You are on page 1of 6

Joint Sparse Channel Estimation and Data Detection

for Underwater Acoustic Channels Using Partial


Interval Demodulation
Arunkumar K.P., Chandra R. Murthy and Venkatesh Elango

AbstractWe present a scheme for joint sparse-channel recov- paths [4]. This property has been used in [1] to characterize
ery and data detection in cyclic-prefix orthogonal frequency divi- the channel impulse response by using a path-based model and
sion multiplex (CP-OFDM) communication over doubly-spread estimate the exact ICI for a zero-padded OFDM system. For
underwater acoustic channels. Inter-carrier interference (ICI),
caused by path-dependent Doppler, results in a non-diagonal channels selective in both time and frequency, [1] also shows
channel mixing matrix that makes recovery difficult. To combat that by constructing a dictionary of the signals parameterized
the effect of ICI, we consider the sequence of observations from by a selection of possible parameter values, one can formulate
partial interval demodulators, and using a path-based channel channel estimation as a sparse signal recovery problem. In [5],
model, cast them into a data model amenable for sparse channel an ICI equalization scheme was proposed wherein an iterative
recovery. We then propose a two-stage algorithm for joint channel
estimation and data detection. In the first stage, we recover the receiver progressively increases the ICI span and the maximum
channel from pilot only observations and estimate the unknown Doppler spread to improve the channel estimation in severe ICI
data symbols from post-combined partial interval demodulator conditions. Both [1] and [5] require additional pilot subcarriers
outputs. In the second stage, we use the data symbols estimated in to estimate the ICI, and hence the unknown channel vector,
the first stage to reconstruct the dictionary matrix corresponding in high Doppler spread environments. An alternative approach
to a full interval demodulator, re-estimate the channel using
the entire observations including data subcarriers, and use it to ICI mitigation using several partial interval demodulators,
to detect the unknown data symbols. We also iterate between instead of a full interval demodulator, was proposed in [6]; the
channel estimation and data detection in each stage, refining the authors develop a recursive algorithm to compute the weights
dictionary in every iteration, to further reduce the detection error. for combining the outputs of partial interval demodulators
Our simulation studies show that initial sparse channel recovery so as to make the post-combined channel matrix close to
from the outputs of partial-interval demodulators considerably
improves data detection performance, in terms of bit error rate, diagonal. However, the inherent sparsity of the channel is not
over that from a traditional full length demodulator output, in exploited in [6].
highly Doppler distorted scenarios. In this paper, we consider a cyclic-prefix OFDM (CP-
Index TermsUnderwater acoustic communications, sparse OFDM) system, similar to [6]. Instead of combining the
channel recovery, partial FFT demodulator. outputs of partial interval demodulators and obtaining a least
squares estimate of the post-combined channel matrix as in [6],
I. I NTRODUCTION we exploit channel sparsity to estimate the path-dependent
Orthogonal frequency division multiplexing (OFDM) delay, Doppler and amplitude parameters of the channel from
achieves high data rates, despite large delay spreads, due to pilot only observations of the raw (uncombined) partial inter-
better resilience to inter symbol interference (ISI). However, val demodulator ouputs. An approximate dictionary, initially
the orthogonality between subcarriers is lost when the signal constructed using only the pilot symbols, is used to initiate the
propagates through a Doppler distorted channel, leading to channel recovery for the first stage. We refine the dictionary
inter carrier interference (ICI). Special receiver side signal using improved estimates of the data symbols and iterate be-
processing techniques are needed to recover the transmitted tween channel estimation and data detection until the channel
data symbols in the presence of severe ICI. estimates cease to change significantly between successive
Sparsity based channel recovery techniques are well known iterations in the first stage. The data symbols detected at the
to produce significantly improved channel estimates, and end of the first stage are then used to initialize a second
hence lead to better symbol detection on data subcarriers [1] stage that makes use of the entire observation vector at the
[3]. Underwater acoustic (UWA) channels typically have large output of the full interval demodulator. At every iteration of
delay and Doppler spreads but have only a few dominant this second stage, the dictionary matrix for the full interval
demodulator output is reconstructed using improved estimates
This paper was financially supported by a research grant from the Defence of the unknown symbols on data subcarriers obtained from the
Research & Development Organisation, Govt. of India.
Arunkumar K. P. is with the Naval Physical Oceanographic Labo- previous iteration.
ratory, Kochi 682 021, India. Chandra R. Murthy is with the Dept. The proposed scheme performs joint channel estimation
of ECE, Indian Institute of Science, Bangalore 560 012, India (e-mail: and data detection akin to [7], but applied to the more
cmurthy@ece.iisc.ernet.in). Venkatesh Elango is at the Dept. of ECE, Uni-
versity of California, San Diego. challenging case of underwater channels with a non-diagonal
978-1-5090-1749-2/16/$31.00 c 2016 IEEE channel mixing matrix arising from severe ICI. Unlike [5]
that uses only a single measurement from a full interval The resampled signal is equivalent to a received signal ob-
demodulator, we make use of multiple measurements emerging tained through a channel with Doppler rate . The resampling
from several partial interval demodulators within a symbol removes the nonzero mean of and this lets us assume that
duration in the first stage. We therefore obtain a good channel the Doppler spread is centered around zero [1].
estimate and accurate data symbol detection using fewer pilot We divide the OFDM interval [0, ] into consecutive
()
subcarriers, compared to [5], in the pilot assisted mode. Also, partial intervals of duration / each. The output on the
compared to [6], which uses a least squares channel estimation, subcarrier, by performing CP-OFDM demodulation for the
th

