OctoInformBeta

OpenAI preprint, September 2026 · §4. Transverse configurations and determinant signs · §4

Lemma 4.5 [Local transverse approximation]: The gap is at most a sum of pair and triple errors whose mean tends to 00 as q→∞q\to\infty when r≥2r\ge2

On transverse faces the face polynomial approximates the local average.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

The graph HH of the counterexample to Sidorenko’s conjecture is bipartite: thirteen points on one side, twenty-two faces on the other, each face joined to the three points it contains, one in each of its slots 00, 11, 22. In a weighted homomorphism count of HH, once the points carry values, the count factors over the faces: each face contributes the average, over its own value, of the product of the weights of its three edges.

Here, for a natural number rr, the values are symmetric 2r×2r2r\times2r matrices over a finite field. The edge from a face with matrix YY to its point in slot kk, with matrix XkX_k, has as weight a function of Xk−YX_k-Y that depends on the rank rr and on a sign ξk\xi_k assigned to slot kk, for the slots kk in a set TT of active slots; at the other slots the weight is 11. The factor of the face is then its local average FF, a function of X0X_0, X1X_1, X2X_2. The lemma compares FF with an explicit polynomial GG in the quadratic characters of the determinants det⁡(Xi−Xj)\det(X_i-X_j), on faces whose three differences Xi−XjX_i-X_j are nonsingular, and bounds both.

Symmetric Matrices and Their Signs

Let KK be a finite field and χ\chi its quadratic character. For a natural number DD, the space SDS_D of symmetric matrices of size DD over KK is the subspace of the D×DD\times D matrices MM over KK with MT=MM^{T}=M.

An odd prime is a prime q≠2q\neq2, and 𝔽q\mathbb{F}_q is the field of integers modulo qq.✦ A limit as q→∞q\to\infty runs through the odd primes: it is taken along the sets of odd primes that contain every odd prime beyond some bound.✦

Let BB be a symmetric bilinear form on a finite-dimensional vector space VV over KK, and let RR be its radical, the subspace of the vectors x∈Vx\in V with B(x,y)=0B(x,y)=0 for all y∈Vy\in V. The radical quotient form of BB is the bilinear form B‾\bar B on V/RV/R given by B‾(x+R,y+R)=B(x,y)\bar B(x+R,y+R)=B(x,y); it is nondegenerate.✦ The sign of BB is sgn⁡B=χ(det⁡MB‾),\operatorname{sgn}B=\chi\bigl(\det M_{\bar B}\bigr),(1)✦ where MB‾M_{\bar B} is the Gram matrix of B‾\bar B in a basis of V/RV/R; it equals 11 when BB has rank zero. The sign sgn⁡Z\operatorname{sgn}Z of a matrix Z∈SDZ\in S_D is the sign of the symmetric form (x,y)↦xTZy(x,y)\mapsto x^{T}Zy on KDK^D.✦

For natural numbers rr, DD and an integer σ\sigma, the signed rank layer Lr,σL_{r,\sigma} is the set of matrices in SDS_D of rank rr and sign σ\sigma✦, and its layer mass is its share of SDS_D: μr,σ=|Lr,σ||SD|.\mu_{r,\sigma}=\frac{|L_{r,\sigma}|}{|S_D|}.(2)✦ The rank kernel κr,σ\kappa_{r,\sigma} of rank rr and sign σ\sigma is a real function on SDS_D whose value at ZZ depends on whether ZZ lies in Lr,σL_{r,\sigma}.

The uniform mean 𝔼a∈Af(a)\mathbb{E}_{a\in A}f(a) of a real function ff on a finite set AA is 1|A|∑a∈Af(a)\frac{1}{|A|}\sum_{a\in A}f(a).✦ From here on D=2rD=2r and the rank is rr: we write LσL_\sigma, μσ\mu_\sigma, κσ\kappa_\sigma for Lr,σL_{r,\sigma}, μr,σ\mu_{r,\sigma}, κr,σ\kappa_{r,\sigma}, and 𝔼A\mathbb{E}_A for the uniform mean over A∈S2rA\in S_{2r}.

