Positive Definite Completions and Continuous Graphical Models

This thesis concerns the theory of positive-definite completions and its mutually beneficial connections to the statistics of function-valued or continuously-indexed random processes, better known as functional data analysis. In particular, it dwells upon the reproducing kernel character of covariances.

In the introduction, we attempt to summarize the basic ideas and thoughts upon which the thesis is built.

Chapter 1 deals with the problem of covariance completion and develops an intuitive and interpretable approach to the problem of covariance estimation for fragmented functional data based on the concept of canonical completion. For a suitably restricted class of domains, we describe how the canonical completion may be constructed and use it to produce a characterization of the set of all completions. Furthermore, we identify necessary and sufficient conditions for uniqueness of completion and for the exact recovery of the true covariance. In doing so, we settle many fundamental questions concerning covariance estimation with fragmented functional data.

Chapter 2 considers the problem of positive-definite completion in its generality and represents a purely mathematical treatment of the subject compared to Chapter 1. We study the problem for many classes of domains and present results concerning existence and uniqueness of solutions, their characterization and the existence and uniqueness of a special solution called canonical completion. We prove many new variational and algebraic characterizations of the canonical completion. Most importantly, we show the existence of canonical completion for the class of band-like domains. This leads to the existence of a canonical extension in the context of the classical problem of extensions of positive-definite functions, which is shown to correspond to a strongly continuous one-parameter semigroup and consequently, to an abstract Cauchy problem.

Chapter 3 presents a rigorous generalization of undirected Gaussian graphical models to arbitrary, possibly uncountable index sets. We prove an inverse zero characterization for these models, analogous the one known for multivariate graphical models and develop a procedure for their estimation based on the notion of resolution. The utility of the concept and method is illustrated using many real data applications and simulation studies.

Chapter 4 concerns the problem of recovering conditional independence relationships between a finite number of jointly distributed second-order Hilbertian random elements in a sparse high-dimensional regime with a particular interest in multivariate functional data. We propose an infinite-dimensional generalization of the multivariate graphical lasso and prove model selection consistency under natural assumptions. The method can be motivated from a coherent maximum likelihood philosophy.

Chapter 5 discusses ways in which the results of this thesis can be strengthened or completed and its ideas and conclusions extended. With