we exploit channel sparsity in the delay-Doppler plane to th partial interval, ( 1) , = 1, 2, , ,
obtain better channel estimates and hence more accurate data is given by
detection. We show, through simulations, that the resulting bit
() 1
error rate of our algorithm is considerably lower than these = ()2
existing methods in the high Doppler spread environment. (1)


II. S YSTEM M ODEL () ()
= , []2 [] + ,
We consider a CP-OFDM system as in [6]. We let denote =1
the OFDM symbol duration, the guard interval, and the (7)
total OFDM block duration. When using a carrier frequency
()
and subcarriers, the th subcarrier is at frequency where is the additive noise, and
( )
()
= + /, = /2, . . . , /2 1. (1) sin , / ()
()
, [] = , (21) , (8)
The transmitted symbol at the th subcarrier is denoted by ()
, /
[]. The disjoint sets of data subcarriers , pilot subcar-
() 1
riers , and null subcarriers partition the available , = ( ) + . (9)

subcarriers. So, the transmitted signal is
{[ ] } By stacking the received symbols across all the subcarriers
1 into z() , the data symbols into s , and the noise
2 2

() = [] , [0, ].
into v() , we get the channel input-output equation as:

(2) z() = H() s + v() , (10)
The impulse response of the UWA channel is modeled as [8]
for = 1, . . . , , where the channel matrix H() can be
expressed as
(, ) = () ( ()) , (3)
=1

H() = ()
. (11)
where () and () are the time-varying amplitude and =1
delay, respectively, of the th path, and is the number of () () ()
significant paths in the channel. As in [1], we assume that, In the above, is a matrix with [ ], = , [], and
within the OFDM symbol duration , the path amplitudes are is a diagonal matrix with (, )th entry [ ], = 2 .
constant, that is, () = , and that the time variation of The output from a full interval demodulator is obtained by
the path delays due to Doppler rate is approximated as summing up all z() , = 1, 2, , :


() = . (4)
z= z() = Hs + v, (12)
Then, the received signal simplifies to =1


where H = =1 H() and v = =1 v() .
() =
([1 + ] ) +
(), (5) If the additive noise in (5) is zero mean circularly symmetric
=1 white Gaussian distributed, then the noise in (12) is zero mean
where () is the additive noise. In [1], [8], the Doppler Gaussian with covariance 0 I, where 0 is the variance of
effect is removed by first resampling the received signal each component of v and I is the identity matrix.
(), with a resampling factor corresponding to a coarse The noise in (10) is also zero mean Gaussian with covariance
Doppler rate estimate, to obtain (). Next, a fine Doppler given by:
shift compensation is applied on (), the baseband version of { } ( )
() () 0 2()(21)
(), to obtain ()2 , where is( the estimated residual = 2 sinc , (13)
)


mean Doppler shift. Defining = 1+ , the resampled
and { }
signal () is given by ( ) ( )
1 2 = 0 for 1 = 2 . (14)
( )
It is straightforward to show that for a time-invariant channel
() =
((1 + ) ) + . (6)
=1
1+ (i.e., when = 0), the channel matrix H, as seen by the full
interval demodulator, is diagonal. When s are nonzero and recovery algorithms. The representative values of ( , ) are
high, due to heavy Doppler spread, the channel matrix is no chosen as { }
longer diagonal. This results in the mixing of the symbols 2
, , . . . , , (19)
at the output corresponding to each subcarrier. If the time-
variation due to can be neglected within a partial interval
{max , max + , . . . , max } . (20)
[6], and can be approximated by , where = (2

1) 2 is the mid-point of the th partial interval, then we may The time resolution for is chosen as a multiple, , of the
approximate the channel matrices as sampling time /, resulting in = / candidate
delays. For , the assumption is that they are spread around

H() = J() () zero after compensation by , with = 2max /() + 1
, (15)
=1 candidate Doppler rates. Defining the coefficient vector corre-
sponding to all delays associated with Doppler scale as
()
where is a diagonal matrix whose (, )th entry is given [ ]
() () ()
by x = 1 , . . . , , (21)
()
[ ], = 2 ( ) , (16)
the stacked coefficient vector corresponding to all candidate
and J() is a matrix with delays and Doppler rates is given by
( ) [(
() 1 2()(21) ) ( ) ]
J, = 2 sinc , (17) x = x
(1) ( )
, . . . , x . (22)

for , = 1, . . . , . Hence, (18) now becomes
Henceforth, we shall consider the channel model for the
(1)
partial interval demodulator outputs expressed by (10), where [ x
]
= J() 1 s, . . . , s + v() , (23)
() () ..
the channel matrix is given by (15) and the noise vector has z() .
zero-mean and a covariance matrix with elements given by ( )
(13) and (14). Given the partial interval demodulator outputs x
in (10), we wish to detect the data symbols [] for . where = is the total number of grid points used in
We do not know the channel parameters, and hence the channel the delay-Doppler plane. Equation (23) is of the form:
mixing matrices H() . We next describe the sparse recovery
framework for channel estimation and the MMSE receiver for z() = A() x + v() , (24)
data symbol detection.
where,
III. C HANNEL E STIMATION AND DATA D ETECTION [ ]
() ()
A() = J() 1 s, . . . , s , = 1, . . . , , (25)
In the formulation above, the channel matrix H() as seen
by each partial interval demodulator output is defined by and hence is amenable to sparse signal recovery algorithms.
triplets ( , , ). In this section, we cast the problem of However, construction of the matrix A() requires knowledge
estimating these triplets as a sparse channel recovery of the complete transmitted symbol vector s. We now describe
problem, given the sequence of partial interval demodulator the two-stage algorithm to recover the channel vector x and
outputs z() , = 1, 2, , , and describe the proposed two the data vector s through an iterative process.
stage algorithm for joint channel estimation and data detection.
By rewriting the received symbol vector as A. STAGE-1: Joint Channel Estimation and Data Detection
using Pilot Subcarriers
[ ] 1
() We start by constructing a vector s from s ,
z() = J() () s, . . . , s ... + v() , (18) by setting the unknown data symbols to zero1 and placing
the known pilot symbol at pilot locations in . Next, the
null subcarriers are grouped around every pilot subcarrier such
the observed signal is a linear combination of an unknown
that they form clusters of 2 + 1 consecutive subcarriers with
number of signals, each characterized by a triplet of unknown
known (zero or pilot) data, and with a pilot at the middle
parameters. This problem can be solved by constructing a
location of each cluster [1]. This placement makes it possible
dictionary, consisting of the signals parameterized by a rep-
for the receiver to estimate the ICI from subcarriers; using
resentative selection of possible parameter values [1]. Since
= 0 corresponds to ICI-ignored receiver processing. By
parameter values which are not part of the solution will have ()
defining z to be a sub-vector of z()
the corresponding coefficient as zero and a large number
of parameter values are needed to construct an accurate 1 We started by constructing s with the unknown data symbols set to zero.
dictionary, the vector of coefficients is sparse. Hence, the However, we find in our simulation studies that a random initialization of the
coefficient vector can be estimated by using sparse signal data symbols works equally well.
l l
that collects the symbols corresponding to the pilot subcarrier 80
80 90 100 110 120
80
80 90 100 110 120

