Matrix Padding Method for Sparse Signal Reconstruction

Sabna N & P. R. Saseendran Pillai
Signal Processing: An International Journal (SPIJ), Volume (9) : Issue (1) : 2015 1
Matrix Padding Method for Sparse Signal Reconstruction
Sabna N sabnan@yahoo.com
Department of Electronics,
Cochin University of Science and Technology,
Cochin 682 022, India
P. R. Saseendran Pillai prspillai@cusat.ac.in
Department of Electronics,
Cochin University of Science and Technology,
Cochin 682 022, India
Abstract
Compressive sensing has been evolved as a very useful technique for sparse reconstruction of
signals that are sampled at sub-Nyquist rates. Compressive sensing helps to reconstruct the
signals from few linear projections of the sparse signal. This paper presents a technique for the
sparse signal reconstruction by padding the compression matrix for solving the underdetermined
system of simultaneous linear equations, followed by an iterative least mean square
approximation. The performance of this method has been compared with the widely used
compressive sensing recovery algorithms such as l1_ls, l1-magic, YALL1, Orthogonal Matching
Pursuit, Compressive Sampling Matching Pursuit, etc.. The sounds generated by 3-blade engine,
music, speech, etc. have been used to validate and compare the performance of the proposed
technique with the other existing compressive sensing algorithms in ideal and noisy
environments. The proposed technique is found to have outperformed the l1_ls, l1-magic,
YALL1, OMP, CoSaMP, etc. as elucidated in the results.
Keywords: Compressive Sensing, Greedy Algorithms, LMS Approximation, Relaxation Methods,
Sparse Recovery, Sub-Nyquist Rate.
1. INTRODUCTION
Compressive Sensing [1]-[3] is a new paradigm that gained the attention of researchers in signal
processing, communication as well as mathematics. It helps in reconstructing the signal from far
less samples than that required by the sampling theorem, which paves the way for saving
memory and low data rate requirements in communication applications.
Traditional methods make use of signal representations conforming to the sampling theorem that
makes compression a necessity before storage or transmission in situations where the memory
space and bandwidth are scarce resources. A signal with only a few non zero coefficients in any
transform domain is called a sparse signal and a signal which can be approximated by a few non
zero coefficients in any transform domain is called a compressible signal. For sparse or
compressible signals, the compressive sensing technology is a paradigm shift.
The recovery of the signal is carried out by using certain optimization techniques. The recovery
becomes more difficult when the signal to be compressed is corrupted with noise. For resorting
to compressive sensing, it is required that the compression and reconstruction techniques should
be capable of transforming the data into a suitable representation domain. Many natural and
manmade signals have underlying sparse representations in some basis functions [4]. Basis
functions like Discrete Fourier Transform, Discrete Cosine Transform, Wavelets, etc. can be
used, depending on the information and type of the signal.

