acceptodds
Desk rejected at ICLR 2027

Structural Compatibility Enables Exact Solvability in Wasserstein Proximal Dynamics

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

Abstract

Finite-dimensional Wasserstein updates replace an optimization over probability measures by one over a structured family. We ask when this restriction is mathematically lossless rather than merely convenient. The key property is structural compatibility: on every statistic fiber, the same canonical representative must minimize the energy and saturate the Wasserstein transport lower bound. We show that this compatibility is equivalent to all-step fiberwise exactness. For mean–covariance fibers, this yields full-space finite-step Gaussian and matched -Gaussian closure for Shannon and power entropies under quadratic confinement. Our main converse result is a rigidity theorem: within regular fixed-generator radial location–scatter families, all-step exactness forces the internal energy into the Shannon/positive-power class and fixes the radial generator. When exactness fails, the normal Wasserstein gradient gives the first-order distance to the family in one-dimensional quantile geometry; a strongly convex counterexample shows that vanishing first-order defect still does not imply finite-step closure. The associated -dimensional radial defect is directly estimable. Across 375 frozen Monte Carlo configurations with , every mismatch has positive estimated defect, matched cases are numerically zero, the mean relative error at is , and all six preregistered bootstrap intervals contain the analytic target. Together, the results separate exact finite-dimensional solvability, infinitesimal compatibility, and structurally misspecified approximation.

Primary area:
optimization
Keywords:
Wasserstein Optimization, JKO Schemes, Exact Solvability, Finite-Dimensional Reduction

Related papers

Discussion (0)

Sign in to comment.