Data assimilation using a global Girsanov nudged particle filter

TL;DR

This paper introduces a novel particle filtering algorithm using Girsanov perturbations for stochastic models in infinite-dimensional spaces, improving response to extreme events and maintaining ensemble diversity.

Contribution

The paper develops a Girsanov nudged particle filter with an optimization framework that allows parallel computation and better ensemble management in high-dimensional stochastic systems.

Findings

Responds more quickly to extreme events

Maintains better ensemble diversity

Compared favorably with temper-jitter filter

Abstract

We present a particle filtering algorithm for stochastic models on infinite dimensional state space, making use of Girsanov perturbations to nudge the ensemble of particles into regions of higher likelihood. We argue that the optimal control problem needs to couple control variables for all of the particles to maintain an ensemble with good effective sample size (ESS). We provide an optimisation formulation that separates the problem into three stages, separating the nonlinearity in the ESS term in the functional with the nonlinearity due to the forward problem, and allowing independent parallel computation for each particle when calculations are performed over control variable space. The particle filter is applied to the stochastic Kuramoto-Sivashinsky equation, and compared with the temper-jitter particle filter approach. We observe that whilst the nudging filter is over spread compared to the temper-jitter filter, it responds to extreme events in the assimilated data more quickly and robustly.