Lean FRO and Mathlib certify Alpöge's Lean proof of the Riemann Hypothesis: what was checked, and by what
On November 21 the Lean FRO and the Mathlib Initiative published a joint statement on Levent Alpöge's Lean proof of the Riemann Hypothesis: "We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct." The @leanprover account posted it at 15:12 UTC and added in a reply that it is a joint statement of the Lean and Mathlib communities, endorsed by the two organizations.
The Lean FRO and Mathlib Initiative have released a joint statement on the Lean verification of the Riemann Hypothesis.
On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. Over the past day, the Lean and Mathlib… pic.twitter.com/pYR61ZaREc
A day earlier, when I wrote about the release, the 175,187-file project was still compiling and nobody had a verdict. The statement says the proof now meets all four tests Kevin Buzzard set out on November 12 in "…and the proof is checked in Lean": the Lean statement is the right one, the code compiles, it proves that statement and not a lookalike, and it doesn't rely on a bug in Lean's kernel. The evidence behind it is mostly on two public pages: the verification files of Andrew Sutherland at MIT, whose checks were run with Claude Code, and the notes in leanprover-community/riemann-hypothesis, a Comparator setup that GitHub user Deicyde started on November 20 and that now sits in the Lean community's organization. I read both against the statement.
What ran
Sutherland's page, last updated at 12:55 UTC on November 21, reports that:
- all 175,187 modules compile under Lean v4.33.1 with zero errors, and the compiled Mathlib files are byte-identical to the official cache;
#print axiomson the five theorems inMain.leanlists onlypropext,Classical.choiceandQuot.sound;leancheckerreplayed every module without a rejection;- the exported proof, 61,570,595,788 bytes unpacked, was accepted by nanoda, an independent kernel written in Rust (27 minutes on 80 threads), and by con-leche (1,468,418 declarations, 3 minutes 15 seconds on 64 threads);
- con-leche's
--comparemode confirmed that the five proved theorems match a challenge file that imports only Mathlib; - lean4lean accepted every compiled module, five of them after its step limit was raised.
Lean's default and paranoid kernels on the export, and an end-to-end lake comparator run, were still going at that update. In the comparator run the statement comparison and axiom check had already passed.
The page also lets anyone repeat the main check without building 17 million lines. With Lean v4.36.0, you download the 7.7 GB compressed export and the 47 MB challenge, compare hashes and run the checker. Sutherland puts it at about 100 GB of RAM for the full lake comparator; con-leche alone needs about 50 GB and took under four minutes on 64 threads.
Where the statement says more than the logs
The statement says the export "has been checked by several proof checkers, including Lean's official kernel, the con-leche and con-ron checkers", and that these two "come with a formal proof, relative to set theory, that they cannot accept a false proof".
The published logs are narrower. The con-leche and con-ron binaries bundled with Lean v4.36.0 exited with code 2 on this export. That is a refusal, not a rejection: they don't support a mutual inductive block with a reflexive member, and the proof has one, FunctorialResolution.EmbeddedResolution.PrincipalizationCertificate. The community notes reproduced the refusal on a small test project. The con-leche that accepted is a build of its master branch at cd2eceab, a commit from November 17, three days before the proof appeared, so it was not patched for this case. I found no published log of con-ron accepting the export. Because of the refusals, the community repository's comparator.json names only Lean's paranoid kernel and nanoda as extra checkers.
The statement also says "dozens of people have reported correct compilation of the codebase, reproducing the same raw proof export file." The repository's verification notes, last changed at 16:32 UTC, document one second build that reproduced the export's checksum (by GitHub user ishaan2006, in a different directory) and say that "none of these builds was sandboxed". Reports on the Lean Zulip may have run ahead of the notes; I didn't check Zulip.
Neither point suggests a fault in the proof. Two independent kernel designs, Lean's and nanoda's, plus a con-leche development build, accept the same export. But the con-ron part of the claim has no log behind it that I could find, and the formally verified checker that did accept is a master build, not the version shipped with Lean.
Two things my first article understated
In #191 I wrote that the text scans found two kinds of metaprogramming. The community notes, corrected after issue #2, list more. A tactic called kexactB, used through small macros in 67 files, "elaborates a term and assigns it to the goal without the elaborator checking that its type matches"; kcopy rebuilds terms field by field in the same way; and run_tac appears 33 times. For those proofs the kernel is the only check. That is sound, because the kernel checks every declaration, and the notes give it as the reason the repository re-checks the export with independent kernels.
The second is the Lean version. The project pins v4.33.1, which predates two runtime fixes in 4.34.1 (#15241 and #15289) that, per the notes' reading of the release notes, could potentially be turned into a proof of False in the official kernel with deliberately constructed input. So a successful build counts, in the notes' words, "as limited evidence here", and the judging is left to Lean 4.36.0 and the independent checkers. The release notes say kernels not built on the Lean runtime are unaffected; nanoda, written in Rust, is one of them.
What exactly is proved
The checked statements are Mathlib's RiemannHypothesis and four statements about Dirichlet L-functions. "GRH" in this project covers every Dirichlet character, and nothing beyond: not Dedekind zeta functions, not other L-functions.
One gap people raised is that the Dirichlet statements only talk about zeros with positive real part. In issue #34 a contributor proved in Lean that the project's PrimitiveDirichletLZerosOnLine is equivalent to GRH.generalized_riemann_hypothesis from Google DeepMind's formal-conjectures repository, a statement written independently in November 2025. The bridge uses only the three standard axioms, so nothing is lost on the left half of the plane.
Sutherland also checked a consequence that doesn't depend on Mathlib's definition of the zeta function. He combined the proof with kimihiro64/Robin1984, an independent formalization of Robin's and Lagarias's criteria, and got Robin's inequality, σ(n) < e^γ · n · log log n for every n > 5040, as a Lean theorem that never mentions zeta. nanoda and con-leche accept the combined 61.75 GB export.
What is still open
The certification covers the formal proof. It says nothing about whether anyone understands it, and the statement doesn't pretend otherwise: it ends by looking forward to "all the activity in the mathematical community that will help explain the new ideas in this proof." Andrew Granville told Nature that Alpöge's four-page note is "garbled", and Alpöge has said he wants to simplify the argument and write it up. Meanwhile contributors are drawing maps of the 17 million lines in the repository's issues, from call paths to a table of the classical theorems buried in the archive.
- Lean FRO and Mathlib Initiative: Lean verification of the Riemann Hypothesis (joint statement)
- @leanprover: joint statement on the Lean verification of the Riemann Hypothesis
- Andrew Sutherland: Verification files for the Lean proof of the Riemann Hypothesis (12:55 UTC update)
- leanprover-community/riemann-hypothesis: Comparator setup
- riemann-hypothesis verification notes
- Issue #34: PrimitiveDirichletLZerosOnLine is equivalent to formal-conjectures' GRH (machine-checked)
- Issue #2: Source-level import cone
- Kevin Buzzard (Xena): "...and the proof is checked in Lean"
- con-leche commit cd2eceab (Nov 17, toolchain v4.36.0)
- Nature: Mathematician says he and AI solved Riemann hypothesis
Comments
No comments yet. Be the first.