OctoInformBeta

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

Lemma 7.4 [Profiles not separated by the first cost]: For every even dimension above a threshold D0≥1642D_0\ge 1642, a total gain reaching the profile cost forces all cj=0c_j=0 and all he∈{0,2}h_e\in\{0,2\}

In large even dimension, a gain reaching the cost forces rigidity.

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 is the incidence graph of 13 points and 22 faces, each face a triple of points: it joins each point to the faces containing it. A face jj has three point pairs as its sides, held in three slots i∈{0,1,2}i\in\{0,1,2\}, and ej,ie_{j,i} denotes the side in slot ii.✦ There are 33 point pairs, each a side of exactly two faces.✦ Each face carries a fixed label in {0,1}3\{0,1\}^3.✦ The separation count of a point pair ee is the number of the three coordinates at which the labels of the two faces through ee differ, and aea_e is one third of it, so that ae∈{0,13,23,1}a_e\in\{0,\frac13,\frac23,1\}.✦

Each face jj has a number cjc_j and each point pair ee has a number heh_e, in an even dimension D=2rD=2r; each face also carries a bit saying whether it is active and a gain, both defined below. The numbers cjc_j and heh_e form a profile (c,h)(c,h), which has a cost 𝒞D(c,h)\mathcal C_D(c,h); the labels enter the cost through the numbers aea_e. The question is which profiles carry a total gain, the global tail gain defined below, at least as large as their cost once DD is large.

Profiles and Their Cost

Let DD be a natural number. A geometric profile in dimension DD 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 such that, for every face jj, cj≤Dc_j\le D, cj≤hej,ic_j\le h_{e_{j,i}} for each slot ii, and the profile sizes sj,i=hej,i−cjs_{j,i}=h_{e_{j,i}}-c_j satisfy sj,0+sj,1+sj,2≤D−cj.s_{j,0}+s_{j,1}+s_{j,2}\le D-c_j.(1)✦

The face natural cost of a face jj, for a profile (c,h)(c,h) in dimension DD, is the natural number 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},(2)✦ and the profile cost of (c,h)(c,h) is the real number 𝒞D(c,h)=12∑jNj−12∑e(1−ae)b(he),\mathcal C_D(c,h)=\frac12\sum_j N_j-\frac12\sum_e(1-a_e)\,b(h_e),(3)✦ where jj runs over the faces, ee over the point pairs, and b(n)=n(n+1)/2b(n)=n(n+1)/2 is the triangular function.✦

Tail Tuples and Their Gains

An activity pattern ε\varepsilon assigns to each face jj a bit εj∈{0,1}\varepsilon_j\in\{0,1\}, the face being active when εj=1\varepsilon_j=1.

A tail tuple τ=(c,u,v,p,t)\tau=(c,u,v,p,t) consists of three natural numbers cc, uu, vv and two 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.✦ Its tail sizes are si=pi+tis_i=p_i+t_i✦ and its tail heights hi=c+sih_i=c+s_i✦, with maximal tail height H=max⁡ihiH=\max_i h_i✦; its tail offset is the real number δ=2u−c\delta=2u-c✦, and its tail sum is P=p0+p1+p2P=p_0+p_1+p_2.✦

The planted gain of real numbers cc, dd, δ\delta, vv and two real triples pp, tt is γ(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{split} \gamma(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{split}(4)✦ with sums over i∈{0,1,2}i\in\{0,1,2\}. The tail gain of a tail tuple τ\tau in dimension DD with a bit ε\varepsilon is gD(τ,ε)={γ(c,D−c,δ,v,p,t)if ε=1,b(⌊H/2⌋)if ε=0.g_D(\tau,\varepsilon)= \begin{cases} \gamma(c,D-c,\delta,v,p,t) & \text{if }\varepsilon=1,\\ b(\lfloor H/2\rfloor) & \text{if }\varepsilon=0. \end{cases}(5)✦

