OpenAI claims 3D restriction and Bochner–Riesz conjectures
Awaiting review
Importance: major (4 of 5)The takeaway
Beside quasi-RH and Unique Games, OpenAI’s 6 Oct catalogue claims a whole cluster of central harmonic-analysis conjectures in three dimensions: Stein’s restriction conjecture for the sphere, the diagonal extension conjecture, Bochner–Riesz, Sogge’s local smoothing and the planar endpoint of Carleson’s Schrödinger problem. None has a Lean formalization and no expert review has been published; the first arXiv paper to cite one appeared on 15 Oct.
Status
- Claim
Awaiting review
- Our reporting
- Medium confidence
- Verification
- Preprint only
- Importance
- Major (4 of 5)
- Last verified
- 19 October 2026
Your AI and this story
- GPT-6 Astra159 days after its cutoff
- Claude Opus 5.598 days after its cutoff
- Gemini 4 Argon189 days after its cutoff
- Grok 4.7128 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 98 days before it.
Key facts
- Family 077 (catalogue): ‘Proves the diagonal Fourier extension conjecture for positively curved surfaces in three dimensions … bounded from Lᵖ(Σ) to Lᵖ(ℝ³) for every p > 3.’ Two papers dated 24 Sept 2026: ‘Elliptic capacity propagation and Fourier restriction to the sphere’ (103 pp.; ‘the Fourier extension operator maps L∞(S²) boundedly into Lᵖ(ℝ³) for every p > 3. This is the full conjectured open range, and the threshold p = 3 is sharp’) and ‘Diagonal Fourier extension for positively curved surfaces in three dimensions’ (25 pp.)
- Prior state of the art, as the sphere paper itself summarises it: Guth’s polynomial partitioning gave the bounded-data estimate for p > 13/4, Hong Wang’s broom method p > 42/13, and Wang–Wu the diagonal estimate for p > 22/7. The claim closes the remaining gap to p > 3
- Dependencies: the diagonal paper ‘uses two theorem inputs’: the packet propagation theorem of the sphere paper and ‘The Kakeya maximal theorem [12, Theorem 1.1]’, i.e. OpenAI’s family 074 manuscript ‘The Kakeya maximal conjecture in three dimensions’ (23 Sept). The restriction claim therefore stands or falls with the Kakeya maximal claim, which Levent Alpöge linked to a scooping dispute
- Family 078: ‘Resolves the three-dimensional Bochner–Riesz conjecture in its strict range’ (paper ‘Bochner–Riesz multipliers in three dimensions’, 114 pp.; main estimate: Tδ bounded on L³(ℝ³) for every δ > 0, then interpolation and duality)
- Family 079: ‘Resolves Sogge’s local smoothing conjecture for the Euclidean wave equation in three spatial dimensions’ (paper ‘Critical local smoothing for the three-dimensional wave equation’, 165 pp.; critical L³ estimate with every positive Sobolev loss)
- Family 080: ‘Proves almost-everywhere convergence e^{itΔ}f → f … for every f ∈ H^{n/(2(n+1))}(ℝⁿ) and every dimension n ≥ 2’, the sharp Sobolev endpoint of Carleson’s problem (planar paper ‘Endpoint convergence for the planar Schrödinger equation’, 83 pp.)
