OctoInformBeta

OpenAI preprint, September 2026 · §5. Lagrangian counts and ordered triple orbits · §5

Lemma 5.2 [Chart comparison]: The rank kernel at A−YA-Y is at most (πD,r/pξ)J(ΓA,ΓY)(\pi_{D,r}/p_\xi)\,J(\Gamma_A,\Gamma_Y), and face averages and singular tails obey bounds of the same kind

Lagrangian kernels bound rank kernels up to constant factors.

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, with 13 points on one side and 22 faces, triples of points, on the other; each point is joined to the faces containing it. A corner is a face together with one of its three slots, and the slot holds one of the face’s points. The 33 point pairs are the pairs of points that share a face.

In the matrix model for HH, every point carries a symmetric matrix over a finite field and every face averages over a center matrix YY of its own. Some corners are marked active, and each active corner contributes a rank kernel at the difference between its point’s matrix and YY, a weight that depends only on whether this difference has a fixed rank rr and a given sign. Configurations in which the two points of some pair carry matrices with singular difference make up the singular tail.

The item bounds this model, corner by corner, face by face and on the singular tail, by a model on Lagrangian subspaces, in which a symmetric matrix AA is replaced by its graph ΓA\Gamma_A and the rank kernel by a weight that depends only on the dimension of an intersection of two Lagrangian subspaces.

Signs and Rank Kernels

Let 𝔽\mathbb{F} be a finite field with quadratic character χ\chi. For an integer D≥0D\ge0, the space SDS_D of symmetric D×DD\times D matrices over 𝔽\mathbb{F} is the subspace of the space of all D×DD\times D matrices over 𝔽\mathbb{F} consisting of the matrices MM with MT=MM^{T}=M.

Let BB be a symmetric bilinear form on a finite-dimensional space VV over 𝔽\mathbb{F}, with radical RR. The radical quotient form of BB is the form B‾(x+R,y+R)=B(x,y)\bar B(x+R,y+R)=B(x,y) induced by BB on V/RV/R, which is nondegenerate.✦ The sign sgn⁡B\operatorname{sgn}B of BB is χ(det⁡G)\chi(\det G), where GG is the Gram matrix of B‾\bar B in a basis of V/RV/R; at rank zero it is 11.✦ A matrix Z∈SDZ\in S_D has as its sign sgn⁡Z\operatorname{sgn}Z the sign of the form (x,y)↦xTZy(x,y)\mapsto x^{T}Zy on 𝔽D\mathbb{F}^D.✦

Fix r≥0r\ge0. For an integer ξ\xi, the signed rank layer SDr,ξS_D^{r,\xi} is the set of matrices in SDS_D of rank rr and sign ξ\xi✦, and the layer mass of sign ξ\xi is the proportion of SDS_D that this layer takes up, pξ=|SDr,ξ||SD|.p_\xi=\frac{|S_D^{r,\xi}|}{|S_D|}.(1)✦ The rank kernel KξK_\xi of rank rr and sign ξ\xi is a real function on SDS_D whose value at a matrix ZZ depends only on whether ZZ lies in the signed rank layer SDr,ξS_D^{r,\xi}.

Lagrangian Subspaces

Let EE be a finite-dimensional vector space over 𝔽\mathbb{F} with a bilinear form ω\omega. A Lagrangian subspace of EE is a subspace LL of EE equal to its orthogonal {x∈E:ω(y,x)=0 for all y∈L}\{x\in E:\omega(y,x)=0\text{ for all }y\in L\}; Λ(E)\Lambda(E) denotes the set of them.✦ For A∈Λ(E)A\in\Lambda(E), the stratum Λr(A)\Lambda^r(A) is the set of L∈Λ(E)L\in\Lambda(E) with dim⁡(L∩A)=r\dim(L\cap A)=r.✦ For a triple L=(L0,L1,L2)L=(L_0,L_1,L_2) in Λ(E)\Lambda(E) and a set T⊆{0,1,2}T\subseteq\{0,1,2\} of slots, the active center set Λr(L;T)\Lambda^r(L;T) is the set of Y∈Λ(E)Y\in\Lambda(E) with dim⁡(Li∩Y)=r\dim(L_i\cap Y)=r for every i∈Ti\in T.✦ The active face density of LL on TT, relative to AA, is ρTA(L)=|Λr(L;T)||Λ(E)|(|Λ(E)||Λr(A)|)|T|.\rho^A_T(L)=\frac{|\Lambda^r(L;T)|}{|\Lambda(E)|} \left(\frac{|\Lambda(E)|}{|\Lambda^r(A)|}\right)^{|T|}.(2)✦

