Efficient spatio-temporal Gaussian regression via Kalman filtering

We study the non-parametric reconstruction of spatio-temporal dynamical processes via Gaussian Processes (GPs) regression from sparse and noisy data. GPs have been mainly applied to spatial regression where they represent one of the most powerful estimation approaches also thanks to their universal representing properties. Their extension to dynamical processes has been instead elusive so far since classical implementations lead to unscalable algorithms or require some sort of approximation. We propose a novel procedure to address this problem by coupling GPs regression and Kalman filtering. In particular, assuming space/time separability of the covariance (kernel) of the process and rational time spectrum, we build a finite-dimensional discrete-time state-space process representation amenable to Kalman filtering. With sampling over a finite set of fixed spatial locations, our major finding is that the current Kalman filter state represents a sufficient statistic to compute the minimum variance estimate of the process at any future time over the entire spatial domain. In machine learning, a representer theorem states that an important class of infinite-dimensional variational problems admits a computable and finite-dimensional exact solution. In view of this, our result can be interpreted as a novel Dynamic Representer Theorem for GPs. We then extend the study to situations where the spatial input locations set varies over time. The proposed algorithms are tested on both synthetic and real field data, providing comparisons with standard GP and truncated GP regression techniques.