Equispaced Fourier representations for efficient Gaussian process regression from a billion data points

arXiv2210.10210MaRDI QIDQ6414386

Author name not available (Why is that?)

Publication date: 18 October 2022

Abstract: We introduce a Fourier-based fast algorithm for Gaussian process regression in low dimensions. It approximates a translationally-invariant covariance kernel by complex exponentials on an equispaced Cartesian frequency grid of $M$ nodes. This results in a weight-space $MimesM$ system matrix with Toeplitz structure, which can thus be applied to a vector in $mathcalO(MlogM)$ operations via the fast Fourier transform (FFT), independent of the number of data points $N$. The linear system can be set up in $mathcalO(N+MlogM)$ operations using nonuniform FFTs. This enables efficient massive-scale regression via an iterative solver, even for kernels with fat-tailed spectral densities (large $M$). We provide bounds on both kernel approximation and posterior mean errors. Numerical experiments for squared-exponential and Mat'ern kernels in one, two and three dimensions often show 1-2 orders of magnitude acceleration over state-of-the-art rank-structured solvers at comparable accuracy. Our method allows 2D Mat'ern-$$frac{3}{2}$$ regression from $N=10^9$ data points to be performed in 2 minutes on a standard desktop, with posterior mean accuracy $10^{-3}$. This opens up spatial statistics applications 100 times larger than previously possible.

Has companion code repository: https://github.com/flatironinstitute/gp-shootout

This page was built for publication: Equispaced Fourier representations for efficient Gaussian process regression from a billion data points

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6414386)