You are on page 1of 14

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO.

6, JULY 2011 2569


Semiblind Sparse Channel Estimation
for MIMO-OFDM Systems
Feng Wan, Member, IEEE, Wei-Ping Zhu, Senior Member, IEEE, and M. N. S. Swamy, Fellow, IEEE
AbstractIn this paper, a semiblind algorithm is presented for
the estimation of sparse multiple-inputmultiple-output orthog-
onal frequency-division multiplexing (MIMO-OFDM) channels.
An analysis of the second-order statistics of the signal that was
received through a sparse MIMO channel is rst conducted, show-
ing that the correlation matrices of the received signal can be ex-
pressed in terms of the most signicant taps (MSTs) of the sparse
channel. This relationship is used to derive a blind constraint for
the effective channel vector that corresponds to the MST position.
The blind constraint is then combined with the training-based
least square criterion to develop a semiblind approach for the
estimation of MSTs of the sparse channel. A signal perturbation
analysis of the proposed approach is conducted, showing that the
new semiblind solution is not subject to the signal perturbation
error when the sparse channel is a decimated version of a full
nite impulse response channel. Furthermore, the proposed sparse
semiblind algorithm has been extended for the estimation of chan-
nels in the upsampling domain for MIMO-OFDM systems with
pulse shaping. A number of computer-simulation-based experi-
ments for various sparse channels are carried out to conrm the
effectiveness of the proposed semiblind approach.
Index TermsMost signicant taps (MSTs), multiple-input
multiple-output (MIMO) linear prediction, multiple-input
multiple-output orthogonal frequency-division multiplexing
(MIMO-OFDM), semiblind estimation, sparse channel estimation.
NOMENCLATURE
Pseudoinverse.
Kronecker product.
T Transpose.
H Complex conjugate transpose.
Circular convolution.
| |
F
Frobenius norm.
() Delta function.
vec() Stacking of the columns of the involved matrix into a
vector.
Manuscript received February 12, 2010; revised August 13, 2010,
November 29, 2010 and March 12, 2011; accepted April 7, 2011. Date of
publication May 12, 2011; date of current version July 18, 2011. This work was
supported in part by the Natural Sciences and Engineering Research Council of
Canada. The review of this paper was coordinated by Prof. K. Fukawa.
F. Wan was with the Center for Signal Processing and Communications,
Department of Electrical and Computer Engineering, Concordia University,
Montreal, QC H3G 1M8, Canada. He is now with MobilePeak Systems
Inc., San Diego, CA 92121 USA (e-mail: fwan@mobilepeak.com).
W.-P. Zhu and M. N. S. Swamy are with the Center for Signal Processing
and Communications, Department of Electrical and Computer Engineering,
Concordia University, Montreal, QC H3G 1M8, Canada (e-mail: weiping@
ece.concordia.ca; swamy@ece.concordia.ca).
Digital Object Identier 10.1109/TVT.2011.2153218
I. INTRODUCTION
M
ULTIPLE-inputmultiple-output orthogonal frequency-
division multiplexing (MIMO-OFDM) is a promising
technology for future wireless communication systems [1].
Channel estimation is of crucial importance to MIMO-OFDM
systems. Broadly speaking, MIMO-OFDM channel estimation
techniques can be categorized into the following three classes:
1) training-based methods; 2) blind methods; and 3) semiblind
methods. First, training-based methods, e.g., the least square
(LS), maximum-likelihood (ML), and minimum mean square
error (MMSE) methods, employ known training signals to ren-
der an accurate channel estimation and thus require a consump-
tion of spectral resources [2][7]. Blind MIMO-OFDM channel
estimation algorithms, e.g., the approaches proposed in [8]
[11], which exploit the second-order stationary statistics, correl-
ative coding, and other properties, normally have better spectral
efciency. With a small number of training symbols, semiblind
methods have been proposed to estimate the channel ambi-
guity matrix in MIMO-OFDM systems [12][14]. Note that
most of the existing blind and semiblind methods for MIMO-
OFDM channel estimation, except for several algorithms that
are proposed for orthogonal spacetime-coded systems, e.g.,
[15] and [16], are based on the second-order statistics of a
long vector, whose size is equal to or larger than the number
of subcarriers. To reliably estimate the correlation matrix, these
techniques need a large number of OFDM symbols and are not
suitable for fast time-varying channels. In addition, because the
matrices that are involved in these algorithms are of huge size,
their computational complexity is extremely high. In contrast,
a linear-prediction-based semiblind algorithm that is based on
the second-order statistics of a short vector with a size that
is only slightly larger than the channel length has been found
to be much more efcient than the conventional LS methods
for the estimation of frequency-selective MIMO channels [17],
[18]. This method has then been extended for MIMO-OFDM
systems in [19] and [20].
A wireless channel can often be modeled as a sparse channel,
in which the delay spread could be very large, but the number
of signicant (nonzero) paths is normally very small [21][29].
Based on the sparsity assumption of the equivalent discrete-
time channel, where only a few taps in the long tapped delay
line are considered most signicant, the sparse structure of
the channel has been employed to improve channel estimation
for OFDM [21][23], [27] and code-division multiple access
CDMA systems [26]. Note that almost all of the sparse chan-
nel estimation methods in the literature utilize a training se-
quence and comply with the following two steps: 1) Detect the
0018-9545/$26.00 2011 IEEE
2570 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
positions of the most signicant taps (MSTs), which are also
referred to as nonzero taps in some literature, and 2) obtain an
estimate of effective channels by exploiting the position of the
MSTs.
The common problem of the aforementioned sparse channel
estimation methods is that a large number of pilots are needed
to render an accurate MST detection and effective channel
estimation. To increase the spectral efciency, the available
information of user data could be applied to both the MST
detection and the channel estimation. Unfortunately, very little
work on blind MST detection and blind sparse channel es-
timation is found in the existing literature. In [27], a cyclic
prex (CP) has been utilized to blindly estimate the MSTs.
However, this detection scheme needs a large number of OFDM
symbols and a large CP length to obtain precise MST positions.
Based on second-order statistics of the received signal, a very
efcient semiblind MST detection algorithm that requires only
a small number of OFDM symbols and pilot subcarriers has
been developed in our previous work [30]. The MST positions
detected are then utilized, along with an LS algorithm to obtain
the sparse channel estimate.
In this paper, our objective is to develop, for MIMO-OFDM
systems, an efcient semiblind sparse channel estimation ap-
proach. Through the analysis of the second-order statistics of
the received signal that passes through a sparse channel, a
blind algorithm for obtaining a constraint on the sparse channel
vector with respect to the MSTs is developed. By combining
this blind constraint with a training-based sparse LS criterion, a
semiblind solution to the estimation of the effective channel is
then obtained.
The rest of this paper is organized as follows. Section II
introduces the signal and sparse channel models for MIMO-
OFDM systems and gives a brief review of linear-prediction-
based semiblind MIMO-OFDM channel estimation. Section III
presents a semiblind algorithm for the estimation of sparse
channels based on an analysis of second-order statistics of the
received signal through a sparse MIMO channel. Section IV
conducts a perturbation analysis of the proposed sparse semi-
blind approach, justifying that a blind constraint that is not
affected by signal perturbation can be obtained for an M-rate
sparse channel. Section V extends the proposed sparse semi-
blind algorithm for the estimation of channels in the upsam-
pling domain for MIMO-OFDM systems with pulse shaping.
Section VI comprises a number of experiments that validate the
proposed approach and show the signicant advantage of the
sparse semiblind solution over the sparse LS method, as well as
the regular LS and semiblind techniques, in terms of the mean
square error (MSE) of the channel estimate. Finally, Section VII
concludes this paper by highlighting some of the contributions
presented.
II. PRELIMINARIES
A. Data Model
Consider a MIMO-OFDM system with the vertical Bell Lab-
oratories layered spacetime (V-BLAST) structure, in which
there are N
T
independent links, each connected to a transmit
antenna and containing both the pilot and information data. The
mth OFDM symbol in the i
T
th link can be written as a vector
of the frequency-domain signals, i.e.,
X
i
T
(m)

