\chapter{From a window estimate to the gap of the block dynamics}\label{s:block} In this section, we prove that the block dynamics of \zcref{s:quantum-block}, with blocks of side $b = 2M + 1$, has spectral gap at least $1/2$ for all sufficiently large $M$, uniformly in $L, N \ge b + 4$. This is \zcref{p:blockgap}. We prove it under Hypothesis~(W) below, which asks that the laws of certain sets of spins in a large box have a component of positive mass that is common to all boundary configurations. Hypothesis~(W) is proved in \zcref{t:window}, and \zcref{p:blockgap} is combined with \zcref{t:comparison} in the proof of \zcref{t:main} in \zcref{s:computer-proofmain}. The heuristics are explained in \zcref{s:ov-block}. We follow the proof of \cite[Proposition~2.1]{MOS94}, where the single-spin space is finite. For a finite single-spin space and positive weights, two conditional laws on a set of bounded size have an overlap that is bounded below uniformly in the boundary conditions. For the continuous spins here, whose densities can vanish, we prove the corresponding bound in \zcref{l:overlap}. Throughout this section, $M \ge 1$ is an integer, and we set \begin{equation*} b = 2M + 1 , \qquad \ell = M^{1/4} , \qquad \mathcal C_r = \{ x \in \Z^2 : |x|_\infty \le r \} \quad \text{for } r \ge 0 . \end{equation*} We call $\mathcal C_M$ the box; it has $b^2$ sites. For configurations $n$ and $n'$ on a set $D$, we let $d_{\mathrm H}(n, n') = |\{ x \in D : n_x \neq n'_x \}|$ denote their Hamming distance, and for $D' \subset D$, we say that $n$ and $n'$ agree on $D'$ if $n_x = n'_x$ for all $x \in D'$. Given a finite set $A \subset \Z^2$, a configuration $n$ on $A$, and a boundary configuration $\eta$ of $A$, we let $n \eta \in (\Sph)^{\Z^2}$ denote the configuration with $(n \eta)_x = n_x$ for $x \in A$ and $(n \eta)_x = \eta_x$ for $x \notin A$. \section{Two properties of the local measures}\label{s:block-local} We begin with the consistency of the local measures \eqref{e:muA} under resampling. \begin{lemma}\label{l:resample} Let $A \subset \Z^2$ be finite, let $F \subset A$, and let $\eta$ be a boundary configuration of $A$. Let $n$ be a random configuration with law $\mu_A^\eta$, and let $\mathcal G$ be a $\sigma$-algebra such that $n$ is $\mathcal G$-measurable. Let $\hat n$ be a random configuration on $A$ such that $\hat n_{A \setminus F} = n_{A \setminus F}$ and the conditional law of $\hat n_F$ given $\mathcal G$ is $\mu_F^{(n \eta)_{\Z^2 \setminus F}}$. Then $\hat n$ has law $\mu_A^\eta$. \end{lemma} \begin{proof} We write the product of the factors in \eqref{e:muA} as $w_1(n) \, w_2(n)$, where $w_1$ is the product of the factors that involve a site of $F$ and $w_2$ is the product of the other factors. Then $w_2(n)$ is a function of $n_{A \setminus F}$. As a function of $n_F$, the product $w_1(n)$ is the product of the factors of $\mu_F^{(n \eta)_{\Z^2 \setminus F}}$ in \eqref{e:muA}, so that \begin{equation}\label{e:resample-Z} \int w_1(n) \prod_{x \in F} \dd n_x = Z_F^{(n \eta)_{\Z^2 \setminus F}} . \end{equation} Let $f$ be a bounded measurable function on $(\Sph)^A$, and for a configuration $m$ on $F$, let $(m, n_{A \setminus F})$ denote the configuration on $A$ that agrees with $m$ on $F$ and with $n$ on $A \setminus F$. By the definition of $\hat n$ and Fubini's theorem, we have \begin{align*} \E [f(\hat n)] &= \frac{1}{Z_A^\eta} \int \frac{w_2(n)}{Z_F^{(n \eta)_{\Z^2 \setminus F}}} \Bigl( \int w_1(n) \prod_{x \in F} \dd n_x \Bigr) \\ &\qquad \times \Bigl( \int f(m, n_{A \setminus F}) \, w_1(m, n_{A \setminus F}) \prod_{x \in F} \dd m_x \Bigr) \prod_{x \in A \setminus F} \dd n_x \\ &= \frac{1}{Z_A^\eta} \int f(m, n_{A \setminus F}) \, w_1(m, n_{A \setminus F}) \, w_2(n) \prod_{x \in F} \dd m_x \prod_{x \in A \setminus F} \dd n_x = \mu_A^\eta(f) , \end{align*} where we used \eqref{e:resample-Z} in the second equality. \end{proof} We will apply \zcref{l:resample} to pairs of configurations. Fix a finite set $A \subset \Z^2$, a subset $D \subset A$, and boundary configurations $\eta$ and $\eta'$ of $A$. Let $(n, n')$ be a random pair of configurations on $A$ that is measurable with respect to a $\sigma$-algebra $\mathcal G$. We say that a random pair $(\hat n, \hat n')$ is obtained from $(n, n')$ by a \emph{coupled resampling} of $D$ if $\hat n_{A \setminus D} = n_{A \setminus D}$, $\hat n'_{A \setminus D} = n'_{A \setminus D}$, and the conditional law of $(\hat n_D, \hat n'_D)$ given $\mathcal G$ is a coupling of $\mu_D^{(n \eta)_{\Z^2 \setminus D}}$ and $\mu_D^{(n' \eta')_{\Z^2 \setminus D}}$. If $n$ has law $\mu_A^\eta$ and $n'$ has law $\mu_A^{\eta'}$, then \zcref{l:resample}, applied to $\hat n$ and to $\hat n'$, shows that $\hat n$ has law $\mu_A^\eta$ and $\hat n'$ has law $\mu_A^{\eta'}$. Next, we bound the overlap of two local measures on a set of bounded size. Let $\lambda$ be a measure, let $p$ and $p'$ be probability densities