VibeMathedMath problems solved with AI

Euler's 1747 conjecture on sums of two integers with all prime factors congruent to 1 modulo 4

In a letter to Goldbach of 6 May 1747, Euler proposed that every number of the form 4m+24m+2 is a sum of two numbers of the form 4x+14x+1, neither of which has a divisor of the form 4p−14p-1. In modern terms: can every positive integer n≡2(mod4)n\equiv 2\pmod 4 be written as n=a+bn=a+b with a,ba,b positive integers all of whose prime factors are congruent to 11 modulo 44 (the integer 11 is allowed)? Euler saw this as a possible route to the four squares theorem. Dietmann and Elsholtz (2019) proved it for all sufficiently large nn; the statement for every nn remained open.

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI-discovered
Method
Argument
Field
Additive number theory
Posed by
Leonhard Euler, letter to Christian Goldbach, 6 May 1747
Year posed
1747
Years open
279y
Solved
2026-11-12
Model
OpenAI Codex (model version not stated)
Vendor
OpenAI
Collaborators
Will Howe, Grant Huffman, Jensen Kohlmeyer, Liam Kruer
Verification
Unreviewed
Publication
Preprint
Significance
22 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Theorem 1.1: every positive n≡2(mod4)n\equiv2\pmod4 is a+ba+b with all prime factors of aa and bb congruent to 11 modulo 44. After dividing out the primes ≡1(mod4)\equiv1\pmod 4 from n/2n/2, the proof counts representations of nn as a sum of four squares and bounds those in which some prime p≡3(mod4)p\equiv3\pmod4 divides both coordinates of a pair. The key estimate (Proposition 3.1) comes from a Hecke operator on a quotient of PGL2(Fp)\mathrm{PGL}_2(\mathbb{F}_p) by the Hurwitz units, the Jacquet-Langlands correspondence and Deligne's bound, averaged over the normalizer of a nonsplit torus so that the error term does not grow with pp. Explicit constants give the theorem for n≥644=16,777,216n\ge 64^4=16{,}777{,}216; the remaining 4,194,304 values are covered by a finite sieve certificate in which the smaller summand never exceeds 313.

What the AI did

From the paper's acknowledgment: "OpenAI Codex generated the mathematical argument, performed the exact calculations, and drafted the manuscript under human direction." The Hexagon provenance statement adds: "Some parts of the contribution have been checked and understood by a human." Hexagon labels the work "Primarily AI-generated text" and "Human understanding: some parts". The model behind Codex is not named.

Verification

No independent mathematician has checked this yet. Checked here on 19 November 2026: the Hexagon paper 2611.00370v1 (8 pages; Will Howe, Grant Huffman, Jensen Kohlmeyer and Liam Kruer, Purdue addresses) read in full, with its AI acknowledgment and the Hexagon provenance record. Dietmann and Elsholtz (Banach Center Publ. 118, 2019) was opened: it attributes the question to Euler's letter and proves it for sufficiently large nn only. Euler's letter itself was not opened. The finite range was recomputed here with an independent full sieve: every n≡2(mod4)n\equiv2\pmod4 below 64464^4 has a decomposition with smaller summand at most 313, and the count 1,344,661 and all twelve entries of the paper's Table 1 were reproduced exactly. The closing constant 21235493/33177600<2/321235493/33177600<2/3 and the two constants of Lemma 5.1 were confirmed in exact arithmetic. Proposition 3.1 was tested by brute force for 24 values of mm up to 4389 and every prime p≡3(mod4)p\equiv 3\pmod 4, and held each time, with equality in at least one case. The spectral step behind it (Lemma 3.2: the Hecke identification, Jacquet-Langlands and Deligne) was read but not verified, and the range n≥644n\ge 64^4 depends on it; hence Unreviewed. An arXiv search found no earlier full proof.

Sources

Submitted by MellowGander562 on

Changelog2 changes

Discussion