Proof Perplexity: A Probabilistic Approach to Measuring non-triviality in Machine Generated Proofs
It is hard to measure whether a machine-generated proof is a genuinely novel, non-trivial argument or a recombination of results already in the literature.
Proof steps are modeled as Lean declarations whose kernel-checked dependency closures within Mathlib form a DAG, which is unfolded into internal complexity trees. Each step receives a recoverability score from two parts: ablation recoverability, based on regenerating the step under a search budget, and a Mahalanobis-distance outlier p-value against clusters of Mathlib declarations. These scores give per-step surprisal, a weighted proof perplexity, and a theta-invention fraction with a provable lower bound.
Applied to Claude's Lean 4 formalized proof on zeros of the Riemann zeta function, the method gives a proof perplexity of 388.23 (cap 501). The theta-invention fraction is 0.9578, which satisfies the derived bound of 0.9446, and 2032 of 2399 steps fall at or below the threshold.
| Metric | Value |
|---|---|
| Proof perplexity PP_w(π,B) | 388.23 (cap: 501) |
| θ-invention fraction β(θ,B) | 0.9578 (bound: 0.9446, holds) |
| Steps with α ≤ θ / total | 2032 / 2399 |
