OctoInformBeta

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

Proposition 4.7 [Transverse limit]: The gap between kernel mean and model splits into a transverse difference and two singular parts, and vanishes as q→∞q\to\infty for signs ±1\pm1

Kernel means approach a Boolean sign model in large dimension.

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

Sidorenko’s conjecture asks whether t(H,G)≥p(G)|E(H)|t(H,G)\ge p(G)^{|E(H)|} for every bipartite graph HH and every graph GG, where t(H,G)t(H,G) is the homomorphism density of HH in GG and p(G)=2|E(G)|/|V(G)|2p(G)=2|E(G)|/|V(G)|^2 is the edge density of GG. Here HH is the bipartite graph between 13 points and 22 faces, each face a triple of points, that joins each point to the faces containing it. For a function KK on the symmetric matrices over a finite field, the homomorphism density of HH in the weighted graph on these matrices with weight K(X−Y)K(X-Y) between XX and YY is the mean, over a matrix XiX_i at each point ii and a matrix YjY_j at each face jj, of the product of the weights K(Xi−Yj)K(X_i-Y_j) over the edges of HH.

The points are numbered 0,…,120,\dots,12 and the faces 0,…,210,\dots,21. A fixed table lists the points v0(j)<v1(j)<v2(j)v_0(j)<v_1(j)<v_2(j) of each face jj in its slots 00, 11, 22, and a corner is a pair (j,a)(j,a) of a face and a slot, at the point va(j)v_a(j).✦ A second table holds the sides {v0(j),v1(j)}\{v_0(j),v_1(j)\}, {v0(j),v2(j)}\{v_0(j),v_2(j)\} and {v1(j),v2(j)}\{v_1(j),v_2(j)\} of each face jj in slots 00, 11, 22.✦ The sides run over 33 point pairs, each of them a side of exactly two faces, and a pair is written e={i,k}e=\{i,k\} with its left point ii smaller than its right point kk.

Proposition 4.7 takes such means with a kernel of prescribed sign at each of a chosen set of corners, the active corners, and compares them, as the finite field 𝔽q\mathbb{F}_q of odd prime order qq grows, with a model that sees the 13 point matrices only through the quadratic characters of the determinants of their differences along the 33 sides. The comparison is split along the event that all 33 differences are nonsingular.

Signs and Rank Kernels

Let 𝔽\mathbb{F} be a finite field with quadratic character χ\chi. The space SDS_D of symmetric matrices of size DD over 𝔽\mathbb{F} is the subspace of the D×DD\times D matrices MM over 𝔽\mathbb{F} with MT=MM^{T}=M.✦ For a symmetric bilinear form BB on a finite-dimensional space VV over 𝔽\mathbb{F} with radical RR, the radical quotient form is the nondegenerate form B‾(x+R,y+R)=B(x,y)\bar B(x+R,y+R)=B(x,y) on V/RV/R✦, and the discriminant sign of BB is χ\chi of the determinant of a Gram matrix of B‾\bar B, which is 11 when BB has rank zero.✦ The sign of a matrix Z∈SDZ\in S_D is the discriminant sign of the form (x,y)↦xTZy(x,y)\mapsto x^{T}Zy on 𝔽D\mathbb{F}^D.✦

The signed rank layer LξL_\xi of rank rr and sign ξ\xi is the set of matrices in SDS_D of rank rr and sign ξ\xi✦, and its layer mass is the proportion |Lξ|/|SD||L_\xi|/|S_D|.✦ The rank kernel KξK_\xi of rank rr and sign ξ\xi is a real function on SDS_D whose value at ZZ is determined by whether ZZ lies in LξL_\xi. Throughout, D=2rD=2r, every kernel has rank rr, and c0=χ((−1)r)c_0=\chi((-1)^r).

