\chapter{Overview of the proof}\label{s:overview} This section explains the proof of \zcref{t:main}. It is written to be read on its own, with the notation of \zcref{s:intro-notation}. The precise statements and proofs occupy \zcref{s:quantum,s:block,s:cut,s:walks,s:contraction,s:computer} and the appendices; a roadmap is given in \zcref{s:ov-organization}. \section{The strategy}\label{s:ov-strategy} The proof is a chain of four reductions. First, we bound the gap of $H_{L, N}$ from below by the gap of a classical Markov process, whose stationary measure is the squared modulus of the VBS state, written as a function of spin coherent states, on configurations of unit vectors. Second, we bound the gap of this process through the coupling argument of Martinelli, Olivieri, and Schonmann \cite{MOS94}. Its only input is a weak form of spatial mixing, which asks that, for every boundary condition of a large box, the law of a segment of spins far from the boundary dominate a fixed positive multiple of one probability measure. Third, we deduce this property from the exponential decay of certain differences between correlations with and without boundary conditions, by removing a set of sites between the segment and the boundary. Fourth, we prove the required decay by an expansion of correlations into walks, in the spirit of Kennedy, Lieb, and Tasaki \cite{KLT88}, combined with a contraction argument whose numerical inputs we verify by computer. The first, third, and fourth steps are new. The comparison with a classical dynamics replaces the finite-size criteria used for the hexagonal lattice, which require a lower bound for the gap of a finite quantum system above an explicit threshold; in our argument, a finite quantum system enters only through the positivity of its gap. The removal of a separating set replaces the weak mixing hypothesis of \cite{MOS94} by a bound on the cost of removing sites, and this cost needs to be bounded, not small. The contraction argument replaces the positivity behind the expansion of \cite{KLT88}, which holds on bipartite tori and for N\'eel boundary conditions, by a comparison of the expansions with and without boundary conditions, walk by walk; the local inequalities that this comparison requires are new as well. The second step follows the proof of \cite[Proposition~2.1]{MOS94}, with an elementary overlap bound in place of the finiteness of the spin space. \section{From the quantum gap to a classical dynamics}\label{s:ov-quantum} We use the polynomial representation of spin $2$ \cite{AKLT88,AAH88}, in which we identify the states at one site with the homogeneous polynomials of degree four in a variable $z \in \C^2$, and the states of the torus with the polynomials in variables $z_x \in \C^2$, $x \in \Lambda$, that are homogeneous of degree four in each $z_x$. For a bond $\{x, y\}$, we call $z_{x, 1} z_{y, 2} - z_{x, 2} z_{y, 1}$ its \emph{singlet polynomial}. The kernel of the projection onto total spin $4$ on the bond consists of the polynomials divisible by this singlet polynomial. We let $\VBS$ denote the product of the singlet polynomials over all bonds of the torus; this is the VBS state. Let $\sigma = (\sigma_1, \sigma_2, \sigma_3)$ denote the vector of Pauli matrices. For $a \in \Sph$, we let $\zeta(a) \in \C^2$ denote a unit vector with $\zeta(a) \zeta(a)^* = (1 + a \cdot \sigma)/2$, which is a spin-$\frac12$ coherent state. We choose $\zeta$ measurably as in \zcref{s:quantum-poly}. Evaluation at the coherent states $\zeta(s_x)$ identifies the space of states isometrically with a subspace of $L^2((\Sph)^\Lambda, \prod_{x\in\Lambda} 5\,\mathrm d s_x)$. Let $\mathcal E$ denote the set of bonds of the torus, and define $\mu_{L, N}$ to be the probability measure on $(\Sph)^\Lambda$ given by \begin{equation}\label{e:ov-mu} \mu_{L, N}(\dd s) = \frac{1}{Z_{L, N}} \prod_{\{x, y\} \in \mathcal E} (1 - s_x \cdot s_y) \prod_{x \in \Lambda} \dd s_x, \end{equation} where $Z_{L, N}$ is the normalizing constant. The singlet polynomial of a bond $\{x, y\}$, evaluated at coherent states, has squared modulus $(1 - s_x \cdot s_y)/2$. Therefore the squared modulus of $\VBS$, as a function of $s$, is proportional to the density of $\mu_{L, N}$. This classical representation of the VBS state goes back to \cite{AAH88,KLT88}. We now explain the main observation of this step. Let $\varphi$ be a state orthogonal to $\VBS$, and let $F$ denote the ratio of $\varphi$ and $\VBS$, as functions of $s$. Then $\mu_{L, N}(F) = 0$ and $\Var_{\mu_{L, N}}(F) = \|\varphi\|^2/\|\VBS\|^2$. Fix