= [X
i
T
(m, 0), X
i
T
(m, 1), . . . , X
i
T
(m, K 1)]
T
where K denotes the number of subcarriers. The inverse
discrete Fourier transform (IDFT) processing gives the time-
domain OFDM signal, which is denoted as
x
i
T
(m)

= [x
i
T
(m, 0), x
i
T
(m, 1), . . . , x
i
T
(m, K 1)]
T
.
After adding a CP, each OFDM symbol is then sent out by the
corresponding transmit antenna.
Suppose that there are N
R
receive antennas in the MIMO-
OFDM receiver. After removing the CP in each link, the signal
that is received at the i
R
th antenna can be described as
y
i
R
(m)

= [y
i
R
(m, 0), y
i
R
(m, 1), . . . , y
i
R
(m, K 1)]
T
. (1)
Then, the received frequency-domain signal after the discrete
Fourier transform (DFT) processing is given by
Y
i
R
(m)

= [Y
i
R
(m, 0), Y
i
R
(m, 1), . . . , Y
i
R
(m, K 1)]
T
.
Considering that MIMO-OFDM systems are designed for
broadband wireless communications, the signal bandwidth is
always larger than the coherence bandwidth, implying that the
channel is frequency selective. Therefore, the channel can be
considered a combination of L
c
paths, i.e.,
H
c
(t) =
L
c
1

l=0

l
(t t
l
)
where t
l
is the delay of the lth path, and
l
is an N
R

N
T
attenuation matrix. Most of the existing channel estima-
tion methods work on the equivalent discrete-time channel,
i.e., the sampled version of the continuous-time channel re-
sponse [31] Thus, each element of the discrete-time multiple-
inputmultiple-output nite impulse response (MIMO-FIR)
channel is an L-tap FIR lter. Assuming that the channel is
constant during a number of consecutive OFDM symbols, the
channel matrix for the lth tap can be written as
H(l)

h
1,1
(l) h
1,2
(l) . . . h
1,N
T
(l)
h
2,1
(l) h
2,2
(l) . . . h
2,N
T
(l)
.
.
.
.
.
.
.
.
.
.
.
.
h
N
R
,1
(l) h
N
R
,2
(l) . . . h
N
R
,N
T
(l)

C
N
R
N
T
where h
i
R
,i
T
(l), 0 l L 1, represents the composite
channel response between the i
R
th receive antenna and the i
T
th
transmit antenna for the lth tap. If the length of the CP is not less
than the channel length L, the time-domain signal model for the
frequency-selective fading channel is given by
y
i
R
(m, n) =
N
T

i
T
=1
h
i
R
,i
T
(n) x
i
T
(m, n) +v
i
R
(m, n)
m 0, . . . , g 1 (2)
where g is the number of OFDM symbols within which
the channel remains unchanged, and v
i
R
(m, n) C is a
WAN et al.: SEMIBLIND SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 2571
spatiotemporally uncorrelated noise with zero mean and vari-
ance
2
v
. The aforementioned channel and signal models will be
used in the next section to develop a semiblind sparse channel
estimation algorithm.
B. Linear-Prediction-Based Semiblind MIMO-OFDM
Channel Estimation
We now briey review our previous work on the linear-
prediction-based semiblind MIMO-OFDM channel estimation
[19]. The key is to employ the second-order statistics-based
MIMO linear prediction method to obtain a blind constraint on
the channel vector, i.e.,
h

=
_
h
T
1
, . . . , h
T
N
R

T
where h
i
R

= [h
i
R
,1
(0), . . . , h
i
R
,1
(L1), . . . , h
i
R
,N
T
(L1)]
T
.
By dening
y(n)

= [y
1
(n), . . . , y
N
R
(n)]
T
(3)
y
P
(n 1)

=
_
y
T
(n 1), . . . , y
T
(n P)

T
(4)
we have the autocorrelation matrix of y
P
(n 1) and the cross-
correlation matrix of y
P
(n 1) and y(n) as

R
n1

=E
_
y
P
(n 1)y
H
P
(n 1)
_
(5)

R
n

=E
_
y(n)y
H
P
(n 1)
_
. (6)
Then, the MIMO linear predictor can be written as [19], [32]
P
P

= [P
P
(1), P
P
(2), . . . , P
P
(P)] =

R
n

R
1
n1
(7)
where P is the length of linear predictor, and P
P
(n), n =
1, . . . , P, is an N
R
N
R
matrix that represents the nth tap of
the prediction lter. In the aforementioned equations, the index
m has been omitted for notational convenience. By using the
properties of the MIMO linear predictor, a blind constraint for
the channel vector h can be derived as [19]
B = (I P

)E
P
(8)
where E
P
is a known permutation matrix, and P

is a matrix
that is determined by a block Toeplitz matrix that consists of
P
P
(n), n = 1, . . . , P, and the null subspace of H(0), which
can be estimated from the covariance matrix of the prediction
error
2
y,P
= R(0) P
P

R
H
n
.
By combining the blind constraint with a training-based
LS criterion, a semiblind channel estimation problem can be
formulated as
min

h
= |Y
pilot


A

h|
2
F
+|

h|
2
F
(9)
where

A is a pilot signal matrix that can be constructed as
shown in [30, Sec. IV], Y
pilot
is the corresponding received
signal vector,

Bis an estimate of the blind constraint, and > 0
is a weighting factor. The solution to this optimization problem
is given by

h = (

A
H

A+

B
H

B)


A
H
Y
pilot
(10)
where the value of can be determined using the scheme
proposed in [19]. In the aforementioned semiblind approach
and in several existing MIMO-OFDMchannel estimation meth-
ods, e.g., the techniques proposed in [2], [3], [5], [8], [9],
and [20], the sparse case of the wireless channel has not
been considered. Thus, the semiblind solution obtained is not
efcient when the channel is sparse. In the next section, we will
propose a new semiblind algorithm for the estimation of sparse
channels that contain only a small number of signicant or non-
zero taps.
III. PROPOSED SEMIBLIND SPARSE CHANNEL
ESTIMATION ALGORITHM
A. Second-Order Statistics of the Signal Received Through
Sparse MIMO
It is known that a wireless channel can very often be modeled
as a sparse channel that contains several zero taps in the uniform
delay line [21][29]. In this paper, we consider point-to-point
MIMO systems, in which the transmit and receive antennas are
colocated. In this case, the propagation delay is roughly the
same for all transmitreceive antenna pairs [33], [34], which
has been considered in the generations of multipath MIMO
channels in practical channel models, e.g., the spatial channel
model (SCM) [35]. Because the delay in the FIR channel is
uniquely determined by the propagation delay of the multi-
paths, when the channel is sparse, the MSTs of the channel
should happen at the same positions for all transmitreceive
antenna pairs [33]. Therefore, the MIMO channel matrix with
respect to the dth (d = 0, 1, . . . , D 1) MST can be ex-
pressed as
Z(d) = H(l
d
) (11)
where l
d
(d = 0, 1, . . . , D 1) are integers, with 0 = l
0
<
l
1
< < l
D1
, and H(l
d
) is, in general, considered Rayleigh
distributed. To distinguish from H(l), Z(d) is referred to as the
effective channel matrix throughout this paper.
The correlation matrix of the received signal vector y(n) can
be, in general, dened as
R(l)

= E
_
y(n)y
H
(n l)
_
, (l = 0, 1, . . . , P). (12)
Obviously, (12) includes the autocorrelation matrix of y(n) as a
special case when l = 0. It has been proved in [30] that, for the
noise-free case, R(l) can be expressed in terms of the effective
sparse channel matrix Z(d), d = 0, 1, . . . , D 1. Using (2),
(3), and (11) in (12), we obtain
R(l) = Z
A
R
x,D
(l)Z
H
A
(13)
where
Z
A

= [ Z(0) Z(1) Z(D1) ] (14)


R
x,D
(l)

=E

x(n)
x(nl
1
)
.
.
.
x(nl
D1
)

x(nl)
x(nl
1
l)
.
.
.
x(nl
D1
l)

(15)
2572 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
with x(n)

= [x
1
(n), x
2
(n), . . . , x
N
T
(n)]
T
. By assuming a unit
signal variance, i.e.,
2
x
= 1, we can further prove [30]
R(l) =
D1

i=0
D1

j=0
(l l
i
+l
j
)Z(i)Z
H
(j). (16)
In the following discussion, we disclose the detailed expression
of R(l) for a few small values of D.
Let us consider the simplest case when D = 2. In this case,
there are only two nonzero effective channel matrices Z(0) and
Z(1) that correspond to H(0) and H(l
1
), respectively. Assum-
ing a unit signal variance, i.e.,
2
x
= 1, according to (16), we
can nd that there are only two nonzero matrices of R(l), i.e.,
R(0) =Z(0)Z
H
(0) +Z(1)Z
H
(1) (17)
R(l
1
) =Z(1)Z
H
(0) (18)
which means that the most signicant lag(MSL) positions of
R(l) are l = 0, l
1
.
When D = 3, R
x,3
(l) has different sparse structures, de-
pending on the relationship between l
1
and l
2
, which leads R(l)
to have different expressions. Using (16), we can obtain the
following results.
Case D3.1. If l
2
= 2l
1
, we have
R(l) =

i=0
Z(i)Z
H
(i), if l = 0
Z(1)Z
H
(0) +Z(2)Z
H
(1), if l = l
1
Z(2)Z
H
(0), if l = l
2
0, otherwise.
(19)
Case D3.2. If l
2
,= 2l
1
, we have
R(l) =

i=0
Z(i)Z
H
(i), if l = 0
Z(1)Z
H
(0), if l = l
1
Z(2)Z
H
(1), if l = l
2
l
1
Z(2)Z
H
(0), if l = l
2
0, otherwise.
(20)
It is shown that the main difference between the two
cases D3.1 and D3.2 lies in the number of MSLs, in addition
to the expression of R(l). The two common MSLs for the
two cases are R(0) and R(l
2
). Other MSLs depend on the
relationship between l
1
and l
2
.
In the case of D = 4, we can easily obtain R(l) with respect
to the rst and the last MSLs as
R(0) =
3

i=0
Z(i)Z
H
(i) (21)
R(l
3
) =Z(3)Z
H
(0) (22)
which is similar to the cases of D = 2 and D = 3. All other
MSLs can be determined from the relationship among l
1
, l
2
,
l
2
l
1
, l
3
l
1
, and l
3
l
2
. Because some of these values can
be identical, as discussed for the case of D = 3, the number of
the MSLs of R(l) can be different. As such, the position of the
MSLs and the expression of R(l) depend on the values of l
1
, l
2
,
l
3
, and their differences. Using (16), we can have a total of eight
possible cases, which can be summarized in a table, as shown
in [30, Tab. I], except that we replace r
i,j
= z(i)z

(j) with
R
i,j
= Z(i)Z
H
(j). When D > 4, a similar table can easily be
designed by a simple computer programming.
By utilizing the relationship between the MSTs and the
MSLs, we have proposed in [30] a highly efcient semiblind
algorithm for the rst step of sparse channel estimation, i.e., the
MST detection. However, for the second step, the estimation
of the effective channel, e.g., the method in [30] and most of
the existing sparse channel estimation methods, e.g., in [21]
[24], [26], [28], [36], and [37], relies on training-based esti-
mation. In the following discussion, we would like to propose a
more efcient semiblind estimation algorithm for sparse MIMO
channels, which requires only a few OFDM symbols plus a very
small number of pilots.
B. Sparse Semiblind Estimation Idea
In this section, we extend our previously developed MIMO-
OFDM semiblind approach [19] for the sparse channel estima-
tion. Prior to presenting the new idea, we propose the following
theorem.
Theorem 1: Correlation Matrix as a Function of Channel
Matrix: In the absence of noise, the correlation matrix R(l)
of the received signal can be expressed in terms of channel
matrices H(l) as
R(l) =

L1

i=l
H(i)H
H
(i l), l = 0, 1, . . . , L 1
0, l > L 1.
(23)
Proof: Letting
H
A

=[H(0), H(1), . . . , H(L 1)]


x
L
(n)

=
_
x
T
(n) x
T
(n L + 1)

T
, (n=0, 1, . . . , K 1)
where x(n) = x(K +n) for n < 0, the circular convolution
(2) in the noise-free case can be rewritten in matrix form as
y(n) = H
A
x
L
(n). (24)
Substituting (24) into (12) yields
R(l) = H
A
R
x,L
(l)H
H
A
(25)
where
R
x,L
(l) = E
_
x
L
(n)x
H
L
(n l)
_
. (26)
When l = 0, we can easily verify that
R
x,L
(0) = I
N
T
L
. (27)
Using (27) into (25) gives
R(0) =
L1

i=0
H(i)H
H
(i). (28)
WAN et al.: SEMIBLIND SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 2573
When l = 1, (26) reduces to
R
x,L
(1) =
_
0
N
T
(L1)N
T
0
N
T
N
T
I
(L1)N
T
(L1)N
T
0
(L1)N
T
N
T
_
(29)
leading (25) to
R(1) =
L1

i=1
H(i)H
H
(i 1). (30)
In a similar manner, we can derive
R(l) =
L1

i=l
H(i)H
H
(i l), for l = 2, . . . , L 1. (31)
Obviously, when l > L 1, we get R(l) = 0, because
R
x,L
(l) = 0. Thus, (23) is proved.
According to (4)(6) and (12), we can get

R
n1
=

R(0) R(1) R(P 1)


R(1) R(0) R(P 2)
.
.
.
.
.
.
.
.
.
.
.
.
R(1 P) R(2 P) R(0)

(32)

R
n
= [ R(1) R(2) R(P) ] . (33)
The aforementioned two equations indicate that the linear pre-
dictor (7) is determined by R(l), which, in turn, constitutes the
blind constraint for the channel vector h according to (8).
Inspired by the relationship (23) revealed in the aforemen-
tioned theorem, let us dene
R
Z
(d)

=
D1

i=d
Z(i)Z
H
(i d)
=
D1

i=d
R
i,id
, (d = 0, 1, . . . , D 1). (34)
Note that R
Z
(d) can be regarded as the autocorrelation matrix
of an output signal that corresponds to the effective sparse
MIMO channel. By comparing (23) and (34), it is not difcult
to nd that, as long as R
Z
(d) is known, a blind constraint on
the effective channel matrix Z(d) can be obtained by following
the linear prediction process in Section II-B.
Letting

R
Z,n1

R
Z
(0) R
Z
(1) R
Z
(D 2)
R
Z
(1) R
Z
(0) R
Z
(D 3)
.
.
.
.
.
.
.
.
.
.
.
.
R
Z
(2 D) R
Z
(3 D) R
Z
(0)

(35)

R
Z,n

= [ R
Z
(1) R
Z
(2) R
Z
(D 1) ] (36)
a constraint B
Z
on Z(d) (d = 0, 1, . . . , D 1) can be obtained
by using the linear prediction method. Therefore, once the
matrices R
Z
(d) (d = 0, 1, . . . , D 1) have been available, a
semiblind estimation problem for the effective channel matrix
Z(d) (d = 0, 1, . . . , D 1) can be formulated as
min
z
= |Y
pilot


A
Z
z|
2
F
+|B
Z
z|
2
F
(37)
where

A
Z
= I A
/
, with A
/
being a pilot signal matrix similar
to

A in (9). The solution to this optimization problem is
given by
z =
_

A
H
Z

A
Z
+B
H
Z
B
Z
_


A
H
Z
Y
pilot
. (38)
It will be shown in Section VI that the performance of the
proposed sparse channel estimation approach is signicantly
superior to the sparse LS and the original semiblind methods.
Clearly, the complexity of the sparse estimation method is
much lower than the original version for nonsparse channels
due to D L. Typically, the computational complexity of the
original semiblind algorithm is in the order of O(L
3
), whereas
the computational complexity for the proposed sparse approach
is only O(D
3
).
C. Computation of R
Z
(d)
We now discuss the estimation of R
Z
(d), d = 0, 1, . . . ,
D 1, with respect to different values of D by utilizing the
second-order statistics analysis in Section III-A. Our objective
here is to express

