Applied Mathematics and Computation 283 (2016) 22–28
Contents lists available at ScienceDirect
Applied Mathematics and Computation
journal homepage: www.elsevier.com/locate/amc
Statistical tracking behavior of affine projection algorithm for
unity step size
Yongfeng Zhi , Yunyi Yang , Xi Zheng , Jun Zhang , Zhen Wang
∗
School of Automation, Northwestern Polytechnical University, Xi’an 710072, China
a r t i c l e i n f o
Keywords:
Affine projection algorithm
Tracking behavior
System identification
Step size
a b s t r a c t
Since unity step size could guarantee the fastest convergence and more detailed analysis
for the affine projection (AP) algorithm, a statistical tracking behavior of AP algorithm is
discussed in this paper. Deterministic recursive equations are derived for the mean weight
error and mean-square error. All the possible correlations between the adaptive filtering
coefficients and the past measurement noise are considered as well.
©2016 Elsevier Inc. All rights reserved.
1. Introduction
Over the past decades, many computationally efficient, rapidly converging adaptive filtering algorithms have been nu-
merously proposed across a myriad of engineering realms. Among all these algorithms, the affine projection (AP) algorithm,
discovered from the geometric viewpoint of affine subspace projections, attracts particular attention, both theoretical and
experimental [1] . Based on the direction vector, a new definition for the AP algorithm was presented when the unity step
size was incorporated [2] . To achieve both the fastest convergence rate and the lowest steady-state error, the optimal step
size AP algorithm was also discussed [3] . By setting the weight error to be zero in the direction of the adaptive weight
update, the optimal step size was obtained for the pseudo-AP (PAP) algorithm [4] . By considering the measurement noise
influence on the steady-state error, the first moment estimation of the measurement noise was used to modify the variable
step size update, and then a modified variable step size affine projection sign algorithm was proposed in [5] . Along this
framework, there still exist a great number of excellent progresses, such as normalized least mean square algorithm with
orthogonal factors [6] , AP with direction error algorithm [7, 8] , AP algorithm dynamically selecting input vectors [9] , to name
yet a few.
Except for the accumulated achievements of updating of algorithms itself, analysis of its statistical properties also be-
comes a hot issue recently. Here we can mention some examples for capturing a clearer knowledge. In Ref. [10] , a statistical
analytical model for predicting the stochastic behavior of the AP algorithm has been provided for autoregressive (AR) inputs.
To extend applicability of input processes that is suitable AR as well as autoregressive-moving average, a new statistical
analysis framework was added to the AP algorithm with unity step size [11] . If the step size of PAP algorithm was less that
one, deterministic recursive equations were derived for the mean weight error and for the mean-square error (MSE) [12,13] .
Moreover, a quantitative analysis of the AP algorithm was presented [14] , which analyzed the mean weight error and the
MSE based on an independent and identically distributed input signal. In [15,16] , the correlations between the weight error
and the past measurement noise was presented to analyze the AP algorithm.
∗
Corresponding author.
E-mail addresses: yongfeng@nwpu.edu.cn (Y. Zhi), zhangjun@nwpu.edu.cn (J. Zhang), zhenwang0@gmail.com (Z. Wang).
http://dx.doi.org/10.1016/j.amc.2016.02.003
0 096-30 03/© 2016 Elsevier Inc. All rights reserved.