OctoInformBeta

OpenAI preprint, September 2026 · §3. Constructing a negative kernel correlation · §3

§3 (setting): the rank-layer kernels K_ξ and the sign model: Signs of symmetric forms, signed rank layers of SDS_D and their kernels Kr,ξK_{r,\xi}, and layer and sign-model correlations on corners of HH

Kernels from the rank and sign of symmetric matrices.

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 here is the bipartite graph HH whose vertices are 1313 points 0,…,120, \dots, 12 and 2222 faces F0,…,F21F_0, \dots, F_{21}, each face a set of three points, a point being joined to the faces that contain it.

For a real function ff on a finite set AA, 𝔼a∈Af(a)=|A|−1∑a∈Af(a)\mathbb{E}_{a \in A} f(a) = |A|^{-1} \sum_{a \in A} f(a) is its uniform average. A nonnegative function WW on a finite abelian group GG is a weighted bipartite graph between two copies of GG, with weight W(x−y)W(x - y) on the pair (x,y)(x, y), and the density of HH in it is t(H,W)=𝔼X∈G13,Y∈G22∏j=021∏i∈FjW(Xi−Yj).t(H, W) = \mathbb{E}_{X \in G^{13},\ Y \in G^{22}} \prod_{j=0}^{21} \prod_{i \in F_j} W(X_i - Y_j).(1) Sidorenko’s conjecture for HH asks that t(H,W)≥(𝔼W)e(H)t(H, W) \ge (\mathbb{E}\, W)^{e(H)} for every such WW, where e(H)e(H) is the number of edges of HH and 𝔼W\mathbb{E}\, W is the average of WW over GG.

For a finite field 𝔽\mathbb{F} and a natural number DD, the space SD(𝔽)S_D(\mathbb{F}) of symmetric D×DD \times D matrices over 𝔽\mathbb{F} is the subspace of the D×DD \times D matrices over 𝔽\mathbb{F} made of those ZZ with Z⊤=ZZ^{\top} = Z.

Signs of Symmetric Forms

Let 𝔽\mathbb{F} be a finite field, χ\chi its quadratic character, and VV a finite-dimensional vector space over 𝔽\mathbb{F}. The radical quotient β‾\bar\beta of a symmetric bilinear form β\beta on VV is the bilinear form on the quotient of VV by the radical rad⁡β={v∈V:β(v,w)=0 for all w∈V}\operatorname{rad}\beta = \{v \in V : \beta(v, w) = 0 \text{ for all } w \in V\} given by β‾(v+rad⁡β,w+rad⁡β)=β(v,w)(v,w∈V).\bar\beta(v + \operatorname{rad}\beta,\ w + \operatorname{rad}\beta) = \beta(v, w) \qquad (v, w \in V).(2)✦

The sign sgn⁡β\operatorname{sgn}\beta of a symmetric bilinear form β\beta on VV is the quadratic character of the determinant of its radical quotient, sgn⁡β=χ(det⁡M),\operatorname{sgn}\beta = \chi(\det M),(3)✦ where MM is the matrix of β‾\bar\beta in a basis of V/rad⁡βV/\operatorname{rad}\beta; the zero form has sign 11.

The sign of a matrix Z∈SD(𝔽)Z \in S_D(\mathbb{F}) is the sign of the bilinear form it defines on 𝔽D\mathbb{F}^D: sgn⁡Z=sgn⁡βZ,βZ(x,y)=x⊤Zy.\operatorname{sgn} Z = \operatorname{sgn}\beta_Z, \qquad \beta_Z(x, y) = x^{\top} Z y.(4)✦

Signed Rank Layers

For a natural number rr and an integer ξ\xi, the signed rank layer Lr,ξL_{r,\xi} of SD(𝔽)S_D(\mathbb{F}) is the set of symmetric matrices of rank rr and sign ξ\xi: Lr,ξ={Z∈SD(𝔽):rank⁡Z=r,sgn⁡Z=ξ}.L_{r,\xi} = \{Z \in S_D(\mathbb{F}) : \operatorname{rank} Z = r,\ \operatorname{sgn} Z = \xi\}.(5)✦

The mass mr,ξm_{r,\xi} of the layer Lr,ξL_{r,\xi} is the probability that a uniformly random symmetric matrix lies in it: mr,ξ=|Lr,ξ||SD(𝔽)|.m_{r,\xi} = \frac{|L_{r,\xi}|}{|S_D(\mathbb{F})|}.(6)✦

Corners and Point Pairs

The 2222 faces are the following triples of points: F0,…,F5:{0,1,3},{0,1,9},{0,2,3},{0,2,9},{1,3,10},{1,7,10},F6,…,F11:{1,7,12},{1,9,12},{2,3,4},{2,4,9},{3,4,11},{3,10,11},F12,…,F17:{4,8,9},{4,8,11},{5,6,7},{5,6,11},{5,7,10},{5,10,11},F18,…,F21:{6,7,12},{6,8,11},{6,8,12},{8,9,12}.\begin{aligned} F_0, \dots, F_5 &: \{0,1,3\},\ \{0,1,9\},\ \{0,2,3\},\ \{0,2,9\},\ \{1,3,10\},\ \{1,7,10\},\\ F_6, \dots, F_{11} &: \{1,7,12\},\ \{1,9,12\},\ \{2,3,4\},\ \{2,4,9\},\ \{3,4,11\},\ \{3,10,11\},\\ F_{12}, \dots, F_{17} &: \{4,8,9\},\ \{4,8,11\},\ \{5,6,7\},\ \{5,6,11\},\ \{5,7,10\},\ \{5,10,11\},\\ F_{18}, \dots, F_{21} &: \{6,7,12\},\ \{6,8,11\},\ \{6,8,12\},\ \{8,9,12\}. \end{aligned}(7)✦

The set CC of corners is the set of pairs (i,j)(i, j) of a point ii and a face index jj with i∈Fji \in F_j, C={(i,j)∈{0,…,12}×{0,…,21}:i∈Fj},C = \{(i, j) \in \{0, \dots, 12\} \times \{0, \dots, 21\} : i \in F_j\},(8)✦ the 6666 edges of HH.

