OpenAI Claims an All-Dimensions Local Smoothing Proof. Here Is What the Artifacts Say.

OpenAI has published a mathematical claim large enough to require more than a benchmark chart: an all-dimensions proof of local smoothing for the Euclidean wave equation, with consequences for Kakeya geometry and Fourier restriction. The useful news is not simply that a model produced persuasive mathematics. There is a 59-page manuscript and a public Lean development that specialists can inspect.

That does not make independent review optional. OpenAI itself describes these manuscripts as raw research data, rather than the finished scientific understanding that researchers still need to develop. That distinction should govern the headline and the procurement conversation alike.

What the paper actually states

The October 25 manuscript states its main theorem for spatial dimensions n ≥ 2 and every finite p ≥ 2. It bounds the space-time Lp norm of a half-wave over the interval [1, 2] using a Sobolev norm of the initial data. The required regularity is max(0, (n−1)/2 − n/p) plus any positive epsilon; at p = 2, epsilon can be zero.

In ordinary language, a wave can become extremely concentrated at a particular instant. The conjecture asks how much better its behavior becomes when measured across time instead of only at that worst instant. The gain is quantitative, not a claim that waves stop focusing.

The epsilon qualification matters. A result with arbitrarily small positive loss is not automatically a zero-loss endpoint estimate. The paper also discusses obstructions to endpoint smoothing. A headline that compresses this into “every wave problem solved” would erase exactly the mathematical boundaries that make the theorem meaningful.

Why Kakeya appears in a wave-equation story

The manuscript connects local smoothing to estimates for Fourier extension, then to how thin tubes in many directions can overlap. It claims Kakeya maximal estimates with an arbitrarily small scale loss and, consequently, full Hausdorff dimension for every Euclidean Kakeya set.

A Kakeya set contains a unit line segment pointing in every direction. Asking how small such a set can be is a geometric problem, but its connections with wave packets and Fourier analysis mean a strong analytic estimate can have geometric consequences. The manuscript presents those connections as deductions, not as three unrelated model discoveries.

The paper also claims Bochner–Riesz bounds, Stein square-function bounds, and Fourier restriction estimates for the sphere, compact paraboloid, and truncated cone in the interior of the conjectured range. “Interior” is another important boundary: this is not a blanket assertion about every limiting case of every restriction formulation.

The prior work has names and dates

This research does not begin with an AI system. An Annals of Mathematics paper by Larry Guth, Hong Wang, and Ruixiang Zhang proves a sharp square-function estimate and, consequently, local smoothing in two spatial dimensions plus time.

Its publication record is specific: received October 9, 2019, accepted June 22, 2020, and published online September 9, 2020. OpenAI refers to the result as a 2019 advance; the journal record provides the distinction between when the work entered the publication process and when the paper appeared.

The new claim is an extension across dimensions, not the discovery that local smoothing is an important question. Crediting the earlier analytic framework is essential to understanding what would actually be new if the proof survives scrutiny.

What the Lean release adds

OpenAI says it obtained the natural-language output on October 18, 2026, and formalized it in Lean the following day. Its scope document lists the wave estimate and several consequences, including full Hausdorff dimension for Kakeya sets.

We inspected both the standalone reference statements and the proof-side Comparator file. The latter supplies theorem bodies by referring into the published mathematical development. That is materially different from a PDF saying that a formal proof exists somewhere.

There is a trap for casual repository readers: the reference challenge file contains sorry placeholders because it specifies targets, rather than supplying the actual solution. Seeing those placeholders alone does not establish that the proof-side development has missing proofs. Conversely, reading the final export declarations is not a complete audit of their dependencies.

Developer312 did not rebuild this Lean development, check its complete axiom dependencies, or independently verify that every formal definition matches the intended mathematics. Those are outstanding verification tasks, not accomplishments we can infer from a repository listing.

Publication is the start of examination

For this result, OpenAI says an outside expert recommended releasing the model output publicly and allowing researchers to organize their own explanation efforts. That differs from its physics collaboration, where artifacts are being withheld while an academic group develops a writeup.

Public release makes the local smoothing claim more inspectable. It does not tell us how quickly specialists will establish correctness, identify a limitation, or find a clearer argument. Independent contextualization matters even if a formal checker accepts a statement: humans still have to assess the statement, assumptions, significance, and relationship to existing literature.

The builder takeaway is about evidence

This is not evidence that a general chat session can reliably replace a research mathematician, nor does the release establish a commercial performance metric. The announcement does not provide a deployment recipe or price for reproducing this particular achievement.

It does illustrate a better standard for scientific AI claims: publish the exact theorem, the readable argument, the formal target, and the proof artifacts. Then separate vendor-reported verification from independently reproduced verification.

For teams evaluating research agents, ask for that evidence chain rather than extrapolating from a dramatic headline. The substantive signal here is a potentially major mathematical result with public materials. Whether it becomes established mathematics depends on examination of those materials, not on the confidence of the announcement.

AI assistance disclosure: This draft was researched and written with a GPT-class language model. Human editorial review is pending.