with respect to $\lambda$, and set $\gamma = \int \min(p, p') \dd \lambda$. We call the law of the following random pair $(m, m')$ the \emph{maximal coupling} of $p \lambda$ and $p' \lambda$. With probability $\gamma$, we draw $m$ from the density $\min(p, p')/\gamma$ and set $m' = m$. With probability $1 - \gamma$, we draw $m$ and $m'$ independently from the densities $(p - \min(p, p'))/(1 - \gamma)$ and $(p' - \min(p, p'))/(1 - \gamma)$. We omit any branch of probability zero. This is a coupling of $p \lambda$ and $p' \lambda$ under which $m = m'$ with probability at least $\gamma$. When $p = p'$ almost everywhere, we have $\gamma = 1$ and $m = m'$ almost surely. For an integer $K \ge 1$, we set \begin{equation}\label{e:overlap} \varepsilon_K = \frac12 (16 K)^{-4K} . \end{equation} \begin{lemma}\label{l:overlap} Let $K \ge 1$ be an integer, and let $A \subset \Z^2$ be a finite set with $|A| \le K$. Let $\eta$ and $\eta'$ be boundary configurations of $A$, and let $p^\eta$ and $p^{\eta'}$ denote the densities of $\mu_A^\eta$ and $\mu_A^{\eta'}$ with respect to $\prod_{x \in A} \dd n_x$. Then \begin{equation*} \int \min(p^\eta, p^{\eta'}) \prod_{x \in A} \dd n_x \ge \varepsilon_K . \end{equation*} In particular, under the maximal coupling of $\mu_A^\eta$ and $\mu_A^{\eta'}$, the two configurations are equal with probability at least $\varepsilon_K$. \end{lemma} \begin{proof} We may assume that $A \neq \emptyset$. The product of the factors in \eqref{e:muA} has $k$ factors of the form $1 + n_x \cdot n_y$ or $1 + n_x \cdot \eta_y$, with values in $[0, 2]$, where $k \le 4 |A| \le 4K$, since each factor involves a site of $A$ and each site of $A$ has four neighbors. This gives $Z_A^\eta \le 2^k$, and the same holds for $\eta'$. Under the product measure $\prod_{x \in A} \dd n_x$, each factor is uniformly distributed on $[0, 2]$. Indeed, by Archimedes' theorem, the inner product of a uniform random point of $\Sph$ with a unit vector is uniformly distributed on $[-1, 1]$, and for a factor $1 + n_x \cdot n_y$ with $x, y \in A$, we apply this conditionally on $n_y$. We set $t = (8K)^{-1}$. Let $\mathcal Y$ denote the set of configurations $n \in (\Sph)^A$ at which all factors of $p^\eta$ and of $p^{\eta'}$ are at least $t$. The two densities have at most $8K$ factors in total, and each of them is smaller than $t$ on a set of measure $t/2$. By a union bound, the complement of $\mathcal Y$ has measure at most $8K \cdot t/2 = 1/2$. (The union bound does not require the factors to be independent.) On $\mathcal Y$, we have \begin{equation*} p^\eta(n) \ge \frac{t^k}{2^k} \ge \Bigl( \frac{t}{2} \Bigr)^{4K} = (16 K)^{-4K} , \end{equation*} and the same bound holds for $p^{\eta'}$. Integrating over $\mathcal Y$, we obtain \begin{equation*} \int \min(p^\eta, p^{\eta'}) \prod_{x \in A} \dd n_x \ge (16 K)^{-4K} \int \1_{\mathcal Y}(n) \prod_{x \in A} \dd n_x \ge \frac12 (16 K)^{-4K} = \varepsilon_K . \qedhere \end{equation*} \end{proof} \section{Shells}\label{s:block-shells} We now introduce the sets on which the coupling of \zcref{s:block-coupling} operates. They are a variant of the shells in the proof of \cite[Proposition~2.1]{MOS94}. For $v \in \Z^2$ and an integer $r \ge 1$, we call $\{ x \in \Z^2 : |x - v|_\infty = r \}$ the \emph{sphere} of radius $r$ around $v$. It is the boundary of the square $v + [-r, r]^2$, and we call the four segments $\{v_1 \pm r\} \times [v_2 - r, v_2 + r]$ and $[v_1 - r, v_1 + r] \times \{v_2 \pm r\}$ of lattice points its \emph{sides}. Fix $y \in \Z^2$ with $\dist_\infty(y, \mathcal C_M) = 1$, and set $d_0 = 24$. For integers $i$ with $2 \le i \le \ell$, we set $k_i = \lfloor i \ell \rfloor$ and \begin{equation}\label{e:shells} \begin{aligned} F_i &= \{ x \in \mathcal C_M : |x - y|_\infty > (i - 1) \ell \} , & \Sigma_i &= \{ x \in \mathcal C_M : i \ell - 1 < |x - y|_\infty \le i \ell \} , \\ \Sigma_i^\partial &= \{ x \in \Sigma_i : \dist_\infty(x, \Z^2 \setminus \mathcal C_M) \le d_0 \} , & W_i &= \Sigma_i \setminus \Sigma_i^\partial . \end{aligned} \end{equation} We call $\Sigma_i$ the $i$-th \emph{shell}, $\Sigma_i^\partial$ its \emph{ends}, and $W_i$ the \emph{window} of the shell $\Sigma_i$. Since $|x - y|_\infty$ is an integer, we have \begin{equation*} \Sigma_i = \{ x \in \mathcal C_M : |x - y|_\infty = k_i \} , \end{equation*} which is the part in $\mathcal C_M$ of the sphere of radius $k_i$ around $y$. Observe that $\Sigma_i \subset F_i$, since $k_i > i \ell - 1 \ge (i - 1) \ell$. We let $I_{\mathrm{corner}}$ denote the set of integers $i$ with $2 \le i \le \ell$ such that $\dist_\infty(v, \Sigma_i) < \ell/3$ for some corner site $v \in \{-M, M\}^2$ of $\mathcal C_M$. These sets are illustrated in \zcref{f:shells}. \begin{figure}[ht] \centering \begin{tikzpicture}[scale=0.5] % F_i: the box minus the square of radius (i - 1) ell around y \fill[gray!20] (0,0) -- (4.5,0) -- (4.5,2) -- (9.5,2) -- (9.5,0) -- (14,0) -- (14,10) -- (0,10) -- cycle; % sites