OctoInformBeta

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

Lemma 3.1 [Rank-layer limit]: As q→∞q\to\infty through odd primes, once DD reaches a threshold D*≥1642D_\ast\ge1642; negating a rank kernel’s argument multiplies its sign by χ((−1)r)\chi((-1)^r)

Layer correlations approach the sign model in large even 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

The counterexample to Sidorenko’s conjecture is built on a configuration of 1313 points, numbered 0,…,120,\dots,12, and 2222 faces, numbered 0,…,210,\dots,21, the face jj being a set FjF_j of three points. In order, the faces are {0,1,3}\{0,1,3\}, {0,1,9}\{0,1,9\}, {0,2,3}\{0,2,3\}, {0,2,9}\{0,2,9\}, {1,3,10}\{1,3,10\}, {1,7,10}\{1,7,10\}, {1,7,12}\{1,7,12\}, {1,9,12}\{1,9,12\}, {2,3,4}\{2,3,4\}, {2,4,9}\{2,4,9\}, {3,4,11}\{3,4,11\}, {3,10,11}\{3,10,11\}, {4,8,9}\{4,8,9\}, {4,8,11}\{4,8,11\}, {5,6,7}\{5,6,7\}, {5,6,11}\{5,6,11\}, {5,7,10}\{5,7,10\}, {5,10,11}\{5,10,11\}, {6,7,12}\{6,7,12\}, {6,8,11}\{6,8,11\}, {6,8,12}\{6,8,12\} and {8,9,12}\{8,9,12\}.✦

The graph HH is the incidence graph of the configuration: the bipartite graph on the points and the faces that joins each point to the faces containing it. A corner is a pair (v,j)(v,j) of a point vv and a face jj with v∈Fjv\in F_j, and CC is the set of corners.✦

Fix an odd prime qq, write 𝔽q=ℤ/qℤ\mathbb F_q=\mathbb Z/q\mathbb Z, and fix an even dimension DD. Every point vv receives a symmetric D×DD\times D matrix XvX_v over 𝔽q\mathbb F_q and every face jj such a matrix YjY_j, all independent and uniform. A set BB of corners spans a subgraph of HH, and each of its corners (v,j)(v,j) is weighted by a kernel of rank D/2D/2, with a sign ±1\pm1 chosen for that corner, evaluated at Xv−YjX_v-Y_j. The expected product of these weights is a homomorphism density of the subgraph, each edge carrying its own kernel. The question is how this density compares with the sign model correlation as q→∞q\to\infty with DD fixed, and how a kernel changes when Xv−YjX_v-Y_j is replaced by Yj−XvY_j-X_v.

Symmetric Matrices, Signs and Rank Kernels

Let 𝔽\mathbb F be a finite field and χ\chi its quadratic character. The space SD(𝔽)S_D(\mathbb F) of symmetric matrices is the subspace of the D×DD\times D matrices over 𝔽\mathbb F formed by the matrices ZZ with Z𝖳=ZZ^{\mathsf T}=Z, and SDS_D is SD(𝔽q)S_D(\mathbb F_q).✦

For a symmetric bilinear form β\beta on a finite-dimensional vector space VV over 𝔽\mathbb F, the radical rad⁡β\operatorname{rad}\beta is the subspace of the vectors xx with β(x,y)=0\beta(x,y)=0 for every y∈Vy\in V, and the radical quotient form β‾\bar\beta is the nondegenerate form that β\beta induces on V/rad⁡βV/\operatorname{rad}\beta.✦ The sign of β\beta is sgn⁡β=χ(det⁡G),\operatorname{sgn}\beta=\chi(\det G),(1)✦ where GG is the Gram matrix of β‾\bar\beta in a basis of V/rad⁡βV/\operatorname{rad}\beta; it is 11 when β=0\beta=0. The sign sgn⁡Z\operatorname{sgn}Z of a matrix Z∈SD(𝔽)Z\in S_D(\mathbb F) is the sign of the form (x,y)↦x𝖳Zy(x,y)\mapsto x^{\mathsf T}Zy on 𝔽D\mathbb F^D.✦

For a natural number rr and an integer ξ\xi, the signed rank layer Lr,ξL_{r,\xi} is the set of matrices Z∈SD(𝔽)Z\in S_D(\mathbb F) of rank rr with sgn⁡Z=ξ\operatorname{sgn}Z=\xi✦, and its mass is μr,ξ=|Lr,ξ||SD(𝔽)|.\mu_{r,\xi}=\frac{|L_{r,\xi}|}{|S_D(\mathbb F)|}.(2)✦ For a natural number rr and an integer ξ\xi, the rank kernel Kr,ξK_{r,\xi} of rank rr and sign ξ\xi is the real function on SD(𝔽)S_D(\mathbb F) attached to the layer Lr,ξL_{r,\xi}.

