acceptodds
Desk rejected at ICLR 2027

Statistical Bias and Effective Dynamics of JKO Schemes with Estimated Ground Geometry

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Can smaller time steps remove errors caused by estimating a transport geometry from finite data? In smooth finite-dimensional local models, optimal-plan response contributes to a bias that can survive exact solves and vanishing steps. A candidate-dependent frozen-plan comparison isolates its endpoint coefficient; the local transport tensor determines its velocity limit. A two-support example proves nonvanishing at fixed entropic regularization. Under independent refreshment, finite-horizon averaging separates drift, fluctuations, and numerical error. In a five-support test, the independently computed coefficient predicts endpoint contrasts within and probability-mass contrasts within at the smallest perturbation. Plan response partially cancels the frozen response. Split-sample correction reduces leading coordinate-mean bias; on two frozen-feature tasks it improves finite-bank reference tracking by – while worsening prediction. Flow selection, accurate solution, and predictive quality are distinct objectives.

Primary area:
optimization
Keywords:
Wasserstein Gradient Flows, JKO Schemes, Stochastic Ground Metrics, Variational Bias, Optimal Transport

Related papers

Discussion (0)

Sign in to comment.