The face baseline of real numbers DD and cc and real sizes s=(s0,s1,s2)s=(s_0,s_1,s_2) is β(D,c,s)=12((D+1)c+∑ib(si))−16∑ib(c+si),\beta(D,c,s)=\frac12\Bigl((D+1)\,c+\sum_i b(s_i)\Bigr)-\frac16\sum_i b(c+s_i),(6)✦ and the tail slack of τ\tau in dimension DD with the bit ε\varepsilon is σD(τ,ε)=β(D,c,s)−gD(τ,ε)\sigma_D(\tau,\varepsilon)=\beta(D,c,s)-g_D(\tau,\varepsilon), with cc and the sizes ss those of τ\tau.✦ The separation excess of an assignment hh of a natural number heh_e to each point pair ee is E(h)=16∑e(3ae−1)b(he).E(h)=\frac16\sum_e(3a_e-1)\,b(h_e).(7)✦

A tail tuple τ\tau is tail feasible in dimension DD for a bit ε\varepsilon when c≤Dc\le D and s0+s1+s2≤D−cs_0+s_1+s_2\le D-c and, if ε=1\varepsilon=1, also u≤cu\le c, ti≤pi+δt_i\le p_i+\delta and v≤pi+δv\le p_i+\delta for every ii, and 12(D−c−3δ)−P+v≥0\frac12(D-c-3\delta)-P+v\ge0.✦ A family (τj)j(\tau_j)_j of tail tuples, one for each face, is globally tail feasible in dimension DD for an activity pattern ε\varepsilon and an assignment hh on the point pairs when each τj\tau_j is tail feasible for εj\varepsilon_j and the tail heights of τj\tau_j are the values of hh on the sides of jj, the height in slot ii being hej,ih_{e_{j,i}}.✦

Tail Choices

Let DD, rr and cc be natural numbers and s=(s0,s1,s2)s=(s_0,s_1,s_2) a triple of natural numbers. The common parameter set U(D,r,c)U(D,r,c) is the set of integers u∈{0,…,c}u\in\{0,\dots,c\} with u≤ru\le r and r−u≤D−cr-u\le D-c.✦ For a natural number dd, a quadruple θ=(θ0,θ1,θ2,θ3)∈{0,…,d}4\theta=(\theta_0,\theta_1,\theta_2,\theta_3)\in\{0,\dots,d\}^4 is a choice of feasible base parameters for rr and ss when r≤θ3r\le\theta_3, θi+θk≤r\theta_i+\theta_k\le r for distinct i,k∈{0,1,2}i,k\in\{0,1,2\}, θ0+θ1+θ2+θ3≤3r\theta_0+\theta_1+\theta_2+\theta_3\le 3r, and, for each i∈{0,1,2}i\in\{0,1,2\}, θi≤si\theta_i\le s_i, si+2r≤d+2θis_i+2r\le d+2\theta_i and 2r≤θ3+θi2r\le\theta_3+\theta_i✦; these quadruples form the base parameter set Θ(d,r,s)\Theta(d,r,s).✦

A full tail choice for DD, rr, cc, ss is a pair (u,θ)(u,\theta) with u∈U(D,r,c)u\in U(D,r,c) and θ∈Θ(D−c,r−u,s)\theta\in\Theta(D-c,r-u,s).✦ The planted tail tuple of a full tail choice (u,θ)(u,\theta) for DD, rr, cc, ss is the tail tuple (c,u,D−c−θ3,(θ0,θ1,θ2),(s0−θ0,s1−θ1,s2−θ2))(c,\,u,\,D-c-\theta_3,\,(\theta_0,\theta_1,\theta_2),\,(s_0-\theta_0,s_1-\theta_1,s_2-\theta_2)), and the lower tail tuple of cc and ss is the tail tuple (c,0,0,s,(0,0,0))(c,0,0,s,(0,0,0)). A face tail choice for a bit ε\varepsilon is a full tail choice when ε=1\varepsilon=1 and the single trivial choice when ε=0\varepsilon=0✦; its face tail tuple is the planted tail tuple in the first case and the lower tail tuple in the second.✦

