\chapter{Introduction}\label{s:intro} \section{Background}\label{s:intro-background} Haldane predicted that the antiferromagnetic Heisenberg chain has a spectral gap above its ground state when the spin per site is an integer \cite{Haldane83a,Haldane83b}. The first rigorous example of an isotropic model with this behavior was found by Affleck, Kennedy, Lieb, and Tasaki \cite{AKLT87,AKLT88}. They constructed a spin-$1$ chain whose ground state is built from valence bonds, and they proved that it has a unique ground state, a spectral gap, and exponentially decaying correlations. The same construction gives a rotation-invariant, frustration-free Hamiltonian on any lattice, with spin equal to half the number of neighbors at each site, and an explicit ground state of this Hamiltonian, the \emph{valence-bond-solid} (VBS) state. These Hamiltonians are now called AKLT models. In \cite{AKLT88}, models in two and higher dimensions were presented that ``are expected to have the same properties'' as the chain. In particular, the AKLT models on the hexagonal lattice, with spin $3/2$, and on the square lattice, with spin $2$, are expected to have a unique ground state in the thermodynamic limit and a spectral gap (see \cite[Section~1.1]{JNY26} for this reading of \cite{AKLT88}). Beyond their role in the study of Haldane's prediction, two-dimensional AKLT states have been studied as resources for measurement-based quantum computation \cite{WAR11}. Kennedy, Lieb, and Tasaki \cite{KLT88} proved that the spin-$3/2$ model on the hexagonal lattice has a unique ground state in the infinite-volume limit, in which all truncated correlation functions decay exponentially. They also characterized the finite-volume ground states of the VBS models in every dimension, and they proved that the two-point function of the spin-$2$ model on the square lattice decays exponentially on bipartite tori. Their bound on the staggered magnetization with N\'eel boundary conditions decays exponentially in the distance to the boundary, with a prefactor proportional to the boundary size. These results use the representation of the VBS state, due to Arovas, Auerbach, and Haldane \cite{AAH88}, as the Boltzmann weight of a classical model of unit vectors with a factor $1 - s_x \cdot s_y$ for each bond. On the square lattice, the proof in \cite{KLT88} expands correlations into walks whose weights are nonnegative on bipartite tori and for N\'eel boundary conditions (in which the boundary spins are fixed in a staggered pattern), and it bounds them by a weighted count of bond-self-avoiding walks. Kennedy, Lieb, and Tasaki pointed out that the exponential decay for other boundary conditions and the uniqueness of the infinite-volume ground state on the square lattice were open questions. Jackson, Nachtergaele, and Young \cite[Section~1.2]{JNY26} left open the question of uniqueness among frustration-free states. The spectral gap of two-dimensional AKLT models has been studied mostly through \emph{finite-size criteria}, which bound the gap of a frustration-free Hamiltonian from below, uniformly in the volume, once the gap of a fixed finite system exceeds an explicit threshold \cite{Knabe88,GossetMozgunov16,LemmMozgunov19,Lemm20}. For the hexagonal lattice, Lemm, Sandvik, and Wang \cite{LSW20} and Pomata and Wei \cite{PomataWei20} combined such criteria with numerical computations for finite systems, by the density matrix renormalization group and by exact and Lanczos diagonalization, respectively, and obtained a gap. However, the numerical steps of these works are not certified, and Lemm, Sandvik, and Wang state that their result ``may not be considered a rigorous mathematical proof as a matter of principle'' \cite{LSW20}. Recently, Kastoryano \cite{Kastoryano26} gave an analytic proof of a uniform gap for the hexagonal model on a family of rectangular regions, through the boundary-state criterion of Kastoryano, Lucia, and P\'erez-Garc\'ia \cite{KLPG19}, with a scalar estimate verified in the proof assistant Lean. For decorated lattices, in which the bonds of a lattice are replaced by chains of spin-$1$ sites, gaps were proved in \cite{ALLNY20,PomataWei19,LuciaYoung23}, and gaps for two further planar lattices with spin $2$ at every site were obtained in \cite{GPW21}. Jackson, Nachtergaele, and Young \cite{JNY26} proved the uniqueness of the frustration-free infinite-volume state, local indistinguishability of finite-volume ground states, and the stability of the gap for the hexagonal lattice and for the decorated square lattices (see also \cite{LMY24}). For the spin-$2$ model on the square lattice, the existence of a spectral gap has remained open \cite{MKL20,GPW21}. Numerical studies indicate a gap of order $10^{-2}$ \cite{GMW13}, and Martyn, Kato, and Lucia \cite{MKL20} gave numerical evidence for the boundary-state criterion of \cite{KLPG19}. In \cite{GPW21}, the gap on the square lattice is reduced to a finite-size problem for a weighted Hamiltonian, which was left open in that work. The goal of this paper is to prove that the spin-$2$ AKLT model on the square lattice has a spectral gap bounded below uniformly when both side lengths of the torus are sufficiently large. The proof is computer-assisted. \section{Main result}\label{s:intro-result} We now define the model. For integers $L, N \ge 3$, let $\Lambda = \Lambda_{L, N} = \Z_L \times \Z_N$ denote the discrete torus, and let $\mathcal E$ denote its set of nearest-neighbor bonds. At every site $x \in \Lambda$ we place a spin $2$, that is, a copy $V_x$ of the five-dimensional irreducible representation of $\mathrm{SU}(2)$, and we let $\mathcal H_\Lambda = \bigotimes_{x \in \Lambda} V_x$. For a bond $e = \{x, y\} \in \mathcal E$, let $P_e$ denote the orthogonal projection of $V_x \otimes V_y$ onto its subspace of total spin $4$, extended by the identity to $\mathcal H_\Lambda$. We call \begin{equation}\label{e:H} H_{L, N} = \sum_{e \in \mathcal E} P_e \end{equation} the \emph{spin-$2$ AKLT Hamiltonian} on $\Lambda_{L, N}$. For a nonnegative self-adjoint operator $H$ on a finite-dimensional space that is not zero, we let $\gap(H)$ denote its smallest nonzero eigenvalue. Our main result is the following theorem. \begin{theorem}\label{t:main} Set \begin{equation}\label{e:mainconstants} L_0 = 2 \bigl\lceil \exp(10^{40}) \bigr\rceil + 5, \qquad c = \exp \Bigl[ -\exp \bigl( \exp(10^{41}) \bigr) \Bigr]. \end{equation} For all integers $L, N \ge L_0$, the kernel of $H_{L, N}$ is one-dimensional and \begin{equation*} \gap(H_{L, N}) \ge c. \end{equation*} \end{theorem} The kernel of $H_{L, N}$ is spanned by the VBS state for all $L, N \ge 3$. This follows from the divisibility argument of \cite[Sections~1--2]{KirillovKorepin89} (see also \cite{KLT88} and \cite[Theorem~2.1]{JNY26}), and we include the short proof in \zcref{s:quantum}, since our comparison argument rests on the same divisibility property. The content of \zcref{t:main} is therefore the lower bound on the gap, which is uniform in the size of the torus and holds for both parities of $L$ and $N$. (For odd $L$ or $N$, the torus is not bipartite, and the classical measure that we use is frustrated; our arguments use the bipartite structure only locally.) We have made no attempt to optimize the constants in \zcref{t:main}. They are far from the expected size of the gap, and they reflect the very large blocks that our argument requires. The proof uses three statements that we verify with the help of a computer: a family of $7{,}962$ polynomial inequalities, which we verify by testing the positive definiteness of integer matrices in exact arithmetic; an exponential bound on a weighted number of walks, which we verify with an integer vector in exact arithmetic; and a bound below one on a weighted sum over returning walks, which we verify in floating-point arithmetic with directed rounding. The programs, the certificates, and the outputs of the computations are included as ancillary files and described in \zcref{s:ancillary}. The rest of the proof is written out in full. The same coupling and coherent-state representation also give uniqueness of the infinite-volume frustration-free state on the square lattice. We prove this in \zcref{s:uniqueness}. \section{Methods and relation to previous work}\label{s:intro-methods} We give an overview of the proof in \zcref{s:overview}. We now describe how its ingredients relate to the literature. Our starting point, proved in \zcref{s:quantum}, is a comparison of the gap of $H_{L, N}$ with the gap of a classical Markov process. The VBS state, written in spin coherent states \cite{Radcliffe71,Perelomov72,Lieb73}, defines a probability measure on configurations of unit vectors, and this is the classical model of \cite{AAH88,KLT88}. We show that the gap of $H_{L, N}$ is bounded below by the spectral gap of a dynamics that resamples large square blocks from their conditional laws under this measure, times $b g_b/(b+1)$, where $b$ is the side length of the block and $g_b$ is the smallest positive eigenvalue of the sum of the bond projections over the bonds that meet one block. Correspondences between classical dynamics satisfying detailed balance and frustration-free quantum Hamiltonians are known for Rokhsar--Kivelson-type Hamiltonians \cite{Henley04,CCMP05,VWPGC06}. In these works, the quantum Hamiltonian is constructed from the classical dynamics on a discrete state space, and its excitation spectrum is that of the classical dynamics. In our setting, however, the Hamiltonian is given, the coherent states are not orthogonal, and the comparison is an inequality, which rests on the divisibility of the local ground states by the bond singlets. We are not aware of an earlier comparison of this kind for AKLT models. In contrast with finite-size criteria, the smallest positive eigenvalue of this part of $H_{L, N}$ enters only through its positivity, so no threshold has to be met by a finite quantum system. The spectral gap of Glauber dynamics for classical spin systems is closely tied to spatial mixing; see \cite{DS85,DS87,MO94,Martinelli99}. In two