behavior-aware particle filtering

Particle filtering is notoriously difficult for beginners to grasp due to its abstract probabilistic foundations and algorithmic complexity. Method: This work proposes a progressive pedagogical framework centered on Bayesian recursion, systematically integrating Monte Carlo sampling, weighted particle representation, and systematic resampling to articulate the full predict–update–resample pipeline. The approach bridges theory, intuition, and implementation via intuitive visualizations, rigorous mathematical derivations, and reproducible Python code. Contribution/Results: To our knowledge, this is the first lightweight, pedagogically structured particle filtering curriculum explicitly designed for nonlinear, non-Gaussian dynamic systems. By preserving algorithmic robustness while simplifying conceptual exposition, it substantially lowers the entry barrier: novices can rapidly understand, implement, and deploy state estimation algorithms in realistic noisy environments.

This work proposes a novel particle filtering framework tailored for discrete-time state-space models under severely degraded or vanishing observation noise, specifically when the observation equation is a linear function of the latent state with degenerate additive noise. The method preserves desirable filtering properties even as the noise approaches its degenerate limit and, for the first time, extends this approach to continuous-time models where the hidden state is driven by a diffusion process. By integrating state-space modeling, specialized handling of degenerate noise, and refined time-discretization analysis, the proposed algorithm demonstrates strong robustness and high accuracy in multiple numerical experiments, particularly under low-noise regimes and fine temporal resolutions.

This work addresses the failure of conventional particle filters in hidden Markov models when observation likelihoods are zero, a scenario that complicates parameter estimation and incurs high computational costs. The authors propose the Frankenfilter method, which constructs a partially activated particle filter by fixing upper and lower bounds on simulated particles while targeting a user-specified number of successful likelihood evaluations. This approach yields unbiased likelihood estimates suitable for embedding within a pseudo-marginal Metropolis–Hastings algorithm. Notably, Frankenfilter is the first to integrate a controllable success target with explicit computational constraints, substantially enhancing both robustness and efficiency: it remains stable under outliers or poor initialization, achieves 2–3× computational speedup, and provides clear guidance for setting the success target—e.g., equal to the number of exact observations \(n\) in precise observation settings.

Modeling periodic dynamics—such as circadian and seasonal rhythms—in animal behavior studies poses challenges for conventional homogeneous hidden Markov models (HMMs), which assume stationarity and fail to capture time-varying latent state distributions. Method: We propose a novel statistical inference framework for periodic non-homogeneous HMMs, analytically deriving the periodic unconditional state distribution and time-varying sojourn-time distribution for the latent state process. Integrating probability theory and stochastic processes, our approach unifies interpretable modeling, rigorous parameter inference, and model diagnostics within a single coherent framework. Contribution/Results: Applied to Drosophila photobehavioral sensor data, the method successfully identifies dynamic reconfiguration of circadian activity patterns under light-environment perturbations. It demonstrates strong validity, robustness to ecological noise, and biological interpretability in real-world sensor-based behavioral analysis—overcoming key limitations of standard HMMs in non-stationary biological time series.

This work addresses smoothing in hidden Markov models (HMMs) via conditional particle MCMC. We theoretically establish the superiority of the conditional backward particle filter (CBPF) over the conditional particle filter (CPF). Under a strong mixing assumption, we provide the first rigorous proof that CBPF achieves optimal mixing time of *O*(log *T*), significantly improving upon CPF’s *O*(*T*²). Consequently, CBPF attains total computational complexity *O*(*T* log *T*), versus *O*(*T*²) for CPF. Our key methodological innovation is the construction of a maximally coupled particle system—realizable via exact coupling of resampling steps—combined with unbiased estimation and a stochastic gradient maximum likelihood framework, yielding unbiased, finite-variance estimators of path functionals. The approach is successfully applied to parameter estimation in financial time-series HMMs. This work establishes a new paradigm for efficient, theoretically grounded smoothing and learning in state-space models.

This work proposes a novel framework that integrates classical particle filtering with learning-based methods to address the high training cost and poor interpretability of end-to-end learning in robotic state estimation. Leveraging the Markov assumption, the approach trains a dynamics model using single-step state transitions and implicitly learns the observation model via denoising score matching, thereby approximating the Bayesian filtering equations step-by-step during inference without requiring end-to-end optimization. A key innovation lies in preserving the modular structure of the filter, which enables flexible incorporation of prior knowledge and external sensor models without retraining. Experiments demonstrate that the method achieves accuracy comparable to well-tuned end-to-end baselines in simulation while significantly reducing training complexity and exhibiting superior generalization and compositional capabilities.

This work addresses the longstanding computational barrier that has hindered the application of non-Gaussian filtering to parameter estimation and Bayesian inference in state-space models, primarily due to the high cost of numerical integration. Leveraging modern computational capabilities, the authors embed unknown parameters into the state vector and integrate self-organizing state-space modeling with deterministic numerical integration to jointly estimate parameters and latent states. This approach overcomes the computational bottleneck of non-Gaussian filtering and yields stable, smooth log-likelihood estimates across low- to moderate-dimensional linear, nonlinear, and radar tracking models. The method demonstrably outperforms particle filters, whose performance is degraded by Monte Carlo noise, thereby affirming the practicality and superiority of non-Gaussian filtering under contemporary hardware conditions.

This work addresses Bayesian filtering for continuous-discrete state-space models where the hidden state evolves according to an Itô stochastic differential equation and observations arrive at discrete time instances. The authors propose a novel constrained particle filter that enforces hard support constraints on the state at each observation time via barrier functions, directly restricting the system dynamics rather than truncating the likelihood, thereby enhancing numerical stability. A unified theoretical analysis establishes convergence and time-uniform error bounds that explicitly account for numerical integration errors arising from the SDE solver. Experimental results on the stochastic Lorenz-96 system demonstrate that the proposed method effectively confines the state exploration range while maintaining high accuracy, significantly outperforming conventional particle filters.

This work addresses the challenges of particle degeneracy in traditional particle filters and the lack of rigorous Bayesian updating in existing generative approaches when assimilating high-dimensional, nonlinear, non-Gaussian data. To overcome these limitations, the authors propose the Flow-based Proposal Particle Filter (FPPF), which, for the first time, integrates a conditional generative model with computable likelihood into the particle filtering framework. By learning an approximation to the optimal proposal distribution that minimizes variance, FPPF steers particles toward high-likelihood regions and enables exact importance weighting for principled Bayesian updating. A localization strategy is further incorporated to ensure scalability in high-dimensional settings. Experimental results demonstrate that FPPF significantly outperforms both conventional and generative baselines across diverse complex dynamical systems, effectively mitigating particle degeneracy and yielding more accurate and stable posterior estimates.

This work addresses the challenges of Bayesian filtering in high-dimensional nonlinear dynamical systems, where particle degeneracy and prohibitive computational costs severely limit scalability. The study proposes a novel approach that integrates a pretrained diffusion model as a training-free generative dynamics simulator within a particle filtering framework. By circumventing the constraints of conventional numerical solvers, this method enables a theoretically optimal filtering variant previously deemed infeasible. Notably, it requires no additional training and can be efficiently deployed in high-dimensional chaotic systems—such as those arising in atmospheric dynamics—yielding substantial improvements in both estimation accuracy and computational scalability.