\section{The free theory} \label{sec:free} We begin with the theory at $\lambda=0$. The results of this section are due to Sundborg \cite{Sundborg} and AMMPV \cite{AMMPV1}; AMMPV trace a special case to Hallin and Persson \cite{HallinPersson}, and the counting of single-trace words is also in Polyakov \cite{Polyakov}. A short review is \cite{AMMPVreview}. We review the free theory because every later section perturbs around it, and because the mechanism can be followed completely. We put the theory on a three-sphere and set the radius to one. Energies are then measured in units of $1/R$. We write $\beta=1/T$ for the inverse temperature and \begin{equation} x=e^{-\beta},\qquad y=\sqrt{x} \end{equation} for the Boltzmann factor of a unit of energy and its square root. The square root appears because fermions on $S^3$ have half-integer energies. Both $x$ and $y$ are parametrisations of the temperature. \subsection{Letters} A free field on $S^3$ is a collection of harmonic oscillators, one for each spherical harmonic. The rotation group of the sphere is $SO(4)=SU(2)_L\times SU(2)_R$, and the harmonics are labelled by a pair of spins $(j_L,j_R)$. Table~\ref{tab:letters} lists the modes of the three kinds of field in $\cN=4$ super Yang--Mills theory. We call a single oscillator mode a \emph{letter}, since the states of the gauge theory will be words made from them. \begin{table}[t] \centering \begin{tabular}{llll} \toprule field & harmonics $(j_L,j_R)$ & energy & single-letter partition function \\ \midrule 6 real scalars & $(j,j)$ & $2j+1$ & $6\times x(1+x)/(1-x)^3$ \\ 4 Weyl fermions & $(j+\half,j)$ and $(j,j+\half)$ & $2j+\tfrac32$ & $4\times 4x^{3/2}/(1-x)^3$ \\ gauge field & $(j+1,j)$ and $(j,j+1)$ & $2j+2$ & $(6x^2-2x^3)/(1-x)^3$ \\ \bottomrule \end{tabular} \caption{The letters of $\cN=4$ super Yang--Mills theory on a unit $S^3$. In each row $j=0,\half,1,\ldots$ The scalars are conformally coupled, which gives the lowest mode energy one. Only the transverse modes of the gauge field are listed. In Coulomb gauge they are the only ones that propagate.} \label{tab:letters} \end{table} The single-letter partition functions in the last column are the sums of $x^{E}$ over the modes in each row. For the scalars, for instance, the harmonic $(j,j)$ has $(2j+1)^2$ components of energy $2j+1$, and $\sum_{n\ge1}n^2x^n=x(1+x)/(1-x)^3$. Adding the rows gives the bosonic and fermionic totals \begin{equation} z_B(x)=\frac{6x+12x^2-2x^3}{(1-x)^3},\qquad z_F(x)=\frac{16x^{3/2}}{(1-x)^3}. \label{eq:zBzF} \end{equation} Their sum has a compact form in terms of $y=\sqrt x$, \begin{equation} z(x)=z_B+z_F=\frac{2y^2(3-y)}{(1-y)^3}. \label{eq:z} \end{equation} There is a second way to read table~\ref{tab:letters}. The theory on $\mathbb{R}\times S^3$ is conformally equivalent to the theory on $\mathbb{R}^4$, and under that map the energy on the sphere becomes the scaling dimension. The letters are then the elementary fields at the origin of $\mathbb{R}^4$ with any number of derivatives acting on them, modulo the equations of motion: $\partial^k\phi$, $\partial^k\psi$, $\partial^kF$. The lowest scalar mode is $\phi$ itself with dimension one, and so on. We will use both languages. \subsection{Gauss's law and the matrix model} Every letter is an $N\times N$ matrix, because the fields are in the adjoint representation of $SU(N)$. If there were no further constraint, the partition function would be that of $N^2$ copies of table~\ref{tab:letters} and there would be no transition. The sphere, however, is compact. Electric flux lines that leave a charge have nowhere to end, so the total charge on $S^3$ must vanish. For a non-abelian theory the statement is that physical states are singlets of the global $SU(N)$. This is the one effect of the gauge field that survives at zero coupling. A projection onto singlets is an average over the group. If $\hat U$ denotes the operator that rotates every letter by the group element $U$, then \begin{equation} Z(x)=\int dU\;\Tr\big(x^{H}\hat U\big), \label{eq:ZintU} \end{equation} where $dU$ is the Haar measure and the trace is over the unconstrained Fock space. In the path integral on $S^3\times S^1$ the same $U$ appears as the holonomy of the gauge field around the thermal circle, $U=P\exp(i\oint A_0)$. It is the one mode of the gauge field that has no quadratic action, which is why it must be integrated exactly. The two descriptions agree. We call $U$ the holonomy, or the Polyakov loop matrix. The trace in \eqref{eq:ZintU} is that of free oscillators in a background $U$ and is elementary. An adjoint letter of energy $E$ contributes a factor $\det_{\rm adj}(1\mp x^EU)^{\mp1}$. Taking the logarithm, expanding it, and using $\tr_{\rm adj}U^n=|\tr\,U^n|^2$ at large $N$ gives \begin{equation} Z(x)=\int dU\,\exp\Big[\sum_{n=1}^\infty\frac{z_n(x)}{n}\,|\tr\,U^n|^2\Big], \qquad z_n(x)=z_B(x^n)+(-1)^{n+1}z_F(x^n). \label{eq:Zfree} \end{equation} The sign $(-1)^{n+1}$ is the usual statistics factor for fermions. The integrand depends on $U$ only through its eigenvalues $e^{i\theta_a}$. At large $N$ they form a continuous density $\rho(\theta)$ on the circle, normalised to one, whose Fourier coefficients are the moments \begin{equation} u_n=\frac1N\tr\,U^n=\int d\theta\,\rho(\theta)e^{in\theta}, \qquad \rho(\theta)=\frac1{2\pi}\Big[1+\sum_{n\ge1}\big(u_ne^{-in\theta}+{\rm c.c.}\big)\Big]. \label{eq:rho} \end{equation} The first moment $u_1$ is the Polyakov loop. The Haar measure, written in terms of the eigenvalues, is a repulsion between them, $\prod_{ax_H$ the coefficient of $|u_1|^2$ in \eqref{eq:Seff0} is negative. The action decreases as $|u_1|$ grows, and as a function of an unconstrained $u_1$ it is unbounded below. The run-away is stopped by the fact that the $u_n$ are not independent variables. They are Fourier coefficients of a density, and a density cannot be negative. With only $u_1$ turned on, $\rho=(1+2u_1\cos\theta)/2\pi$ is non-negative if and only if \begin{equation} |u_1|\le\half . \label{eq:half} \end{equation} Beyond this point the density develops a gap, an interval of $\theta$ around $\pi$ in which there are no eigenvalues. The repulsion among the remaining eigenvalues is then stronger than the simple $|u_1|^2$ of \eqref{eq:Seff0}. We give the exact form in the next section. The free theory therefore behaves as follows. For $T