Abstract
We propose a Standing Wave Decomposition (SWD) approximation to GaussianProcess regression (GP). GP involves a costly matrix inversion operation, whichlimits applicability to large data analysis. For an input space that can beapproximated by a grid and when correlations among data are short-ranged, thekernel matrix inversion can be replaced by analytic diagonalization using theSWD. We show that this approach applies to uni- and multi-dimensional inputdata, extends to include longer-range correlations, and the grid can be in alatent space and used as inducing points. Through simulations, we show that ourapproximate method applied to the squared exponential kernel outperformsexisting methods in predictive accuracy per unit time in the regime where dataare plentiful. Our SWD-GP is recommended for regression analyses where there isa relatively large amount of data and/or there are constraints on computationtime.