VibeMathedMath problems solved with AI

The Proportion of Zeta Zeros on the Critical Line

The Riemann hypothesis asserts that every nontrivial zero of the zeta function lies on the critical line. Short of proving it, the standard measure of progress is the proportion of zeros known unconditionally to lie there: Selberg established a positive proportion, Levinson reached a third in 1974, Conrey two fifths in 1989, and the record stood at 512\tfrac{5}{12} for zeros that are simple and on the line, and 0.66030.6603 for distinct zeros.

Under the Riemann hypothesis, Montgomery deduced 23\tfrac23 simple from the pair-correlation second moment in 1973. His prime-side evaluation was already unconditional; RH entered only to read the zero side as a positive sum over real ordinates. Goldston and Suriajaya isolated that termwise positivity as the remaining obstacle and asked what would follow if it could be removed.

This removes it, proving unconditionally that at least 23\tfrac23 of zeros are simple and on the line and at least 56\tfrac56 are distinct - 67.25%67.25\ldots\% and 0.836250.83625 with the Montgomery-Taylor window.

Result
Proved(see note)
Status
Partial result
AI contribution
AI-discovered
Method
Argument
Field
Analytic number theory
Posed by
Bernhard Riemann (1859) for the hypothesis; the proportion ladder runs from Hardy and Selberg through Levinson and Conrey
Year posed
Years open
Solved
2026-08-10
Model
Claude (unreleased research version)
Vendor
Anthropic
Collaborators
Jarred Sumner, Levent Alpöge, Ralph Furman, Eric Easley
Verification
Lean-checked, statement unaudited
Publication
Preprint
Significance
68 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

An unconditional record, not a resolution: the Riemann hypothesis is untouched, and Anthropic states it does not expect these techniques to lead to a proof of it. The paper is explicit that these are lower bounds only - the remaining third of the zeros are not shown to be off the line, merely not reached by the certificate.

What it does settle is a question that was posed. Goldston and Suriajaya had reduced Montgomery's conditional 23\tfrac23 to a single obstruction, the termwise positivity that fails for zeros off the line, and asked what would follow without it. Theorem A replaces that positivity with a rank-trace inequality on a finite compression of Weil's Hermitian form, with Sylvester's law of inertia handling off-line pairs; reading the negative index of truncations as a count of off-line pairs is Bombieri's device. The paper also proves the bound sharp for this route: improving on 23\tfrac23 this way would need pair-correlation information beyond Fourier support 1.

What the AI did

Claude was asked to take a real stab at the Riemann hypothesis, with the mathematical choices left to it, and the bound improvement came out as a byproduct of failing at that. It generated and discarded roughly 650 ideas in a first session; in a second it coordinated about 60 subagents which ran some 2,400 shell commands, wrote hundreds of scripts, checked numerically against known zeros and refereed one another. Two subagents developed the key ideas, thirteen fed them, thirty tried and failed, thirteen validated, two drafted the paper. Roughly 31 million output tokens across two Claude Code sessions.

The decisive step was combining the unconditional pair-correlation work of Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh with a 2000 paper of Bombieri, treating the whole function space at once with the quadratic form allowed to be non-diagonal rather than splitting it. Claude also proposed writing the result up, checked 54 arXiv papers for prior art, and recommended that a human number theorist validate it.

Verification

A sorry-free Lean 4 / Mathlib formalization accompanies the paper, with the statements built from Mathlib's own riemannZeta and analyticOrderAt rather than an assumed form of the result. This site audited the sources at tag v1.0: no sorry under Zeta23/, the 27 real ones all in the comparator/Challenge files that state the theorems with placeholder proofs by design; no native_decide; no axiom declarations - the two a plain grep reports sit inside a tactic's docstring.

One finding changes the entry's meaning: the headline theorems are unconditional, not modulo literature inputs. Zeta23.thmA0 takes no hypotheses, because Weil's explicit formula is itself proved in Lean from Mathlib's functional equation, as are Riemann-von Mangoldt, Montgomery-Vaughan, Stirling and Chebyshev-Mertens.

Still held at the unaudited rung, for the reason that tier exists: the trusted statement files are written in the same repository by the same team, so nobody independent has checked the Lean statement against the paper's claim. The repository ships a leanprover/comparator challenge-and-solution split built for exactly that audit - running it is what a promotion would take. The kernel was not re-run here; the axiom lines are the repository's own record.

On human review: two Anthropic mathematicians validated the work, and Brian Conrey and Dan Goldston examined the paper. Examination is not endorsement, and Goldston is an author of the prior work this builds on.

Sources

FrontierStep on Zeros on the critical line · >67.25%> 67.25\% · current best

Submitted by SwiftTapir695 on

Changelog9 changes
  • Rasmus Lindahlchanged What was actually shown from An unconditional record, not a resolution: the Riemann hypothesis is untouched, and Anthro… to An unconditional record, not a resolution: the Riemann hypothesis is untouched, and Anthro…, also Source name, Source URL, Statement, Verification note
  • VibeGenecommented
  • Rasmus Lindahlchanged Age note from Deliberately no single posed-year, so the entry cannot claim an age it has not earned. The… to Deliberately no single posed-year, so the entry cannot claim an age it has not earned. The…
  • sapphirecommented
  • CrimsonLynx347commented
  • Lampcommented
  • Rasmus Lindahlcommented
  • Rasmus Lindahlapproved this entry
  • SwiftTapir695submitted this entry

Discussion4

VibeGene19 Aug 2026, 04:59 UTC

For the "open number of years" case, I think partial results can fairly claim the years since the previous best result was known. Wikipedia says, "In 2020, the [RH percentage of zeros on the critical line] estimate was extended to five-twelfths (i.e., 41.6%) by Pratt, Robles, Zaharescu and Zeindler[43]", so it should be justifiable to say this is an improvement on a six-year problem.

0
CrimsonLynx34711 Aug 2026, 14:32 UTC

By reading the paper, it seems that the improvement is not the proportion in itself. Montgomery proved the same proportion but only for "simple" non-trivial zeros. Claude made it general, that is for all non-trivial zeros.

0
Lamp10 Aug 2026, 19:43 UTC

Wow!

0