\chapter{Certificates for the local inequalities}\label{s:local} In this appendix, we prove \zcref{t:H1}. Recall that it asserts that \begin{equation}\label{e:local-goal} \mathcal Z_\pi^{\mathbf a}[I_{\mathrm{cur}} - c_\pi I_{\mathrm{next}}] \ge 0 \qquad \text{for all } \pi \in \Pext \text{ and } \mathbf a \in \Ball^{12}, \end{equation} where $c_\pi = 961/1000$ if $\pi$ is straight and $c_\pi = 187/200$ if $\pi$ is a turn. This is Hypothesis~(H1) at the thresholds $(\cs, \ct) = (961/1000, 187/200)$, one of the three inputs of \zcref{t:inputs}, and we refer to \zcref{s:overview} for the role of these inequalities in the argument. The proof is computer-assisted and uses exact tests of integer matrices. We begin in \zcref{s:local-ring} by integrating out the corners with at most one present ring bond. This reduces \eqref{e:local-goal} to the nonnegativity of a polynomial in at most eleven unit vectors that is affine in each of them. The main idea, in \zcref{s:local-inversion}, is to write this polynomial as the expectation of an explicit Hermitian operator in a product of spin-$1/2$ coherent states, after the signs of some of the vectors have been reversed. The symmetries of the operator then reduce its positivity to a few eigenspaces of the total spin component, and in \zcref{s:local-integer} we turn each of these tests into a test of positive definiteness of a symmetric integer matrix. In \zcref{s:local-computation}, we describe the programs and the certificate file and complete the proof of \zcref{t:H1}. The checks in \zcref{s:local-checks} corroborate the computation but are not part of the proof, and \zcref{s:local-trust} lists what a reader must trust. Throughout this appendix, we fix a pattern $\pi \in \Pext$ and a rational number $c = c_1/c_2$, where $c_1$ and $c_2$ are positive integers; for \zcref{t:H1}, we take $c = c_\pi$. We use the notation of \zcref{s:walks} for the template $\tmpl(0)$, its middle site $0$, its ring $\mathsf R$, which consists of the axial sites $\sE$, $\sN$, $\sW$, and $\sS$ and the corners $(\pm 1, \pm 1)$, its tips, and its slots. Let $o = o_\pi$ be the out-neighbor of $\pi$, and let $f$ and $g$ be the two sites of $\{\sE, \sN, \sW\} \setminus \{o\}$. For a non-removed ring site $y$, we abbreviate $\lambda_y = \lambda_{\mathrm{st}_\pi(y)}$. Superscripts on a vector $b \in \R^3$ denote its coordinates, $b = (b^1, b^2, b^3)$, and we set $b^0 = 1$. For $Q \in \mathrm O(3)$ and $\mathbf b = (b_1, \dots, b_k) \in (\R^3)^k$, we set $Q \mathbf b = (Q b_1, \dots, Q b_k)$. \section{The ring polynomial and leaf corners}\label{s:local-ring} Recall from \zcref{s:walks} that every ring bond joins an axial site to a corner; in particular, the corners are pairwise non-adjacent. Recall also that the integrand of the ring functional \eqref{e:ringfunctional} is the product of $h$, of the factors $1 + n_y \cdot n_z$ of the present ring bonds $\{y, z\}$, of the factors $1 + (3/5) u_z \cdot n_z$ of the axial sites $z$ whose tip bond is retained, and of the two slot factors $1 + m_k \cdot n_{z(k)}$ of every non-removed corner $z(k)$, and that it is integrated against the measures $\lambda_y(\mathrm d n_y)$ of the non-removed ring sites $y$. Note that in $I_{\mathrm{cur}}$ and $I_{\mathrm{next}}$, the spin $n_f$ is replaced by $0$ if $f$ is removed, and likewise for $g$. We define the \emph{ring polynomial} \begin{equation}\label{e:local-Phi} \Phi_c(\mathbf a) = \mathcal Z_\pi^{\mathbf a}[I_{\mathrm{cur}} - c I_{\mathrm{next}}], \qquad \mathbf a \in (\R^3)^{12}, \end{equation} where for $\mathbf a \notin \Ball^{12}$ the right side is defined by the formula \eqref{e:ringfunctional}. In this notation, \zcref{t:H1} asserts that $\Phi_{c_\pi} \ge 0$ on $\Ball^{12}$ for all $\pi \in \Pext$. \begin{lemma}\label{l:ringpoly} Let $c \in \R$. Then $\Phi_c$ is a polynomial in the coordinates of $\mathbf a$ that is affine in each of the twelve vectors of $\mathbf a$, and $\Phi_c(Q \mathbf a) = \Phi_c(\mathbf a)$ for all $Q \in \mathrm O(3)$ and $\mathbf a \in (\R^3)^{12}$. \end{lemma} \begin{proof} Each vector of $\mathbf a$ occurs in at most one factor of the integrand in \eqref{e:ringfunctional}, which is affine in it, and $I_{\mathrm{cur}}$ and $I_{\mathrm{next}}$ do not involve $\mathbf a$. Expanding the integrand, we see that $\Phi_c(\mathbf a)$ is a finite sum of moments of the product measure $\prod_y \lambda_y$, each multiplied by a product of coordinates of the vectors of $\mathbf a$ that contains at most one coordinate of each vector. The first assertion follows. To prove the invariance, let $Q \in \mathrm O(3)$, let $\mathbf n$ denote the family of spins of the non-removed ring sites, and let $\mathcal G(\mathbf n, \mathbf a)$ denote the integrand of \eqref{e:ringfunctional} with $h = I_{\mathrm{cur}} - c I_{\mathrm{next}}$. Every factor of $\mathcal G$ is a function of scalar products between spins and vectors of $\mathbf a$, so that $\mathcal G(Q \mathbf n, Q \mathbf a) = \mathcal G(\mathbf n, \mathbf a)$. Using the invariance of $\lambda_0$ and $\lambda_1$ under $Q$ and changing variables $n_y \mapsto Q n_y$ for every non-removed ring site $y$, we obtain \begin{equation*} \Phi_c(Q \mathbf a) = \int \mathcal G(\mathbf n, Q \mathbf a) \prod_y \lambda_y(\mathrm d n_y) = \int \mathcal G(Q \mathbf n, Q \mathbf a) \prod_y \lambda_y(\mathrm d n_y) = \Phi_c(\mathbf a). \qedhere \end{equation*} \end{proof} We will use the following elementary fact several times. \begin{lemma}\label{l:affine} Let $k \ge 1$, and let $f$ be a real polynomial on $(\R^3)^k$ that is affine in each of its $k$ arguments. If $f \ge 0$ on $(\Sph)^k$, then $f \ge 0$ on $\Ball^k$. \end{lemma} \begin{proof} We show by induction on $l \in \{0, \dots, k\}$ that $f \ge 0$ on $\Ball^l \times (\Sph)^{k - l}$. The case $l = 0$ is the assumption. Suppose that the claim holds for $l - 1$, and let $\mathbf b \in \Ball^l \times (\Sph)^{k - l}$. Choose a unit vector $e$ such that $b_l = |b_l| e$, and let $\mathbf b_+$ and $\mathbf b_-$ be obtained from $\mathbf b$ by replacing $b_l$ with $e$ and $-e$, respectively. Since $f$ is affine in its $l$-th argument and \begin{equation*} b_l = \frac{1 + |b_l|}{2} \, e + \frac{1 - |b_l|}{2} \, (-e), \end{equation*} we have \begin{equation*} f(\mathbf b) = \frac{1 + |b_l|}{2} f(\mathbf b_+) + \frac{1 - |b_l|}{2} f(\mathbf b_-) \ge 0, \end{equation*} where the inequality follows from the induction hypothesis, applied to $\mathbf b_+$ and $\mathbf b_-$. \end{proof} We now integrate out the spin of a corner that has at most one present ring bond. \begin{lemma}\label{l:leaf} Let $s \in \{0, 1\}$ and $m, m' \in \Sph$, and set $r = 1/2$ if $s = 0$ and $r = 1/3$ if $s = 1$. Then \begin{equation}\label{e:local-leafa} \int (1 + m \cdot n)(1 + m' \cdot n) \, \lambda_s(\mathrm d n) > 0, \end{equation} and there exist $A > 0$ and $b \in \Ball$, depending only on $s$, $m$, and $m'$, such that \begin{equation}\label{e:local-leafb} \int (1 + n' \cdot n)(1 + m \cdot n)(1 + m' \cdot n) \, \lambda_s(\mathrm d n) = A (1 + r \, b \cdot n') \qquad \text{for all } n' \in \R^3. \end{equation} \end{lemma} \begin{proof} Set $t = m \cdot m'$. Recall that the moments of $\dd n$ of odd degree vanish and that \begin{equation*} \int (a \cdot n)(a' \cdot n) \dd n = \frac{a \cdot a'}{3}, \qquad a, a' \in \R^3. \end{equation*} Expanding the products and using these facts, we find for all $n' \in \R^3$ that \begin{equation}\label{e:local-leaf0} \int (1 + m \cdot n)(1 + m' \cdot n) \dd n = A_0, \qquad \int (1 + n' \cdot n)(1 + m \cdot n)(1 + m' \cdot n) \dd n = A_0 + V \cdot n', \end{equation} where \begin{equation*} A_0 = 1 + \frac t3, \qquad V = \frac{m + m'}{3}. \end{equation*} Since $\lambda_1 = (3/5) \dd n + (2/5) \delta_0$, the same two integrals with $\lambda_1$ in place of $\dd n$ are \begin{equation}\label{e:local-leaf1} A_1 \quad \text{and} \quad A_1 + \frac35 V \cdot n', \qquad \text{where } A_1 = \frac{3 A_0 + 2}{5}. \end{equation} The bound $t \ge -1$ implies that $A_0 \ge 2/3$ and $A_1 \ge 4/5$, and \eqref{e:local-leafa} follows. We now prove \eqref{e:local-leafb}. Using $|m + m'|^2 = 2 + 2t$, which holds since $m$ and $m'$ are unit vectors, we find \begin{equation*} A_0^2 - 4 |V|^2 = \Big( 1 + \frac t3 \Big)^2 - \frac{8 + 8t}{9} = \frac{(1 - t)^2}{9} \ge 0, \end{equation*} so that $|V| \le A_0/2$. For $s = 0$, we set \begin{equation*} A = A_0, \qquad b = \frac{2V}{A_0}. \end{equation*} Then $|b| \le 1$, and \eqref{e:local-leaf0} yields \eqref{e:local-leafb} with $r = 1/2$. For $s = 1$, we set \begin{equation*} A = A_1, \qquad b = \frac{9V}{3 A_0 + 2}, \qquad \text{so that} \qquad A r b = \frac{3 A_0 + 2}{15} \cdot \frac{9V}{3 A_0 + 2} = \frac35 V \quad \text{for } r = \frac13. \end{equation*} Then \eqref{e:local-leaf1} gives \eqref{e:local-leafb}. Further, using $|V| \le A_0/2$, we have \begin{equation*} |b| \le \frac{9 A_0}{2 (3 A_0 + 2)} \le 1, \end{equation*} where the last inequality is equivalent to $A_0 \le 4/3$, which holds since $t \le 1$. \end{proof} We call a non-removed corner an \emph{isolated corner} if it has no present ring bond, and a \emph{leaf corner} if it has exactly one present ring bond. Recall that every non-removed corner carries two slot factors in \eqref{e:ringfunctional}. Using \zcref{l:leaf}, we can integrate out the spins of the isolated corners and of the leaf corners, and this leads to the following two representations of $\Phi_c$. \begin{definition}\label{d:local-rep} The \emph{unreduced representation} of $\pi$ consists of the following \emph{groups} of vectors in $\R^3$, each vector with a radius and a ring site that carries it. \begin{enumerate} \item For each non-removed axial site $z$ whose tip bond is retained, one vector, carried by $z$, with radius $3/5$. \item For each non-removed corner $z$ that is not isolated, two vectors, carried by $z$, with radius $1$. \end{enumerate} The \emph{reduced representation} of $\pi$ is obtained from the unreduced one by replacing, for each leaf corner $z$, the two vectors of $z$ by one vector, carried by the axial site that is joined to $z$ by its present ring bond, with radius $1/2$ if $z$ has status $0$ and radius $1/3$ if $z$ has status $1$. In both representations, we call a group of one vector a \emph{single vector} and a group of two vectors a \emph{pair}, and we say that the corner $z$ of item (2) carries its pair. We number the vectors $1, \dots, k$, and we let $r_i$ denote the radius of the vector $i$ and $z_i$ the site that carries it. The \emph{retained sites} are the non-removed ring sites other than the isolated corners and, in the reduced representation, the leaf corners. \end{definition} For $\mathbf b = (b_1, \dots, b_k) \in (\R^3)^k$ and a function $h$ of the spins of the retained sites, we define \begin{equation}\label{e:local-hatZ} \hat Z[h](\mathbf b) = \int h \prod_{\{y, z\}} (1 + n_y \cdot n_z) \prod_{i = 1}^k (1 + r_i \, b_i \cdot n_{z_i}) \prod_{y} \lambda_y(\mathrm d n_y), \end{equation} where the first product runs over the present ring bonds whose ends are retained and the last over the retained sites. Every non-removed axial site is retained, so that $I_{\mathrm{cur}}$ and $I_{\mathrm{next}}$ are functions of the spins of the retained sites. We set $\hat Z_{\mathrm{cur}} = \hat Z[I_{\mathrm{cur}}]$ and $\hat Z_{\mathrm{next}} = \hat Z[I_{\mathrm{next}}]$, and we define \begin{equation}\label{e:local-hatPhi} \hat\Phi_c = \hat Z_{\mathrm{cur}} - c \hat Z_{\mathrm{next}}. \end{equation} These polynomials on $(\R^3)^k$ depend on the representation, which will always be clear from the context. The proof of \zcref{l:ringpoly} applies verbatim to \eqref{e:local-hatZ}. It shows that $\hat Z_{\mathrm{cur}}$ and $\hat Z_{\mathrm{next}}$ are affine in each of the $k$ vectors and that $\hat Z[h](Q \mathbf b) = \hat Z[h](\mathbf b)$ for $h \in \{I_{\mathrm{cur}}, I_{\mathrm{next}}\}$, $Q \in \mathrm O(3)$, and $\mathbf b \in (\R^3)^k$. \begin{corollary}\label{c:reduce} Fix one of the two representations of $\pi$. If $\hat\Phi_c(\mathbf b) \ge 0$ for all $\mathbf b \in (\Sph)^k$, then $\Phi_c(\mathbf a) \ge 0$ for all $\mathbf a \in \Ball^{12}$. \end{corollary} \begin{proof} By \zcref{l:ringpoly}, the ring polynomial $\Phi_c$ is affine in each vector of $\mathbf a$. By \zcref{l:affine}, applied with $k = 12$ and $f = \Phi_c$, it suffices to show that $\Phi_c(\mathbf a) \ge 0$ for all $\mathbf a \in (\Sph)^{12}$. Fix such an $\mathbf a$. We begin by integrating out the spins of the corners that are not retained. The spin of a corner $z$ occurs in \eqref{e:ringfunctional} only in the two slot factors of $z$ and in the factors of the present ring bonds at $z$, since $I_{\mathrm{cur}}$ and $I_{\mathrm{next}}$ involve only spins of axial sites. The corners are pairwise non-adjacent, so that we may integrate out their spins one corner at a time. Let $z$ be a corner that is not retained, and let $m, m' \in \Sph$ be its slot vectors. By \eqref{e:local-leafa}, an isolated corner $z$ contributes the factor \begin{equation*} A_z = \int (1 + m \cdot n)(1 + m' \cdot n) \, \lambda_z(\mathrm d n) > 0. \end{equation*} In the reduced representation, let $z$ be a leaf corner, with present ring bond $\{z, y\}$. Applying \eqref{e:local-leafb} with $n' = n_y$, we have \begin{equation*} \int (1 + n_y \cdot n)(1 + m \cdot n)(1 + m' \cdot n) \, \lambda_z(\mathrm d n) = A_z (1 + r \, b_z \cdot n_y) \end{equation*} for some $A_z > 0$ and $b_z \in \Ball$, where $r$ is the radius of the vector that replaces $z$ in the reduced representation. Now let $\mathbf b \in \Ball^k$ be the family with the entry $u_z$ for the vector of item (1) of \zcref{d:local-rep} at an axial site $z$, the entry $b_z$ for the vector that replaces a leaf corner $z$ in the reduced representation, and the corresponding slot vectors for the vectors of a pair. Comparing with \eqref{e:local-hatZ}, we find \begin{equation*} \mathcal Z_\pi^{\mathbf a}[h] = A \, \hat Z[h](\mathbf b), \qquad h \in \{I_{\mathrm{cur}}, I_{\mathrm{next}}\}, \qquad \text{where } A = \prod_z A_z > 0, \end{equation*} and the product runs over the corners that are not retained ($A = 1$ if there are none). The factor $A$ is the same for $h = I_{\mathrm{cur}}$ and $h = I_{\mathrm{next}}$, so that \begin{equation*} \Phi_c(\mathbf a) = A \, \hat\Phi_c(\mathbf b). \end{equation*} By assumption, $\hat\Phi_c \ge 0$ on $(\Sph)^k$, and \zcref{l:affine}, applied with $f = \hat\Phi_c$, extends this to $\Ball^k$. (For $k = 0$, the polynomial $\hat\Phi_c$ is a constant.) The claim follows. \end{proof} \section{The inversion certificate}\label{s:local-inversion} Recall from \zcref{s:quantum} the Pauli matrices $\sigma = (\sigma_1, \sigma_2, \sigma_3)$ and the spin-$1/2$ coherent states $\zeta(a) \in \C^2$, $a \in \Sph$, which satisfy \eqref{e:zeta}. We let $\sigma_0 = \Id$ be the $2 \times 2$ identity matrix, and for $a \in \R^3$ we set $a \cdot \sigma = \sum_{j = 1}^3 a^j \sigma_j$. In the standard basis $e_\uparrow = (1, 0)$, $e_\downarrow = (0, 1)$ of $\C^2$ (the eigenvectors of $\sigma_3$ for the eigenvalues $1$ and $-1$), we have \begin{equation*} \sigma_1 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}, \qquad \sigma_2 = \begin{pmatrix} 0 & -\iu \\ \iu & 0 \end{pmatrix}, \qquad \sigma_3 = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}. \end{equation*} The inner product $\langle \varphi, \psi \rangle = \varphi^* \psi$ is linear in the second argument. On $(\C^2)^{\otimes k}$, we use the \emph{standard basis}, which consists of the tensor products of $e_\uparrow$ and $e_\downarrow$. For an operator $A$ on $\C^2$ and $1 \le i \le k$, we let $A^{(i)}$ denote the operator on $(\C^2)^{\otimes k}$ that acts as $A$ on the $i$-th factor and as the identity on the others. For $\mathbf b \in (\Sph)^k$, we set $\zeta(\mathbf b) = \zeta(b_1) \otimes \dots \otimes \zeta(b_k)$. Let $k \ge 0$, and let $f$ be a real polynomial on $(\R^3)^k$ that is affine in each argument. Expanding $f$ in one argument at a time, we see that there exist real numbers $f_{\mathbf j}$, indexed by $\mathbf j = (j_1, \dots, j_k) \in \{0, 1, 2, 3\}^k$, such that \begin{equation}\label{e:local-expand} f(\mathbf b) = \sum_{\mathbf j} f_{\mathbf j} \prod_{i = 1}^k b_i^{j_i}, \qquad \mathbf b \in (\R^3)^k. \end{equation} The numbers $f_{\mathbf j}$ are unique, since the products on the right side are distinct monomials in the coordinates of $\mathbf b$. We call them the \emph{coefficients} of $f$, and we define the Hermitian operator \begin{equation}\label{e:local-O} \mathcal O(f) = \sum_{\mathbf j} f_{\mathbf j} \, \sigma_{\mathbf j} \quad \text{on } (\C^2)^{\otimes k}, \qquad \text{where } \sigma_{\mathbf j} = \sigma_{j_1} \otimes \dots \otimes \sigma_{j_k}. \end{equation} For $k = 0$, the polynomial $f$ is a constant, $(\C^2)^{\otimes 0} = \C$, and $\mathcal O(f) = f$. For a set $\mathcal I \subset \{1, \dots, k\}$, which we call an \emph{inversion set}, let $\mathbf b^{\mathcal I}$ be obtained from $\mathbf b$ by replacing $b_i$ with $-b_i$ for all $i \in \mathcal I$, and set $f^{\mathcal I}(\mathbf b) = f(\mathbf b^{\mathcal I})$. Observe that, by \eqref{e:local-expand}, the coefficients of $f^{\mathcal I}$ are \begin{equation}\label{e:local-invcoef} f^{\mathcal I}_{\mathbf j} = (-1)^{|\{i \in \mathcal I : j_i \ne 0\}|} f_{\mathbf j}. \end{equation} \begin{lemma}\label{l:inversion} Let $k \ge 0$, let $f$ be a real polynomial on $(\R^3)^k$ that is affine in each argument, and let $\mathcal I \subset \{1, \dots, k\}$. Then \begin{equation}\label{e:local-inversion} f(\mathbf b) = \big\langle \zeta(\mathbf b^{\mathcal I}), \mathcal O(f^{\mathcal I}) \, \zeta(\mathbf b^{\mathcal I}) \big\rangle \qquad \text{for all } \mathbf b \in (\Sph)^k. \end{equation} In particular, if $\mathcal O(f^{\mathcal I})$ is positive semidefinite, then $f \ge 0$ on $\Ball^k$. \end{lemma} \begin{proof} Let $a \in \Sph$ and $j \in \{0, 1, 2, 3\}$. Using \eqref{e:zeta} and the relations $\Tr(\sigma_j \sigma_l) = 2 \delta_{jl}$ for $j, l \in \{0, 1, 2, 3\}$, we have \begin{equation}\label{e:local-expect1} \langle \zeta(a), \sigma_j \zeta(a) \rangle = \Tr \big( \sigma_j \zeta(a) \zeta(a)^* \big) = \frac12 \Tr \Big( \sigma_j \sigma_0 + \sum_{l = 1}^3 a^l \sigma_j \sigma_l \Big) = a^j. \end{equation} Now let $\tilde f$ be a real polynomial on $(\R^3)^k$ that is affine in each argument, and let $\mathbf b' \in (\Sph)^k$. Applying \eqref{e:local-expect1} in each factor of the tensor product, we find \begin{equation*} \langle \zeta(\mathbf b'), \mathcal O(\tilde f) \, \zeta(\mathbf b') \rangle = \sum_{\mathbf j} \tilde f_{\mathbf j} \prod_{i = 1}^k \langle \zeta(b'_i), \sigma_{j_i} \zeta(b'_i) \rangle = \sum_{\mathbf j} \tilde f_{\mathbf j} \prod_{i = 1}^k (b'_i)^{j_i} = \tilde f(\mathbf b'). \end{equation*} We apply this with $\tilde f = f^{\mathcal I}$ and $\mathbf b' = \mathbf b^{\mathcal I}$, which lies in $(\Sph)^k$. Then \eqref{e:local-inversion} follows from \begin{equation*} f^{\mathcal I}(\mathbf b^{\mathcal I}) = f\big( (\mathbf b^{\mathcal I})^{\mathcal I} \big) = f(\mathbf b). \end{equation*} For the last assertion, note that \eqref{e:local-inversion} writes $f(\mathbf b)$ as the expectation of $\mathcal O(f^{\mathcal I})$ in a unit vector. When $\mathcal O(f^{\mathcal I})$ is positive semidefinite, we therefore have $f \ge 0$ on $(\Sph)^k$, and for $k \ge 1$, \zcref{l:affine} extends this to $\Ball^k$. \end{proof} The inversion is what makes \zcref{l:inversion} useful, already for two vectors. To see this, let $\mathsf X$ be the operator on $\C^2 \otimes \C^2$ that exchanges the two factors, $\mathsf X(\varphi \otimes \psi) = \psi \otimes \varphi$. It is straightforward to check on the standard basis that \begin{equation}\label{e:local-swap} \sum_{j = 0}^3 \sigma_j \otimes \sigma_j = 2 \mathsf X. \end{equation} Recall that $\mathsf X$ has the eigenvalue $1$ on the symmetric subspace $\mathrm{Sym}^2 \C^2$, which is spanned by $e_\uparrow \otimes e_\uparrow$, $e_\uparrow \otimes e_\downarrow + e_\downarrow \otimes e_\uparrow$, and $e_\downarrow \otimes e_\downarrow$, and the eigenvalue $-1$ on the antisymmetric subspace $\bigwedge^2 \C^2$, which is spanned by $e_\uparrow \otimes e_\downarrow - e_\downarrow \otimes e_\uparrow$. \begin{example} Let $k = 2$ and $f(b_1, b_2) = 1 + b_1 \cdot b_2$, which is nonnegative on $(\Sph)^2$ and vanishes when $b_2 = -b_1$. By \eqref{e:local-invcoef} and \eqref{e:local-swap}, we have \begin{equation*} \mathcal O(f) = \sigma_0 \otimes \sigma_0 + \sum_{j = 1}^3 \sigma_j \otimes \sigma_j = 2 \mathsf X, \qquad \mathcal O(f^{\{1\}}) = \sigma_0 \otimes \sigma_0 - \sum_{j = 1}^3 \sigma_j \otimes \sigma_j = 2 \sigma_0 \otimes \sigma_0 - 2 \mathsf X. \end{equation*} The operator $\mathcal O(f)$ has the eigenvalue $-2$ on $\bigwedge^2 \C^2$, so that \zcref{l:inversion} with $\mathcal I = \emptyset$ does not apply, although $f \ge 0$. However, $\mathcal O(f^{\{1\}})$ has the eigenvalues $0$ on $\mathrm{Sym}^2 \C^2$ and $4$ on $\bigwedge^2 \C^2$, and \zcref{l:inversion} with $\mathcal I = \{1\}$ gives $f \ge 0$ on $\Ball^2$. (The operator $\sum_{j = 1}^3 \sigma_j \otimes \sigma_j = 2 \mathsf X - \sigma_0 \otimes \sigma_0$ is the Heisenberg coupling of two spins $1/2$, with the eigenvalue $1$ on the triplet $\mathrm{Sym}^2 \C^2$ and $-3$ on the singlet $\bigwedge^2 \C^2$.) \end{example} We now fix one of the two representations of $\pi$, and we apply \zcref{l:inversion} to $\hat\Phi_c$. We will always take the inversion set to be a union of groups, which means that the two vectors of a pair are either both in $\mathcal I$ or both outside $\mathcal I$. For a corner $z$ that carries a pair, with vectors $i$ and $i'$, let $\mathsf X_z$ be the operator on $(\C^2)^{\otimes k}$ that exchanges the factors $i$ and $i'$, and let \begin{equation*} \Pi^{\mathrm{sym}}_z = \frac12 (\Id + \mathsf X_z) \end{equation*} be the orthogonal projection onto the vectors that are invariant under $\mathsf X_z$. On the factors $i$ and $i'$, it acts as the orthogonal projection of $\C^2 \otimes \C^2$ onto $\mathrm{Sym}^2 \C^2$. We will need the following lemma, which collects the properties of the operator that the test uses. Parts (3) and (4) distinguish the two statuses of the corners, since the measure $\lambda_1$ has an atom at $0$. \begin{lemma}\label{l:Ostruct} Fix one of the two representations of $\pi$, let $f$ be a real linear combination of $\hat Z_{\mathrm{cur}}$ and $\hat Z_{\mathrm{next}}$, and let $\mathcal I$ be an inversion set that is a union of groups. Then the following hold. \begin{enumerate} \item The matrix of $\mathcal O(f^{\mathcal I})$ in the standard basis is real and symmetric. \item For all $g \in \mathrm{SU}(2)$, we have $g^{\otimes k} \mathcal O(f^{\mathcal I}) = \mathcal O(f^{\mathcal I}) g^{\otimes k}$. \item If $z$ is a corner of status $0$ that carries a pair, then $\Pi^{\mathrm{sym}}_z \mathcal O(f^{\mathcal I}) = \mathcal O(f^{\mathcal I}) = \mathcal O(f^{\mathcal I}) \Pi^{\mathrm{sym}}_z$. \item If $z$ is a corner of status $1$ that carries a pair, then $\mathsf X_z \mathcal O(f^{\mathcal I}) = \mathcal O(f^{\mathcal I}) \mathsf X_z$. \end{enumerate} \end{lemma} In the proof of \zcref{l:Ostruct}, we will use four properties of the map $f \mapsto \mathcal O(f)$, each of which follows from \eqref{e:local-O} and the uniqueness of the coefficients. First, the map is linear. Second, suppose that $\{1, \dots, k\}$ is the disjoint union of sets $J$ and $J'$, and that $f(\mathbf b) = f_1(\mathbf b_J) f_2(\mathbf b_{J'})$, where $\mathbf b_J = (b_i)_{i \in J}$ and $\mathbf b_{J'} = (b_i)_{i \in J'}$. Then the coefficients of $f$ are the products of those of $f_1$ and $f_2$, and $\mathcal O(f)$ is the tensor product of $\mathcal O(f_1)$, acting on the factors in $J$, and $\mathcal O(f_2)$, acting on the factors in $J'$. In particular, for one vector, $a \in \R^3$, and $r \in \R$, we have \begin{equation}\label{e:local-Oone} \mathcal O(b \mapsto 1 + r \, a \cdot b) = \sigma_0 + r \, a \cdot \sigma. \end{equation} Third, if $f = \int f_n \, \mu(\mathrm d n)$ for a finite measure $\mu$ and polynomials $f_n$ whose coefficients are bounded measurable functions of $n$, then $\mathcal O(f) = \int \mathcal O(f_n) \, \mu(\mathrm d n)$. Fourth, let $f$ and $\mathcal I$ be as in \zcref{l:Ostruct}. As noted after \eqref{e:local-hatPhi}, we have $f(Q \mathbf b) = f(\mathbf b)$ for all $Q \in \mathrm O(3)$. Since $(Q \mathbf b)^{\mathcal I} = Q(\mathbf b^{\mathcal I})$, this implies that \begin{equation}\label{e:local-invariance} f^{\mathcal I}(Q \mathbf b) = f^{\mathcal I}(\mathbf b) \qquad \text{for all } Q \in \mathrm O(3) \text{ and } \mathbf b \in (\R^3)^k. \end{equation} \begin{proof}[Proof of \zcref{l:Ostruct}(2)] Let $g \in \mathrm{SU}(2)$, and recall from \zcref{s:quantum} that its rotation $\bar g \in \mathrm{SO}(3)$ is defined by \begin{equation*} g (a \cdot \sigma) g^* = (\bar g a) \cdot \sigma, \qquad a \in \R^3. \end{equation*} Let $R$ be the $4 \times 4$ matrix, indexed by $\{0, 1, 2, 3\}$, such that $R_{00} = 1$, $R_{0j} = R_{j0} = 0$, and $R_{lj} = \bar g_{lj}$ for $l, j \in \{1, 2, 3\}$. Taking for $a$ the unit vectors of $\R^3$, we obtain \begin{equation}\label{e:local-gsigma} g \sigma_j g^* = \sum_{l = 0}^3 R_{lj} \sigma_l, \qquad j \in \{0, 1, 2, 3\}. \end{equation} Inserting \eqref{e:local-gsigma} into \eqref{e:local-O}, we find \begin{equation}\label{e:local-conj} g^{\otimes k} \mathcal O(f^{\mathcal I}) (g^*)^{\otimes k} = \sum_{\mathbf l} \Big( \sum_{\mathbf j} f^{\mathcal I}_{\mathbf j} \prod_{i = 1}^k R_{l_i j_i} \Big) \sigma_{\mathbf l}, \end{equation} where $\mathbf l$ and $\mathbf j$ run over $\{0, 1, 2, 3\}^k$. On the other hand, for $b \in \R^3$ and $j \in \{0, 1, 2, 3\}$, we have \begin{equation*} (\bar g^{\mathsf T} b)^j = \sum_{l = 0}^3 R_{lj} b^l, \end{equation*} so that \begin{equation*} f^{\mathcal I}(\bar g^{\mathsf T} \mathbf