Hidden Markov Model Where Higher Noise Makes Smaller Errors

TL;DR

This paper analyzes parameter estimation in a linear Gaussian system with small, possibly unequal, noise levels, revealing that larger state noise can paradoxically lead to smaller estimation errors, supported by asymptotic analysis.

Contribution

It demonstrates that both MLE and Bayes estimators are consistent, asymptotically normal, and efficient in a model where increased state noise reduces estimation error, an unusual phenomenon.

Findings

Estimators are consistent and asymptotically normal.

Larger state noise can decrease estimation errors.

Asymptotic properties are derived using Kalman-Bucy filter analysis.

Abstract

We consider the problem of parameter estimation in a partially observed linear Gaussian system with small noises in the state and observation equations. We describe asymptotic properties of the MLE and Bayes estimators in the setting with state and observation noises of possibly unequal intensities. It is shown that both estimators are consistent, asymptotically normal with convergent moments and asymptotically efficient. This model has an unusual feature: larger noise in the state equation yields smaller estimation error. The proofs are based on asymptotic analysis of the Kalman-Bucy filter and the associated Riccati equation in particular.