What was already known
Tao asked for explicit matrices comparable to random ones. Bourgain and colleagues had already exceeded exponent one half by using cancellation. Later deterministic bounds improved that exponent by small amounts.
AI involvement
The authors report that GPT-6 Astra produced the proof of Theorem 1 in one attempt. They developed the exposition, extracted the hybrid norm estimate and studied its limitations in later exchanges with models.
- AI and formalization declarationThe authors’ account of GPT-6 Astra’s original proof and their later work
Importance
The construction reaches exponent two thirds by splitting the Gram matrix by entry size and using graph incidence bounds on the larger entries.
Technical verification
Kümmerle and Mixon say they checked all mathematical claims and computations. They also report that Lean checks the mathematical statements of the paper, including the incidence theorem used in the construction.
Formal checking
Kümmerle and Mixon say they checked all mathematical claims and computations. They also report that Lean checks the mathematical statements of the paper, including the incidence theorem used in the construction.
- Checker
- Lean 4 with Mathlib, as reported by the authors
- Scope
- Checked Theorem 1 and its mathematical dependencies. The construction’s running time is outside the formal scope.
- Basis
- Reported by the linked checker
- Lean development for the matrix constructionReported checking and theorem scope
Human understanding
Kümmerle and Mixon explain how a column-sum bound handles small Gram entries and a graph-based bound handles large ones. Their account shows how a vertex selection using Zahl’s incidence theorem produces the needed sparse subgraphs.
- The authors’ explanation of the hybrid estimateHybrid estimate
Other useful sources
Correction and publication history
Curated the reported result and its documented AI contribution.
Read this published revision