What was already known
De Loera, Haddock and Rademacher showed that the minimum-norm insertion rule can take exponentially many iterations. They left open the worst case for Wolfe’s original linopt rule.
AI involvement
Disser and Mosis report that the central link between Wolfe’s method and the simplex method emerged in their discussions with Claude Opus 5.5. They developed and rewrote the proofs with model assistance.
- Disser and Mosis’s reduction from simplex to Wolfe iterationsAI disclosure and provenance
Importance
The reduction transfers lower bounds for multiple simplex rules and settles the exponential worst case for linopt. It also yields an exponential example for the Lawson–Hanson algorithm.
Technical verification
The authors report verifying all statements and proofs. They also report Lean 4 checking of the main reduction, the first four corollaries and their numbered supporting statements.
Formal checking
The authors report verifying all statements and proofs. They also report Lean 4 checking of the main reduction, the first four corollaries and their numbered supporting statements.
- Checker
- Yann Disser and Nils Mosis, with Lean 4 for the specified statements
- Scope
- Disser and Mosis give a polynomial-time reduction from bounded, nondegenerate linear programs with variables and equality constraints to minimum-norm-point instances on integer points in dimension . The reduction matches sequences of simplex pivots with sequences of Wolfe insertions. It transfers an exponential lower bound from Dantzig’s pivot rule to Wolfe’s original linopt rule.
- Basis
- Reported by the linked checker
- Lean 4 verification archiveProof and documented checking scope
Human understanding
Disser and Mosis explain why feasible simplex bases give Wolfe active sets and how the correspondence preserves the order of additions and removals.
Correction and publication history
Published a scoped research advance with primary evidence of central AI work.
Read this published revision