OctoInformBeta

OpenAI preprint, September 2026 · §6. Local bounds from the planted law · §6

Proposition 6.3 [Three active incidences]: For q≥4cDq\ge 4c_D, first through the centers of the triple’s intersection profile, then by a finite sum of explicit qq-power terms

Bounding the planted face factor of three Lagrangians.

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 has a vertex for each of the 1313 points and each of the 2222 faces, a face being a triple of points, and joins every point to the faces that contain it; each face vertex has degree three. In the homomorphism density t(H,W)t(H,W) of HH in a kernel WW, once values are fixed at the points, a face contributes its face factor, the average, over a value YY at the face, of the product of the three kernel values on its corners.

In the construction behind the counterexample to Sidorenko’s conjecture, the points carry Lagrangian subspaces of a symplectic space over a finite field and each face carries a Lagrangian subspace YY. For Lagrangians L,M,NL,M,N at the three points of a face, the question is how large this face factor can be, as a power of the field size qq, once the dimensions of L∩M∩NL\cap M\cap N, L∩ML\cap M, L∩NL\cap N and M∩NM\cap N are fixed.

Lagrangians and Their Centers

Let KK be a finite field with qq elements and EE a vector space of dimension 2D2D over KK with a symplectic form ω\omega. A Lagrangian subspace of EE is a subspace LL of EE equal to its orthogonal L⟂={y∈E:ω(x,y)=0for allx∈L}L^{\perp}=\{y\in E:\ \omega(x,y)=0\ \text{for all}\ x\in L\}, and ℒ\mathcal L denotes the set of Lagrangian subspaces of EE.✦

The center stratum 𝒮r(Y)\mathcal S_r(Y) of Y∈ℒY\in\mathcal L at rank rr is the set of Lagrangian subspaces XX of EE with dim⁡(X∩Y)=r\dim(X\cap Y)=r.✦ The triple center set 𝒵r(L,M,N)\mathcal Z_r(L,M,N) of subspaces L,M,NL,M,N of EE is the set of Lagrangian subspaces YY of EE that meet each of LL, MM and NN in a subspace of dimension rr, the common centers of rank rr of the three.✦

Fix a reference Lagrangian subspace AA, and for X,Y∈ℒX,Y\in\mathcal L let J(X,Y)J(X,Y) be |ℒ|/|𝒮r(A)||\mathcal L|/|\mathcal S_r(A)| when dim⁡(X∩Y)=r\dim(X\cap Y)=r and 00 otherwise. The triple face density Tr(A;L,M,N)T_r(A;L,M,N) of Lagrangian subspaces L,M,NL,M,N, their planted face factor, is the real number Tr(A;L,M,N)=|𝒵r(L,M,N)||ℒ|2|𝒮r(A)|3=1|ℒ|∑Y∈ℒJ(L,Y)J(M,Y)J(N,Y).T_r(A;L,M,N)=\frac{|\mathcal Z_r(L,M,N)|\,|\mathcal L|^{2}}{|\mathcal S_r(A)|^{3}} =\frac{1}{|\mathcal L|}\sum_{Y\in\mathcal L}J(L,Y)\,J(M,Y)\,J(N,Y).(1)✦

The triple dimension profile 𝒫(c;h1,h2,h3)\mathcal P(c;h_1,h_2,h_3) of natural numbers c,h1,h2,h3c,h_1,h_2,h_3 is the set of ordered triples (L,M,N)(L,M,N) of Lagrangian subspaces of EE with dim⁡(L∩M∩N)=c\dim(L\cap M\cap N)=c, dim⁡(L∩M)=h1\dim(L\cap M)=h_1, dim⁡(L∩N)=h2\dim(L\cap N)=h_2 and dim⁡(M∩N)=h3\dim(M\cap N)=h_3.✦ The profile center set 𝒞r(c;h1,h2,h3)\mathcal C_r(c;h_1,h_2,h_3) is the set of pairs ((L,M,N),Y)\bigl((L,M,N),Y\bigr) formed by a triple in 𝒫(c;h1,h2,h3)\mathcal P(c;h_1,h_2,h_3) and one of its common centers Y∈𝒵r(L,M,N)Y\in\mathcal Z_r(L,M,N) of rank rr.✦

