A Wasserstein index of dependence for random measures

Nonparametric latent structure models provide flexible inference on distinct, yet related, groups of observations. Each component of a vector of $d \ge 2$ random measures models the distribution of a group of exchangeable observations, while their dependence structure regulates the borrowing of information across different groups. Recent work has quantified the dependence between random measures in terms of Wasserstein distance from the maximally dependent scenario when $d=2$. By solving an intriguing max-min problem we are now able to define a Wasserstein index of dependence $I_\mathcal{W}$ with the following properties: (i) it simultaneously quantifies the dependence of $d \ge 2$ random measures; (ii) it takes values in [0,1]; (iii) it attains the extreme values $\{0,1\}$ under independence and complete dependence, respectively; (iv) since it is defined in terms of the underlying L\'evy measures, it is possible to evaluate it numerically in many Bayesian nonparametric models for partially exchangeable data.