A Kalman Filter Algorithm with Process Noise Covariance Update

TL;DR

This paper develops a Kalman Filter variant with an adaptive process noise covariance update, improving state estimation in biomolecular models with state-dependent noise.

Contribution

It introduces a Kalman Filter algorithm that adaptively updates process noise covariance, with theoretical guarantees for linear and certain nonlinear systems.

Findings

Optimal for systems with linear dynamics and square root-dependent noise.

Proved for discrete and continuous-time systems.

Extends to nonlinear dynamics via quadratic approximation.

Abstract

Stochastic models in biomolecular contexts can have a state-dependent process noise covariance. The choice of the process noise covariance is an important parameter in the design of a Kalman Filter for state estimation and the theoretical guarantees of updating the process noise covariance as the state estimate changes are unclear. Here we investigated this issue using the Minimum Mean Square Error estimator framework and an interpretation of the Kalman Filter as minimizing a weighted least squares cost using Newton's method. We found that a Kalman Filter-like algorithm with a process noise covariance update is the best linear unbiased estimator for a class of systems with linear process dynamics and a square root-dependence of the process noise covariance on the state. We proved the result for discrete-time system dynamics and then extended it to continuous-time dynamics using a limiting procedure. For nonlinear dynamics with a general dependence of process noise covariance on the state, we showed that this algorithm minimizes a quadratic approximation to a least squares cost weighted by the noise covariance. The algorithm is illustrated with an example.