a square $R \subset \Lambda$ of side $b$. Let $H_R$ be the sum of the bond projections over the bonds that meet $R$, and let $\psi$ be the orthogonal projection of $\varphi$ onto the kernel of $H_R$. Every polynomial in this kernel is divisible by the singlet polynomials of all bonds that meet $R$, and the same holds for $\VBS$. These singlet polynomials exhaust the degree of every variable $z_x$ with $x \in R$. Hence the ratio of $\psi$ and $\VBS$ does not depend on the spins in $R$. The conditional expectation given the spins outside $R$ is the best mean-square approximation of $F$ by such functions. Let $\Var_R F$ denote the conditional variance of $F$ given the spins outside $R$, and let $g_b$ denote the smallest positive eigenvalue of $H_R$. Then \begin{equation}\label{e:ov-comparison} \mu_{L, N} \bigl( \Var_R F \bigr) \le \frac{\|\varphi - \psi\|^2}{\|\VBS\|^2} \le \frac{\langle \varphi, H_R \varphi \rangle}{g_b \|\VBS\|^2}. \end{equation} We call the Markov dynamics that resamples each translate of $R$, at rate $b^{-2}$, from its conditional law under $\mu_{L, N}$ the \emph{block dynamics}. Up to the factor $b^{-2}$, the left side of \eqref{e:ov-comparison} is the contribution of $R$ to the Dirichlet form of the block dynamics. Summing \eqref{e:ov-comparison} over the translates of $R$, we find that for $L, N \ge b + 4$, the gap of $H_{L, N}$ is at least $b g_b/(b + 1)$ times the spectral gap of the block dynamics. This is \zcref{t:comparison}. In \eqref{e:ov-comparison}, the finite quantum system appears only through $g_b$, and a uniform gap requires only that $g_b$ be positive, which is automatic. However, the size of $g_b$ determines the constant in \zcref{t:main}. The characteristic polynomial of $70 H_R$ has integer coefficients, and it shows that $g_b \ge [140 b (b + 1)]^{-5^{b^2 + 4b}}$, where $5^{b^2 + 4b}$ is the dimension of the space on which the bond projections of $H_R$ act nontrivially. The coupling argument below requires very large blocks. This is why the bound on $g_b$ is extremely small, and it is the source of the triple exponential in \zcref{t:main}. \section{From a window estimate to the gap of the block dynamics}\label{s:ov-block} To bound the gap of the block dynamics, we use a standard contraction argument. Suppose that two boundary conditions of a box of side $b$ differ at a single site, and that the two conditional laws in the box can be coupled so that the expected number of sites where they disagree is small compared with $b$. Then one update of a uniformly chosen block contracts the Lipschitz seminorm with respect to the Hamming distance, and this gives a gap of order one for the block dynamics. After a change of variables that reverses the spins on one sublattice of the box, the conditional laws in a box have the weight $1 + n_x \cdot n_y$ for each bond inside the box and $1 + n_x \cdot \eta_y$ for each bond to a boundary spin $\eta_y$. We work in this form from now on. We build the coupling by following the proof of \cite[Proposition~2.1]{MOS94}. Let $y$ be the site where the boundary conditions differ, and let the box have side $b = 2M + 1$. We call the intersections of the box with the boundaries of the squares of radius $\ell, 2\ell, 3\ell, \dots$ centered at $y$ the \emph{shells}, where $\ell = M^{1/4}$. We try to make the two configurations agree on one shell after another. Once they agree on a shell, the two conditional laws beyond the shell coincide. We then resample the region beyond the shell identically, so that all later disagreements stay within the shell. The two configurations agree on a shell with probability bounded below by a constant $\delta > 0$, provided two requirements hold. We call the part of a shell at $\ell^\infty$ distance more than $24$ from the complement of the box its \emph{window}. The first requirement is the \emph{window estimate}, which states that the law of the window has a component of mass $p_0 > 0$ that is the same for all boundary conditions. The second requirement is that the configurations on the two ends of the shell, at most $192$ sites near the boundary of the box, can be made to agree with probability bounded below. For a finite spin space with a uniform positive lower bound on the conditional weights, the second requirement follows from that lower bound. Here the densities can vanish, and we use an elementary overlap bound (\zcref{l:overlap}), which shows that two conditional laws of $K$ sites overlap in mass at least $(16 K)^{-4K}/2$, uniformly in the boundary conditions. With about $M^{1/8}$ trials, the