\chapter{The cut lemma and the window estimate}\label{s:cut} In this section, we prove \zcref{t:window}, which states that Hypothesis~(W) of \zcref{s:block} follows from Hypothesis~(P) below. Hypothesis~(P) is an exponential bound on the effect of boundary conditions on certain pair correlations in a continuous model with pins. Together with \zcref{p:blockgap,t:comparison}, \zcref{t:window} reduces \zcref{t:main} to Hypothesis~(P), which we deduce in \zcref{s:walks,s:contraction} from the inputs verified in \zcref{s:computer}. The heuristics behind the argument are explained in \zcref{s:overview}. The proof has two parts. The first, and the main idea of this section, is an exact identity, \zcref{l:cut}. Suppose that a set of sites $X$ separates the window $W_i$ from the complement of $F_i$. Then the density of the law of the spins in $W_i$ is, up to a normalizing constant, the product of a density that is the same for all boundary configurations and the exponential of the change in the logarithm of a ratio of partition functions caused by removing $X$. In the second part, we bound this change. We remove the sites of $X$ one at a time, in an order in which every site after the first has a neighbor that was removed before it. Each of these removals concerns a site with at most three remaining neighbors, and we bound its cost by the differences of pair correlations that Hypothesis~(P) controls. In \zcref{s:cut-geometry}, we construct for each shell a set $X$ along which these bounds have a sum that is bounded uniformly in $M$. \section{The model with pins}\label{s:cut-model} Let $G \subset \Z^2$ be finite. We call a pair $(x, z)$ of sites with $x \in G$, $z \notin G$, and $x \sim z$ a \emph{leg} of $G$ at $x$. A \emph{pin configuration} on $G$ is a family $\tau = (\tau_{xz})$ of vectors $\tau_{xz} \in \Sph$ indexed by the legs $(x, z)$ of $G$, and we call $\tau_{xz}$ the \emph{pin} on the leg $(x, z)$. Let $Y \subset G$ and set $D = G \setminus Y$. We call $D$ a \emph{domain} in $G$, and we call the sites of $Y$ \emph{removed}. A removed site carries no spin, and the integrands of the partition functions below are products of one factor for each bond $\{x, x'\} \subset D$ and one factor for each leg of $G$ at a site of $D$. In particular, a removed site does not become a pin. A bond $\{x, z\}$ with $x \in D$ and $z \in Y$ is not a leg of $G$ and contributes no factor. \zcref[S]{d:domain} in \zcref{s:walks} extends the notion of a domain to sets in which some sites have been visited once by a walk. For a domain $D$ in $G$ and a pin configuration $\tau$ on $G$, we define the \emph{pinned partition function} $Z^\tau_D$ and the \emph{free partition function} $Z^0_D$ by \begin{equation}\label{e:Zpin} \begin{aligned} Z^\tau_D &= \int \prod_{\substack{\{x, x'\} \subset D \\ x \sim x'}} (1 + n_x \cdot n_{x'}) \prod_{\substack{(x, z) \text{ leg of } G \\ x \in D}} (1 + \tau_{xz} \cdot n_x) \prod_{x \in D} \dd n_x, \\ Z^0_D &= \int \prod_{\substack{\{x, x'\} \subset D \\ x \sim x'}} (1 + n_x \cdot n_{x'}) \prod_{x \in D} \dd n_x, \end{aligned} \end{equation} with $Z^\tau_\varnothing = Z^0_\varnothing = 1$. The integrands are nonnegative and bounded, and they are positive outside a set of measure zero, since $1 + a \cdot b > 0$ for all $a, b \in \Sph$ with $b \ne -a$. In particular, all these partition functions are positive and finite. We let $\E^\tau_D$ and $\E^0_D$ denote the probability measures on $(\Sph)^D$ whose densities with respect to $\prod_{x \in D} \dd n_x$ are the integrands in \eqref{e:Zpin} divided by $Z^\tau_D$ and $Z^0_D$, respectively, and we use the same symbols for the corresponding expectations. We write $Z^\cdot_D$ and $\E^\cdot_D$ in statements that hold with $\cdot = \tau$ and with $\cdot = 0$. We also define \begin{equation*} \Xi_D(\tau) = \frac{Z^\tau_D}{Z^0_D}. \end{equation*} These quantities depend on $G$ through its legs (we suppress $G$ in the notation, since it will always be clear from the context). The local measures of \zcref{s:quantum-rotated} are special cases of these measures. Indeed, for a finite set $A \subset \Z^2$ and $\eta \in (\Sph)^{\Z^2 \setminus A}$, let $\tau$ be the pin configuration on $A$ with $\tau_{xz} = \eta_z$ for every leg $(x, z)$ of $A$. Comparing \eqref{e:Zpin} for the domain $A$ in $A$ with \eqref{e:muA}, we see that $\mu^\eta_A = \E^\tau_A$ and $Z^\eta_A = Z^\tau_A$. We will use the moments of the uniform measure on $\Sph$. Let $n^\alpha$ denote the $\alpha$-th coordinate of $n \in \Sph$. For $m \ge 1$ and $\alpha_1, \dots, \alpha_{2m} \in \{1, 2, 3\}$, we have \begin{equation}\label{e:moments} \int_{\Sph} n^{\alpha_1} \cdots n^{\alpha_{2m}} \dd n = \frac{1}{3 \cdot 5 \cdots (2m + 1)} \sum_{\pi} \prod_{\{r, r'\} \in \pi} \delta_{\alpha_r \alpha_{r'}}, \end{equation} where the sum runs over the partitions $\pi$ of $\{1, \dots, 2m\}$ into pairs. The moments of odd order vanish, since $\dd n$ is invariant under $n \mapsto -n$. The formula \eqref{e:moments} is standard. It follows from Wick's formula for a standard Gaussian vector $g$ in $\R^3$, since $g / |g|$ has law $\dd n$ and is independent of $|g|$, and $\E |g|^{2m} = 3 \cdot 5 \cdots (2m + 1)$. In particular, we have $\int (a \cdot n)(b \cdot n) \dd n = (a \cdot b)/3$ for all $a, b \in \R^3$. We will need two further elementary properties of the measures $\E^\cdot_D$. \begin{lemma}\label{l:elementary} Let $D$ be a domain in $G$, and let $\tau$ be a pin configuration on $G$. \begin{enumerate} \item Let $f$ be a finite product of scalar products $n_x \cdot n_{x'}$ with $x, x' \in D$. Then $\E^0_D[f] \ge 0$. \item Let $u \in D$. Under $\E^\cdot_D$, the conditional law of $n_u$ given $(n_x)_{x \in D \setminus \{u\}}$ has a density with respect to $\dd n_u$ that is proportional to the product of the factors of the integrand of $Z^\cdot_D$ in \eqref{e:Zpin} that depend on $n_u$. If $u, v \in D$ are distinct and not neighbors, then $n_u$ and $n_v$ are conditionally independent under $\E^\cdot_D$ given $(n_x)_{x \in D \setminus \{u, v\}}$. \end{enumerate} \end{lemma} \begin{proof} We begin with the first part. We expand the product of the bond factors in \eqref{e:Zpin}, and we write each scalar product as a sum over the three coordinates. This expresses $Z^0_D \E^0_D[f]$ as a finite sum of integrals of the form \begin{equation*} \int \prod_{q = 1}^{m} n^{\alpha_q}_{x_q} n^{\alpha_q}_{x'_q} \prod_{x \in D} \dd n_x = \prod_{x \in D} \int \prod_{q \colon x_q = x} n^{\alpha_q} \prod_{q \colon x'_q = x} n^{\alpha_q} \dd n, \end{equation*} where $m \ge 0$, $x_q, x'_q \in D$, and $\alpha_q \in \{1, 2, 3\}$. By \eqref{e:moments}, every factor on the right side is nonnegative. It follows that $\E^0_D[f] \ge 0$, since $Z^0_D > 0$. For the second part, recall that the density of $\E^\cdot_D$ is a product of factors, each of which depends on the spins of two neighbors or on the spin of a single site. Under the conditioning, the factors that do not depend on $n_u$ are constant, which gives the first claim; the normalizing integral is positive because the integrand is positive almost everywhere. Now let $u$ and $v$ be distinct sites that are not neighbors. Then every factor depends on at most one of $n_u$ and $n_v$, and the conditional density of $(n_u, n_v)$ given the other spins is the product of a function of $n_u$ and a function of $n_v$. \end{proof} The first part of \zcref{l:elementary} is in the spirit of the positivity of staggered correlations used for bipartite tori in \cite{KLT88}. Under $\E^\tau_D$, two isolated sites with one pin each pointing in opposite directions have a negative pair correlation. We will also need the following bound on the conditional mean of a spin with at most three factors, which will be used in \zcref{s:walks}. \begin{lemma}\label{l:threefifths} \begin{enumerate} \item Let $v_1, v_2, v_3 \in \Ball$. Then the probability measure on $\Sph$ with density proportional to $\prod_{r = 1}^3 (1 + n \cdot v_r)$ with respect to $\dd n$ has mean of norm at most $3/5$. \item Let $D$ be a domain in $G$, let $\tau$ be a pin configuration on $G$, and let $u \in D$. Suppose that at most three factors of the integrand of $Z^\tau_D$ in \eqref{e:Zpin} depend on $n_u$. Then the conditional expectation of $n_u$ under $\E^\tau_D$ given $(n_x)_{x \in D \setminus \{u\}}$ has norm at most $3/5$. If $u$ has at most three neighbors in $D$, then the same holds under $\E^0_D$. \end{enumerate} \end{lemma} \begin{proof} We first prove the first part when $v_1 = a$, $v_2 = b$, and $v_3 = c$ are unit vectors. We set \begin{equation*} \gamma_1 = b \cdot c, \qquad \gamma_2 = a \cdot c, \qquad \gamma_3 = a \cdot b, \qquad \bar\gamma = \gamma_1 + \gamma_2 + \gamma_3, \end{equation*} and we expand the product $(1 + n \cdot a)(1 + n \cdot b)(1 + n \cdot c)$. By \eqref{e:moments}, the terms of odd degree in $n$ have integral zero, and we have \begin{equation*} \int n \, (n \cdot a) \dd n = \frac{a}{3}, \qquad \int n \, (n \cdot a)(n \cdot b)(n \cdot c) \dd n = \frac{1}{15} \bigl( a \, (b \cdot c) + b \, (a \cdot c) + c \, (a \cdot b) \bigr). \end{equation*} Integrating term by term, we obtain \begin{equation*} \int (1 + n \cdot a)(1 + n \cdot b)(1 + n \cdot c) \dd n = 1 + \frac{\bar\gamma}{3} \end{equation*} and \begin{equation*} 15 \int n \, (1 + n \cdot a)(1 + n \cdot b)(1 + n \cdot c) \dd n = \mathbf q, \qquad \mathbf q = (5 + \gamma_1) a + (5 + \gamma_2) b + (5 + \gamma_3) c. \end{equation*} Dividing the second identity by $15$ times the first, we find that the measure in the first part has mean $\mathbf q / (5 (3 + \bar\gamma))$. Observe that \begin{equation*} 3 + 2 \bar\gamma = |a + b + c|^2 \ge 0, \qquad 3 + \bar\gamma \ge \frac32. \end{equation*} It therefore suffices to show that $|\mathbf q| \le 3 (3 + \bar\gamma)$. Since $a$, $b$, and $c$ are unit vectors, we have \begin{equation*} |\mathbf q|^2 = \sum_{r = 1}^3 (5 + \gamma_r)^2 + 2 (5 + \gamma_1)(5 + \gamma_2) \gamma_3 + 2 (5 + \gamma_1)(5 + \gamma_3) \gamma_2 + 2 (5 + \gamma_2)(5 + \gamma_3) \gamma_1. \end{equation*} Expanding the two sides, one can check the identity \begin{equation}\label{e:threefifthsidentity} 9 (3 + \bar\gamma)^2 - |\mathbf q|^2 = 4 \bigl[ (\gamma_1 - \gamma_2)^2 + (\gamma_1 - \gamma_3)^2 + (\gamma_2 - \gamma_3)^2 \bigr] + 6 (1 - \gamma_1)(1 - \gamma_2)(1 - \gamma_3). \end{equation} The right side of \eqref{e:threefifthsidentity} is nonnegative, since $\gamma_r \le 1$ for all $r$, which completes the proof of the first part for unit vectors. Next, let $v_1, v_2, v_3 \in \Ball$. For $r = 1, 2, 3$, we choose $e_r \in \Sph$ such that $v_r = |v_r| e_r$ (with $e_r$ arbitrary if $v_r = 0$). Then \begin{equation*} 1 + n \cdot v_r = \frac{1 + |v_r|}{2} (1 + n \cdot e_r) + \frac{1 - |v_r|}{2} (1 - n \cdot e_r). \end{equation*} Expanding the product of these three sums, we write the measure $\prod_{r = 1}^3 (1 + n \cdot v_r) \dd n$ as a combination with nonnegative coefficients of the eight measures $\prod_{r = 1}^3 (1 + \epsilon_r n \cdot e_r) \dd n$ with $\epsilon_r \in \{1, -1\}$. After normalization, the probability measure in the first part is a convex combination of the normalizations of these eight measures, whose means have norm at most $3/5$ by the case of unit vectors. Its mean is the corresponding convex combination of these means, and the first part follows. For the second part, \zcref{l:elementary} shows that the conditional law of $n_u$ given the other spins has a density proportional to the product of at most three factors $1 + n_u \cdot w$, where each $w$ is the spin of a neighbor of $u$ in $D$ or a pin at $u$. We add factors $1 = 1 + n_u \cdot 0$ until there are three factors, and we apply the first part. Under $\E^0_D$, the factors that depend on $n_u$ are those of the bonds from $u$ to its neighbors in $D$, and the same argument applies. \end{proof} \section{Deleting a site}\label{s:cut-delete} For a domain $D$ in $G$ and a site $x \in D$, the set $D \setminus \{x\}$ is again a domain in $G$, and we say that it is obtained from $D$ by \emph{deleting} the site $x$. The following lemma computes the effect of a deletion on the partition functions when the deleted site has at most three neighbors in $D$. \begin{lemma}\label{l:delete} Let $D$ be a domain in $G$, and let $x \in D$ be a site whose four neighbors all belong to $G$. Suppose that $x$ has $k \le 3$ neighbors $u_1, \dots, u_k$ in $D$, and set \begin{equation}\label{e:Ix} I_x = 1 + \frac13 \sum_{1 \le a < a' \le k} n_{u_a} \cdot n_{u_{a'}}. \end{equation} Then $I_x \ge 1/2$. Further, for every pin configuration $\tau$ on $G$, we have \begin{equation}\label{e:deleteZ} Z^\cdot_D = Z^\cdot_{D \setminus \{x\}} \, \E^\cdot_{D \setminus \{x\}} [I_x] \end{equation} and \begin{equation}\label{e:deletebound} \begin{aligned} \bigl| \log \Xi_D(\tau) - \log \Xi_{D \setminus \{x\}}(\tau) \bigr| &= \bigl| \log \E^\tau_{D \setminus \{x\}}[I_x] - \log \E^0_{D \setminus \{x\}}[I_x] \bigr| \\ &\le \frac23 \sum_{1 \le a < a' \le k} \bigl| \E^\tau_{D \setminus \{x\}}[n_{u_a} \cdot n_{u_{a'}}] - \E^0_{D \setminus \{x\}}[n_{u_a} \cdot n_{u_{a'}}] \bigr|. \end{aligned} \end{equation} \end{lemma} \begin{proof} Since all neighbors of $x$ belong to $G$, the set $G$ has no leg at $x$. The spin $n_x$ therefore appears in the integrand of $Z^\cdot_D$ only through the $k$ factors $1 + n_x \cdot n_{u_a}$. By \eqref{e:moments}, we have \begin{equation*} \int \prod_{a = 1}^k (1 + n \cdot n_{u_a}) \dd n = 1 + \sum_{1 \le a < a' \le k} \int (n \cdot n_{u_a})(n \cdot n_{u_{a'}}) \dd n = I_x, \end{equation*} where we used that the terms of degrees one and three have integral zero, and that there is no term of degree four, since $k \le 3$. The remaining factors of the integrand of $Z^\cdot_D$ are exactly those of $Z^\cdot_{D \setminus \{x\}}$, since the deleted site $x$ does not become a pin. Integrating out $n_x$ first, we arrive at \eqref{e:deleteZ}. Next, using $|n_{u_a}| = 1$ for all $a$, we have \begin{equation*} \sum_{1 \le a < a' \le k} n_{u_a} \cdot n_{u_{a'}} = \frac12 \Bigl( \Bigl| \sum_{a = 1}^k n_{u_a} \Bigr|^2 - k \Bigr) \ge -\frac{k}{2}, \qquad I_x \ge 1 - \frac{k}{6} \ge \frac12. \end{equation*} The equality in \eqref{e:deletebound} follows from dividing \eqref{e:deleteZ} with $\cdot = \tau$ by \eqref{e:deleteZ} with $\cdot = 0$ and taking logarithms. For the inequality, note that by \eqref{e:Ix}, we have \begin{equation*} \E^\tau_{D \setminus \{x\}}[I_x] - \E^0_{D \setminus \{x\}}[I_x] = \frac13 \sum_{1 \le a < a' \le k} \bigl( \E^\tau_{D \setminus \{x\}}[n_{u_a} \cdot n_{u_{a'}}] - \E^0_{D \setminus \{x\}}[n_{u_a} \cdot n_{u_{a'}}] \bigr). \end{equation*} Since $I_x \ge 1/2$, these expectations belong to $[1/2, \infty)$, where the logarithm is Lipschitz with constant $2$. The inequality in \eqref{e:deletebound} follows. \end{proof} The differences on the right side of \eqref{e:deletebound} are the central objects of the rest of the paper. For a domain $D$ in $G$, a pin configuration $\tau$ on $G$, and sites $u, v \in D$, we call \begin{equation*} \E^\tau_D [n_u \cdot n_v] - \E^0_D [n_u \cdot n_v] \end{equation*} the \emph{response} of the pair $(u, v)$ in $D$. In this language, \zcref{l:delete} states that the pins enter the cost of deleting a site $x$ with $k \le 3$ neighbors in $D$ through the responses of the pairs of neighbors of $x$, taken in the domain $D \setminus \{x\}$. These pairs are at distance $2$, and the deleted site is a removed common neighbor of each of them. Hypothesis~(P) in \zcref{s:cut-removal} bounds responses of this kind. In this section, they bound the cost of removing a set that separates the window from the boundary. In \zcref{s:walks,s:contraction}, we prove Hypothesis~(P) by expanding each response into further responses of the same kind. \section{The cut lemma}\label{s:cut-cutlemma} Throughout this subsection, let $F \subset \Z^2$ be finite, let $W \subset F$ be a set such that every neighbor of a site of $W$ belongs to $F$, and set $G = F \setminus W$. In the application in \zcref{s:cut-window}, we have $F = F_i$ and $W = W_i$, and we call $W$ the window, as in \zcref{s:block}. Recall that $\mu^\eta_{F, W}$ denotes the marginal of $\mu^\eta_F$ on $W$. Every leg $(x, z)$ of $G$ satisfies $z \in W$ or $z \in \Z^2 \setminus F$. For $\eta \in (\Sph)^{\Z^2 \setminus F}$ and $\xi \in (\Sph)^W$, we let $\eta\xi$ denote the pin configuration on $G$ given by \begin{equation*} (\eta\xi)_{xz} = \xi_z \quad \text{if } z \in W, \qquad (\eta\xi)_{xz} = \eta_z \quad \text{if } z \notin F. \end{equation*} We define the internal weight of the window by \begin{equation*} \Theta_W(\xi) = \prod_{\substack{\{w, w'\} \subset W \\ w \sim w'}} (1 + \xi_w \cdot \xi_{w'}), \end{equation*} and, for $X \subset G$, we define the \emph{cut cost} \begin{equation}\label{e:cutcost} \mathrm{cost}_X(\eta, \xi) = \log \Xi_{G \setminus X}(\eta\xi) - \log \Xi_G(\eta\xi), \end{equation} where $G \setminus X$ and $G$ are domains in $G$. We call a set $X \subset G$ a \emph{cut} if $G \setminus X$ is the disjoint union of two sets $G_1$ and $G_2$ such that no bond joins $G_1$ to $G_2$, every neighbor of a site of $G_1$ belongs to $F$, and every neighbor of a site of $G_2$ belongs to $\Z^2 \setminus W$. \begin{lemma}\label{l:cut} Let $X \subset G$ be a cut, and suppose that there exists $C \ge 0$ such that $|\mathrm{cost}_X(\eta, \xi)| \le C$ for all $\eta \in (\Sph)^{\Z^2 \setminus F}$ and $\xi \in (\Sph)^W$. Then there exists a probability measure $\nu$ on $(\Sph)^W$ such that \begin{equation}\label{e:cutratio} \mu^\eta_{F, W} \ge e^{-2C} \nu \qquad \text{for all } \eta \in (\Sph)^{\Z^2 \setminus F}. \end{equation} \end{lemma} \begin{proof} Let $G_1$ and $G_2$ be as in the definition of a cut. We first compute the density $f_\eta(\xi)$ of $\mu^\eta_{F, W}$ with respect to $\prod_{w \in W} \dd \xi_w$. In the density of $\mu^\eta_F$, we set $n_w = \xi_w$ for $w \in W$. The factors of the bonds inside $W$ give $\Theta_W(\xi)$. Since every neighbor of a site of $W$ belongs to $F$, every remaining factor belongs to a bond inside $G$ or to a leg of $G$, and a leg $(x, z)$ of $G$ contributes the factor $1 + (\eta\xi)_{xz} \cdot n_x$. Integrating over $(n_x)_{x \in G}$ yields \begin{equation}\label{e:windowlaw} f_\eta(\xi) = \frac{\Theta_W(\xi) \, Z^{\eta\xi}_G}{Z^\eta_F}. \end{equation} Next, we factorize the partition functions of the domain $G \setminus X$. Since no bond joins $G_1$ to $G_2$, the integrand of $Z^{\eta\xi}_{G \setminus X}$ is the product of the factors of the bonds inside $G_1$, of the bonds inside $G_2$, and of the legs of $G$ at sites of $G_1 \cup G_2$. Every leg of $G$ at a site of $G_1$ ends in $W$, since every neighbor of a site of $G_1$ belongs to $F$. Similarly, every leg of $G$ at a site of $G_2$ ends outside $F$. Let $Z^{\eta\xi}_{G_1}$ and $Z^{\eta\xi}_{G_2}$ denote the partition functions of the domains $G_1 = G \setminus (X \cup G_2)$ and $G_2 = G \setminus (X \cup G_1)$ in $G$. Then \begin{equation}\label{e:factorize} Z^{\eta\xi}_{G \setminus X} = Z^{\eta\xi}_{G_1} Z^{\eta\xi}_{G_2}, \qquad Z^0_{G \setminus X} = Z^0_{G_1} Z^0_{G_2}, \end{equation} where $Z^{\eta\xi}_{G_1}$ is a function of $\xi$ alone and $Z^{\eta\xi}_{G_2}$ is a function of $\eta$ alone. By the definition \eqref{e:cutcost} of the cut cost and by \eqref{e:factorize}, we have \begin{equation*} Z^{\eta\xi}_G = e^{-\mathrm{cost}_X(\eta, \xi)} \, \frac{Z^{\eta\xi}_{G \setminus X} \, Z^0_G}{Z^0_{G \setminus X}} = e^{-\mathrm{cost}_X(\eta, \xi)} \, Z^{\eta\xi}_{G_1} \, \frac{Z^{\eta\xi}_{G_2} \, Z^0_G}{Z^0_{G_1} Z^0_{G_2}}. \end{equation*} We set \begin{equation*} N_1 = \int \Theta_W(\xi) \, Z^{\eta\xi}_{G_1} \prod_{w \in W} \dd \xi_w, \qquad \bar f(\xi) = \frac{\Theta_W(\xi) \, Z^{\eta\xi}_{G_1}}{N_1}, \qquad c_\eta = \frac{N_1 \, Z^{\eta\xi}_{G_2} \, Z^0_G}{Z^0_{G_1} Z^0_{G_2} Z^\eta_F}. \end{equation*} Since $Z^{\eta\xi}_{G_1}$ is a function of $\xi$ alone, $N_1 > 0$ and the probability density $\bar f$ on $(\Sph)^W$ are the same for all $\eta$. Since $Z^{\eta\xi}_{G_2}$ is a function of $\eta$ alone, so is $c_\eta > 0$. Inserting the last two