locations , we get:
() () () ()
z = A x + A x + v , (26) 90 90

where,
[ ]
100 100

k
() () () ()
A = I J 1 s, . . . , s ,
110 110

(27)
() () ()
A = A A ,
[ ] 120 120

() ()
= I J() 1 (s s) . . . , (s s) , Fig. 1. Left Pane: Channel matrix seen at full interval demodulator output.
Right Pane: Post-combined channel matrix at (combined) partial interval
(28) demodulator output. (, ) are (row, column) indices of the channel matrix.
Only indices from 80-120 are shown for clarity. Bright yellow indicates high
where the matrix I is the sub-matrix of the magnitude while dark blue indicates low magnitude.
identity matrix consisting only of its rows indexed
by , i.e., I selects the measurements from subcarriers
corresponding to the pilot locations . We may rewrite (26)
pilots as we estimate the MMSE weights, directly, without
as:
() () () recursion. Since we exploit the sparsity of the channel in
z = A x + e , (29)
the delay-Doppler plane, we obtain accurate channel estimates
() () () with fewer pilots than the least squares channel estimator in
where e A x + v is the effective noise that also
includes the dictionary estimation error. [6], and hence our combining weights yield a post-combined
After the sparse channel vector x in (29) is recovered using channel matrix that is closer to diagonal than that of [6].
one of the sparse signal recovery algorithms,2 the estimated If the weights in (31) are normalized to unit 2-norm, then
channel matrix H() for each partial interval demodulator it follows from (13), (14) and (34) that the
{ variance
} of the
output is constructed using (15). These estimates are weighed noise at the post-combiner output is 2 = 0 . A
and combined so as to make the post-combined channel matrix minimum mean square error (MMSE) receiver is applied for
close to diagonal: data demodulation, as follows:
[( )1 ]

() () s = dec H H + 0 I H z (35)
H,: = H,: , (30)
=1
where z is obtained by stacking , = 1, . . . , , and
[ ]
(1) ( ) dec(.) is the hard-thresholding operation to the signal constel-
where the weights w = are
, ,
lation. We find in our simulation studies that zeroing out all
chosen for each subcarrier , as follows, to minimize the
but diagonal entries of the post-combined channel matrix, H,
interference from other subcarriers [6]:
has negligible effect on data detection accuracy. However, this
w R1
z h, , (31) approximation reduces the computational complexity involved
[ ] in the matrix inversion in (35). Using this approximation,
(1) ( )
h, = H, , , H, , (32) unknown data symbols are estimated as follows:
[ ]
0

Rz = h, h
, + I . (33) dec
,
, , (36)
=1
, 2 + 0

The post-combined observation is obtained as a weighted sum where , is the th diagonal entry of H. Using the
()
of : ()
updated s, we reconstruct the dictionary matrix A , for


=
() ()
. (34) = 1, 2, , , and iterate through channel estimation and
=1
data symbol detection.
It can be shown that using observations from the partial
Figures 1 contrasts an instance of the channel matrix obtained
interval demodulator output in the stage 1, instead of the
by simply summing up partial interval outputs with that
full interval demodulator output, improves channel estimation
obtained by weighted combination. The latter is clearly closer
accuracy in a doubly spread channel. We omit the details due
to diagonal than the former. In [6], combining weights are
to lack of space.
estimated using a recursive least squares algorithm, and hence
Notice that, in stage 1, the channel vector is recovered
extra pilots are needed to bootstrap the recursion and reduce
from pilot only observations at the partial interval demodulator
the estimation error. In our approach, we do not need extra
output, and the data symbols are detected from the post-
2 In this work, we consider orthogonal matching pursuit (OMP) for sparse combined demodulator output. That is, the detection of data
signal recovery. symbols primarily helps in accurately estimating the ICI
TABLE I IV. N UMERICAL S IMULATIONS
S UMMARY OF C OMPUTATIONAL C OMPLEXITY.

