RH · checked in Lean

Status

The proof is settled; the news around it is not. This page is the one place on the site for facts that change quickly: checks still running, responses, peer review and the Clay prize. Every other page links here instead of repeating them.

Where things stand

Last checked: see below.

What to watch next

These are the things that would change this page. When one happens, we update the facts above and the date.

What does not change here. Whether RH is proved does not depend on any item on this page. Machines checked every step, and the Lean and Mathlib communities certified the proof on 21 November 2026. What is still moving is the human side: explanation, review and recognition. See verified, certified.

The record so far

Verified, certified, peer-reviewed: three different things. Verified here means a machine checked every step: Lean's own kernel and independently written checkers (nanoda, con-leche, con-ron on the full export; lean4lean on every compiled module) accepted the proof (MIT verification page). Certified means the Lean and Mathlib communities' formal statement that both the statement and the proof are correct (joint statement, 21 Nov 2026). In plain terms, the community statement certifies the Lean proof and how it was checked; it is not a peer review of the ideas, and it doesn't mean anyone yet understands the argument in full. No journal has published the proof, and mathematicians are still reading it. MIT's Andrew Sutherland: "There is much work to be done to understand the mathematical details of the proof" (Scientific American, 20–21 Nov 2026). For peer review and the Clay prize, see Where things stand above.

What happened

  1. 20 Nov 2026. Alpöge posts a short announcement (remark.pdf) and a Lean library split into 14 archive parts (listed in the announcement's first footnote). He writes that he found the proof "working with Claude", and that he chose to "only present Lean" rather than a very long human-readable write-up, which he intends to produce later.
  2. Within a day. Volunteers compile the library, isolate the statement in a file that imports only Mathlib, and run several independent proof checkers on an exported copy of the proof.
  3. 21 Nov 2026. The Lean and Mathlib communities publish a statement certifying that the four standard criteria for trusting a Lean proof are met. MIT's Andrew Sutherland publishes verification files and logs.

What "verified in Lean" means

Lean is a programming language in which every step of a proof must be justified by earlier steps, down to a handful of basic rules of logic. A small program called the kernel checks every step. You don't have to trust the author, or the AI that helped, or read 17 million lines. You have to trust four things:

  1. the statement really says RH (here it is Mathlib's own, expert-vetted definition);
  2. the code compiles with no errors;
  3. the theorem proved is exactly that statement, and uses only the standard axioms;
  4. the checker is sound. Several independent kernels accepted the proof, and two of them come with a formal proof that they can't accept a false proof.

Developers can read how this works in detail.

What changes, and what doesn't

For mathematics

Many published results were proved "assuming RH" or "assuming GRH". The ones that rest on the cases proved here (RH, and GRH for Dirichlet L-functions) no longer need that assumption. RH has been one of the Clay Mathematics Institute's Millennium Prize Problems since 2000.

Myth: "It breaks encryption"

No. RSA-style systems rely on factoring large numbers being hard, and elliptic-curve cryptography on a different hard problem (discrete logarithms on elliptic curves). RH is about how primes are distributed on average. It gives no fast way to solve either problem, and nothing in this proof is an algorithm for breaking codes.

What does change in computing

The clearest example is Miller's primality test. Miller (1976) proved that, if GRH holds for Dirichlet L-functions, every odd composite \(n\) has a strong witness (a base that the Miller–Rabin test catches) below a constant times \((\ln n)^2\), so testing those bases decides primality; Bach (1990) made it explicit: under that hypothesis, every odd composite \(n\) has a strong witness below \(2(\ln n)^2\). That case of GRH is among the checked theorems, so the guarantee now holds outright. (The step from GRH to the test is Miller's and Bach's classical proof, not part of the Lean check. GRH for number fields' Dedekind zeta functions, which some other algorithms assume, is not covered.) Day to day nothing changes: people already run the fast randomised Miller–Rabin test, and the unconditional deterministic AKS test (2002) is rarely used in practice because it is slow.

What "working with Claude" involved

Only what the primary sources say. In his announcement, Alpöge writes that he guided the work "to the right situation analogous to that toy setting, and then Claude began to leave me in awe", that near the end "Claude far outpaced me", and gives one example: the length of the chain of theta lifts, about \(\sqrt{X}\), was "Claude's suggested" choice. He writes that he was "out of my element" because he has "little mastery of automorphic forms on higher rank groups", and that some parts are ones "in which my vague intuition worked out due to Claude's technical mastery". He writes that "Claude and I have also formalized the arguments" in Lean, and that the step-by-step timeline in the announcement was one "Claude provided after a thorough search". The checking was separate: the community statement describes the Lean and Mathlib communities rallying to verify it with several independent checkers, and MIT's verification runs were done "with Claude Code" (MIT page). The draft guide to the argument was written by AI systems at Michael Rubinstein's direction (Sources).

All times are UTC. Sources: Sources.