b) = \sum_{\mathbf j} f^{\mathcal I}_{\mathbf j} \prod_{i = 1}^k \sum_{l_i = 0}^3 R_{l_i j_i} b_i^{l_i} = \sum_{\mathbf l} \Big( \sum_{\mathbf j} f^{\mathcal I}_{\mathbf j} \prod_{i = 1}^k R_{l_i j_i} \Big) \prod_{i = 1}^k b_i^{l_i}. \end{equation*} By \eqref{e:local-invariance}, the left side coincides with $f^{\mathcal I}(\mathbf b)$, and the uniqueness of the coefficients implies that \begin{equation*} \sum_{\mathbf j} f^{\mathcal I}_{\mathbf j} \prod_{i = 1}^k R_{l_i j_i} = f^{\mathcal I}_{\mathbf l} \qquad \text{for all } \mathbf l \in \{0, 1, 2, 3\}^k. \end{equation*} Inserting this into \eqref{e:local-conj}, we have $g^{\otimes k} \mathcal O(f^{\mathcal I}) (g^*)^{\otimes k} = \mathcal O(f^{\mathcal I})$, which is (2), since $g^* = g^{-1}$. \end{proof} \begin{proof}[Proof of \zcref{l:Ostruct}(1)] The operator $\mathcal O(f^{\mathcal I})$ is Hermitian, since its coefficients are real and the matrices $\sigma_j$ are Hermitian. To show that its matrix is real, observe that \begin{equation*} \bar\sigma_j = \sigma_j \quad \text{for } j \ne 2, \qquad \bar\sigma_2 = -\sigma_2, \end{equation*} where the bar denotes the entrywise complex conjugate. Let $D = \diag(1, -1, 1)$, the element of $\mathrm O(3)$ that reverses the second coordinate. Then the entrywise complex conjugate of the matrix of $\mathcal O(f^{\mathcal I})$ is \begin{equation*} \sum_{\mathbf j} (-1)^{|\{i : j_i = 2\}|} f^{\mathcal I}_{\mathbf j} \, \sigma_{\mathbf j} = \mathcal O\big( \mathbf b \mapsto f^{\mathcal I}(D \mathbf b) \big) = \mathcal O(f^{\mathcal I}), \end{equation*} where we used \eqref{e:local-invariance} in the last step. We conclude that the matrix of $\mathcal O(f^{\mathcal I})$ is real, and since it is Hermitian, it is symmetric. \end{proof} \begin{proof}[Proof of \zcref{l:Ostruct}(3)] Let $z$ be a corner of status $0$ that carries a pair, with vectors $i$ and $i'$. Since $\mathcal I$ is a union of groups, the vectors $i$ and $i'$ are either both in $\mathcal I$ or both outside it, and we let the sign $\pm$ be $-$ in the first case and $+$ in the second. The spin $n_z$ occurs in \eqref{e:local-hatZ} only in the factors of the vectors $i$ and $i'$ and in the factors of the ring bonds at $z$, and the vectors $b_i$ and $b_{i'}$ occur only in the factors $1 + b_i \cdot n_z$ and $1 + b_{i'} \cdot n_z$. Integrating first over all spins other than $n_z$, and recalling that $\lambda_z = \dd n$, we can write \begin{equation}\label{e:local-split} f^{\mathcal I}(\mathbf b) = \int (1 \pm b_i \cdot n)(1 \pm b_{i'} \cdot n) \, H_n(\mathbf b') \dd n, \end{equation} where $\mathbf b' = (b_l)_{l \ne i, i'}$, and $H_n$ is a real polynomial in $\mathbf b'$ that is affine in each argument and whose coefficients are polynomials in $n$. Applying the second and third properties above and \eqref{e:local-Oone} to \eqref{e:local-split}, we find \begin{equation}\label{e:local-Osplit} \mathcal O(f^{\mathcal I}) = \int (\sigma_0 \pm n \cdot \sigma)^{(i)} (\sigma_0 \pm n \cdot \sigma)^{(i')} \, \widetilde{\mathcal O}(H_n) \dd n, \end{equation} where $\widetilde{\mathcal O}(H_n)$ acts as $\mathcal O(H_n)$ on the factors other than $i$ and $i'$ and as the identity on these two. We now show that $\Pi^{\mathrm{sym}}_z$ leaves every integrand in \eqref{e:local-Osplit} unchanged, from the left and from the right. Let $n \in \Sph$. By \eqref{e:zeta} with $a = \pm n$, we have $\sigma_0 \pm n \cdot \sigma = 2 \zeta(\pm n) \zeta(\pm n)^*$, so that \begin{equation*} (\sigma_0 \pm n \cdot \sigma)^{(i)} (\sigma_0 \pm n \cdot \sigma)^{(i')} = 4 P_n, \end{equation*} where $P_n$ is the orthogonal projection onto $\zeta(\pm n) \otimes \zeta(\pm n)$ on the factors $i$ and $i'$, tensored with the identity on the others. The vector $\zeta(\pm n) \otimes \zeta(\pm n)$ is invariant under the exchange of the two factors, which gives \begin{equation*} \Pi^{\mathrm{sym}}_z P_n = P_n = P_n \Pi^{\mathrm{sym}}_z. \end{equation*} Further, $\Pi^{\mathrm{sym}}_z$ commutes with $\widetilde{\mathcal O}(H_n)$, since they act on different factors. The measure $\dd n$ is carried by $\Sph$, and (3) follows from \eqref{e:local-Osplit}. (For a corner of status $1$, the atom of $\lambda_1$ at $0$ contributes a term to the analogue of \eqref{e:local-Osplit} in which the factors $i$ and $i'$ carry $\sigma_0 \otimes \sigma_0$, and (3) fails.) \end{proof} \begin{proof}[Proof of \zcref{l:Ostruct}(4)] Let $z$ be a corner of status $1$ that carries a pair, with vectors $i$ and $i'$, and let the sign $\pm$ be as in the proof of (3). The vectors $b_i$ and $b_{i'}$ occur in $f^{\mathcal I}$ only through the product $(1 \pm b_i \cdot n_z)(1 \pm b_{i'} \cdot n_z)$, which is symmetric in $b_i$ and $b_{i'}$. In particular, exchanging $b_i$ and $b_{i'}$ leaves $f^{\mathcal I}$ unchanged, and the uniqueness of the coefficients implies that $f^{\mathcal I}_{\mathbf j} = f^{\mathcal I}_{\mathbf j'}$, where $\mathbf j'$ is obtained from $\mathbf j$ by exchanging $j_i$ and $j_{i'}$. Since $\mathsf X_z \sigma_{\mathbf j} \mathsf X_z = \sigma_{\mathbf j'}$, we obtain \begin{equation*} \mathsf X_z \mathcal O(f^{\mathcal I}) \mathsf X_z = \sum_{\mathbf j} f^{\mathcal I}_{\mathbf j} \sigma_{\mathbf j'} = \sum_{\mathbf j} f^{\mathcal I}_{\mathbf j'} \sigma_{\mathbf j'} = \mathcal O(f^{\mathcal I}), \end{equation*} which is (4), since $\mathsf X_z^2 = \Id$. \end{proof} \begin{remark}\label{r:local-choice} The second part of \zcref{l:Ostruct} suggests how to choose $\mathcal I$. Let $f$ be as in \zcref{l:Ostruct}, let $e \in \Sph$, and let $\mathbf b \in (\Sph)^k$ be such that $b_i \in \{e, -e\}$ for all $i$ and the two vectors of every pair are equal. Let $\mathcal I = \{i : b_i = -e\}$, a union of groups, so that $\zeta(\mathbf b^{\mathcal I}) = \zeta(e)^{\otimes k}$. Up to a scalar factor, this is the only eigenvector of $(1/2) \sum_{i = 1}^k (e \cdot \sigma)^{(i)}$ for its largest eigenvalue $k/2$. Differentiating the identity of part (2) along $g_t = \exp(\iu t \, e \cdot \sigma/2)$ at $t = 0$, we see that this operator commutes with $\mathcal O(f^{\mathcal I})$. It follows that $\zeta(e)^{\otimes k}$ is an eigenvector of $\mathcal O(f^{\mathcal I})$, with the eigenvalue $f(\mathbf b)$ by \eqref{e:local-inversion}. Consequently, if a minimizer $\mathbf b$ of $\hat Z_{\mathrm{cur}}/\hat Z_{\mathrm{next}}$ on $(\Sph)^k$ has this form, then $\hat\Phi_c(\mathbf b)$ is an eigenvalue of $\mathcal O(\hat\Phi_c^{\mathcal I})$, and this eigenvalue vanishes when $c$ is the minimum of the ratio. The polynomial of the example above is also invariant under simultaneous rotations. Its minimizer $\mathbf b = (e, -e)$ leads to $\mathcal I = \{2\}$, for which $f^{\{2\}} = f^{\{1\}}$. In the computation, the inversion set is chosen in this way from a numerical minimizer, which is only approximately of this form; see \zcref{s:local-computation}. The choice of $\mathcal I$ affects only whether the test succeeds, not its validity. \end{remark} We now use the symmetries of \zcref{l:Ostruct} to reduce the test to small subspaces. Let \begin{equation*} S^3_{\mathrm{tot}} = \frac12 \sum_{i = 1}^k \sigma_3^{(i)} \end{equation*} be the total spin component. Let $\Pi_{\mathcal V}$ be the product of the projections $\Pi^{\mathrm{sym}}_z$ over the corners $z$ of status $0$ that carry a pair, and let $\mathcal V$ be its range. These projections act on different factors, so that $\Pi_{\mathcal V}$ is the orthogonal projection onto $\mathcal V$, and $\mathcal V$ is the tensor product, over the groups, of $\C^2$ for a single vector, $\mathrm{Sym}^2 \C^2$ for the pair of a corner of status $0$, and $\C^2 \otimes \C^2$ for the pair of a corner of status $1$. We call a family $\epsilon = (\epsilon_z)$ of signs $\epsilon_z \in \{1, -1\}$, indexed by the corners $z$ of status $1$ that carry a pair, a \emph{sign assignment}, and we set \begin{equation*} \mathcal V_\epsilon = \{ \psi \in \mathcal V : \mathsf X_z \psi = \epsilon_z \psi \text{ for every corner } z \text{ of status } 1 \text{ that carries a pair} \}. \end{equation*} Equivalently, $\mathcal V_\epsilon$ is the tensor product in which the factor $\C^2 \otimes \C^2$ of such a corner $z$ is replaced by $\mathrm{Sym}^2 \C^2$ if $\epsilon_z = 1$ and by $\bigwedge^2 \C^2$ if $\epsilon_z = -1$. The space $\mathcal V$ is the orthogonal direct sum of the spaces $\mathcal V_\epsilon$. Let $\kappa_\circ = 0$ if the number of single vectors is even and $\kappa_\circ = 1/2$ if it is odd, and let \begin{equation*} \mathcal V_\epsilon^\circ = \{ \psi \in \mathcal V_\epsilon : S^3_{\mathrm{tot}} \psi = \kappa_\circ \psi \}. \end{equation*} Note that for $g \in \mathrm{SU}(2)$, the