Estimation of structured covariance matrices

Covariance matrices from stationary time series are Toeplitz. Multichannel and multidimensional processes have covariance matrices of block Toeplitz form. In these cases and many other situations, one knows that the actual covariance matrix belongs to a particular subclass of covariance matrices. This paper discusses a method for estimating a covariance matrix of specified structure from vector samples of the random process. The theoretical foundation of the method is to assume that the random process is zero-mean multivariate Gaussian, and to find the maximum-likelihood covariance matrix that has the specified structure. An existence proof is given and the solution is interpreted in terms of a minimum-entropy principle. The necessary gradient conditions that must be satisfied by the maximum-likelihood solution are derived and unique and nonunique analytic solutions for some simple problems are presented. A major contribution of this paper is an iterative algorithm that solves the necessary gradient equations for moderate-sized problems with reasonable computational ease. Theoretical convergence properties of the basic algorithm are investigated and robust modifications discussed. In doing maximum-entropy spectral analysis of a sine wave in white noise from a single vector sample, this new estimation procedure causes no splitting of the spectral line in contrast to the Burg technique.