within distance d_0 of the complement of the box \draw[dashed] (0.8,0.8) rectangle (13.2,9.2); % the box \draw[thick] (0,0) rectangle (14,10); \node at (2.1,8.2) {$\mathcal C_M$}; \node at (10.6,7.6) {$F_i$}; % shell i - 1 \draw[line width=1pt] (4.5,0) -- (4.5,2) -- (9.5,2) -- (9.5,0); \node[above] at (7,2) {$\Sigma_{i - 1}$}; % shell i: window and ends \draw[line width=2pt, blue] (3,0.8) -- (3,3.5) -- (11,3.5) -- (11,0.8); \draw[line width=2pt, red] (3,0) -- (3,0.8); \draw[line width=2pt, red] (11,0) -- (11,0.8); \node[above] at (7,3.5) {$W_i$}; \node[below] at (3,0) {$\Sigma_i^\partial$}; \node[below] at (11,0) {$\Sigma_i^\partial$}; % shell i + 1 \draw[line width=1pt] (1.5,0) -- (1.5,5) -- (12.5,5) -- (12.5,0); \node[above] at (7,5) {$\Sigma_{i + 1}$}; % the site y \fill (7,-0.5) circle (3pt); \node[below] at (7,-0.5) {$y$}; % width d_0 \draw[decorate, decoration={brace, amplitude=3pt}] (14.25,0.8) -- (14.25,0) node[midway, right=3pt] {$d_0$}; \end{tikzpicture} \caption{The sets of \eqref{e:shells} for a site $y$ adjacent to the lower side of the box $\mathcal C_M$ (schematic, not to scale). The thick lines are the shells $\Sigma_{i - 1}$, $\Sigma_i$, and $\Sigma_{i + 1}$, the parts in $\mathcal C_M$ of the spheres of radii $k_{i - 1}$, $k_i$, and $k_{i + 1}$ around $y$. The shaded region is $F_i$. The sites between the dashed line and the boundary of $\mathcal C_M$ are those at $\ell^\infty$ distance at most $d_0$ from $\Z^2 \setminus \mathcal C_M$. The shell $\Sigma_i$ consists of its two ends $\Sigma_i^\partial$ (red), which lie in this region, and its window $W_i$ (blue).} \label{f:shells} \end{figure} We set $M_{\mathrm g} = 73^4$. The following lemma collects the properties of the shells that we use below and in \zcref{s:cut}. \begin{lemma}\label{l:shellgeom} Let $M \ge M_{\mathrm g}$, and let $y \in \Z^2$ satisfy $\dist_\infty(y, \mathcal C_M) = 1$. Then the following hold. \begin{enumerate}[label=(\alph*)] \item\label{i:shellgeom-corner} $|I_{\mathrm{corner}}| \le 4$. \item\label{i:shellgeom-sides} Let $i \in \{2, \dots, \lfloor \ell \rfloor\} \setminus I_{\mathrm{corner}}$, let $j \in \{1, 2\}$ and $t \in \{y_j - k_i, y_j + k_i\}$, and suppose that the side \begin{equation*} S = \{ x \in \Z^2 : x_j = t, \ |x_{3 - j} - y_{3 - j}| \le k_i \} \end{equation*} of the sphere of radius $k_i$ around $y$ meets $\mathcal C_M$. Then $|t| \le M + 1 - \ell/3$. Equivalently, $S$ is at $\ell^\infty$ distance at least $\ell/3$ from each of the half-planes $\{x \in \Z^2 : x_j \ge M + 1\}$ and $\{x \in \Z^2 : x_j \le -M - 1\}$. \item\label{i:shellgeom-ends} $|\Sigma_i^\partial| \le 8 d_0 = 192$ for all $i \in \{2, \dots, \lfloor \ell \rfloor\} \setminus I_{\mathrm{corner}}$. \end{enumerate} \end{lemma} \begin{proof} Since $M \ge M_{\mathrm g}$, we have $\ell \ge 73$, which implies that $\lfloor \ell \rfloor > 2\ell/3$ and $\ell/3 > d_0$. For $2 \le i \le \ell$, we also have \begin{equation}\label{e:shell-kbounds} \ell \le \lfloor 2 \ell \rfloor \le k_i \le i \ell \le \ell^2 = M^{1/2} \le M . \end{equation} We begin with \ref{i:shellgeom-corner}. Let $2 \le i < i' \le \ell$. Using $\lfloor a + a' \rfloor \ge \lfloor a \rfloor + \lfloor a' \rfloor$ for real numbers $a$ and $a'$, we find that \begin{equation}\label{e:shell-separation} k_{i'} - k_i \ge \lfloor (i' - i) \ell \rfloor \ge \lfloor \ell \rfloor > \frac{2 \ell}{3} . \end{equation} By the triangle inequality, every site $v$ satisfies $\dist_\infty(v, \Sigma_i) \ge \bigl| |v - y|_\infty - k_i \bigr|$, and the same holds for $i'$. A site $v$ with $\dist_\infty(v, \Sigma_i) < \ell/3$ and $\dist_\infty(v, \Sigma_{i'}) < \ell/3$ would satisfy \begin{equation*} k_{i'} - k_i \le \bigl| |v - y|_\infty - k_i \bigr| + \bigl| |v - y|_\infty - k_{i'} \bigr| < \frac{2 \ell}{3} , \end{equation*} which contradicts \eqref{e:shell-separation}. We conclude that each of the four corner sites of $\mathcal C_M$ is at distance less than $\ell/3$ from at most one shell, and $|I_{\mathrm{corner}}| \le 4$. Next, we prove \ref{i:shellgeom-sides}. The symmetries of $\Z^2$ that preserve $\mathcal C_M$ map the shells, their sides, the corner sites, and $I_{\mathrm{corner}}$ to the corresponding objects for the image of $y$. Using these symmetries, we may assume that $y_2 = -M - 1$ and $|y_1| \le M + 1$. We write $k = k_i$ and consider the four sides in turn. The side with $j = 2$ and $t = y_2 - k$ does not meet $\mathcal C_M$. The side with $j = 2$ and $t = y_2 + k$ satisfies \begin{equation*} -M - 1 + \frac{\ell}{3} \le t = k - M - 1 \le -1 \end{equation*} by \eqref{e:shell-kbounds}, and $|t| \le M + 1 - \ell/3$ follows. Now consider the side with $j = 1$ and $t = y_1 - k$, and suppose that it meets $\mathcal C_M$. Since $|y_1| \le M + 1$, the bounds \eqref{e:shell-kbounds} imply that \begin{equation*} -M \le t \le M + 1 - k \le M + 1 - \frac{\ell}{3} . \end{equation*} The