The Kalman filter is a beautiful machine because it keeps only what the linear-Gaussian world requires: a mean and a covariance.

But the world is often rude to that bargain.

A sensor may observe a square. A radar may see a bearing but not a range. A financial time series may have jumps, heavy tails, and hidden regimes. A robot may know that it is near a wall but not which side of a corridor it is on. In those cases, the posterior is not politely summarized by one center and one ellipse.

Sometimes the posterior has two hills. Sometimes it has a long tail. Sometimes the mean points to a state that is plausible to nobody.

The particle filter starts from a more literal idea:

carry many possible worlds forward
give each world a weight
kill, copy, and perturb them as evidence arrives

That sounds crude. It is crude. Its virtue is that the crudeness is explicit. The posterior is represented by an empirical measure:

\[p(x_t \mid y_{1:t}) \approx \sum_{i=1}^N w_t^{(i)} \delta_{x_t^{(i)}}.\]

The particles are locations. The weights say how much posterior mass each location is carrying. The method is not trying to make the posterior Gaussian. It is trying to keep enough weighted hypotheses alive that the posterior shape can be approximated numerically.

Recursion Under the Crowd

In a hidden Markov model, the latent state evolves by:

\[x_t \sim f(x_t \mid x_{t-1}),\]

and observations arrive by:

\[y_t \sim g(y_t \mid x_t).\]

The exact Bayesian filter has two steps. Prediction:

\[p(x_t \mid y_{1:t-1}) = \int f(x_t \mid x_{t-1})p(x_{t-1}\mid y_{1:t-1})\,dx_{t-1}.\]

Update:

\[p(x_t \mid y_{1:t}) \propto g(y_t \mid x_t)p(x_t \mid y_{1:t-1}).\]

The equations are simple. The integral is the trap. In nonlinear or non-Gaussian models, it usually has no convenient closed form.

The bootstrap particle filter, introduced by Gordon, Salmond, and Smith in 1993, replaces the density by samples: propagate particles through the transition model, weight them by the observation likelihood, normalize, and resample when the weights become too uneven.1 Kitagawa developed related Monte Carlo filtering and smoothing methods for non-Gaussian nonlinear state-space models.2 Liu and Chen later gave a broad sequential Monte Carlo framework for dynamic systems.3

The practical algorithm is almost embarrassingly short:

for each new observation:
  propagate particles through the state dynamics
  compute likelihood weights under the observation model
  normalize weights
  estimate posterior summaries
  resample if the weights have collapsed

The embarrassment fades when the posterior is shaped like something a Gaussian cannot draw.

A Two-Signed Observation Lab

The lab below uses a standard nonlinear benchmark model:

\[x_t = 0.5x_{t-1} + \frac{25x_{t-1}}{1+x_{t-1}^2} + 8\cos(1.2t) + q\epsilon_t,\] \[y_t = \frac{x_t^2}{20} + r\eta_t.\]

The observation sees roughly \(x_t^2\), so it often cannot tell positive state from negative state. A single observation can support two signs. That is the point of the example: a line is not enough.

Try three moves:

  1. set Resample threshold to zero and watch effective sample size collapse;
  2. lower Assumed observation noise and see sharp likelihoods punish almost every particle;
  3. move Inspect time to a moment where the orange observation roots split the posterior into two plausible regions.
True state Posterior mean 80% particle band Observation roots Effective sample size Resampling event

Deterministic synthetic experiment. The state transition and observation model are nonlinear. The filter uses the bootstrap proposal: particles are propagated through the transition model, weighted by the observation likelihood, and systematically resampled when ESS / N falls below the selected threshold.

The orange dots in the state panel are the two signed roots implied by a positive observation:

\[x_t \approx \pm\sqrt{20y_t}.\]

The observation model has thrown away the sign. The dynamics and the particle history have to recover it. When the filter is uncertain, the particle posterior can hold mass on both sides. When the data and dynamics agree, one side wins.

This is the first habit particle filters teach:

do not ask only where the posterior center is
ask what shape of uncertainty the data created

The ESS Alarm Bell

Importance weights decay unevenly. After a surprising observation, a few particles may explain the data well while most explain it terribly. Then the posterior approximation is formally made of many particles but practically made of very few.

The usual diagnostic is effective sample size:

\[\mathrm{ESS} = \frac{1}{\sum_i (w_t^{(i)})^2}.\]

If all \(N\) weights are equal, ESS is \(N\). If one particle has nearly all the weight, ESS is near one. The red panel shows ESS divided by \(N\).

Set the resample threshold to zero. The filter stops copying high-weight particles. The unique initial ancestry metric stays high, but this is a technical victory with bad epistemology: most particles are alive only nominally. They are not carrying posterior mass. The mean drifts because the weighted representation has collapsed.

Raise the threshold. The filter resamples when ESS falls below the selected fraction of \(N\). Weight degeneracy improves. But resampling has its own cost: it duplicates successful particles and deletes weak ones. After many rounds, the final particles may descend from very few early ancestors. That is sample impoverishment.

This is the central tension:

without resampling: many dead weights
with too much resampling: many copied histories

Resampling is not a Bayesian update. It is a representation repair. It changes the particle set so the next update is numerically usable, while trying to leave the represented distribution unchanged.

