sysnote
Analysis

Levent Alpöge says he and Claude proved the Riemann Hypothesis in Lean; the 17.4 million lines are still compiling

Levent Alpöge, a mathematician at Anthropic, posted a four-page note on his website on November 20 that begins: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He says the method also proves the Generalized Riemann Hypothesis, and that "Claude and I have also formalized the arguments without assumptions." There is no paper. The proof is a Lean 4 project, linked in the note's first footnote as 14 archive parts, lean.tar.gz.part01 to part14.

The note explains the choice. Alpöge writes that he is presenting "a complete argument before having mastered a massive neighbourhood around its mathematics", and that rather than post "some massive unreadable travesty of 'slop'" (a footnote says that when he tried, it ran past 250,000 pages), he "decided to only present Lean". So the claim stands or falls, for now, on whether that Lean project compiles and proves what it says. Nobody has finished that check yet. Anthropic hasn't commented, according to The Verge.

What the files state

The joined archive is 239,275,554 bytes; the GitHub copy jsmorph/rh and a Hugging Face mirror report the same SHA-256, 9631b4f0…49c7fc. I worked from the unpacked tree in jsmorph/rh, which says it holds the archive's contents without modification. The project pins Lean v4.33.1 and Mathlib commit 0df444a3.

The top of the project is short. src/Main.lean has five theorems, and the first one is the headline:

theorem riemannHypothesis : RiemannHypothesis := rhOfAll

RiemannHypothesis here is Mathlib's own definition, not one the project wrote. I checked it at the pinned commit: it sits in Mathlib/NumberTheory/LSeries/RiemannZeta.lean, says that every zero of riemannZeta other than the trivial ones at −2, −4, … has real part 1/2, and carries the docstring "constructing a term of this type is worth a million dollars". The other four statements are defined in src/Statements.lean, which imports only Mathlib. They cover Dirichlet L-functions: no zeros with real part above 1/2, and zeros in the critical strip on the line, for all characters and for primitive ones. That is what "Generalized Riemann Hypothesis" means in this project. The Extended Riemann Hypothesis, for the zeta functions of number fields, is what Alpöge says he expects to reach "with more time"; the files don't state it.

The step from the analytic statement to the classical one is also visible. Formalism/GammaFactorNonvanishing/RhOfRight.lean (62 lines) derives RiemannHypothesis from "ζ has no zeros with real part above 1/2" using the functional equation, and CritLine.lean (32 lines) does the same for Dirichlet L-functions. Both lean on one statement, UpperTheory.EveryCharacter: no Dirichlet L-function vanishes to the right of the line. Under a hundred lines turn that into the five headline theorems, so the substance of the claim is the proof of EveryCharacter.

How big it is

By my count the project has 175,187 .lean files and 17,374,215 lines. Mathlib at the pinned commit, all of Lean's shared mathematics library, has 8,311 files and 2,293,877 lines, so the proof is about 7.6 times the size of the library it builds on. I parsed the import lines of every file: all 175,187 are in the import closure of Main.lean. There is no unused section to skip. Checking the theorem means compiling all of it.

The build settings point to a heavy compile. The project raises maxHeartbeats, Lean's limit on how much work one declaration may take, to 4,000,000 by default, 20 times Lean's usual 200,000. Four files switch the limit off.

What a text scan can and can't tell you

The first checks people ran on Friday were text searches. Mine agree with the published ones: no sorry, admit, axiom declarations, native_decide, implemented_by, @[extern], unsafe, #eval or initialize anywhere in the tree, and no debug.skipKernelTC, the option that lets declarations into the environment without the kernel checking them. Two kinds of metaprogramming do appear. Five run_cmd lines in src/BuildAid/ only hide names from Lean's lemma-suggestion tactics. One file, CutCptCopy.lean, defines a command cpt_copy_thms that walks a cone of definitions and theorems, substitutes some constants, and adds the copies under new names with addDecl. That is an unusual way to write a proof, but addDecl sends every copy to the kernel.

A clean scan only rules out the crude ways to fake a Lean proof. The real check is the build, then #print axioms on the five theorems (a sound proof uses only propext, Classical.choice and Quot.sound), then a replay of the proofs through independent kernels.

Who is checking, and when to expect an answer

The fullest public attempt is Hertaspace/alpoge-rh-lean-verification, whose README says the checks "were set up and run by a large language model agent (Claude) under human direction". Its build started at 16:12 UTC on a 256-core machine with 2 TB of RAM. The project has 183,903 build jobs; at about 140 a minute the README estimates roughly 20 hours. At the 17:08 UTC refresh of its status page, 15,868 jobs had run, about 139 a minute over the previous ten minutes, which puts the end of the first pass around midday UTC on Saturday. There was one failure: Statements could not read a compiled Mathlib file, which the record puts down to its own cache problem (2,531 of 8,690 Mathlib cache files failed to unpack because of an open-file limit), not to the project. Because Main imports Statements, a second pass will be needed before the axiom check can run. After that come Lean's comparator against a challenge file that copies the four definitions verbatim, a clean-room compile of the statements, and replays through nanoda, con-leche, con-ron and leanchecker.

So the earliest machine verdict comes on Saturday, after the second pass. Even a clean one would settle what the kernel accepts, not whether mathematicians understand the argument. Alpöge says it is "incredibly long at the moment", that parts of his own timeline paraphrase "a 'slop' summary, which might make zero sense", and that he means to simplify and write it up.

The release lacks its own checking kit

The challenge file had to come from Hertaspace, because the archive has none. It contains only the lakefile, the manifest, the toolchain file and src/: no Challenge.lean, no comparator.json, no README and no license.

That is unlike the other formal work around Alpöge. Anthropic's formal-math repository, which includes the Lean proof for the August paper by Alpöge and Ralph Furman on zeta zeros, gives each project a Mathlib-only Challenge module and a comparator configuration, and its CI runs comparator with the nanoda kernel. The Navier–Stokes blow-up paper Alpöge and Tristan Buckmaster posted the same day does the same with comparator*.json files. Its README also says the numerical certificate is not released yet and that the patched Lean package "was not rebuilt" before posting. For the bigger claim of the two, outside checkers are writing the statement files themselves. When OpenAI and Anthropic posted their ETH refutations in October, both shipped a challenge file and comparator configuration with the proof, and outside checkers replayed both under independent kernels within two weeks.

In the August zeta-zeros work, Claude raised the proven share of zeta zeros on the critical line from 41.6% to 67.2%, Alpöge and Furman checked it, and Anthropic's post said: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." Alpöge's new note says the work behind the full claim ran from October 6 to November 16, by a timeline he says Claude assembled. The note doesn't say which Claude model took part.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.