Flexible bayesian inference for families of random densities.
Loading...
Date
2020
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
A main goal of this thesis is to propose and study novel flexible Bayesian models for setups that entail
families of random densities. Two specific contexts will be examined: one involves phase-varying point
processes, whereas the other involves functional principal component analysis. The common denominator
underlying these contexts is the need to model families of random measures to each of which corresponds a
different data generating process. On both contexts, prior processes will be used so to devise priors on the
target objects of interest.
In more detail, one context entails separating amplitude variation from phase variation in a multiple
point process setting. In this framework, I pioneer the development of priors on spaces of warping maps by
proposing a novel Bayesian semiparametric approach for modeling registration of multiple point processes.
Specifically, I develop induced priors for warp maps via a Bernstein polynomial prior so to learn about the
structural measure of the point process and about the phase variation in the process. Theoretical properties
of the induced prior, including support and posterior consistency, are established under a fairly mild proviso.
Also, numerical experiments are conducted to assess the performance of this new approach; finally, a real
data application in climatology illustrates the proposed methodology.
The other context that will be considered in this thesis involves modeling families of random densities
using functional principal component analysis through the so-called Karhunen–Loève decomposition. For
this, I develop a data-driven prior based on the Karhunen–Loève decomposition which can be used to
borrowing strength across samples. The proposed approach defines a prior on the space of families of
densities. Theoretical properties are developed to ensure that the trajectories from an infinite mixture
belong to L
2 which is a necessary condition for the Karhunen–Loève decomposition to hold. Numerical
experiments are conducted to assess the performance of the proposed approach against competing methods,
and we offer an illustration by revisiting Galton’s height parents dataset.
Description
Tesis (Doctor in Statistics)--Pontificia Universidad Católica de Chile, 2020