expected number of disagreements is bounded by a constant times $M^{3/4}$ for large $M$, which is small compared with $b$. The window estimate asks only for a common component of positive mass, uniformly in all boundary conditions, and not for a common component of mass close to one. This is what makes it accessible to the arguments of the next two steps, which give a mass $p_0$ that is positive but extremely small. \section{The cut lemma}\label{s:ov-cut} The window estimate is still a statement about all boundary conditions, and the influence of an arbitrary boundary condition on the square lattice is the difficulty that was left open in \cite{KLT88}. It is natural to change the boundary spins one at a time and to add up their influences on the window. However, this requires the influence of a single boundary spin to be summable over the boundary, which is the kind of decay that we are trying to prove. We avoid this by changing the region instead of the boundary condition. We call a unit vector attached to a bond that leaves a finite set $G$ a \emph{pin}. A pin $\tau$ on the bond from $x \in G$ contributes the factor $1 + n_x \cdot \tau$, and we describe boundary conditions by pins. We say that a site of $G$ is \emph{removed} when its spin and the factors of all its bonds are deleted. Removed sites do not become pins. We call a connected set $X$ of sites that separates the window from the boundary of the box a \emph{cut}. The density of the window spins is the product of their internal bond weights and the partition function on the part of the region outside the window, where the fixed window spins act as additional pins. The normalizing constant is the partition function of the entire region. When the cut is removed, the partition function on the complement factorizes into a part inside the cut, which depends on the window but not on the boundary condition, and a part outside, which depends on the boundary condition but not on the window. We call the change in the logarithm of the ratio of the partition functions with and without pins, when the cut is removed, the \emph{cut cost}. By the factorization, the density of the window law is, up to normalization, a density independent of the boundary condition times $e^{-\mathrm{cost}_X}$, where $X$ is the cut. A bound on the cut cost by a constant $C$, uniform in the boundary condition and the window configuration, gives the window estimate with $p_0 = e^{-2C}$ (\zcref{l:cut}). In particular, the cost need not be small. Removing a site is a soft modification compared with pinning, because its effect is local and explicit. Let $x$ be a site incident to no pins and with $k \le 3$ remaining neighbors $u_1, \dots, u_k$. Integrating out the spin at $x$ with the moments of the uniform measure on $\Sph$, we find that \begin{equation}\label{e:ov-delete} Z(\text{with } x) = Z(\text{without } x) \, \E_{\text{without } x} \Bigl[ 1 + \frac13 \sum_{a < a'} n_{u_a} \cdot n_{u_{a'}} \Bigr], \end{equation} both with and without pins, where $Z$ denotes the partition function and $\E$ the corresponding expectation (\zcref{l:delete}). We call the difference between the correlations $\E[n_{u_a} \cdot n_{u_{a'}}]$ with and without pins the \emph{response} of the pair $(u_a, u_{a'})$. By \eqref{e:ov-delete}, the pins change the cost of removing $x$ only through the responses of the pairs of neighbors of $x$. Each such pair is at distance $2$ and has the removed site $x$ as a common neighbor. We remove the sites of the cut one at a time, each of them, except the first, next to a site removed earlier. Then every site except the first has at most three remaining neighbors when it is removed, and the cost of its removal is bounded by responses. The first removal costs at most a constant. It remains to show that the responses decay exponentially in the distance to the pins, uniformly in the pins and in the set of removed sites, which is Hypothesis~(P) of \zcref{s:cut}. Given this decay, a cut has bounded cost if it keeps a fixed distance from the pins near the ends of the window and moves away from them linearly elsewhere (see \zcref{f:cut}). \section{Responses and walk expansions}\label{s:ov-walks} Recall that a response is the difference between a correlation with pins and the same correlation without pins. We call a walk on the lattice that uses each bond at most once a \emph{trail}. To prove the decay of responses, we expand a correlation $\E[n_u \cdot n_v]$ into trails. The expansion rests on the identity \begin{equation}\label{e:ov-step} \int (e \cdot