TL;DR
This paper establishes the strong consistency of the Maximum Likelihood Estimator (MLE) for a broad class of Hidden Markov Models (HMMs), even when dealing with complex, non-compact state spaces. The authors introduce a novel information-theoretic approach to prove identifiability, which bypasses the need for explicit relative entropy rate calculations and offers a more general framework for analyzing dependent time series.
What Problem Does This Paper Address?
Hidden Markov Models (HMMs) are powerful tools for modeling sequential data where the underlying states are not directly observable. They find applications across diverse fields like speech recognition, finance, and biology. A fundamental challenge in HMMs is parameter estimation, particularly using the Maximum Likelihood Estimator (MLE). While MLE is a cornerstone of statistical inference, proving its consistency (i.e., that it converges to the true parameter as more data becomes available) for general HMMs, especially those with non-compact state spaces, has been notoriously difficult. Existing methods often rely on restrictive assumptions like uniform positivity of transition densities, which are rarely met in practice. This paper aims to bridge this gap by providing a general consistency result for MLE in HMMs under more relaxed and practical conditions.
Methods
The authors develop a novel proof strategy for MLE consistency in general HMMs. Instead of directly proving the existence and identifiability of the limiting relative entropy rate (a common but challenging step), they focus on showing that the log-likelihood is asymptotically bounded away from the true parameter's log-likelihood. This is achieved through an information-theoretic technique for proving identifiability that avoids explicit relative entropy rate representations. A key assumption enabling this approach is that an iterate of the state transition kernel possesses a density with respect to a $\sigma$-finite measure. The proof also leverages a general concentration inequality for $V$-uniformly ergodic Markov chains, which is of independent interest. The authors then demonstrate the applicability of their main result to linear Gaussian state space models, finite state models, and nonlinear state space models.
Datasets
Not discussed in the paper. The paper focuses on theoretical results and does not present empirical evaluations on specific datasets.
Key Insights and Ideas
- Bypassing Explicit Entropy Rate: The most significant insight is the development of an information-theoretic technique for proving identifiability that doesn't require an explicit representation of the relative entropy rate. This is a major simplification and generalization.
- Focus on Boundedness: Instead of proving convergence of the log-likelihood, the authors show it's sufficient to bound it from above by a sequence that is asymptotically separated from the true parameter's log-likelihood.
- Generalization to Non-Compact Spaces: The framework is designed to handle HMMs where state and observation spaces are complete separable metric spaces, significantly broadening its applicability beyond traditional finite-state models.
- Power of $V$-Uniform Ergodicity: The paper highlights the utility of $V$-uniform ergodicity for Markov chains, showing it's sufficient to establish crucial properties like large deviations and identifiability.
- Foundation for Broader Time Series: The techniques developed are presented as a potential foundation for consistency proofs in other complex dependent and non-Markovian time series models.
Key Results
The main result is the proof of strong consistency for the Maximum Likelihood Estimator (MLE) of the parameter in general Hidden Markov Models. This consistency holds under a set of assumptions that are shown to be satisfied in several important classes of models, including:
- Linear Gaussian state space models.
- Finite state models.
- Nonlinear state space models (e.g., vector ARCH type, stochastic volatility models).
The paper also provides a general concentration inequality for $V$-uniformly ergodic Markov chains.
Limitations and Future Work
- Assumptions: While the assumptions are described as "rather minimal," they still include conditions like the existence of a density for an iterate of the transition kernel and $V$-uniform ergodicity. Verifying these in specific complex models might still require effort.
- Sub-probabilities: The paper focuses on parameters that correspond to probability measures. Generalizing the results to allow for sub-probabilities (which can arise in "infinite" parameter spaces) would be a natural extension.
- Asymptotic Normality: The paper establishes consistency, but the asymptotic normality of the MLE is left for future investigation.
- Broader Time Series: The authors explicitly suggest that the techniques could be extended to other dependent and non-Markovian time series models, which represents a significant avenue for future research.
Summary updated