OctoInformBeta

OpenAI preprint, September 2026 · §7. The singular contribution · §7

§7 (setting): profiles, costs and slack: Intersection dimensions at faces and pairs, the cost as face baselines plus a separation excess on geometric profiles, and tail tuples

Profiles of thirteen Lagrangians, their cost and tail slack.

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 the incidence graph of 13 points and 22 faces: each face is a set of three points, and each point is joined to the faces containing it. Any two points of a face form a side of it, and each of the 33 point pairs that occur as sides lies on exactly two faces.

Take a vector space EE over a field KK with a bilinear form ω\omega, place at each point xx a Lagrangian subspace LxL_x of EE for ω\omega, that is, a subspace equal to its orthogonal for ω\omega, and record how the 13 subspaces meet: at a face jj, the dimension cjc_j of the intersection of its three subspaces; at a pair e={x,y}e=\{x,y\}, the dimension heh_e of Lx∩LyL_x\cap L_y. Numbers of this kind are weighed by a cost, built from them and from a three-bit label on each face, and the cost is compared with a sum of baselines, one for each face. At a single face, tail tuples refine the data further, and the baseline of the face is set against a gain. The library defines all of these and states one identity relating the cost to the baselines.

Faces, Labels and Profiles

The points are 0,…,120,\dots,12, and the faces F0,…,F21F_0,\dots,F_{21} are, in this order, the sets {0,1,3},{0,1,9},{0,2,3},{0,2,9},{1,3,10},{1,7,10},{1,7,12},{1,9,12},{2,3,4},{2,4,9},{3,4,11},{3,10,11},{4,8,9},{4,8,11},{5,6,7},{5,6,11},{5,7,10},{5,10,11},{6,7,12},{6,8,11},{6,8,12},{8,9,12}.\begin{aligned} &\{0,1,3\},\ \{0,1,9\},\ \{0,2,3\},\ \{0,2,9\},\ \{1,3,10\},\ \{1,7,10\},\ \{1,7,12\},\ \{1,9,12\},\\ &\{2,3,4\},\ \{2,4,9\},\ \{3,4,11\},\ \{3,10,11\},\ \{4,8,9\},\ \{4,8,11\},\ \{5,6,7\},\ \{5,6,11\},\\ &\{5,7,10\},\ \{5,10,11\},\ \{6,7,12\},\ \{6,8,11\},\ \{6,8,12\},\ \{8,9,12\}. \end{aligned}(1)✦ If v0<v1<v2v_0<v_1<v_2 are the points of FjF_j, the face jj has the side ej,0={v0,v1}e_{j,0}=\{v_0,v_1\} in slot 00, the side ej,1={v0,v2}e_{j,1}=\{v_0,v_2\} in slot 11 and the side ej,2={v1,v2}e_{j,2}=\{v_1,v_2\} in slot 22.✦

Each face carries a fixed three-bit label. 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, their Hamming distance, a number between 00 and 33; we write ae=σe/3a_e=\sigma_e/3.✦

For a bilinear form ω\omega on a vector space EE over a field KK, a Lagrangian subspace of EE is a subspace LL equal to its orthogonal L⟂ω={w∈E:ω(z,w)=0for allz∈L}L^{\perp_\omega}=\{w\in E:\ \omega(z,w)=0\ \text{for all}\ z\in L\}.✦

A profile (c,h)(c,h) is a pair of functions, cc assigning a natural number cjc_j to each face jj and hh a natural number heh_e to each pair ee. The sizes of a profile (c,h)(c,h) at a face jj are the three numbers sj,i=hej,i−cjs_{j,i}=h_{e_{j,i}}-c_j, one for each slot i∈{0,1,2}i\in\{0,1,2\}.✦

The Claim

Definition 1. Let LxL_x, one for each of the 13 points xx, be Lagrangian subspaces of EE for ω\omega. A profile (c,h)(c,h) is the point profile of these subspaces if, for every face jj and every pair e={x,y}e=\{x,y\}, cj=dim⁡K⋂z∈FjLz,he=dim⁡K(Lx∩Ly).c_j=\dim_K\bigcap_{z\in F_j}L_z,\qquad h_e=\dim_K\,(L_x\cap L_y).(2)✦ For a natural number DD, a profile (c,h)(c,h) is a geometric profile in dimension DD if, for every face jj and every slot 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.(3)✦ The triangular function bb on the real numbers is given by b(n)=n(n+1)2.b(n)=\frac{n(n+1)}{2}.(4)✦ The face cost of a face jj for a profile (c,h)(c,h) in dimension DD is the natural number Cj=(D+1)cj+∑i=02(sj,i+12),C_j=(D+1)\,c_j+\sum_{i=0}^{2}\binom{s_{j,i}+1}{2},(5)✦ and the profile cost of (c,h)(c,h) in dimension DD is the real number 𝒞D(c,h)=12∑jCj−16∑e(3−σe)b(he),\mathcal C_D(c,h)=\frac12\sum_{j}C_j-\frac16\sum_{e}(3-\sigma_e)\,b(h_e),(6)✦ the sums running over the 22 faces and the 33 pairs. The face baseline of real numbers DD, cc and s=(s0,s1,s2)s=(s_0,s_1,s_2) is β(D,c,s)=(D+1)c+∑ib(si)2−∑ib(c+si)6,\beta(D,c,s)=\frac{(D+1)\,c+\sum_i b(s_i)}{2}-\frac{\sum_i b(c+s_i)}{6},(7)✦ and the separation excess of a function hh on the pairs is ℰ(h)=12∑e(ae−13)b(he).\mathcal E(h)=\frac12\sum_{e}\Bigl(a_e-\frac13\Bigr)\,b(h_e).(8)✦

