Large Sample Properties of the Adaptive Gaussian Mixture Filter

Abstract

In high-dimensional dynamic systems, standard Monte Carlo techniques that asymptotically reproduce the posterior distribution are computationally too expensive. Alternative sampling strategies are usually applied and among these the ensemble Kalman filter (EnKF) is perhaps the most popular. However, the EnKF suffers from severe bias if the model under consideration is far from linear. Another class of sequential Monte Carlo methods is kernel-based Gaussian mixture filters, which reduce the bias but maintain the robustness of the EnKF. Although many hybrid methods have been introduced in recent years, not many have been analyzed theoretically. Here it is shown that the recently proposed adaptive Gaussian mixture filter can be formulated in a rigorous Bayesian framework and that the algorithm can be generalized to a broader class of interpolated kernel filters. Two parameters—the bandwidth of the kernel and a weight interpolation factor—determine the filter performance. The new formulation of the filter includes particle filters, EnKF, and kernel-based Gaussian mixture filters as special cases. Techniques from particle filter literature are used to calculate the asymptotic bias of the filter as a function of the parameters and to derive a central limit theorem. The asymptotic theory is then used to determine the parameters as a function of the sample size in a robust way such that the error norm vanishes asymptotically, whereas the normalized error is sample independent and bounded. The parameter choice is tested on the Lorenz 63 model, where it is shown that the error is smaller or equal to the EnKF and the optimal particle filter for a varying sample size.