Just in
TechUbisoft Motherbrain AI Unit Launches as H1 Net Bookings Fall 16.5%AIChatGPT Space Rolls Out to Plus Users, Replacing the Library TabTechAnthropic IPO Closes With $119.5 Billion in Combined Gross ProceedsAIRiemann Hypothesis Proof Certified by Lean and Mathlib CommunitiesTechCairn AI Agents Found a Zero-Day in MONAI Label, Amazon SaysAITasks in Gemini Now Calls Businesses for You, With Duplex Inside
AI

Riemann Hypothesis Proof Certified by Lean and Mathlib Communities

The Lean and Mathlib communities say Levent Alpöge's 17 million line Lean proof of the Riemann Hypothesis is correct after checks by several independent kernels.

USA Publishing Group4 min read
Share
Blackboard filled with handwritten mathematical formulas, evoking the newly certified Riemann Hypothesis proof
Photo: Erwan Hesry / Unsplash

Key takeaways

  • The Lean and Mathlib communities said Nov. 21 that both the statement and the proof in Levent Alpöge's Riemann Hypothesis library are correct.
  • The proof is a Lean library of about 17 million lines across 175,187 modules, and it was accepted by Lean's kernel and several independent checkers.
  • An MIT verification page run with Claude Code said two checks were still in progress as of 9:10 a.m. ET on Nov. 22.
In this story 6 sections
  1. A proof written for a machine to check
  2. How the communities checked it
  3. The numbers in MIT's verification log
  4. More than the classic hypothesis
  5. Where AI fits
  6. What comes next

The Lean and Mathlib communities said on Nov. 21 that they have certified a computer-checked Riemann Hypothesis proof released a day earlier by mathematician Levent Alpöge, calling both the theorem statement and the proof correct. "We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct," the groups wrote in a joint statement endorsed by the Lean FRO and the Mathlib Initiative.

The Riemann Hypothesis, proposed by Bernhard Riemann in 1859, holds that every nontrivial zero of the Riemann zeta function sits on the line where the real part equals 1/2. It has gone unproven for 167 years and ranks among the best known open questions in mathematics.

Alpöge's personal website lists Anthropic (NASDAQ: ANT) as his employer, along with a junior fellowship at Harvard's Society of Fellows. The Claude developer priced its IPO at $83 a share on Nov. 19.

A proof written for a machine to check

The proof came out Nov. 20 as a Lean library of about 17 million lines of code, according to the statement. Lean is a programming language and proof assistant in which every logical step has to be accepted by a small trusted program called a kernel. Mathlib is the community's shared library of formalized mathematics, and the statement says Alpöge's code relies directly on nothing else.

A formal proof changes what reviewers have to ask. Instead of following a long argument line by line, checkers focus on two narrower questions: does the code prove exactly the statement mathematicians care about, and does the proof pass the kernel?

How the communities checked it

The joint statement lists four tests and says all of them were met:

  1. The theorem statement, which already lives in Mathlib, was reviewed by specialists in Lean and number theory.
  2. Dozens of people compiled the library and reported matching output, with no errors or failed builds.
  3. A separate file that imports only Mathlib was compared against Alpöge's theorem using two independent tools, and both found the two statements identical.
  4. The exported proof went through Lean's official kernel plus several outside checkers built in different programming languages. Two of them, con-leche and con-ron, carry their own formal proof, relative to set theory, that they cannot accept a false proof.

The groups added that these checks ran on an inert export of the proof, so any hostile code hidden in the source files could not affect the results.

The numbers in MIT's verification log

A verification page kept by MIT mathematician Andrew Sutherland, who says his checks were run with Claude Code, gave this status as of 14:10 UTC on Nov. 22 (9:10 a.m. ET):

  • All 175,187 modules compiled under Lean v4.33.1 with zero errors.
  • The five headline theorems depend only on Lean's three standard axioms.
  • The independent Rust kernel nanoda accepted the full export in 27 minutes and 42 seconds on 80 threads.
  • The con-leche checker accepted 1,468,418 declarations.
  • A sandboxed rebuild on a separate machine, with no network access, produced an export of 61,570,595,788 bytes that matched the published file exactly.

By our math, 17 million lines spread across 175,187 modules works out to about 97 lines per module, and the export file comes to roughly 61.6 GB.

Sutherland's page also says two checks were still running at that point: the end to end lake comparator run and a replay of the export through Lean's default kernel. In the rebuild, 175,186 of the 175,187 compiled files matched; the page says the lone exception compiles differently from one run to the next.

More than the classic hypothesis

According to the MIT page, the library proves five theorems. Besides the Riemann Hypothesis in Mathlib's own wording, it covers Dirichlet L-functions: for every modulus and every Dirichlet character, any zero inside the critical strip lies on the line with real part 1/2. The page notes that the version for primitive characters matches the way the Generalized Riemann Hypothesis is most often written.

Where AI fits

Google search suggestions for the topic now include phrases such as "riemann hypothesis claude" and "riemann hypothesis ai," a sign of how much attention the AI angle is drawing. The primary documents we were able to open describe AI's part in the checking, not the writing: Sutherland's page credits Claude Code for running the verification. Alpöge posted his announcement as a PDF on his website, which we could not read in full, so this article does not describe his account of how the proof was produced.

For background on the kind of software involved, see our explainer on large language models, and follow more research and model news on our AI page.

What comes next

The certification settles correctness in the formal sense. The Lean and Mathlib groups said they now expect mathematicians to work on explaining "the new ideas in this proof." Human understanding of why the argument works, and any recognition from prize committees, are separate matters the statement does not take up.