A Sieve-based Estimator for Entropic Optimal Transport (2512.21981v1)

Abstract: The entropically regularized optimal transport problem between probability measures on compact Euclidean subsets can be represented as an information projection with moment inequality constraints. This allows its Fenchel dual to be approximated by a sequence of convex, finite-dimensional problems using sieve methods, enabling tractable estimation of the primal value and dual optimizers from samples. Assuming only continuity of the cost function, I establish almost sure consistency of these estimators. I derive a finite-sample convergence rate for the primal value estimator, showing logarithmic dependence on sieve complexity, and quantify uncertainty for the dual optimal value estimator via matching stochastic bounds involving suprema of centered Gaussian processes. These results provide the first statistical guarantees for sieve-based estimators of entropic optimal transport, extending beyond the empirical Sinkhorn approach.