Slow-Particle Filter Strategies

• Slow-particle filter is a family of particle filtering techniques that deliberately increases computational effort to enhance posterior smoothing and robustness.
• These methods balance trade-offs between Monte Carlo variance, weight collapse, and runtime by employing advanced proposals, adaptive resampling, and control-based strategies.
• Variants such as windowed rejection samplers, nudged filters, and localization techniques exemplify practical approaches to improve filter performance in high-dimensional and degenerate settings.

In the cited literature, “slow-particle filter” is not a standardized algorithmic name. This suggests a broader encyclopedic usage: the phrase denotes particle-filtering constructions in which either the filter becomes slow because many particles, expensive likelihoods, or high-dimensional state spaces are required, or the algorithm is made deliberately more computationally intensive in order to obtain better smoothing, stronger robustness to degeneracy, improved rare-event performance, or a lower-dimensional representation of the posterior. Across these variants, the common substrate is sequential Monte Carlo: the posterior $p(x_k\mid z_{1:k})$ is represented by a cloud of weighted hypotheses $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$, with

$$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$

and the central issue is how to manage the trade-off between Monte Carlo variance, weight collapse, structural bias, and runtime (Dhayalkar, 3 Nov 2025, Corcoran et al., 2014).

1. Terminological scope and common Bayesian formulation

Particle filters are sequential Monte Carlo algorithms for nonlinear, non-Gaussian state-space models. In the standard formulation,

$$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$

and Bayesian filtering proceeds by the prediction–update recursion

$$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$

$$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$

In the bootstrap or SIR filter, particles are propagated with

$x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$

weighted by

$$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$

normalized, and then resampled (Dhayalkar, 3 Nov 2025).

Because the phrase “slow-particle filter” is used in several non-equivalent ways across the cited work, the topic is best understood as a family resemblance rather than a single algorithm. Some papers use “slow” in the computational sense: a PF is slow because runtime scales with particle count and with expensive model or likelihood evaluations. Others use it in a design sense: the algorithm performs more work per step—windowed rejection, control, variational guidance, or flow-based transport—to improve sample quality or avoid collapse. Still others reduce the effective state dimension through localization, manifolds, or constrained parametrizations so that an otherwise prohibitively slow PF becomes feasible.

Interpretation	Mechanism	Representative sources
Computationally slow PF	Large $N$, high-dimensional $x_k$, expensive likelihoods	(Dhayalkar, 3 Nov 2025)
Deliberately slower high-quality PF	Windowed rejection, nudging, control	(Corcoran et al., 2014, Karampela et al., 17 Mar 2026)
Weight-stabilized or resampling-free PF	Energy-based weight smoothing, feedback control	(Gilpin et al., 27 Oct 2025, Zhang et al., 2017)
Structure-exploiting PF	Lattice QMC, adaptive blocking, slow manifolds, observation manifolds	(Ormoneit et al., 2013, Bertoli et al., 2014, Qiao et al., 2017, Zhumekenov et al., 13 Jan 2026)

2. Sources of slowness in classical particle filtering

In a basic SIR implementation, per-time-step cost is linear in particle count: $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$0 so the total cost is $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$1 per step, with the dominant constant determined by propagation and likelihood evaluation. The beginner-oriented treatment in “Particle Filter Made Simple: A Step-by-Step Beginner-friendly Guide” emphasizes five recurrent causes of perceived slowness: large particle count $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$2, high-dimensional states $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$3, expensive sensor likelihoods, resampling overhead, and non-vectorized implementations. It also stresses the curse of dimensionality: in $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$4D, $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$5 may suffice, whereas in $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$6D or $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$7D one may need thousands of particles to cover the state space adequately; consequently, naive $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$8D pose tracking can become very slow (Dhayalkar, 3 Nov 2025).

This computational picture explains why PFs are frequently contrasted with Kalman-type methods. Kalman, EKF, and UKF avoid Monte Carlo sampling, so their cost scales with state dimension through matrix operations rather than through $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$9. The trade-off is model class: PFs drop Gaussian and linear assumptions and can represent multimodal, skewed, and heavy-tailed posteriors. The cost of that modeling flexibility is that accuracy is often purchased by increasing $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$0.

The same source also makes clear that “slow” is contextual rather than absolute. In a $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$1D tracking example with $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$2, $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$3, and $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$4 particles, and in a $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$5D position–velocity example with $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$6, vectorized NumPy implementations remain fast. This suggests that the phrase “slow-particle filter” usually refers not to all PFs, but to regimes in which state dimension, proposal mismatch, or implementation choices make the $$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$7 loop operationally costly (Dhayalkar, 3 Nov 2025).