site $v' = (t, -M)$ belongs to this side and to $\mathcal C_M$, and $|v' - y|_\infty = \max(k, 1) = k$, so that $v' \in \Sigma_i$. Since $i \notin I_{\mathrm{corner}}$, the distance $|v' - (-M, -M)|_\infty = t + M$ from $v'$ to the corner site $(-M, -M)$ is at least $\ell/3$. We obtain $t \ge -M + \ell/3$, which together with the upper bound gives $|t| \le M + 1 - \ell/3$. The side with $j = 1$ and $t = y_1 + k$ is treated in the same way, with the corner site $(M, -M)$. The equivalent formulation follows from the fact that $S$ has $\ell^\infty$ distance $M + 1 - t$ from $\{x : x_j \ge M + 1\}$ and $M + 1 + t$ from $\{x : x_j \le -M - 1\}$. Finally, we prove \ref{i:shellgeom-ends}. Recall that $\Sigma_i$ is the part in $\mathcal C_M$ of the sphere of radius $k_i$ around $y$. Let $x \in \Sigma_i^\partial$. Then $x$ lies on a side $S$ of this sphere, with $j$ and $t$ as in \ref{i:shellgeom-sides}, and $S$ meets $\mathcal C_M$. Since $x \in \mathcal C_M$, we have \begin{equation*} \min \bigl( M + 1 - |x_1|, M + 1 - |x_2| \bigr) = M + 1 - |x|_\infty = \dist_\infty(x, \Z^2 \setminus \mathcal C_M) \le d_0 . \end{equation*} Part \ref{i:shellgeom-sides} implies that $M + 1 - |x_j| = M + 1 - |t| \ge \ell/3 > d_0$. The minimum is therefore attained at $3 - j$, which gives \begin{equation*} M + 1 - d_0 \le |x_{3 - j}| \le M . \end{equation*} Thus $x_{3 - j}$ takes at most $2 d_0$ values. Since a site of $S$ is determined by its coordinate $x_{3 - j}$, each of the four sides contains at most $2 d_0$ sites of $\Sigma_i^\partial$, and $|\Sigma_i^\partial| \le 8 d_0$. \end{proof} \section{The window estimate}\label{s:block-window} We now state the hypothesis under which we prove the gap of the block dynamics. Given a family of probability measures on a measurable space, a probability measure $\nu$ on the same space, and $p \in (0, 1]$, we say that $p \nu$ is a \emph{common component} of the family if every measure of the family is bounded below by $p \nu$. \begin{hypothesis}{W} There exist $p_0 \in (0, 1]$ and $M_\star \ge 1$ such that the following holds for all integers $M \ge M_\star$, all $y \in \Z^2$ with $\dist_\infty(y, \mathcal C_M) = 1$, and all integers $i$ with $2 \le i \le \ell$ and $i \notin I_{\mathrm{corner}}$. There exists a probability measure $\nu_{M, y, i}$ on $(\Sph)^{W_i}$ such that \begin{equation*} \mu^\eta_{F_i, W_i} \ge p_0 \, \nu_{M, y, i} \qquad \text{for every boundary configuration } \eta \text{ of } F_i . \end{equation*} \end{hypothesis} We call Hypothesis~(W) the \emph{window estimate}. It states that $p_0 \nu_{M, y, i}$ is a common component of the laws $\mu^\eta_{F_i, W_i}$ of the window, over all boundary configurations $\eta$ of $F_i$. Note that $\nu_{M, y, i}$ may depend on $M$, $y$, and $i$, but not on $\eta$. Assuming Hypothesis~(W), we set \begin{equation}\label{e:discrepancy} \delta = p_0 \, \varepsilon_{192} , \qquad J = \lfloor M^{1/8} \rfloor , \qquad \Delta_M = 9 M^{1/2} (J + 5)^2 + b^2 (1 - \delta)^J . \end{equation} \section{The coupling}\label{s:block-coupling} We now describe the coupling of two local measures on the box whose boundary configurations differ at one site. Until the end of the proof of \zcref{p:coupling}, we assume Hypothesis~(W), we fix $M \ge \max(M_\star, M_{\mathrm g})$ and a site $y \in \Z^2$ adjacent to $\mathcal C_M$, and we fix boundary configurations $\eta$ and $\eta'$ of $\mathcal C_M$ with $\eta_u = \eta'_u$ for all $u \neq y$. Since $\dist_\infty(y, \mathcal C_M) = 1$, the sets \eqref{e:shells} and $I_{\mathrm{corner}}$ are defined, and \zcref{l:shellgeom} applies. \paragraph{The shells used.} Since $M \ge M_{\mathrm g} \ge 3^8$, we have $M^{1/8} \ge 3$ and \begin{equation*} \ell - 6 = (M^{1/8})^2 - 6 \ge 3 M^{1/8} - 6 \ge M^{1/8} , \end{equation*} which implies that $J \le \lfloor \ell \rfloor - 6$. By \zcref{l:shellgeom}\ref{i:shellgeom-corner}, the set $\{2, \dots, \lfloor \ell \rfloor - 1\} \setminus I_{\mathrm{corner}}$ has at least $\lfloor \ell \rfloor - 6 \ge J$ elements. Let $i_1 < i_2 < \dots < i_J$ denote its $J$ smallest elements. For $k \le J$, the set $\{2, \dots, k + 5\}$ is contained in $\{2, \dots, \lfloor \ell \rfloor - 1\}$ and contains at least $k$ elements outside $I_{\mathrm{corner}}$, so that \begin{equation}\label{e:indexbound} i_k \le k + 5 , \qquad 1 \le k \le J . \end{equation} \paragraph{The trials.