dimensions, Martinelli, Olivieri, and Schonmann \cite{MOS94} proved, for finite-range systems with a finite single-spin space, that weak mixing implies strong mixing on squares. The heart of their proof is a coupling that, after a change of the boundary condition at one site, makes the two configurations agree on successive shells around that site, with a probability bounded below at each shell. In \zcref{s:block}, we follow the proof of \cite[Proposition~2.1]{MOS94} for our continuous spins. In their proof, the finiteness of the spin space and the positivity of the weights give a lower bound for the probability that the two configurations agree near the boundary of the box. We replace this bound by an elementary overlap estimate for densities that may vanish. The step from a contracting coupling to a spectral gap is the standard argument of Chen \cite{Chen98} (see \cite[Theorem~13.1]{LPW17} for finite state spaces, and \cite{Ollivier09} for general ones), which we write out for our block dynamics. The input of the coupling argument is a statement about the laws of shell segments far from the boundary of the box, uniformly in all boundary conditions. Uniformity in the boundary condition is the difficulty that was left open in \cite{KLT88} for the square lattice. We obtain it in \zcref{s:cut} through a cut lemma: once we remove sites that separate the segment from the boundary, the law of the segment no longer depends on the boundary condition, and the cost of the removal is controlled by the changes that the boundary condition causes in certain pair correlations, which we call \emph{responses}. In \zcref{s:walks,s:contraction}, we bound the responses through an expansion of correlations into bond-self-avoiding walks (trails), of the type used in \cite{KLT88}. With arbitrary boundary conditions, the terms of this expansion have no sign. Instead of positivity, we compare, for each walk, its contributions with and without boundary conditions, and the differences are again responses. The resulting linear inequality for the responses closes as a contraction in a weighted supremum norm, provided a weighted sum over returning trails is smaller than one. The three computer-assisted statements are proved in \zcref{s:computer} and \zcref{s:local,s:tail}. Each polynomial inequality is verified through an explicit Hermitian operator whose expectations in products of spin-$\frac12$ coherent states reproduce the polynomial; its positivity reduces to the positive definiteness of integer matrices, which we test by fraction-free elimination \cite{Bareiss68}. The bound on the weighted number of trails uses a transfer matrix over the last six steps of a trail, in the spirit of the finite-memory bounds for self-avoiding walks \cite{Alm93,PonitzTittmann00}, with an exact integer certificate. The bound on returning trails combines an exact enumeration of short trails with a function that satisfies a one-step inequality for a walk that dominates the longer ones, and this inequality is verified in floating-point arithmetic with directed rounding \cite{IEEE754} (see \cite{Tucker11} for rigorous computation with directed rounding). In each case we state precisely what the programs compute and what a reader must trust. \section{Notation}\label{s:intro-notation} We write $x, y, z, u, v$ for sites of $\Z^2$ and $d(x, y) = |x_1 - y_1| + |x_2 - y_2|$ for the graph distance between them, and we let $|x - y|_\infty = \max(|x_1 - y_1|, |x_2 - y_2|)$ denote their $\ell^\infty$ distance. We write $x \sim y$ if $d(x, y) = 1$, and we call a set $\{x, y\}$ with $x \sim y$ a \emph{bond}. For $A \subset \Z^2$ and a site $x$, we let $\dist(x, A)$ and $\dist_\infty(x, A)$ denote the minimum of $d(x, a)$ and of $|x - a|_\infty$ over $a \in A$, with both distances equal to $+\infty$ when $A$ is empty. We use the same notation on a torus $\Z_L \times \Z_N$, with the distances taken in the torus. We let $|A|$ denote the cardinality of a finite set $A$, and we set $(t)_+ = \max(t, 0)$ for $t \in \R$. For $a, b \in \R^3$, we let $a \cdot b$ denote the Euclidean inner product and $|a|$ the Euclidean norm. We let $\Sph$ denote the unit sphere of $\R^3$, $\dd n$ the uniform probability measure on $\Sph$, and $\Ball$ the closed unit ball of $\R^3$. We call the elements of $\Sph$ \emph{spins}, and we write $n = (n_x)_x$ for a configuration of spins. For a probability measure $\mu$ and an integrable function $f$, we write $\mu(f) = \int f \dd\mu$. Let $\mu$ be a probability measure on a measurable space, and let $\mathcal L$ be a bounded, self-adjoint, nonpositive operator on $L^2(\mu)$ that annihilates the constant functions, such as the generator of a Markov dynamics reversible with respect to $\mu$ with bounded jump rates. We call \begin{equation*} \gap(\mathcal L) = \inf \biggl\{ \frac{\langle F, -\mathcal L F \rangle_{L^2(\mu)}}{\Var_\mu(F)} : F \in L^2(\mu) \text{ not constant} \biggr\} \end{equation*} the \emph{spectral gap} of $\mathcal L$, where $\Var_\mu(F) = \mu(|F - \mu(F)|^2)$. All constants in this paper are explicit.