The Face Local Average and the Face Polynomial

A triple X=(X0,X1,X2)X=(X_0,X_1,X_2) of matrices in S2rS_{2r}, indexed by the three slots of a face, is transverse if the differences X0−X1X_0-X_1, X0−X2X_0-X_2 and X1−X2X_1-X_2 are nonsingular.✦

Fix a set T⊆{0,1,2}T\subseteq\{0,1,2\} of active slots and a triple ξ=(ξ0,ξ1,ξ2)\xi=(\xi_0,\xi_1,\xi_2) of integers. The face local average of a triple XX of matrices in S2rS_{2r} is F(X)=𝔼Y∏k∈Tκξk(Xk−Y).F(X)=\mathbb{E}_{Y}\prod_{k\in T}\kappa_{\xi_k}(X_k-Y).(3)✦

Let c0=χ((−1)r)c_0=\chi\bigl((-1)^r\bigr). For integers σ\sigma, τ\tau and B∈S2rB\in S_{2r}, the pair polynomial is gστ(B)=1+c0χ(det⁡B)στg_{\sigma\tau}(B)=1+c_0\,\chi(\det B)\,\sigma\tau.✦ The face polynomial of XX is the product of pair polynomials over the pairs of active slots, G(X)=∏i<ji,j∈Tgξiξj(Xi−Xj),G(X)=\prod_{\substack{i<j\\ i,j\in T}}g_{\xi_i\xi_j}(X_i-X_j),(4)✦ equal to 11 when TT contains no pair of slots.

Pair, Triple and Face Errors

For integers σ\sigma, τ\tau, ω\omega and B,C∈S2rB,C\in S_{2r}, the pair and triple local averages are Pστ(B)=𝔼Aκσ(A)κτ(A−B),Pστω(B,C)=𝔼Aκσ(A)κτ(A−B)κω(A−C).P_{\sigma\tau}(B)=\mathbb{E}_A\,\kappa_\sigma(A)\,\kappa_\tau(A-B),\qquad P_{\sigma\tau\omega}(B,C)=\mathbb{E}_A\,\kappa_\sigma(A)\,\kappa_\tau(A-B)\,\kappa_\omega(A-C).(5)

A matrix BB matches σ,τ\sigma,\tau if χ(det⁡B)=c0στ\chi(\det B)=c_0\sigma\tau✦, and a pair (B,C)(B,C) matches σ,τ,ω\sigma,\tau,\omega if BB matches σ,τ\sigma,\tau, CC matches σ,ω\sigma,\omega and C−BC-B matches τ,ω\tau,\omega.✦ The pair error and the triple error are δστ(B)={|Pστ(B)−2⋅𝟏[B matches σ,τ]|,det⁡B≠0,0,det⁡B=0,\delta_{\sigma\tau}(B)= \begin{cases} \bigl|P_{\sigma\tau}(B)-2\cdot\mathbf{1}[B\text{ matches }\sigma,\tau]\bigr|, & \det B\neq0,\\ 0, & \det B=0, \end{cases}(6)✦ δστω(B,C)={|Pστω(B,C)−8⋅𝟏[(B,C) matches σ,τ,ω]|,B,C,C−B nonsingular,0,otherwise.\delta_{\sigma\tau\omega}(B,C)= \begin{cases} \bigl|P_{\sigma\tau\omega}(B,C)-8\cdot\mathbf{1}[(B,C)\text{ matches }\sigma,\tau,\omega]\bigr|, & B,\ C,\ C-B\text{ nonsingular},\\ 0, & \text{otherwise}. \end{cases}(7)✦

The face error of a triple XX of matrices in S2rS_{2r}, for a triple ξ\xi of integers, is Δξ(X)=δξ0ξ1(X0−X1)+δξ0ξ2(X0−X2)+δξ1ξ2(X1−X2)+δξ0ξ1ξ2(X0−X1,X0−X2).\begin{aligned} \Delta_\xi(X)={}&\delta_{\xi_0\xi_1}(X_0-X_1)+\delta_{\xi_0\xi_2}(X_0-X_2) +\delta_{\xi_1\xi_2}(X_1-X_2)\\ &+\delta_{\xi_0\xi_1\xi_2}(X_0-X_1,\,X_0-X_2). \end{aligned}(8)✦

The Face Bound

For a vector space VV over KK with dual space V*V^{*}, the canonical form on V×V*V\times V^{*} is the bilinear form ⟨(v,f),(w,g)⟩=g(v)+f(w)\langle(v,f),(w,g)\rangle=g(v)+f(w).✦ A self-orthogonal subspace of V×V*V\times V^{*} is a subspace LL equal to its own orthogonal complement for the canonical form; when VV is finite-dimensional, these are the isotropic subspaces of half the dimension of V×V*V\times V^{*}.✦

