Clustered Particle Filter Methods

• Clustered Particle Filters are sequential Monte Carlo methods that organize particles into structurally distinct groups to better capture multimodal posteriors and mitigate weight collapse.
• They incorporate methods like Gaussian mixtures, k-means clustering, block partitioning, and ancestry-tree analysis to locally manage resampling and variance.
• Empirical results demonstrate that these clustering strategies improve performance in high-dimensional and degenerate likelihood scenarios by enhancing sample efficiency and preserving mode diversity.

A clustered particle filter is a family of sequential Monte Carlo constructions in which the posterior is not managed as a single undifferentiated set of weighted Dirac particles, $$p(x_k\mid z_{1:k}) \approx \sum_i w_k^{(i)}\delta(x_k-x_k^{(i)})$$, but is organized through clusters, components, blocks, or other structured subpopulations. The literature uses this idea in several non-equivalent ways. Propagated particles may be clustered into Gaussian mixture components before measurement update, particles may be clustered before resampling to preserve multimodality, state variables may be clustered into blocks to control high-dimensional variance, and current particles may be clustered by ancestry-tree topology rather than by geometry (Veettil et al., 2016, Granade et al., 2016, Min et al., 2022, Vallivaara et al., 28 Sep 2025). This suggests that “clustered particle filter” is best read as a structural theme within particle filtering rather than as the name of a single canonical algorithm.

1. Conceptual basis and recurring objectives

The common motivation is that ordinary particle filtering is often reliable only when the particle cloud can compete globally without catastrophic weight collapse. In high-dimensional settings, the importance weights become extremely concentrated, so only a tiny fraction of particles remain compatible with the observation, and the required particle count grows essentially exponentially with dimension (Min et al., 2022). In multimodal problems, a different failure mode appears: a global resampler built from one mean and one covariance can collapse several genuine modes toward a low-probability intermediate region, especially in degenerate or aliased likelihoods (Granade et al., 2016).

Clustered formulations address these pathologies by introducing local structure into the posterior approximation. One recurring strategy is to replace many nearby particles by a smaller number of local parametric summaries. Another is to partition the filtering problem into blocks or components so that correction and resampling occur locally rather than only globally. A third is to preserve competing modes explicitly during resampling, instead of letting a single high-weight region dominate the entire particle population. The literature therefore associates clustering with at least three distinct design objectives: sample-efficiency, multimodal robustness, and dimensionality control.

A further theme is that clustering need not be geometric. In some methods, the “cluster” is a Gaussian component; in others it is a block of state variables, a subtree in an ancestry graph, or a Bernoulli-labelled component in a multitarget posterior. This makes the term structurally broad but technically precise only after the clustering variable is specified: particles, state dimensions, labels, or genealogical lineages.

2. Gaussian-component formulations

The clearest particle-space clustered construction in the provided literature is the Particle Gaussian Mixture filter. Its recursion begins with a Gaussian-mixture posterior

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$

draws an ensemble from that posterior, propagates the samples through the transition model, clusters the propagated particles, and reconstructs a predicted Gaussian mixture from cluster statistics (Veettil et al., 2016). If the propagated sample set is partitioned into clusters $S_1,\dots,S_M$, the cluster parameters are computed as

$n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$

$$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$

Each cluster then undergoes a Kalman- or least-squares-style measurement update, and the mixture weights are updated by

$$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$

The implementation uses $k$-means for clustering, supports adaptive selection of the number of clusters up to a prescribed $M_{\max}$, and adds an optional Gaussian-component merging step. The paper proves weak convergence of the PGM density to the true filter density under exponential forgetting of initial conditions, assuming a perfect clustering algorithm and adequate componentwise Gaussian update (Veettil et al., 2016).

