Sunday, Nov 22 | --:--
Back to home

AI-Aided Navier–Stokes Blowup Claim Drops the Force; Its Key Certificate Is Unreleased

Tristan Buckmaster of New York University and Levent Alpöge of Anthropic posted a 79-page, heavily AI-assisted paper on November 20 claiming that smooth solutions of the three-dimensional Navier–Stokes equations can blow up in finite time with no external force, the case OpenAI's September result left open. The proof depends on a numerical certificate that the authors have not yet released, and the computer check of that certificate is outside their Lean formalization. Terence Tao says the strategy is broadly clear but a proper write-up is still needed.

5 min read New York View as Markdown
A handmade editorial collage on torn ivory paper in cobalt and vermilion: a large still cut-paper whirlpool of layered cobalt and ivory spirals tightens toward a torn point at its centre, with a torn ivory paper flap lying over the very centre and hiding the endpoint; faint pencil marks sit in the surrounding paper, and a small vermilion scrap and a small cobalt scrap rest in the far corner; an official Anthropic logo is glued on a torn-paper clipping whose top edge is tucked under the whirlpool's outer lower ivory edge, with a soft shadow from the whirlpool falling on the clipping, the full logo visible.

The second Millennium-scale claim from Levent Alpöge in one day is the more conventional one, and that makes it easier to test. On November 20, Alpöge, a mathematician at Anthropic, and Tristan Buckmaster of New York University posted a paper claiming that the three-dimensional Navier–Stokes equations can develop a singularity in finite time with no external force acting on the fluid. Their argument follows a known route, which means specialists know what to check. The file the argument stands on has not been published yet.

What they posted#

The paper, "Finite-time blow-up for the Navier–Stokes equations," runs to 79 pages and is dated November 20. Its abstract says the authors prove that smooth, compactly supported, divergence-free starting data in open space, and smooth divergence-free data on the torus, produce solutions of the equations "without force" that blow up in finite time. In the formulation of the Clay Mathematics Institute's Millennium Prize problem by Charles Fefferman, they write, this disproves statements (A) and (B) and so proves (C) and (D).

The repository on GitHub also holds a Lean 4 formalization, which its README describes as about 2,270 files and 600,000 lines. Buckmaster announced the release on Mastodon, writing, according to The Next Web, that "after last time, we couldn't risk waiting, so the Navier-Stokes write-up is unpolished and heavily LLM-assisted."

"Last time" refers to September, when OpenAI published its own Lean-checked Navier–Stokes blowup, a result that relied on a smooth force stirring the fluid. Scientific American reports that many experts saw that force as a loophole in the problem's wording, and that Buckmaster had accused OpenAI of trying to scoop his and Alpöge's work, which OpenAI denied. The new claim removes the force.

On the same day Alpöge, Buckmaster and NYU's Matei Coiculescu released a separate paper proving finite-time blowup for the Euler equations, which describe fluids without viscosity, also without external force. Its repository says the Lean formalization corresponds one-to-one with the paper's proof and contains no unproved placeholders, with one cited criterion left unformalized.

What is missing#

The repository is candid about the gaps:

  • The certificate. The proof turns on an approximate numerical solution that satisfies explicit inequalities, which the authors call a certificate. The paper gives its hash and says it "will be released soon." The repository states that it is not included.
  • The check of the certificate. The README says a rigorous interval-arithmetic computation, at 128 bits of precision using the Arb library through FLINT, verifies that the certificate meets the paper's conditions. That computation is not part of the Lean formalization and is not in the repository either.
  • The build. The Lean package had a patch applied on November 20 to some statements and the proofs that depend on them, and the README says the patched package "was not rebuilt" when the repository was assembled.

The repository's commit history showed no new files after November 20 when we checked on November 22.

Claims and checks#

The blowup claim is the authors' own, and nobody has yet reproduced the numerical step. The closest thing to an outside reading is from UCLA's Terence Tao, who wrote on November 20 that the Navier–Stokes paper is "technically human-readable" and follows "one of the expected paths to this result": numerically build an approximate self-similar solution, then use rigorous perturbation theory to turn it into an exact one. He added that "more work needs to be done to create a proper, professionally written proof." He ranked the Euler paper as the most digested of Alpöge's three announcements and the Lean-only Riemann Hypothesis proof as the least.

Our read: unlike the Riemann claim, this one can be settled by ordinary means once the certificate is out. A matching hash, an independent rerun of the interval computation and a rebuilt Lean package would leave the analytical argument as the remaining question, and Tao's comments suggest experts recognise its shape. Until then the Lean line count does not substitute for the missing data.

Limits#

  • The certificate, its verification run and a rebuilt Lean package were not public when we checked; correctness has not been independently confirmed.
  • Buckmaster's Mastodon post is quoted as reported by The Next Web; we did not open the original.
  • The account of the September dispute is Scientific American's; OpenAI denied wrongdoing.
  • We did not assess the mathematics. Tao's comments are an early reading, not a review.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading