Remarks on filtering error due to quantisation of a 2-state hidden Markov model
Authors:
- L. ShueLSL. Shue
- S. DeySDS. Dey
- B.D.O. AndersonBAB.D.O. Anderson
- F. De BruyneFBF. De Bruyne
- 22 January 2003
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 4, 4123-4124
Abstract
This paper considers the problem of quantisation from the perspective of minimising filtering error when quantised in - stead of continuous measurements are used as inputs to the nonlinear filter, specialising to a discrete-time 2-state h id- den Markov model (HMM) with continuous output. In communication systems, a fundamental problem is the question of how much degradation in performance a re- ceiver suffers when incoming data are quantised. In this brief note we will consider the effects of quantisation on the performance of a nonlinear filter, with the signal source assuming the form Yk h Xk vk. Here Xk is a 2-state Markov chain, with h Xk 1 1 and vk is independent and identically distributed (i.i.d.) noise. The function o f the filter is to estimate Xk given quantised versions of Yk. Much of previous work dealing with quantisation aims to reduce the distortion between the original and quantised signals under some minimum mean-squared-error criteria (2, 3, 6). In contrast, the primary focus here is to minimise the probability of error of a nonlinear filter with quantised observations as inputs. Some of the questions we address are: (1) Is there an optimal strategy for choosing the quan- tisation levels which minimises the filtering error? (2) How does filtering error vary with noise for different number of quantisation levels M? At this stage we have only partial answers to both questions. For the first question, we provide two quantisation schemes, based on the maximisation of the mutual information and Kullback-Leibler (KL) divergence measures respectively. A comparison of these schemes will be made by simulations. We also provide an analysis for computing the maximum a posteriori (MAP) error as a function of signal-to-noise ratio (SNR), for a 2-state discrete-time HMM with continuous- range output (this is a discrete-time version of a continuou s- time result obtained by Wonham (5)).Keywords
This publication has 6 references indexed in Scilit:
- Elements of Information Theory2001
- A tutorial on hidden Markov models and selected applications in speech recognitionProceedings of the IEEE, 1989
- On universal quantizationIEEE Transactions on Information Theory, 1985
- Least squares quantization in PCMIEEE Transactions on Information Theory, 1982
- Some Applications of Stochastic Differential Equations to Optimal Nonlinear FilteringJournal of the Society for Industrial and Applied Mathematics Series A Control, 1964
- Quantizing for minimum distortionIEEE Transactions on Information Theory, 1960