Let Gr⁡(r,K2r)\operatorname{Gr}(r,K^{2r}) be the set of subspaces of K2rK^{2r} of dimension rr, and Λr\Lambda_r the set of self-orthogonal subspaces of Kr×(Kr)*K^r\times(K^r)^{*}. For integers σ\sigma, τ\tau, ω\omega, the pair and triple coefficients are αστ=|Gr⁡(r,K2r)||S2r|μσμτ,βστω=|Λr||S2r|μσμτμω,\alpha_{\sigma\tau}=\frac{|\operatorname{Gr}(r,K^{2r})|}{|S_{2r}|\,\mu_\sigma\mu_\tau}, \qquad \beta_{\sigma\tau\omega}=\frac{|\Lambda_r|}{|S_{2r}|\,\mu_\sigma\mu_\tau\mu_\omega},(9) and the face bound of a triple ξ\xi of integers is Cξ=1+αξ0ξ1+αξ0ξ2+αξ1ξ2+βξ0ξ1ξ2.C_\xi=1+\alpha_{\xi_0\xi_1}+\alpha_{\xi_0\xi_2}+\alpha_{\xi_1\xi_2}+\beta_{\xi_0\xi_1\xi_2}.(10)✦

The Claim

Theorem 1 (Local transverse approximation). Let KK be a finite field, rr a natural number, TT a subset of the slots {0,1,2}\{0,1,2\}, ξ=(ξ0,ξ1,ξ2)\xi=(\xi_0,\xi_1,\xi_2) a triple of integers and X=(X0,X1,X2)X=(X_0,X_1,X_2) a triple of matrices in S2rS_{2r}. Then F(X)≥0.F(X)\ge0.(11)✦ Suppose from now on that each ξk\xi_k is 11 or −1-1. Then |G(X)|≤8.|G(X)|\le8.(12)✦ If moreover KK has odd characteristic, r≥1r\ge1 and XX is transverse, then |F(X)−G(X)|≤Δξ(X)|F(X)-G(X)|\le\Delta_\xi(X)(13)✦ and F(X)≤Cξ.F(X)\le C_\xi.(14)✦ Let K=𝔽qK=\mathbb{F}_q for an odd prime qq, let each of σ\sigma, τ\tau, ω\omega be 11 or −1-1, write 𝔼B\mathbb{E}_B and 𝔼C\mathbb{E}_C for the uniform means over B∈S2rB\in S_{2r} and C∈S2rC\in S_{2r}, and let q→∞q\to\infty through the odd primes. If r≥1r\ge1, then limq→∞𝔼Bδστ(B)=0\lim_{q\to\infty}\mathbb{E}_B\,\delta_{\sigma\tau}(B)=0(15)✦ and limq→∞Cξ=29.\lim_{q\to\infty}C_\xi=29.(16)✦ If r≥2r\ge2, then limq→∞𝔼B𝔼Cδστω(B,C)=0,\lim_{q\to\infty}\mathbb{E}_B\,\mathbb{E}_C\,\delta_{\sigma\tau\omega}(B,C)=0,(17)✦ and, for any three distinct points v0v_0, v1v_1, v2v_2 of HH, writing 𝔼M\mathbb{E}_M for the mean over all assignments p↦Mpp\mapsto M_p of a matrix Mp∈S2rM_p\in S_{2r} to each of the thirteen points pp, limq→∞𝔼MΔξ(Mv0,Mv1,Mv2)=0.\lim_{q\to\infty}\mathbb{E}_M\,\Delta_\xi(M_{v_0},M_{v_1},M_{v_2})=0.(18)✦

In the paper, the result reads: for r>1r>1 and each face jj, with 𝒯j\mathcal{T}_j the event that the three point differences of jj are nonsingular, 𝔼[𝟏𝒯j|Fj−Gj|]=O(q−1/2)\mathbb{E}\bigl[\mathbf{1}_{\mathcal{T}_j}|F_j-G_j|\bigr]=O(q^{-1/2}) uniformly over activity and signs; moreover 0≤Fj≤CD0\le F_j\le C_D on 𝒯j\mathcal{T}_j with CDC_D independent of qq, and 0≤Gj≤80\le G_j\le8 everywhere; here FjF_j and GjG_j are FF and GG at the matrices on the points of jj, the activity is TT and D=2rD=2r.

The library replaces the rate O(q−1/2)O(q^{-1/2}) by the limit (18) and the constant CDC_D by the face bound CξC_\xi of (14).

What Is Checked

The claim is stated over the space S2rS_{2r} of symmetric matrices over a finite field, the sign of a symmetric matrix through its radical quotient form, the signed rank layers with their masses and rank kernels, transverse triples, the face local average and the pair and face polynomials, the pair, triple and face errors with the local averages and matching conditions they are built from, the face bound with its pair and triple coefficients over the subspaces of dimension rr of K2rK^{2r} and the self-orthogonal subspaces of Kr×(Kr)*K^r\times(K^r)^{*}, and limits as q→∞q\to\infty through the odd primes.