Multimarginal Optimal Transport (MOT) has recently attracted significant interest due to applications in machine learning, statistics, and the sciences. However, in most applications, the success of MOT is severely limited by a lack of efficient algorithms. Indeed, in general, MOT requires exponential time in the number of marginals k and their support sizes n. This paper develops a general theory about "structural properties" that make MOT solvable in poly(n,k) time. We identify two such properties: decomposability of the cost into either (i) local and simple global interactions; or (ii) low-rank and sparse components. These two structures encompass many--if not most--current applications of MOT. In addition to providing the first poly(n,k)-time algorithms for a wide range of MOT problems, our results also provide better algorithms for MOT problems that are already known to be tractable: Our algorithms compute solutions which are exact and sparse. (Previous algorithms can do neither.) We demonstrate our results theoretically and numerically on popular applications in machine learning, statistics, and fluid dynamics.
View on arXiv