operator $g \otimes g$ commutes with the exchange of the two factors of $\C^2 \otimes \C^2$, so that $g^{\otimes k}$ commutes with every $\mathsf X_z$ and every $\Pi^{\mathrm{sym}}_z$. Therefore $\mathcal V$ and every $\mathcal V_\epsilon$ are invariant under $g^{\otimes k}$ and, by the same argument, under $S^3_{\mathrm{tot}}$. \begin{lemma}\label{l:sector} Let $\epsilon$ be a sign assignment, and let $\mathcal O'$ be a Hermitian operator on $\mathcal V_\epsilon$ such that $\mathcal O' g^{\otimes k} \psi = g^{\otimes k} \mathcal O' \psi$ for all $g \in \mathrm{SU}(2)$ and $\psi \in \mathcal V_\epsilon$. Then $\mathcal O'$ maps $\mathcal V_\epsilon^\circ$ into itself, and the restriction of $\mathcal O'$ to $\mathcal V_\epsilon^\circ$ has the same set of eigenvalues as $\mathcal O'$. In particular, $\mathcal O'$ is positive definite if and only if its restriction to $\mathcal V_\epsilon^\circ$ is positive definite. \end{lemma} \begin{proof} We first determine the eigenvalues of $S^3_{\mathrm{tot}}$ on $\mathcal V_\epsilon$. Recall that $\sigma_3 e_\uparrow = e_\uparrow$ and $\sigma_3 e_\downarrow = -e_\downarrow$, so that the operator $(\sigma_3 \otimes \sigma_0 + \sigma_0 \otimes \sigma_3)/2$ on $\C^2 \otimes \C^2$ has the eigenvalues $1$, $0$, and $-1$ on the vectors $e_\uparrow \otimes e_\uparrow$, $e_\uparrow \otimes e_\downarrow + e_\downarrow \otimes e_\uparrow$, and $e_\downarrow \otimes e_\downarrow$ of $\mathrm{Sym}^2 \C^2$, and the eigenvalue $0$ on $\bigwedge^2 \C^2$. Taking tensor products of these eigenvectors and of $e_\uparrow$, $e_\downarrow$ for the single vectors, we obtain a basis of $\mathcal V_\epsilon$ consisting of eigenvectors of $S^3_{\mathrm{tot}}$. Each eigenvalue is a sum of one number $\pm 1/2$ for each single vector and one integer for each pair, so that every eigenvalue of $S^3_{\mathrm{tot}}$ on $\mathcal V_\epsilon$ lies in $\kappa_\circ + \Z$. Next, the operators $\sigma_3^{(i)}$ commute, so that \begin{equation*} g_t = \exp(\iu t \sigma_3/2) \in \mathrm{SU}(2) \qquad \text{satisfies} \qquad g_t^{\otimes k} = \exp(\iu t S^3_{\mathrm{tot}}), \qquad t \in \R. \end{equation*} By assumption, $\mathcal O'$ commutes with the restriction of $g_t^{\otimes k}$ to $\mathcal V_\epsilon$. Differentiating at $t = 0$, we find that $\mathcal O'$ commutes with the restriction of $S^3_{\mathrm{tot}}$ to $\mathcal V_\epsilon$. In particular, $\mathcal O'$ maps the eigenspace $\mathcal V_\epsilon^\circ$ into itself, and every eigenvalue of the restriction of $\mathcal O'$ to $\mathcal V_\epsilon^\circ$ is an eigenvalue of $\mathcal O'$. Conversely, let $\theta$ be an eigenvalue of $\mathcal O'$, and let $E \subset \mathcal V_\epsilon$ be its eigenspace. The space $E$ is invariant under $g^{\otimes k}$ for all $g \in \mathrm{SU}(2)$, since $\mathcal O'$ commutes with $g^{\otimes k}$. Let $W \subset E$ be an irreducible invariant subspace with $W \ne \{0\}$. By the standard classification of the irreducible representations of $\mathrm{SU}(2)$, there exists $j \in \{0, 1/2, 1, 3/2, \dots\}$ such that the eigenvalues of $S^3_{\mathrm{tot}}$ on $W$ are $j, j - 1, \dots, -j$; here we used that $S^3_{\mathrm{tot}}$ is $-\iu$ times the derivative of $g_t^{\otimes k}$ at $t = 0$. These eigenvalues lie in $\kappa_\circ + \Z$ by the first step. Therefore $j - \kappa_\circ$ is a nonnegative integer, and $\kappa_\circ$ is one of the eigenvalues $j, j - 1, \dots, -j$. It follows that $W \cap \mathcal V_\epsilon^\circ \ne \{0\}$, and every nonzero vector of $W \cap \mathcal V_\epsilon^\circ$ is an eigenvector of the restriction of $\mathcal O'$ to $\mathcal V_\epsilon^\circ$ with eigenvalue $\theta$. This completes the proof. \end{proof} We can now state the certificate. \begin{proposition}\label{p:certificate} Fix one of the two representations of $\pi$, and let $\mathcal I$ be an inversion set that is a union of groups. Suppose that for every sign assignment $\epsilon$, \begin{equation}\label{e:local-certhyp} \big\langle \psi, \mathcal O(\hat\Phi_c^{\mathcal I}) \, \psi \big\rangle > 0 \qquad \text{for all } \psi \in \mathcal V_\epsilon^\circ \setminus \{0\}. \end{equation} Then $\Phi_c(\mathbf a) \ge 0$ for all $\mathbf a \in \Ball^{12}$. \end{proposition} \begin{proof} Let $\mathcal O = \mathcal O(\hat\Phi_c^{\mathcal I})$. We apply \zcref{l:Ostruct} with $f = \hat\Phi_c$, which is a real linear combination of $\hat Z_{\mathrm{cur}}$ and $\hat Z_{\mathrm{next}}$. Applying part (3) to every corner of status $0$ that carries a pair, we obtain \begin{equation}\label{e:local-OPi} \mathcal O = \Pi_{\mathcal V} \mathcal O \Pi_{\mathcal V}, \end{equation} so that $\mathcal O$ maps $\mathcal V$ into itself. By part (4), we have $\mathcal O \mathsf X_z = \mathsf X_z \mathcal O$ for every corner $z$ of status $1$ that carries a pair, so that $\mathcal O$ maps every $\mathcal V_\epsilon$ into itself. Let $\epsilon$ be a sign assignment, and let $\mathcal O_\epsilon$ be the restriction of $\mathcal O$ to $\mathcal V_\epsilon$. The operator $\mathcal O_\epsilon$ is Hermitian. Using part (2) and the invariance of $\mathcal V_\epsilon$ under $g^{\otimes k}$, we see that $\mathcal O_\epsilon$ satisfies the assumption of \zcref{l:sector}. By that lemma, $\mathcal O_\epsilon$ maps $\mathcal V_\epsilon^\circ$ into itself. Its restriction to $\mathcal V_\epsilon^\circ$ is positive definite by \eqref{e:local-certhyp}, and \zcref{l:sector} shows that $\mathcal O_\epsilon$ is positive definite. The space $\mathcal V$ is the orthogonal direct sum of the spaces $\mathcal V_\epsilon$, and $\mathcal O$ maps each of them into itself. We conclude that the restriction of $\mathcal O$ to $\mathcal V$ is positive definite. Now let $\mathbf b \in (\Sph)^k$. We apply \zcref{l:inversion} with $f = \hat\Phi_c$ and the inversion set $\mathcal I$, and we obtain \begin{equation*} \hat\Phi_c(\mathbf b) = \big\langle \zeta(\mathbf b^{\mathcal I}), \mathcal O \, \zeta(\mathbf b^{\mathcal I}) \big\rangle = \big\langle \Pi_{\mathcal V} \zeta(\mathbf b^{\mathcal I}), \mathcal O \, \Pi_{\mathcal V} \zeta(\mathbf b^{\mathcal I}) \big\rangle \ge 0, \end{equation*} where we used \eqref{e:local-OPi} in the second step and the positive definiteness of $\mathcal O$ on $\mathcal V$ in the last. By \zcref{c:reduce}, this implies that $\Phi_c \ge 0$ on $\Ball^{12}$. \end{proof} (The proof works equally well if the strict inequality in \eqref{e:local-certhyp} is replaced by $\ge$; the program checks the strict inequality.) The reduction to the spaces $\mathcal V_\epsilon^\circ$ makes the test much cheaper. For the patterns of \zcref{s:local-computation}, the space $\mathcal V$ has dimension up to $2^{11} = 2048$, while the spaces $\mathcal V_\epsilon^\circ$ have dimension at most $245$. \section{The test in integer arithmetic}\label{s:local-integer} We now turn \eqref{e:local-certhyp} into a statement about integer matrices. Let $\mathfrak d$ be a positive integer such that the polynomials $3 \mathfrak d \hat Z_{\mathrm{cur}}$ and $5 \mathfrak d \hat Z_{\mathrm{next}}$ have integer coefficients. Such an integer exists, since the coefficients of $\hat Z_{\mathrm{cur}}$ and $\hat Z_{\mathrm{next}}$ are rational. (By \eqref{e:local-hatZ}, they are finite sums of products of the radii $r_i$, of the coefficients $1/3$ and $1/5$ of $I_{\mathrm{cur}}$ and $I_{\mathrm{next}}$, and of moments of $\lambda_0$ and $\lambda_1$, which are rational by \eqref{e:local-moments} below.) The value of $\mathfrak d$ that the programs use is given in \zcref{s:local-computation}. By \eqref{e:local-hatPhi}, we have \begin{equation}\label{e:local-intscale} 15 \mathfrak d c_2 \, \hat\Phi_c = 5 c_2 \, (3 \mathfrak d \hat Z_{\mathrm{cur}}) - 3 c_1 \, (5 \mathfrak d \hat Z_{\mathrm{next}}). \end{equation} Let $\mathcal B$ be the set of the tensor products, over the groups, of one vector for each group, chosen among $e_\uparrow$ and $e_\downarrow$ for a single vector, among \begin{equation*} e_\uparrow \otimes e_\uparrow, \qquad e_\uparrow \otimes e_\downarrow + e_\downarrow \otimes e_\uparrow, \qquad e_\downarrow \otimes e_\downarrow \end{equation*} for the pair of a corner of status $0$, and among these three vectors and $e_\uparrow \otimes e_\downarrow - e_\downarrow \otimes e_\uparrow$ for the pair of a corner of status $1$. Then $\mathcal B$ is a basis of $\mathcal V$, its elements have integer coordinates in the standard basis, and each of them is an eigenvector of $S^3_{\mathrm{tot}}$ and of every $\mathsf X_z$. For a sign assignment $\epsilon$, the space $\mathcal V_\epsilon^\circ$ is, by its definition, a joint eigenspace of these commuting