There are many compressive sensing solvers like l1-magic, l1_ls, YALL1, etc., which come under
the category of relaxation methods and Orthogonal Matching Pursuit (OMP), Compressive
Sampling Matching Pursuit (CoSaMP), etc., which come under the category of greedy methods
for sparse signal recovery. A study of compressive sensing recovery algorithms has been
performed in [5] and concluded that although the relaxation algorithms have high computational
complexity and slow response time, they work better in terms of relative error between the
original and reconstructed signals. However, the authors do not seem to have made any effort to
study the noise immunization capability of the algorithms.
Greedy algorithms are faster and have simpler implementation. For many applications OMP does
not offer adequate performance, which led to the introduction of CoSaMP. It is faster and more
effective for compressive sensing problems but is usually less efficient than algorithms based on
convex optimization or relaxation methods. Hence, a method which offers an acceptable signal
reconstruction and fast response has been proposed in this paper. Moreover, the robustness of
various compressive sensing recovery algorithms are compared with the proposed technique
under noisy and ideal environments.
The paper is organized as follows. Section 2 of the paper discusses the fundamental concepts in
compressive sensing while section 3 discusses some of the most widely used compressive
sensing algorithms. Section 4 gives a method for sparse signal reconstruction based on matrix
padding and the iterative least mean squares approximation. Section 5 makes a comparison of
the performances of the algorithms by estimating the signal-to-noise ratio, correlation and mean
squared error. Section 6 throws light on some of the prospective applications of the proposed
technique.
2. BACKGROUND
A discrete time signal x(n) with N elements, can be viewed as an N x 1 vector with n=1, 2, .... N.
Consider a basis function ψ, which provides K sparse representations (i.e., ||x||0 ≤ K) of x(n), with
K < N. x(n) can be represented in the matrix form as x=ψf, where ψ is the basis matrix of order
N x N and f is the weighting coefficient vector of order N x 1. The vector y=ϕx, where ϕ is the
measurement matrix of order M x N with M < N, is the linear projection of the signal x(n).
Recovering the original signal x requires solving an underdetermined system of simultaneous
linear equations. Given the knowledge that x is sparse, the system regenerates the actual signal
from the acquired small number of non-adaptive linear projections of the signal.
2.1 Data Recovery
To recover the sparse signal, the condition that should be satisfied is
minimize ||x||0 subject to y=ϕx (1)
where ||x||0 is the number of non-zero elements of x, which is also called l0 norm. Computing l0
norm is an NP-hard problem [6] which led to making use of the basis pursuit relaxation or convex
optimization [7]. Such approaches have led to the establishment of an l1 - l0 equivalence [8] and
hence, (1) can be represented as,
minimize ||x||1 subject to y=ϕx (2)
where ||x||1 is the sum of the absolute values of the elements in x, which is also being referred to
as the l1-norm.
In the case of signals, which are contaminated with noise, the equality constraint is relaxed to
allow some error tolerance є ≥ 0 [9], such that
minimize ||x||1 subject to ||૖‫ܠ‬ − ‫||ܡ‬ଶ ≤ є (3)
will help to reconstruct x with insignificant error [10],[11].

2.2 Foundation of Compressive Sensing
Compressive sensing is based on sparsity and incoherence. Incoherence expresses the idea
that the signals that are spread out in the domain in which they are acquired, may have a sparse
representation in another domain. For effective reconstruction, it is mandatory that ϕ has to be
incoherent with ψ. Incoherence implies that the mutual coherence or the maximum magnitude of
entries of the product matrix ϕψ is relatively small. Coherence is measured according to
μሺ߶, ߰ሻ = √ܰ ݉ܽ‫ݔ‬ห〈߶௞, ߰ ௝〉ห 1 ≤ ݇, ݆ ≤ ܰ (4)
If ϕ and ψ contain correlated elements, the coherence is large [1]. Coherence takes the values in
the range [1, √ܰ]. The random measurement matrices are largely incoherent with any fixed basis
function ψ. Hence the sparsity basis function need not even be known, when designing the
measurement system.
The necessary and sufficient condition for the sparse signals to be uniquely determined is that the
matrix ϕ should satisfy the Restricted Isometry Property of order K [9], such that
ሺ1 − ߜ௄ሻ‖‫‖ܠ‬ଶ
ଶ
≤‖ࣘ‫‖ܠ‬ଶ
ଶ
≤ ሺ1 + ߜ௄ሻ‖‫‖ܠ‬ଶ
ଶ
(5)
where ߜ௄ is the isometry constant of the matrix ߶ and K is the sparsity of the signal. Evaluating
RIP for a given matrix being computationally complex, this property is normally verified by
computing the coherence of the matrix ϕ. Certain random matrices such as Gaussian and
Bernoulli matrices are known to obey RIP [12].
A signal x is compressible if its sorted coefficient magnitudes ߙn in the transform domain ψ
observe a power law decay [11], according to which,
|ߙ௡| ≤ ܴ݊ି௤
, n=1, 2, … (6)
where R and q are constants with q ≥ 1.
3. COMPRESSIVE SENSING ALGORITHMS
The general block diagram for Compressive Sensing and recovery is shown in Fig. 1. The data
from the wave file is segmented into frames followed by transforming the signal into suitable
domain for making it sparse. It is then compressed with the help of a measurement matrix.
These values can either be stored or transmitted, depending on the requirement. These signals
are reconstructed with the help of a compressing sensing algorithm, followed by transforming it to
the domain in which the data was acquired and the frames are reassembled to regenerate the
signal.
FIGURE 1: Block Diagram for Compressive Sensing and Recovery.
Solution to the sparse recovery problem can be achieved with Relaxation Methods or with Greedy
Algorithms. Relaxation methods replace the original sparse recovery problem with a convex
optimization problem, whereas Greedy Algorithms focus on finding the non-zero values of x at
their respective locations, which are determined iteratively. The Greedy Algorithms accumulate
the approximations by making locally optimal choices iteratively.
Regenerated
Signal
Compression
Reassembling
of frames
Input
Signal
Segmentation Transformation
Reconstruction
Inverse
Transformation