The Proposal Decides Where Hope Goes

The lab uses the bootstrap proposal:

\[x_t^{(i)} \sim f(x_t \mid x_{t-1}^{(i)}).\]

That is convenient because it samples from the transition model and then lets the observation likelihood decide which particles mattered. It is also wasteful when observations are informative. If the sensor likelihood is sharp, most particles may be propagated to places the new observation immediately rejects.

Better proposals use the new observation when proposing particles. Auxiliary particle filters, unscented particle filters, Rao-Blackwellized particle filters, and many modern sequential Monte Carlo variants are answers to the same complaint:

do not spend particles where the next observation already says they are useless

Doucet, de Freitas, and Gordon’s edited volume helped consolidate these methods under the sequential Monte Carlo name.4 Arulampalam, Maskell, Gordon, and Clapp’s tutorial remains a practical entry point for online nonlinear and non-Gaussian tracking.5 Doucet and Johansen later gave a broader tutorial emphasizing particle filtering and smoothing as special cases of a general SMC framework.6

The important design question is not “particle filter or Kalman filter?” It is:

what proposal distribution keeps posterior mass represented
at the computational budget I actually have?

The Mean Can Point Nowhere

In the lab, the posterior mean is sometimes a useful estimate. It is also sometimes a compromise between incompatible stories. If particles sit near \(-8\) and \(+8\), the mean may sit near zero, where neither mode believes the state is likely.

This is not a minor visualization issue. Downstream decisions often consume a single number: estimated location, latent volatility, demand state, default risk, hidden intent. If the posterior is multimodal, a single posterior mean can send the decision maker into a region of low posterior density.

Particle filters let you ask better downstream questions:

  • what is the probability the state is positive?
  • what is the tail probability above a danger threshold?
  • how much mass lies in each regime?
  • what action minimizes posterior expected loss, not squared error by default?

The posterior representation is richer than the mean. Use it.

Ways the Crowd Collapses

Particle filters do not defeat dimension. In high-dimensional state spaces, the number of particles needed to represent the posterior can become brutal. This is one reason naive particle filters struggle in weather models, large latent factor systems, and many deep state-space models.

The likelihood can be too sharp. If observation noise is assumed too small, the filter becomes overconfident and murders almost every particle at each step. In the lab, lower Assumed observation noise and watch ESS fall.

The transition model can be wrong. A particle filter is not robust merely because it uses samples. If the true state moves in ways the transition model does not propose, the filter may never put particles where the posterior should be.

Parameter learning is harder than state filtering. Filtering \(x_t\) given known parameters is already approximate. Learning static parameters online can suffer from degeneracy because parameters do not naturally rejuvenate through the state transition. Particle MCMC, SMC squared, iterated filtering, and related methods exist because this problem is real.

Finally, resampling frequency is a policy. Too little resampling leaves dead weights. Too much resampling loses diversity. Adding roughening noise, move steps, better proposals, or analytic marginalization can help, but each fix is a modeling choice, not a free lunch.

Questions Before Trusting the Crowd

Before I trust a particle filter, I want to know:

  • what state-space model generated the transition and observation densities;
  • whether the proposal uses the latest observation or only the transition;
  • how ESS behaves through time, not only the final RMSE;
  • what resampling scheme is used and at what threshold;
  • whether posterior intervals have empirical coverage in simulation;
  • whether the posterior is multimodal and whether decisions use that fact;
  • how many distinct ancestors survive after repeated resampling;
  • whether the filter was tested under misspecified observation noise;
  • how performance changes with particle count;
  • whether high-dimensional structure was exploited rather than sampled blindly.

The particle filter is not magic dust sprinkled on nonlinear models. It is an accounting system for uncertainty under simulation. Every particle is a claim about where the hidden state could be. Every weight is a likelihood receipt. Every resampling step is an admission that the receipts have become too uneven to keep carrying forward.

The posterior is not always a bell curve. Sometimes it is a moving particle system, and the most honest thing you can do is keep the particles visible.

Primary Sources

  1. N. J. Gordon, D. J. Salmond, and A. F. M. Smith, “Novel approach to nonlinear/non-Gaussian Bayesian state estimation”, IEE Proceedings F, 1993. ↩

  2. Genshiro Kitagawa, “Monte Carlo Filter and Smoother for Non-Gaussian Nonlinear State Space Models”, Journal of Computational and Graphical Statistics, 1996. ↩

  3. Jun S. Liu and Rong Chen, “Sequential Monte Carlo Methods for Dynamic Systems”, Journal of the American Statistical Association, 1998. ↩

  4. Arnaud Doucet, Nando de Freitas, and Neil Gordon, eds., Sequential Monte Carlo Methods in Practice, Springer, 2001. ↩

  5. M. Sanjeev Arulampalam, Simon Maskell, Neil Gordon, and Tim Clapp, “A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking”, IEEE Transactions on Signal Processing, 2002. ↩

  6. Arnaud Doucet and Adam M. Johansen, “A Tutorial on Particle Filtering and Smoothing: Fifteen Years Later”, in The Oxford Handbook of Nonlinear Filtering, 2011. Repository metadata is also available through Warwick WRAP. ↩