On The Use Of Gradient Information In Gaussian Process Quadratures
Jakub Prüher, University of West Bohemia
Simo Särkkä, Aalto University

Gaussian process quadrature is a promising alternative Bayesian approach to numerical integration, which offers attractive advantages over its well-known classical counterparts. We show how Gaussian process quadrature can naturally incorporate gradient information about the integrand. These results are applied for the design of transformation of means and covariances of Gaussian random variables. We theoretically analyze connections between our proposed moment transform and the linearization transform based on Taylor series. Numerical experiments on common sensor network nonlinearities show that adding gradient information improves the resulting estimates.