A configuration is a family X=(X0,…,X12)X=(X_0,\dots,X_{12}) of matrices in S2rS_{2r}, one at each point. Means are uniform: 𝔼X\mathbb{E}_X averages over all configurations and 𝔼Y\mathbb{E}_Y over all Y∈S2rY\in S_{2r}.✦

An activity pattern TT assigns to each face jj a set Tj⊆{0,1,2}T_j\subseteq\{0,1,2\} of active slots, and ξ=(ξja)\xi=(\xi_{ja}) assigns an integer to each corner. The face local mean of jj at a configuration XX is Fj(X)=𝔼Y∏a∈TjKξja(Xva(j)−Y),F_j(X)=\mathbb{E}_Y\prod_{a\in T_j}K_{\xi_{ja}}\bigl(X_{v_a(j)}-Y\bigr),(1)✦ and the configuration product is ∏jFj(X)\prod_jF_j(X), over the 22 faces.✦

For B∈S2rB\in S_{2r} and integers ξ\xi and ζ\zeta, the pair polynomial is Pξ,ζ(B)=1+c0χ(det⁡B)ξζ.P_{\xi,\zeta}(B)=1+c_0\,\chi(\det B)\,\xi\zeta.(2)✦ The face polynomial Gj(X)G_j(X) of a face jj at XX is the product of one pair polynomial for each pair of active slots of jj, Gj(X)=∏a<b,a,b∈TjPξja,ξjb(Xva(j)−Xvb(j)),G_j(X)=\prod_{a<b,\ a,b\in T_j}P_{\xi_{ja},\xi_{jb}}\bigl(X_{v_a(j)}-X_{v_b(j)}\bigr),(3) and the configuration polynomial is ∏jGj(X)\prod_jG_j(X).✦

A configuration XX is fully transverse when det⁡(Xi−Xk)≠0\det(X_i-X_k)\ne0 for each of the 33 pairs {i,k}\{i,k\}; 𝒯\mathcal T is the event that XX is fully transverse, and 𝒯c\mathcal T^{c} its complement.✦ The pair sign of XX at a pair e={i,k}e=\{i,k\} with i<ki<k is χ(det⁡(Xi−Xk))\chi(\det(X_i-X_k)), which is ±1\pm1 when the difference is invertible and 00 otherwise.✦ The actual signs of XX form the vector ηX∈{±1}33\eta^X\in\{\pm1\}^{33} with ηeX=1\eta^X_e=1 when the pair sign at ee is 11 and ηeX=−1\eta^X_e=-1 otherwise.✦

For a real number cc and a sign vector η∈{±1}33\eta\in\{\pm1\}^{33}, the Boolean configuration is Φc(η)=∏j∏a<b,a,b∈Tj(1+cηej(a,b)ξjaξjb),\Phi_c(\eta)=\prod_j\ \prod_{a<b,\ a,b\in T_j}\bigl(1+c\,\eta_{e_j(a,b)}\,\xi_{ja}\,\xi_{jb}\bigr),(4) where ej(a,b)e_j(a,b) is the side of jj through va(j)v_a(j) and vb(j)v_b(j); the product over the pairs a<ba<b in TjT_j is the Boolean face polynomial of jj. The model at a configuration XX is Φc0(ηX)\Phi_{c_0}(\eta^X).

Transverse and Singular Parts

The transverse difference Δ𝒯\Delta_{\mathcal T} is the mean of the indicator of 𝒯\mathcal T times the difference of the configuration product and the configuration polynomial: Δ𝒯=𝔼X[𝟏𝒯(X)(∏jFj(X)−∏jGj(X))].\Delta_{\mathcal T}=\mathbb{E}_X\Bigl[\mathbf 1_{\mathcal T}(X)\Bigl(\prod_jF_j(X)-\prod_jG_j(X)\Bigr)\Bigr].(5) The matrix singular tail and the Boolean singular part are the means of the configuration product and of the model off 𝒯\mathcal T: Σmat=𝔼X[𝟏𝒯c(X)∏jFj(X)],\Sigma_{\mathrm{mat}}=\mathbb{E}_X\Bigl[\mathbf 1_{\mathcal T^{c}}(X)\prod_jF_j(X)\Bigr],(6)✦ Σbool=𝔼X[𝟏𝒯c(X)Φc0(ηX)].\Sigma_{\mathrm{bool}}=\mathbb{E}_X\bigl[\mathbf 1_{\mathcal T^{c}}(X)\,\Phi_{c_0}(\eta^X)\bigr].(7)✦

