Separable correlation and maximum likelihood

Abstract

We consider estimation of the covariance matrix of a multivariate normal distribution when the correlation matrix is separable in the sense that it factors as a Kronecker product of two smaller matrices. A computationally convenient coordinate descent-type algorithm is developed for maximum likelihood estimation. Simulations indicate our method often gives smaller estimation error than some common alternatives when correlation is separable, and that correctly sized tests for correlation separability can be obtained using a parametric bootstrap. Using dissolved oxygen data from the Upper Mississippi River, we illustrate how our model can lead to interesting scientific findings that may be missed when using competing models.