Computation Stage 1 Stage 2)


In the simulations, we consider uncoded 16 QAM transmis-
(
Dictionary Matrix ( ) 2 sion in a CP-OFDM system. The carrier frequency, bandwidth,
OMP (( ) ) (( )) number of subcarriers, inter carrier spacing, and symbol in-
Channel Matrix ( 2 ) 2 terval specifications are given in Table II. We construct the
Combining Weights ( 3 ) -
sparse channel using = 4 discrete paths, where the inter-
Post-Combined Channel Matrix 2 (- )
Data Demodulation () 3 arrival times are distributed exponentially with a mean of
1 ms. The path amplitudes are Rayleigh distributed with the
average power decreasing exponentially with delay, where the
difference between the beginning and the end of the guard time
caused by the heavy Doppler spread in the channel, leading is 20 dB. The residual Doppler rate for each path is uniformly
to better channel estimates. However, the observations on the distributed in [max , max ], where max = 5 104 (similar
data subcarriers are not used in stage 1. Next, we make use to the numerical studies in [1], [5], [6] but with a higher
of the measurements from locations corresponding to data spread). The grids are constructed using = 5 points for
subcarriers too, in addition to the pilot subcarriers, in the the Doppler rate and = 30 points for the delay resolution
stage 2. (we consider = 2 for oversampling the delay grid). These
CP-OFDM parameters and the channel parameters are a scaled
B. STAGE-2: Joint Channel Estimation and Data Detection version of the specification used in [1].
using Pilot and Data Subcarriers We use the OMP algorithm [11], with number of non-zero
entries to be recovered set to = 10, for sparse channel
We start by constructing the dictionary matrix: recovery. The number of propagation paths need not be known
precisely; as long as we set > , simulation studies show

good channel recovery. Pilot symbols are placed regularly at
A= A() , (37)
every = 16 subcarriers. Two null subcarriers are placed
=1
on either side of each pilot symbol. Therefore, a total of
corresponding to the full length demodulator, using the data 80 subcarriers carry known symbols on them. Note that the
symbols s estimated at the end of the stage 1 iteration. Then fraction of pilot and null subcarriers we use in an OFDM
we proceed to re-estimate the channel vector from: symbol are less than in [1] for the high Doppler spread
scenario. Also, the number of pilots are much fewer than that

required in [6] to yield a converged estimate of the combining
z= z() = Ax + v, (38) weights using the recursive least squares algorithm proposed
=1 in that work.
Figure 2 shows the bit error rate (BER) of the proposed
as in stage 1, but now using the full set of observations includ-
dual stage algorithm with SNR ranging from 0 dB to 25 dB,
ing data subcarriers at the output of full length demodulator.
obtained through Monte Carlo runs averaged over 100 channel
From the estimated channel vector we construct the channel
instantiations. We iterate for iter = 3 through the stage 1
matrices, H() , = 1, . . . , , using (15), and apply the
that uses the partial interval demodulator output and through
MMSE receiver in (35) for data demodulation, with
stage 2 that uses the full interval demodulator output. We use

= 4 partial interval demodulators in our simulations, i.e.,
H = H() . (39) the OFDM symbol duration is split into = 4 equal partial
=1 intervals (refer to [6] for a discussion on the effect of on
system performance). Also shown in Figure 2, for comparison,
We iterate between channel estimation and data detection, as are the BER versus SNR curves for algorithms in [1], [5],
in stage 1, until convergence or till a fixed number of iterations the least squares based channel estimation and data detection,
(iter ) have elapsed. and genie-aided data detection which has perfect channel state
Table I shows the computational complexity, based on float- information. The algorithm in [1] is non-iterative in nature
ing point operation (FLOP) count for matrix-vector operations and uses the full interval demodulator output for pilot only
[9], [10], per iteration, of stage 1 and stage 2. Note that measurements. The algorithm in [5] iterates for iter = 6 upon
the near diagonal nature of post-combined channel matrix the initial estimate from pilot only measurements using the
was utilized to reduce the computational complexity of data full interval demodulator outputs. The dual stage iterative al-
demodulation
( in stage 1. We find that,
) stage 1 has a complexity gorithm proposed in this paper clearly outperforms both these
of 2 + 3 +( ) per iteration and stage 2 algorithms, and the least squares channel estimation based data
has a complexity of 3 + 2 per iteration. detection algorithm, at all SNRs. Figures 3 shows the reduction
In the next section, we present the results of numerical in BER for various algorithms that iterate between channel
simulations evaluating the performance of our algorithm. estimation and data detection, as the number of iterations is
TABLE II 10 0
CP-OFDM PARAMETERS U SED IN THE S IMULATION . Berger et al. [1]
Huang et al. [5]
Carrier frequency ( ) 13 kHz
Stage-1
Bandwidth () 2.44 kHz
Stage-2
No. subcarriers () 256
Symbol duration ( ) 104.86 ms
Subcarrier spacing ( ) 9.54 Hz
Guard interval ( ) 6.15 ms

