Zig-Zag And Replacement Product Expander Graphs
For Compressive Sensing
Zhenghua Wu, Qiang Wang, Member, IEEE,Yi Shen,Member, IEEE, Jie Liu, Member, IEEE
Department of Control Science and Engineering, Harbin Institute of Technology, Harbin, China
Phone: 86-0451-86413411; Fax: 86-0451-86418378; E-mail: zhenghuahitchina@gmail.com
Abstract—Compressive Sensing (CS) asserts that one can recover
a sparse signal from a limited number of random or deterministic
projections exactly if the measurement matrix satisfies the so-
called RIP. People try to design deterministic matrices for CS
because of the small storage, high efficiency and low complexity
compared with the random matrices in practical applications.
Recent works explore expander graphs for efficient CS
reconstruction, but the existing expander graphs for CS are
either difficult to obtain or restricted on the number of the
vertices. In this paper, we propose an algorithm named zig-zag
and replacement product expander graphs whose main idea is to
produce another expander graph or an explicit family of
expander graphs using two or more known expander graphs.
Based on the proposed algorithm, the expander graphs are easy
to obtain and the vertices of the graphs, corresponding to the
length of the original signal and the measurement times, aren’t
restricted too much. Finally, numerical simulations are
conducted to verify the better performance of the zig-zag product
matrices compared with the random matrices.
Keywords-Compressive Sensing, Expander Graph, Zig-zag and
Rreplacement Product , Measurement Matrix, RIP
I. INTRODUCTION
In the last few years, Compressive Sensing or Compressive
Sensing (CS) [1]-[3] has received a great amount of attention
for it provides a more efficient data acquisition framework to
replace conventional sampling. The goal of Compressive
Sensing or Compressive Sensing is to recover a sparse signal
from a limited number of random or deterministic projections.
Two conditions make CS possible. On the one hand, signals of
interest must have a sparse representation over an appropriate
basis, orthonormal basis or overcomplete dictionary. On the
other hand, the measurement matrix is expected to satisfy
Restricted Isometry Property (RIP) or the largest correlation
between any two columns of the measurement matrix is as
small as possible. In words, the measurement matrix must
preserve the Euclidean length of sparse signals. In Compressive
Sensing, the originally approach was through the use of dense
random matrices, which satisfy RIP with high probability. For
example, Gaussian random matrix, whose elements are
independent and identically distributed (iid) random variables
from a zero-mean,
V
-variance Gaussian density, has the RIP
with high probability if the number of the measurements
satisfies certain conditions. Although random matrices have
many theoretical advantages, they have the disadvantages of
large storage, low efficiency and high complexity in practical
applications. On the contrary, deterministic matrices allow fast
processing, having low complexity and good performance on
signal recovery. Therefore, it is very important to design
deterministic matrices for Compressive Sensing.
So far, the main technique of designing deterministic
matrices is based on well known codes and sequences e.g.
chirp sequences [4], second order Reed-Muller codes [5] and
dual BCH codes [6]. The foundations of other methods for
deterministic construction are some particular theories such as
finite fields, cyclic difference sets and representation theory.
Recent works [7] explore expander graphs for efficient CS
reconstruction. The expander we care is an unbalanced bipartite
simple graph which has quite sparse density, but the vertices of
it are highly connected. Informally, an expander graph has the
quality that arbitrary small subset of the left vertices has a
sufficient large set of neighbors on the right side. With this
simple property, the expander graph is very useful for
measurement matrix designing because the adjacency matrix of
an expander graph satisfies the RIP-1 and guarantees the
uniqueness of sparse reconstruction [8]. Consequently, a good
expander graph signifies an efficient measurement matrix for
CS.
Although the expander graph has theoretical guidance for
CS, it is difficult to obtain the efficient explicit construction.
Recently, Guruswami et al. proved that there exists a class of
expander graphs with explicit construction which satisfy the
definition of expander graphs. Unfortunately, the degree of the
left vertices and the number of right side vertices are both
restricted, that’s to say, the measurement times that we can
choose are limited. What’s more, it’s difficult to get the
expander graph because the algorithm is complex and the
algorithm based on the theory of finite field. In this paper, we
aim to obtain a good expander graph from another two
expander graphs based on a technique named zig-zag and
replacement product [9]. An expander graph with very small
set of vertices is easy to design, and then the technique will
lead to an iterative construction of an explicit family of
expander graphs. Both theoretical analysis and simulation
results show the products are efficient for measurement matrix
designing. This is the main contribution of this paper. This idea
is shown as Fig.1, we get the Zig-Zag and replacement product
from two expander graphs, and the incidence matrix of the
product, which is the measurement matrix for CS, is picked up
by the proposed algorithm.
978-1-4577-1772-7/12/$26.00 ©2012 IEEE