n) \prod_{i = 1}^k (1 + n \cdot w_i) \dd n = \frac13 \sum_{i = 1}^k (e \cdot w_i) \int \prod_{i' \neq i} (1 + n \cdot w_{i'}) \Bigl( \frac35 \dd n + \frac25 \delta_0 \Bigr), \end{equation} valid for vectors $e, w_1, \dots, w_k \in \R^3$ with $k \le 4$, where $\delta_0$ denotes the point mass at $0 \in \R^3$; it follows by comparing the moments of both sides. Consider the correlation of $e \cdot n_u$ with a function of the other spins. Integrating out $n_u$ with \eqref{e:ov-step} replaces $e \cdot n_u$ by $(e \cdot n_{u'})/3$, summed over the neighbors $u'$ of $u$. It also uses up the bond $\{u, u'\}$ and gives $u$ the modified measure in parentheses, whose moments of positive degree are $3/5$ times those of $\dd n$. Repeating the step at $u'$ and so on, we obtain a sum over trails from $u$ that stop at $v$ or in the boundary layer $\partial_2 G$. A trail of length $m$ carries the weight $3^{-m}$, an extra factor $3/5$ for each site it leaves for the second time, and the ratio of the partition functions after and before the trail is traced. On bipartite tori and with N\'eel boundary conditions, the expansions in \cite{KLT88} have nonnegative terms. The authors bound the two-point function on the torus and the boundary-induced staggered magnetization by weighted numbers of bond-self-avoiding walks, obtaining exponential decay. With arbitrary pins, the terms have no sign, and the ratio of partition functions along a trail can exceed one. To overcome this obstacle, we compare the expansions with and without pins trail by trail. They have the same trails and the same weights $3^{-m} (3/5)^r$, where $r$ is the number of sites that the trail leaves for the second time. The ratios of partition functions can differ, as can the terminal correlations of trails that end in the boundary layer. Without pins, every term is nonnegative and every ratio is bounded by one. At a step that leaves a site for the second time, the ratio equals one with and without pins. At a step that leaves a site for the first time, we show that the ratio with pins is bounded by $1/0.961$ when the trail goes straight through the site and by $1/0.935$ when it turns. These bounds are the local inequalities of \zcref{s:ov-computer}. Given them, the trails that end in $\partial_2 G$ contribute an amount that is exponentially small in the distance to $\partial_2 G$, provided the weighted number of trails of length $m$, with the step factors above, decays exponentially in $m$. We prove this decay with the rate $2499/2500$, which leaves very little room. Next, we treat the trails that return to $v$. We write the difference of the two products of ratios along such a trail as a sum over its steps. At one step, the difference of the logarithms of the two ratios is a combination, with coefficients at most $1/6$ and $9/25$, of the responses of the pairs of remaining neighbors of the site that the trail leaves, computed in the region where that site is removed (\zcref{l:response}). These pairs are again at distance $2$ with a removed common neighbor, and their responses are of the same kind as the one we started from. For an integer $d$, let $r(d)$ denote the largest absolute value of the response of a pair of this kind at distance at least $d$ from $\partial_2 G$. We obtain a linear inequality of the form \begin{equation}\label{e:ov-contraction} r(d) \le \sum_{w} (\text{weight of } w) \, r \bigl( d - |w| \bigr) + C \gamma^d, \end{equation} where the sum is over the trails $w$ from $u$ that return to $v$, $|w|$ denotes the length of $w$, $\gamma < 1$ is the rate of decay of the weighted number of trails, and $C$ is a constant. There is an apparent circularity here, since we use responses to bound responses. The way out is to work in a finite set $G$ and to consider the supremum of all responses in $G$, weighted by $\theta^{-d}$ for a rate $\theta < 1$ and by a weight for each of the two shapes of pairs. (We call a pair \emph{collinear} if its two sites and their common neighbor lie on a line, and \emph{diagonal} otherwise.) The supremum is over all regions obtained from $G$ by removing sites and tracing trails, and over all pins. Since $G$ is finite, this supremum is finite. Removed sites never become pins, so the distance to the boundary of $G$ is the same for all these regions. The inequality \eqref{e:ov-contraction} then shows that the supremum is at most a constant $K_\theta$ times itself plus a constant independent of $G$. Therefore the supremum is bounded uniformly in $G$ as soon as $K_\theta < 1$. The constant $K_\theta$ is the larger