A Rapid Method of Computing Empirical Orthogonal Functions from a Large Dataset

Abstract

A rapid method of computing the EOFs for a problem with a large covariance matrix is presented. The resulting EOFs are not the ?true? ones because they do not guarantee that the maximum variance is explained in the dependent sample. However, in an independent sample they turn out to be very efficient when compared with the ?true? EOFs computed from a very large dataset. The differences between different EOF sets do not necessarily indicate essential differences in the corresponding sets of the source data. A rotation for secondary reasons may produce very noticeable differences between the EOFs.