displays into \eqref{e:windowlaw}, we find \begin{equation}\label{e:windowfactor} f_\eta(\xi) = c_\eta \, \bar f(\xi) \, e^{-\mathrm{cost}_X(\eta, \xi)}. \end{equation} Integrating \eqref{e:windowfactor} over $\xi$ and using the assumption that $|\mathrm{cost}_X| \le C$, we find \begin{equation*} 1 = c_\eta \int \bar f(\xi) \, e^{-\mathrm{cost}_X(\eta, \xi)} \prod_{w \in W} \dd \xi_w \le c_\eta \, e^{C}. \end{equation*} We conclude that $c_\eta \ge e^{-C}$, and by \eqref{e:windowfactor}, we have $f_\eta(\xi) \ge e^{-2C} \bar f(\xi)$ for all $\xi$. This is \eqref{e:cutratio}, with $\nu$ the probability measure with density $\bar f$. \end{proof} A pin fixes the spin at the other end of a leg, whereas deleting a site with at most three neighbors removes the factors of its bonds and nothing else; by \zcref{l:delete}, the effect of such a deletion on $\log \Xi$ is a combination of responses, which Hypothesis~(P) below makes small far from the pins. This is why \zcref{l:cut} compares the boundary configurations through the cost of removing a cut, and not by changing the pins on $\Z^2 \setminus F$ one at a time. \section{Hypothesis (P) and the removal lemma}\label{s:cut-removal} For a finite set $G \subset \Z^2$ and a site $x \in \Z^2$, we call \begin{equation*} \dep_G(x) = \dist(x, \Z^2 \setminus G) \end{equation*} the \emph{depth} of $x$ in $G$, and we define the boundary layer \begin{equation*} \partial_2 G = \{ z \in G : \dep_G(z) \le 2 \}. \end{equation*} By the triangle inequality, we have $|\dep_G(x) - \dep_G(z)| \le d(x, z)$ for all $x, z \in \Z^2$. Further, $x \in G$ if and only if $\dep_G(x) \ge 1$, and all four neighbors of $x$ belong to $G$ if and only if $\dep_G(x) \ge 2$. In particular, every leg of $G$ starts at a site of $\partial_2 G$. \begin{hypothesis}{P} There exist $A \ge 2$ and $\theta \in (0, 1)$ such that the following holds. Let $G \subset \Z^2$ be finite, let $\tau$ be a pin configuration on $G$, let $Y \subset G \setminus \partial_2 G$, and let $u, v \in G \setminus Y$ be sites with $d(u, v) = 2$ that have a common neighbor in $Y$. Then \begin{equation}\label{e:P} \bigl| \E^\tau_{G \setminus Y} [n_u \cdot n_v] - \E^0_{G \setminus Y} [n_u \cdot n_v] \bigr| \le A \, \theta^{\min(\dist(u, \partial_2 G), \, \dist(v, \partial_2 G))}. \end{equation} \end{hypothesis} The left side of \eqref{e:P} is the absolute value of the response of the pair $(u, v)$ in the domain $G \setminus Y$. Since $|n_u \cdot n_v| \le 1$, every response has absolute value at most $2$, and the condition $A \ge 2$ makes \eqref{e:P} trivial when $u \in \partial_2 G$ or $v \in \partial_2 G$. Note that the distance in \eqref{e:P} is measured to $\partial_2 G$ and not to the removed set $Y$, although $u$ and $v$ are neighbors of a site of $Y$. \zcref[S]{t:contraction} proves Hypothesis~(P) from Hypotheses~(H1), (H2), and (H3). We now bound the cost of removing a connected set of sites far from $\Z^2 \setminus G$. \begin{lemma}\label{l:removal} Assume Hypothesis~(P), with constants $A$ and $\theta$. Let $G \subset \Z^2$ be finite, and let $\tau$ be a pin configuration on $G$. Let $X \subset G$ be a nonempty set, and let $x_1, \dots, x_{|X|}$ be an enumeration of $X$ such that $\dep_G(x_j) \ge 3$ for all $j$ and such that $x_j$ has a neighbor in $\{x_1, \dots, x_{j - 1}\}$ for all $j \ge 2$. Set $C_\star = \log 8192$. Then \begin{equation}\label{e:removal} \bigl| \log \Xi_{G \setminus X}(\tau) - \log \Xi_G(\tau) \bigr| \le C_\star + 2 A \sum_{j = 2}^{|X|} \theta^{\dep_G(x_j) - 3}. \end{equation} \end{lemma} \begin{proof} For $0 \le j \le |X|$, we set $Y_j = \{x_1, \dots, x_j\}$ and $D_j = G \setminus Y_j$, so that $D_0 = G$, $D_{|X|} = G \setminus X$, and $D_j$ is obtained from $D_{j - 1}$ by deleting $x_j$. Then \begin{equation}\label{e:telescope} \log \Xi_G(\tau) - \log \Xi_{G \setminus X}(\tau) = \sum_{j = 1}^{|X|} \bigl( \log \Xi_{D_{j - 1}}(\tau) - \log \Xi_{D_j}(\tau) \bigr). \end{equation} Note that all four neighbors of $x_j$ belong to $G$, since $\dep_G(x_j) \ge 3$. They all belong to $D_0 = G$ when $j = 1$, while at most three of them belong to $D_{j - 1}$ when $j \ge 2$, since $x_j$ has a neighbor in $Y_{j - 1}$. We first bound the summand with $j = 1$. Let $u_1, \dots, u_4$ be the neighbors of $x_1$, and set \begin{equation*} \tilde I_{x_1} = \int \prod_{a = 1}^4 (1 + n \cdot n_{u_a}) \dd n. \end{equation*} As in the proof of \zcref{l:delete}, integrating out $n_{x_1}$ first gives \begin{equation*} Z^\cdot_{D_0} = Z^\cdot_{D_1} \, \E^\cdot_{D_1}[\tilde I_{x_1}]. \end{equation*} Each of the four factors in $\tilde I_{x_1}$ is at most $2$, which implies $\tilde I_{x_1} \le 16$. For a unit vector $b$, the variable $1 + n \cdot b$ is uniformly distributed on $[0, 2]$ under $\dd n$ (this is Archimedes' theorem). In particular, it is smaller than $1/4$ with probability $1/8$. By a union bound over the four factors, the set of $n$ on which every factor is at least $1/4$ has measure at least $1/2$, which gives $\tilde I_{x_1} \ge 4^{-4}/2$. Therefore $\E^\tau_{D_1}[\tilde I_{x_1}]$ and $\E^0_{D_1}[\tilde I_{x_1}]$ belong to $[1/512, 16]$, and \begin{equation*} \bigl| \log \Xi_{D_0}(\tau) - \log \Xi_{D_1}(\tau) \bigr| = \bigl| \log \E^\tau_{D_1}[\tilde I_{x_1}] - \log \E^0_{D_1}[\tilde I_{x_1}] \bigr| \le \log (16 \cdot 512) = C_\star. \end{equation*} Next, let $j \ge 2$. We apply \zcref{l:delete} with $D = D_{j - 1}$ and $x = x_j$, whose neighbors $u_1, \dots, u_k$ in $D_{j - 1}$ number $k \le 3$. This gives \begin{equation}\label{e:removalstep} \bigl| \log \Xi_{D_{j - 1}}(\tau) - \log \Xi_{D_j}(\tau) \bigr| \le \frac23 \sum_{1 \le a < a' \le k} \bigl| \E^\tau_{D_j}[n_{u_a} \cdot n_{u_{a'}}] - \E^0_{D_j}[n_{u_a} \cdot n_{u_{a'}}] \bigr|. \end{equation} We bound each of the at most three responses on the right side by Hypothesis~(P), applied with $Y = Y_j$, $u = u_a$, and $v = u_{a'}$. Its assumptions are satisfied. Indeed, we have $Y_j \subset G \setminus \partial_2 G$ since $\dep_G(x_r) \ge 3$ for all $r$; the sites $u_a$ and $u_{a'}$ belong to $D_j$ and satisfy $d(u_a, u_{a'}) = 2$; and $x_j \in Y_j$ is a common neighbor of $u_a$ and $u_{a'}$. To bound the exponent in \eqref{e:P}, let $z \in \partial_2 G$. Then \begin{equation*} d(u_a, z) \ge \dep_G(u_a) - \dep_G(z) \ge \dep_G(x_j) - 1 - 2, \end{equation*} where we used that $\dep_G$ is Lipschitz with constant $1$, that $\dep_G(u_a) \ge \dep_G(x_j) - 1$, and that $\dep_G(z) \le 2$. It follows that $\dist(u_a, \partial_2 G) \ge \dep_G(x_j) - 3$, and the same bound holds for $u_{a'}$. Since $\theta < 1$, Hypothesis~(P) bounds each response in \eqref{e:removalstep} by $A \theta^{\dep_G(x_j) - 3}$ in absolute value. Inserting this into \eqref{e:removalstep}, we find \begin{equation*} \bigl| \log \Xi_{D_{j - 1}}(\tau) - \log \Xi_{D_j}(\tau) \bigr| \le \frac23 \cdot 3 \cdot A \theta^{\dep_G(x_j) - 3} = 2 A \theta^{\dep_G(x_j) - 3}. \end{equation*} Summing over $j$ in \eqref{e:telescope}, and using the bound for $j = 1$, we arrive at \eqref{e:removal}. \end{proof} The first removal costs up to $C_\star = \log 8192 < 10$, which is not small. This is harmless, since \zcref{l:cut} requires a bounded cut cost, not a small one. \section{The cut for a shell}\label{s:cut-geometry} In this subsection, we construct a cut for the window of a shell, and we bound the depths of the sites of the cut. Throughout this subsection, we fix $M \ge M_{\mathrm g}$, a site $y$ with $\dist_\infty(y, \mathcal C_M) = 1$, and an index $i \in \{2, \dots, \lfloor \ell \rfloor\} \setminus I_{\mathrm{corner}}$, and we write $k = k_i = \lfloor i \ell \rfloor$. Recall the sets $F_i$, $\Sigma_i$, $\Sigma_i^\partial$, and $W_i$ from \zcref{s:block}. We set \begin{equation*} F = F_i, \qquad W = W_i, \qquad G = F \setminus W. \end{equation*} Recall that $M_{\mathrm g} = 73^4$, so that $\ell = M^{1/4} \ge 73$. Note that $|y|_\infty = M + 1$, that $2 \ell - 1 \le k \le \ell^2$, and that $\Sigma_i = \{ x \in \mathcal C_M : |x - y|_\infty = k \}$, since $|x - y|_\infty$ is an integer. For $x \in \Z^2$, we define the \emph{offset} of $x$ from the shell and the \emph{height} of $x$ above the boundary of the box by \begin{equation*} \off(x) = \bigl| |x - y|_\infty - k \bigr|, \qquad \hgt(x) = \dist_\infty(x, \Z^2 \setminus \mathcal C_M). \end{equation*} We have $\hgt(x) = M + 1 - |x|_\infty$ for $x \in \mathcal C_M$, and $\hgt(x) = 0$ for $x \notin \mathcal C_M$. By the triangle inequality for $| \cdot |_\infty$, we have \begin{equation}\label{e:lipschitz} |\off(x) - \off(z)| \le |x - z|_\infty, \qquad |\hgt(x) - \hgt(z)| \le |x - z|_\infty \qquad \text{for all } x, z \in \Z^2. \end{equation} Recall that $d_0 = 24$. In this notation, $\Sigma_i = \{ x \in \mathcal C_M : \off(x) = 0 \}$ and $W = \{ x \in \Sigma_i : \hgt(x) \ge d_0 + 1 \}$. We define \begin{equation}\label{e:U} U = \bigl\{ x \in \mathcal C_M : \off(x) \le \min \bigl( \hgt(x) - d_0/2, \, \ell/12 \bigr) \bigr\}. \end{equation} The set $U$ is a band of half-width $\ell/12$ around the shell, which near the boundary of $\mathcal C_M$ tapers linearly and ends at height $d_0/2 = 12$ (see \zcref{f:cut}). We call two sites $x$ and $z$ \emph{$*$-neighbors} if $|x - z|_\infty = 1$. A sequence of sites in which consecutive sites are $*$-neighbors is a \emph{$*$-path}, and a set $A \subset \Z^2$ is \emph{$*$-connected} if any two of its sites are joined by a $*$-path in $A$. A set is \emph{connected} if any two of its sites are joined by a nearest-neighbor path in it. For $A \subset \Z^2$, we let \begin{equation*} \partial_\infty A = \{ z \in \Z^2 \setminus A : \dist_\infty(z, A) = 1 \} \end{equation*} denote the outer $\ell^\infty$ boundary of $A$. We define \begin{equation}\label{e:cutsets} \begin{aligned} X_0 &= \partial_\infty U, & X &= \{ z \in \Z^2 : \dist(z, X_0) \le 1 \}, \\ G_1 &= U \setminus (W \cup X), & G_2 &= G \setminus (U \cup X). \end{aligned} \end{equation} The set $X_0$ surrounds $U$, but it need not be connected. The set $X$ adds to $X_0$ the nearest neighbors of its sites, and \zcref{l:Xconn} shows that $X$ is connected. We need this in order to remove $X$ in an order to which \zcref{l:removal} applies. The two topological facts that we use, \zcref{l:boundaryconn,l:thicken}, are proved in \zcref{s:topology}, independently of this section. \begin{figure}[t] \centering \begin{tikzpicture}[scale=0.85] \fill[gray!12] (-1.7, -2.6) rectangle (0, 2.6); \draw[thick] (0, -2.6) -- (0, 2.6); \node[font=\small] at (-0.85, 2.15) {$\Z^2 \setminus \mathcal C_M$}; \draw[line width=7pt, gray!45, line join=miter] (11, 1.6) -- (2.54, 1.6) -- (0.94, 0) -- (2.54, -1.6) -- (11, -1.6); \draw[dashed] (11, 1.4) -- (2.6, 1.4) -- (1.2, 0) -- (2.6, -1.4) -- (11, -1.4); \draw[thick] (0, 0) -- (2.4, 0); \draw[line width=2.4pt] (2.5, 0) -- (11, 0); \node[font=\small] at (0.42, 0.36) {$\Sigma_i^\partial$}; \node[font=\small, above] at (6.5, 0.05) {$W_i$}; \node[font=\small] at (11.45, 0) {$\Sigma_i$}; \node[font=\small] at (11.45, 1.6) {$X$}; \node[font=\small] at (11.45, -1.6) {$X$}; \node[font=\small] at (4.6, 0.75) {$G_1$}; \node[font=\small] at (4.6, -0.75) {$G_1$}; \node[font=\small] at (6.5, 2.2) {$G_2$}; \node[font=\small] at (6.5, -2.2) {$G_2$}; \draw[<->] (9.4, 0.08) -- (9.4, 1.32); \node[font=\small, right] at (9.45, 0.7) {$\ell/12$}; \draw[dotted] (1.2, 0) -- (1.2, -2.35); \draw[<->] (0, -2.35) -- (1.2, -2.35); \node[font=\small, below] at (0.6, -2.35) {$d_0/2$}; \end{tikzpicture} \caption{A schematic picture, not to scale, of one end of the shell $\Sigma_i$ near the boundary of the box $\mathcal C_M$. The shell is the horizontal segment; its part within $\ell^\infty$ distance $d_0$ of $\Z^2 \setminus \mathcal C_M$ is $\Sigma_i^\partial$, and the rest is the window $W_i$ (thick). The region $U$ (dashed boundary) is a band of half-width $\ell/12$ around the shell that tapers linearly and ends at height $d_0/2$. The cut $X$ (shaded) surrounds $U$ and separates $G_1$, which lies in $U$, from $G_2$, which contains the sites of $G$ near the boundary of the box.} \label{f:cut} \end{figure} The shape of $U$ is dictated by the ends of the window. Away from the ends, a band of constant half-width $\ell/12$ keeps the cut at depth of order $\ell$. However, the window comes within $\ell^\infty$ distance $d_0 + 1$ of the pins on $\Z^2 \setminus \mathcal C_M$, and near its ends the cut must pass at a bounded depth. Because $U$ tapers, the depth of a site of $X_0$ near the boundary of $\mathcal C_M$ grows linearly with its offset, and \zcref{l:spherering} shows that a bounded number of sites of $X_0$ have a given offset. The sum of the bounds of \zcref{l:removal} over $X$ is therefore bounded uniformly in $M$, which is the content of \zcref{l:cutgeom}. This count near the ends of the window is the only delicate point of the construction. For the counting, recall from \zcref{s:block-shells} the sphere $\{ x \in \Z^2 : |x - a|_\infty = r \}$ of radius $r$ around a site $a$ and its four sides, the segments $\{a_1 \pm r\} \times [a_2 - r, a_2 + r]$ and $[a_1 - r, a_1 + r] \times \{a_2 \pm r\}$ of lattice points. Thus $\Sigma_i$ is the part in $\mathcal C_M$ of the sphere of radius $k$ around $y$, and for $1 \le h \le M$, the set $\{ x \in \mathcal C_M : \hgt(x) = h \}$ is the sphere of radius $M + 1 - h$ around $0$. \begin{lemma}\label{l:spherering} Let $j$ and $h$ be integers with $0 \le j \le \ell/12$ and $j + 10 \le h \le j + 11$. Then