Elsevier

Signal Processing

Volume 81, Issue 5, May 2001, Pages 1053-1059
Signal Processing

Short communication
Convergence behavior of the LMS algorithm in subband adaptive filtering

https://doi.org/10.1016/S0165-1684(00)00263-2Get rights and content

Abstract

Subband adaptive filtering with an LMS-type algorithm has been considered as a possible alternative to the conventional fullband adaptive filtering. However, its convergence behavior has not yet been sufficiently analyzed. In this paper, an approximate expression for the convergence behavior of the cancellation error of the critically sampled subband adaptive digital filters (ADF) is presented. It is represented in the frequency domain and enables us to infer whether the subband ADF is efficient in terms of convergence speed. Computer simulations are presented to see the validity of the expression.

Introduction

In applications such as adaptive echo canceler and adaptive noise control, adaptive filtering with an LMS-type algorithm suffers from slow convergence and high computational complexity [4], since the adaptive filter requires hundreds or thousands of taps and the input signal is often highly correlated. Subband adaptive filtering has been proposed to avoid these drawbacks [1], [2], [3].

The convergence speed of the subband adaptive digital filters (ADF) with LMS-type algorithms was investigated, respectively, in [6], [9] by means of the condition numbers of correlation matrices of the subband tap-input vectors. It is said that, since the condition numbers of correlation matrices of the subband tap-input vectors are smaller than those for the corresponding fullband ADF, the subband ADF would converge faster and hence the subband ADF is attractive in convergence speed. However, it has also been shown by some computer simulations that the cancellation error of the subband ADF does not always decay faster than that of the fullband ADF.

To get more insights into the convergence speed of the subband ADF, we derive a theoretical but approximate expression that describes the convergence behavior of the cancellation error of the critically sampled subband ADF. It is shown that the critically sampled subband ADF do not always have advantages over the fullband ADF in convergence speed. Computer simulations are presented to see the validity of the expression.

Section snippets

A review of subband adaptive digital filters

First of all, let us review subband ADF with the LMS algorithm. Throughout this paper, signals and coefficients of filters are assumed to be real. We only consider the critically sampled subband ADF with M-channel paraunitary filter banks [8]. Let us denote the delay associated to the filter bank as D and express the z-transform of a sequence denoted by a small letter by the corresponding capital letter.

Fig. 1 shows a block diagram of the subband ADF where the decimation rate is two for

Convergence behavior of critically sampled subband ADF

Let us consider the fullband ADF. We assume that the unknown system has a finite impulse response of length L and denote the correlation function and the spectral density of the input signal as r(n) and as S(ejω), respectively.

The tap-weight vector w(n) is adjusted byw(n+1)=w(n)+με(n)x(n),where x(n) is the tap-input vector at time n, μ is a convergence parameter, ε(n) is the adaptation (cancellation) error given byε(n)=d(n)−wT(n)x(n).

Let us assume that the convergence parameter is small enough

Numerical examples

Simulation results and numerical examples are presented to confirm our analysis. We considered the normalized LMS algorithm with μ̃=μ̃0=⋯=μ̃M−1=0.5. The frequency response of the unknown system F(z) is plotted in Fig. 2. The length of the impulse response of the unknown system was 512. We put the length of the fullband adaptive filter to be 512. As input signals, first-order AR processes with coefficients 0.9 and −0.5 driven by white processes were used. The output signals of the unknown system

Conclusions

An approximate expression that describes the convergence behavior of the cancellation error of the critically sampled subband ADF has been derived. From the expression, the convergence speed of the subband ADF has been investigated. It has been shown theoretically and empirically that the critically sampled subband ADF does not always have advantages over the fullband ADF in convergence speed. Computer simulations have also been presented to see the validity of the expression.

Acknowledgements

The authors wish to thank the anonymous reviewers for their valuable comments and suggestions.

References (9)

  • Y.G. Yang et al.

    On the performance analysis and applications of the subband adaptive digital filter

    Signal Processing.

    (February 1995)
  • A. Gilloire, M. Vetterli, Adaptive filtering in subbands, Proceedings of the International Conference on Acoustics and...
  • A. Gilloire et al.

    Adaptive filtering in subband with critical sampling: analysis, experiments and application to acoustic echo cancellation

    IEEE Trans. Signal Process.

    (August 1992)
  • W. Kellermann, Analysis and design of multirate systems for cancellation of acoustical echoes, Proceedings of the...
There are more references available in the full text version of this article.

Cited by (12)

  • A brief survey of time- and frequency-domain adaptive filters

    2016, IISA 2016 - 7th International Conference on Information, Intelligence, Systems and Applications
  • A review of advances in subband adaptive filtering

    2013, World Applied Sciences Journal
View all citing articles on Scopus
View full text