Layer Correlations

The uniform mean of a real function ff on a finite set AA is 𝔼a∈Af(a)=|A|−1∑a∈Af(a)\mathbb E_{a\in A}f(a)=|A|^{-1}\sum_{a\in A}f(a).✦ For a natural number DD, an odd prime qq, a set BB of pairs (v,j)(v,j) of a point and a face, and an integer function ξ\xi on such pairs, the layer correlation is the uniform mean ΛD(q;B,ξ)=𝔼X,Y∏(v,j)∈BK⌊D/2⌋,ξ(v,j)(Xv−Yj)\Lambda_D(q;B,\xi)=\mathbb E_{X,Y}\prod_{(v,j)\in B}K_{\lfloor D/2\rfloor,\xi(v,j)}(X_v-Y_j)(3)✦ over the families X=(Xv)X=(X_v) of matrices in SDS_D indexed by the points and Y=(Yj)Y=(Y_j) indexed by the faces. For a natural number DD, an odd prime qq, a set BB of pairs of a point and a face and an integer function ξ\xi on such pairs, the sign model correlation MD(q;B,ξ)M_D(q;B,\xi) is the real number that the sign model, the comparison model for the layer correlation, assigns to these data.

A real function ff on the odd primes tends to 00 as q→∞q\to\infty through odd primes when for every ε>0\varepsilon>0 there is a bound QQ with |f(q)|<ε|f(q)|<\varepsilon for every odd prime q≥Qq\ge Q.

The Singular Tail Threshold

A point pair is a set of two points lying together in some face.✦ There are 3333 point pairs ee, numbered 0,…,320,\dots,32, and each lies in exactly two faces. The sides of the face jj, with Fj={v0,v1,v2}F_j=\{v_0,v_1,v_2\} and v0<v1<v2v_0<v_1<v_2, are ej,0={v0,v1}e_{j,0}=\{v_0,v_1\}, ej,1={v0,v2}e_{j,1}=\{v_0,v_2\} and ej,2={v1,v2}e_{j,2}=\{v_1,v_2\}.✦

Each face jj carries a three-bit label λj∈{0,1}3\lambda_j\in\{0,1\}^3, from an explicit table that runs from λ0=000\lambda_0=000 and λ1=010\lambda_1=010 to λ21=111\lambda_{21}=111. The separation count σ(e)\sigma(e) of a point pair ee is the Hamming distance between the labels of the two faces containing ee, a number from 00 to 33, and ae=σ(e)/3a_e=\sigma(e)/3.✦

A profile is a pair (c,h)(c,h) of a natural number cjc_j for each face jj and a natural number heh_e for each point pair ee; write sj,i=hej,i−cjs_{j,i}=h_{e_{j,i}}-c_j for i=0,1,2i=0,1,2. The profile is geometric in dimension DD when, for every face jj and every ii, cj≤D,cj≤hej,i,sj,0+sj,1+sj,2≤D−cj.c_j\le D,\qquad c_j\le h_{e_{j,i}},\qquad s_{j,0}+s_{j,1}+s_{j,2}\le D-c_j.(4)✦

In dimension DD, the natural cost of the face jj is Nj=(D+1)cj+∑i=02(sj,i+12)N_j=(D+1)c_j+\sum_{i=0}^{2}\binom{s_{j,i}+1}{2}✦, and the profile cost is Cost⁡D(c,h)=12∑jNj−12∑e(1−ae)b(he),\operatorname{Cost}_D(c,h)=\frac12\sum_jN_j-\frac12\sum_e(1-a_e)\,b(h_e),(5)✦ where b(x)=x(x+1)/2b(x)=x(x+1)/2 is the triangular function✦.