Exponents and Constants

The triangular function bb sends a real number xx to b(x)=x(x+1)/2b(x)=x(x+1)/2.✦ The triple defect of subspaces L,M,NL,M,N of EE, with c=dim⁡(L∩M∩N)c=\dim(L\cap M\cap N) and h1,h2,h3h_1,h_2,h_3 the dimensions of L∩ML\cap M, L∩NL\cap N and M∩NM\cap N, is the natural number ΔD(L,M,N)=(D+1)c+b(h1−c)+b(h2−c)+b(h3−c).\Delta_D(L,M,N)=(D+1)\,c+b(h_1-c)+b(h_2-c)+b(h_3-c).(2)✦ The isotropic exponent of natural numbers DD and cc is the natural number ι(D,c)=b(D)+c(D−c)−b(D−c),\iota(D,c)=b(D)+c(D-c)-b(D-c),(3)✦ which equals 2Dc−(3c2−c)/22Dc-(3c^2-c)/2 when c≤Dc\le D.

The Lagrangian constant of a natural number nn is cn=(n+1)2nc_n=(n+1)\,2^n.✦ The pair orbit constant of DD is oD=8⋅23DcD3=8(D+1)326Do_D=8\cdot 2^{3D}c_D^{3}=8(D+1)^3\,2^{6D}✦, and the orbit mass constant is mD=cDoD=8(D+1)427Dm_D=c_D\,o_D=8(D+1)^4\,2^{7D}✦.

Parameters and Planted Terms

The subspace triple profile with parameters k,p1,p2,p3k,p_1,p_2,p_3 of a finite-dimensional vector space VV over a field is the set of triples (U1,U2,U3)(U_1,U_2,U_3) of kk-dimensional subspaces of VV with U1∩U2∩U3=0U_1\cap U_2\cap U_3=0, dim⁡(U1∩U2)=p1\dim(U_1\cap U_2)=p_1, dim⁡(U1∩U3)=p2\dim(U_1\cap U_3)=p_2 and dim⁡(U2∩U3)=p3\dim(U_2\cap U_3)=p_3.✦ The span profile with span dimension nn of the subspace triple profile with parameters k,p1,p2,p3k,p_1,p_2,p_3 of VV is the set of those triples (U1,U2,U3)(U_1,U_2,U_3) with dim⁡(U1+U2+U3)=n\dim(U_1+U_2+U_3)=n.✦

The common parameter set 𝒰D,r,c\mathcal U_{D,r,c} of natural numbers D,r,cD,r,c is the set of integers uu with 0≤u≤c0\le u\le c, u≤ru\le r and r−u≤D−cr-u\le D-c.✦

Let d,kd,k be natural numbers and s=(s1,s2,s3)s=(s_1,s_2,s_3) a triple of natural numbers. A base parameter for d,k,sd,k,s is a quadruple θ=(p1,p2,p3,n)\theta=(p_1,p_2,p_3,n) of integers between 00 and dd such that, for every ii and every j≠ij\ne i in {1,2,3}\{1,2,3\}, k≤n,pi+pj≤k,n+p1+p2+p3≤3k,pi≤si,si+2k≤d+2pi,2k≤n+pi;k\le n,\quad p_i+p_j\le k,\quad n+p_1+p_2+p_3\le 3k,\quad p_i\le s_i,\quad s_i+2k\le d+2p_i,\quad 2k\le n+p_i;(4)✦ the base parameters form the base parameter set ℬd,k(s)\mathcal B_{d,k}(s).✦

The reduced planted constant of a quadruple θ=(p1,p2,p3,n)\theta=(p_1,p_2,p_3,n) of integers between 00 and dd is βk,s(θ)=2n+p1+p2+p3+3k+∑i(si−pi)\beta_{k,s}(\theta)=2^{\,n+p_1+p_2+p_3+3k+\sum_i(s_i-p_i)}✦, and its reduced planted exponent is ρd,k,s(θ)=−(d−2k)∑ipi−∑ipi2−(d−n)(3k−n−∑ipi)+b(d−n)−∑ib(si−pi).\rho_{d,k,s}(\theta)=-(d-2k)\sum_ip_i-\sum_ip_i^2-(d-n)\Bigl(3k-n-\sum_ip_i\Bigr)+b(d-n) -\sum_ib(s_i-p_i).(5)✦ For natural numbers d,kd,k and a triple ss of natural numbers, the reduced planted sum Rd,k(s;q)R_{d,k}(s;q) at a real number qq is Rd,k(s;q)=∑θ∈ℬd,k(s)βk,s(θ)qρd,k,s(θ).R_{d,k}(s;q)=\sum_{\theta\in\mathcal B_{d,k}(s)}\beta_{k,s}(\theta)\,q^{\rho_{d,k,s}(\theta)}.(6)

