Gromov-Wasserstein Transfer Operators

Abstract

Gromov-Wasserstein (GW) transport is inherently invariant under isometric transformations of the data. Having this property in mind, we propose to estimate dynamical systems by transfer operators derived from GW transport plans, when merely the initial and final states are known. We focus on entropy regularized GW transport, which allows to utilize the fast Sinkhorn algorithm and a spectral clustering procedure to extract coherent structures. Moreover, the GW framework provides a natural quantitative assessment on the shape-coherence of the extracted structures. We discuss fused and unbalanced variants of GW transport for labelled and noisy data, respectively. Our models are verified by three numerical examples of dynamical systems with governing rotational forces.