A tail tuple τ=(c,u,w,p,t)\tau=(c,u,w,p,t) consists of natural numbers cc, uu, ww and triples p=(p0,p1,p2)p=(p_0,p_1,p_2) and t=(t0,t1,t2)t=(t_0,t_1,t_2) of natural numbers.✦ For a tail tuple τ=(c,u,w,p,t)\tau=(c,u,w,p,t), the sizes are si(τ)=pi+tis_i(\tau)=p_i+t_i, the heights are hi(τ)=c+si(τ)h_i(\tau)=c+s_i(\tau), the maximal height is hmax(τ)=max⁡ihi(τ)h_{\max}(\tau)=\max_ih_i(\tau) and the offset is the real number δ(τ)=2u−c\delta(\tau)=2u-c.

The planted gain of real numbers c,d,δ,wc,d,\delta,w and real triples p,tp,t is Φ(c,d,δ,w,p,t)=(d+c2+c+1)c+δ2−2(c+δ2)2+∑ib(pi+ti)−δ∑ipi−∑ipi2−w(d−3δ2−∑ipi+w)−∑ib(ti)+b(w).\begin{aligned} \Phi(c,d,\delta,w,p,t)={}&\Bigl(\frac{d+c}{2}+c+1\Bigr)\frac{c+\delta}{2} -2\Bigl(\frac{c+\delta}{2}\Bigr)^2+\sum_ib(p_i+t_i)-\delta\sum_ip_i-\sum_ip_i^2\\ &-w\Bigl(\frac{d-3\delta}{2}-\sum_ip_i+w\Bigr)-\sum_ib(t_i)+b(w). \end{aligned}(6)✦ In dimension DD, the full tail gain of a tail tuple τ=(c,u,w,p,t)\tau=(c,u,w,p,t) is Φ(c,D−c,δ(τ),w,p,t)\Phi(c,D-c,\delta(\tau),w,p,t), and its lower tail gain is b(⌊hmax(τ)/2⌋)b(\lfloor h_{\max}(\tau)/2\rfloor).✦

Let DD and rr be natural numbers, cc a natural number and s=(s0,s1,s2)s=(s_0,s_1,s_2) a triple of sizes. The lower tail tuple of cc and ss is (c,0,0,s,(0,0,0))(c,0,0,s,(0,0,0)).✦ A common parameter is an integer uu with 0≤u≤c0\le u\le c, u≤ru\le r and r−u≤D−cr-u\le D-c.✦ For a rank ρ\rho and a dimension dd, a base parameter is a quadruple η=(η0,η1,η2,η3)\eta=(\eta_0,\eta_1,\eta_2,\eta_3) of integers in {0,…,d}\{0,\dots,d\} that is feasible✦, which means that for i=0,1,2i=0,1,2 ρ≤η3,η0+η1,η0+η2,η1+η2≤ρ,η0+η1+η2+η3≤3ρ,ηi≤si,(si−ηi)+2ρ≤d+ηi,(d−η3)+2ρ≤d+ηi.\begin{gathered} \rho\le\eta_3,\qquad \eta_0+\eta_1,\ \eta_0+\eta_2,\ \eta_1+\eta_2\le\rho,\qquad \eta_0+\eta_1+\eta_2+\eta_3\le3\rho,\\ \eta_i\le s_i,\qquad (s_i-\eta_i)+2\rho\le d+\eta_i,\qquad (d-\eta_3)+2\rho\le d+\eta_i. \end{gathered}(7)✦ The planted tail tuple of a common parameter uu and a base parameter η\eta for the rank r−ur-u in dimension D−cD-c is (c,u,D−c−η3,(η0,η1,η2),(s0−η0,s1−η1,s2−η2))(c,u,D-c-\eta_3,(\eta_0,\eta_1,\eta_2),(s_0-\eta_0,s_1-\eta_1,s_2-\eta_2)).✦