R
Z
(d), which is the estimate of R
Z
(d), in
terms of

R(l), which is the estimate of R(l). For the simplest
case, when D = 2, according to (17) and (18), it is clear that

R
Z
(0) =

R(0) (39)

R
Z
(1) =

R(l
1
). (40)
When D = 3, utilizing (19) and (20), we can estimate
R
Z
(d) (d = 0, 1, 2) as follows:

R
Z
(0) =

R(0) (41)

R
Z
(1) =

R(l
1
) +

R(l
2
l
1
) [1 (l
2
2l
1
)] (42)

R
Z
(2) =

R(l
2
). (43)
In a similar manner, when D = 4, based on (21) and (22),
we have

R
Z
(0) =

R(0) (44)

R
Z
(3) =

R(l
3
). (45)
The estimation of

R
Z
(1) =

R
1,0
+

R
2,1
+

R
3,2
and

R
Z
(2) =

R
2,0
+

R
3,1
in terms of

R(l) with respect to different cases
of D = 4 can be made, as shown in Table I. Note that for,
cases D4.2, D4.3, and D4.7, aside from

R(l),

R
1,0
or

R
2,0
is
required. In our method, both

R
1,0
and

R
2,0
can be calculated
based on the estimate of Z(d) using the sparse LS method in
[30]. The aforementioned discussion can easily be extended to
the case of a larger value of D.
IV. PERTURBATION ANALYSIS OF THE PROPOSED
SPARSE CHANNEL ESTIMATION
It is known that the linear-prediction- or subspace-based
channel estimation methods are always perturbed by various
sources, e.g., nite data length and measurement noise [38],
[39]. In our previous work [19], [40], we have successfully
2574 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
TABLE I
CORRELATION MATRIX R
Z
(d) WITH RESPECT TO DIFFERENT CASES OF D = 4
applied the perturbation theory to the analysis of MIMO and
MIMO-OFDM channel estimation. In this section, we evaluate
the performance of the proposed sparse channel estimation
approach from the perspective of the perturbation analysis.
We consider only the perturbation due to the nite data length
in the computation of correlation matrices. First, we introduce
the estimated version

R(l) of R(l). The correlation matrix of
y(n) can be estimated as

R(l) =
1
K
K1

n=0
y(n)y
H
(n l) (46)
where y(n) = y(K +n) for n < 0. By letting
x
D
(n)

=
_
x
T
(n l
0
), x
T
(n l
1
), . . . , x
T
(n l
D1
)

T
(n = 0, 1, . . . , K 1)
where x(n) = x(K +n) for n < 0, the circular convolution
(2) in the noisy case can be rewritten in matrix form as
y(n) = Z
A
x
D
(n) +v(n) (47)
where v(n)

= [v
1
(n), v
1
(n), . . . , v
N
R
(n)]
T
. Substituting (47)
into (46) and using (13), we have

R(l) = R(l) +Z
A
R
x,D
(l)Z
H
A
+R
v
(48)
where
R
x,D
(l)

=
1
K
K1

n=0
x
D
(n)x
H
D
(n l) R
x,D
(l) (49)
represents the signal perturbation, and R
v
is the perturbation
error that is introduced by the noise. Obviously, the rst term
in the right-hand side (RHS) of (48) is an ideal correlation
matrix of the received signal vector y(n) with neither the signal
perturbation nor the noise corruption, whereas the second and
the third terms are the errors that are introduced by the signal
and noise perturbations, respectively.
In the noise-free case, (48) can be rewritten as

R(l) = Z
A
R
x,D
(l)Z
H
A
+Z
A
R
x,D
(l)Z
H
A
(50)
where R
x,D
(l) is given by (49), which represents an error
term that is introduced by the signal perturbation. Based on
(50), we would like to investigate the signal perturbation error
of the blind constraint

B
Z
in the proposed semiblind approach.
To this end, we need to derive the perturbation term R
Z
(d) of
R
Z
(d) by employing R
x,D
(l). Let us rst consider the case
of an M-rate sparse channel case, i.e., the sparse channel is
a decimated version of the full-length FIR channel by a factor
of M. We have
Z(d) = H(dM). (51)
In this case, (49) is still applicable, except that x
D
(n) is
replaced by
x
D
(n) =
_
x
H
(n), x
H
(n M), . . . x
H
(n (D 1)M)

H
.
(52)
On the other hand, based on (23), (34), and (51), we have
R
Z
(d) = R(dM), (d = 0, 1, . . . , D 1). (53)
Using (50) and (53), we obtain
R
Z
(d) = R(dM) = Z
A
R
x,D
(dM)Z
H
A
. (54)
By dening x
M
(n)

= x(nM) and using (13), (49), and (52),
(54) can be rewritten as
R
Z
(d) = Z
A
_
1
K
K1

n=0
x
MD
(n)x
H
MD
(n d)
_
Z
H
A
R
Z
(d)
(55)
where
x
MD
(n) =
_
x
H
M
(n), x
H
M
(n 1), . . . , x
H
M
(n D + 1)

H
.
(56)
Interestingly, (55) and (56) indicate that R
Z
(d) can be
viewed as the perturbation term of the correlation matrix of
the received signal through a regular FIR channel characterized
by Z(l), l = 0, 1, . . . , D 1, with respect to the transmitted
signal x
M
(n). Thus, the perturbation analysis of the regular
channel estimation method in [19] can directly be applied to the
M-rate sparse channel case. As a result, an ideal nulling con-
straint on the effective channel vector z is obtained, leading to
a signal-perturbation-free sparse channel estimation. Moreover,
the practical scheme of computing the weighting factor , as
suggested in [19], can also be used.
In the non-M-rate sparse channel case, because the expres-
sion of R
Z
(d) as given in (55) and (56) is not available,
an ideal nulling constraint on the channel vector z cannot be
obtained, even in the absence of noise. Therefore, the non-
WAN et al.: SEMIBLIND SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 2575
Fig. 1. Discrete-time channel model with pulse shaping.
M-rate sparse channel estimation is, in general, subject to sig-
nal perturbation error. Recall that, to determine the weighting
factor , MSE
B
needs to be calculated, which, as shown in
[19], can easily be implemented online by using a closed-
form expression in signal-perturbation-free channel estimation.
In the non-M-rate sparse channel estimation, it is difcult to
obtain a closed-form expression for MSE
B
due to the existence
of the signal perturbation error. Considering that the MSE
B
only slightly varies with the change of the channel, however,
we can estimate MSE
B
using a training signal and then compute
the weighting factor ofine for the semiblind sparse channel
estimation.
V. SPARSE CHANNEL ESTIMATION FOR PULSE-SHAPED
MULTIPLE-INPUTMULTIPLE-OUTPUTORTHOGONAL
FREQUENCY-DIVISION MULTIPLEXING SYSTEMS
A. Sparse Semiblind Algorithm With Pulse Shaping
Because pulse-shaping and matched lters are commonly
used in digital communication systems, we consider the sparse
channel estimation problem for pulse-shaped systems in this
section. First, we introduce a channel modeling of pulse-shaped
MIMO-OFDM systems [41]. In digital communications, the
pulse-shaping lter is often realized in the upsampled domain
by a discrete-time version off a raised-cosine FIR lter, i.e.,
g(t) = sinc
_
t
T
_
cos
_
t
T
_
1
_
2t
T
_
2
where is the roll-off factor, and T is the symbol period.
Therefore, for a pulse-shaped MIMO-OFDM systems, a com-
posite channel model should include the pulse-shaping lter,
the analog multipath channel H
c
(t), and the matched lter,
as shown in Fig. 1. In this model, the upsampling is imple-
mented by inserting M 1 zeros between any two consecutive
input samples prior to pulse shaping. The transmit g
t
(t) and
receive lters g
r
(t) are replaced by two-root raised-cosine FIR
lters g
t
(n) and g
r
(n), whose sampling period is T/M. In
the upsampling domain, i.e., the discrete-time domain with a
sampling duration of T/M, the combination of the digital-
to-analog(D/A) converter, the multipath channel H
c
(t), and
the analog-to-digital (A/D) converter can be represented by an
equivalent discrete-time multipath channel H
e
(n). In gen-
eral, the channel path may not arrive at the exact sampling
time, but it can be considered an equivalent path that occurs
at the sampling instant synchronized to H
e
(T/M), because the
waveform of the D/A converter can normally be assumed as
p(t) = 1, 0 t < (T/M). Thus, in the case of L
d
paths, the
discrete-time channel H
e
(n) can be represented by
H
e
(n) =
L
d
1

