The Wasserstein cost of the Girsanov Theorem

Simon Coste

For a diffusion with bounded Lipschitz drift, Coste reweights independent Brownian paths by their Girsanov likelihoods and normalises the weights to sum to 1. For each fixed p≥1p \geq 1, the pth power of the Wasserstein distance from this empirical law to the diffusion law has order (log⁡n)−p/2(\log n)^{-p/2} in expectation and almost surely, using the supremum norm on paths.

What was already known

Earlier quantization estimates and empirical-measure bounds give the logarithmic order for Wiener measure. Coste adapts these arguments to the weighted diffusion law.

AI involvement AI-centered

Coste started the project. He credits GPT-6 Astra with extracting the relevant arguments and providing the steps that adapt them to the Girsanov setting. He wrote the main text, and Grok supplied simulation code.

Importance Notable

The change of measure preserves the logarithmic Wasserstein rate for the stated class of drifts.

Technical verification Formal (Lean 4)

Coste reports Lean 4 formalization of the results using MathFin for stochastic calculus and model audits of correspondence with the written statements.

Formal checking

Coste reports Lean 4 formalization of the results using MathFin for stochastic calculus and model audits of correspondence with the written statements.

Checker
Lean 4 in Coste’s formalization
Scope
For a diffusion with bounded Lipschitz drift, Coste reweights independent Brownian paths by their Girsanov likelihoods and normalises the weights to sum to 1. For each fixed p≥1p \geq 1, the pth power of the Wasserstein distance from this empirical law to the diffusion law has order (log⁡n)−p/2(\log n)^{-p/2} in expectation and almost surely, using the supremum norm on paths.
Basis
Reported by the linked checker

Human understanding Explained

Coste explains how the proof adapts Brownian quantization estimates to the normalised weighted empirical measure.

Correction and publication history