Abstract
dc:description.abstractIn many cases, physical dynamical systems present a plurality of characteristic scales in space and time. Differential equations developed to model the interaction of the disparate scales may further be augmented with stochastic noise to reflect uncertainty, nondeterminacy, etc. The resulting multiscale stochastic differential equations are in fact well represented in the literature in a wide range of applied fields, including physics, chemistry, and biology [5, 20, 43], neuroscience [21], meteorology [28], and econometrics and mathematical finance [11, 42]. This work is concerned with the problem of parametric statistical inference in the asymptotic regime in which the amplitude of the stochastic noise is taken to approach zero. Such is the regime of interest in studies of, e.g., rare transition events among equilibrium states of multiscale dynamical systems [8, 20, 43], small stochastic perturbations of multiscale dynamical systems [12, 21], and small-time asymptotics of multiscale models [9, 10, 35]. Both continuous and discrete data are considered. Chapter 2 considers the problem of estimation based on continuous-time (complete) observation of both slow and fast processes. This part of the work focuses on extending the results of [36] on maximum likelihood estimation to yet more general stochastic models in the averaging regime. Besides allowing general coefficients in the equations governing fast and slow processes, the work also relaxes compactness conditions such as boundedness and periodic dependence, making use of recent results of [30, 31] on solutions of certain Poisson equations. The maximum likelihood estimator is shown to be consistent, asymptotically normal, and asymptotically statistically efficient, the limiting variance having been calculated explicitly. A more convenient simplified estimator is then constructed, which is shown to be consistent as well. Chapter 3 considers the problem of estimation based on discrete-time observation of the slow process only; the fast process is not assumed to be observed. This part of the work constructs and studies minimum contrast estimators in both averaging and homogenization regimes. The particular form of these estimators is suggested the principle of maximum likelihood applied to a misspecified model, which is in turn motivated by a second-order stochastic Taylor expansion of the slow process. For a fixed number of observations, the estimators are shown to be consistent, asymptotically normal, and asymptotically statistically efficient in the sense that the limiting variance, as a function of the number of observations n, attains the Cramér-Rao bound as n → ∞. More convenient and robust simplified estimators are then constructed, which are shown to be consistent and asymptotically normal as well, although not in general efficient. Finally, sufficient conditions are given for the results to hold in the limit of frequent observation, in which the number of observations is taken to infinity at the same time that the asymptotic parameters in the model are taken to approach zero. This draws a connection with the subsampling prescribed in [29, 32] for a similar estimation problem in the asymptotic regime of long time.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics & Statistics
- Grantor
- Boston University
- Year dc:date.issued
- 2018
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gailus, Siragan
- Advisor dc:contributor.advisor
-
- Spiliopoulos, Konstantinos
Subjects
dc:subject × 9Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53534