abstract: The celebrated Sinkhorn algorithm is used as the go-to algorithm to compute the optimal coupling between two probability measures: it is an alternating dual ascent on the entopic penalization of the Kantorovich problem.
While it has been shown in several ways that it converges exponentially fast to the correct solution (also in the multimarginal case), the proven rate of convergence has a big discrepancy to the real one observed in simulations. We try to close the gap, in the regime of small penalization parameter, under the assumption of measures which satisfy a Poincaré inequality.