Wasserstein Stability of Contracting Flows: Effective Rates, Euler Self-Correction, and Noise Tightening

What the math gives to ML

The paper provides a distribution-level contraction analysis that replaces a global worst-case contraction rate with a displacement-weighted average of local contraction rates along the evolving coupling. This is directly transferable to flow matching, diffusion ODEs, and neural ODEs, where it can yield tighter Wasserstein stability monitors and adaptive integration or training schedules. It also characterizes a non-monotone Euler error under contraction: discretization error initially grows, reaches a universal peak, and then decays because the dynamics self-correct perturbations. Finally, the noise-tightening result suggests designing nonlinear restoring drifts whose state-dependent contraction is stronger away from equilibrium, providing a principled route to lower stationary sampling variance than a linear drift with the same worst-case rate.