3.1 Relaxation Methods
l1-norm based sparse recovery problems can be solved using a variety of existing solvers such as
l1-magic, YALL1, l1_ls, etc.. l1-magic involves recovery of a sparse vector x from a small number
of linear measurements y=ϕx or y=ϕx+e using primal-dual method [13], [14], where e is the
measurement noise,. YALL1 is another solver that can be applied to l1-optimization, which is a
collection of fast l1-minimization algorithms based on the dual alternating direction method [15],
[16]. It is a first order primal-dual algorithm, as it explicitly updates both primal and dual variables
at every iteration.
l1_ls solves an optimization problem of the form
minimize ‖૖‫ܠ‬ − ‫‖ܡ‬ଶ
ଶ
+ ߣ‖‫‖ܠ‬ଵ (7)
where λ ˃ 0 is the regularization parameter. Equation (7) is referred to as an l1-regularized least
squares problem. l1_ls is a specialized interior-point method [17] that uses the preconditioned
conjugate gradients algorithm to compute the search direction.
3.2 Greedy Algorithms
Greedy Algorithms include Orthogonal Matching Pursuit (OMP), Compressive Sampling Matching
Pursuit (CoSaMP), etc.. Orthogonal Matching Pursuit, which is one of the earliest methods for
sparse approximation, provides simple and fast implementation [18]. It is an iterative algorithm
that selects at each step the dictionary element best correlated with the residual part of the signal
[19]. New approximation is generated by projecting the signal on to the dictionary elements that
have already been selected and solving a least squares problem. The residual is updated in
every iteration. The number of iterations can be made equal to the sparsity of the signal or the
stopping criteria can be based on the magnitude of the residual [9]. Compressive Sampling
Matching Pursuit [20] which selects multiple columns per iteration [9], is an enhancement to
OMP. Each iteration of CoSaMP reduces the error in the current signal approximation.
4. PROPOSED APPROACH FOR SPARSIFICATION AND RECOVERY
In view of the overwhelming limitations of the existing compressive sensing algorithms, in terms
of computational complexities, response time, overall reconstruction capabilities, etc., there was a
constant search for high performance, more capable and reliable techniques for the sparsification
and recovery of audio, speech and natural images. Hence, a new method has been proposed in
this paper which involves padding the matrix ϕ during compression phase for the purpose of
solving the underdetermined system of simultaneous linear equations, followed by least mean
square based adaptation during the reconstruction phase. The solution is obtained with the help
of a pseudo-inverse matrix by matrix padding, which is then corrected using iterative least mean
square based adaptation.
The signal has to be converted to the domain in which it is sparse, depending on the information
and type of the signal. In the simulation studies, the domain chosen is Discrete Cosine Transform
(DCT). Conversion of the signal to DCT results in a signal which is sparse with real valued
coefficients, thus making the reconstruction easier [21]. The advantages of the Discrete Cosine
Transform over Discrete Fourier Transform lies on the fact that it is real-valued, has better energy
compaction and as such a sizeable fraction of the signal energy can be represented by a few
initial coefficients [22]. The DCT of a 1-D sequence f(x) of length N is
ܿሺ‫ݑ‬ሻ = ߙሺ‫ݑ‬ሻ ∑ ݂ሺ‫ݔ‬ሻ ܿ‫ݏ݋‬ ቂ
గሺଶ௫ାଵሻ௨
ଶே
ቃேିଵ
௫ୀ଴ (8)
for u = 0,1,2,…,(N −1).
where
ߙሺ‫ݑ‬ሻ =
‫ە‬
‫۔‬
‫ۓ‬ට
ଵ
ே
, ݂‫ݎ݋‬ ‫ݑ‬ = 0
ට
ଶ
ே
, ‫ݐ݋‬ℎ݁‫݁ݏ݅ݓݎ‬
(9)

