The Kato-Temple inequality and eigenvalue concentration

In this paper we use the setting of inhomogeneous random graphs to present an extension of the Kato-Temple inequality for bounding eigenvalues. We prove under mild assumptions that the top eigenvalues (i.e. largest in magnitude and ordered) of the adjacency matrix concentrate around the corresponding top eigenvalues of the edge probability matrix with high probability, even when the latter have multiplicity. The proof techniques we employ do not strictly depend upon the underlying random graph setting and are readily extended to yield high-probability bounds for perturbations of singular values of rectangular matrices in a quite general random matrix setting.