The space ED=𝔽D×(𝔽D)*E_D=\mathbb{F}^D\times(\mathbb{F}^D)^* carries the canonical symplectic form ω((v,φ),(w,ψ))=ψ(v)−φ(w)\omega((v,\varphi),(w,\psi))=\psi(v)-\varphi(w)✦, and ΛD=Λ(ED)\Lambda_D=\Lambda(E_D) is the set of its Lagrangian subspaces✦. The vertical Lagrangian is the vertical subspace W={0}×(𝔽D)*W=\{0\}\times(\mathbb{F}^D)^*, the kernel of the projection of EDE_D onto 𝔽D\mathbb{F}^D.✦ The matrix Lagrangian of M∈SDM\in S_D is the graph ΓM={(v,w↦vTMw):v∈𝔽D}\Gamma_M=\{(v,\,w\mapsto v^{T}Mw):v\in\mathbb{F}^D\} of the bilinear form of MM.✦

The canonical stratum mass is the proportion of Lagrangian subspaces of EDE_D that meet WW in dimension rr, πD,r=|Λr(W)||ΛD|,\pi_{D,r}=\frac{|\Lambda^r(W)|}{|\Lambda_D|},(3)✦ and the canonical rank kernel on ΛD\Lambda_D is J(A,Y)={πD,r−1if dim⁡(A∩Y)=r,0otherwise.J(A,Y)=\begin{cases}\pi_{D,r}^{-1}&\text{if }\dim(A\cap Y)=r,\\ 0&\text{otherwise.}\end{cases}(4)✦ In dimension nn, the chart ratio is |Λn|/|Sn||\Lambda_n|/|S_n|✦ and the Lagrangian constant is cn=(n+1)2n.c_n=(n+1)\,2^{n}.(5)✦

Faces and Configurations

Throughout the rest of the page the dimension is D=2rD=2r, where r≥0r\ge0 is the rank fixed above.

The points of face jj, in increasing order, are v0(j)<v1(j)<v2(j)v_0(j)<v_1(j)<v_2(j), held in its slots 00, 11, 22; the corner (j,i)(j,i) is slot ii of face jj.✦ 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 set T⊆{0,1,2}T\subseteq\{0,1,2\} of active slots, integer signs ξ=(ξ0,ξ1,ξ2)\xi=(\xi_0,\xi_1,\xi_2) and matrices X0,X1,X2∈SDX_0,X_1,X_2\in S_D, the face local average is the mean over a center matrix YY, ΦT,ξ(X0,X1,X2)=𝔼Y∈SD∏k∈TKξk(Xk−Y).\Phi_{T,\xi}(X_0,X_1,X_2)=\mathbb{E}_{Y\in S_D}\prod_{k\in T}K_{\xi_k}(X_k-Y).(6)✦

Let T=(Tj)T=(T_j) choose a set Tj⊆{0,1,2}T_j\subseteq\{0,1,2\} of active slots at each face, let ξ=(ξj,i)\xi=(\xi_{j,i}) put an integer sign at each corner, and write ξj=(ξj,0,ξj,1,ξj,2)\xi_j=(\xi_{j,0},\xi_{j,1},\xi_{j,2}). For an assignment XX of a matrix Xi∈SDX_i\in S_D to each point ii, the configuration product is the product over the 22 faces PT,ξ(X)=∏jΦTj,ξj(Xv0(j),Xv1(j),Xv2(j)).P_{T,\xi}(X)=\prod_{j}\Phi_{T_j,\xi_j}\bigl(X_{v_0(j)},X_{v_1(j)},X_{v_2(j)}\bigr).(7)✦ For an assignment LL of a Lagrangian subspace Li∈ΛDL_i\in\Lambda_D to each point ii, the canonical configuration is ΨT(L)=∏jρTjW(Lv0(j),Lv1(j),Lv2(j)),\Psi_T(L)=\prod_{j}\rho^W_{T_j}\bigl(L_{v_0(j)},L_{v_1(j)},L_{v_2(j)}\bigr),(8)✦ and ΓX\Gamma_X denotes the assignment i↦ΓXii\mapsto\Gamma_{X_i}.

An assignment XX of matrices in SDS_D to the 13 points is fully transverse when det⁡(Xu−Xw)≠0\det(X_u-X_w)\neq0 for each of the 33 point pairs {u,w}\{u,w\}, u<wu<w.✦ The matrix singular tail is the mean of the configuration product over all such assignments, the fully transverse ones counted as zero, ΘT,ξ=𝔼X[𝟏[X not fully transverse]PT,ξ(X)].\Theta_{T,\xi}=\mathbb{E}_{X}\bigl[\mathbf{1}[X\text{ not fully transverse}]\,P_{T,\xi}(X)\bigr].(9)✦ The Lagrangian singular tail ΘTΛ\Theta^{\Lambda}_T is the mean, over all assignments of a Lagrangian subspace of EDE_D to the 13 points, of the product over the 22 faces jj of the active face densities ρTjW\rho^W_{T_j} of the three Lagrangian subspaces at the face; two Lagrangian subspaces are transverse when they meet only in {0}\{0\}; only the assignments in which the Lagrangian subspaces at the two points of some pair are not transverse contribute.

