Supplement paper to ‘Nonparametric estimation in hidden Markov models’

Thierry Dumont, Sylvain Le Corff2012

TL;DR

This paper provides supplementary proofs and algorithmic details for a previously published work on nonparametric estimation in hidden Markov models. It delves into the mathematical underpinnings of estimating an unknown function within a partially observed Markov chain, offering rigorous justifications for the proposed methods.

What Problem Does This Paper Address?

The paper addresses the challenge of estimating an unknown target function, $f$, that maps from a compact set $K$ to $\mathbb{R}^\ell$. This estimation occurs in the context of observing a process $\{Y_k\}_{k \in \mathbb{N}}$, where each observation $Y_k$ is a noisy version of $f$ applied to a hidden Markov state $X_k$, plus some Gaussian noise $\epsilon_k$. The core difficulty lies in the fact that the states $X_k$ are not directly observed, making standard function estimation techniques insufficient. This is a common problem in fields like signal processing, econometrics, and bioinformatics where underlying states influence observable data.

Methods

This paper is primarily a theoretical supplement, focusing on providing detailed proofs for results presented in a prior publication. The core methodology revolves around:

  • Recalling the Model and Assumptions: Section 1 meticulously restates the statistical model, including the definition of the observed process $\{Y_k\}$, the hidden Markov chain $\{X_k\}$, the noise distribution, and the properties of the target function $f$ and the state space $K$. Key assumptions like $K$ being compact and homeomorphic to a convex subset, and the properties of the transition kernel $Q_a$, are crucial.
  • Mathematical Proofs: Section 2 is dedicated to proving technical results, particularly those related to concentration inequalities for estimators. This involves leveraging tools from probability theory, such as the Stein Theorem and results from Adamczak and Bednorz (2012), to bound the error of the estimation process. The proofs often involve intricate manipulations of expectations, suprema, and probability measures.
  • Algorithm Details: Section 3 elaborates on the Expectation-Maximization (EM) algorithm used for estimation. It details how the algorithm iteratively refines an estimate of the target function $f$ by maximizing a Q-function. This section includes the derivation of differential equations that guide the update steps of the EM algorithm and discusses numerical approximations for solving these equations.

Datasets

Not discussed in the paper. The paper focuses on theoretical proofs and algorithmic descriptions rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Rigorous Justification for Nonparametric Estimation: The paper provides the mathematical backbone for estimating complex, unknown functions in the presence of hidden states. This is crucial for building confidence in the reliability of such methods.
  • Concentration Inequalities as a Cornerstone: The proofs heavily rely on establishing strong concentration inequalities. These inequalities quantify how likely the estimated function is to be close to the true function, which is fundamental for statistical inference.
  • The Power of EM for Hidden Models: The detailed explanation of the EM algorithm highlights its adaptability to complex statistical models with latent variables, demonstrating how it can be tailored for nonparametric function estimation in HMMs.
  • Bridging Theory and Computation: By providing both theoretical proofs and algorithmic details, the paper connects the abstract mathematical framework to practical implementation, which is essential for applied researchers.

Key Results

The paper's primary contribution is the provision of detailed proofs for key theoretical results, including:

  • Proposition 2.1: Establishes concentration inequalities for the estimator, showing that the error is bounded and decays appropriately with sample size. This is a critical result for establishing the statistical properties of the estimation procedure.
  • Lemma 2.2 & 2.3: These lemmas provide essential bounds on intermediate quantities (like envelope functions) that are used in the proofs of Proposition 2.1.
  • Algorithm Derivations: Section 3 derives the specific update rules for the EM algorithm, including the differential equations and their solutions, which are necessary for implementing the nonparametric estimation.

Quantitative results are presented in the form of inequalities and bounds (e.g., $P_z[\sup_{g \in G_M} |\nu_n(g)| \geq c E[\sup_{g \in G_M} |\nu_n(g)|] + M t] \leq K_1 e^{-K_2 t^2} + e^{-K_2 t}$), demonstrating the theoretical guarantees of the method.

Limitations and Future Work

  • Focus on Theoretical Foundations: The paper is a supplementary document, meaning its primary goal is to provide proofs for existing work. It does not introduce new empirical results or extensive numerical experiments.
  • Assumptions on the Model: The proofs rely on specific assumptions about the target function $f$ (e.g., belonging to a Sobolev space), the state space $K$, and the noise distribution. The performance of the method in scenarios violating these assumptions is not explored.
  • Computational Complexity: While Section 3 details the EM algorithm, the practical computational cost and scalability for very large datasets or high-dimensional problems are not deeply analyzed.

Future work, as implied by the nature of such supplementary papers, would likely involve extending these theoretical results to more complex scenarios, exploring alternative estimation algorithms, and conducting more extensive empirical studies to validate the method's performance in real-world applications.

Summary updated

Abstract

This document is a supplementary material to the article “Nonparametric estimation using partially observed Markov chains”. It provides additional proofs of some technical results given in the original paper. Section 1 recalls the model, the definitions and the assumptions used in the paper. Section 2 provides proofs of some results stated in the paper and Section 3 gives details on the algorithm used to perform the Expectation-Maximization based estimation. 1 Model and definitions In this section, we recall the model and the assumptions given in Dumont and Le Corff [2012]. Comments on these assumptions can be found in [Dumont and Le Corff, 2012, Section 2]. Let ℓ and m be positive integers and K be a subset of R m. The main statistical problem considered in this paper is the estimation of an unknown target function f ⋆ : K → R ℓ when observing a process {Yk} k∈N such that for any k ≥ 0, Yk belongs to R ℓ and satisfies Yk def = f⋆(Xk) + ɛk. {ɛk} k∈N is assumed to be an i.i.d Gaussian process with common distribution N (0, σ 2 Iℓ), Iℓ being the identity matrix of size ℓ and σ 2 a fixed positive parameter. Denote by ϕ the probability distribution of ɛ0, i.e. ∀z ∈ R ℓ, ϕ(z) def
BibTeX
@misc{dumont2012supplement,
  author = {Thierry Dumont and Sylvain Le Corff},
  title = {{Supplement paper to ‘Nonparametric estimation in hidden Markov models’}},
  year = {2012},
}
Figure