Reviews: Screening Sinkhorn Algorithm for Regularized Optimal Transport

This paper proposes a reformulation of the dual of the entropy regularized wassserstein distance problem that is amenable to screening techniques. Such techniques allow to reduce the dimension of optimization problems hence reducing the computational costs. Here static screening rule is proposed, meaning that the variables are screened before running the solver which is here an L-BFGS-B quasi-Newton method. Two screening techniques are proposed, either using a fixed threshold or a fixed budget. The latter appearing easier to use. A theorem quantifying the error on the orignal problem induced by approximation and screening is provided.