at most $16$ sites $x \in \mathcal C_M$ satisfy $\off(x) = j$ and $\hgt(x) = h$. \end{lemma} \begin{proof} Let $Q'$ denote the sphere of radius $M + 1 - h$ around $0$. Since $\ell \ge 73$, we have \begin{equation*} 1 \le h \le \frac{\ell}{12} + 11 \le M, \qquad Q' = \{ x \in \mathcal C_M : \hgt(x) = h \}. \end{equation*} Every site $x$ in the lemma belongs to $Q'$ and to one of the spheres of radius $k - j$ and $k + j$ around $y$. It therefore suffices to show that, for $r \in \{k - j, k + j\}$, the sphere $Q$ of radius $r$ around $y$ meets $Q'$ in at most $8$ sites. Note that \begin{equation}\label{e:radiusbounds} r \ge k - j \ge 2 \ell - 1 - \ell/12 \ge \ell/4, \qquad r \le k + j \le \ell^2 + \ell/12 \le M. \end{equation} We claim that it suffices to show that no side of $Q$ that meets $\mathcal C_M$ lies on a line containing a side of $Q'$. Indeed, suppose that this holds. A side of $Q$ that does not meet $\mathcal C_M$ does not meet $Q' \subset \mathcal C_M$. A side of $Q$ that meets $\mathcal C_M$ is parallel to two sides of $Q'$ and lies on a different line from each of them. It meets these two sides in no site. Further, it meets each of the two perpendicular sides of $Q'$ in at most one site. Hence every side of $Q$ meets $Q'$ in at most two sites, and $Q$ meets $Q'$ in at most $8$ sites. To verify the condition, we use the symmetries of the problem. The rotations and reflections of $\Z^2$ that fix $0$ preserve $\mathcal C_M$, $\hgt$, and $\ell^\infty$ distances, and they map spheres to spheres and sides to sides. The set $I_{\mathrm{corner}}$ and the conclusion of part (b) of \zcref{l:shellgeom} are invariant under them (see the proof of \zcref{l:shellgeom}). Since $|y|_\infty = M + 1$, we may therefore assume that $y_2 = -M - 1$ and $|y_1| \le M + 1$. For $c \in [-M - 1, M + 1]$, the vertical line $\{x_1 = c\}$ has $\ell^\infty$ distance $M + 1 - c$ from the half-plane $\{x_1 \ge M + 1\}$ and $M + 1 + c$ from the half-plane $\{x_1 \le -M - 1\}$. It contains a side of $Q'$ if and only if one of these two distances is $h$. The same holds for horizontal lines, with $x_2$ in place of $x_1$. We also have \begin{equation}\label{e:hrange} j < h \le \ell/12 + 11 < \ell/4, \end{equation} where we used $\ell \ge 73$ in the last inequality. It suffices to show that every side of $Q$ that meets $\mathcal C_M$ has distance at most $j$ or at least $\ell/4$ from each of the two half-planes whose boundary lines are parallel to it. We consider the four sides in turn. The bottom side of $Q$ lies on the line $\{x_2 = -M - 1 - r\}$, so it does not meet $\mathcal C_M$. The top side lies on the line $\{x_2 = -M - 1 + r\}$, whose distances from the half-planes $\{x_2 \le -M - 1\}$ and $\{x_2 \ge M + 1\}$ are $r$ and $2M + 2 - r$. By \eqref{e:radiusbounds}, these distances are at least $\ell/4$. The right side of $Q$ lies on the line $\{x_1 = y_1 + r\}$. Its distance from the half-plane $\{x_1 \le -M - 1\}$ is $y_1 + r + M + 1 \ge r \ge \ell/4$. To bound its distance $M + 1 - y_1 - r$ from the half-plane $\{x_1 \ge M + 1\}$, we consider the right side of the sphere of radius $k$ around $y$, which lies on the line $\{x_1 = y_1 + k\}$, and we distinguish two cases. Suppose first that $y_1 + k \le M$. Since $y_1 \ge -M - 1$ and $k \ge 1$, we have \begin{equation*} -M \le y_1 + k \le M, \qquad (y_1 + k, -M) \in \mathcal C_M, \end{equation*} and the right side $\{ x \in \Z^2 : x_1 = y_1 + k, \ |x_2 - y_2| \le k \}$ of the sphere of radius $k$ contains the site $(y_1 + k, -M)$. In particular, this side meets $\mathcal C_M$, and part (b) of \zcref{l:shellgeom}, which applies since $i \notin I_{\mathrm{corner}}$, gives $|y_1 + k| \le M + 1 - \ell/3$, that is, $M + 1 - y_1 - k \ge \ell/3$. Using $r \le k + j$ and $j \le \ell/12$, we obtain \begin{equation*} M + 1 - y_1 - r \ge M + 1 - y_1 - k - j \ge \frac{\ell}{3} - \frac{\ell}{12} = \frac{\ell}{4}. \end{equation*} Suppose next that $y_1 + k \ge M + 1$. Then \begin{equation*} M + 1 - y_1 - r \le M + 1 - y_1 - k + j \le j. \end{equation*} In both cases, the distance is at most $j$ or at least $\ell/4$. The reflection $x_1 \mapsto -x_1$ preserves our normalization $y_2 = -M - 1$. It maps the left side of $Q$ to the right side of the sphere of radius $r$ around $(-y_1, y_2)$. The argument for the right side therefore applies to the left side as well. This completes the proof. \end{proof} \begin{lemma}\label{l:cutgeom} Let $M$, $y$, and $i$ satisfy the standing assumptions of this subsection, and let $U$, $X_0$, $X$, $G_1$, and $G_2$ be given by \eqref{e:U} and \eqref{e:cutsets}. \begin{enumerate} \item We have $W \subset U \subset F$. Every neighbor of a site of $W$ belongs to $U$, and every neighbor of a site of $U$ belongs to $F$. \item Every $x \in X_0$ satisfies $\dep_G(x) \ge 4$, every $x \in X$ satisfies $\dep_G(x) \ge 3$, and for all $t \in (0, 1)$, we have \begin{equation}\label{e:cutsum} \sum_{x \in X} t^{\dep_G(x)} \le 240 \, t^3 \Bigl( 7 + \frac{1}{1 - t} \Bigr) + 5 (2M + 1)^2 \, t^{\ell/12 - 4}. \end{equation} \item We have $X \subset G$, and $X$ is a cut, with $G \setminus X$ the disjoint union of the sets $G_1$ and $G_2$ in \eqref{e:cutsets}. \end{enumerate} \end{lemma} \begin{proof} We prove the three parts in turn. Recall from \eqref{e:U} that every $x' \in U$ satisfies $x' \in \mathcal C_M$, $\off(x') \le \ell/12$, and $\hgt(x') \ge \off(x') + 12$. \emph{The first part.} Let $x \in W$. Then $\off(x) = 0$ and $\hgt(x) \ge 25$, which implies $x \in U$. Let $z$ be a neighbor of $x$. By \eqref{e:lipschitz}, we have $\off(z) \le 1$ and $\hgt(z) \ge 24$, so that $z \in \mathcal C_M$ and \begin{equation*} \off(z) \le 1 \le \min \bigl( \hgt(z) - 12, \, \ell/12 \bigr), \end{equation*} where we used $\ell \ge 12$. This shows that $z \in U$. Now let $x \in U$. Then $\hgt(x) \ge 12$, and every neighbor of $x$ belongs to $\mathcal C_M$. Every site $z \in \mathcal C_M \setminus F$ satisfies \begin{equation}\label{e:CminusF} |z - y|_\infty \le (i - 1) \ell = i \ell - \ell < k + 1 - \ell, \end{equation} where we used $k > i \ell - 1$. Suppose that $x$ had a neighbor $z \in \mathcal C_M \setminus F$. Then \eqref{e:lipschitz} and \eqref{e:CminusF} would give \begin{equation*} \off(x) \ge k - |x - y|_\infty \ge k - |z - y|_\infty - 1 > \ell - 2, \end{equation*} which contradicts $\off(x) \le \ell/12$. Hence every neighbor of $x$ belongs to $F$. Finally, we have \begin{equation*} |x - y|_\infty \ge k - \off(x) \ge k - \ell/12 > i \ell - 1 - \ell/12 > (i - 1) \ell, \end{equation*} which shows that $x \in F$. This proves the first part. \emph{A lower bound for the depth.