New Methods for the Efficient Optimization of
Cumulant-based Contrast Functions
Wei Zhao, Yuehong Shen, Jiangong Wang, Zhigang Yuan and Wei Jian
College of Communications Engineering, PLA University of Science and Technology, Nanjing, China
Email: gmajyie@126.com
Abstract—This paper deals with efficient optimization of the
cumulant-based contrast functions. Such a problem occurs in the
blind source separation framework, where contrast functions are
criteria to be maximized to retrieve the sources. Inspired from
the recently proposed reference contrast functions, a similar one
called new kurtosis contrast function is put forward. Based on
this criterion, new efficient optimization methods are proposed.
They are similar in spirit to the classical algorithms based on the
kurtosis contrast function, but differ in the fact that they show a
cubic dependence with respect to the searched parameters.
Therefore, the main advantage of these new methods consists in
the significant improvement of computational speed, which is
particularly striking with large number of samples. Simulations
validate the performance of these methods and also show
experimentally that they are much quicker than some classical
and other corresponding methods.
Keywords-cumulant-based contrast functions; reference
contrast functions;, kurtosis contrast function
I. INTRODUCTION
For the last decades, blind source separation (BSS) has
been applied in a wide variety of fields such as array
processing, passive sonar, seismic exploration, speech
processing, multi-user wireless communications, etc[1]. In the
case of a linear multi-input/multi-output (MIMO) instantaneous
system, BSS corresponds to independent component analysis
(ICA), which is now a well recognized concept[2]. In this
contribution, we are mainly concerned with optimization
algorithms in the ICA framework, where sources are
statistically mutually independent and linearly and
instantaneously mixed.
In a linear multi-input/multi-output (MIMO) instantaneous
system, the problem of BSS has found interesting solutions
through the optimization of so-called contrast functions[1],
which are generally treated as separation criteria. Many
separation criteria rely on higher-order statistics (e.g., the
kurtosis contrast function[2],[3]) or can be linked to higher-
order statistics (e.g., the constant modulus contrast function[4]).
These criteria are known to provide good results. Recently,
some novel contrast functions referred to as “reference-based”
have been proposed in [6] and [7]. They are based on cross-
statistics or cross-cumulants between the estimated outputs and
reference signals [8]-[10]. And they have an appealing feature
in common: the corresponding optimization algorithms are
quadratic with respect to the searched parameters. Taking
advantage of the quadratic reference contrast functions, some
novel methods have been proposed recently.
A maximization algorithm based on singular value
decomposition (SVD) has been proposed in [6] and [11], and
was shown to be significantly quicker than other maximization
algorithms. However, the method often suffers from the need
to have a good knowledge of the filter orders due to its
sensitivity on the rank estimation [5]. The drawback of the
SVD based method is well overcome when replaced by the
gradient optimization method proposed in [12], in which the
reference signals are fixed during the whole optimization
process. Similarly, a relevant gradient algorithm with the
reference signals updating after each one-dimensional
optimization has been proposed in [13], which shows better
performance. Based on the algorithms in [12] and [13], an
improved method is proposed to adjust between performance
and speed of it by introducing a new iterative updating
parameter in [5], and simultaneously a detailed proof of the
global convergence of the algorithm to a stationary point is
performed.
Inspired from [6] and [7], where quadratic higher-order
cumulant based contrast functions are proposed, we come up
with a cubic fourth-order cumulant based contrast function
called new kurtosis contrast function in this paper. Based on
this criterion, new optimization algorithms are proposed, which
show a cubic dependence with respect to the searched
parameters. The papers most directly linked to our approaches
are [7] on the one hand and [6] on the other hand. The former
has introduced the cumulant based quadratic contrast functions,
which inspires us to think of a similarly cubic one. The latter
takes the quadratic fourth-order cumulant based contrast
function as the maximization criterion to restore sources, which
contributes to the proposal of new algorithms presented in our
paper. Besides the algorithms based on the reference contrast
functions in [5] and [13], our methods provide another new
approaches to the efficient optimization of cumulant-based
contrast functions. To our knowledge, it has not been
investigated yet despite its simplicity.
Section II describes the model and assumptions we consider
in this paper. In section III, the separation criteria we adopt are
presented. Our proposed new algorithm can be found in
Section IV. Simulation results are given in Section V and
Section VI concludes this paper.