The Claim

Theorem 1. Let 𝔽\mathbb{F} be a finite field, r≥0r\ge0 and D=2rD=2r.

  1. Let ξ\xi be an integer, A,Y∈SDA,Y\in S_D, and C≥0C\ge0 with πD,r/pξ≤C\pi_{D,r}/p_\xi\le C. Then Kξ(A−Y)≤CJ(ΓA,ΓY).K_\xi(A-Y)\le C\,J(\Gamma_A,\Gamma_Y).(10)✦

  2. Suppose r≥1r\ge1 and ξ∈{1,−1}\xi\in\{1,-1\}. For every sufficiently large odd prime qq, the stratum mass and the layer mass over the field 𝔽q\mathbb{F}_q with qq elements satisfy π2r,rpξ≤4⋅2r.\frac{\pi_{2r,r}}{p_\xi}\le4\cdot2^{r}.(11)✦

  3. Let T⊆{0,1,2}T\subseteq\{0,1,2\}, let ξ0,ξ1,ξ2\xi_0,\xi_1,\xi_2 be integers and X0,X1,X2∈SDX_0,X_1,X_2\in S_D, and let C≥1C\ge1 with πD,r/pξi≤C\pi_{D,r}/p_{\xi_i}\le C for every i∈Ti\in T. Then ΦT,ξ(X0,X1,X2)≤C3cDρTW(ΓX0,ΓX1,ΓX2).\Phi_{T,\xi}(X_0,X_1,X_2)\le C^{3}c_D\, \rho^W_T\bigl(\Gamma_{X_0},\Gamma_{X_1},\Gamma_{X_2}\bigr).(12)✦

  4. Let TT choose a set of active slots at each face, ξ\xi an integer sign at each corner and XX a matrix in SDS_D at each point, and let C≥1C\ge1 with πD,r/pξj,i≤C\pi_{D,r}/p_{\xi_{j,i}}\le C at every corner (j,i)(j,i). Then PT,ξ(X)≤(C3cD)22ΨT(ΓX).P_{T,\xi}(X)\le\bigl(C^{3}c_D\bigr)^{22}\,\Psi_T(\Gamma_X).(13)✦

  5. Under the hypotheses on TT, ξ\xi and CC of the preceding part, ΘT,ξ≤(C3cD)22cD13ΘTΛ.\Theta_{T,\xi}\le\bigl(C^{3}c_D\bigr)^{22}c_D^{13}\,\Theta^{\Lambda}_T.(14)✦

  6. For every n≥0n\ge0, the chart ratio is at most the Lagrangian constant: |Λn||Sn|≤cn.\frac{|\Lambda_n|}{|S_n|}\le c_n.(15)✦

In the paper, the result reads: for a set BB of active corners, a corner joining point ii to face jj with sign ξij\xi_{ij}, and for every nonnegative function FF of the 35 Lagrangian subspaces at the points and faces, 𝔼[F((ΓXi),(ΓYj))∏BKξij(Xi−Yj)]≤βD−35∏BπD,rpξij𝔼[F((Li),(Mj))∏BJ(Li,Mj)],\mathbb{E}\Bigl[F\bigl((\Gamma_{X_i}),(\Gamma_{Y_j})\bigr)\prod_{B}K_{\xi_{ij}}(X_i-Y_j)\Bigr] \le\beta_D^{-35}\prod_{B}\frac{\pi_{D,r}}{p_{\xi_{ij}}}\, \mathbb{E}\Bigl[F\bigl((L_i),(M_j)\bigr)\prod_{B}J(L_i,M_j)\Bigr],(16) where XiX_i and YjY_j are the matrices at point ii and face jj, LiL_i and MjM_j the Lagrangian subspaces in their place, and the multiplier βD−35\beta_D^{-35} in front is bounded in terms of DD for large qq.

What Is Checked

The claim is stated over the symmetric matrices of size 2r2r over a finite field, their signs, signed rank layers, layer masses (1) and rank kernels; the Lagrangian subspaces of E2rE_{2r} for the canonical symplectic form, the vertical Lagrangian WW, the matrix Lagrangians ΓA\Gamma_A, the stratum mass (3), the canonical rank kernel (4) and the active face densities (2); the face local averages (6), configuration products (7), canonical configurations (8), the matrix singular tail (9) and its Lagrangian counterpart; the Lagrangian constant (5); and, in (11), the fields 𝔽q\mathbb{F}_q as qq runs through the odd primes to infinity. Beyond the claim, the library establishes that the vertical subspace WW and the graph ΓM\Gamma_M of every symmetric matrix MM are Lagrangian subspaces of EDE_D.