} We now turn to the second part. Let $x \in X_0$, and set $j = \off(x)$. We have \begin{equation*} \Z^2 \setminus G = (\Z^2 \setminus \mathcal C_M) \cup (\mathcal C_M \setminus F) \cup W, \end{equation*} and the graph distance dominates the $\ell^\infty$ distance. Therefore $\dep_G(x)$ is at least the smallest of the $\ell^\infty$ distances from $x$ to these three sets, which satisfy \begin{equation*} \begin{gathered} \dist_\infty(x, \Z^2 \setminus \mathcal C_M) = \hgt(x), \qquad \dist_\infty(x, \mathcal C_M \setminus F) > \ell - 1 - j, \\ \dist_\infty(x, W) \ge \max(j, 25 - \hgt(x)). \end{gathered} \end{equation*} The first identity is the definition of $\hgt$. The second bound follows from \eqref{e:CminusF}, since $|x - z|_\infty \ge |x - y|_\infty - |z - y|_\infty$ and $|x - y|_\infty \ge k - j$. The third follows from \eqref{e:lipschitz}, since every $w \in W$ satisfies $\off(w) = 0$ and $\hgt(w) \ge 25$. We conclude that \begin{equation}\label{e:depthmin} \dep_G(x) \ge \min \bigl\{ \max \bigl( j, \, 25 - \hgt(x) \bigr), \, \hgt(x), \, \ell - 1 - j \bigr\}. \end{equation} \emph{The neighbor in $U$.} Since $x \in \partial_\infty U$, there exists $x' \in U$ such that $|x - x'|_\infty = 1$. Using $\off(x') \le \ell/12$ and $\hgt(x') \ge \off(x') + 12$, together with \eqref{e:lipschitz}, we obtain \begin{equation}\label{e:fromU} j \le \ell/12 + 1, \qquad \hgt(x) \ge \hgt(x') - 1 \ge \off(x') + 11 \ge j + 10. \end{equation} In particular, $\hgt(x) > 0$, and $X_0 \subset \mathcal C_M$. We now distinguish two cases. \emph{Sites far from the shell.} Suppose that $j > \ell/12$. By \eqref{e:fromU}, we have $\hgt(x) > \ell/12 + 10$ and \begin{equation*} \ell - 1 - j \ge \ell - 2 - \ell/12 > \ell/12. \end{equation*} Inserting these bounds into \eqref{e:depthmin}, we find $\dep_G(x) > \ell/12$. Since $X_0 \subset \mathcal C_M$, there are at most $(2M + 1)^2$ such sites, and \begin{equation}\label{e:farsum} \sum_{\substack{x \in X_0 \\ \off(x) > \ell/12}} t^{\dep_G(x)} \le (2M + 1)^2 \, t^{\ell/12}. \end{equation} \emph{Sites near the shell.} Suppose that $j \le \ell/12$. Since $x \in \mathcal C_M \setminus U$ and $\off(x) \le \ell/12$, the definition \eqref{e:U} implies that $\off(x) > \hgt(x) - 12$, that is, $\hgt(x) \le j + 11$. Together with \eqref{e:fromU}, this gives \begin{equation}\label{e:hgtnear} \hgt(x) \in \{ j + 10, \, j + 11 \}. \end{equation} Since $j \le \ell/12$ and $\ell \ge 12$, we have \begin{equation*} \ell - 1 - j \ge \frac{11 \ell}{12} - 1 \ge \frac{\ell}{12} + 6. \end{equation*} Inserting this bound and \eqref{e:hgtnear} into \eqref{e:depthmin}, we find \begin{equation}\label{e:depthnear} \dep_G(x) \ge \min \bigl\{ \max(j, 14 - j), \, j + 10, \, \ell/12 + 6 \bigr\} \ge \max(j, 6). \end{equation} We now count these sites. By \eqref{e:hgtnear} and \zcref{l:spherering}, applied with $h = j + 10$ and $h = j + 11$, at most $32$ sites of $X_0$ have offset $j$, for each integer $j \in [0, \ell/12]$. Combining this with \eqref{e:depthnear}, we find \begin{equation}\label{e:nearsum} \sum_{\substack{x \in X_0 \\ \off(x) \le \ell/12}} t^{\dep_G(x)} \le 32 \sum_{j \ge 0} t^{\max(j, 6)} = 32 \Bigl( 7 t^6 + \frac{t^7}{1 - t} \Bigr) \le 32 \, t^6 \Bigl( 7 + \frac{1}{1 - t} \Bigr). \end{equation} \emph{Conclusion of the second part.} In the two cases, we have shown that $\dep_G(x) \ge 6$ for all $x \in X_0$. Every site of $X$ is within graph distance $1$ of a site of $X_0$, and $\dep_G$ is Lipschitz with constant $1$, which implies $\dep_G(z) \ge 5$ for all $z \in X$. Adding \eqref{e:farsum} and \eqref{e:nearsum}, we find \begin{equation*} \sum_{x \in X_0} t^{\dep_G(x)} \le 32 \, t^6 \Bigl( 7 + \frac{1}{1 - t} \Bigr) + (2M + 1)^2 \, t^{\ell/12}. \end{equation*} The set $X_0 \subset \mathcal C_M$ is finite, and the second part of \zcref{l:thicken} gives \begin{equation*} \sum_{z \in X} t^{\dep_G(z)} \le 5 t^{-1} \sum_{x \in X_0} t^{\dep_G(x)} \le 160 \, t^5 \Bigl( 7 + \frac{1}{1 - t} \Bigr) + 5 (2M + 1)^2 \, t^{\ell/12 - 1}. \end{equation*} Since $0 < t < 1$, this implies \eqref{e:cutsum}. The bounds on the depths in the second part hold as well. \emph{The third part.} By the second part, every site of $X$ has positive depth, which means that $X \subset G$. By the first part, $U \subset F$, and therefore $U \cap G = U \setminus W$. It follows that $G_1 = (U \cap G) \setminus X$ and $G_2 = (G \setminus U) \setminus X$ partition $G \setminus X$. Let $x \in G_1$, and let $z$ be a neighbor of $x$ with $z \notin U$. Then $\dist_\infty(z, U) = 1$, so that $z \in X_0 \subset X$ and $z \notin G_2$. This shows that no bond joins $G_1$ to $G_2$. By the first part, every neighbor of a site of $G_1 \subset U$ belongs to $F$, and every neighbor of a site of $W$ belongs to $U$. Since $G_2$ is disjoint from $U$, every neighbor of a site of $G_2$ belongs to $\Z^2 \setminus W$, and $X$ is a cut. \end{proof} The proof gives better constants than those in \eqref{e:cutsum}. We state \eqref{e:cutsum} in the form used in \zcref{s:explicit}. \begin{lemma}\label{l:Xconn} Let $M$, $y$, and $i$ satisfy the standing assumptions of this subsection. Then the set $X$ of \eqref{e:cutsets} is connected. \end{lemma} \begin{proof} We will show that $U$ is nonempty and $*$-connected, and that $\Z^2 \setminus U$ is connected, hence $*$-connected. Since $U \subset \mathcal C_M$ is finite, \zcref{l:boundaryconn} then shows that $X_0 = \partial_\infty U$ is nonempty and $*$-connected, and the first part of \zcref{l:thicken} shows that $X$ is connected. Observe that \begin{equation}\label{e:Ualt} U = \bigl\{ x \in \Z^2 : \off(x) \le \ell/12, \ \hgt(x) - \off(x) \ge 12 \bigr\}, \end{equation} since $\hgt(x) \ge 12$ implies $x \in \mathcal C_M$. We set $m = M - 11$, so that $\hgt(x) \ge 12$ if and only if $|x|_\infty \le m$. Then \begin{equation}\label{e:kbounds} 12 \le 2 \ell - 1 \le k \le \ell^2 < m. \end{equation} \emph{Moving towards the shell.} Let $x \in U$. Note that $x \ne y$, since $y \notin \mathcal C_M$. Suppose that $|x - y|_\infty > k$. We move every coordinate $x_a$ with $|x_a - y_a| = |x - y|_\infty$ one step towards $y_a$. This is a $*$-step that decreases $|x - y|_\infty$ by one. Suppose instead that $|x - y|_\infty < k$. We move one coordinate $x_a$ with $|x_a - y_a| = |x - y|_\infty$ one step away from $y_a$. This is a nearest-neighbor step that increases $|x - y|_\infty$ by one. After each of these steps, $\off$ has decreased by one and $\hgt$ has changed by at most one. The difference $\hgt - \off$ has not decreased, and the new site belongs to $U$ by \eqref{e:Ualt}. Repeating these steps, we join $x$ inside $U$ to a site of \begin{equation*} K = \{ x \in \Z^2 : |x - y|_\infty = k, \ |x|_\infty \le m \} = \{ x \in \Z^2 : |x - y|_\infty = k, \ \hgt(x) \ge 12 \}, \end{equation*} and $K \subset U$ by \eqref{e:Ualt}. It remains to show that $K$ is nonempty and connected. \emph{The set $K$.