Year: 2006
Type: article
Abstract: The spectral representation of Gaussian processes via the Fourier basis provides a computationally efficient specification of spatial surfaces and nonparametric regression functions in various statistical models. I describe the representation in detail and introduce the spectralGP R library for computations. Because of the large number of basis coefficients, some form of shrinkage is necessary; I focus on a natural Bayesian approach via a particular parameterized prior structure that approximates stationary Gaussian processes. I review several alternative parameterizations in the literature, suggest a simple modification suitable for exponential family data, and provide example code demonstrating MCMC sampling using the spectralGP library. I note that mixing can be slow in certain situations for reasons I describe, and provide some suggestions for MCMC techniques to improve mixing, also with example code. Note that I do not attempt to exhaustively compare parameterizations or MCMC techniques, but hope to provide a range of alternatives that may be useful for various models as well as some general recommendations grounded in experience.
Author: Paciorek, Christopher J.
Institution: Harvard University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W54348983
