An Operator-Theoretic Analysis of Nonlinear Filtering under Model Misspecification

Abstract

We study Bayesian filtering under misspecified model dynamics. The optimal filter is governed by predictive operators $K=\{K_t\}_{t\ge 1}$ associated with the true dynamics, while the approximate filter uses perturbed operators $\widehat K=\{\widehat K_t\}_{t\ge 1}$. Both filters share the same update operators $U=\{U_t\}_{t\ge 1}$ induced by the observation model. We work in a fully probabilistic setting in which observations are random variables, rather than fixed realizations. This viewpoint yields a linear recursion for the mean measures of the optimal filter and makes the propagation of filtering errors depend on two ingredients: the contraction of the predictive operators, and the accumulated model discrepancy. Within this framework, we derive explicit bounds for the discrepancy between the optimal and misspecified filters. The bounds separate the effect of the initial-condition error from the perturbation introduced by the dynamical mismatch $\widehat K-K$. Under suitable contractivity assumptions, the cumulative error stabilizes at a steady-state level determined by the size of the misspecification and by the Dobrushin coefficients of the predictive operators. We also give analytical examples in which the assumptions can be verified directly. In particular, if the transition functions are bounded and the transition noise belongs to a family of elliptically symmetric distributions, including Gaussian, Student-$t$ and Laplace laws, then the filter error stabilizes as predicted by the theory. The results provide a theoretical basis for quantifying the robustness of nonlinear filters under random observations and misspecified dynamics.