of two weighted sums, one for each shape of pairs, over the trails from $u$ to $v$ in $\Z^2$ that avoid the common neighbor of $u$ and $v$, in which each trail carries its weight from the expansion, the response coefficients at its first visits, and the step factors of the local inequalities. By continuity in $\theta$, it suffices to prove $K_1 < 1$. With the weights $1$ and $5/4$ for collinear and diagonal pairs, we prove that the sums for the two shapes are at most $0.9232$ and $0.8994$. The margin is small, and it requires the exact geometry of the trails that return. Our bound for the weighted number of all trails of length $m$ is $10 \cdot (2499/2500)^m$, whose sum over $m$ is $25{,}000$. This bound on all trails is too large to prove $K_1 < 1$. \section{The computer-assisted inputs}\label{s:ov-computer} Three statements are verified by computer. The second and the third use the first only through the two step factors $1/0.961$ and $1/0.935$, and the third also uses the response coefficients $1/6$ and $9/25$. The first is the family of local inequalities. When a trail leaves a site $x$ for the first time, we bound the ratio of the pinned partition functions after and before the step by an inequality on the sites at distance at most $2$ from $x$. The local pattern records the site statuses and the used bonds at these sites. After the spin at $x$ and the spins at the four sites at distance $2$ along the axes are integrated out, the rest of the system appears only through at most twelve vectors of norm at most one. This leaves $7{,}962$ local configurations, after the reflection and identifications of \zcref{s:local-computation}, and for each of them a polynomial inequality, affine in each of at most eleven unit vectors. We certify each inequality by an explicit Hermitian operator on $(\C^2)^{\otimes k}$ whose expectations in products of spin-$\frac12$ coherent states are the values of the polynomial. After the signs of a suitable subset of the vectors are reversed, positivity of this operator implies the inequality. The operator commutes with simultaneous rotations, so in each exchange-symmetry sector it suffices to test one eigenspace of the total spin component. Each test is the positive definiteness of a symmetric integer matrix of size at most $245$, which we verify by fraction-free Gaussian elimination \cite{Bareiss68} in exact arithmetic. The second is the exponential decay of the weighted number of all trails. We verify it with a positive integer vector $\mathsf v$ on the $916$ possible sequences of the last six steps of a trail. The weighted transfer matrix $\mathsf A_6$ satisfies $\mathsf A_6 \mathsf v \le \gamma_{\mathrm o} \mathsf v$ entrywise, as we check in exact integer arithmetic. The third is the bound $K_1 < 1$. We enumerate the trails of length at most $20$ exactly. For a longer trail, we follow the first $20$ steps exactly and dominate the continuation by a weighted walk on $\Z^2$ that remembers its position and its last $10$ steps. We bound the total weight of its paths to $v$ by a function of its states that satisfies a one-step inequality, and we check this inequality state by state in floating-point arithmetic with directed rounding. The programs and their outputs are provided as ancillary files, which we describe in \zcref{s:ancillary}. \section{Organization of the paper}\label{s:ov-organization} \zcref[S]{s:quantum} recalls the polynomial representation and the VBS state and proves \zcref{t:comparison}, the comparison between the quantum gap and the gap of the block dynamics. In \zcref{s:block}, we prove that the window estimate implies a gap for the block dynamics, through the coupling argument of \cite{MOS94}. \zcref[S]{s:cut} introduces the model with pins and removed sites, proves the cut lemma, and deduces the window estimate from the decay of responses. \zcref[S]{s:walks} sets up the walk expansion and states the local inequalities and the decay of weighted trail counts as hypotheses, and \zcref{s:contraction} proves the decay of responses from these hypotheses and the bound on returning trails. \zcref[S]{s:computer} verifies the three hypotheses with computer assistance and completes the proof of \zcref{t:main}. \zcref[S]{s:topology} proves that the cut is connected, \zcref{s:local,s:tail} contain the certificates for the local inequalities and for the returning trails, \zcref{s:explicit} tracks the constants, \zcref{s:ancillary} describes the ancillary files, and \zcref{s:uniqueness} proves uniqueness of the infinite-volume frustration-free state.