Adaptive approximation of higher order posterior statistics

Type
Article

Authors
Lee, Wonjung

KAUST Grant Number
KUK-C1-013-04

Date
2014-02

Abstract
Filtering is an approach for incorporating observed data into time-evolving systems. Instead of a family of Dirac delta masses that is widely used in Monte Carlo methods, we here use the Wiener chaos expansion for the parametrization of the conditioned probability distribution to solve the nonlinear filtering problem. The Wiener chaos expansion is not the best method for uncertainty propagation without observations. Nevertheless, the projection of the system variables in a fixed polynomial basis spanning the probability space might be a competitive representation in the presence of relatively frequent observations because the Wiener chaos approach not only leads to an accurate and efficient prediction for short time uncertainty quantification, but it also allows to apply several data assimilation methods that can be used to yield a better approximate filtering solution. The aim of the present paper is to investigate this hypothesis. We answer in the affirmative for the (stochastic) Lorenz-63 system based on numerical simulations in which the uncertainty quantification method and the data assimilation method are adaptively selected by whether the dynamics is driven by Brownian motion and the near-Gaussianity of the measure to be updated, respectively. © 2013 Elsevier Inc.

Citation
Lee W (2014) Adaptive approximation of higher order posterior statistics. Journal of Computational Physics 258: 833–855. Available: http://dx.doi.org/10.1016/j.jcp.2013.11.015.

Acknowledgements
The author thanks Dr. Chris Farmer for helpful discussions and suggestions. The author also thanks King Abdullah University of Science and Technology (KAUST) Award No. KUK-C1-013-04 for its financial support of this research.

Publisher
Elsevier BV

Journal
Journal of Computational Physics

DOI
10.1016/j.jcp.2013.11.015

Permanent link to this record