The first coefficient, being the average value of the sample sequence, is referred to as the dc
coefficient, while all other transform coefficients are called the ac coefficients. Similarly, the
inverse DCT is defined as
݂ሺ‫ݔ‬ሻ = ∑ ߙሺ‫ݑ‬ሻ ܿሺ‫ݑ‬ሻ ܿ‫ݏ݋‬ ቂ
గሺଶ௫ାଵሻ௨
ଶே
ቃேିଵ
௨ୀ଴ (10)
for x = 0, 1, 2,… (N −1).
The proposed technique converts the signal into sparse domain by applying DCT, followed by
compressing it using a modified measurement matrix. This modification of the measurement
matrix has been effected by padding it with a suitable sub matrix for resolving the singularity
problems, while solving the underdetermined system of simultaneous linear equations. Making
use of this technique, a computationally efficient sparse signal reconstruction can be achieved.
4.1 Matrix Padding
The data from the wave file is divided into N' frames of N samples and these frames are then
converted to the frequency domain by using DCT which resulted in a sparse data representation.
The compression matrix used is a random Gaussian measurement matrix ϕ of size M x N with M
< N. In order to make the computation of the matrix inverse feasible during the reconstruction
phase at the receiver, it is padded with (N-M) x N ones which makes the matrix size to N x N.
Operation of this modified matrix ϕ' upon the framed sparse data results in a signal matrix y that
has two sub matrices of which the first sub matrix yc gives the data pertaining to the matrix
operation y=ϕx, while the other sub matrix yam provides certain redundant data consequent to the
process of matrix padding. Removing the redundant data from yam results in a vector yav of size 1
x N'. The matrix yc of order M x N' and the vector yav of order 1 x N' are to be transmitted or
stored separately. The algorithmic procedure for the compression is given in ALGORITHM I.
ALGORITHM I: Compression Procedure
Begin
Read the wave file
Convert it into N' frames of N samples
Create a matrix x(N,N') with the N' frames of samples
Generation of modified Measurement matrix
Generate Gaussian random measurement matrix ϕ
Generate modified Measurement matrix ϕ'
Compression Phase
Multiply x with DCT matrix ψ
Compress the signal using ϕ'
Generate yc (compressed data) and yam (auxiliary matrix)
Store / transmit yc and yav (auxiliary vector) separately
End
The signal is reconstructed at the receiver by generating the signal matrix y' by appending the
received yc' with yam', which is generated from the received yav' by performing the reverse of the
operations carried out at the transmitter. The Moore Penrose inverse of ϕ' is taken and multiplied
with y' and the data so obtained is converted back to the time domain by the Inverse DCT
operation to generate the initial solution, which is refined further by iterative LMS adaptation. The
procedure for the recovery of the signal at a later stage is furnished in ALGORITHM II.

4.2 Initial Solution
The problem in sparse recovery addresses the selection of the right solution from the feasible set.
In order to find an appropriate solution, consider
minimize J(x) subject to y=Ax (11)
Choosing J(x) as the squared Euclidean norm ‖‫‖ܠ‬ଶ
ଶ
and using Lagrange multipliers,
खሺ‫,ܠ‬ ߣሻ = ‖‫‖ܠ‬ଶ
ଶ
+ ߣ்ሺ‫ܠۯ‬ − ‫ܡ‬ሻ (12)
where ߣ is the Lagrange multiplier or dual variable. Derivative of (12) with respect to x,
∇‫ܠ‬ख = 2‫ܠ‬ + ‫ۯ‬୘
ߣ (13)
Equating (13) to 0 yields,
‫ܠ‬୧୬୧୲୧ୟ୪ = −
ଵ
ଶ
‫ۯ‬୘
ߣ (14)
Substituting (14) into the constraint y=Ax of (11) gives,
‫ܠۯ‬୧୬୧୲୧ୟ୪ = −
ଵ
ଶ
‫ۯۯ‬୘
ߣ = ‫ܡ‬ ⟹ ߣ = −2ሺ‫ۯۯ‬୘ሻିଵ
‫ܡ‬ (15)
Putting the value of ߣ from (15) into (14) yields,
‫ܠ‬୧୬୧୲୧ୟ୪ = ‫ۯ‬୘ሺ‫ۯۯ‬୘ሻିଵ
‫ܡ‬ = ‫ۯ‬ற
‫ܡ‬ (16)
which is the pseudo-inverse solution. As the initial solution is not sparse, an LMS based
adaptation has been used to refine the result.
ALGORITHM II: Recovery Procedure
Begin
Generate the initial solution
Retrieve / receive the signals yc' and yav'
Generate yam' from yav'
Append yam' to yc' and generate y'
Generate Moore Penrose inverse of ϕ' and multiply it with y'
Perform Inverse DCT operation and reassembling of frames
Perform LMS based adaptation for signal refinement
For n=1,……Length of x(n)
Compute the output y(n)
Compute the error signal e(n)
Update the filter coefficients w(n+1)
End
End
4.3 LMS Based Adaptation
An adaptive filter is a self-designing one, which relies on a recursive algorithm for its operation
that makes it possible for the filter to perform satisfactorily in an environment where complete
knowledge of the relevant signal characteristics is not available [23]. For descending towards the
minimum on the mean-square-error performance surface of the adaptive filter, least-mean-square
or LMS algorithm can be used, which is simple and has less computational complexity. LMS filter
is used in a large number of applications like echo cancellation, channel equalization, signal
prediction, etc..
If y(n) is the n
th
sample of the observed output signal [24], then

