Forgetting of the initial distribution for nonergodic Hidden Markov Chains

Randal Douc, Élisabeth Gassiat, Benoit Landelle, Éric MoulinesThe Annals of Applied Probability 2010 · DOI

TL;DR

This paper tackles the problem of "forgetting" the initial state in non-ergodic Hidden Markov Models (HMMs). The authors introduce a novel set of conditions that generalize existing results, proving that the filter's output eventually becomes independent of its starting point, both almost surely and in expectation. This is crucial for reliable inference in complex systems where initial conditions are often unknown or uncertain.

What Problem Does This Paper Address?

Hidden Markov Models (HMMs) are powerful tools for modeling systems where underlying states are not directly observable, but influence observable outputs. A key property for HMM filters (which estimate the hidden state from observations) is "forgetting the initial distribution." This means that after a sufficient amount of time, the filter's estimate should no longer depend on its initial guess of the hidden state. This is particularly important in non-ergodic HMMs, where the underlying Markov chain doesn't necessarily converge to a stationary distribution. Existing methods often require strong mixing conditions that are not always met in practical scenarios, especially for non-linear state-space models. This paper aims to provide more general conditions under which this forgetting property holds, extending previous work and covering a wider range of models.

Methods

The core of the paper lies in developing a new set of conditions for establishing the forgetting property. The authors introduce the concept of a "Local Doeblin set function" (LD-set function), which is a generalization of LD-sets that accounts for the dependence on successive observations. They then derive two main theorems:

  • Theorem 5: Establishes almost sure convergence of the total variation distance between filters initialized with different distributions. This means the difference between the filters' probability distributions vanishes with high probability as time goes to infinity.
  • Theorem 6: Provides a bound for the expectation of this total variation distance, leading to convergence in expectation.

These theorems are supported by detailed mathematical proofs involving probability theory, analysis of Markov chains, and inequalities like the Birkhoff contraction inequality. The authors also analyze specific examples of nonlinear state-space models to illustrate the applicability of their conditions.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis and general models rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Generalizing Forgetting Conditions: The introduction of LD-set functions is a key innovation, allowing for more flexible conditions than previous approaches that relied on strict mixing properties.
  • Pathwise vs. Expectation Convergence: The paper demonstrates that the filter's forgetting property can be established both in an almost sure sense (pathwise convergence) and in expectation, providing a more complete understanding of filter behavior.
  • Applicability to Non-Ergodic Models: The work explicitly addresses non-ergodic HMMs, which are common in real-world applications like tracking and financial modeling, where traditional ergodic assumptions fail.
  • Weakening Assumptions: The proposed conditions are shown to be more general than those in prior work, particularly for nonlinear state-space models, making the results applicable to a broader class of problems.

Key Results

The paper establishes theoretical results for the forgetting property of HMM filters under novel conditions. Specifically:

  • Theorem 5 guarantees that for a broad class of non-ergodic HMMs, the total variation distance between filters initialized with different distributions converges to zero almost surely.
  • Theorem 6 provides a bound on the expected total variation distance, leading to convergence in expectation.
  • The paper demonstrates the application of these results to nonlinear state-space models, showing that forgetting occurs even with dependent state noise and when observations do not perfectly match the model.

Limitations and Future Work

  • Abstract Conditions: While general, the conditions for forgetting are presented in an abstract mathematical form. The paper suggests that for specific filtering equations, these conditions can be made more concrete, which is explored in Section 3.
  • Model Mismatch: The paper touches upon the case where observations might not be generated by the assumed model, indicating robustness, but a deeper exploration of model mismatch could be a future direction.
  • Geometric Rates: While Theorem 6 provides convergence in expectation, it doesn't directly yield a geometric rate. Section 3 aims to derive such rates for specific models.
  • Further Examples: The paper presents examples of nonlinear state-space models. Exploring other classes of non-ergodic HMMs and demonstrating the applicability of the new conditions would be valuable future work.

Summary updated

Abstract

In this paper, the forgetting of the initial distribution for a nonergodic Hidden Markov Models (HMM) is studied. A new set of conditions is proposed to establish the forgetting property of the filter. Both a pathwise and mean convergence of the total variation distance of the filter started from two different initial distributions are obtained. The results are illustrated using a generic nonergodic state-space model for which both pathwise and mean exponential stability is established.
BibTeX
@article{douc2010forgetting,
  author = {Randal Douc and Élisabeth Gassiat and Benoit Landelle and Éric Moulines},
  title = {{Forgetting of the initial distribution for nonergodic Hidden Markov Chains}},
  journal = {The Annals of Applied Probability},
  year = {2010},
  doi = {10.1214/09-aap632},
}
Figure