operators. It follows that the set $\mathcal B_\epsilon$ of the elements of $\mathcal B$ that lie in $\mathcal V_\epsilon^\circ$ is a basis of $\mathcal V_\epsilon^\circ$. We define the matrices \begin{equation*} \mathsf A^{\mathrm{cur}}_\epsilon = \Big( \big\langle \beta, \mathcal O\big( (3 \mathfrak d \hat Z_{\mathrm{cur}})^{\mathcal I} \big) \beta' \big\rangle \Big)_{\beta, \beta' \in \mathcal B_\epsilon}, \qquad \mathsf A^{\mathrm{next}}_\epsilon = \Big( \big\langle \beta, \mathcal O\big( (5 \mathfrak d \hat Z_{\mathrm{next}})^{\mathcal I} \big) \beta' \big\rangle \Big)_{\beta, \beta' \in \mathcal B_\epsilon}. \end{equation*} Their entries lie in $\Z[\iu]$, since the coefficients of the polynomials are integers and the entries of $\sigma_0, \dots, \sigma_3$, $\beta$, and $\beta'$ lie in $\Z[\iu]$. By \zcref{l:Ostruct}(1), applied with $f = 3 \mathfrak d \hat Z_{\mathrm{cur}}$ and with $f = 5 \mathfrak d \hat Z_{\mathrm{next}}$, the entries are real and the matrices are symmetric, so that $\mathsf A^{\mathrm{cur}}_\epsilon$ and $\mathsf A^{\mathrm{next}}_\epsilon$ are symmetric integer matrices. We set \begin{equation}\label{e:local-Meps} \mathsf M_\epsilon = 5 c_2 \mathsf A^{\mathrm{cur}}_\epsilon - 3 c_1 \mathsf A^{\mathrm{next}}_\epsilon. \end{equation} By \eqref{e:local-intscale} and the linearity of $f \mapsto \mathcal O(f^{\mathcal I})$, we have \begin{equation}\label{e:local-Mform} \mathsf M_\epsilon = 15 \mathfrak d c_2 \Big( \big\langle \beta, \mathcal O(\hat\Phi_c^{\mathcal I}) \beta' \big\rangle \Big)_{\beta, \beta' \in \mathcal B_\epsilon}. \end{equation} \begin{lemma}\label{l:integertest} Fix one of the two representations of $\pi$, let $\mathcal I$ be an inversion set that is a union of groups, and let $\epsilon$ be a sign assignment. Then \eqref{e:local-certhyp} holds for this $\epsilon$ if and only if the symmetric integer matrix $\mathsf M_\epsilon$ is positive definite. \end{lemma} \begin{proof} Every $\psi \in \mathcal V_\epsilon^\circ$ can be written uniquely as \begin{equation*} \psi = \sum_{\beta \in \mathcal B_\epsilon} x_\beta \beta, \qquad x = (x_\beta)_{\beta \in \mathcal B_\epsilon} \in \C^{\mathcal B_\epsilon}, \end{equation*} and $\psi \ne 0$ if and only if $x \ne 0$. Write $x = x' + \iu x''$ with $x'$ and $x''$ real. Using \eqref{e:local-Mform} and the fact that $\mathsf M_\epsilon$ is real and symmetric, we have \begin{equation*} 15 \mathfrak d c_2 \big\langle \psi, \mathcal O(\hat\Phi_c^{\mathcal I}) \, \psi \big\rangle = x^* \mathsf M_\epsilon x = (x')^{\mathsf T} \mathsf M_\epsilon x' + (x'')^{\mathsf T} \mathsf M_\epsilon x''. \end{equation*} The right side is positive for all nonzero $x$ if and only if $\mathsf M_\epsilon$ is positive definite. \end{proof} Combining \zcref{p:certificate} and \zcref{l:integertest}, we see that once a representation and an inversion set have been chosen, the inequality \eqref{e:local-goal} for $\pi$ follows from the positive definiteness of the matrices $\mathsf M_\epsilon$ for all sign assignments $\epsilon$. We test positive definiteness with a standard criterion. For a symmetric $n \times n$ matrix $\mathsf M$ and $1 \le j \le n$, let $\mathsf M_{[j]}$ denote its upper left $j \times j$ submatrix, and set $\det \mathsf M_{[0]} = 1$. Recall that, by Sylvester's criterion \cite{HornJohnson13}, a real symmetric matrix $\mathsf M$ is positive definite if and only if $\det \mathsf M_{[j]} > 0$ for $j = 1, \dots, n$. To compute these minors, we use fraction-free Gaussian elimination \cite{Bareiss68}. Let $\mathsf M$ be a symmetric $n \times n$ integer matrix, and set $\mathsf M^{(0)} = \mathsf M$. For $l = 1, \dots, n - 1$ and $l < a, b \le n$, define \begin{equation}\label{e:local-bareiss} \mathsf M^{(l)}_{ab} = \frac{\mathsf M^{(l - 1)}_{ll} \mathsf M^{(l - 1)}_{ab} - \mathsf M^{(l - 1)}_{al} \mathsf M^{(l - 1)}_{lb}}{\det \mathsf M_{[l - 1]}}. \end{equation} By Sylvester's determinant identity \cite{Bareiss68}, as long as the minors $\det \mathsf M_{[1]}, \dots, \det \mathsf M_{[l - 1]}$ are nonzero, $\mathsf M^{(l)}_{ab}$ is the determinant of the submatrix of $\mathsf M$ with rows $1, \dots, l, a$ and columns $1, \dots, l, b$. In particular, the pivots are the leading principal minors, \begin{equation}\label{e:local-pivots} \mathsf M^{(j - 1)}_{jj} = \det \mathsf M_{[j]}, \qquad 1 \le j \le n, \end{equation} every division in \eqref{e:local-bareiss} is exact, and every $\mathsf M^{(l)}_{ab}$ is an integer. Further, $\mathsf M^{(l)}_{ab} = \mathsf M^{(l)}_{ba}$, since the two submatrices are transposes of each other. The program computes the pivots \eqref{e:local-pivots} in the order $j = 1, \dots, n$ with unbounded integers, updates only the entries $\mathsf M^{(l)}_{ab}$ with $a \le b$, and rejects $\mathsf M$ at the first pivot that is not positive. When it accepts $\mathsf M$, all leading principal minors are positive, and $\mathsf M$ is positive definite. \section{The computation}\label{s:local-computation} In this section, we describe the patterns, the programs, and the certificate file, and we complete the proof of \zcref{t:H1}. All files named below are ancillary files of this paper in the directory \texttt{anc/h1}. \paragraph{Patterns.} The reflection $(z_1, z_2) \mapsto (-z_1, z_2)$ of the template maps $\Pext$ onto itself. It fixes $\sS$ and $\sN$, exchanges $\sE$ and $\sW$, and preserves the type of a pattern. It maps the ring functional of a pattern to that of its image, after the corresponding relabeling of the twelve vectors of $\mathbf a$, so that it suffices to treat the straight patterns and the turns with $o = \sE$ (the turns with $o = \sW$ are their reflections). Further, we call two patterns \emph{duplicates} if they have the same out-neighbor, the same removed sites among $\sE$, $\sN$, and $\sW$, the same present ring bonds, the same retained tip bonds, and the same status at every ring site that is a non-removed site among $\sE$, $\sN$, and $\sW$ or an end of a present ring bond or of a retained tip bond. The functions $I_{\mathrm{cur}}$ and $I_{\mathrm{next}}$ of two duplicates are the same. Their ring functionals differ only in the factors of the remaining non-removed ring sites, those without present ring bonds and retained tip bonds. Such a site is either a corner, which contributes the factor $1 + m \cdot m'/3$ or $1 + m \cdot m'/5$ if it has status $0$ or $1$ and slot vectors $m$ and $m'$, or the arrival site $\sS$, which contributes the factor $1$. We conclude that the ring polynomials of two duplicates differ by a factor that is a positive function of $\mathbf a \in \Ball^{12}$. The file \texttt{patterns.json} lists 7,962 patterns, of which 2,696 are straight and 5,266 are turns with $o = \sE$. For each of them, it records the out-neighbor, the non-removed sites among $\sE$, $\sN$, and $\sW$, the present ring bonds, the retained tip bonds, and the statuses of the ring sites at which these bonds end. The program \texttt{enumerate\_patterns.py}, which was written from \zcref{d:pattern} independently of the program that produced \texttt{patterns.json}, finds 106,101 assignments of statuses and used bonds that satisfy the four numbered conditions of \zcref{d:pattern}, before choosing an out-neighbor. For each assignment, it adds each choice of a non-removed site among $\sE$, $\sN$, and $\sW$ as an out-neighbor. The resulting patterns have exactly the 7,962 classes in \texttt{patterns.json} modulo the reflection and duplicates. Consequently, for every pattern of $\Pext$, the pattern or its reflection is a duplicate of a pattern of \texttt{patterns.json} of the same type. \paragraph{The coefficients.} The module \texttt{ring.py} reads a pattern from \texttt{patterns.json} and reads the tip radius from the environment variable \texttt{TIPRAD}, with the default value $3/5$. It also contains a floating-point evaluator and minimizer of $\hat Z_{\mathrm{cur}}/\hat Z_{\mathrm{next}}$, which only \texttt{run\_prod.py} uses. The module \texttt{cert.py} constructs the retained sites, bonds, and groups for the chosen representation, computes the integer coefficients of $3 \mathfrak d \hat Z_{\mathrm{cur}}$ and $5 \mathfrak d \hat Z_{\mathrm{next}}$, and forms and tests the matrices $\mathsf M_\epsilon$ for the chosen inversion set. For each retained site $y$, it computes the moment tensor of $\lambda_y$, with one index in $\{0, 1, 2, 3\}$ for each factor of \eqref{e:local-hatZ} at $y$ and, if $y \in \{\sE, \sN, \sW\}$, for the bond from $y$ to the middle site. It uses exact rational arithmetic and the formula \begin{equation}\label{e:local-moments} \int (n^1)^{p_1} (n^2)^{p_2} (n^3)^{p_3} \dd n = \frac{(p_1 - 1)!! \, (p_2 - 1)!! \, (p_3 - 1)!!