Variance reduction for particle filters of systems with time-scale separation

We present a particle filter construction for a system that exhibits time-scale separation. The separation of time-scales allows two simplifications that we exploit: i) The use of the averaging principle for the dimensional reduction of the system needed to solve for each particle and ii) the factorization of the transition probability which allows the Rao-Blackwellization of the filtering step. Both simplifications can be implemented using the coarse projective integration framework. The resulting particle filter is faster and has smaller variance than the particle filter based on the original system. The method is tested on a multiscale stochastic differential equation and on a multiscale pure jump diffusion motivated by chemical reactions.

💡 Research Summary

The paper addresses the computational and statistical challenges inherent in particle filtering for systems that exhibit a pronounced separation of time scales. Traditional particle filters applied directly to the full high‑dimensional stochastic model suffer from two major drawbacks: (i) the need to propagate each particle through expensive fine‑scale dynamics, and (ii) large variance in the importance weights caused by sampling noise, which often leads to particle impoverishment after resampling. The authors propose a novel particle filtering framework that simultaneously mitigates both issues by exploiting the mathematical structure of multiscale systems.

The first key ingredient is the averaging principle. When a system contains fast variables that evolve on a much shorter time scale than the slow variables, the fast dynamics can be replaced by their invariant (or stationary) distribution conditioned on the slow state. This yields a reduced‑order “effective” model that only involves the slow variables, while preserving the correct marginal dynamics. By integrating this reduced model instead of the full system, each particle’s prediction step becomes dramatically cheaper: only the slow dynamics need to be simulated, and the fast component is accounted for analytically through its averaged effect.

The second ingredient is a Rao‑Blackwellization of the weighting step. The transition density of the original system can be factorized into a product of a slow‑component kernel and a fast‑component kernel. Because the fast component has been averaged out, its conditional expectation given the slow state can be computed either analytically or via a short, high‑resolution micro‑simulation. Substituting this conditional expectation for the sampled fast trajectory eliminates the stochastic variability associated with the fast variables in the weight calculation. Consequently, the variance of the importance weights is reduced, leading to more stable particle sets and fewer resampling operations.

Implementation is carried out within the coarse projective integration (CPI) framework. CPI performs brief bursts of fine‑scale simulation to estimate the necessary statistics of the fast dynamics, then projects the slow variables forward over a much larger macro‑time step using the averaged dynamics. In the context of particle filtering, each particle undergoes a short micro‑burst to evaluate the conditional expectations required for Rao‑Blackwellization, after which the macro‑step prediction proceeds with the reduced model. This hybrid approach retains the accuracy of the full multiscale dynamics while achieving substantial computational savings.

The authors validate their methodology on two benchmark problems. The first is a multiscale stochastic differential equation comprising a fast Ornstein‑Uhlenbeck process coupled to a slow nonlinear drift. The second is a multiscale pure‑jump diffusion model inspired by chemical reaction networks, where rapid reaction events (jumps) coexist with slower diffusion of species concentrations. For both cases, the proposed filter is compared against a conventional particle filter that directly simulates the full system. Performance metrics include mean‑square error (MSE) of the state estimate, variance of the importance weights, and total runtime.

Results show that the averaged model reproduces the slow‑state statistics with negligible bias, even when the fast dynamics exhibit strong fluctuations. Rao‑Blackwellization reduces the weight variance by roughly 30–50 %, which translates into fewer resampling steps and a marked decrease in particle degeneracy. Overall computational cost is lowered by a factor of 2–4, depending on the stiffness of the fast subsystem and the chosen macro‑time step size. The experiments also demonstrate that the method remains effective for pure‑jump processes, where the fast component is discontinuous, confirming the broad applicability of the approach.

In summary, the paper makes three principal contributions: (1) it integrates the averaging principle into particle filtering to achieve dimensional reduction for multiscale stochastic systems; (2) it introduces a Rao‑Blackwellized weighting scheme that leverages the factorized transition density to cut down sampling variance; and (3) it embeds both techniques within a coarse projective integration scheme, providing a practical algorithm that is both faster and more statistically efficient than standard particle filters. The methodology is particularly relevant for real‑time data assimilation in fields such as chemical kinetics, climate modeling, and quantitative finance, where systems often display widely separated time scales and where computational resources are at a premium.