3. Degeneracy, sample impoverishment, and low-noise collapse

The canonical pathology behind slow-particle filtering is weight degeneracy. In its extreme form, a single weight becomes nearly one while the rest are nearly zero, so the empirical posterior collapses to an almost single-point mass. A standard diagnostic is the effective sample size

$$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$8

or, equivalently in SIR notation,

$$p(x_k \mid z_{1:k}) \approx \sum_{i=1}^{N} w_k^{(i)}\, \delta(x_k - x_k^{(i)}),$$9

When one particle dominates, $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$0; when weights are uniform, $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$1. Closely related diagnostics are the maximum weight $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$2 and the variance of log-weights $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$3 (Gilpin et al., 27 Oct 2025, Corcoran et al., 2014).

Resampling combats degeneracy but introduces sample impoverishment. In the windowed rejection sampling study, SIR produced only roughly $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$4 distinct $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$5-dimensional trajectories in one example, and for the smoothing marginal $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$6 only about $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$7 of values were distinct. The same work emphasizes a third issue: the resulting draws are dependent rather than i.i.d., and smoothing is especially fragile because early ancestral states collapse onto a few trajectories unless one uses additional forward–backward machinery (Corcoran et al., 2014).

A distinct but related collapse mechanism occurs under low or degenerate observational noise. For Gaussian observation noise with variance scaled by $$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$8, bootstrap weights take the form

$$x_k \sim p(x_k \mid x_{k-1},u_k), \qquad z_k \sim p(z_k \mid x_k),$$9

As $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$0, the likelihood becomes extremely concentrated: only particles almost exactly on the observation manifold retain non-negligible weight. In the degenerate case $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$1, the posterior is supported on

$$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$2

so standard PFs in the ambient space become numerically unstable and require enormous particle counts (Zhumekenov et al., 13 Jan 2026).

A common misconception is that degeneracy is merely a resampling problem. The cited work shows that it is more fundamental: it arises from proposal mismatch, dimensionality, peaked likelihoods, and rare-event structure. Resampling addresses the symptom of skewed weights; it does not by itself solve path collapse, ancestral dependence, or observation-manifold singularity.

4. Deliberately slower but higher-quality particle filters

One important meaning of slow-particle filtering is a method that spends more computation per particle to improve posterior quality. The clearest example is the Windowed Rejection Sampler (WRS). Instead of importance sampling plus resampling, WRS applies rejection sampling to local windows $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$3, using a proposal built from prior dynamics and an acceptance rule based on block likelihoods. For fixed window length $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$4, WRS yields i.i.d. samples from exact block-level targets, produces full trajectories in a single pass, and turns smoothing into direct empirical marginalization over accepted paths. The approximation arises because each window conditions only on local future observations rather than on the entire $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$5; accuracy improves as $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$6 increases. WRS can be slower because each accepted block may require several proposals and because larger $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$7 reduces acceptance probability, but in the reported C++ experiments runtimes were often comparable to SIR and sometimes better: in a linear-Gaussian example WRS finished in under $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$8 seconds while SIR took $$p(x_k \mid z_{1:k-1}) = \int p(x_k \mid x_{k-1})\,p(x_{k-1}\mid z_{1:k-1})\,dx_{k-1},$$9–$$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$0 seconds, and in a highly nonlinear example WRS took about $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$1 seconds while SIR took about $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$2 minute. In the dynamic Tobit example, both were on the order of $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$3 minutes (Corcoran et al., 2014).

A second family is the nudged particle filter (nPF) and its variational pseudo-observation guided extension. The nPF replaces pure prior propagation by a controlled SDE,

$$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$4

where the control is derived from an optimal control problem with terminal cost given by the negative log-likelihood of the upcoming observation. The price is high computational burden: on each observation interval one must approximate Feynman–Kac expectations repeatedly to obtain controls. The variational extension constructs a strong-constraint 4D-Var path $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$5, defines pseudo-observations $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$6, and replaces one long-horizon control problem by many short fixed-horizon ones. In stochastic Lorenz-63 with $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$7 particles, $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$8, and $$p(x_k \mid z_{1:k}) \propto p(z_k \mid x_k)\,p(x_k \mid z_{1:k-1}).$$9 control subintervals, reported runtimes were $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$0 s for nPF and $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$1 s for Var-nPF in one representative experiment; over $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$2 Monte Carlo runs the PF with $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$3 particles had average RMSE $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$4, nESS about $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$5, and runtime $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$6 s, whereas Var-nPF with $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$7 particles had average RMSE $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$8, nESS about $x_k^{(i)} \sim p(x_k \mid x_{k-1}^{(i)},u_k),$9, and runtime $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$0 s (Karampela et al., 17 Mar 2026).

A broader control-theoretic reinterpretation appears in the controlled particle filter for global optimization. There, particles evolve with equal weights under

$$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$1

and the control is chosen so that the empirical law follows a KL-gradient flow for

$$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$2

The selected feedback $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$3 is characterized by a weighted Poisson equation and is the minimum-$$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$4 solution among controls that realize the desired divergence constraint. Resampling, birth, and death of particles are avoided entirely. Although this construction targets optimization rather than state estimation, it exemplifies a “slow” PF architecture in which smooth controlled transport is preferred to abrupt resampling because of variance and stability considerations (Zhang et al., 2017).