i=0
D(i)(n l
ui
) (57)
where D(i) and l
ui
are the channel matrix and the delay
with respect to the ith path. Now, the composite discrete-
time channel H(n) can be regarded as a downsampled version
of the convolution of the transmit pulse-shaping lter g
t
(n),
the discrete-time multipath channel H
e
(n), and the received
matched lter g
r
(n). Note that the delay l
ui
can be determined
prior to channel estimation. In particular, in advanced wireless
networks, the times of arrival (TOAs) are often estimated at
the start of communication and are periodically updated [42],
[43]. For example, the TOA estimation is conducted by using
the ranging techniques for the uplink synchronization phase of
the orthogonal frequency-division multiple access (OFDMA)
systems [42] or for some geolocation applications [43]. On the
other hand, TOA is known to be a slow-fading parameter, com-
pared with the fast-fading parameter (i.e., complex fading am-
plitude), which means that, once an estimate of TOA has been
obtained, it can be used to estimate the fading amplitude for
a relatively larger period of time. In the following discussion,
based on the knowledge of TOAs, i.e., l
ui
, i = 0, 1, . . . , L
d
1,
we develop a sparse semiblind algorithm for the estimation of
the upsampling duration-based effective channel matrix, i.e.,
D(i), i = 0, 1, . . . , L
d
1.
In contrast to the semiblind estimation problemin Section III,
here, we focus on the estimation of the upsampling duration-
based effective channel vector, i.e.,
d

= vec(

D) (58)
where

D

=[d
1
, . . . , d
N
R
]
d
i
R

=
_
d
T
i
R
,1
, . . . , d
T
i
R
,N
T

T
d
i
R
,i
T
= [d
i
R
,i
T
(0), . . . , d
i
R
,i
T
(L
d
1)]
T
. (59)
Similar to obtaining (37), we derive a new optimization scheme
over the upsampling duration-based effective channel vector as
min

d
= |Y
pilot


A
D

d|
2
F
+|B
D

d|
2
F
(60)
where B
D
is a blind constraint matrix, and the matrix

A
D
= I
N
R
A
D
, with A
D
being obtained by following the
LS estimation for upsampling duration-based channels given in
[41]. The solution to the aforementioned optimization problem
is given by

d =
_

A
H
D

A
D
+B
H
D
B
D
_


A
H
D
Y
pilot
. (61)
2576 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
The remaining task is to apply the linear prediction method to
obtain the blind constraint B
D
.
B. Blind Constraint on Sparse Channel in the
Upsampling Domain
Let the overall channel in the upsampling domain, including
the pulse-shaping and matched lters, be denoted by H
u
(n),
whose (i
R
, i
T
)th element is given by
h
u,i
R
,i
T
(n) = g(n) h
e,i
R
,i
T
(n) (62)
where h
e,i
R
,i
T
is the (i
R
, i
T
)th element of H
e
(n), and g(n) =
g
t
(n) g
r
(n). Then, we have the upsampling domain version
of Threorem 1 as follows.
Theorem 2: Correlation Matrix as a Function of Channel
Matrix in the Upsampling Domain: In the absence of noise, the
correlation matrix of the received signal in the upsampling do-
main y
u
(n), R
u
(l)

= Ey
u
(n)y
H
u
(n l) can be expressed
in terms of channel matrices H
u
(l) as
R
u
(l) =

1
M
L
u
1

i=l
H
u
(i)H
H
u
(i l), l = 0, 1, . . . , L
u
1
0, l > L
u
1
(63)
where L
u
is the length of H
u
(l).
Proof: Letting
x
L
u
(n)

=
_
x
T
u
(n) x
T
u
(nL
u
+1)

T
, (n=0, 1, . . . , K1)
(64)
where x
u
(n) is the transmitted signal in the upsampling do-
main, x
u
(n) = x
u
(MK +n) for n < 0 due to the circular
convolution, then the received signal in the upsampling domain
y
u
(n) can be written as
y
u
(n) = H
uA
x
L
u
(n) (65)
where
H
uA

= [H
u
(0), H
u
(1), . . . , H
u
(L
u
1)] . (66)
Note that one transmitted sample in the sampling domain
x(n) contributes to M samples in the upsampling do-
main, i.e., [x(n), 0, . . . , 0]. Dene a correlation matrix of
x
L
u
(n) as
R
x
u
,L
u
(l)

= E
_
x
L
u
(n)x
H
L
u
(n l)
_
. (67)
Considering that
R
x
u
(l)

= E
_
x(n)x
H
(n l)
_
=
1
M
I
N
T
(l) (68)
we have
R
x
u
,L
u
(0) =
1
M
I
N
T
L
u
(69)
R
x
u
,L
u
(1) =
1
M
_
0
N
T
(L
u
1)N
T
0
N
T
N
T
I
(L
u
1)N
T
(L
u
1)N
T
0
(L
u
1)N
T
N
T
_
.
(70)
Following the proof of Threorem 1 and using (64)(70), (63)
can be proved.
If an ideal pulse-shaping scheme is used, we have g(n) =
(n), and
R
u
(l) =R
e
(l)
=

1
M
L
u
1

i=l
H
e
(i)H
H
e
(i l), l = 0, 1, . . . , L
u
1
0, l > L
u
1.
(71)
By comparing (71) and (57) in the upsampling-domain case,
with (23) and (11) in the original sampling domain, we can nd
that the two versions are identical in principle, except that, in
(71), there is a xed coefcient 1/M. Thus, the proposed sparse
semiblind algorithm for the sampling-domain channels in
Section III can directly be applied to the upsampling-
domain channel estimation. In particular, by utilizing l
ui
, i =
0, 1, . . . , L
d
1, we can identify the MSLs of correlation
matrices R
e
(m), i.e., m = l
ui
l
uj
, i j : i, j = 0, 1, . . . ,
L
d
1. Then, the semiblind sparse algorithm in Section III can
be used for the estimation of D(i), i = 0, 1, . . . , L
d
1.
Now, we consider the practical case where the convolution
of the pulse-shaping and matched lters is given by g(n),
n = 0, 1, . . . , L
g
1. Because R
e
(l) is different fromR
u
(l) in
this case, we develop a scheme to estimate R
e
(m = l
ui
l
uj
),
i j : i, j = 0, 1, . . . , L
d
1, by utilizing the knowledge of
both R
u
(l) and g(n). Because the effect of g(n) on R
u
(l) is
the same for all antenna links, for notational simplicity, we use
r
u
(l), r
e
(l), h
u
(l), and h
e
(l) to represent R
u
(l), R
e
(l), H
u
(l),
and H
e
(l), respectively, at any (i
R
, i
T
)th transmit receive link.
Dening
g
/
(n)

=g(n) g

(n) (72)
h
/
e
(n)

=h
e
(n) h

e
(n) (73)
h
/
u
(n)

=h
u
(n) h

u
(n) (74)
and noting that h
u
(n) = g(n) h
e
(n), we have
h
/
u
(n)

