Distance multivariance: new dependence measures for random vectors (Q2328059)

Two new measures for the dependence of random variables \(X_1,\dots,X_n\) \((n\ge 2)\) are introduced, namely the distance multivariance \(M_{\rho}(X_1,\dots,X_n)\) and the total distance multivariance \(\overline{M}_{\rho}(X_1,\dots,X_n)\). Both measures are defined for random variables with values in \(\mathbb{R}^{d_i}\) for \(d_i\ge 1\) for every \(i=1,\dots,n\) (i.e., eventually, the spaces have arbitrary dimensions). The introduced measures are closely related to distance covariance, as introduced in [\textit{G. J. Székely} et al., ibid. 35, No. 6, 2769--2794 (2007; Zbl 1129.62059); \textit{G. J. Székely} and \textit{M. L. Rizzo}, Ann. Appl. Stat. 3, No. 4, 1236--1265 (2009; Zbl 1196.62077)], and inherit many of its features. Moreover, both measures are based on the weighted \(L^2\)-distance of quantities related to the characteristic functions of the underlying random variables. As a relevant property, \(\overline{M}_{\rho}(X_1,\dots,X_n)=0\) if, and only if, the involved random variables are independent. Thus, that total distance multivariance can be used to detect the independence of \(n\) random variables and, under some mild moment conditions, a test for independence of multiple random vectors has been provided, which is consistent against all alternatives. Various properties of the proposed measures have been discussed, together with their connections with the related literature. For the immediate use of distance multivariance in applications, the R package \texttt{multivariance} is avalaible online.