LPOT: Locality-Preserving Gromov–Wasserstein Discrepancy for Nonrigid Point Set Registration

Abstract

The Gromov-Wasserstein (GW) distance, rooted in optimal transport (OT) theory, provides a natural framework for aligning heterogeneous datasets. Alas, statistical estimation of the GW distance suffers from the curse of dimensionality and its exact computation is NP hard. To circumvent these issues, entropic regularization has emerged as a remedy that enables parametric estimation rates via plug-in and efficient computation using Sinkhorn iterations. Motivated by further scaling up entropic GW (EGW) alignment methods to data dimensions and sample sizes that appear in modern machine learning applications, we propose a novel neural estimation approach. Our estimator parametrizes a minimax semi-dual representation of the EGW distance by a neural network, approximates expectations by sample means, and optimizes the resulting empirical objective over parameter space. We establish non-asymptotic error bounds on the EGW neural estimator of the alignment cost and optimal plan. Our bounds characterize the effective error in terms of neural network (NN) size and the number of samples, revealing optimal scaling laws that guarantee parametric convergence. The bounds hold for compactly supported distributions, and imply that the proposed estimator is minimax-rate optimal over that class. Numerical experiments validating our theory are also provided.