= h
/
e
(n) g
/
(n). (75)
On the other hand, based on (63), we can get the stacking of
h
/
u
(n) due to the convolution of h
u
(n), n = 0, 1, . . . , L
u
1,
as given by
h
/
u
= [r
u
(1 L
u
), r
u
(2 L
u
), . . . , r
u
(0)
. . . , r
u
(L
u
2), r
u
(L
u
1)]
T
.
Similarly, we have the stacking of h
/
e
(n), as given by the
following vector:
h
/
e
= [r
e
(1 L
e
), r
e
(2 L
e
), . . . , r
e
(0)
. . . , r
e
(L
e
2), r
e
(L
e
1)]
T
where L
e
is the length of h
e
(n). Then, using (75), we can obtain
the relationship between h
/
u
and h
/
e
as
G
/
h
/
e
= h
/
u
(76)
WAN et al.: SEMIBLIND SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 2577
where G
/
is a Toeplitz matrix that consists of g
/
(n). Letting
m
l
, l = 0, 1, 2, . . . ,

M, where m
0
= 0, be the MSL of a sparse
channel that was calculated from known delays l
ui
, i = 0,
1, . . . , L
d
1, we can extract the nonzero elements of h
/
e
as
h
//
e
=
_
r
e
(m

M
), r
e
(m

M1
), . . . , r
e
(0)
. . . , r
e
(m

M1
), r
e
(m

M
)

T
.
Thus, (76) can be rewritten as
G
//
h
//
e
= h
/
u
(77)
where G
//
is a matrix that sequently consists of the ((L
u

1/2) +j)th column of the matrix G
/
, j = m

M
, m

M1
,
. . . , 0, . . . , m

M1
, m

M
. Note that G
/
is a full-rank matrix of
(2L
u
1) (2

M + 1), where L
u


M for a sparse channel.
Because an estimate of h
/
u
,

h
/
u
, can be obtained by estimating
the correlation function r
u
(l) from the received signal in the
upsampling domain, we can estimate h
//
e
as

h
//
e
= [G
//
]

h
/
u
. (78)
As a result, the estimate of r
e
(m
i
) can be obtained as
r
e
(m
i
) =

h
//
e
(

M + 1 +i), (i = 0, 1, . . . ,

M).
By performing the aforementioned calculation for all the N
R

N
T
transmit receive links, all the signicant correlation ma-
trices R
e
(m
i
), i = 0, 1, . . . ,

M can be estimated. In other
words, R
e
(m
i
), i = 0, 1, . . . ,

M, as in an ideal pulse-shaping
scheme can be obtained. Therefore, the blind sparse approach
in Section III can be used to derive a blind constraint of the
effective channel vector d in the frequency domain.
VI. SIMULATION RESULTS
We consider a MIMO-OFDM system with two transmit and
four receive antennas. The number of subcarriers is set to 1024,
the length of CP is 30, and the length of the linear predictor
in the semiblind algorithm is P = L. In our simulation, the
quaternary phase-shift keying (QPSK) modulation is used, and
a sparse Rayleigh channel with a unit power that is modeled
by a three- or four-nonzero-tap MIMO-FIR lter is assumed, in
which each tap corresponds to a 4 2 random matrix whose
elements are independent and identically distributed (i.i.d.)
complex Gaussian variables with zero mean. For the non-
M-rate sparse channel, the value of MSE
B
is calculated ofine
based on the true channel matrix by 50 Monte Carlo iterations.
The parameter K
e
in the MST detection algorithm in [30] is set
to 0.8.
For comparison, the channel vector h is rst estimated by
the original LS and semiblind methods. With regard to the
estimation of the effective channel vector z, we consider the
proposed sparse LS and semiblind estimation methods, both
with the MST detection. An ideal sparse LS method and an
ideal semiblind method with the knowledge of the true MST
information are also simulated for comparison. For easy ci-
tation, we call these four methods as the sparse LS, sparse
semiblind, ideal sparse LS, and ideal sparse semiblind methods.
The estimation performance is evaluated in terms of the MSE
of the estimate of the channel matrix as given by
MSE =
1
N
MC
N
MC

n=1
|

h
n
h
n
|
2
where N
MC
is the number of Monte Carlo iterations, and h
n
and

h
n
are the true and the estimated channel vectors with
respect to the nth Monte Carlo iteration, respectively.
A simulation study with comparison is conducted for each of
the following sparse channel models.
1) Channel A (Case D3.2): three-tap sparse channel with
l
0
= 0, l
1
= 5, and l
2
= 12;
2) Channel B (Case D4.4): four-tap sparse channel with l
0
=
0, l
1
= 6, l
2
= 14, and l
3
= 20;
3) Channel C (Case D4.2): four-tap sparse channel with
l
0
= 0, l
1
= 5, l
2
= 10, and l
3
= 20;
4) Channel D (Case D4.1): four-tap sparse channel with
l
0
= 0, l
1
= 4, l
2
= 8, and l
3
= 12;
5) Channel E (Case D3.2): three-tap sparse channel with
l
0
= 0, l
1
= 4, and l
2
= 12;
6) Channel F (Case D4.4): four-tap sparse channel in the
upsampling domain, with l
0
= 0, l
1
= 15, l
2
= 25, and
l
2
= 40 for an upsampling rate of 4;
7) Channel G: four-tap fractional channel with l
0
= 0, l
1
=
3.78, l
2
= 6.28, and l
3
= 10.09.
Experiment 1Channel A: In the rst experiment, the chan-
nel estimation performance in terms of the MSE as a function
of the signal-to-noise ratio (SNR) is investigated for channel A.
Here, we consider 15 OFDM symbols and use 30 subcarriers
at the rst symbol as pilot for training. Fig. 2 shows the MSE
plots for the proposed and the reference sparse methods along
with two original methods from 1000 Monte Carlo iterations. It
is shown that the sparse semiblind method is highly consistent
with its ideal version. It is also noted that the proposed sparse
semiblind method signicantly outperform the sparse LS and
the two original methods. In particular, the sparse semiblind
method is superior to the original LS method by nearly 8.9 and
7.3 dB when the SNR is 3 and 19 dB, respectively. Moreover,
the performance of the sparse semiblind method is superior to
the sparse LS method by 2.7 and 1.1 dB at the two SNR levels,
respectively. It implies that the sparse semiblind method is more
advantageous for a lower SNR.
We now investigate the performance of the proposed sparse
channel estimation methods versus the number of OFDM sym-
bols. Here, the same pilot scheme as in the aforementioned
experiment is used. Fig. 3 shows the MSE plots from 1000
Monte Carlo iterations for an SNR of 7 dB. It is shown that
the performance of both semiblind algorithms is improved with
increasing the number of OFDM symbols. In particular, the
sparse semiblind method can achieve a gain of 2.1 2.8 dB
over the sparse LS method.
Experiment 2Channel B: Now, we examine the channel
estimation performance as a function of the SNR for channel B.
The simulation involves 1000 Monte Carlo runs of the transmis-
sion of 15 OFDM symbols, in which 30 subcarriers at the rst
symbol are used as pilot for training. Fig. 4 shows the channel
2578 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
Fig. 2. MSE versus SNR for channel A.
Fig. 3. MSE versus the number of OFDM symbols for channel A.
estimation results of the four sparse and the LS methods. It is
shown that the sparse semiblind method can achieve a gain of
2.7 0.8 dB over the sparse LS method. Similar to Experiment
1, it is observed that the gain of the sparse semiblind method
over the sparse LS method decreases with increasing SNR. This
case is explained as follows. Because the effect of the noise
on the channel estimation decreases much at a higher SNR
level, the effect of the signal perturbation error becomes more
signicant. It conrms our analysis result in Section IV that
the non-M-rate sparse channel estimation is subject to signal
perturbation error.
Experiment 3Channel C: Using the same condition as in
Experiment 2, the channel estimation result is shown in Fig. 5.
We can nd that the sparse semiblind method can achieve a
0.3 1.2 dB gain over the sparse LS method, indicating the
effectiveness of calculating

