1: \begin{abstract}

2: We introduce a principled way of computing the Wasserstein distance between two distributions in a federated manner.

3: Namely, we show how to estimate the Wasserstein distance between two samples stored and

4:  kept on different devices/clients whilst a central entity/server orchestrates the computations 

5:  (again, without having access to the samples). To achieve this feat, we take advantage of the geometric 

6: properties of the Wasserstein distance -- in particular, the triangle inequality -- 

7:  and that of the associated {\em geodesics}: our algorithm, \ours (for Federated Wasserstein Distance), iteratively approximates 

8:  the Wasserstein distance by manipulating and exchanging distributions from the

9:   space of geodesics in lieu of the input samples. 

10:   In addition to establishing the convergence properties of \ours,

11:    we provide empirical results on federated coresets and federate 

12:    optimal transport dataset distance, that we respectively exploit for

13:    building a novel federated model and for boosting performance of popular federated learning algorithms.

14: \end{abstract}

15: