The singular values and vectors of low rank perturbations of large rectangular random matrices (Q444963): Difference between revisions

The \(n \times m\) signal-plus-noise data or measurement matrix formed by stacking the \(m\) samples or measurement of \(n \times 1\) observation vectors alongside each other can be modeled as \( \tilde X = \sum_{i=1}^r \sigma_i u_i v_i^* + X\), where \(u_i\) and \(v_i\) are left and right `signal' column vectors, \(\sigma_i\) are the associated `signal' values and \(X\) is the noise-only matrix of random noises. This model is widely used in signal processing with the aim to determine the signal subspace of a set of vectors \(u_i\) and \(v_i\) that contain signal energy. This is accomplished by computing the singular value decomposition (SVD) of \( \tilde X\) and extracting the \(r\) largest values and the associated singular values of \( \tilde X\). In this paper, under the assumption that \(n\) and \(m\) are large and \(r\) is known, the SVD is used to form estimates of \(\{ \sigma_i \}\), \(\{ u_i \}_{i=1}^r\) and \(\{ v_i \}_{i=1}^r\). A characterization of the relationship between the estimated extreme singular values of \( \tilde X\) and the underlying `signal' singular values \(\sigma_i\) is presented. Results are very general in terms of possible distributions for the noise model \(X\). As a special case, for the Gaussian setting, a new characterization for the right singular vectors, or equivalently, the eigenvectors of \({\tilde X}^* \tilde X \) is provided.