VibeMathedMath problems solved with AI

The polynomial removal conjecture for ordered binary matrices

For a fixed binary k×kk\times k matrix HH, an ordered copy in a binary n×nn\times n matrix AA is a choice of rows r1<⋯<rkr_1<\dots<r_k and columns c1<⋯<ckc_1<\dots<c_k with A(ri,cj)=H(i,j)A(r_i,c_j)=H(i,j), zeros included. Alon, Ben-Eliezer and Fischer proved qualitative ordered removal: if AA is ϵ\epsilon-far from HH-free (at least ϵn2\epsilon n^2 entry changes needed), it has at least δ(ϵ)n2k\delta(\epsilon)n^{2k} copies. Alon and Ben-Eliezer (Problem 1.4) asked whether δ\delta can be polynomial, and Gishboliner and Shapira conjectured it for each single HH. Are there cH,CH>0c_H,C_H>0 with NH(A)≥cHϵCHn2kN_H(A)\ge c_H\epsilon^{C_H}n^{2k} whenever AA is ϵ\epsilon-far from HH-free?

Result
Disproved(see note)
Status
Candidate (review pending)
AI contribution
AI-discovered
Method
Construction
Field
Extremal combinatorics; property testing
Posed by
Noga Alon and Omri Ben-Eliezer (Problem 1.4, 2020); Lior Gishboliner and Asaf Shapira (Conjecture 4.5, 2025); question raised by Alon, Fischer and Newman (2007)
Year posed
2007
Years open
19y
Solved
2026-09-25
Model
Unreleased internal OpenAI model
Vendor
OpenAI
Collaborators
—
Verification
Lean-checked, statement unaudited
Publication
Announced
Collection
OpenAI math release (October 2026), version adc7f12
Significance
15 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Theorem 1.1: there is a fixed binary 66×6666\times66 matrix HH such that for every h≥1h\ge1, with nh=(386h+2)2hn_h=(386h+2)2^h and ϵh=(386h+2)−2\epsilon_h=(386h+2)^{-2}, some AhA_h is ϵh\epsilon_h-far from HH-free yet NH(Ah)≤ϵh2−hnh132N_H(A_h)\le\epsilon_h2^{-h}n_h^{132}; hence no polynomial removal bound holds for this HH, and the finite-family question fails for one square pattern. Corollary: canonical row-column sampling testers need exp⁡(Ω(ϵ−1/2))\exp(\Omega(\epsilon^{-1/2})) samples along this sequence. Not shown: which patterns do admit polynomial removal, or a counterexample smaller than 66×6666\times66.

What the AI did

The release README says every result was produced by an unreleased internal OpenAI model with a fixed procedure of roughly three hours of ChatGPT Pro thinking compute per result. This family is not among the README's exceptions (the Hodge conjecture for CM abelian varieties, and the Re(s) > 11/12 zero-free region, whose write-up was human-edited). The manuscripts are authored 'OpenAI' and name no human author. The family is a single manuscript dated September 25, 2026.

Verification

No independent mathematician has checked this yet. Checked here: the introduction and Theorem 1.1 were read against the conjecture as the manuscript quotes it. Lean: formalization.yaml lists ComparatorChallenges/MatrixRemoval.json, declaration OAI.Problem348.no_polynomial_removal_bound in OAI/Combinatorics/MatrixRemoval/Main.lean. The challenge statement was read: for one explicitly defined 66x66 Boolean matrix and all c, C > 0, there are n, 0 < eps < 1 and an n x n binary matrix whose normalized edit distance to H-freeness (edits in both directions) is at least eps and whose ordered-copy count is below c eps^C n^132. That is the headline claim. Not rebuilt here.

Sources

Changelog1 change

Discussion