VibeMathedMath problems solved with AI

Record Lower Bounds for the Shannon Capacity of Odd Cycles

Determine the Shannon capacities of odd cycles beyond C5C_5, or improve the best explicit bounds. Lovasz's theta function settled C5C_5 in 1979 and every longer odd cycle has stayed open since. The current records, all obtained with model assistance and formally verified, are Θ(C7)3.258805369885\Theta(C_7) \ge 3.258805369885, Θ(C11)5.294502522149\Theta(C_{11}) \ge 5.294502522149, Θ(C13)6.302455083464\Theta(C_{13}) \ge 6.302455083464, Θ(C15)7.301600534487\Theta(C_{15}) \ge 7.301600534487, Θ(C19)9.357192705918\Theta(C_{19}) \ge 9.357192705918, Θ(C21)10.342455853338\Theta(C_{21}) \ge 10.342455853338 and Θ(C23)11.328224257774\Theta(C_{23}) \ge 11.328224257774.

Result
Proved(see note)
Status
Partial result
AI contribution
AI-discovered
Method
Construction
Field
Zero-error information theory
Posed by
Claude Shannon
Year posed
1956
Years open
70y
Solved
2026-07-31
Model
ChatGPT 5.6 Sol Pro, ChatGPT 5.6 Sol, Claude Opus 5
Vendor
OpenAI / Anthropic
Collaborators
Pjotr Buys, Sven Polak, Jeroen Zuiddam, Yu Gao, Nathaniel Itty, Christopher D. Rosin, Chase Carstensen, Daniel Reichman
Verification
Lean-verified
Publication
Preprint
Significance
35 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

record lower bounds for seven odd cycles; the exact capacities remain open for every odd cycle beyond C5

What the AI did

Three model-assisted papers in eleven days, each beating the last. Itty, Rosin, Carstensen and Reichman had ChatGPT-5.6 Sol Pro generate and run search programs across repeated prompts, returning explicit independent sets in strong graph powers that the authors checked. Gao then improved C7C_7 with a recursive construction and states that ChatGPT 5.6 Sol implemented all the code and expanded the proofs. Buys, Polak and Zuiddam followed both methods using ChatGPT 5.6 Sol Pro and Claude Opus 5, beat every previous bound, added three more cycles, and formalised the lot in Lean.

Verification

The current records are formalised in Lean 4 at the linked repository, one base tuple per bound, so the seven stated inequalities are machine-checked rather than author-checked. We have not compiled it. Gao's intermediate record ships exact-integer verification code pinned to a fixed commit, and the earlier Itty-Rosin-Carstensen-Reichman constructions came with public data, prompts and checking code. All three are arXiv preprints; none is peer-reviewed.

Sources

FrontierStep on Shannon capacity of C₁₁ · 5.294502522149 · a step on the frontier, since improved

Changelog1 change
  • Rasmus Lindahlchanged Verification from unreviewed to lean-verified, also Statement

Discussion