} We construct the coupling by a finite sequence of coupled resamplings of subsets of $\mathcal C_M$, with $A = \mathcal C_M$ and the boundary configurations $\eta$ and $\eta'$. We start from a pair $(n, n')$ of independent configurations with laws $\mu^\eta_{\mathcal C_M}$ and $\mu^{\eta'}_{\mathcal C_M}$, and we write $(n, n')$ for the current pair at every stage. In each resampling, $\mathcal G$ is the $\sigma$-algebra generated by all random variables constructed before it, and the conditional law of the new pair given $\mathcal G$ is a function of the current pair. All the kernels used below depend measurably on the current pair, since the densities of the local measures are jointly measurable in the configuration and the boundary configuration. In every coupled resampling of a set $D$ below, when the laws $\mu_D^{(n \eta)_{\Z^2 \setminus D}}$ and $\mu_D^{(n' \eta')_{\Z^2 \setminus D}}$ coincide, we sample identically, drawing $\hat n_D$ from their common law and setting $\hat n'_D = \hat n_D$. Recall from \zcref{s:quantum-rotated} that these two laws coincide when $(n \eta)_u = (n' \eta')_u$ for all sites $u \notin D$ adjacent to $D$. For $i = i_1, i_2, \dots, i_J$, in this order, we perform the following three coupled resamplings, which we call the \emph{trial} at $i$. \begin{enumerate}[label=(\alph*)] \item\label{i:trial-F} A coupled resampling of $F_i$. When the two laws coincide, we sample identically. Otherwise, we write $\xi = (n \eta)_{\Z^2 \setminus F_i}$ and $\xi' = (n' \eta')_{\Z^2 \setminus F_i}$, which are boundary configurations of $F_i$, and $\nu = \nu_{M, y, i}$. By Hypothesis~(W), which applies since $M \ge M_\star$ and $i \notin I_{\mathrm{corner}}$, we have $\mu^\xi_{F_i, W_i} \ge p_0 \nu$ and $\mu^{\xi'}_{F_i, W_i} \ge p_0 \nu$. With probability $p_0$, we draw $w$ from $\nu$ and set $w' = w$. With probability $1 - p_0$ (which is positive only if $p_0 < 1$), we draw $w$ and $w'$ independently from $(\mu^\xi_{F_i, W_i} - p_0 \nu)/(1 - p_0)$ and $(\mu^{\xi'}_{F_i, W_i} - p_0 \nu)/(1 - p_0)$. We then set $\hat n_{W_i} = w$ and $\hat n'_{W_i} = w'$, and we draw $\hat n_{F_i \setminus W_i}$ and $\hat n'_{F_i \setminus W_i}$ independently from $\mu_{F_i \setminus W_i}^{w \xi}$ and $\mu_{F_i \setminus W_i}^{w' \xi'}$. Here $w \xi$ is the boundary configuration of $F_i \setminus W_i$ that agrees with $w$ on $W_i$ and with $\xi$ off $F_i$. \item\label{i:trial-ends} A coupled resampling of $\Sigma_i^\partial$, by the maximal coupling of the two laws. \item\label{i:trial-next} A coupled resampling of $F_{i + 1}$. When the two laws coincide, we sample identically. Otherwise, we sample independently. \end{enumerate} In \ref{i:trial-F}, the window configuration $w$ has law $\mu^\xi_{F_i, W_i}$. By \zcref{l:resample} with $A = F_i$, $F = F_i \setminus W_i$, and the boundary configuration $\xi$, drawing $w$ from $\mu^\xi_{F_i, W_i}$ and then the configuration on $F_i \setminus W_i$ from $\mu_{F_i \setminus W_i}^{w \xi}$ produces a configuration on $F_i$ with law $\mu_{F_i}^\xi$. Hence $\hat n_{F_i}$ has law $\mu_{F_i}^\xi$, and likewise $\hat n'_{F_i}$ has law $\mu_{F_i}^{\xi'}$, so that \ref{i:trial-F} is a coupled resampling of $F_i$. We call the trial at $i$ \emph{successful} if the two configurations agree on $\Sigma_i$ after \ref{i:trial-ends}. By the discussion after \zcref{l:resample} and induction over the resamplings, the final pair $(n, n')$ is a coupling of $\mu^\eta_{\mathcal C_M}$ and $\mu^{\eta'}_{\mathcal C_M}$. The following proposition, which is the main step of this section, bounds its expected Hamming distance. \begin{proposition}\label{p:coupling} Assume Hypothesis~(W), and let $M \ge \max(M_\star, M_{\mathrm g})$. Let $y \in \Z^2 \setminus \mathcal C_M$, and let $\eta$ and $\eta'$ be boundary configurations of $\mathcal C_M$ such that $\eta_u = \eta'_u$ for all $u \neq y$. Then there exists a coupling $(n, n')$ of $\mu^\eta_{\mathcal C_M}$ and $\mu^{\eta'}_{\mathcal C_M}$ such that \begin{equation*} \E \, d_{\mathrm H}(n, n') \le \Delta_M . \end{equation*} Further, $\Delta_M / b \to 0$ as $M \to \infty$. \end{proposition} \begin{proof} If $y$ is not adjacent to $\mathcal C_M$, then $\mu^\eta_{\mathcal C_M} = \mu^{\eta'}_{\mathcal C_M}$ by \zcref{s:quantum-rotated}, and we take $n = n'$. From now on, we assume that $y$ is adjacent to $\mathcal C_M$, and we let $(n, n')$ be the final pair of the construction above. \emph{The success probability.} We claim that, conditionally on all random variables constructed before the trial at $i$, the trial at $i$ is successful with probability at least $\delta$. Recall that $\mu_A^\eta$ is determined by the spins of $\eta$ at the sites adjacent to $A$. Suppose first that the two laws in \ref{i:trial-F} coincide, so that the configurations agree on $F_i$ after \ref{i:trial-F}. Let $u \notin \Sigma_i^\partial$ be adjacent to a site $x \in \Sigma_i^\partial$. Since $\ell \ge 2$, we have \begin{equation*} |u - y|_\infty \ge k_i - 1 > i \ell - 2 \ge (i - 1) \ell , \end{equation*} and $u \neq y$, since $|x - y|_\infty = k_i \ge 2$. Thus $u$ belongs to $F_i$, where the configurations agree, or to $\Z^2 \setminus (\mathcal C_M \cup \{y\})$, where $\eta_u = \eta'_u$. It follows that the laws in \ref{i:trial-ends} coincide, \ref{i:trial-ends} samples identically, and the trial is successful. Suppose now that the two laws in \ref{i:trial-F} differ. Then \ref{i:trial-F} makes the configurations agree on $W_i$ with conditional