A closely related but not identical construction appears in the Gaussian particle filter for multi-target sensing. There, the primitive object is already a Gaussian particle,

$$P_k^i=\{w^i_k,\mathcal{N}(\mathbf{x}^i_k\mid\bm{\mu}^i_k,\bm{\Sigma}^i_k)\},$$

rather than a Dirac particle later grouped into a cluster (Li, 2015). In the paper’s interpretation, $w_i$ is the probability that a target exists, and the estimated number of targets is

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$0

Prediction uses the target dynamic function, and visible particles are updated through a modified Kalman step conditioned on Boolean subset hypotheses

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$1

restricted by

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$2

The method also merges nearby Gaussian particles using a Mahalanobis-distance criterion and sets the new weight to the sum of the constituent weights. This makes the method “clustered” in spirit, but the paper explicitly does not derive its Gaussians by clustering ordinary point particles after propagation; it is better described as a Gaussian-mixture-style, existence-weighted, Kalman-updated particle representation than as a canonical sample-clustering PF (Li, 2015).

Taken together, these formulations show two distinct meanings of clustering in particle filtering. In PGM, clusters are extracted from a propagated point cloud. In the Gaussian particle filter, the cluster-like Gaussian component is the particle itself, and the explicit clustering operation appears only in the later merging rule.

3. Clustering before resampling and tree-structured posterior organization

Structured filtering is the most explicit “cluster particles before resampling” method in the provided literature. It starts from a standard SMC approximation

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$3

and replaces the usual single-cloud resampling picture by a dynamically learned collection of particle subfilters organized in a tree (Granade et al., 2016). Weighted $$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$4-means is used to partition particles: $$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$5 with weighted centroids

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$6

The posterior is represented by a rooted tree with three node types: filter nodes containing particles, mixture nodes representing weighted mixtures of child subtrees, and decision nodes representing model selection among competing clustering hypotheses. A particle’s effective weight is the product of its local weight and all edge weights on the path from the root to the leaf filter node. Resampling is triggered by the local effective sample size

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$7

and when a leaf is resampled, the algorithm may replace it by a decision node whose children correspond to competing cluster counts. Clusters are then resampled locally, typically with Liu–West, rather than globally across modes.

This organization allows the method to preserve multimodal posteriors that global Liu–West resampling would collapse. The paper formalizes branch pruning through a “champion rule”: if the largest outgoing edge weight is $$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$8, then the node is pruned to that child when

$$\pi_n(x)=\sum_{i=1}^{M(n)}\omega_i(n)\,\mathcal N(x;\mu_i(n),P_i(n)),$$9

Mixture weights at split time are inherited from cluster masses,

$S_1,\dots,S_M$0

The method is therefore a clustered PF in a strong sense: clustering is not merely an auxiliary diagnostic, but the mechanism that determines where resampling occurs and how posterior structure is represented (Granade et al., 2016).

The experimental evidence targets explicitly degenerate physics likelihoods. In randomized gap estimation, ordinary Liu–West filtering saturates after about 200 experiments, while structured filtering continues learning and reaches numerical precision within about 1000 experiments with mean canonical loss fitted by

$S_1,\dots,S_M$1

In collapse-free phase estimation, the reported asymptotic fits are

$S_1,\dots,S_M$2

for Liu–West and

$S_1,\dots,S_M$3

for structured filtering (Granade et al., 2016). These results are specific to the reported experiments, but they illustrate the central rationale for clustered resampling: preserving mode separation can matter more than preserving one global covariance.

4. Clustering of state dimensions and localized high-dimensional filtering

A different branch of the literature clusters not particles but coordinates of the state vector. In the constrained spectral-clustering block particle filter, the state space $S_1,\dots,S_M$4 is partitioned into non-overlapping blocks

$S_1,\dots,S_M$5

and the filter approximates the full posterior by a product of block marginals,

$S_1,\dots,S_M$6

This is possible under the local likelihood factorization

$S_1,\dots,S_M$7

with blockwise weights

$S_1,\dots,S_M$8

The paper formulates the partitioning problem as clustering over state variables, using the absolute correlation matrix estimated from predicted particles,

$S_1,\dots,S_M$9

as the similarity matrix in constrained spectral clustering (Min et al., 2022).

The clustering stage computes the normalized symmetric Laplacian

$n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$0

extracts the eigenvectors associated with the $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$1 smallest eigenvalues, and then performs constrained $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$2-means with a maximum block size

$n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$3

This maximum-size constraint is essential because BPF inherits a bias–variance tradeoff: smaller blocks reduce variance but break dependencies and increase blocking bias. The paper quotes the Rebeschini–Van Handel error form

$n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$4

Computational overhead is reported as $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$5 for correlation estimation, $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$6 for eigenvector computation, and an expected minimum-cost-flow complexity for the constrained clustering stage (Min et al., 2022).

