Wasserstein convergence of Girsanov-weighted path samples
To sample the law of a diffusion with drift, one can instead sample Brownian paths and give each path a weight that accounts for the drift. How quickly does this weighted sample approximate the diffusion law when distance between paths is measured by their largest separation over time?
Why it matters and what counts as progress
Why it matters: The rate tells us how much a finite sample can reveal about the distribution of whole paths. It also shows whether reweighting changes the convergence rate for Brownian samples.
What counts as progress: Establish matching bounds on the distance for weighted samples under stated conditions on the drift. Determining a limiting constant or extending the conditions would refine the result.
How the problem developed
- 2014 · Wiener measure rate [3]
Emmanuel Boissard and Thibaut Le Gouic Gave expected Wasserstein convergence bounds for empirical measures, including Wiener measure.
Simon Coste with GPT-6 Astra Extended the logarithmic order to Girsanov-weighted Brownian samples for bounded Lipschitz drift.
AI contribution
The Wasserstein cost of the Girsanov Theorem
For an Itô diffusion starting at zero with unit Brownian noise and bounded drift that is globally Lipschitz in space and uniformly Lipschitz in time, Coste reweights independent Brownian paths by their Girsanov likelihoods and normalises the weights to sum to 1. For each fixed , the pth power of the Wasserstein distance from this empirical law to the diffusion law has order in expectation and almost surely, using the supremum norm on paths.
Sources
Publication history
Daily emails by default, only when there’s an update.