probability at least $p_0$. Further, \zcref{l:shellgeom}\ref{i:shellgeom-ends} implies that $|\Sigma_i^\partial| \le 192$. Applying \zcref{l:overlap} with $K = 192$ and $A = \Sigma_i^\partial$, we find that \ref{i:trial-ends} makes the configurations agree on $\Sigma_i^\partial$ with conditional probability at least $\varepsilon_{192}$, given all random variables constructed before \ref{i:trial-ends}. The resampling \ref{i:trial-ends} leaves the configurations on $W_i$ unchanged, since $W_i \cap \Sigma_i^\partial = \emptyset$. Combining these two bounds, we conclude that the trial is successful with conditional probability at least $p_0 \varepsilon_{192} = \delta$, as claimed. Multiplying the conditional probabilities of failure over the $J$ trials, we obtain \begin{equation}\label{e:nosuccess} \Prob(\text{no trial is successful}) \le (1 - \delta)^J . \end{equation} \emph{Agreement after a success.} Suppose that the trial at $i$ is successful. We show that the two configurations agree on $\Sigma_i \cup F_{i + 1}$ from the end of this trial on. Let $u \notin D$ be adjacent to a set $D \subset F_{i + 1}$. We claim that \begin{equation}\label{e:boundaryD} u \in F_{i + 1} \cup \Sigma_i \qquad \text{or} \qquad u \in \Z^2 \setminus (\mathcal C_M \cup \{y\}) . \end{equation} Indeed, a site $u \in \mathcal C_M \setminus F_{i + 1}$ adjacent to a site of $F_{i + 1}$ satisfies $i \ell - 1 < |u - y|_\infty \le i \ell$, which is the defining condition of $\Sigma_i$. If $u \notin \mathcal C_M$, then $u \neq y$, since the sites of $F_{i + 1}$ are at $\ell^\infty$ distance more than $i \ell \ge 2$ from $y$. By \eqref{e:boundaryD}, if the configurations agree on $\Sigma_i \cup F_{i + 1}$ before a coupled resampling of $D$, then the two laws in it coincide, it samples identically, and the configurations agree on $\Sigma_i \cup F_{i + 1}$ after it. Before \ref{i:trial-next} at $i$, the configurations agree on $\Sigma_i$. By \eqref{e:boundaryD} with $D = F_{i + 1}$, every site $u \notin F_{i + 1}$ adjacent to $F_{i + 1}$ belongs to $\Sigma_i$ or to $\Z^2 \setminus (\mathcal C_M \cup \{y\})$. The two laws in \ref{i:trial-next} coincide, and after \ref{i:trial-next} the configurations agree on $\Sigma_i \cup F_{i + 1}$. Every later resampling is a coupled resampling of a set $D \in \{F_{i'}, \Sigma_{i'}^\partial, F_{i' + 1}\}$ with $i' > i$, and $D \subset F_{i'} \subset F_{i + 1}$. By the previous paragraph and induction over the later resamplings, the configurations agree on $\Sigma_i \cup F_{i + 1}$ until the end of the construction. \emph{The expected discrepancy.} Suppose that some trial is successful, and let $i$ be the smallest index at which a trial is successful. Then \eqref{e:indexbound} implies that $i \le i_J \le J + 5$. Since the final configurations agree on $F_{i + 1}$, they can differ only at sites $x \in \mathcal C_M$ with $|x - y|_\infty \le i \ell$. Using $i \le J + 5$ and $(J + 5) \ell \ge 1$, we see that the number of these sites is at most \begin{equation*} \bigl( 2 (J + 5) \ell + 1 \bigr)^2 \le 9 (J + 5)^2 \ell^2 = 9 M^{1/2} (J + 5)^2 . \end{equation*} On the event that no trial is successful, we use the bound $d_{\mathrm H}(n, n') \le |\mathcal C_M| = b^2$. Combining these bounds with \eqref{e:nosuccess}, we obtain $\E \, d_{\mathrm H}(n, n') \le \Delta_M$. \emph{The limit.} We have $J + 5 \le 6 M^{1/8}$ and $J \ge M^{1/8} - 1$. Using $b \le 3M$ and $1 - \delta \le e^{-\delta}$, we find that \begin{equation*} \frac{\Delta_M}{b} \le \frac{9 M^{1/2} \cdot 36 M^{1/4}}{2M} + b \, e^{-\delta (M^{1/8} - 1)} \le 162 M^{-1/4} + 3M e^{-\delta (M^{1/8} - 1)} . \end{equation*} Since $\delta = p_0 \varepsilon_{192} > 0$, the right side tends to zero as $M \to \infty$. \end{proof} \section{The gap of the block dynamics}\label{s:block-gap} We now return to the torus and prove the main result of this section. \begin{proposition}\label{p:blockgap} Assume Hypothesis~(W). Let $M \ge \max(M_\star, M_{\mathrm g})$ be an integer, set $b = 2M + 1$, and let $L, N \ge b + 4$. Then the generator $\Lbl$ of \eqref{e:blockgen} satisfies \begin{equation}\label{e:blockgap} \gap(\Lbl) \ge 1 - \frac{4 \Delta_M}{b} . \end{equation} In particular, there exists $M_{\mathrm{bl}} \ge \max(M_\star, M_{\mathrm g})$, depending only on $p_0$ and $M_\star$, such that $\gap(\Lbl) \ge 1/2$ for all integers $M \ge M_{\mathrm{bl}}$ and all integers $L, N \ge 2M + 5$. \end{proposition} \begin{proof} The second claim follows from the first, since $\Delta_M$ is determined by $p_0$ and $M$, and $\Delta_M / b \to 0$ as $M \to \infty$ by \zcref{p:coupling}. To prove \eqref{e:blockgap}, we write $\mu = \mu_{L, N}$, and we set \begin{equation*} \mathcal K = \frac{1}{|\Lambda|} \sum_{x \in \Lambda} \mathcal Q_{B_x} , \qquad \gamma_{\mathrm{bl}} = 1 - \frac{b^2 - 4 b \Delta_M}{|\Lambda|} . \end{equation*} The operator $\mathcal K$ resamples a uniformly chosen block, and \eqref{e:blockgen} can