Empirically, the method is evaluated on linear Gaussian systems with $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$7 and on Lorenz–96 with $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$8. In the linear Gaussian case with block-diagonal $n_i = \sum_{j=1}^N \mathbbm{1}(x_j\in S_i), \qquad w_i = \frac{n_i}{N},$9, sufficiently strong within-block correlation can lead the learned partition to recover the exact partition with ARI reaching 1. In Lorenz–96, the proposed adaptive SC-BPF performs significantly better whenever $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$0, with best performance typically around $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$1 (Min et al., 2022). This is a clustered PF only in the sense that dimensions, not particles, are clustered.

The Finkelstein algorithm is a related localized alternative rather than a standard clustered PF. It constructs each new particle as a hybrid particle whose coordinates may come from different propagated parent particles, controlled by an MCMC over source indices

$$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$2

with acceptance ratios localized to neighborhoods

$$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$3

The localized acceptance ratio

$$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$4

replaces global products by neighborhood-restricted products, and the sampled-history variant further reduces the cost from $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$5 to $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$6 when $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$7 (1901.10543). The paper explicitly presents the method as a localized hybrid-particle alternative to block PFs, with lower bias and higher variance than block PF, and with more spatially homogeneous error because fixed block boundaries are avoided (1901.10543).

5. Genealogical clustering and ancestry-tree diversity maintenance

Ancestry Tree Topology clustering introduces a non-geometric interpretation of clustered particle filtering. Instead of clustering particles by Euclidean or Mahalanobis distance, it clusters them by their position in the ancestry tree induced by repeated resampling (Vallivaara et al., 28 Sep 2025). The maintained object is a minimal tree obtained by removing dead branches and recursively merging nodes with only one child. In this minimized form, the maintained tree has at most $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$8 nodes, which supports the paper’s linear-time claim.

For any node $$\mu_i = \frac{1}{n_i}\sum_{x_j\in S_i} x_j, \qquad C_i = \frac{1}{n_i-1}\sum_{x_j\in S_i}(x_j-\mu_i)(x_j-\mu_i)^T.$$9, the subtree weight is defined recursively as

$$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$0

Given a target cluster size $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$1, a node $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$2 is a cluster root if

$$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$3

A cluster is then the set of alive leaves descending from that root. Particles not belonging to any sufficiently large subtree are assigned to the special set $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$4. This construction makes “close” particles those with nearby lineage rather than nearby state.

The full ATOG-FS method combines ancestry clustering with two diversity controls. First, cluster-dependent selection boosts particles in $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$5: $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$6 with $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$7 in the reported experiments. Second, intra-cluster fitness sharing normalizes competition within clusters: $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$8 The paper also defines an inheritance-tax mechanism through a tax cluster $$\omega_i(n)=\frac{\omega_i^-(n)l_i(n)}{\sum_j \omega_j^-(n)l_j(n)}.$$9, with particles included independently with probability $k$0, so that dead-end clusters are not protected indefinitely (Vallivaara et al., 28 Sep 2025).

The reported results are unusually explicit. In the Square robotics environment at $k$1, the success rates are 0.30 for PF, 0.98 for FDS, 1.00 for DR-SIR, 0.56 for PGM-II, 1.00 for ATOG-FS, and 1.00 for ATOG-CDS. In the real indoor positioning experiment, the success rates are 0.55 for PF, 0.35 for FDS, 0.40 for DR-SIR, 1.00 for ATOG-FS, and 0.85 for ATOG-CDS, with RMSE over all runs reported as $k$2 for ATOG-FS and $k$3 for ATOG-CDS (Vallivaara et al., 28 Sep 2025). These results support the paper’s main claim that lineage-based clustering can maintain multimodality without using any spatial or domain-specific similarity metric.

