Particle Filter¶

Approximate a recursive hidden-state posterior with a weighted particle population that is propagated through a state model, corrected by observation likelihoods, and selectively resampled to control weight degeneracy.

Version: v2 · 2026-09-06 · History
Domain-specific #: 2455
Origin domain: statistics
Subdomain: sequential bayesian filtering
Aliases: Particle Filtering, Sequential Monte Carlo Filter, Smc Filter

Core Idea¶

A particle filter is an online Monte Carlo method for estimating the changing, partly hidden state of a dynamical system from noisy observations. It represents the filtering distribution

\[ p(x_t\mid y_{1:t}) \]

not by a closed-form density, a fixed grid, or only a mean and covariance, but by a finite weighted empirical distribution

\[ \widehat p_N(dx_t\mid y_{1:t})=\sum_{i=1}^{N}w_t^{(i)}\,\delta_{x_t^{(i)}}(dx_t), \qquad \sum_i w_t^{(i)}=1. \]

Here each particle is a candidate state, and its normalized weight records how much posterior support that candidate receives. When a new time step arrives, the filter performs two logically indispensable operations: it proposes or propagates candidate states using a transition-informed proposal, and it weights them by the observation and by any correction between the proposal and the target. Selective resampling is the characteristic and ordinarily necessary practical control that reassigns computation from negligible-weight particles to supported regions before importance weights become unusably concentrated. Pure sequential-importance-sampling formulations can omit resampling, but repeated updates then incur severe weight degeneracy.

Scope of Application¶

Nonlinear and non-Gaussian state estimation. This is the central habitat: recursively estimating a hidden Markov state where linear-Gaussian closure is unavailable. The bootstrap filter, Monte Carlo filter, and later generalized importance-sampling formulations all preserve the same weighted-population recursion.
Target tracking and signal processing. Bearings-only tracking, maneuvering-target models, channel tracking, and related signal-estimation tasks reuse the transition, likelihood, weighted-particle, and resampling roles. The 1993 bootstrap-filter paper used nonlinear bearings-only tracking as a principal demonstration.

Clarity¶

Particle filtering clarifies an otherwise vague claim—“the tracker maintains several guesses”—by forcing each guess into an explicit inferential role. What distribution should the population approximate? What state does a particle encode? Which model predicts descendants? Which likelihood scores the new observation? Which proposal actually generated each descendant, and where is the proposal-to-target correction? When does effective support become too concentrated, and which resampling scheme renews it?

Manages Complexity¶

The exact nonlinear filtering recursion operates on full probability distributions. Grid methods face rapidly increasing storage with state dimension; analytic filters require closure assumptions; carrying every possible state trajectory is impossible. A particle filter compresses that distributional object into \(N\) weighted representatives and reuses the same population from one observation to the next. Computation is concentrated in regions the current model and data regard as plausible rather than allocated uniformly over a state-space grid.

Abstract Reasoning¶

The mechanism licenses a compact set of predictive and intervention inferences:

If weights concentrate before resampling, effective sample size falls. Increase particle count, improve the proposal, temper the likelihood, reduce the information entering one update, or resample adaptively; merely normalizing the same weights cannot restore coverage.
If resampling is too frequent, low-weight but legitimate modes may disappear. Raise the resampling threshold, use a lower-variance resampler, add move/rejuvenation steps, or improve proposal diversity.

Knowledge Transfer¶

The method transfers literally across Bayesian filtering practices when the roles remain intact. In radar tracking, a particle is a possible target state; in robot localization, a possible pose; in computer vision, a possible contour or motion state; in econometrics, a possible latent regime or volatility state. Transition models, observations, likelihoods, weights, and resampling all keep the same mathematical meaning. This is exact in-domain transfer even though the physical sensors and state coordinates differ.

Relationships to Other Abstractions¶

Current abstraction Particle Filter Domain-specific

Parents (3) — more general patterns this builds on

Particle Filter is a kind of Monte Carlo Simulation Prime

monte_carlo_simulation (confirmed; primary parent). A particle filter is a kind of Monte Carlo simulation: it replaces an analytically difficult probability distribution with estimates computed from random samples.
Particle Filter is part of Bayesian Updating Prime

bayesian_updating (confirmed; constitutive relation). The observation-weighting step implements repeated Bayesian correction: the predicted distribution supplies the prior for time \(t\), the observation model supplies the likelihood.
Particle Filter presupposes State and State Transition Prime

state_and_state_transition (confirmed; prerequisite relation). Particle filtering presupposes a state representation and a transition rule that propagates it.

Hierarchy paths (10) — routes to 6 parentless roots

Particle Filter → Monte Carlo Simulation → Approximation → Representation → Abstraction

Show alternative paths (9)

Neighborhood in Abstraction Space¶

Particle Filter sits in a sparse region of the domain-specific corpus (65^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08