For natural numbers DD and rr and a profile (c,h)(c,h), a choice of tails ω\omega marks each face jj full or lower, and gives each full face jj a common parameter uju_j for cjc_j and a base parameter ηj\eta_j for the rank r−ujr-u_j, the dimension D−cjD-c_j and the sizes sj,0,sj,1,sj,2s_{j,0},s_{j,1},s_{j,2}.✦ The face jj then carries the tail tuple τj\tau_j: the planted tail tuple of (uj,ηj)(u_j,\eta_j) if jj is full, the lower tail tuple of cjc_j and (sj,0,sj,1,sj,2)(s_{j,0},s_{j,1},s_{j,2}) if not. The global tail gain Gain⁡D(ω)\operatorname{Gain}_D(\omega) is the sum over the faces jj of the full tail gain of τj\tau_j at full faces and of its lower tail gain at lower ones.✦

The singular tail threshold D*D_\ast is a fixed natural number with D*≥1642D_\ast\ge1642 such that, for every even D=2r≥D*D=2r\ge D_\ast, every profile (c,h)(c,h) geometric in dimension DD and every choice of tails ω\omega for it, Cost⁡D(c,h)>Gain⁡D(ω)orcj=0 for all j,he∈{0,2} for all eandGain⁡D(ω)=23∑eb(he).\begin{gathered} \operatorname{Cost}_D(c,h)>\operatorname{Gain}_D(\omega)\quad\text{or}\\ c_j=0\text{ for all }j,\quad h_e\in\{0,2\}\text{ for all }e\quad\text{and}\quad \operatorname{Gain}_D(\omega)=\frac23\sum_eb(h_e). \end{gathered}(8)✦

The Claim

Theorem 1 (Rank-layer limit).

  1. There is a positive integer D0D_0 with the following property. Let D≥D0D\ge D_0 be even, let B⊆CB\subseteq C be a set of corners and let ξ\xi be an integer function on the pairs of a point and a face with ξ(v,j)∈{1,−1}\xi(v,j)\in\{1,-1\} for every (v,j)∈B(v,j)\in B. Then, as q→∞q\to\infty through odd primes, ΛD(q;B,ξ)−MD(q;B,ξ)→0.\Lambda_D(q;B,\xi)-M_D(q;B,\xi)\to0.(9)✦

  2. The threshold can be taken to be the singular tail threshold D*≥1642D_\ast\ge1642: for every natural number rr with 2r≥D*2r\ge D_\ast and every BB and ξ\xi as in (1), as q→∞q\to\infty through odd primes, Λ2r(q;B,ξ)−M2r(q;B,ξ)→0.\Lambda_{2r}(q;B,\xi)-M_{2r}(q;B,\xi)\to0.(10)✦

  3. Let 𝔽\mathbb F be a finite field, DD and rr natural numbers, ξ\xi an integer and Z∈SD(𝔽)Z\in S_D(\mathbb F). Then Kr,ξ(−Z)=Kr,χ((−1)r)ξ(Z).K_{r,\xi}(-Z)=K_{r,\chi((-1)^r)\,\xi}(Z).(11)✦

In the paper, the result reads: there is D0D_0 such that for every fixed even D≥D0D\ge D_0, as q→∞q\to\infty through odd primes, for independent uniform Xv,Yj∈SDX_v,Y_j\in S_D, every B⊆CB\subseteq C and all signs ξvj∈{±1}\xi_{vj}\in\{\pm1\} on BB, the mean 𝔼∏(v,j)∈BKξvj(Xv−Yj)\mathbb E\prod_{(v,j)\in B}K_{\xi_{vj}}(X_v-Y_j) equals the value of the sign model up to o(1)o(1), uniformly in BB and the signs, and the same holds with Yj−XvY_j-X_v in place of Xv−YjX_v-Y_j, with the page’s KD/2,ξK_{D/2,\xi} for the paper’s KξK_\xi. The second form, with Yj−XvY_j-X_v, is answered in the box by the reflection rule (11).

What Is Checked

The rank-layer limit (9) and its form (10) are stated over the corners of HH, the symmetric matrices over 𝔽q\mathbb F_q with their ranks and signs, the rank kernels of rank D/2D/2, the layer and sign model correlations, limits as q→∞q\to\infty through odd primes and, in (10), the singular tail threshold D*D_\ast; the reflection rule (11) is stated over the rank kernels on the symmetric matrices over a finite field. Beyond the claim, the library proves the global tail dichotomy: a natural number at least 16421642 with the property (8) asked of D*D_\ast exists.✦