5. Variance reduction and efficiency-enhancing variants

Not all responses to slow-particle filtering add more computation; many seek to make each particle more useful. At the algorithmic level, three elementary levers recur: a better proposal distribution $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$5 that uses the current measurement, adaptive resampling based on $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$6, and vectorized or parallel implementations. In the SIR setting, the beginner guide recommends resampling only when $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$7, often $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$8, and notes that systematic resampling is $$w_k^{(i)} \propto w_{k-1}^{(i)}\,p(z_k\mid x_k^{(i)}),$$9 and less noisy than naive multinomial resampling. It also emphasizes that vectorized prediction and likelihood evaluation can turn modest-dimensional PFs into real-time algorithms (Dhayalkar, 3 Nov 2025).

A more explicit anti-collapse device is the modified particle filter that reduces weight collapse. After classical Bayesian weights

$N$0

are computed, the method performs a single explicit Euler step of an interacting-weight gradient flow: $N$1 with a Morse-type potential, projection to $N$2 if necessary, and renormalization. The added step is $N$3, but in Lorenz-63 it substantially reduced $N$4, increased $N$5, and improved state estimates. For $N$6, mean analysis RMS error fell from $N$7 to $N$8; for $N$9, it fell from $x_k$0 to $x_k$1. Mean $x_k$2 increased from $x_k$3 to $x_k$4 at $x_k$5 and from $x_k$6 to $x_k$7 at $x_k$8 (Gilpin et al., 27 Oct 2025).

The lattice particle filter (LPF) attacks slowness by replacing i.i.d. proposal uniforms with randomized low-discrepancy lattice points. The bootstrap structure is preserved—resampling and likelihood weighting remain standard—but propagation uses a shifted Korobov lattice rule with per-time-step random permutations. Theoretical motivation comes from QMC error $x_k$9 versus Monte Carlo $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$00 under suitable conditions, while randomized shifting preserves unbiasedness. In a synthetic $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$01D tracking problem, the LPF was equivalent to a conventional PF that had between $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$02 and $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$03 more particles, depending on particle sparsity in state space. Across $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$04–$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$05 particles, LPF reduced variance by at least $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$06, and the conventional PF needed between $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$07 and $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$08 more particles to match LPF performance. In a $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$09D human-motion problem, the variance of estimated posterior means across repeated runs was $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$10–$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$11 lower for LPF (Ormoneit et al., 2013).

These variants underscore an important point: “slow-particle filter” need not mean a filter that literally executes slowly. It can also denote a regime in which a naive PF would require so many particles that the practical remedy is to reduce variance, regularize weights, or improve space-filling so that the same accuracy is obtained with fewer particles.

6. Localization, slow manifolds, and constrained-manifold formulations

A different response to slow-particle filtering is to reduce the effective filtering problem. In large-scale dynamic random fields, a standard PF is prohibitive because approximation error is exponentially dependent on model dimension, so an enormous number of particles is required. The adaptively blocked particle filter localizes the recursion on blocks of sites and cycles among several partitions $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$12. Blocking makes the variance term depend mainly on block size rather than on global dimension, but it introduces a spatially varying bias that depends on distance to the block boundary. The adaptive extension averages this bias through

$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$13

and for a certain class of random field achieves a completely spatially uniform bound on the bias; more generally it significantly reduces spatial inhomogeneity compared with fixed blocking (Bertoli et al., 2014).

In slow–fast stochastic systems, slowness can be reduced by filtering only the slow coordinates. For

$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$14

the cited work constructs a random slow manifold

$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$15

derives a reduced low-dimensional system on that manifold, and proves that the low-dimensional filter approximates the original filter in a suitable metric. The approximation error is controlled by the time-scale separation parameter $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$16 and by the initial distance to the manifold. This is, in a literal sense, a slow-particle filter: particles need only represent the slow component, with fast variables slaved to the manifold relation (Qiao et al., 2017).

The most explicit constrained-manifold treatment appears in filtering with low and degenerate observational noise. When

$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$17

and $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$18 is small, standard ambient-space PFs collapse because the posterior concentrates near the observation manifold. The proposed remedy is to parametrize particles directly on the appropriate manifold. In the degenerate case $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$19,

$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$20

and one writes

$\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$21

In the low-noise case, an extended state–noise manifold is used instead. The resulting PFs are constructed so that, as $\{x_k^{(i)},w_k^{(i)}\}_{i=1}^N$22, the low-noise PF provably inherits the properties of the degenerate-case PF. The framework is extended to diffusion-driven hidden states by combining endpoint manifold parametrization with guided diffusion bridges, producing algorithms robust to both low noise and fine time discretization (Zhumekenov et al., 13 Jan 2026).

Taken together, these developments show that slow-particle filtering is best understood as a design space organized around one question: when direct SMC in the ambient state space is too costly, too unstable, or too degenerate, should one spend more computation for better proposals and smoothing, or should one reformulate the state space so that fewer, better-targeted particles suffice? The literature supports both answers, and the term “slow-particle filter” plausibly spans both.