What was already known
Majdoddin bounded a class of grid recursions that rewrites local constraints and merges terms built as products. He identified decompositions with correlated terms as a case outside that argument.
AI involvement
Vaibhav Tulsyan says Claude drafted the mathematical arguments and Lean proofs. Tulsyan chose the problems, directed the work and reviewed the theorem statements and the scope of the checks.
- Tulsyan’s research statement and AI disclosureProvenance statement
- Tulsyan’s manuscript, Theorem 4.11
Importance
The bound applies to arbitrary summands of this grid tensor. A single contraction gives a rank estimate for both cuts, even when the terms do not factor as products.
Technical verification
Tulsyan reports Lean kernel checking for the rank inequality in Theorem 4.11. The pinned declaration covers arbitrary finite sums over a field and a nonempty finite alphabet. The separate step from matrix ranks to the cost of a network is given on paper.
Formal checking
Tulsyan reports Lean kernel checking for the rank inequality in Theorem 4.11. The pinned declaration covers arbitrary finite sums over a field and a nonempty finite alphabet. The separate step from matrix ranks to the cost of a network is given on paper.
- Checker
- Lean 4 kernel, reported by Vaibhav Tulsyan
- Scope
- Theorem 4.11 rank inequality for the grid
- Basis
- Reported by the linked checker
- Lean declaration for the rank boundGrid33.rankSum_ge
Human understanding
Tulsyan reports reviewing the statements and the scope of verification. The source credits Claude with drafting the contraction argument and does not identify a human explanation of it.
Correction and publication history
Source-reviewed lower bound on the smaller-cut rank sum, with central Claude work and a scoped Lean check.
Read this published revision