News · 2026-11-21
Independent Lean audit backs claimed proof of RH and Dirichlet GRH
An independent audit reports that a Lean proof submitted by Levent Alpöge passes multiple formal checks for the Riemann Hypothesis and its generalization to Dirichlet L-functions. The checks provide substantial evidence for the precise encoded theorems, while some verification stages and the work of explaining the mathematical ideas remain unfinished.
Key facts
- Andrew Sutherland’s report was updated November 21 at 12:55 UTC in the dossier.
- The report says all 175,187 Lean modules compile and multiple independent kernels accept the exported proof.
- Five headline statements were checked against the challenge; the generalization covers Dirichlet L-functions.
- Primary source: Sutherland’s verification report and logs.
The Riemann Hypothesis concerns the locations of zeros of a function deeply connected to the distribution of prime numbers. Its Dirichlet generalization concerns a family of related functions. The exact family matters: the checked claim is not a proof of every statement mathematicians might call a generalized Riemann hypothesis.
Alpöge’s announcement credits work with Claude and points to a large archive of formal proof files. It sketches a possible path through several areas of advanced mathematics, but is an announcement rather than a conventional, readable derivation. Alpöge says he does not fully understand parts of the argument. The dossier quotes Sutherland’s description of the proof as using only Lean’s standard axioms, including Classical.choice.
A proof assistant checks a chain of formal deductions against explicit rules. Think of it as an exceptionally strict inspector checking every joint in a long bridge: it can establish that the pieces fit its specification, but someone must also check that the specification describes the bridge people intended to build. For that reason Sutherland’s independent comparison of the five statements is important alongside the checks of the derivation itself. The audit reports that the independent nanoda and con-leche kernels accepted the full exported proof, and that lean4lean accepted all compiled modules after timed-out checks were rerun with higher limits.
This is more evidence than a viral announcement alone. But the report recorded other kernel replays and a full comparator run as still in progress at its stated update. Formal verification also does not automatically turn a giant artifact into an explanation mathematicians can teach, reuse or extend. The audit covers the zeta function and Dirichlet L-functions, not the full range of Hecke, Dedekind or automorphic cases.
Terence Tao distinguishes the Lean-only claim from human-readable results and discusses the time needed for mathematical digestion. Scott Aaronson is optimistic while acknowledging uncertainty; his view is not an independent kernel audit. Hacker News discussion reflects the attention and questions, rather than supplying proof.
Alpöge attributes important contributions to Claude and to his own direction of the work. Neither the artifact nor the audit independently divides credit between human and model. For readers, the immediate significance is a precise, substantial formal result with an unusually large remaining interpretive job. The proof assistants lesson explains why checking a formal derivation and understanding its mathematical strategy are related but distinct.
The three axioms listed in the audit are propext, Classical.choice and Quot.sound. The point of naming them is to make the logical foundations inspectable. The audit did not report an extra assumption simply asserting the conclusion; it reported standard Lean axioms. Still, interpreting an encoded theorem requires attention to the definitions on which the statement depends. A correct derivation of the wrong formal statement would not resolve the question mathematicians intended to ask.
The cost of reproducing the artifact is another practical issue. The audit describes a build environment needing roughly 100 GB of memory and 70 GB of disk, a substantial barrier for casual independent replication. The report identifies the archive and software versions so that others can check the work. Successful independent kernels reduce reliance on any single proof checker, but the report’s running jobs and subsequent specialist review remain separate steps.
That distinction matters for reporting the scale of the claim. Saying that no meaningful evidence exists would ignore the reported checks; saying that all imaginable versions of the generalized hypothesis are settled would go beyond the statements actually compared. The strong, testable claim is that the encoded RH and Dirichlet-GRH statements have passed the checks Sutherland reports at the stated time.
Key questions
What has the audit checked?
Does this cover every generalized Riemann hypothesis?
Did Claude independently prove it?
Cite this
APA
Ground Truth. (2026, November 21). Independent Lean audit backs claimed proof of RH and Dirichlet GRH. Ground Truth. https://groundtruth.day/news/lean-audit-dirichlet-grh.html
BibTeX
@misc{groundtruth:lean-audit-dirichlet-grh,
title = {Independent Lean audit backs claimed proof of RH and Dirichlet GRH},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/lean-audit-dirichlet-grh.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.