VibeMathedMath problems solved by AI

Dittert's Conjecture in Dimension Five

The dimension-five case asks whether, for every nonnegative 5×55\times5 real matrix AA whose entries sum to 55, the Dittert functional Φ(A)=iri+jcjper(A)\Phi(A)=\prod_i r_i+\prod_j c_j-\operatorname{per}(A) is uniquely maximized at U5=J5/5U_5=J_5/5. The submitted artifact claims the stronger quantitative bound Φ(A)12266251625AU5F2,\Phi(A)\leq \frac{1226}{625}-\frac{1}{625}\lVert A-U_5\rVert_F^2, which implies uniqueness.

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI-discovered
Method
Computation
Field
Linear algebra; permanents; exact sum-of-squares certificates
Posed by
Eberhard Dittert
Year posed
1983
Years open
43y
Solved
2026-08-06
Model
GPT-5.6 Sol (Ultra)
Vendor
OpenAI
Collaborators
Arthur Moisés da Costa Borges
Verification
Lean-checked, statement unaudited
Publication
Announced
Significance
15 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

Dimension 5 only; public AI-generated candidate with no independent specialist review.

What the AI did

Operating through OpenAI Codex, GPT-5.6 Sol with the Ultra reasoning-effort setting selected the problem after literature triage, developed the symmetry-reduced sum-of-squares approach, ran numerical discovery and rational recovery, produced the exact certificate and mechanically separate verifier, formalized the quantitative bound and equality characterization in Lean 4, audited the artifacts, and wrote the manuscript. Human mathematical supervision was minimal. Arthur Moisés da Costa Borges defined the broad objective, authorized execution and publication decisions, supplied factual metadata, and maintains the artifact, but did not derive or independently validate its technical content.

Verification

The public artifact contains a Lean 4.30.0-rc1 formalization of the n=5 quantitative bound and equality characterization, with no sorry, admit, or user-declared axioms. Lean checks the included exact rational SOS witness directly. Large finite equalities use native_decide; the trusted base therefore includes Lean's native compiler and runtime, not the kernel alone. A separate Python/FLINT verifier checks 54/54 orbital identities, 425/425 kernel constraints, and 420/420 positive leading principal minors; deterministic generators reproduce the PSD witness and 41 Lean data modules. No independent specialist has yet checked the informal-to-formal correspondence, historical or novelty claims, or the overall argument. Treat this as a public AI-generated candidate, not an established or peer-reviewed result. Tier: the same system produced both the proof and its Lean formalization, and no independent party has audited the informal-to-formal correspondence.

Source

Public GitHub research artifact (commit 91920c4)

Submitted by LuckyMongoose479 on

Changelog2 changes
  • Rasmus Lindahlapproved this entry
  • LuckyMongoose479submitted this entry

Discussion