High-dimensional Gaussian linear processes: Marchenko-Pastur beyond simultaneous diagonalizability

Except for trivial cases, the eigenvectors of spectral density matrices f () corresponding to stationary Gaussian process depend explicitly on the frequency 0, 2. The most commonly used estimator of the spectral density matrix is the smoothed periodogram, which takes the form of sample covariance matrices YYT for data-matrices Y with iid columns that each have differing underlying covariance structure. However, the covariance matrices of the columns need not be simultaneously diagonalizable and therefore, such sample covariance matrices are out of reach for the current state of random matrix theory. In this paper, we derive a Marchenko-Pastur law in this non-simultaneously diagonalizable case. On the technical level, we make the following two major contributions: - Except for negligible approximations we construct a matrix with the same eigenvalue distribution as the smoothed periodogram, but which is much more accessible. To this aim, we introduce a generalization of graph-theoretical methods specific to Gaussian random matrices, which allow for the exploitation of independence properties in-between either columns or rows. - By means of the Lagrange inversion formula, we draw a novel connection between trace moment expansions and the Marchen\-ko-Pastur equation, which enables us to formulate deviation inequalities for the corresponding l-th moments for all l N. The Marchenko-Pastur law emerges when the dimension d of the process and the smoothing span m of the smoothed periodogram grow at the same rate, which is slower than the number of observations n.