Wasserstein Multivariate Autoregressive Models for distributional time series and its applications in graph learning

In this work, we propose a new autoregressive model for multivariate distributional time series. We consider a collection of N series of probability measures supported over a bounded interval in R, which are indexed by distinct time instants. Especially, we wish to develop such a model which can identify the dependency structure in the temporal evolution of the measures. To this end, we adopt the Wasserstein metric. We establish the regression model in the Tangent space of the Lebesgue measure by first "centering" all the raw measures so that their Fréchet means turn Lebesgue. The uniqueness and stationarity results are provided. We also propose a consistent estimator for the model coefficient. In addition to the simulated data, the proposed model is illustrated on a real data set of age distributions of countries.