Relative perturbation bounds with applications to empirical covariance operators (Q2111217)

The authors establish relative perturbation bounds for the empirical covariance operator, which they then use to prove limit theorems, concentration inequalities and inconsistency results. One of the main ingredients in these perturbation bounds is the relative rank operator, measuring the separation of a given eigenvalue from the rest of the spectrum, and defined for a covariance operator \(\Sigma\) with eigenvalues \(\lambda_1,\lambda_2,\ldots\) by \[ r_j(\Sigma)=\sum_{k\not=j}\frac{\lambda_k}{|\lambda_j-\lambda_k|}+\frac{\lambda_j}{g_j}\,, \] where \(g_j\) is the distance of \(\lambda_j\) from the rest of the spectrum of \(\Sigma\). Under certain boundedness conditions on this relative rank, the authors establish perturbation expansions for empirical eigenvalues, eigenvectors and spectral projectors. These expansions are then used to establish limit theorems, including versions of the law of large numbers and central limit theorem for empirical eigenvalues under mild moment conditions. The authors similarly establish high-probability concentration bounds, and also investigate inconsistency issues, including a phase transition between (relative) consistency and inconsistency of the leading empirical eigenvalue, with analogous results for spectral projectors. These results are illustrated with various examples and applications.