A framework for measuring association of random vectors via collapsed random variables (Q2001082)

A quite general approach to measuring association of random vectors via collapsed random variables is proposed. Let's consider two random vectors, a \(p\)-dimensional \(X=(X_1,\dots,X_p)\) and a \(q\)-dimensional \(Y=(Y_1,\dots,Y_q)\), both with continuous marginal distribution functions. Collapsing function maps random vector \(X\) on a single random variable \(S_X(X)\). The following functions are the examples: \(S(X)=\sum_{i=1}^p w_i X_i\), weighted sum; \(S(X)=\max_{1\le i \le p} X_i\), maximum; \(S(X)=\min_{1\le i \le p} X_i\), minimum; \(S(X)=F_X(X)\), probability integral transform. A \(2p\)-variate collapsing function \(S_X(X,X')\), where \(X'\) is an independent copy of \(X\), is also considered, e.g., \(S_X(X,X')=I(X\le X')\), multivariate rank, where the inequality \(X \le X'\) is understood component-wise. Collapsing functions \(S_Y(Y)\) and \(S_Y(Y,Y')\) for \(Y\) are defined in a similar way. As a measure of association, Pearson's correlation coefficient \(\rho(S_X(X),S_Y(Y))\) is mainly considered. Notice that Kendall's tau and Spearman's rho, which are well known measures of association, are described by the proposed collapsing functions. The properties of the proposed collapsing functions are investigated, in particular, a multivariate extension of the Kendall distribution and corresponding Kendall copula are derived. For the measure of association of the large sample \((X,Y)_1,\dots,(X,Y)_n\) with \(n \to \infty\) the asymptotic normality is established. Applications to bioinformatics and finance are presented.