A universal network strategy for lightspeed computation of entropy-regularized optimal transport

Yong Shi, Lei Zheng, Pei Quan, Yang Xiao, Lingfeng Niu

Abstract: Highlights•A universal network strategy is proposed for fast transport matrix estimation.•It is suitable for arbitrary cost functions and varying marginal distributions.•Network is devised in log domain with the dual form to reduce numerical instability.•Error bound of the resulting algorithm for approximate inputs is analyzed in theory.•The new method is extended to the computation of robust OT and barycenter.