The common planted term of natural numbers D,c,r,h1,h2,h3D,c,r,h_1,h_2,h_3 and u∈𝒰D,r,cu\in\mathcal U_{D,r,c}, at a real number qq, is Γu(q)=2ucD−u23(D−c)qu(c−u)+b(D−u)+3b(D−c)−3b(r−u)RD−c,r−u(h1−c,h2−c,h3−c;q).\Gamma_u(q)=2^u\,c_{D-u}\,2^{3(D-c)}\,q^{\,u(c-u)+b(D-u)+3b(D-c)-3b(r-u)}\, R_{D-c,\,r-u}(h_1-c,h_2-c,h_3-c;\,q).(7)✦ For a triple ss of natural numbers and θ∈ℬD−c,r−u(s)\theta\in\mathcal B_{D-c,\,r-u}(s), the full planted exponent is f(u,θ)=(D+1)c+∑ib(si)+3b(r)−4b(D)+ι(D,c)+u(c−u)+b(D−u)+3b(D−c)−3b(r−u)+ρD−c,r−u,s(θ).f(u,\theta)=(D+1)\,c+\sum_ib(s_i)+3b(r)-4b(D)+\iota(D,c)+u(c-u)+b(D-u)+3b(D-c)-3b(r-u) +\rho_{D-c,\,r-u,\,s}(\theta).(8)✦

Tail Tuples

A tail tuple a=(c,u,v,p,t)a=(c,u,v,p,t) consists of natural numbers c,u,vc,u,v and two triples p=(p1,p2,p3)p=(p_1,p_2,p_3) and t=(t1,t2,t3)t=(t_1,t_2,t_3) of natural numbers.✦ The tail tuple aa has tail sizes si=pi+tis_i=p_i+t_i, tail heights hi=c+sih_i=c+s_i with largest value HH, tail offset δ=2u−c\delta=2u-c and tail sum P=p1+p2+p3P=p_1+p_2+p_3.

The planted gain of real numbers c,d,δ,vc,d,\delta,v and real triples p,tp,t is G(c,d,δ,v,p,t)=gC+∑ib(pi+ti)−δP−∑ipi2−vw−∑ib(ti)+b(v),G(c,d,\delta,v,p,t)=g_C+\sum_ib(p_i+t_i)-\delta P-\sum_ip_i^2-vw-\sum_ib(t_i)+b(v),(9)✦ where P=∑ipiP=\sum_ip_i, w=d−3δ2−P+vw=\tfrac{d-3\delta}{2}-P+v and gC=(d+c2+c+1)c+δ2−2(c+δ2)2g_C=\bigl(\tfrac{d+c}{2}+c+1\bigr)\tfrac{c+\delta}{2}-2\bigl(\tfrac{c+\delta}{2}\bigr)^2. The full tail gain gainD(a)\mathrm{gain}_D(a) of a tail tuple aa in dimension DD is G(c,D−c,δ,v,p,t)G(c,D-c,\delta,v,p,t).✦ The short tail gain of aa is b(⌊H/2⌋)b(\lfloor H/2\rfloor).

A tail tuple aa is feasible in dimension DD when c≤Dc\le D and s1+s2+s3≤D−cs_1+s_2+s_3\le D-c, and fully feasible when moreover, with ww as in (9) at d=D−cd=D-c, u≤c,ti≤pi+δandv≤pi+δ(i=1,2,3),w≥0;u\le c,\qquad t_i\le p_i+\delta\ \text{and}\ v\le p_i+\delta\ \ (i=1,2,3),\qquad w\ge 0;(10)✦ ℱD\mathcal F_D denotes the set of tail tuples fully feasible in dimension DD.

The planted tail tuple of natural numbers D,r,cD,r,c, a triple ss of natural numbers, u∈𝒰D,r,cu\in\mathcal U_{D,r,c} and θ=(p1,p2,p3,n)∈ℬD−c,r−u(s)\theta=(p_1,p_2,p_3,n)\in\mathcal B_{D-c,\,r-u}(s) is a(u,θ)=(c,u,D−c−n,(p1,p2,p3),(s1−p1,s2−p2,s3−p3)).a(u,\theta)=\bigl(c,\ u,\ D-c-n,\ (p_1,p_2,p_3),\ (s_1-p_1,\,s_2-p_2,\,s_3-p_3)\bigr).(11)✦

