A formally verified convergence theorem for the tug-of-war scheme of the normalized $p$-Laplacian
p-harmonious functions solve the dynamic programming principle of tug-of-war with noise. As the step size goes to zero they converge to the viscosity solution of the Dirichlet problem for the normalized p-Laplacian, 2 ≤ p < ∞. The goal is a complete machine-checked proof of this convergence theorem.
The proof follows the Barles–Souganidis argument. It is formalized in Lean 4 on Mathlib, for domains satisfying the exterior sphere condition. The development has four layers: the discrete scheme, the viscosity framework, the asymptotic expansion of the scheme with an explicit error rate, and relaxed limits with comparison, boundary estimates and convergence. Aleksandrov's theorem and Jensen's lemma, which were not available in Mathlib, are also formalized.
The development comprises about 29,400 lines in 225 files. The main theorem depends only on propext, Classical.choice and Quot.sound. It also includes existence and uniqueness for the discrete problem, the comparison principle, and uniqueness of viscosity solutions.
| Layer | Content | Lines (approx.) |
|---|---|---|
| 1 | scheme: operator, measurability, existence, uniqueness, comparison | 3300 |
| 2 | viscosity framework | 1100 |
| 3 | asymptotic expansion of T_ε | 5500 |
| 4 | relaxed limits, comparison, boundary, convergence | 19100 |
| total (225 files) | 29400 |