Given a profile (c,h)(c,h) and an activity pattern ε\varepsilon, a global tail choice xx for DD and rr assigns to each face jj a face tail choice xjx_j for DD, rr, cjc_j, the profile sizes sj=(sj,0,sj,1,sj,2)s_j=(s_{j,0},s_{j,1},s_{j,2}) and the bit εj\varepsilon_j.✦ The global tail tuple τj(x)\tau_j(x) of xx at jj is the face tail tuple of xjx_j✦, and the global tail gain of xx is G(x)=∑jgD(τj(x),εj).G(x)=\sum_j g_D\bigl(\tau_j(x),\varepsilon_j\bigr).(8)✦

The Claim

Theorem 1. (a) There is an integer D0D_0 such that the following holds for every natural number D≥D0D\ge D_0. Let ε\varepsilon be an activity pattern, hh an assignment of a natural number to each point pair and (τj)j(\tau_j)_j, with τj=(cj,uj,vj,pj,tj)\tau_j=(c_j,u_j,v_j,p_j,t_j), a family of tail tuples that is globally tail feasible in dimension DD for ε\varepsilon and hh and satisfies ∑jσD(τj,εj)+E(h)≤0\sum_j\sigma_D(\tau_j,\varepsilon_j)+E(h)\le0. Then cj=0c_j=0 and σD(τj,εj)=0\sigma_D(\tau_j,\varepsilon_j)=0 for every face jj, he∈{0,2}h_e\in\{0,2\} for every point pair ee, E(h)=0E(h)=0, and ∑jgD(τj,εj)=23∑eb(he).\sum_j g_D(\tau_j,\varepsilon_j)=\frac23\sum_e b(h_e).(9)✦

(b) There is an integer D0≥1642D_0\ge1642 such that the following holds. Let rr be a natural number with D=2r≥D0D=2r\ge D_0, let (c,h)(c,h) be a geometric profile in dimension DD, ε\varepsilon an activity pattern and xx a global tail choice for DD and rr on (c,h)(c,h) and ε\varepsilon. Then 𝒞D(c,h)−G(x)>0or(cj=0for every j,he∈{0,2}for every e,G(x)=23∑eb(he)).\mathcal C_D(c,h)-G(x)>0 \quad\text{or}\quad \Bigl(c_j=0\ \text{for every }j,\ \ h_e\in\{0,2\}\ \text{for every }e,\ \ G(x)=\tfrac23\textstyle\sum_e b(h_e)\Bigr).(10)✦

(c) Let D*D_*, the singular tail threshold, be a fixed choice of an integer D0D_0 as in (b).✦ For every natural number rr with D=2r≥D*D=2r\ge D_* and every (c,h)(c,h), ε\varepsilon and xx as in (b), 𝒞D(c,h)−G(x)>0or(he∈{0,2}for every e,G(x)=23∑eb(he)).\mathcal C_D(c,h)-G(x)>0 \quad\text{or}\quad \Bigl(h_e\in\{0,2\}\ \text{for every }e,\ \ G(x)=\tfrac23\textstyle\sum_e b(h_e)\Bigr).(11)✦

In the paper, the result reads: there is an even-dimensional threshold such that, for every even DD above it, every feasible profile and activity set with ∑jgj≥𝒞\sum_j g_j\ge\mathcal C has cj=0c_j=0 for all jj, he∈{0,2}h_e\in\{0,2\} for all ee and ∑jgj≤23∑eb(he)\sum_j g_j\le\frac23\sum_e b(h_e), where 𝒞\mathcal C is the profile cost 𝒞D(c,h)\mathcal C_D(c,h), the activity set is the set of active faces and ∑jgj\sum_j g_j is the total gain G(x)G(x).

What Is Checked

The three parts of the theorem are stated over the configuration of 13 points and 22 faces, with the sides of each face, the two faces through each point pair and the face labels with their separation counts; over geometric profiles, their face natural costs and the profile cost; over tail tuples with their tail feasibility, planted and tail gains, face baseline, tail slack and the separation excess; and over global tail choices, built from the common and base parameter sets through the planted and lower tail tuples, with their global tail gain.