}{(p_1 + p_2 + p_3 + 1)!!} \end{equation} for even integers $p_1, p_2, p_3 \ge 0$, where $(-1)!! = 1$; the integral vanishes if one of the $p_l$ is odd. The moments of $\lambda_1$ of positive degree are $3/5$ times those of $\dd n$. (Formula \eqref{e:local-moments} follows by writing a standard Gaussian vector of $\R^3$ as the product of its norm and an independent uniform unit vector.) The program multiplies each moment tensor by the least common denominator of its entries. For a single vector from a retained tip bond, it multiplies the factor $1 + (3/5) \, b_i \cdot n$ by $5$. For a reduced leaf, it represents the single-vector field on the neighboring axial site by a diagonal transfer tensor at the leaf. If its radius is $r_i = 1/q$ with $q \in \{2, 3\}$, the tensor after scaling is $\diag(q, 1, 1, 1)$. It joins the index of the leaf vector to that of the ring bond, giving $q(1 + r_i \, b_i \cdot n)$ upon contraction. The integer $\mathfrak d$ is the product of the local scales. Since $1 + n_y \cdot n_z = \sum_{j = 0}^3 n_y^j n_z^j$, the coefficients of $\mathfrak d \hat Z[1]$ are obtained by contracting the integer tensors along the present ring bonds, in the order $\sE$, $(1, 1)$, $\sN$, $(-1, 1)$, $\sW$, $(-1, -1)$, $\sS$, $(1, -1)$ of the ring, with the indices of the bonds to the middle site set to $0$. For a pair $\{a', a''\}$ of non-removed sites in $\{\sE, \sN, \sW\}$, the coefficients of $\mathfrak d \hat Z[n_{a'} \cdot n_{a''}]$ are obtained in the same way, with the indices of the bonds from $a'$ and $a''$ to the middle site set to a common value $j \in \{1, 2, 3\}$ and summed over $j$. By \eqref{e:IJ}, this gives \begin{equation*} 3 \mathfrak d \hat Z_{\mathrm{cur}} = 3 \mathfrak d \hat Z[1] + \sum_{\{a', a''\}} \mathfrak d \hat Z[n_{a'} \cdot n_{a''}], \qquad 5 \mathfrak d \hat Z_{\mathrm{next}} = 5 \mathfrak d \hat Z[1] + \mathfrak d \hat Z[n_f \cdot n_g], \end{equation*} where the sum runs over the pairs of non-removed sites in $\{\sE, \sN, \sW\}$, and the last term is absent if $f$ or $g$ is removed. We bound the integers that occur in these contractions. A moment tensor is of degree at most four. By \eqref{e:local-moments}, its least common denominator divides $15$ for $\lambda_0$ and $25$ for $\lambda_1$. Its scaled entries are nonnegative and at most $25$, since the moments have absolute value at most one. The transfer tensor of a reduced leaf has entries of size at most $3$. Each of the at most four retained tip factors contributes an additional scale of at most $5$. Thus any product of local entries is at most $25^8 5^4$. Fix the exterior indices and the indices of the bonds to the middle site in a coefficient. At a ring site, once the index of the incoming ring bond is fixed, the parity of the three coordinate exponents determines at most one index of the outgoing bond for which the moment is nonzero. Indeed, index $0$ leaves all parities unchanged, and each of indices $1$, $2$, and $3$ changes a different parity. A missing bond has index $0$. The transfer tensor of a reduced leaf satisfies the same rule, since its leaf-vector index equals its single ring-bond index. For the closed ring this leaves at most four contributing assignments, one for each possible index on the first bond. For a partial path contraction, the index at its first end is fixed, and the same rule determines the rest of the assignment. Since the summands are nonnegative, all intermediate products and sums are bounded by \begin{equation*} M_{\mathrm{coeff}} = 4 \cdot 25^8 5^4 . \end{equation*} In the coefficients of $3 \mathfrak d \hat Z_{\mathrm{cur}}$ there are three copies of the partition coefficient and at most nine pair-index contributions. The coefficients of $5 \mathfrak d \hat Z_{\mathrm{next}}$ use five copies and at most three contributions. Both sums are bounded by $12 M_{\mathrm{coeff}} < 2^{53}$, including every partial sum. This proves that the $64$-bit integer tensor contractions and coefficient sums are free of overflow. The checker also guards their magnitudes by run-time assertions using binary64 copies of the contractions. The bounds above ensure exact integer representation in these copies. \paragraph{The matrices.} The program then forms the matrices of \zcref{s:local-integer}. It changes the signs of the coefficients according to \eqref{e:local-invcoef}, replaces each index $j$ by the matrix $\sigma_j$, with real and imaginary parts kept separately, and passes to the basis $\mathcal B$ by an integer change of basis on the factors of each pair. This is again done in $64$-bit integers. Before each single-factor Pauli substitution, the magnitude of each input is checked to be less than $2^{61}$, and each output is a signed sum of at most two inputs. Before each pair change of basis, the magnitude is checked to be less than $2^{58}$, and each resulting entry is a signed sum of at most four inputs. If a magnitude assertion fails, the program repeats the assembly with unbounded integers. The program checks exactly that the imaginary parts vanish and that the matrices are symmetric. For every corner of status $0$ that carries a pair, it checks that the vector $e_\uparrow \otimes e_\downarrow - e_\downarrow \otimes e_\uparrow$ of that pair is annihilated from the left and from the right. The program converts the maximum absolute residual of each of these three checks to binary64 for comparison with zero. These residuals are integers, and the conversion preserves whether they vanish. It also checks that the entries joining two elements of $\mathcal B$ with different eigenvalues of $S^3_{\mathrm{tot}}$ or of some $\mathsf X_z$ vanish. These checks follow from \zcref{l:Ostruct} and serve to catch errors in the assembly. Finally, for every sign assignment $\epsilon$, the program extracts the rows and columns of $\mathcal B_\epsilon$, forms $\mathsf M_\epsilon$ by \eqref{e:local-Meps} with unbounded integers, and tests it by the elimination of \zcref{s:local-integer}. \paragraph{Production and checking.} We produced the certificates with the program \texttt{run\_prod.py}. For each pattern, it minimizes the ratio $\hat Z_{\mathrm{cur}}/\hat Z_{\mathrm{next}}$ of the unreduced representation numerically, by coordinate descent. In each step, the ratio is a quotient of two affine functions of one vector, which the program minimizes over that vector in closed form. The program then lets $\mathcal I$ consist of the groups whose vectors have a negative total component along the principal axis of the minimizer, as in \zcref{r:local-choice}. Next, it estimates the largest $c$ for which the forms in \eqref{e:local-certhyp} are positive semidefinite. When this value is not sufficiently above the threshold, it changes $\mathcal I$ by a greedy search over the groups or passes to the unreduced representation. The numerical minimization and eigenvalue estimates guide these choices. The program \texttt{verify\_all.py} is the checker. It reads each record of the certificate file and recomputes the integer coefficients, the matrices $\mathsf M_\epsilon$, and the test with \texttt{cert.py}, from the stored data alone and without any optimization. It refuses to run if assertions are disabled. Its output is the number of certificates verified and the list of failures. The program \texttt{check\_certificate\_file.py} checks the relation between the certificate file and \texttt{patterns.json}, which \texttt{verify\_all.py} does not check; see \zcref{s:local-trust}. The program \texttt{closed\_exact.py} is imported by \texttt{verify\_all.py} and \texttt{run\_prod.py}. It implements a different kind of certificate, which does not occur in the certificate file. \paragraph{The certificate file.