Testing in high-dimensional spiked models (Q2196218)

In multivariate statistical analysis, it is well-known that, if the \(p\times p\) matrix-variate random variables \(A\) and \(B\) are independently gamma distributed with the same scale parameter matrix, then \(A+B\) is gamma distributed and the ratios \(U=(A+B)^{-1/2}A(A+B)^{-1/2}\) and \(V=B^{-1/2}AB^{-1/2}\) are, respectively, \(p\times p\) matrix-variate type-1 and type-2 beta distributed random variables. A large number of likelihood ratio criteria for testing hypotheses on the parameters of \(p\)-variate Gaussian populations are based on the eigenvalues of \(U\) and \(V\). The eigenvalues of \(U\) are also the same as those of \((A+B)^{-1}A\) and the eigenvalues of \(V\) are the same as the eigenvalues of \(B^{-1}A\). Wishart matrices are special cases of gamma distributed matrices. Then, a study of the eigenvalues of \(W=C^{-1}A\), equivalently, the study of the determinantal equation \(|A-\lambda C|=0\), where \(C\) and \(A\) are Wishart matrices or at least \(A\) is a Wishart matrix, is very important in multivariate statistical analysis. \textit{A. T. James} [Ann. Math. Stat. 35, 475--501 (1964; Zbl 0121.36605)] categorized the study of the eigenvalue problem connected with \(W\) into five different classes of testing statistical hypotheses when the population is \(p\)-variate real Gaussian. These five classes are shown to be connected to the hypergeometric functions of matrix argument \({_0F_0}\), \({_1F_0}\), \({_0F_1}\), \({_1F_1}\), \({_2F_1}\). \vskip.2cm In the present paper, the authors re-examine these five classes and do an asymptotic analysis of the log-likelihood function when the dimensionality parameter \(p\) and the sample sizes are proportionately large. Spiked model alternative hypothesis is taken in the following sense: If the null hypothesis is that a parameter matrix \(M=O\) (null matrix), then the alternative is taken as \(M=\theta \mu\mu'\) where \(\theta\) is a real scalar parameter and \(\mu\) is an unknown \(p\times 1\) vector, \(\mu'\) its transpose, so that \(\mu\mu'\) is a \(p\times p\) matrix of rank 1. Thus, the alternative is a rank one matrix or a single-spiked situation characterized by the real scalar parameter \(\theta\). Then the null distribution of the test statistic is associated with \(\theta=0\) situation and the non-null distribution corresponds to nonzero \(\theta\). It is shown that in single-spiked models the theory of scalar variable Mellin and Laplace transforms are applicable. Several asymptotic results are also obtained.