} For $a \in \{1, 2\}$, the interval $[y_a - k, y_a + k]$ meets $[-m, m]$, since $|y_a| \le M + 1$ and $k \ge 12 = M + 1 - m$. Further, at least one of the endpoints $y_a - k$ and $y_a + k$ belongs to $[-m, m]$. Otherwise we would have $y_a - k < -m$ and $y_a + k > m$, which implies $k > m$ and contradicts \eqref{e:kbounds}. We choose such an endpoint $c_a$ for $a = 1, 2$. Then $c = (c_1, c_2)$ is a corner of the square $y + [-k, k]^2$, and $c \in K$. The intersection of the square $y + [-k, k]^2$ with $[-m, m]^2$ is the product of the two intervals $[y_a - k, y_a + k] \cap [-m, m]$, each of which contains $c_a$. The two sides of the square through $c$ meet $[-m, m]^2$ in segments that contain $c$. Each of the two other sides meets $[-m, m]^2$, if at all, in a segment that contains its common corner with one of the first two sides, and this corner belongs to $K$. Thus $K$ is a union of at most four segments, each of which meets one of the two segments through $c$, and $K$ is connected. \emph{The complement of $U$.} The set $\Z^2 \setminus \mathcal C_M$ is connected and disjoint from $U$. Let $x \in \mathcal C_M \setminus U$. We join $x$ to $\Z^2 \setminus \mathcal C_M$ by a nearest-neighbor path in $\Z^2 \setminus U$, and we consider three cases. Suppose first that $|x - y|_\infty > k + \ell/12$. We repeatedly move a coordinate $x_a$ with $|x_a - y_a| = |x - y|_\infty$ one step away from $y_a$. Each step increases $|x - y|_\infty$, and $\off$ stays larger than $\ell/12$ until the path leaves $\mathcal C_M$. In particular, the path stays outside $U$. Suppose next that $|x - y|_\infty < k - \ell/12$. We repeatedly move a coordinate $x_a$ with $|x_a - y_a| = |x - y|_\infty$ one step towards $y_a$. The value of $|x - y|_\infty$ never increases, and $\off$ stays larger than $\ell/12$ until the path reaches $y \notin \mathcal C_M$. In particular, the path stays outside $U$. Otherwise, we have $\off(x) \le \ell/12$, and \eqref{e:Ualt} gives $\hgt(x) - \off(x) < 12$. Then $x \ne 0$. Indeed, if $\off(0) \le \ell/12$, then \begin{equation*} \hgt(0) - \off(0) \ge M + 1 - \frac{\ell}{12} > 12. \end{equation*} We repeatedly move a coordinate $x_a$ with $|x_a| = |x|_\infty$ one step away from $0$. Until the path leaves $\mathcal C_M$, each step decreases $\hgt$ by one and changes $\off$ by at most one. Thus $\hgt - \off$ does not increase, and the path stays outside $U$ by \eqref{e:Ualt}. In all three cases, the path ends in $\Z^2 \setminus \mathcal C_M$, and we conclude that $\Z^2 \setminus U$ is connected. This completes the proof. \end{proof} \section{Proof of the window estimate}\label{s:cut-window} We now prove the main result of this section. \begin{theorem}\label{t:window} Assume Hypothesis~(P), with constants $A \ge 2$ and $\theta \in (0, 1)$, and set \begin{equation}\label{e:Ccut} C_{\mathrm{cut}} = C_\star + 480 A \Bigl( 7 + \frac{1}{1 - \theta} \Bigr) + 10 A, \end{equation} with $C_\star = \log 8192$ as in \zcref{l:removal}. Then Hypothesis~(W) holds with $p_0 = e^{-2 C_{\mathrm{cut}}}$ and with a constant $M_\star$ that depends only on $\theta$. \end{theorem} \begin{proof} Since $(2M + 1)^2 \theta^{M^{1/4}/12 - 7} \to 0$ as $M \to \infty$, there exists $M_1$, depending only on $\theta$, such that \begin{equation}\label{e:Mone} (2M + 1)^2 \, \theta^{M^{1/4}/12 - 7} \le 1 \qquad \text{for all } M \ge M_1. \end{equation} We set \begin{equation*} M_\star = \max(M_{\mathrm g}, M_1). \end{equation*} Let $M \ge M_\star$, let $y$ satisfy $\dist_\infty(y, \mathcal C_M) = 1$, and let $i \in \{2, \dots, \lfloor \ell \rfloor\} \setminus I_{\mathrm{corner}}$, where $\ell = M^{1/4}$. Hypothesis~(W) asks for a probability measure $\nu_{M, y, i}$ on $(\Sph)^{W_i}$ such that $\mu^\eta_{F_i, W_i} \ge p_0 \nu_{M, y, i}$ for all $\eta \in (\Sph)^{\Z^2 \setminus F_i}$. Since $M \ge M_{\mathrm g}$, the standing assumptions of \zcref{s:cut-geometry} hold. We set $F = F_i$, $W = W_i$, and $G = F \setminus W$, and we define $U$, $X_0$, $X$, $G_1$, and $G_2$ by \eqref{e:U} and \eqref{e:cutsets}. By the first part of \zcref{l:cutgeom}, every neighbor of a site of $W$ belongs to $F$, and the setting of \zcref{s:cut-cutlemma} applies. By the third part of \zcref{l:cutgeom}, the set $X$ is a cut. We now bound the cut cost. Fix $\eta \in (\Sph)^{\Z^2 \setminus F}$ and $\xi \in (\Sph)^W$. By \zcref{l:Xconn}, the finite set $X$ is connected. We fix $x_1 \in X$, and we enumerate the sites of $X$ as $x_1, \dots, x_{|X|}$ in order of nondecreasing length of the shortest nearest-neighbor path in $X$ from $x_1$ (a breadth-first order). Then every $x_j$ with $j \ge 2$ has a neighbor among $x_1, \dots, x_{j - 1}$, and $\dep_G(x_j) \ge 3$ for all $j$ by the second part of \zcref{l:cutgeom}. Applying \zcref{l:removal} with $\tau = \eta\xi$, we obtain \begin{equation*} |\mathrm{cost}_X(\eta, \xi)| \le C_\star + 2 A \sum_{j = 2}^{|X|} \theta^{\dep_G(x_j) - 3} \le C_\star + 2 A \theta^{-3} \sum_{x \in X} \theta^{\dep_G(x)}. \end{equation*} Using \eqref{e:cutsum} with $t = \theta$, together with \eqref{e:Mone}, we find \begin{equation}\label{e:costbound} |\mathrm{cost}_X(\eta, \xi)| \le C_\star + 480 A \Bigl( 7 + \frac{1}{1 - \theta} \Bigr) + 10 A (2M + 1)^2 \, \theta^{\ell/12 - 7} \le C_{\mathrm{cut}}. \end{equation} By \eqref{e:costbound}, the cut $X$ satisfies the assumption of \zcref{l:cut} with $C = C_{\mathrm{cut}}$. This lemma implies that there exists a probability measure $\nu$ on $(\Sph)^W$ such that $\mu^\eta_{F, W} \ge e^{-2 C_{\mathrm{cut}}} \nu$ for all $\eta$. Setting $\nu_{M, y, i} = \nu$ proves Hypothesis~(W) with $p_0 = e^{-2 C_{\mathrm{cut}}}$. The constant $M_\star$ depends only on $\theta$, since $M_{\mathrm g}$ is an absolute constant. \end{proof} \begin{remark}\label{r:Pweak} The proof of \zcref{t:window} uses the exponential form of \eqref{e:P} only through the sum over $X$. Suppose that the right side of \eqref{e:P} is replaced by $f(\min(\dist(u, \partial_2 G), \dist(v, \partial_2 G)))$ for a nonincreasing function $f \colon \{0, 1, 2, \dots\} \to [0, \infty)$. Then the proofs of \zcref{l:removal,t:window} bound the cut cost by $C_\star + 2 \sum_{x \in X} f(\dep_G(x) - 3)$. By the proof of \zcref{l:cutgeom}, the sites of $X$ within distance $1$ of a site $x \in X_0$ with $\off(x) \le \ell/12$ contribute a bounded amount if $f$ is summable. The remaining sites, at most $5 (2M + 1)^2 \le 45 \ell^8$ in number, have depth larger than $\ell/12 - 1$, and they contribute a bounded amount if $f(d) = O(d^{-8})$. \zcref[S]{t:contraction} gives exponential decay, so we do not need this generality. \end{remark}