Navier–Stokes Lost In Translation

TL;DR

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

A new arXiv preprint by Alexander Bastounis argues that the Lean formalization associated with OpenAI’s announced Navier–Stokes blow-up proof does not faithfully represent the natural-language argument. The paper says this mismatch shows that mechanically checking a formal proof does not, by itself, verify the original written claim; its findings have not been independently established here as a peer-reviewed result.

A new arXiv preprint argues that the Lean formalization associated with OpenAI’s announced proof of blow-up for the Navier–Stokes equations does not represent the natural-language argument it is meant to check. The paper, submitted by Alexander Bastounis on October 6, 2026, raises a basic reliability issue for AI-assisted mathematics: a formal proof can be mechanically verified while still failing to establish the claim expressed in the original prose.

Bastounis’s 25-page paper examines autoformalisation, the process of translating mathematical writing into a formal system such as Lean. The author argues that this translation must preserve the meaning of the source text before a proof assistant’s verification can provide evidence about the original argument. The preprint says ambiguity in mathematical natural language can make that task exceptionally difficult.

The abstract places the difficulty in the Solvability Complexity Index and arithmetical hierarchies, claiming that semantically faithful translation can require solving problems of arbitrarily high complexity. It contrasts that task with the halting problem, which the abstract describes as having SCI = 1. These are claims made by the preprint; the supplied material does not include an independent assessment of the technical argument.

The paper says it gives several examples of mismatches between natural-language statements or proofs and their Lean versions, including the OpenAI-announced Navier–Stokes proof. Its central case-specific claim is that the Lean proof does not correspond to the written proof of blow-up. That is a claim about the relationship between the two representations, not a reported independent resolution of the underlying Navier–Stokes problem.

When a Checked Proof Changes Meaning

The issue matters because formal verification is often presented as a way to make mathematical reasoning more reliable. A proof assistant can check whether a formal statement follows from specified premises under the system’s rules. But that check only answers a limited question: whether the encoded proof establishes the encoded proposition. If the encoding changes the meaning of the prose, the mechanical check cannot, on its own, validate the original claim.

For readers evaluating AI-generated mathematics, the paper therefore draws a distinction between a verified formal object and a faithful translation of a human-readable argument. The distinction is relevant beyond this one example: researchers, developers and journalists need to know what proposition was formalized and how it relates to the claim being reported. The preprint does not show that formal verification is useless; rather, it challenges the assumption that verification of a translation automatically verifies its source.

From Natural Language to Lean

Lean is a proof assistant in which mathematical definitions and arguments are represented in a formal language. Once a proposition and proof have been encoded, software can check the formal derivation. Autoformalisation aims to automate or assist the conversion from ordinary mathematical writing into that representation, a capability of growing interest as AI systems produce or help assess proofs.

The preprint’s immediate setting is OpenAI’s announced proof concerning blow-up of Navier–Stokes solutions. The supplied arXiv abstract does not provide details of the original announcement, the exact formal statement, or the proof files needed to compare the two versions. It says the article includes practical mistranslation examples and identifies the Navier–Stokes case among them. The submission is listed as a 25-page paper with four figures, in analysis of partial differential equations, artificial intelligence and logic.

The paper was posted to arXiv on October 6, 2026, as version one. An arXiv posting is a public preprint submission, not evidence by itself that the work has passed peer review. The listing says its DOI registration is pending.

“The formalised Lean proof does not correspond to the NL proof of blow-up of solutions to the Navier–Stokes equations.”

— Alexander Bastounis, author of the arXiv preprint

The Formal and Written Proofs

The supplied source material does not include the paper’s detailed comparison, the relevant Lean code, or a response from OpenAI. It is therefore not possible from the abstract alone to assess exactly where the alleged mismatch occurs, whether the formalization was intended to encode the full written argument, or how the proof’s authors characterize the relationship between them.

The technical result about the complexity of faithful translation is also presented here as the preprint’s conclusion, not as an independently confirmed consensus. The source identifies the submission as an arXiv preprint and gives no peer-review status. The broader mathematical status of the announced Navier–Stokes blow-up result is not settled by the abstract, and the material provided does not establish whether the result has been accepted by the mathematical community.

Scrutiny of the Translation Claim

The next steps for readers and researchers are to examine the full preprint’s examples, definitions and comparison of the natural-language argument with the Lean formalization. A meaningful assessment will require access to the precise formal statement and proof, alongside the written claim, so that specialists can determine what is encoded and whether any discrepancy changes the mathematical conclusion.

Further developments may include responses from OpenAI or the authors of the announced proof, independent analysis by mathematicians and formal-methods researchers, and later revisions or peer review of Bastounis’s paper. The supplied material does not report a response or set a date for any such review, so the status remains developing.

Key Questions

What is the new paper claiming?

The preprint argues that the Lean formalization associated with OpenAI’s announced Navier–Stokes blow-up proof does not correspond to the natural-language proof. It also argues that translating mathematical prose faithfully can be exceptionally difficult.

Does the paper disprove the Navier–Stokes result?

The abstract does not establish that the underlying mathematical claim is false. Its stated criticism is that the formal Lean proof does not match the written argument; the supplied material does not settle the result’s broader mathematical status.

What does Lean verification establish?

A Lean checker can verify a proof of the proposition encoded in Lean, under the system’s rules and assumptions. The preprint’s point is that this check does not by itself establish that the encoded proposition faithfully represents the original prose.

Has the preprint been peer reviewed?

The supplied listing identifies it as an arXiv preprint submitted October 6, 2026. It does not report peer review or acceptance by a journal.

Source: hn