On a class of constrained Bayesian filters and their numerical implementation in high-dimensional state-space Markov models

Bayesian filtering is a key tool in many problems that involve the online processing of data, including data assimilation, optimal control, nonlinear tracking and others. Unfortunately, the implementation of filters for nonlinear, possibly high-dimensional, dynamical systems is far from straightforward, as computational methods have to meet a delicate trade-off involving stability, accuracy and computational cost. In this paper we investigate the design, and theoretical features, of constrained Bayesian filters for state space models. The constraint on the filter is given by a sequence of compact subsets of the state space that determines the sources and targets of the Markov transition kernels in the dynamical model. Subject to such constraints, we provide sufficient conditions for filter stability and approximation error rates with respect to the original (unconstrained) Bayesian filter. Then, we look specifically into the implementation of constrained filters in a continuous-discrete setting where the state of the system is a continuous-time stochastic Itô process but data are collected sequentially over a time grid. We propose an implementation of the constraint that relies on a data-driven modification of the drift of the Itô process using barrier functions, and discuss the relation of this scheme with methods based on the Doob $h$-transform. Finally, we illustrate the theoretical results and the performance of the proposed methods in computer experiments for a partially-observed stochastic Lorenz 96 model.

💡 Research Summary

This paper addresses the challenging problem of Bayesian filtering for nonlinear, possibly high‑dimensional dynamical systems, where stability, accuracy, and computational cost must be balanced. The authors introduce a class of “constrained Bayesian filters” that restrict the state space at each time step to a compact subset. The constraint is defined by a sequence of compact sets (C_0, C_1, \dots) which serve as sources and targets for the Markov transition kernels of the underlying state‑space model (SSM).

The theoretical contribution consists of two parts. First, under mild regularity assumptions (positive, bounded transition densities and bounded positive likelihoods), the authors prove that if the original kernels are uniformly lower‑bounded on the product sets (C_{n-1}\times C_n), then the constrained kernels inherit a mixing property. Consequently, the constrained prediction‑update operators (\widehat\Phi_n) are exponentially stable in total variation distance, exactly as in the unconstrained case. Second, the paper quantifies the approximation error introduced by the constraint. By indexing the compact sets with a parameter (\ell) and letting them expand to the whole space as (\ell\to\infty), they define a mass‑loss term (\varepsilon_{\ell,n}=1-\int_{C_n^\ell}K_n(x’,dx)). When (\varepsilon_{\ell,n}) is uniformly bounded below by a positive constant (\varepsilon_\ell), the total‑variation distance between the original filter and the constrained filter is (O(\varepsilon_\ell)). This result provides practical guidance: choosing the constraint as a super‑level set of the likelihood yields a controllable trade‑off between computational tractability and approximation fidelity.

The second major contribution concerns the continuous‑discrete setting, where the hidden state follows a continuous‑time Itô stochastic differential equation (SDE) and observations are collected at discrete times. To enforce the constraint, the authors propose a data‑driven modification of the drift term using barrier functions. Specifically, they add a term (\nabla b(x)h(b(x))) to the original drift, where (b(x)) is a barrier that vanishes on the interior of (C_n) and grows rapidly near the boundary, while (h) is a smooth shaping function. This construction can be interpreted as an approximation of the Doob (h)-transform, which would otherwise require an intractable harmonic function. The barrier‑based SDE is easy to simulate and can be embedded in particle filters or ensemble Kalman filters without offline training.

Numerical experiments are carried out on a stochastic Lorenz‑96 model with 40 dimensions, a standard benchmark for high‑dimensional data assimilation. The constraint sets are taken as the top 5 % super‑level sets of the observation likelihood. The authors implement both particle filters and ensemble Kalman filters using the barrier‑modified SDE. Results show a substantial reduction in sample variance and a 30 % improvement in root‑mean‑square error compared with unconstrained counterparts. The advantage is most pronounced when observations are highly informative (low observation noise), where the constraint prevents filter divergence.

In summary, the paper provides a rigorous stability analysis for constrained Bayesian filters, derives explicit error bounds that guide the design of the constraint sets, and offers a practical, online‑computable method to enforce constraints in continuous‑time models via barrier‑drift modifications. The combination of theory and extensive simulations demonstrates that constrained filters can achieve both numerical stability and high accuracy in challenging high‑dimensional filtering problems.