The set PP of point pairs is the set of two-element sets of points contained in some face: P=⋃j=021{e⊆Fj:|e|=2}.P = \bigcup_{j=0}^{21} \{e \subseteq F_j : |e| = 2\}.(9)✦

The Claim

Definition 1. Let 𝔽\mathbb{F} be a finite field, DD and rr natural numbers and ξ\xi an integer. The rank-layer kernel Kr,ξK_{r,\xi} is the function on SD(𝔽)S_D(\mathbb{F}) equal to the indicator of the layer Lr,ξL_{r,\xi} divided by its mass: Kr,ξ(Z)=𝟏[Z∈Lr,ξ]mr,ξ.K_{r,\xi}(Z) = \frac{\mathbf{1}[Z \in L_{r,\xi}]}{m_{r,\xi}}.(10)✦ Let qq be an odd prime, 𝔽q\mathbb{F}_q the field of integers modulo qq, DD a natural number and r=⌊D/2⌋r = \lfloor D/2 \rfloor; let BB be a set of pairs (i,j)(i, j) of a point ii and a face index jj, and ξ\xi an assignment of an integer ξij\xi_{ij} to each such pair. The layer correlation of BB and ξ\xi is the average, over X∈SD(𝔽q)13X \in S_D(\mathbb{F}_q)^{13} and Y∈SD(𝔽q)22Y \in S_D(\mathbb{F}_q)^{22}, of the product of kernels over BB: Λq,D(B,ξ)=𝔼X,Y∏(i,j)∈BKr,ξij(Xi−Yj).\Lambda_{q,D}(B, \xi) = \mathbb{E}_{X, Y} \prod_{(i, j) \in B} K_{r, \xi_{ij}}(X_i - Y_j).(11)✦ With χ\chi the quadratic character of 𝔽q\mathbb{F}_q and c0=χ((−1)r)c_0 = \chi\bigl((-1)^r\bigr), the sign model correlation of BB and ξ\xi is the average over sign vectors η=(ηe)e∈P\eta = (\eta_e)_{e \in P} in {1,−1}P\{1, -1\}^P: Σq,D(B,ξ)=𝔼η∏j=021∏{i,k}⊆Fj,i≠k(i,j),(k,j)∈B(1+c0η{i,k}ξijξkj).\Sigma_{q,D}(B, \xi) = \mathbb{E}_{\eta} \prod_{j=0}^{21} \prod_{\substack{\{i,k\} \subseteq F_j,\ i \neq k \\ (i,j),\,(k,j) \in B}} \bigl(1 + c_0\, \eta_{\{i,k\}}\, \xi_{ij}\, \xi_{kj}\bigr).(12)✦

In the paper, the result reads: for an odd prime qq, the sign of a symmetric matrix of positive rank is the quadratic character of the determinant of the form it induces on the quotient by its radical; with D=2rD = 2r, the kernel is Kξ(Z)=𝟏[rank⁡Z=r,sgn⁡Z=ξ]/ℙ(rank⁡=r,sgn⁡=ξ)K_\xi(Z) = \mathbf{1}[\operatorname{rank} Z = r,\ \operatorname{sgn} Z = \xi] / \mathbb{P}(\operatorname{rank} = r,\ \operatorname{sgn} = \xi) and c0=χ((−1)r)c_0 = \chi((-1)^r); and the sign model is 𝔼η∏j∏e⊂Fj(1+c0ηeξijξkj)\mathbb{E}_\eta \prod_j \prod_{e \subset F_j} (1 + c_0 \eta_e \xi_{ij} \xi_{kj}), the inner product over the pairs e={i,k}e = \{i, k\} with both corners (i,j)(i, j) and (k,j)(k, j) in BB, with independent uniform signs ηe\eta_e.

What Is Checked

The sign, the signed rank layers, their masses and the kernels are stated over an arbitrary finite field 𝔽\mathbb{F}, dimension DD and rank rr, with 𝔼\mathbb{E} the uniform average; the layer correlation and the sign model are stated over 𝔽q\mathbb{F}_q for an odd prime qq and any DD, with the faces (7) and the point pairs (9), for any set BB of point-face pairs, the corners (8) among them, and any integer labels ξij\xi_{ij}.

About these constructions the library proves that the kernels have mean one: for a finite field 𝔽\mathbb{F} of characteristic other than 22, natural numbers rr and DD with 0<r≤D0 < r \le D, and ξ∈{1,−1}\xi \in \{1, -1\}, 𝔼Z∈SD(𝔽)Kr,ξ(Z)=1.\mathbb{E}_{Z \in S_D(\mathbb{F})} K_{r,\xi}(Z) = 1.(13)✦