A Variational Formulation of Optimal Nonlinear Estimation

Abstract

We propose a variational method to solve all three estimation problems for nonlinear stochastic dynamical systems: prediction, filtering, and smoothing. Our new approach is based upon a proper choice of cost function, termed the {\it effective action}. We show that this functional of time-histories is the unique statistically well-founded cost function to determine most probable histories within empirical ensembles. The ensemble dispersion about the sample mean history can also be obtained from the Hessian of the cost function. We show that the effective action can be calculated by a variational prescription, which generalizes the ``sweep method'' used in optimal linear estimation. An iterative numerical scheme results which converges globally to the variational estimator. This scheme involves integrating forward in time a ``perturbed'' Fokker-Planck equation, very closely related to the Kushner-Stratonovich equation for optimal filtering, and an adjoint equation backward in time, similarly related to the Pardoux-Kushner equation for optimal smoothing. The variational estimator enjoys a somewhat weaker property, which we call ``mean optimality''. However, the variational scheme has the principal advantage---crucial for practical applications---that it admits a wide variety of finite-dimensional moment-closure approximations. The moment approximations are derived reductively from the Euler-Lagrange variational formulation and preserve the good structural properties of the optimal estimator.