- Verification: the catalogue entries for 074, 077, 078, 079 and 080 carry no Lean link (073 Falconer and 075 L log L Fourier convergence do), lean/docs/077.md returns 404, and neither 077 PDF mentions Lean. Korobenko–Sawyer’s remark that OpenAI proved the diagonal case ‘with L∃∀N certification’ is not supported by the repository as checked on 19 Oct
- First follow-up: Lyudmila Korobenko (Reed College) and Eric T. Sawyer (McMaster), ‘Probabilistic Fourier extension and the Schrödinger equation’, arXiv 2610.22634 (submitted 15 Oct, 8 pp.): ‘Just recently, OpenAI has claimed to prove the diagonal inequality p = q in the case of d = 3 dimensions’; their averaged Sobolev regularity result for the Schrödinger equation is the probabilistic counterpart of ‘the deterministic analogue in three dimensions … obtained recently by OpenAI’. The paper has no AI-use statement
Show 1 more
- Coverage gap: our main release entry listed Kakeya maximal and Falconer but not these four families; a viral 7 Oct summary post had listed ‘Fourier restriction conjecture — PARTIAL’. No press article, Hacker News thread or expert blog post on the restriction or Bochner–Riesz claims was found on 19 Oct
What happened
OpenAI’s 6 Oct 2026 release (see the main entry) is so large that some of its biggest claims went unnoticed here for two weeks. Families 077–080 sit in the analysis section of the catalogue, next to the Falconer (073) and Kakeya (074) families, and claim most of the classical “Stein programme” in three dimensions:
- Fourier restriction (077). Elias Stein’s restriction problem (late 1960s) asks for which exponents the Fourier extension operator of the sphere is bounded. In ℝ³ the conjecture is that L∞(S²) → Lᵖ(ℝ³) holds for every p > 3. The best human results, as OpenAI’s own paper lists them, were p > 13/4 (Guth, polynomial partitioning), p > 42/13 (Hong Wang, “brooms”) and the diagonal estimate for p > 22/7 (Wang and Wu). OpenAI’s 103-page paper claims the full range, and a 25-page companion claims the diagonal Lᵖ(Σ) → Lᵖ(ℝ³) bound for every compact positively curved surface and every p > 3.
- Bochner–Riesz (078). A 114-page paper claims boundedness of the Bochner–Riesz means on L³(ℝ³) for every positive order, which by interpolation gives the conjectured range.
- Local smoothing (079). A 165-page paper claims Sogge’s local smoothing conjecture for the wave equation in three space dimensions.
- Carleson’s Schrödinger problem (080). The catalogue claims almost-everywhere convergence at the sharp Sobolev exponent n/(2(n+1)) in every dimension; the planar paper (83 pages) treats H^(1/3)(ℝ²).
The diagonal restriction paper says it depends on the sphere paper and on OpenAI’s 3D Kakeya maximal theorem (family 074). These four families have no Lean formalization in the repository, unlike the release’s showcase results.
On 15 Oct Lyudmila Korobenko and Eric Sawyer posted the first arXiv paper we found that cites one of these claims. They prove a probabilistic (“martingale transform”) version of Fourier extension estimates for the Schrödinger equation, and describe OpenAI’s diagonal theorem as the deterministic analogue that has been “claimed”. They do not check it.
Why it matters
If they hold, these papers settle, in dimension three, problems that have organised harmonic analysis since the 1970s, and that Bourgain, Wolff, Tao, Guth, Wang and others advanced step by step. They are long, unformalized and unreviewed, and they rest on the equally unreviewed Kakeya maximal manuscript. They are a natural next target for the kind of expert check that confirmed OpenAI’s Brennan proof. Readers should treat them as claims.
What is disputed or not yet verified
| Verification | Unverified manuscripts; no Lean formalization listed for families 077–080; one arXiv follow-up (Korobenko–Sawyer) cites the diagonal claim without checking it |
|---|
Sources
7 sources from 2 sites. Numbers match the chips in the text.
7 sources: 7 primary
Primary
- OpenAI math catalogue (CONTENTS.md, families 074–080)github.com, paper
- OpenAI: Elliptic capacity propagation and Fourier restriction to the sphere (24 Sept 2026, PDF)github.com, paper
- OpenAI: Diagonal Fourier extension for positively curved surfaces in three dimensions (PDF)github.com, paper
- OpenAI: Bochner–Riesz multipliers in three dimensions (PDF)github.com, paper
- OpenAI: Critical local smoothing for the three-dimensional wave equation (PDF)github.com, paper
- OpenAI: Endpoint convergence for the planar Schrödinger equation (PDF)github.com, paper
- arXiv 2610.22634: Probabilistic Fourier extension and the Schrödinger equation (Korobenko, Sawyer)arxiv.org, paper
Changes
- Filed (catalogue and the five OpenAI PDFs read)