VibeMathedMath problems solved with AI

Optimal Exponent Relating Sumsets and Difference Sets

For every finite set AZA\subset\mathbb Z with A2|A|\ge 2, define
C(A)=log(A+A/A)log(AA/A).C(A)=\frac{\log\left(|A+A|/|A|\right)} {\log\left(|A-A|/|A|\right)}.
Determine the largest possible value of C(A)C(A), equivalently the least universal exponent cc such that
A+AA(AAA)c\frac{|A+A|}{|A|} \le \left(\frac{|A-A|}{|A|}\right)^c
for every such set AA. The result proves that the supremum is exactly 22, although no individual admissible set attains it.

Result
Proved(see note)
Status
Resolved
AI contribution
AI-discovered
Method
Argument
Field
Additive combinatorics
Posed by
Year posed
Years open
Solved
2026-07-29
Model
Hy3
Vendor
Tencent Hunyuan
Collaborators
Haowei Lin, Shanda Li
Verification
Lean-verified
Publication
Preprint
Significance
20 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

New arXiv preprint with an author-provided Lean formalization; not yet peer-reviewed.

What the AI did

Tencent Hunyuan’s Hyra research agent, powered by the Hy3 model, was used to explore and optimize finite-set constructions. During an approximately 24-hour run, Hyra produced the construction underlying the paper after moving from finite numerical searches toward natural-language proposals of general constructions and supporting arguments.

The human authors independently checked the construction, corrected and rewrote the exposition, and prepared the final mathematical proof manually. GPT-5.6 Sol was used as an exploration judge and later helped translate the natural-language argument into a Lean 4 formalization. The language-model judgments were not used as proof certificates.

Verification

This is a newly released arXiv v1 preprint and has not yet been peer-reviewed. It contains an explicit, self-contained mathematical construction and proof.

The authors also provide a Lean 4/mathlib formalization. The repository reports that `lake build` completes successfully with no `sorry` declarations or warnings. The principal asymptotic and supremum results use three `native_decide` certificates for elementary finite computations concerning a 12-element base-39 digit block. Consequently, those parts additionally trust Lean’s compiler and native execution, rather than relying exclusively on kernel reduction.

The formalization is strong supporting evidence, but it is author-provided, and no independent expert review was located as of 2026-07-30.

Source

Submitted by matthew on

Changelog4 changes

Discussion2

matthew30 Jul 2026, 09:34 UTC

Indeed it is!

0
Rasmus Lindahl30 Jul 2026, 09:20 UTC

Quite exciting to see a smaller model (295B) make contributions!

0