FunCalc4Stats · Functional Calculus for Computational Statistics

Computations with spatial or spatiotemporal data occur in various areas such as the social and environmental sciences. The technological progress in data collection and storage capacities in recent years has caused a tremendous growth in data, and thereby exposed the limitations of state-of-the-art computational methods for statistical inference and predictions. Achieving computational feasibility severely limits the dependency structures of underlying random processes in currently deployed models. Above all, there is a strong need for methodologies that capture intricate spatial or spatiotemporal dependencies.This project seeks to characterize flexible dependency structures by means of functional calculus and to exploit this characterization for efficient computations with Gaussian processes. My overall aim consists of a systematic construction and approximation of covariance operators, corresponding to sophisticated nonlocal dependencies, via the action of functions on local linear differential operators.To establish numerical methods for operator functions, I will consider contour integral representations. Expedient transformations of the contours will facilitate operator-valued quadratures based on sinc approximations, which I will combine with uniformly stable discretizations for variational formulations of partial differential equations (PDEs) such as elliptic diffusion, parabolic and Stokes equations. To derive sharp rates of convergence, I will employ discrete inf-sup theory. Thereby I will address specific classes of functions and of differential operators suitable for inference from data. A particular focus will be on spatiotemporal statistical models based on fractional powers of non-autonomous parabolic operators, for which I will analyse space-time finite element methods.This project will considerably strengthen the interaction between Numerical Analysis and Statistics by enabling PDE-based methodologies for efficient computations with spatial data.