Each of these depends on 𝔽\mathbb{F}. A limit as q→∞q\to\infty is taken through the odd primes, the primes q≠2q\ne2, with 𝔽=𝔽q\mathbb{F}=\mathbb{F}_q.✦

The Singular Tail Threshold

A profile (c,h)(c,h) assigns a natural number cjc_j to each face jj and heh_e to each pair ee; it is geometric in dimension DD when, for every face jj, cj≤Dc_j\le D, cj≤hec_j\le h_e for each side ee of jj, and the profile sizes sj,a=he−cjs_{j,a}=h_e-c_j, with ee the side of jj in slot aa, satisfy sj,0+sj,1+sj,2≤D−cjs_{j,0}+s_{j,1}+s_{j,2}\le D-c_j.✦

A fixed table gives each face a three-bit label, and the separation count σe\sigma_e of a pair ee is the number of bits at which the labels of the two faces containing ee differ.✦ With the triangular function b(n)=n(n+1)/2b(n)=n(n+1)/2, the profile cost of (c,h)(c,h) in dimension DD is κD(c,h)=12∑jNj−16∑e(3−σe)b(he),Nj=(D+1)cj+∑a=02(sj,a+12),\kappa_D(c,h)=\frac12\sum_jN_j-\frac16\sum_e(3-\sigma_e)\,b(h_e),\qquad N_j=(D+1)\,c_j+\sum_{a=0}^{2}\binom{s_{j,a}+1}{2},(8)✦ where NjN_j is the face natural cost of jj.

A tail tuple (c,u,v,p,t)(c,u,v,p,t) consists of natural numbers cc, uu, vv and triples pp, tt of natural numbers✦; its tail sizes are pa+tap_a+t_a✦, its tail heights c+pa+tac+p_a+t_a✦, and its tail offset is δ=2u−c\delta=2u-c.✦ Take a geometric profile in dimension D=2rD=2r and any set of faces, whose members are called the full faces. A global tail choice xx picks at each full face jj a common parameter uu, an integer with 0≤u≤min⁡(cj,r)0\le u\le\min(c_j,r) and r−u≤D−cjr-u\le D-c_j, and feasible base parameters θ=(θ0,θ1,θ2,θ3)\theta=(\theta_0,\theta_1,\theta_2,\theta_3): integers between 00 and d=D−cjd=D-c_j such that, with r′=r−ur'=r-u, for every aa and every b≠ab\ne a in {0,1,2}\{0,1,2\}, r′≤θ3,θa+θb≤r′,∑m=03θm≤3r′,θa≤sj,a,sj,a+2r′≤d+2θa,2r′≤θ3+θa.r'\le\theta_3,\quad \theta_a+\theta_b\le r',\quad \sum_{m=0}^{3}\theta_m\le3r',\quad \theta_a\le s_{j,a},\quad s_{j,a}+2r'\le d+2\theta_a,\quad 2r'\le\theta_3+\theta_a.(9)✦

The global tail tuple of xx at a full face jj is the planted tail tuple (cj,u,d−θ3,(θa)a,(sj,a−θa)a)(c_j,u,d-\theta_3,(\theta_a)_a,(s_{j,a}-\theta_a)_a), and at any other face the lower tail tuple (cj,0,0,(sj,a)a,(0,0,0))(c_j,0,0,(s_{j,a})_a,(0,0,0)).✦ The tail gain of a full face is the planted gain of its tuple, g=(d+c2+c+1)c+δ2−2(c+δ2)2+∑ab(pa+ta)−δ∑apa−∑apa2−v(d−3δ2−∑apa+v)−∑ab(ta)+b(v),\begin{split} g={}&\Bigl(\frac{d+c}{2}+c+1\Bigr)\frac{c+\delta}{2}-2\Bigl(\frac{c+\delta}{2}\Bigr)^{2} +\sum_a b(p_a+t_a)-\delta\sum_a p_a-\sum_a p_a^{2}\\ &-v\Bigl(\frac{d-3\delta}{2}-\sum_a p_a+v\Bigr)-\sum_a b(t_a)+b(v), \end{split}(10)✦ with d=D−cd=D-c, and the tail gain of any other face is b(⌊m/2⌋)b(\lfloor m/2\rfloor), where mm is the largest tail height of its tuple. The global tail gain γ(x)\gamma(x) is the sum of the 22 tail gains.✦