‫ܡ‬ሺnሻ = ‫ܠ‬୘ሺnሻ ‫ܟ‬ሺnሻ (17)
where ‫ܟ‬ሺnሻ = ሾw଴ሺnሻ wଵሺnሻ … … w୐ିଵሺnሻሿ୘
and xሺnሻ = ሾxሺnሻ xሺn − 1ሻ … … xሺn − L + 1ሻሿ୘
denote the
filter coefficient vector and input vector respectively and L is the filter length. The error of the
adaptive filter output with respect to the desired response signal d(n) is
‫܍‬ሺnሻ = ‫܌‬ሺnሻ − ‫ܠ‬୘ሺnሻ ‫ܟ‬ሺnሻ (18)
By minimizing the cost function which is the mean squared error [25], the filter coefficients are
updated iteratively, so that
‫ܟ‬ሺn + 1ሻ = ‫ܟ‬ሺnሻ + ߤ ‫܍‬ሺnሻ ‫ܠ‬ሺnሻ (19)
where ߤ is the adaptation step size. If R is the autocorrelation matrix of the input vector ‫ܠ‬ሺnሻ and
λmax, its maximum eigen value, the condition for convergence for the LMS is [24]
0 ˂ ߤ ˂
ଵ
ఒ೘ೌೣ
(20)
The structure of adaptive LMS FIR filter is shown in Fig. 2. It is known that the optimum weight
vector, which is the point at the bottom of the performance surface, is
‫܅‬∗
= ‫܀‬ିଵ
‫۾‬ (21)
where R is the input correlation matrix given by
‫܀‬ = Eሾ‫ܠ‬ሺnሻ‫ܠ‬ሺnሻ୘ሿ (22)
and
‫۾‬ = Eሾ‫܌‬ሺnሻ‫ܠ‬ሺnሻሿ (23)
Taking expectation on both sides of (19),
Eሾ‫ܟ‬ሺn + 1ሻሿ = Eሾ‫ܟ‬ሺnሻሿ + ߤ Eሾ‫܍‬ሺnሻ‫ܠ‬ሺnሻሿ
= Eሾ‫ܟ‬ሺnሻሿ + ߤ Eሾ‫܌‬ሺnሻ‫ܠ‬ሺnሻ − ‫ܠ‬ሺnሻ ‫ܠ‬୘ሺnሻ ‫ܟ‬ሺnሻሿ
= Eሾ‫ܟ‬ሺnሻሿ + ߤሺ‫۾‬ − Eሾ‫ܠ‬ሺnሻ ‫ܠ‬୘ሺnሻ ‫ܟ‬ሺnሻሿሻ
= Eሾ‫ܟ‬ሺnሻሿ + ߤሺ‫۾‬ − Eሾ‫ܠ‬ሺnሻ ‫ܠ‬୘ሺnሻሿ Eሾ‫ܟ‬ሺnሻሿሻ
(∵ Coefficient vector ‫ܟ‬ሺnሻ is independent of input vector ‫ܠ‬ሺnሻ)
Eሾ‫ܟ‬ሺn + 1ሻሿ = Eሾ‫ܟ‬ሺnሻሿ + ߤሺ‫۾‬ − ‫܀‬ Eሾ‫ܟ‬ሺnሻሿሻ (24)
FIGURE 2: Tapped Delay Line Structure of the LMS Filter.
∑
Weight
Update
∑
W1 W2 WL-1
W0
Z-1
Z-1
Z-1
x(n)
d(n)
y(n)
e(n)
+
-

