Paperlane

no. 8 · Wednesday 4 November 2026

Path-Kernel Method for Differentiating Unstable Diffusions

Angxiu Ni

math.PR · v3 · 32 pp

Abstract

We derive and rigorously prove the path-kernel formula for the linear response of stochastic differential equations, namely, the derivative of averaged statistics with respect to a model parameter. The parameter may affect the diffusion coefficient, as well as the drift coefficient and initial condition. No hyperbolicity assumption is required. The proof introduces a single change of variables for the state path that transfers part of the unstable path perturbation to differentiation of the probability kernel. We also derive a pathwise Monte Carlo algorithm for estimating the response and demonstrate it on the 40-dimensional noisy Lorenz–96 system.

comments
32 pages, 7 figures. Accepted for publication in Statistics and Computing
fields
Probability, Dynamical Systems, Numerical Analysis
pdf ↗ page arXiv:2503.00718 ↗