BER
-1
10
10 0

10 -1

10 -2
BER

0 5 10 15 20 25
Berger et al. [1] Iterations
Huang et al. [5] (N = 6)
iter Fig. 3. BER versus number of iterations iter for the different sparse channel
10 -2 recovery algorithms at SNR = 20 dB. Doppler spread max = 0.5 103 .
Stage-1 (N iter = 3)

Stage-2 (N iter = 3)

LS R EFERENCES
CSI
[1] C. R. Berger, S. Zhou, J. C. Preisig, and P. Willett, Sparse channel
10 -3 estimation for multicarrier underwater acoustic communication: From
0 5 10 15 20 25
SNR (dB)
subspace methods to compressed sensing, IEEE Trans. Signal Process.,
vol. 57, no. 5, pp. 29412965, May 2011.
[2] J.-Z. Huang, C. R. Berger, S. Zhou, and H. J., Comparison of basis
Fig. 2. Simulation results comparing BER of various OMP based sparse pursuit algorithms for sparse channel estimation in underwater acoustic
channel recovery algorithms with the baseline LS estimator and the genie- OFDM, in Proc. MTS/IEEE OCEANS Conf., May 2010.
aided perfect CSI lower bound. Doppler spread max = 0.5 103 . [3] Y. Huang, L. Wan, S. Zhou, Z. H. Wang, and J.-Z. Huang, Comparison
of sparse recovery algorithms for channel estimation in underwater
acoustic OFDM with data-driven sparsity learning, Elsevier Physical
Communication, vol. 13, no. 3, pp. 156167, Dec. 2014.
increased to iter = 24, at 20 dB SNR. It is seen that the [4] W. Li and J. Preisig, Estimation of rapidly time-varying sparse chan-
proposed dual stage algorithm, at the end of stage 2, has settled nels, IEEE J. Ocean. Eng., vol. 32, no. 4, pp. 927939, Oct. 2007.
in about iter = 3 iterations. [5] J. Z. Huang, S. Zhou, J. Huang, C. R. Berger, and P. Willett, Progressive
inter-carrier interference equalization for OFDM transmission over time-
V. C ONCLUSIONS varying underwater acoustic channels, IEEE J. Sel. Topics Signal
Process., vol. 5, no. 8, pp. 15241536, Dec. 2011.
In this paper, we considered joint sparse channel estimation [6] S. Yerramalli, M. Stojanovic, and U. Mitra, Partial FFT demodulation:
and data detection in a time-varying underwater acoustic chan- A detection method for highly Doppler distorted OFDM systems, IEEE
Trans. Signal Process., vol. 60, no. 11, pp. 59065918, Nov. 2012.
nel for a CP-OFDM system. Our proposed two-stage algorithm [7] R. Prasad, C. R. Murthy, and B. D. Rao, Joint approximately sparse
makes use of the partial interval demodulator outputs and a channel estimation and data detection in OFDM systems using sparse
sparse recovery technique in the first stage to estimate the bayesian learning, IEEE Trans. Signal Process., vol. 62, no. 14, pp.
35913603, Jul. 2014.
channel, followed by dictionary refinement via joint channel [8] B. Li, S. Zhou, M. Stojanovic, L. Freitag, and P. Willett, Multicarrier
estimation and data detection using the full demodulator communication over underwater acoustic channels with nonuniform
outputs in the second stage. Simulation results showed that the Doppler shifts, IEEE J. Ocean. Eng., vol. 33, no. 2, pp. 16381649,
Apr. 2008.
proposed two-stage approach significantly reduces the bit error [9] R. Hunger, Floating point operations in matrix-vector calculus, Tech-
rate in time-varying channels compared to existing approaches. nical Report, Version 3, https://mediatum.ub.tum.de/, Associate Institute
In this work, we considered the OMP algorithm for sparse for Signal Processing, Munich Univ. of Technol., Germany, 2007.
[10] B. L. Sturm and M. G. Christensen, Comparison of orthogonal
signal recovery. Future work will consider other, possibly matching pursuit implementations, 20th European Signal Processing
better performing algorithms such as sparse Bayesian learning Conference (EUSIPCO 2012), pp. 220224, Aug. 2012.
and variational Bayes methods. [11] M. Elad, Sparse and redundant representations: From theory to appli-
cations in signal and image processing. Springer, 2010.

You might also like