Let the error in coefficient vector s(n) be,
‫ܛ‬ሺnሻ = ‫ܟ‬ሺnሻ − ‫܅‬∗
(25)
Substituting (25) in (24),
Eሾ‫ܛ‬ሺn + 1ሻሿ = Eሾ‫ܛ‬ሺnሻሿ + ߤ൫‫۾‬ − ‫܀‬ ሺEሾ‫ܛ‬ሺnሻሿ + ‫܅‬∗ሻ൯
= Eሾ‫ܛ‬ሺnሻሿ + ߤሺ‫۾‬ − ‫܀‬ Eሾ‫ܛ‬ሺnሻሿ − ‫܅܀‬∗ሻ
= Eሾ‫ܛ‬ሺnሻሿ − ߤ ‫܀‬ Eሾ‫ܛ‬ሺnሻሿ ሺ∵ ‫۾‬ = ‫܅܀‬∗
ሻ
= ሺ۷ − ߤ ‫܀‬ሻ Eሾ‫ܛ‬ሺnሻሿ (26)
This implies that the mean error in filter coefficients at instant n+1 depend on step size,
autocorrelation of the input vector and the mean error in filter coefficients at the instant n.
5. RESULTS AND DISCUSSIONS
In the simulation studies, the number of samples per frame is chosen to be 2048, which resulted
in 22 frames for the test signal and the tapped delay line structure has been used for the LMS
adaptation with 32 weights.
The proposed approach for compressive sensing has been simulated under noiseless and noisy
environments and the performance of this approach has been vis-a-vis compared with a few of
the widely used compressive sensing recovery methods like l1-magic, l1_ls, YALL1, OMP and
CoSaMP. Comparison of the performances of various algorithms assuming x’ as the recovered
signal and x as the original signal, has been performed in terms of signal-to-noise ratio,
correlation and mean squared error.
Signal-to-Noise Ratio is computed using
SNR = 10 log
∑ ୶మ
∑ሺ୶ି୶ᇱሻమ (27)
Correlation is computed as
R୶୶ᇱ =
∑ ୶ ୶ᇱ
ඥ∑ ୶మ ∑ ୶ᇱమ
(28)
Mean Squared Error is computed using
MSE =
ଵ
୬
∑ሺx − x′ሻଶ
(29)
5.1 Performance Comparison under Noiseless Scenario
Figs. 3, 4 and 5 show the results of comparison of the performances of the proposed method with
the widely used compressive sensing recovery algorithms at 50% compression. The plots show
the values of signal-to-noise ratio, correlation and mean squared error for the sparse recovery
algorithms under consideration.

FIGURE 3: Signal-to-noise ratio Performance of the proposed method is compared with the other
compressive sensing methods.
These plots reveal that the l1_ls, l1-magic, YALL1, and the proposed method give comparable
and good performance under noiseless scenarios, whereas the signal-to-noise ratio, correlation
and mean squared error values of OMP and CoSaMP show that they do not guarantee adequate
performance.
FIGURE 4: Comparison of the Correlation Performance of the proposed method with the other compressive
sensing methods.
FIGURE 5: Comparison of the Mean Squared Error Performance of the proposed method with the other
15
20
25
30
35
l
1-m
agic
l1-ls
yall1
Proposed
m
ethod
O
M
P
C
oSaM
P
Signal-to-noiseratio(dB)
0.8
0.85
0.9
0.95
1
l1-m
agic
l1-ls
yall1
Proposed
m
ethod
O
M
P
C
oSaM
P
Correlation
0
0.5
1
1.5
2
x 10
-3
l
1-m
agic
l1-ls
yall1
Proposed
m
ethod
O
M
P
C
oSaM
P
MeanSquaredError

