B1172
Title: Smooth covariance functions for multiscale spatiotemporal models
Authors: Marco Ferreira - Virginia Tech (United States)
Thais C O Fonseca - Universidade Federal do Rio de Janeiro (Brazil) [presenting]
Abstract: Spatial covariance functions resulting from multiscale models are step functions given the discrete partition of the spatial domain of interest. However, the original process might be smooth, and step functions might not be a good characterization of spatial dependence. In that context, we derive smooth spatiotemporal covariance structures, which are obtained by averaging spatially shifted versions of the original multiscale model. To illustrate the proposal, we consider a dyadic tree data structure with L levels of resolution and present the smooth original covariance functions and the suggested mixture approximation. In particular, we consider several specifications of the model parameters in order to illustrate the possible behaviours of the implied spatiotemporal covariance functions. Posterior independence of model parameters for different tree guarantee scalability for large data analysis.
