Flexible Gaussian Processes via Convolution
Herbert K. H. Lee, Christopher H. Holloman, Catherine A. Calder, and
Dave M. Higdon
Duke University (and LANL)
June 2002
Spatial and spatio-temporal processes are often described with a
Gaussian process model. This model can be represented as a
convolution of a white noise process and a smoothing kernel. We
expand upon this model by considering convolutions of non-{\it iid}
background processes. We highlight two particular models based on
convolutions of Markov random fields and of time-varying processes.
These models are illustrated using examples from hydrology and
atmospheric science.
Key Words: Bayesian Dynamic Spatial Model, Forward
Filtering Backward Sampling, Permeability, Ozone
The manuscript is available in postscript and
pdf formats.