Particle Learning Algorithm

• Particle learning is a sequential Monte Carlo method for joint state and parameter inference, leveraging recursive sufficient statistics.
• The method employs an exact resample-propagate cycle to update states, compute predictive likelihoods, and smoothly integrate additive functionals.
• Advanced extensions like PaRIS and PPG enable bias reduction and uniform ergodicity, outperforming traditional particle filters in efficiency and stability.

Particle learning (PL) is a fully-adapted sequential Monte Carlo (SMC) methodology that provides joint state and parameter filtering as well as smoothing in general state-space models. It achieves simultaneously low computational complexity, flexibility in model specification, and high Monte Carlo efficiency by augmenting the standard particle filter with recursive conditional sufficient statistics for static parameters and by employing an exact resample-propagate order. PL forms the core for a family of algorithms spanning both classical settings (latent Markov models, state-space models, hidden Markov models), online or parallelized learning scenarios, and recent developments in interactive or bias-reduced particle systems (Carvalho et al., 2010, McAlinn et al., 2012, Cardoso et al., 2023, Marks et al., 14 Oct 2025).

1. Formal Framework and Model Assumptions

Consider a state-space model defined by latent Markov states $x_{1:T}$, observed data $y_{1:T}$, and a static parameter vector $\theta$. The generative process is: $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$ The target is the sequence of filtering distributions $p(x_t, \theta \mid y^{1:t})$ and the joint smoothing posterior $p(x_{1:T},\theta \mid y_{1:T})$.

PL assumes the existence of low-dimensional conditional sufficient statistics $s_t$ for $\theta$ (common in exponential family and conjugate models), along with the ability to evaluate predictive likelihoods $p(y_{t+1}\mid x_t,\theta)$ and perform exact conditional simulation from $p(x_{t+1}\mid x_t, \theta, y_{t+1})$ (Carvalho et al., 2010).

2. Algorithmic Structure and Key Update Steps

At each time $y_{1:T}$0, particle learning maintains a sample of $y_{1:T}$1 triples $y_{1:T}$2 approximating $y_{1:T}$3. The algorithm proceeds as follows (Carvalho et al., 2010, McAlinn et al., 2012):

1. Initialization ($y_{1:T}$4): For each $y_{1:T}$5, draw $y_{1:T}$6, $y_{1:T}$7, and initialize sufficient statistics $y_{1:T}$8.
2. Resample: For the new observation $y_{1:T}$9, compute predictive weights

$\theta$0

then resample the particle set with probabilities proportional to these weights, yielding $\theta$1.

1. Propagate State: For each resampled particle,

$\theta$2

2. Update Sufficient Statistics: Update

$\theta$3

where $\theta$4 is a model-specific recursive mapping.

3. Parameter Update: Draw

$\theta$5

Each "resample–propagate" cycle yields equal-weighted particles for the filtering distribution, with $\theta$6 computational cost per step.

3. Smoothing, Additive Functionals, and Extensions

PL enables pathwise smoothing via a backward-sampling step. Post-filtering, draw $\theta$7 by index sampling; for $\theta$8, select $\theta$9 from $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$0 with weights $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$1, thereby generating a single sample exactly from $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$2 (Carvalho et al., 2010). This backward-sampling requires $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$3 time naively, though optimized implementations can achieve $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$4 using advanced ancestor tracing.

For smoothed additive functionals (e.g., score increments, EM updates), specialized schemes such as PaRIS (Particle Rapid Incremental Smoother) efficiently approximate expected sums $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$5 online, using per-particle backward statistics $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$6 based on sampled mini-ancestors from the backward kernel. The Parisian Particle Gibbs (PPG) extension embeds PaRIS inside a conditional SMC framework, producing bias-reduced, uniformly ergodic smoothing estimates with exponentially decaying bias in the number of sweeps $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$7 and variance scaling as $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$8 (Cardoso et al., 2023).

4. Parallelization and Computational Considerations

Full parallelization of PL is feasible and highly effective for large-scale models and devices such as GPUs. Algorithmic innovations include parallel prefix-sums for CDF construction, cut-point based parallel multinomial resampling, and fully vectorized propagation and parameter updates. This enables a complete PL cycle to be implemented as a sequence of GPU kernels, allowing all computations to remain on device and minimizing host-device data transfer bottlenecks (McAlinn et al., 2012).

