Abstract
We devise a theoretical framework and a numerical method to infertrajectories of a stochastic process from snapshots of its temporal marginals.This problem arises in the analysis of single cell RNA-sequencing data, whichprovide high dimensional measurements of cell states but cannot track thetrajectories of the cells over time. We prove that for a class of stochasticprocesses it is possible to recover the ground truth trajectories from limitedsamples of the temporal marginals at each time-point, and provide an efficientalgorithm to do so in practice. The method we develop, Global Waddington-OT(gWOT), boils down to a smooth convex optimization problem posed globally overall time-points involving entropy-regularized optimal transport. We demonstratethat this problem can be solved efficiently in practice and yields goodreconstructions, as we show on several synthetic and real datasets.