5.2 Performance Comparison under Gaussian Noise
FIGURE 6: Signal-to-noise ratio Performance of the proposed method is compared with the other
Figs. 6, 7 and 8 show the comparison of the performances of the proposed method with the
widely used compressive sensing recovery methods, l1-magic, l1_ls, YALL1, OMP and CoSaMP,
under noisy environment. The plots show the variations of output signal-to-noise ratio, correlation
and mean squared error with respect to the SNR variation at the input.
At high input signal-to-noise ratios, all the methods show comparable performances. But, as the
input signal-to-noise ratio decreases the performance of the proposed method is much better
than that of the rest of the methods. OMP and CoSaMP also perform better compared to l1-
magic, l1_ls and YALL1 as the SNR at the input decreases. Thus, the noise immunization of the
proposed method is better compared to the other recovery algorithms considered.
FIGURE 7: Comparison of the Correlation Performance of the proposed method with the other compressive
sensing methods.
0 5 10 15 20 25 30
5
10
15
20
25
SNR Input (dB)
Signal-to-noiseratio(dB) l1
-magic
l1-ls
OMP
CoSaMP
Yall1
Proposed method
0 5 10 15 20 25 30
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
SNR Input (dB)
Correlation
l1
-magic
l1-ls
OMP
CoSaMP
Yall1
Proposed method

FIGURE 8: Comparison of the Mean Squared Error Performance of the proposed method with the other
The simulation studies further demonstrate the feasibility of improving sparse recovery using the
proposed matrix padding technique in both ideal and noisy environments. The mean square error
performance of this method is found to be negligibly small when compared to the other
compressive sensing algorithms. The proposed sparse recovery algorithms can be effectively
used in practical communication scenarios. For undoing the channel effects such as multipath,
intersymbol interference, etc., suitable equalization procedures need to be devised. Certain
channel characteristics can be estimated with much less overhead using compressive sensing
algorithms. The matrix padding sparse reconstruction algorithm will be used for the sparse
channel estimation towards nullifying the channel effects.
6. CONCLUSIONS
Compressive sensing recently gained immense attention due to the commendable advantages
the technique offers in signal manipulation at comparatively low bit rate requirements. With the
help of compressive sensing, the salient information in a signal can be preserved in a relatively
small number of linear projections. Compressive sensing has applications in signal processing, in
areas such as coding, signal level enhancement, source separation, etc. [6]. For example, a
sparse representation has only a few non-zero values, which necessitates encoding only these
values to transmit or store the signal. This paper proposed a method for sparse recovery which is
robust even in the presence of noise. In the practical scenario, noise cannot be eliminated and
hence, the proposed robust signal recovery method is a good choice for the enhancement of
corrupted signal. The simulation studies demonstrate that the proposed algorithm can effectively
improve sparse recovery with the help of matrix padding and LMS based adaptation in both ideal
and noisy environments.
7. ACKNOWLEDGEMENTS
The authors gratefully acknowledge the Department of Electronics, Cochin University of Science
and Technology, for extending all the facilities for carrying out this work. The authors also
acknowledge the Kerala State Council for Science, Technology and Environment as well as the
University Grants Commission for the financial assistances.
8. REFERENCES
[1] E.J. Candes and M.B. Wakin. “An Introduction to Compressive Sampling.” IEEE Signal
Process. Mag., pp. 21-30, Mar. 2008.
0 5 10 15 20 25 30
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
SNR Input (dB)
MeanSquaredError
l1
-magic
l1-ls
OMP
CoSaMP
Yall1
Proposed method

