Optimal copula transport: clustering multivariate time series via Wasserstein distance

FrnkNlsn · x · 2026-08-21

A recommendation of the ICASSP 2016 paper "Optimal Copula Transport for Clustering Multivariate Time Series" (arXiv:1509.08144, by Gautier Marti, Frank Nielsen, Philippe Donnat). Copulas fully capture the dependence of multivariate distributions by factorizing the joint density with 1D marginals. The paper uses optimal transport (Wasserstein distance) between copulas to encode both intra-dependence of a multivariate time series and inter-dependence between two series, defining two distances for clustering. The proposed multivariate dependence coefficient is robust to noise, deterministic, and can target specified dependencies.

Original post →

More from Research

Research channel →