VibeMathedMath problems solved with AI

Ulam's floating-body problem for origin-symmetric bodies

Ulam (Scottish Book, Problem 19) asked whether a convex body of uniform relative density ρ∈(0,1)\rho\in(0,1) that floats in equilibrium in every orientation must be a Euclidean ball. Ryabogin (2022) answered this negatively with non-centrally-symmetric bodies of revolution at ρ=12\rho=\frac12 in every dimension d≥3d\ge3. For origin-symmetric bodies the answer at ρ=12\rho=\frac12 is yes (Schneider and Falconer in dimension 3; Florentin, Schütt, Werner and Zhang in all dimensions), and Ryabogin noted that no results were known for other densities. Must an origin-symmetric convex body in Rd\mathbb R^d, d≥3d\ge3, that floats in equilibrium in every orientation at some density ρ≠12\rho\neq\frac12 be a ball?

Result
Disproved(see note)
Status
Candidate (review pending)
AI contribution
AI-discovered
Method
Construction
Field
Convex geometry
Posed by
Stanisław Ulam (Scottish Book, Problem 19); the origin-symmetric case at densities other than 1/2 stayed open after Schneider and Falconer
Year posed
—
Years open
—
Solved
2026-10-29
Model
GPT-6 Pro (ChatGPT); GPT-6.1 Sol and GPT-6 Astra (Codex)
Vendor
OpenAI
Collaborators
Liam Price, Tom de Groot
Verification
Unreviewed
Publication
Preprint
Significance
18 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Theorem 1 and Corollary 2: in every dimension d≥3d\ge3 there are nonspherical, origin-symmetric, strictly convex bodies of revolution with real-analytic boundary that float in equilibrium in every orientation. Their densities lie near the cap fraction of the unit ball at height (d+4)−1/2(d+4)^{-1/2} (about 0.2300.230 for d=3d=3), and the same bodies float at the complementary density. The bodies exist on closed parameter sets of asymptotically full measure near an isolated degree-four resonance. Not addressed: dimension 2 (floating logs), and exactly which densities admit such bodies.

What the AI did

From the paper's disclosure: the original argument, for dimension d=4d=4, was developed by GPT-6 Pro in ChatGPT. GPT models working in Codex (Hexagon's provenance record names GPT-6.1 Sol and GPT-6 Astra) extended it to every dimension d≥3d\ge3 and completed the proof, and assisted with drafting, revising and proofreading. Hexagon labels the text 'a mix of human-written and AI-generated text', with human understanding of 'some parts'.

Verification

Checked here on 15 November 2026. The Hexagon paper (2610.00629v1, de Groot and Price, 29 October 2026) was read through Theorem 1, Corollary 2 and Lemma 3, and its disclosure of AI contributions. The posed question was checked in Ryabogin, Ann. of Math. 195 (2022), whose introduction says that no results were known at densities other than 1/2, and in Florentin-Schütt-Werner-Zhang (arXiv 2010.09006). An exact computation written here confirmed Lemma 3 as used (the second derivative of the zonal harmonic vanishes at h=(d+4)−1/2h=(d+4)^{-1/2} for degree 4 and for no other even degree up to 80, d=3,…,40d=3,\dots,40) and that the limiting density for d=3d=3 is 12−577≈0.230\frac12-\frac{5}{7\sqrt7}\approx0.230. The analytic part (small divisors via Matveev's bound, Newton iteration) was not checked. No Lean formalization.

Sources

Submitted by MellowGander562 on

Changelog2 changes

Discussion