[2] R.G. Baraniuk. “Compressive sensing.” IEEE Signal Process. Mag., vol. 24, no. 4, pp. 118–
120,124, 2007.
[3] D.L. Donoho. “Compressed Sensing.” IEEE Trans. Inform. Theory, vol. 52, no. 4, pp. 1289-
1305, Apr. 2006.
[4] R.G. Baraniuk, V. Cevher and M.B. Wakin. “Low-Dimensional Models for Dimensionality
Reduction and Signal Recovery: A Geometric Perspective.” Proc. IEEE, vol. 98, no. 6, pp.
959-971, Jun. 2010.
[5] L. Vidya, V. Vivekananad, U. ShyamKumar, D. Mishra and R. Lakshminarayanan,
“Feasibility Study of Applying Compressed Sensing Recovery Algorithms for Launch
Vehicle Telemetry,” in IEEE Int. Conf. Microelectronics, Commun. Renewable Energy,
2013.
[6] M.D. Plumbey, T. Blumensath, L. Daudet, R. Gribonval and M. Davis. “Sparse
Representations in Audio and Music: From Coding to Source Separation.” Proc. IEEE, vol.
98, no. 6, pp. 995-1005, Jun. 2010.
[7] E.J. Candes and T. Tao. “Decoding by Linear Programming.” IEEE Trans. Inform. Theory,
vol. 51, no. 12, pp. 4203-4215, Dec. 2005.
[8] R.G. Baraniuk, E. Candes, M. Elad and Y. Ma. “Applications of Sparse Representation and
Compressive Sensing.” Proc. IEEE, vol. 98, no. 6, pp. 906-912, Jun. 2010.
[9] J.A. Tropp and S.J. Wright. “Computational Methods for Sparse Solution of Linear Inverse
Problems.” Proc. IEEE, vol. 98, no. 6, pp. 948-958, Jun. 2010.
[10] E.J. Candès. “Compressive sampling,” in Proc. Int. Congr. Mathematicians, Madrid, Spain,
vol. 3, 2006, pp. 1433–1452.
[11] E. Candes, J. Romberg and T. Tao. “Stable Signal Recovery from Incomplete and
Inaccurate Measurements.” Commun. Pure Applied Math., vol. 59, no. 8, pp. 1207-1223,
Aug. 2006.
[12] R. Baraniuk, M. Davenport, R. DeVore and M. Wakin. “A simple proof of the Restricted
Isometry Property for Random Matrices.” Constructive Approximation, vol. 28, no. 3, pp.
253–263, Dec. 2008.
[13] S. Boyd and L. Vandenberghe. Convex Optimization, Cambridge University Press, 2004.
Available: https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf
[14] E. Cande`s and J. Romberg. “l1-MAGIC: Recovery of Sparse Signals via Convex
Programming.” California Inst. Technol., Pasadena, CA, Tech. Rep., Oct. 2005. Available:
http://users.ece.gatech.edu/~justin/l1magic/downloads/l1magic.pdf
[15] J. Yang and Y. Zhang. “Alternating Direction Algorithms For l1-Problems In Compressive
Sensing.” SIAM J. Scientific Computing, vol. 33, no. 1, pp. 250–278, 2011.
[16] Y. Zhang, J. Yang and W. Yin. “User’s guide for YALL1: Your algorithms for L1
Optimization.” Tech. Rep., 2010. [Online]. Available:
http://www.caam.rice.edu/~optimization/L1/YALL1/User_Guide/YALL1v1.0_User_Guide.pdf
[17] S.J. Kim, K. Koh, M. Lustig, S. Boyd and D. Gorinevsky. “An Interior-Point method for
Large-Scale l1-Regularized Least Squares.” IEEE J. Select. Topics Signal Process., vol. 1,
no. 4, pp. 606–617, Dec. 2007.

[18] J.A. Tropp. “Greed is Good: Algorithmic Results for Sparse Approximation.” IEEE Trans.
Inform. Theory, vol. 50, no. 10, pp. 2231-2242, Oct. 2004.
[19] J.A. Tropp and A.C. Gilbert. “Signal Recovery From Random Measurements Via
Orthogonal Matching Pursuit.” IEEE Trans. Inform. Theory, vol. 53, no. 12, pp. 4655-4666,
Dec. 2007.
[20] D.Needell and J.A.Tropp. “CoSaMP: Iterative Signal Recovery from Incomplete and
Inaccurate Samples.” Applied Computational Harmonic Anal., vol. 26, no. 3, pp. 301–321,
May 2009.
[21] Moreno-Alvarado and M. Martinez-Garcia. “DCT-Compressive Sampling of Frequency-
sparse Audio Signals,” in Proc. World Congr. Eng. 2011, vol. II, London, UK, Jul. 2011.
[22] K. Rao & P. Yip. Discrete Cosine Transform - Algorithms, Advantages, Applications. 1st
Edition, Elsevier, Aug. 1990.
[23] S. Haykin. Adaptive Filter Theory. 3rd Edition, Prentice Hall, 1996.
[24] Y. Chen, Y. Gu and A.O. Hero III. “Sparse LMS for system identification,” in IEEE Int. Conf.
Acoustics, Speech and Signal Process, Apr. 2009, pp. 3125 – 3128.
[25] Y. Gu, J. Jin and S. Mei. “l0 Norm Constraint LMS Algorithm For Sparse System
Identification.” IEEE Signal Process. Letters, vol. 16, no. 9, pp. 774-777, Sept. 2009.

Matrix Padding Method for Sparse Signal Reconstruction

More Related Content

What's hot (18)

Similar to Matrix Padding Method for Sparse Signal Reconstruction (20)

Recently uploaded (20)

Matrix Padding Method for Sparse Signal Reconstruction