Stability of Sequential Monte Carlo in General State Spaces

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 227 papers32 conceptsfrozen 2026-10-04

Sequential Monte Carlo (SMC) methods, or particle filters, are the workhorse of non-linear state-space modeling. They allow us to track hidden variables in complex systems by maintaining a "cloud" of particles that represent the probability distribution of the state. However, for decades, a gap persisted between the practical success of these algorithms and our theoretical understanding of why they work.

Most classical proofs of stability—the guarantee that errors won't explode over time—rely on "strong mixing" assumptions. In essence, these proofs assume the system is "well-behaved": it must stay within a bounded (compact) region and forget its initial starting point almost instantly. But real-world systems, from tracking a satellite in open space to modeling financial markets, often involve unbounded variables and heavy-tailed noise. This research direction focuses on extending the theoretical foundations of SMC to these "general" state spaces, ensuring that our algorithms remain reliable even when the math gets messy.

The Problem of Memory and Mixing

The fundamental requirement for a stable particle filter is that it must "forget" its initial distribution. If a small error in the starting guess persists forever, the filter will eventually diverge from the truth.

In On the use of sequential Monte Carlo methods for approximating smoothing functionals, Jimmy Olsson, Eric Moulines, and colleagues point out that standard convergence proofs break down when transition densities aren't bounded away from zero. If a state can wander off toward infinity, the system might not "mix" fast enough to reset its hereditary effects. Nicolas Chopin formalized this in Central limit theorem for sequential Monte Carlo methods and its application to Bayesian inference, showing that in unbounded domains, the variance of particle estimates can diverge, potentially invalidating long-term stability.

Breaking the Compactness Constraint

To bridge this gap, researchers have developed more flexible mathematical frameworks. A landmark contribution by Randal Douc, Eric Moulines, and others in Forgetting of the initial distribution for Hidden Markov Models introduced the Local Doeblin property. Instead of requiring the system to mix uniformly across the entire universe, they proved that as long as the system behaves well "locally," it can still achieve geometric convergence to the true state.

Similarly, Patrice Bertail utilized the "regenerative method" in Sharp Bounds for the Tails of Functionals of Markov Chains to derive exponential tail inequalities. This allowed for proofs that mirror classic independent-data results without requiring the restrictive "uniform ergodicity" that previously limited theory to bounded boxes.

Stability in Smoothing and Derivatives

While filtering (estimating the current state) is difficult, smoothing (estimating the entire trajectory) and parameter learning (estimating the model itself) are even more sensitive to instability.

In Non-asymptotic deviation inequalities for smoothed additive functionals in nonlinear state-space models, Cyrille Dubarry demonstrated that for popular smoothing algorithms like FFBS (Forward Filtering Backward Smoothing), the error scales with the ratio of time to the number of particles ($T/N$). This was a major improvement over previous bounds, but as Dubarry notes, these guarantees still often hinge on the system's ability to forget its past quickly.

The challenge extends to the derivatives of these filters, which are essential for maximum likelihood estimation. Pierre Del Moral and Sumeetpal S. Singh provided the first rigorous proof of "uniform stability" for filter derivatives in Uniform Stability of a Particle Approximation of the Optimal Filter Derivative. They showed that by using a backward-simulation approach, one can prevent the errors in the gradient from blowing up as the data sequence grows longer.

Modern Frontiers: Amortization and Online Learning

Recent work has moved toward making these theoretical guarantees compatible with modern machine learning architectures like Variational Autoencoders (VAEs). Mathis Chagneux and Sylvain Le Corff addressed this in Amortized backward variational inference in nonlinear state-space models. By employing a "backward factorization," they proved that estimation errors in deep sequential models grow only linearly over time, even in complex non-linear systems.

Furthermore, the development of algorithms like Parisian Particle Gibbs (PPG) by Gabriel Cardoso, Jimmy Olsson, and colleagues in State and parameter learning with PARIS particle Gibbs represents the cutting edge of this direction. These "many-body" samplers combine variance-reduction techniques with iterative refinement, providing non-asymptotic guarantees for parameter learning in Hidden Markov Models that were previously thought to be too complex for rigorous proof.

The overarching goal of this field remains the same: to ensure that as our models move into higher dimensions (Ramon van Handel, Can local particle filters beat the curse of dimensionality?) and more complex domains, the mathematical floor beneath our algorithms remains solid.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 17 papers
  2. Version 1 · 2026-10-04 · 27 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_strong_mixing_assumption_502aead8,
  title = {strong mixing assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strong-mixing-assumption/rdr_502aead8dbc04d04abc6a7d65d0c1972},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_stability_of_sequential_monte_carlo_in_g_502aead8_v1,
  title = {Stability of Sequential Monte Carlo in General State Spaces},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stability-of-sequential-monte-carlo-in-general-state-spaces/rdr_502aead8dbc04d04abc6a7d65d0c1972/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}