Empirical benchmarks demonstrate 20–30× speedup for particle learning cycles on GPUs (e.g., NVIDIA GTX580 vs. quad-core CPU), with speedups up to 242× for CDF construction and ~45× for propagation and update phases when $$x_1 \sim p(x_1\mid\theta), \quad x_{t+1} \sim p(x_{t+1}\mid x_t, \theta), \quad y_{t+1} \sim p(y_{t+1}\mid x_{t+1}, \theta).$$9–$p(x_t, \theta \mid y^{1:t})$0. Even double-precision execution on GPU remains 5–10× faster than single-precision on CPU for these tasks.

5. Variants and Advanced Interacting Particle Systems

The particle learning paradigm extends to general “interacting particle” approaches for posterior or marginal likelihood estimation, particularly in models with intractable posteriors. Algorithms such as Interacting Particle Langevin Dynamics (Energy-Based IPLA) define particle evolution by kernelized overdamped Langevin SDEs: $p(x_t, \theta \mid y^{1:t})$1 where $p(x_t, \theta \mid y^{1:t})$2 is a smooth kernel and $p(x_t, \theta \mid y^{1:t})$3 the log-posterior gradient (Marks et al., 14 Oct 2025). Discretization via the Euler–Maruyama scheme and repeated updates yields particle-based approximations of the smoothing distribution and unbiased maximum marginal likelihood gradients. The theoretical guarantees include $p(x_t, \theta \mid y^{1:t})$4 mean-field convergence and $p(x_t, \theta \mid y^{1:t})$5 complexity for achieving $p(x_t, \theta \mid y^{1:t})$6-accurate gradient estimation.

Particle learning also accommodates bias reduction through conditional trajectories as in PPG, yielding sub-Gaussian deviation tails and uniform ergodicity (Cardoso et al., 2023).

6. Performance, Efficiency, and Comparisons

PL achieves substantially higher effective sample size (ESS) and lower estimator variance compared to classical particle filters with naive or kernel-shrinkage parameter inclusion (e.g., Liu–West, bootstrap, auxiliary particle filters). Parameter learning in standard particle filters quickly suffers from particle impoverishment ("freezing"), while PL’s use of sufficient statistics and resample–propagate ordering mitigates degeneracy (Carvalho et al., 2010).

Compared to MCMC-based forward-filter backward-sample (FFBS) algorithms, PL requires only a single forward and backward pass, conferring substantial computational advantages without the need for convergence diagnostics. Empirical and theoretical analyses consistently demonstrate superior scaling and stability for PL in both state and parameter learning contexts.

7. Practical Implementation, Limitations, and Tuning

PL requires:

• Conditional sufficient statistics $p(x_t, \theta \mid y^{1:t})$7 for static parameters,
• Ability to compute predictive likelihoods and perform exact propagation,
• Sufficient particle count $p(x_t, \theta \mid y^{1:t})$8, typically $p(x_t, \theta \mid y^{1:t})$9–$p(x_{1:T},\theta \mid y_{1:T})$0 for low-dimensional problems,
• Model-specific update functions $p(x_{1:T},\theta \mid y_{1:T})$1.

If conjugate updates are unavailable, one can embed a Gibbs or Metropolis–Hastings step inside each particle’s parameter update, at some cost to computational efficiency. For non-conjugate or auxiliary-variable models (mixtures), the methodology extends by sampling auxiliary variables synchronously in the resample or propagate steps.

Approximate or numerically-integrated densities introduce adaptation imperfections and potential variance inflation. Scaling to high-dimensional states may require increasing $p(x_{1:T},\theta \mid y_{1:T})$2 proportionally to state dimension and inverse signal-to-noise ratio.

---

In summary, particle learning and its descendants offer a principled SMC framework for joint online state and parameter inference in state-space models, with smoothing provided as an immediate by-product and efficiency benefits over both standard particle filters and MCMC methodologies. Extensions to parallel and advanced particle systems further expand their applicability and computational efficiency (Carvalho et al., 2010, McAlinn et al., 2012, Cardoso et al., 2023, Marks et al., 14 Oct 2025).