} The file \texttt{certificates.jsonl} contains one record for each of the 7,962 patterns. A record consists of the index of the pattern in \texttt{patterns.json}, the threshold $c$ as a fraction $c_1/c_2$, the representation, the inversion set $\mathcal I$, the tip radius $3/5$, and the value \texttt{covered} of a field that names the ring functional \eqref{e:ringfunctional}, in which every non-removed corner carries two slot factors. The reduced representation is used for all patterns except two, which use the unreduced one. The number of vectors is $k \le 11$, there are at most $16$ sign assignments for a pattern, and the matrices $\mathsf M_\epsilon$ have size at most $245$. A record also contains floating-point information, such as the value of $c$ at which some $\mathsf M_\epsilon$ becomes singular, which the checker does not use. The program \texttt{verify\_all.py} accepts all 7,962 certificates, in about $15$ to $30$ CPU-minutes, and its output is recorded in \texttt{anc/h1/verify\_all\_output.log}. Producing the certificates took about $34$ CPU-minutes. \begin{proof}[Proof of \zcref{t:H1}] Let $\pi \in \Pext$. By the paragraph on patterns, there exists a pattern $\pi'$ of \texttt{patterns.json} of the same type as $\pi$ such that $\pi$ or its reflection is a duplicate of $\pi'$. The ring polynomial of $\pi$ at $\mathbf a$ is then a positive multiple of the ring polynomial of $\pi'$ at a relabeling of $\mathbf a$, and $c_\pi = c_{\pi'}$. It therefore suffices to prove \eqref{e:local-goal} for the patterns of \texttt{patterns.json}. The program \texttt{check\_certificate\_file.py} verifies that \texttt{certificates.jsonl} contains exactly one record for each pattern $\pi'$ of \texttt{patterns.json}, with the threshold $c_{\pi'}$, the tip radius $3/5$, and the ring functional \eqref{e:ringfunctional}. For each record, \texttt{verify\_all.py} verifies that $\mathsf M_\epsilon$ is positive definite for every sign assignment $\epsilon$, with the representation and the inversion set of the record and with $c = c_{\pi'}$. By \zcref{l:integertest}, the hypothesis \eqref{e:local-certhyp} of \zcref{p:certificate} holds, and \zcref{p:certificate} gives $\mathcal Z_{\pi'}^{\mathbf a}[I_{\mathrm{cur}} - c_{\pi'} I_{\mathrm{next}}] \ge 0$ for all $\mathbf a \in \Ball^{12}$. This completes the proof. \end{proof} \section{Checks that are not part of the proof}\label{s:local-checks} We also made the following checks. They are not part of the proof, and their programs are not among the ancillary files, but they corroborate the computation of \zcref{s:local-computation} at the points where an error would be hardest to see. The first check concerns the polynomials themselves. An independent evaluator, written from the definitions, starts from the template before the middle site and the tips are integrated out. It keeps these sites as spins, with three slot vectors at each tip whose bond is present. Every site is integrated with a randomly rotated product quadrature rule that is exact for polynomials of degree at most $4$, with an additional node at $0$ for the atom of $\lambda_1$. At three random configurations of the slot vectors for every pattern, it agrees with the ratio $\hat Z_{\mathrm{cur}}/\hat Z_{\mathrm{next}}$ computed from the integer coefficients of \texttt{cert.py} in the representation of the certificate, to a relative error of at most $2.3 \cdot 10^{-15}$. This comparison checks the model, namely the sites, bonds, measures, and radii that enter the polynomials, including the integration of the tips in \zcref{l:local}. Where this was tested, changing the status of a single site changed the ratio by at least $5 \cdot 10^{-5}$ in relative terms. However, a change of one unit in a single integer coefficient can be invisible at this precision when $\mathfrak d$ is large. The second check concerns the exactness of the coefficients. For all 7,962 certificates, the integer matrices of \texttt{cert.py} on $\mathcal V$, written in the basis $\mathcal B$, commute exactly with the raising and lowering operators of the total spin, written in the same basis. This verifies the hypothesis of \zcref{l:sector} directly, in addition to its proof in \zcref{l:Ostruct}. In $30$ random trials, a single coefficient of $3 \mathfrak d \hat Z_{\mathrm{cur}}$ that multiplies a nonconstant monomial was changed by $1$, and in all $30$ trials the perturbed operator failed the exact checks of \zcref{s:local-computation} or this commutation. The third check concerns \zcref{l:sector}. For $63$ patterns, the matrices were tested exactly on every eigenspace of $S^3_{\mathrm{tot}}$ in every $\mathcal V_\epsilon$, and they were positive definite on all of them. These are the $13$ tightest patterns and $50$ patterns with corners of status $1$ that carry a pair, among them the $10$ patterns with the largest operators. The fourth check consists of negative controls. The stored certificates of the tightest patterns, which are accepted at their threshold, are rejected when the threshold is raised $2 \cdot 10^{-5}$ above an independent numerical minimum of the ratio. The same holds for $20$ random patterns. Finally, a second implementation, written independently of the programs of \zcref{s:local-computation}, confirms all 7,962 certificates, with the thresholds determined from the patterns themselves. It shares with them only the definitions, the order of the vectors needed to read the stored inversion sets, and the idea of the inversion. It computes the polynomials by exact symbolic integration, site by site, with the moment formula \eqref{e:local-moments} in integer arithmetic. It assembles the operators on the whole space $(\C^2)^{\otimes k}$ from exchange operators, using \eqref{e:local-swap}, and checks exactly that their entries between different eigenvalues of $S^3_{\mathrm{tot}}$ vanish. On every eigenspace of $S^3_{\mathrm{tot}}$, it proves positivity by an exact integer congruence. It rounds a scaled inverse of a floating-point Cholesky factor of the block to an integer triangular matrix, computes the congruent matrix exactly, and checks that it is strictly diagonally dominant with a positive diagonal. For the patterns with a corner of status $0$ that carries a pair, it checks exactly that the vectors $e_\uparrow \otimes e_\downarrow - e_\downarrow \otimes e_\uparrow$ of these pairs are annihilated, and it proves positive definiteness on the orthogonal complement. It therefore shows that $\mathcal O(\hat\Phi_c^{\mathcal I})$ is positive semidefinite on all of $(\C^2)^{\otimes k}$, so that its conclusion rests on \zcref{l:inversion,c:reduce} alone, and not on \zcref{l:Ostruct,l:sector}. \section{What must be trusted}\label{s:local-trust} A reader who accepts the lemmas of this appendix must trust the following points. \begin{enumerate} \item The list \texttt{patterns.json} represents $\Pext$ modulo the reflection and duplicates. Two independent enumerations from the definition of the class, one of which is \texttt{enumerate\_patterns.py}, confirm this, class by class. \item The certificate file contains each pattern of \texttt{patterns.json} exactly once, with the threshold of its type, the tip radius $3/5$, and the ring functional \eqref{e:ringfunctional}. The checker verifies each record at the threshold written in it, without checking these properties; the program \texttt{check\_certificate\_file.py} checks them. \item The program \texttt{cert.py} computes the coefficients of $3 \mathfrak d \hat Z_{\mathrm{cur}}$ and $5 \mathfrak d \hat Z_{\mathrm{next}}$ correctly. This part has about $130$ lines, together with the parts of \texttt{ring.py} that read a pattern. Its output agrees with the independent evaluator of \zcref{s:local-checks} and passes the exact commutation check there. \item The program \texttt{cert.py} assembles the matrices $\mathsf M_\epsilon$ and selects the bases $\mathcal B_\epsilon$ correctly. This part has about $110$ lines, and its output satisfies the exact checks of \zcref{s:local-computation}. \item The fraction-free elimination of \zcref{s:local-integer} is implemented correctly. It has about $20$ lines. It agrees with an exact $LDL^{\mathsf T}$ decomposition in rational arithmetic on $600$ random symmetric integer test matrices, which include positive definite, singular positive semidefinite, and indefinite matrices with entries up to $2^{200}$. \item The integer and rational arithmetic of Python and the integer tensor contraction of the library \texttt{numpy} are correct in the absence of overflow. The coefficient bound above excludes overflow in the contractions and coefficient sums. The signed operator assembly uses separate integer magnitude assertions and falls back to unbounded arithmetic when needed. The checker must be run with assertions enabled (without the option \texttt{-O} of Python). The program \texttt{verify\_all.py} refuses to run otherwise, and \texttt{check\_certificate\_file.py} checks this as well. \item The tip radius, a parameter of the programs, must be left at its value $3/5$, which means that the environment variable \texttt{TIPRAD} must be unset or equal to $3/5$. With a smaller value, the checker would verify a weaker inequality. The program \texttt{check\_certificate\_file.py} checks this. \end{enumerate} The proof is valid for any representation and inversion set for which the checker accepts the certificate. In particular, it remains valid if the numerical minimization misses the true minimum.