be written as \begin{equation}\label{e:LK} \Lbl = \frac{|\Lambda|}{b^2} (\mathcal K - \Id) . \end{equation} Recall from \zcref{s:quantum-block} that $\mathcal Q_{B_x}(s, \cdot)$ is defined for every configuration $s \in (\Sph)^\Lambda$. Therefore $\mathcal K f$ is a bounded measurable function, defined at every point, for every bounded measurable function $f$ on $(\Sph)^\Lambda$. For such $f$, we set \begin{equation*} \Lip(f) = \sup_{s \neq s'} \frac{|f(s) - f(s')|}{d_{\mathrm H}(s, s')} \le 2 \| f \|_\infty , \end{equation*} where the supremum in $\Lip(f)$ is over all pairs of distinct configurations $s, s' \in (\Sph)^\Lambda$ and $\| f \|_\infty = \sup_{s \in (\Sph)^\Lambda} |f(s)|$ is the pointwise supremum. \emph{Contraction of the Lipschitz constant.} We claim that \begin{equation}\label{e:Lipcontract} \Lip(\mathcal K f) \le \gamma_{\mathrm{bl}} \Lip(f) \end{equation} for every bounded measurable function $f$ on $(\Sph)^\Lambda$. Two configurations at Hamming distance $k$ are joined by a sequence of $k + 1$ configurations in which consecutive configurations differ at exactly one site. It therefore suffices to show that $|\mathcal K f(s) - \mathcal K f(s')| \le \gamma_{\mathrm{bl}} \Lip(f)$ when $s$ and $s'$ differ only at one site $z$. Fix such $s$ and $s'$, and fix $x \in \Lambda$. For every coupling $(\xi, \xi')$ of $\mathcal Q_{B_x}(s, \cdot)$ and $\mathcal Q_{B_x}(s', \cdot)$, we have \begin{equation}\label{e:onecoupling} \bigl| \mathcal Q_{B_x} f(s) - \mathcal Q_{B_x} f(s') \bigr| = \bigl| \E [f(\xi) - f(\xi')] \bigr| \le \Lip(f) \, \E \, d_{\mathrm H}(\xi, \xi') . \end{equation} By \eqref{e:QR}, the law of the spins in $B_x$ under $\mathcal Q_{B_x}(s, \cdot)$ is determined by the spins of $s$ at the sites adjacent to $B_x$. We choose the coupling according to the position of $z$, in three cases. \begin{itemize} \item When $z \in B_x$, we have $\mathcal Q_{B_x}(s, \cdot) = \mathcal Q_{B_x}(s', \cdot)$, and the identical coupling gives $d_{\mathrm H}(\xi, \xi') = 0$. Since $L, N \ge b$, this case occurs for exactly $b^2$ sites $x$. \item When $z \notin B_x$ and $z$ is not adjacent to $B_x$, the laws of the block spins under $\mathcal Q_{B_x}(s, \cdot)$ and $\mathcal Q_{B_x}(s', \cdot)$ coincide. The identical coupling of the block spins gives $d_{\mathrm H}(\xi, \xi') = 1$, since $\xi_z = s_z$ and $\xi'_z = s'_z$. \item When $z \notin B_x$ and $z$ is adjacent to $B_x$, we use \zcref{p:coupling}, as described in the next paragraph, to obtain a coupling with $\E \, d_{\mathrm H}(\xi, \xi') \le \Delta_M + 1$. In this case, $z = x + u$ for a site $u$ of the patch outside $\{1, \dots, b\}^2$, and $x = z - u$ is determined by $u$. Since there are $4b$ such sites $u$, this case occurs for at most $4b$ sites $x$. \end{itemize} We now construct the coupling in the third case. We apply \zcref{l:rotated} with $c = (-M - 1, -M - 1)$, for which $T = c + \{1, \dots, b\}^2 = \mathcal C_M$. Let $\eta$ and $\eta'$ be the boundary configurations of $\mathcal C_M$ such that \begin{equation*} \eta_u = (-1)^{u_1 + u_2} s_{x + u - c} , \qquad \eta'_u = (-1)^{u_1 + u_2} s'_{x + u - c} \end{equation*} for all $u \notin \mathcal C_M$ adjacent to $\mathcal C_M$, and such that $\eta_u = \eta'_u$ is the same arbitrary element of $\Sph$ at all other sites $u \notin \mathcal C_M$. Since the map $u \mapsto x + u - c$ is injective on the sites at distance at most $1$ from $\mathcal C_M$, the configurations $\eta$ and $\eta'$ differ at most at one site $y \notin \mathcal C_M$, namely the site adjacent to $\mathcal C_M$ with $x + y - c = z$. By \zcref{p:coupling}, there exists a coupling $(n, n')$ of $\mu^\eta_{\mathcal C_M}$ and $\mu^{\eta'}_{\mathcal C_M}$ with $\E \, d_{\mathrm H}(n, n') \le \Delta_M$. We define $\xi$ by \begin{equation*} \xi_{x + u - c} = (-1)^{u_1 + u_2} n_u \quad \text{for } u \in \mathcal C_M , \qquad \xi_v = s_v \quad \text{for } v \notin B_x , \end{equation*} and we define $\xi'$ in the same way from $n'$ and $s'$. The map that sends a configuration $\tilde s$ equal to $s$ off $B_x$ to $\bigl( (-1)^{u_1 + u_2} \tilde s_{x + u - c} \bigr)_{u \in \mathcal C_M}$ is a bijection onto $(\Sph)^{\mathcal C_M}$, and $\xi$ is the preimage of $n$ under this map. Therefore \zcref{l:rotated} shows that $\xi$ has law $\mathcal Q_{B_x}(s, \cdot)$, and likewise $\xi'$ has law $\mathcal Q_{B_x}(s', \cdot)$. Further, $\xi_v = \xi'_v$ for $v \in B_x$ if and only if $n_u = n'_u$ for the corresponding $u \in \mathcal C_M$, and off $B_x$ the configurations $\xi$ and $\xi'$ differ only at $z$. We conclude that \begin{equation*} \E \, d_{\mathrm H}(\xi, \xi') = \E \, d_{\mathrm H}(n, n') + 1 \le \Delta_M + 1 . \end{equation*} We now average \eqref{e:onecoupling} over $x$. Let $\mathcal X_z$ denote the set of sites $x$ in the third case, so that $|\mathcal X_z| \le 4b$, and the second case occurs for $|\Lambda| - b^2 - |\mathcal X_z|$ sites $x$. Using the three cases, we find that \begin{align*} |\mathcal K f(s) - \mathcal K f(s')| &\le \frac{\Lip(f)}{|\Lambda|} \Bigl( |\mathcal X_z| (\Delta_M + 1) + |\Lambda| - b^2 - |\mathcal X_z| \Bigr) \\ &\le \Lip(f) \Bigl( 1 - \frac{b^2 - 4 b \Delta_M}{|\Lambda|} \Bigr) = \gamma_{\mathrm{bl}} \Lip(f) , \end{align*} which proves \eqref{e:Lipcontract}. (Each coupling above is used for one fixed pair $(s, s')$, so no measurable dependence of the couplings on the pair is needed.) \emph{From contraction to the spectrum.