Matrix Spencer conjecture proved; authors credit GPT-5.6 Sol Pro with 'the heavy-lifting' for the key lemma
Emrullah Akbas and Suvrit Sra posted a proof of the Matrix Spencer conjecture (arXiv 2608.28816, 28 Aug 2026). For symmetric n×n matrices A_1..A_n with operator norm at most 1, one can efficiently find signs x in {±1}^n with ||Σ x_i A_i|| = O(√n). They credit GPT-5.6 Sol (Pro) with finding the full proof of the key hereditary small-ball lemma, after a human suggestion on the route, and with 'almost all the calculations'.
Key facts
- Result: for symmetric n×n matrices with ||A_i|| ≤ 1, an efficiently computable coloring x with ||Σ x_i A_i|| = O(√n), the matrix analogue of Spencer's 'six standard deviations' theorem
- Technique: matrix small-ball estimates with a log-barrier determinantal weight, matrix-weighted Poincaré inequalities and dimension-free constants; the new hereditary small-ball estimate for Gaussian series is of independent interest
- AI statement: 'This work made extensive use of GPT-5.6 Sol (Pro) over a period of months'; the route (log-barriers, Schur complements) was suggested to the model by Sra; 'GPT found a full proof after some failed attempts … we attribute the heavy-lifting for that lemma and for almost all the calculations presented in this writeup to GPT'; GPT also wrote the first rough draft, which the authors polished and verified
- Sra also wrote 'GPT, the Counterexample Machine' (arXiv 2608.29595), a catalogue of 15+ counterexamples he found with GPT Pro over 12 months
- Status: preprint (47 pages), not peer-reviewed or formalised
Science result
- Field
- mathematics / discrepancy theory / random matrices
- Problem
- Matrix Spencer conjecture
- Result
- O(√n) discrepancy for n symmetric n×n matrices of norm at most 1, achievable efficiently.
- AI system
- GPT-5.6 Sol Pro
- Human role
- AI-assisted: humans chose the strategy and verified and polished; the model produced the key lemma's proof and most calculations
- Verification
- Preprint; not peer-reviewed
- Status
- pending
- Why surprising
- A well-known open problem in matrix discrepancy was closed through months of dialogue with a chatbot.
What happened
After months of working with GPT-5.6 Sol Pro, Akbas and Sra obtained the hereditary small-ball lemma that completed their approach to Matrix Spencer. They credit the model with most of the technical work.
Why it matters
Matrix Spencer was one of the headline open problems in discrepancy theory. The paper is candid that the model did the "heavy-lifting".
Changelog
- 2026-09-30: created
Related events
- OpenAI broadly releases GPT-5.6 (Sol, Terra, Luna) after government-gated preview ★★★★
- GPT-5.5 Pro finds counterexample disproving McKean's 1966 conjecture and the Gaussian completely monotone conjecture ★★★
- Summer 2026 flood: dozens of named conjectures settled on arXiv with disclosed AI help (July–September catalogue) ★★★★
Sources (2)
- paperarXiv 2608.28816: A Proof of the Matrix Spencer Conjecture (Akbas, Sra)
- paperarXiv 2608.29595: GPT, the Counterexample Machine (Sra)
id: 2026-08-28-matrix-spencer-conjecture-proved · updated 2026-09-30 · open in the interactive timeline