Theorem 1. Let DD be a natural number and (c,h)(c,h) a profile with cj≤hej,ic_j\le h_{e_{j,i}} for every face jj and every slot ii. Then 𝒞D(c,h)=∑jβ(D,cj,(sj,0,sj,1,sj,2))+ℰ(h).\mathcal C_D(c,h)=\sum_{j}\beta\bigl(D,c_j,(s_{j,0},s_{j,1},s_{j,2})\bigr)+\mathcal E(h).(9)✦

Definition 2. A tail tuple is a quintuple a=(c,u,v,p,t),c,u,v∈ℕ,p=(p0,p1,p2),t=(t0,t1,t2)∈ℕ3.a=(c,u,v,p,t),\qquad c,u,v\in\mathbb N,\qquad p=(p_0,p_1,p_2),\ t=(t_0,t_1,t_2)\in\mathbb N^3.(10)✦ The sizes of a tail tuple aa are si=pi+tis_i=p_i+t_i✦ and its heights hi=c+sih_i=c+s_i✦, for i∈{0,1,2}i\in\{0,1,2\}; its maximal height is H=max⁡ihiH=\max_i h_i✦, its offset the real number δ=2u−c\delta=2u-c✦ and its tail sum P=p0+p1+p2P=p_0+p_1+p_2✦. For a natural number DD and a flag f∈{0,1}f\in\{0,1\}, the tail tuple aa is feasible in dimension DD with flag ff if c≤D,s0+s1+s2≤D−c,and, when f=1,u≤c,ti≤pi+δ and v≤pi+δ(i=0,1,2),12(D−c−3δ)−P+v≥0.\begin{aligned} &c\le D,\qquad s_0+s_1+s_2\le D-c,\qquad\text{and, when } f=1,\\ &u\le c,\qquad t_i\le p_i+\delta\ \text{ and }\ v\le p_i+\delta\ \ (i=0,1,2),\qquad \tfrac12(D-c-3\delta)-P+v\ge 0. \end{aligned}(11)✦ The planted gain of real numbers cc, dd, δ\delta, vv and real triples pp, tt is Π(c,d,δ,v,p,t)=(d+c2+c+1)c+δ2−2(c+δ2)2+∑ib(pi+ti)−δ∑ipi−∑ipi2−v(d−3δ2−∑ipi+v)−∑ib(ti)+b(v),\begin{aligned} \Pi(c,d,\delta,v,p,t)={}&\Bigl(\frac{d+c}{2}+c+1\Bigr)\frac{c+\delta}{2} -2\Bigl(\frac{c+\delta}{2}\Bigr)^{2}+\sum_i b(p_i+t_i)-\delta\sum_i p_i-\sum_i p_i^{2}\\ &-v\Bigl(\frac{d-3\delta}{2}-\sum_i p_i+v\Bigr)-\sum_i b(t_i)+b(v), \end{aligned}(12)✦ the sums running over i∈{0,1,2}i\in\{0,1,2\}. The tail gain of aa in dimension DD with flag ff is gDf(a)={Π(c,D−c,δ,v,p,t)if f=1,b(⌊H/2⌋)if f=0,g^f_D(a)=\begin{cases}\Pi(c,\,D-c,\,\delta,\,v,\,p,\,t)&\text{if } f=1,\\ b\bigl(\lfloor H/2\rfloor\bigr)&\text{if } f=0,\end{cases}(13)✦ The tail slack of aa in dimension DD with flag ff is SDf(a)=β(D,c,s)−gDf(a).S^f_D(a)=\beta(D,c,s)-g^f_D(a).(14)✦

What Is Checked

The definitions and the theorem are stated over the configuration fixed above: the 22 faces (1) on the 13 points, the sides of each face in its three slots, the two faces of each pair and the three-bit labels that give the separation counts, and, for the point profile, Lagrangian subspaces for a bilinear form on a vector space over a field.