The Claim

Theorem 1. Suppose KK has odd characteristic and q≥4cDq\ge 4c_D, let r≤Dr\le D be a natural number and A,L,M,NA,L,M,N Lagrangian subspaces of EE, and put c=dim⁡(L∩M∩N)c=\dim(L\cap M\cap N), h1=dim⁡(L∩M)h_1=\dim(L\cap M), h2=dim⁡(L∩N)h_2=\dim(L\cap N) and h3=dim⁡(M∩N)h_3=\dim(M\cap N). Then Tr(A;L,M,N)≤mDcD2qΔD(L,M,N)+3b(r)−4b(D)|𝒞r(c;h1,h2,h3)|,T_r(A;L,M,N)\le m_D\,c_D^{2}\,q^{\Delta_D(L,M,N)+3b(r)-4b(D)}\, \bigl|\mathcal C_r(c;h_1,h_2,h_3)\bigr|,(12)✦ and, for a constant CC that depends on the dimensions alone, Tr(A;L,M,N)≤CqΔD(L,M,N)+3b(r)−4b(D)+ι(D,c)∑u∈𝒰D,r,cΓu(q).T_r(A;L,M,N)\le C\,q^{\Delta_D(L,M,N)+3b(r)-4b(D)+\iota(D,c)} \sum_{u\in\mathcal U_{D,r,c}}\Gamma_u(q).(13)✦ Every triple in the span profile with parameters k,p1,p2,p3k,p_1,p_2,p_3 and span dimension nn of a finite-dimensional vector space VV satisfies n≤dim⁡V,k≤n,pi+pj≤k(i≠j),n+p1+p2+p3≤3k.n\le\dim V,\qquad k\le n,\qquad p_i+p_j\le k\ \ (i\ne j),\qquad n+p_1+p_2+p_3\le 3k.(14)✦ Finally, let rr and cc be natural numbers with D=2rD=2r and c≤Dc\le D, let ss be a triple of natural numbers, and let u∈𝒰D,r,cu\in\mathcal U_{D,r,c} and θ∈ℬD−c,r−u(s)\theta\in\mathcal B_{D-c,\,r-u}(s). If s1+s2+s3≤D−cs_1+s_2+s_3\le D-c, the planted tail tuple is fully feasible, a(u,θ)∈ℱD,a(u,\theta)\in\mathcal F_D,(15)✦ and, with or without that condition, its full tail gain is the full planted exponent: gainD(a(u,θ))=f(u,θ).\mathrm{gain}_D\bigl(a(u,\theta)\bigr)=f(u,\theta).(16)✦

In the paper, the result reads: for a local profile (h1,h2,h3,c)(h_1,h_2,h_3,c), the face factor with three active incidences is O(qg)O(q^{g}) uniformly over all triples in the profile, where gg is the maximum, over a finite set of parameter tuples, of gC+∑ib(si)−δP−∑ipi2−vw−∑ib(ti)+b(v)g_C+\sum_ib(s_i)-\delta P-\sum_ip_i^2-vw-\sum_ib(t_i)+b(v), and g=0g=0 when that set is empty, and every tuple satisfies the necessary constraints 0≤u≤min⁡(c,r)0\le u\le\min(c,r), d−δ=2r′d-\delta=2r', |δ|≤c|\delta|\le c, P+T≤dP+T\le d, 0≤ti,v≤pi+δ0\le t_i,\,v\le p_i+\delta and w≥0w\ge 0, with d=D−cd=D-c, r′=r−ur'=r-u and T=t1+t2+t3T=t_1+t_2+t_3.

In the library, the right side of (13) expands into one term for each u∈𝒰D,r,cu\in\mathcal U_{D,r,c} and θ∈ℬD−c,r−u(s)\theta\in\mathcal B_{D-c,\,r-u}(s) with si=hi−cs_i=h_i-c, the power of qq in that term is f(u,θ)f(u,\theta), and for D=2rD=2r this exponent is, by (16), the tail gain of the planted tail tuple, which (15) makes fully feasible when s1+s2+s3≤D−cs_1+s_2+s_3\le D-c.

What Is Checked

The claim is stated over the triple face density of Lagrangian subspaces of a symplectic space over a finite field, the profile center set, the triple defect and the isotropic exponent, the constants cDc_D and mDm_D, the common planted terms with their reduced planted sums over common and base parameters, the span profile of three subspaces, and the planted tail tuples with their full feasibility, tail gain and full planted exponent.