Approximate inference for spatial functional data on massively parallel processors

With continually increasing data sizes, the relevance of the big n problem of classical likelihood approaches is greater than ever. The functional mixed-effects model is a well established class of models for analyzing functional data. Spatial functional data in a mixed-effects setting is considered, and so-called operator approximations for doing inference in the resulting models are presented. These approximations embed observations in function space, transferring likelihood calculations to the functional domain. The resulting approximated problems are naturally parallel and can be solved in linear time. An extremely efficient GPU implementation is presented, and the proposed methods are illustrated by conducting a classical statistical analysis of 2D chromatography data consisting of more than 140 million spatially correlated observation points.1