MathCompressionBench

Worked examples

The a + b = b + a example is on the home page →

Mathlib197: a new construction and three general lemmas

The matrix biproduct is rebuilt from identity-matrix entries, with no case splits or index transport. Three general lemmas: ordered-partition sums for two and three, for the Faà di Bruno cases, and induction on the free commutative ring, for Ring.DirectLimit.exists_of.

These changes in Lean, before and after
The run, and the declarations that changed most

A multi-step compression, for demonstration only, not ranked. Its new route: one parameter bound for all rational formulas (division by zero is zero) covers division-free ones too; derivations bound the dimension.

The route changes in Lean, before and after

Ramsey96: bounds re-derived from a sharper theorem

Its k log k bounds on the triangle Ramsey number follow in a few lines from a sharper theorem already in the file.

This change in Lean, before and after
The run, and the declarations that changed most

Compactness: small lemmas

Three brute-force case analyses became short arguments through small new lemmas, cycle_map, kTemplateRelation_color_ne and retained_indicator_mul, measured only together with other edits.

These changes in Lean, before and after
The run, and the declarations that changed most

All seven tasks at a glance

A range spans the two measured orders.

Toy targets one at a time or together



  

The fundamental theorem of arithmetic

A multi-step compression, for demonstration only, not ranked.

The agent's runs and checks

The agent's final: its terms and its Lean source
The natural original's largest auto-generated terms
The statements (the same in every version)
Every declaration: natural original and agent final
The hand compression: the version written long on purpose and its compressed form

It is long on purpose, with repeated re-derivations and automation instead of library lemmas.

Whole file difference