James Maynard, Small gaps between primes
The multidimensional sieve, its first- and second-moment estimates, the variational quantity Mk, and the deduction of H1 ≤ 600.
Axiom Math · Lean 4
A unified formalization blueprint
A Lean 4 account of the proof that infinitely many consecutive prime gaps are at most 246. The argument combines Maynard’s multidimensional sieve with the parts of Polymath8b needed to reach the sharper bound.
Write pn for the nth prime. The quantity H1 = lim inf (pn+1 − pn) measures how close consecutive primes come infinitely often. Thus H1 ≤ 246 means that gaps of size at most 246 occur infinitely many times.
Maynard’s multidimensional sieve gave the bound 600. Polymath8b brought it down to 246 by enlarging the simplex in the variational problem and supplying an explicit finite certificate.
The interactive blueprint breaks the argument into definitions, lemmas, and theorems. It shows what each result depends on and points to the corresponding Lean declaration.
PrimeGapsLib
contains both sides of the proof: the general sieve theory and the
explicit 50-dimensional certificate. The certificate is checked in
Lean and combined with the theory in PrimeGaps.Bounded246,
which proves that Bombieri–Vinogradov implies prime gaps bounded by
246 infinitely often.
The proof draws on two papers. We formalize the Maynard argument in detail, but only the portion of Polymath8b used for the bound 246.
The multidimensional sieve, its first- and second-moment estimates, the variational quantity Mk, and the deduction of H1 ≤ 600.
We use the enlarged-simplex refinement of Theorem 3.12(i), the certificate argument from Theorem 3.13(i) and §6.2, and the 50-tuple endgame. The other results in Polymath8b are not part of this project.
We first wrote the proof as a blueprint, with a separate entry for each definition and lemma. The dependency graph then gave us an order in which to formalize it.
Give every definition, lemma, and theorem a label, a precise statement, and a list of the earlier results it uses.
Arrange the graph in layers. By the time we reach a node, all of its dependencies have already been formalized.
AxiomProver is Axiom Math’s AI system for mathematical research through formal proof. We gave it one layer at a time, together with the completed dependencies. It produced machine-checkable Lean 4 proofs built on Mathlib.
Read the proof node by node and follow its dependencies into Lean.
Download the complete blueprint as a PDF.
The public Lean library, including the theory and the 246 certificate.
The project discussion link will be added when its Zulip topic opens.
* Mathematical contributors † Engineering contributors