The singular tail threshold D0D_0 is an integer, at least 16421642, for which the following global tail dichotomy holds.✦ Let D=2r≥D0D=2r\ge D_0, let (c,h)(c,h) be a geometric profile in dimension DD, and let xx be a global tail choice for any set of full faces. Then γ(x)<κD(c,h)or(c=0,he∈{0,2}for every pair e,γ(x)=23∑eb(he)).\gamma(x)<\kappa_D(c,h)\quad\text{or}\quad \Bigl(c=0,\ \ h_e\in\{0,2\}\ \text{for every pair } e,\ \ \gamma(x)=\tfrac23\sum_e b(h_e)\Bigr).(11)✦

The Claim

Theorem 1. Let rr be a natural number, TT an activity pattern and ξ\xi a family of integers on the corners. Over every finite field 𝔽\mathbb{F}, 𝔼X∏jFj(X)−𝔼XΦc0(ηX)=Δ𝒯+Σmat−Σbool.\mathbb{E}_X\prod_jF_j(X)-\mathbb{E}_X\,\Phi_{c_0}(\eta^X) =\Delta_{\mathcal T}+\Sigma_{\mathrm{mat}}-\Sigma_{\mathrm{bool}}.(12)✦ Suppose now that every ξja\xi_{ja} is 11 or −1-1, and let q→∞q\to\infty through the odd primes, with 𝔽=𝔽q\mathbb{F}=\mathbb{F}_q. Then the Boolean singular part tends to 00, limq→∞Σbool=0;\lim_{q\to\infty}\Sigma_{\mathrm{bool}}=0;(13)✦ if r>1r>1, the transverse difference tends to 00, limq→∞Δ𝒯=0;\lim_{q\to\infty}\Delta_{\mathcal T}=0;(14)✦ and if 2r≥D02r\ge D_0, the mean of the configuration product and the mean of the model have the same limit, limq→∞(𝔼X∏jFj(X)−𝔼XΦc0(ηX))=0.\lim_{q\to\infty}\Bigl(\mathbb{E}_X\prod_jF_j(X)-\mathbb{E}_X\,\Phi_{c_0}(\eta^X)\Bigr)=0.(15)✦

In the paper, the result reads: for every fixed even D≥12D\ge12, every activity pattern BB and all incidence signs, 𝔼[𝟏𝒯∏(i,j)∈BKξij(Xi−Yj)]=𝔼η∏j∏{i,k}⊆Aj(1+c0ηikξijξkj)+o(1),\mathbb{E}\Bigl[\mathbf 1_{\mathcal T}\prod_{(i,j)\in B}K_{\xi_{ij}}(X_i-Y_j)\Bigr] =\mathbb{E}_\eta\prod_j\prod_{\{i,k\}\subseteq A_j}\bigl(1+c_0\,\eta_{ik}\,\xi_{ij}\,\xi_{kj}\bigr)+o(1),(16) uniformly over patterns and signs, where the paper’s DD is the dimension 2r2r, BB is the set of active corners, AjA_j is the set of points of jj at its active corners, ξij\xi_{ij} is the sign at the corner of jj at the point ii, YjY_j is the matrix at the face jj, and η=(ηik)\eta=(\eta_{ik}) is a family of signs on the point pairs.

The library states the result for fixed rr, TT and ξ\xi, as limits through the odd primes, and compares the mean of the configuration product with the mean of the model Φc0(ηX)\Phi_{c_0}(\eta^X) at the actual signs of the configuration rather than with a mean over η\eta.

What Is Checked

The claim is stated over the rank kernels of rank rr on S2rS_{2r}, the face local means FjF_j and face polynomials GjG_j of an activity pattern, the Boolean configuration Φc0\Phi_{c_0} at the actual signs ηX\eta^X, the event 𝒯\mathcal T of fully transverse configurations, uniform means over configurations, limits as q→∞q\to\infty through the odd primes, and the singular tail threshold D0D_0.