Limitations on low rank approximations for covariance matrices of spatial data

An approximation in which observations are split into contiguous blocks and independence across blocks is assumed often provides a much better approximation to the likelihood than a low rank approximation requiring similar memory and calculations.