW11]{Winter}
Runyao Duan, Simone Severini, and Andreas Winter.
\newblock Zero-error communication via quantum channels and a quantum
{L}ov\'asz $\vartheta$ function.
\newblock In {\em Proc. Int. Symp. Inf. Theory}, pages 64--68, 2011.
\newblock \href {http://arxiv.org/abs/1002.2514} {\path{arXiv:1002.2514}},
\href {http://dx.doi.org/10.1109/ISIT.2011.6034211}
{\path{doi:10.1109/ISIT.2011.6034211}}.
\bibitem[FIG11]{qchrom-aqis}
Junya Fukawa, Hiroshi Imai, and Fran\c{c}ois~Le Gall.
\newblock Quantum coloring games via symmetric {SAT} games.
\newblock In {\em Asian Conference on Quantum Information Science (AQIS'11)},
2011.
\bibitem[FR87]{frankrod}
Peter Frankl and Vojt{\v{e}}ch R{\"o}dl.
\newblock Forbidden intersections.
\newblock {\em Trans. Amer. Math. Soc.}, 300(1):259--286, 1987.
\newblock \href {http://dx.doi.org/10.2307/2000598}
{\path{doi:10.2307/2000598}}.
\bibitem[God03]{interestinggraphs}
Chris Godsil.
\newblock Interesting graphs and their colourings.
\newblock Unpublished notes, 2003.
\bibitem[GR01]{AGT}
Chris Godsil and Gordon Royle.
\newblock {\em Algebraic graph theory}, volume 207 of {\em Graduate Texts in
Mathematics}.
\newblock Springer-Verlag, New York, 2001.
\newblock \href {http://dx.doi.org/10.1007/978-1-4613-0163-9}
{\path{doi:10.1007/978-1-4613-0163-9}}.
\bibitem[GW02]{Galliard02}
Viktor Galliard and Stefan Wolf.
\newblock Pseudo-telepathy, entanglement, and graph colorings.
\newblock In {\em IEEE International Symposium on Information Theory}, page
101, 2002.
\newblock \href {http://dx.doi.org/10.1109/ISIT.2002.1023373}
{\path{doi:10.1109/ISIT.2002.1023373}}.
\bibitem[HN04]{Nesetril}
Pavol Hell and Jaroslav Ne\v{s}et\v{r}il.
\newblock {\em Graphs and homomorphisms}.
\newblock Oxford University Press, 2004.
\bibitem[HT97]{Tardif}
Ge\v{n}a Hahn and Claude Tardif.
\newblock Graph homomorphisms: structure and symmetry.
\newblock In {\em Graph symmetry}, volume 497 of {\em NATO Adv. Sci. Inst. Ser.
C Math. Phys. Sci.}, pages 107--166. Kluwer Acad. Publ., 1997.
\bibitem[KMS98]{chivec}
David Karger, Rajeev Motwani, and Madhu Sudan.
\newblock Approximate graph coloring by semidefinite programming.
\newblock {\em J. ACM}, 45(2):246--265, 1998.
\newblock \href {http://arxiv.org/abs/cs/9812008} {\path{arXiv:cs/9812008}},
\href {http://dx.doi.org/10.1145/274787.274791}
{\path{doi:10.1145/274787.274791}}.
\bibitem[LMM{\etalchar{+}}12]{LMMOR}
Debbie Leung, Laura Man\v{c}inska, William Matthews, Maris Ozols, and Aidan
Roy.
\newblock Entanglement can increase asymptotic rates of zero-error classical
communication over classical channels.
\newblock {\em Commun. Math. Phys.}, 311:97--111, 2012.
\newblock \href {http://arxiv.org/abs/1009.1195} {\path{arXiv:1009.1195}},
\href {http://dx.doi.org/10.1007/s00220-012-1451-x}
{\path{doi:10.1007/s00220-012-1451-x}}.
\bibitem[Lov79]{lovasz}
L{\'a}szl{\'o} Lov{\'a}sz.
\newblock On the {S}hannon capacity of a graph.
\newblock {\em IEEE Trans. Inform. Theory}, 25(1):1--7, 1979.
\newblock \href {http://dx.doi.org/10.1109/TIT.1979.1055985}
{\path{doi:10.1109/TIT.1979.1055985}}.
\bibitem[MSS12]{MSS}
Laura Man\v{c}inska, Giannicola Scarpa, and Simone Severini.
\newblock A generalization of {K}ochen-{S}pecker sets relates quantum coloring
to entanglement-assisted channel capacity.
\newblock 2012.
\newblock \href {http://arxiv.org/abs/1207.1111} {\path{arXiv:1207.1111}}.
\bibitem[SS12]{SS12}
Giannicola Scarpa and Simone Severini.
\newblock Kochen-{S}pecker sets and the rank-1 quantum chromatic number.
\newblock {\em IEEE Trans. Inf. Theory}, 58(4):2524--2529, 2012.
\newblock \href {http://arxiv.org/abs/1106.0712} {\path{arXiv:1106.0712}},
\href {http://dx.doi.org/10.1109/TIT.2011.2178018}
{\path{doi:10.1109/TIT.2011.2178018}}.
\bibitem[WBC11]{calculus}
Christopher~J. Wood, Jacob~D. Biamonte, and David~G. Cory.
\newblock Tensor networks and graphical calculus for open quantum systems.
\newblock 2011.
\newblock \href {http://arxiv.org/abs/1111.6950} {\path{arXiv:1111.6950}}.
\end{thebibliography}