Asymptotic linear spectral statistics for spiked Hermitian random matrices (Q2516098): Difference between revisions

This paper deals with linear spectral statistics of Wishart and \(F\) matrices, i.e., sums of functions of eigenvalues of such matrices. Specifically, it is assumed that the underlying population covariance matrix have one ``spike'' eigenvalue strictly larger than one (the other eigenvalues being equal to one). The asymptotic regime considered is that the ratio of sample size and data dimension tend to a strictly positive limit. The main results are CLTs for the linear spectral statistics with explicit integral representations of the limit mean and variance. In particular, one has explicit expressions for the effect of the spike on the mean of the limit. The paper discusses three applications to statistics problems: a likelihood ratio test that the population covariance matrix is the identity, under a rank-one spiked alternative; the capacity analysis of multi-antenna wireless communication systems with a line-of-sight transmission path; and a multiple sample significance testing problem. The asymptotic analysis is based on contour integral representations of the joint eigenvalue probability density functions (PDF). In the case of a spiked multivariate \(F\) distribution, the integral representation that is used in the paper is new: If \(\mathbf{F} = \mathbf{W}_1 \mathbf{W}_2^{-1}\), where \(\mathbf{W}_1\) is a complex Wishart matrix of size \(n\) with parameters \(m_1 > n\), \(\Sigma\) and rank-one noncentrality matrix \(\Theta\) with spike eigenvalue \(n \nu\) for \(\nu \geq 0\), and \(\mathbf{W}_2\) has parameters \(m_2\), \(\Sigma\), and \(\Theta\) is the null matrix, and if \(x_j\) denote the eigenvalues of \(\mathbf{F}\) and \(\mathbf{f}_j = x_j/(1+ x_j)\), then the joint (PDF) of the \(\mathbf{f}_j\) is up to a constant equal to \[ \oint_C\;_1F_1(m_1 + m_2 - n + 1, m_1 - n + 1, n \nu z) \prod_{j=1}^n \frac{\mathbf{f}_j^{m_1 - n} (1 - \mathbf{f}_j)^{m_2 - n}}{z - \mathbf{f}_j} \prod_{1 \leq j < k \leq n} (\mathbf{f}_k - \mathbf{f}_j)^2 dz, \] where the contour \(C\) encloses the \(\mathbf{f}_k\) counter-clockwise in its interior.