Backward Smoothing versus Fixed-Lag Smoothing in Particle Filters

TL;DR

The paper addresses the practical trade-offs in particle smoothing for nonlinear and non-Gaussian state-space models by comparing backward smoothing (FFBS and FFBSm) with fixed-lag smoothing in a trend-model setting. It introduces $\,\mathcal{O}(m)$ approximations of FFBSm via subsampling and local neighborhoods and evaluates them under Gaussian and heavy-tailed noise, measuring accuracy with a grid-based divergence. The key findings show that FFBS and FFBSm provide higher accuracy at a fixed particle count, but fixed-lag smoothing often achieves better accuracy within the same CPU time, due to lower per-step cost; localization helps FFBSm substantially in Gaussian regimes and is less advantageous for heavy-tailed dynamics. The work yields practical guidelines: use FFBSm with localization for light-tailed dynamics to gain accuracy efficiently, while fixed-lag smoothing may be preferable when computational budgets limit particle numbers and runtime.

Abstract

Particle smoothing enables state estimation in nonlinear and non-Gaussian state-space models, but its practical use is often limited by high computational cost. Backward smoothing methods such as the Forward Filter Backward Smoother (FFBS) and its marginal form (FFBSm) can achieve high accuracy, yet typically require quadratic computational complexity in the number of particles. This paper examines the accuracy--computational cost trade-offs of particle smoothing methods through a trend-estimation example. Fixed-lag smoothing, FFBS, and FFBSm are compared under Gaussian and heavy-tailed (Cauchy-type) system noise, with particular attention to O(m) approximations of FFBSm based on subsampling and local neighborhood restrictions. The results show that FFBS and FFBSm outperform fixed-lag smoothing at a fixed particle number, while fixed-lag smoothing often achieves higher accuracy under equal computational time. Moreover, efficient FFBSm approximations are effective for Gaussian transitions but become less advantageous for heavy-tailed dynamics.

Figures (7)

• Figure 1: Test data used for the empirical study (Kitagawa 1987, Kitagawa 1996, Kitagawa 2014).
• Figure 2: "True" smoothed distributions for the Gaussian and Cauchy models, obtained by the Kalman smoother and by a fixed-lag particle smoother with $m=10{,}000{,}000$ particles, respectively.
• Figure 3: Smoothing accuracy comparison for the Gaussian model. (Left) Accuracy versus particle number $m$ (FFBS and fixed-lag smoothing). (Right) Accuracy versus CPU time for the same experiments. FFBS with $m=10^5$ is computed only once and hence denoted by light blue.
• Figure 4: Smoothing accuracy for the Gaussian model. (Left) NS-FFBSm accuracy versus $m_s$ for different particle numbers $m$, with fixed-lag smoothing shown for comparison. (Right) Accuracy versus CPU time for the same experiments.
• Figure 5: Smoothing accuracy for the truncated Cauchy model. (Left) Accuracy versus $m_s$ on a log--log scale. (Right) Accuracy versus CPU time.