} For a finite state space, the passage from \eqref{e:Lipcontract} to a lower bound on the spectral gap is a standard argument (see \cite[Theorem~13.1]{LPW17}), which applies \eqref{e:Lipcontract} to an eigenfunction $f$ of $\mathcal K$ with eigenvalue $\lambda$ and obtains $|\lambda| \Lip(f) \le \gamma_{\mathrm{bl}} \Lip(f)$. (A version for reversible chains on general metric spaces is given in \cite{Ollivier09}.) Here the state space is continuous, $\mathcal K$ need not have eigenfunctions, and an element of $L^2(\mu)$ has no Lipschitz constant. We therefore apply \eqref{e:Lipcontract} to the iterates of bounded measurable functions, which are defined at every point because the version of $\mathcal Q_{B_x}$ from \zcref{s:quantum-block} is defined for every exterior configuration, and we pass to $L^2(\mu)$ through the spectral theorem. We may assume that $\gamma_{\mathrm{bl}} < 1$, since otherwise the right side of \eqref{e:blockgap} is nonpositive and \eqref{e:blockgap} holds trivially. Let $f$ be a bounded measurable function with $\mu(f) = 0$, and let $t \ge 0$ be an integer. Recall that each $\mathcal Q_{B_x} g$ is a version of a conditional expectation under $\mu$. In particular, $\mu(\mathcal Q_{B_x} g) = \mu(g)$ for every bounded measurable $g$, and $\mu(\mathcal K^t f) = \mu(f) = 0$. Iterating \eqref{e:Lipcontract} gives $\Lip(\mathcal K^t f) \le 2 \gamma_{\mathrm{bl}}^t \| f \|_\infty$. Since $d_{\mathrm H}(s, s') \le |\Lambda|$, we obtain, for every $s \in (\Sph)^\Lambda$, \begin{equation}\label{e:iterate} |\mathcal K^t f(s)| = \Bigl| \int \bigl( \mathcal K^t f(s) - \mathcal K^t f(s') \bigr) \, \mu(\dd s') \Bigr| \le |\Lambda| \Lip(\mathcal K^t f) \le 2 |\Lambda| \gamma_{\mathrm{bl}}^t \| f \|_\infty . \end{equation} We now pass to $L^2(\mu)$. The function $\mathcal K^t f$, defined pointwise, represents the $t$-th power of the operator $\mathcal K$ on $L^2(\mu)$ applied to $f$, since each $\mathcal Q_{B_x}$ maps functions that agree almost everywhere to functions that agree almost everywhere. Each $\mathcal Q_{B_x}$ is an orthogonal projection on $L^2(\mu)$, so $\mathcal K$ is self-adjoint and $0 \le \mathcal K \le \Id$. For $g \in L^2(\mu)$, let $m_g$ denote the spectral measure of $g$ with respect to $\mathcal K$, which is the finite measure on $[0, 1]$ such that \begin{equation*} \langle g, h(\mathcal K) g \rangle_{L^2(\mu)} = \int h \dd m_g \qquad \text{for all bounded Borel functions } h . \end{equation*} Let $\gamma' \in (\gamma_{\mathrm{bl}}, 1]$. Using \eqref{e:iterate}, we find that \begin{equation*} (\gamma')^{2t} \, m_f \bigl( (\gamma', 1] \bigr) \le \int \lambda^{2t} \, m_f(\dd \lambda) = \| \mathcal K^t f \|^2_{L^2(\mu)} \le 4 |\Lambda|^2 \gamma_{\mathrm{bl}}^{2t} \| f \|_\infty^2 . \end{equation*} Letting $t \to \infty$, we obtain $m_f((\gamma', 1]) = 0$, and letting $\gamma' \downarrow \gamma_{\mathrm{bl}}$, we obtain $m_f((\gamma_{\mathrm{bl}}, 1]) = 0$. Let $E = \1_{(\gamma_{\mathrm{bl}}, 1]}(\mathcal K)$ denote the spectral projection of $\mathcal K$ on $(\gamma_{\mathrm{bl}}, 1]$, and let $L^2_0(\mu)$ denote the space of functions $g \in L^2(\mu)$ with $\mu(g) = 0$. The previous paragraph shows that \begin{equation*} \| E f \|^2_{L^2(\mu)} = m_f \bigl( (\gamma_{\mathrm{bl}}, 1] \bigr) = 0 \end{equation*} for every bounded measurable $f$ with $\mu(f) = 0$. The bounded measurable functions with mean zero are dense in $L^2_0(\mu)$, as one sees by truncating and recentering, and $E$ is bounded. Therefore $E g = 0$ for all $g \in L^2_0(\mu)$. Equivalently, the spectral measure $m_g$ of every $g \in L^2_0(\mu)$ is supported on $[0, \gamma_{\mathrm{bl}}]$, which gives \begin{equation}\label{e:Kcontract} \langle g, \mathcal K g \rangle_{L^2(\mu)} = \int \lambda \, m_g(\dd \lambda) \le \gamma_{\mathrm{bl}} \| g \|^2_{L^2(\mu)} , \qquad g \in L^2_0(\mu) . \end{equation} \emph{Conclusion.} Let $F \in L^2(\mu)$ be nonconstant, and set $g = F - \mu(F)$. Since $\Lbl$ is self-adjoint and annihilates the constant functions, \eqref{e:LK} and \eqref{e:Kcontract} imply that \begin{align*} \langle F, -\Lbl F \rangle_{L^2(\mu)} &= \frac{|\Lambda|}{b^2} \Bigl( \| g \|^2_{L^2(\mu)} - \langle g, \mathcal K g \rangle_{L^2(\mu)} \Bigr) \\ &\ge \frac{|\Lambda|}{b^2} (1 - \gamma_{\mathrm{bl}}) \Var_\mu(F) = \Bigl( 1 - \frac{4 \Delta_M}{b} \Bigr) \Var_\mu(F) . \end{align*} Taking the infimum over $F$ proves \eqref{e:blockgap}. \end{proof}