6. Distributed, multitarget, and label-structured variants

Clustered organization also appears in distributed and multitarget filters, although often in forms that are only partially aligned with the narrow particle-clustering interpretation. In the distributed computation particle PHD filter, the particle population is partitioned across $k$4 processing elements, each processing element holds $k$5 particles so that the total count is

$k$6

each processing element uses the full measurement set, and a central unit fuses local target estimates using measurement labels and majority voting (Junjie et al., 2015). Local estimation relies on the STPHD decomposition

$k$7

local target count is estimated as

$k$8

and particles are exchanged in a ring between neighboring processing elements. The architecture is therefore clustered by particle subsets and compute nodes rather than by posterior geometry.

The consensus/fusion distributed particle filter is best described as a clustered analogue rather than as a direct clustered PF. Each node runs a local filter and a separate fusion filter, with the global posterior factorized as

$k$9

Local posterior and prediction densities are summarized by Gaussians only for distributed product computation, while the fusion filter itself remains particle-based (Mohammadi et al., 2011). This can be read as a clustered architecture over sensing nodes, with explicit compensation for common past information and modified fusion rules when consensus does not converge between observation times.

In multitarget track-before-detect filtering, a different component structure appears through the labelled multi-Bernoulli approximation. An LMB posterior is specified by label-wise existence probabilities and single-target densities,

$M_{\max}$0

The generalised parallel partition PF samples label sets from

$M_{\max}$1

samples target states componentwise conditional on predicted states of the other labels, and then projects the resampled particle cloud back to an LMB by collecting all subparticles carrying each label, with existence probability

$M_{\max}$2

An optional ILMB-MCMC stage performs label-switching moves to improve the fit of the posterior to an LMB factorization when close targets remain together for a long time (García-Fernández, 2016). This is not spatial clustering, but it is a strong example of component-structured particle filtering in which the “clusters” are Bernoulli-labelled components.

7. Terminological boundaries, limitations, and unresolved issues

The literature also makes clear that not every structured particle method is a clustered particle filter in the same sense. The Gaussian particle filter for multi-target sensing is strongly related to clustered PFs because it consolidates mass into Gaussian components and merges nearby hypotheses, but it does not begin with ordinary point particles and then cluster them (Li, 2015). The constrained spectral-clustering block PF is directly relevant to the clustered-PF theme, yet it clusters state dimensions rather than particles (Min et al., 2022). The Finkelstein algorithm localizes filtering through coordinate-wise MCMC recombination over neighborhoods, but explicitly does not use fixed clusters or blocks as its main construction (1901.10543). The controlled particle filter for global optimization is a mean-field controlled particle flow with

$M_{\max}$3

and avoids importance weights and resampling entirely; it is therefore not a clustered PF in the usual sense (Zhang et al., 2017). The collective feedback particle filter for many anonymous targets estimates the empirical distribution of hidden states from the empirical distribution of measurements, again without explicit cluster management of particles or hypotheses (Kim et al., 2020).

Several limitations recur across the genuinely clustered methods. Sample-space clustering by $M_{\max}$4-means is natural for well-separated compact modes but can be poor for elongated, overlapping, or non-convex structures (Veettil et al., 2016, Granade et al., 2016). State-dimension clustering depends on sufficiently accurate correlation estimation from predicted particles, and its performance remains sensitive to the choice of $M_{\max}$5 and the maximum block size $M_{\max}$6 (Min et al., 2022). Genealogical clustering requires ancestry maintenance and may identify lineages rather than current geometric modes, which is advantageous in some problems but indirect in others (Vallivaara et al., 28 Sep 2025). Some of the most interesting constructions, including the Gaussian particle filter for aggregate sensing, have incomplete experimental validation or gaps between the methodological model and the simulation model (Li, 2015).

These boundaries matter because “clustered particle filter” can otherwise conceal substantive algorithmic differences. The term may refer to explicit clustering of particles into resampling groups, compression of a point cloud into Gaussian components, adaptive partition of state coordinates, distributed subdivision of a particle population across processors, or factorization of a multitarget posterior into Bernoulli-labelled components. The most stable technical characterization is therefore structural: a clustered particle filter is any particle-filtering method in which posterior mass is managed through explicitly organized substructures rather than through one globally resampled particle cloud.