Modelling function-valued processes with complex structure

Abstract

Existing approaches to functional principal component analysis (FPCA) usually rely on nonparametric estimation of the covariance structure. When function-valued processes are observed on a multidimensional domain, the nonparametric estimation suffers from the curse of dimensionality, forcing FPCA methods to make restrictive assumptions such as covariance separability. In this thesis, we discuss a general Bayesian framework on modelling function-valued processes by using a Gaussian process (GP) as a prior, enabling us to handle nonseparable and/or nonstationary covariance structure. The nonstationarity is introduced by a convolution-based approach through a varying kernel, whose parameters vary along the input space and are estimated via a local empirical Bayesian method. For the varying anisotropy matrix, we propose to use a spherical parametrisation, leading to unconstrained and interpretable parameters and allowing for interaction between coordinate directions in the covariance function. The unconstrained nature allows the parameters to be modelled as a nonparametric function of time, spatial location and even additional covariates. In the spirit of FPCA, the Bayesian framework can decompose the function-valued processes using the eigenvalues and eigensurfaces calculated from the estimated covariance structure. A finite number of the eigensurfaces can be used to extract some of the most important information involved in data with complex covariance structure. We also extend the methods to handle multivariate function-valued processes. The estimated covariance structure is shown to be important to analyse joint variation in the data and is further used in our proposed multiple functional partial least squares regression model. We show that the interaction between the scalar response variable and function-valued covariates can be explained by fewer terms than in a regression model which uses multivariate functional principal components. Simulation studies and applications to real data show that our proposed approaches provide new insights into the data and excellent prediction results.