Relationships among classes of spherical matrix distributions (Q1062712)

Let \({\mathcal O}(n)\) be the set of \(n\times n\) orthogonal matrices and \(''x^{\underset {=} d}y\) denotes that x and y have the same distribution. Let \(X=(x_ 1,...,x_ p)\) be an \(n\times p\) random matrix, \(F_ 1=\{X: PX^{\underset {=} d}X\) for every \(P\in {\mathcal O}(n)\}\), \(F_ 2=\{X:(P_ 1x_ 1,...,P_ px_ p)^{\underset {=} d}(x_ 1,...,x_ p)\) for every \(P_ i\in {\mathcal O}(n),i=1,...,p\}\). \(F_ 3=\{X: P(vec X)^{\underset {=} d}vec Xfor\) every \(P\in {\mathcal O}(np)\}\) and \(F_ s=\{X: PXQ^{\underset {=} d}Xfor\) every \(P\in {\mathcal O}(n)\) and \(Q\in {\mathcal O}(p)\}\). A distribution of X belonging to \(F_ 1,F_ 2,F_ 3\) or \(F_ s\) is called a spherical matrix distribution. In this paper the relationships among the above classes are discussed in the following aspects: the characteristic functions and the stochastic representations in these classes, coordinate transformations, marginal distributions, marginal densities, and sphericity. Some sufficient conditions that the distribution of a statistic is invariant in these classes are given. The invariance of nine useful statistics (such as the Wilks statistics, the Hotelling \(T^ 2\)- statistic, the canonical correlation coefficient, etc.) in the different classes is listed for applications.