R
2,0
by using the sparse LS channel
estimate.
Experiment 4Channel D: Here, we investigate the esti-
mation performance of the proposed method for an M-rate
sparse channel. Fig. 6 shows the channel estimation perfor-
Fig. 4. MSE versus SNR for channel B.
Fig. 5. MSE versus SNR for channel C.
mance versus the SNR obtained using the same condition
as in Experiment 2. It is shown that the sparse semiblind
method can achieve a gain of 3.7 5.3 dB over the sparse
LS method. Moreover, the performance improvement of the
sparse semiblind method over the sparse LS method slightly
increases with increasing SNR. This case is different from
the previous non-M-rate sparse channel examples, where the
performance improvement decreases with increasing SNR. This
phenomenon conrms our analysis that the sparse semiblind
method is not subject to the signal perturbation error in the
M-rate sparse channel case.
Experiment 5Channel E: With the same simulation con-
dition as in Experiment 2, Fig. 7 shows the channel estimation
results of the aforementioned ve methods together with the
following two reference sparse methods: 1) the sparse LS
method and 2) the semiblind method using channel D as the
MST detection result. It is shown that the sparse LS method
consistently outperforms the same method using channel D as
the MST detection result by about 1.2 dB. We can also observe
that the gain of the sparse semiblind method using channel
WAN et al.: SEMIBLIND SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 2579
Fig. 6. MSE versus SNR for channel D.
Fig. 7. MSE versus SNR for channel E.
D over the sparse semiblind method increases with increasing
SNR. In particular, the gain is about 4 dB when the SNR is
19 dB. This interesting outcome suggests that, if a non-M-rate
sparse channel can be converted to a M-rate channel, by adding
several zero taps, then using the MST detection result from the
M-rate channel may yield a much better channel estimation
performance.
Experiment 6Channel F: In this experiment, we investi-
gate the estimation of sparse channels for pulse-shaped MIMO-
OFDM systems. We consider channel F with four path delays,
l
0
= 0, l
1
= (15/4)T, l
2
= (25/4)T, and l
3
= 10T. More-
over, the channel has an exponentially decaying prole p(l) =
exp(0.2 l), and the channel is Rayleigh distributed at each
path delay. A square-root raised-cosine lter with an order 16,
an oversampling rate of 4, and a roll-off factor of 0.15 is used
for the pulse-shaping and matched lters. Here, we simulate the
sparse semiblind algorithm with pulse shaping and the sparse
LS algorithm with pulse shaping for an upsampling duration-
based channel. The simulation involves 1500 Monte Carlo runs
of the transmission of two OFDM symbols, in which 32 sub-
Fig. 8. MSE versus SNR for channel F in the upsampling domain.
carriers are used as pilots. For comparison, we also consider the
regular LS method with 50 symbols, in which 64 subcarriers are
used as pilots. Fig. 8 shows the channel estimation result of two
sparse algorithms and of the regular LS method. It is shown that
the sparse semiblind algorithm with pulse shaping signicantly
outperforms the other two algorithms. In particular, the sparse
semiblind algorithm with pulse shaping can achieve a gain of
3.5 1 dB over the sparse LS algorithm with pulse shaping,
whereas the latter method consistently outperforms the regular
LS algorithm by about 2 dB.
Experiment 7Channel G: In this experiment, we examine
the performance of our proposed algorithm on the estimation
of a fractional MIMO channel. Using the same condition
as in Experiment 6, we now consider channel G with four
path delays, l
0
= 0, l
1
= 3.78T, l
2
= 6.28T, and l
3
= 10.09T.
Here, the channel also has an exponentially decaying prole
p(l) = exp(0.2 l) and is Rayleigh distributed at each path
delay. Note that, in this simulation, to realize the convolution
of the aforementioned fractional channel with the transmitted
signal, a Nyquist interpolator with an upsampling rate of 100
is applied, because for a pulse-shaped MIMO-OFDM system
with an upsampling rate of 4, as considered in Experiment 6,
we have only the knowledge of the channel delay in units of
T/4. The simulation involves 1000 Monte Carlo runs of the
transmission of two OFDM symbols, in which 32 subcarriers
are used as pilots. For comparison, we also consider the regular
LS method with 50 symbols, in which 64 subcarriers are used
as pilots. Fig. 9 shows the channel estimation result of two
sparse algorithms and the regular LS method. It is shown that
the sparse semiblind algorithm with pulse shaping signicantly
outperforms the other two algorithms. In particular, the sparse
semiblind algorithm with pulse shaping can achieve a gain of
4 1.7 dB over the sparse LS algorithm with pulse shaping.
By comparing the results with the results in Experiment 6, we
can conclude that the proposed sparse semiblind approach is
more benecial for a fractional channel case.
Experiment 8BER Versus SNR: Now, the BER perfor-
mance is investigated by using the estimated channel matrix and
an ordered V-BLAST decoder. The same simulation condition
2580 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
Fig. 9. MSE versus SNR for channel G (a fractional channel).
Fig. 10. MSE versus BER for channel D.
as in Experiment 2 is used. The simulation involves 1000 Monte
Carlo runs for channel D. Fig. 10 shows the bit error rate (BER)
performance of the proposed methods and the LS method.
Clearly, the performance of the sparse semiblind method is the
best. In particular, the gain of the sparse semiblind method over
the sparse LS method is about 0.7 1 dB.
VII. CONCLUSION
A semiblind sparse channel estimation algorithm has been
proposed for MIMO-OFDM systems. Based on the second-
order statistics of the received signal that passes through a
sparse channel, an algorithm for obtaining a blind constraint
on the sparse channel vector with respect to the MSTs has been
developed. By formulating a semiblind problem that combines
the blind constraint with a training-based sparse LS criterion,
a semiblind solution to the estimation of the effective channel
has been obtained. The perturbation analysis has also shown
that the proposed semiblind solution is not subject to the signal
perturbation error when the sparse channel is a decimated
version of a full channel. Moreover, the semiblind algorithmhas
been extended for the estimation of sparse channels in the up-
sampling domain for MIMO-OFDM systems with pulse shap-
ing. Computer simulations based on various sparse channels
have conrmed that the proposed sparse semiblind approach
signicantly outperforms the sparse LS method and the regular
LS and semiblind techniques.
REFERENCES
[1] G. Stuber, J. R. Barry, S. W. Mclaughlin, Y. Li, M. A. Ingram, and
T. G. Pratt, Broadband MIMO-OFDM wireless communications, Proc.
IEEE, vol. 92, no. 2, pp. 271294, Feb. 2004.
[2] I. Barhumi, G. Leus, and M. Moonen, Optimal training design for
MIMO OFDM systems in mobile wireless channels, IEEE Trans. Signal
Process., vol. 51, no. 6, pp. 16151624, Jun. 2003.
[3] X. Ma, L. Yang, and G. B. Giannakis, Optimal training for MIMO
frequency-selective fading channels, IEEE Trans. Wireless Commun.,
vol. 4, no. 2, pp. 453466, Mar. 2005.
[4] M. Shin, H. Lee, and C. Lee, Enhanced channel estimation technique
for MIMO-OFDM systems, IEEE Trans. Veh. Technol., vol. 53, no. 1,
pp. 261265, Jan. 2004.
[5] J. Zhang, Z. He, and Y. Huang, TSK fuzzy approach to channel esti-
mation for MIMO-OFDM systems, IEEE Signal Process. Lett., vol. 14,
no. 6, pp. 381384, Jun. 2007.
[6] H. Li, C. K. Ho, J. W. M. Bergmans, and F. M. J. Willems, Pilot-aided
angle-domain channel estimation techniques for MIMO-OFDM systems,
IEEE Trans. Veh. Technol., vol. 57, no. 2, pp. 906920, Mar. 2008.
[7] J.-T. Kim and J.-T. Lim, MAP-based channel estimation for MIMO
OFDM over fast Rayleigh fading channels, IEEE Trans. Veh. Technol.,
vol. 57, no. 3, pp. 19631968, May 2008.
[8] C. Shin, R. W. Heath, and E. J. Powers, Blind channel estimation
for MIMO-OFDM systems, IEEE Trans. Veh. Technol., vol. 56, no. 2,
pp. 670685, Mar. 2007.
[9] F. Gao and A. Nallanathan, Blind channel estimation for MIMO OFDM
systems via nonredundant linear precoding, IEEE Trans. Signal Process.,
vol. 55, no. 2, pp. 784789, Jan. 2007.
[10] F. Gao, Y. Zeng, A. Nallanathan, and T. Ng, Robust subspace blind
channel estimation for cyclic prexed MIMO ODFM systems: Algorithm,
identiability and performance analysis, IEEE J. Sel. Areas Commun.,
vol. 26, no. 2, pp. 378388, Feb. 2008.
[11] C. Shin, R. W. Heath, and E. J. Powers, Nonredundant precoding-
based blind and semiblind channel estimation for MIMO block transmis-
sion with a cyclic prex, IEEE Trans. Signal Process., vol. 56, no. 6,
pp. 25092523, Jun. 2008.
[12] Y. Zeng, W. H. Lam, and T. S. Ng, Semiblind channel estimation and
equalization for MIMO spacetime coded OFDM, IEEE Trans. Circuits
Syst. I: Reg. Papers, vol. 53, no. 2, pp. 463474, Feb. 2006.
[13] F. Gao and A. Nallanathan, Resolving multidimensional ambiguity in
blind channel estimation of MIMO-FIR systems via block precoding,
IEEE Trans. Veh. Technol., vol. 57, no. 1, pp. 1121, Jan. 2008.
[14] J. Kim and J. Lim, Subspace-based iterative semiblind channel esti-
mation for MIMO-OFDM considering residual error, IEEE Trans. Veh.
Technol., vol. 58, no. 8, pp. 46604665, Oct. 2009.
[15] T.-H. Chang, W.-K. Ma, and C.-Y. Chi, Maximum-likelihood detection
of orthogonal spacetime block-coded OFDM in unknown block fading
channels, IEEE Trans. Signal Process., vol. 56, no. 4, pp. 16371649,
Apr. 2008.
[16] N. Sarmadi, S. Shahbazpanahi, and A. B. Gershman, Blind channel
estimation in orthogonally coded MIMO-OFDM systems: A semide-
nite relaxation approach, IEEE Trans. Signal Process., vol. 57, no. 6,
pp. 23542364, Jun. 2009.
[17] A. Medles, D. T. M. Slock, and E. De Carvalho, Linear prediction
based semiblind estimation of MIMO FIR channels, in Proc. IEEE 3rd
Workshop Signal Process. Adv. Wireless Commun., 2001, pp. 5861.
[18] A. Medles and D. T. M. Slock, Semiblind channel estimation for MIMO
spatial multiplexing systems, in Proc. IEEE Veh. Technol. Conf., 2001,
vol. 2, pp. 12401244.
[19] F. Wan, W.-P. Zhu, and M. N. S. Swamy, A semiblind channel estima-
tion approach for MIMO-OFDM systems, IEEE Trans. Signal Process.,
vol. 56, no. 7, pp. 28212834, Jul. 2008.
[20] F. Wan, W.-P. Zhu, and M. N. S. Swamy, Semiblind channel estimation of
MIMO-OFDM systems with pulse shaping, in Proc. IEEE ISCAS, 2008,
pp. 125128.
WAN et al.: SEMIBLIND SPARSE CHANNEL ESTIMATION FOR MIMO-OFDM SYSTEMS 2581
[21] H. Minn and V. K. Bhargava, An investigation into time-domain ap-
proach for OFDM channel estimation, IEEE Trans. Broadcast., vol. 46,
no. 4, pp. 240248, Dec. 2000.
[22] M. R. Raghavendra and K. Giridhar, Improving channel estimation in
OFDM systems for sparse multipath channels, IEEE Signal Process.
Lett., vol. 12, no. 1, pp. 5255, Jan. 2005.
[23] J. K. Hwang, R. L. Chung, M. F. Tsai, and J. H. Deng, Highly ef-
cient sparse multipath channel estimator with Chu-sequence preamble
for frequency-domain MIMO DFE receiver, IEICE Trans. Commun.,
vol. E90B, no. 8, pp. 21032110, Aug. 2007.
[24] C. Carbonelli, S. Vedantam, and U. Mitra, Sparse channel estimation
with zero-tap detection, IEEE Trans. Wireless Commun., vol. 6, no. 5,
pp. 17431763, May 2007.
[25] G. Z. Karabulut and A. Yongacoglu, Sparse channel estimation using
orthogonal matching pursuit algorithm, in Proc. IEEE Veh. Technol.
Conf., 2004, vol. 6, pp. 38803884.
[26] D. K. Borah, Estimation of frequency-selective CDMA channels with
large possible delay and Doppler spreads, IEEE Trans. Veh. Technol.,
vol. 55, no. 4, pp. 11261136, Jul. 2006.
[27] C. R. N. Athaudage and A. D. S. Jayalath, Delay-spread estimation
using cyclic prex in wireless OFDM systems, Proc. Inst. Elect. Eng.
Commun., vol. 151, no. 6, pp. 559566, Dec. 2004.
[28] C. Wu and D. W. Lin, Sparse channel estimation for OFDM transmission
based on representative subspace tting, in Proc. IEEE Veh. Technol.
Conf., 2005, vol. 1, pp. 495499.
[29] S. M. S. Sadough, M. M. Ichir, P. Duhamel, and E. Jaffrot,
Wavelet-based semiblind channel estimation for ultrawideband OFDM
systems, IEEE Trans. Veh. Technol., vol. 58, no. 3, pp. 13021314,
Mar. 2009.
[30] F. Wan, W.-P. Zhu, and M. N. S. Swamy, Semiblind most signicant tap
detection for sparse channel estimation of OFDM systems, IEEE Trans.
Circuits Syst. I: Reg. Papers, vol. 57, no. 3, pp. 703713, Mar. 2010.
[31] J. G. Proakis, Digital Communications, 3rd ed. New York: McGraw-
Hill, 1995.
[32] A. Gorokhov and P. Loubaton, Blind identication of MIMO-FIR sys-
tems: A generalized linear prediction approach, Signal Process., vol. 73,
no. 1/2, pp. 105124, Feb. 1999.
[33] M. A. Khojastepour, K. Gomadam, and X. Wang, Pilot-assisted channel
estimation for MIMO OFDM systems using theory of sparse signal recov-
ery, in Proc. IEEE Int. Conf. Acoust., Speech Signal Process., Apr. 2009,
pp. 26932696.
[34] O. Simeone, Y. Bar-Ness, and U. Spagnolini, Linear and nonlinear pree-
qualization/equalization for MIMO systems with long-term channel state
information at the transmitter, IEEE Trans. Wireless Commun., vol. 3,
no. 2, pp. 373378, Mar. 2004.
[35] J. Salo, G. Del Galdo, J. Salmi, P. Kyosti, M. Melojevic, P. Laselva,
and C. Schneider, MATLAB implementation of the 3GPP spatial channel
model (3GPP TR 25.996), Jan. 2005. [Online]. Available: http://www.tkk.
/Units/Radio/scm
[36] C. Wu and D. W. Lin, A group matching pursuit algorithm for sparse
channel estimation for OFDM transmission, in Proc. IEEE Int. Conf.
Acoust., Speech Signal Process., 2006, vol. 4, pp. 429432.
[37] J. Homer, I. Mareels, R. R. Bitmead, B. Wahlberg, and A. Gustafsson,
LMS estimation via structural detection, IEEE Trans. Signal Process.,
vol. 46, no. 10, pp. 26512663, Oct. 1998.
[38] Z. Xu, Perturbation analysis for subspace decomposition with applica-
tions in subspace-based algorithms, IEEE Trans. Signal Process., vol. 50,
no. 11, pp. 28202830, Nov. 2002.
[39] F. Wan, W.-P. Zhu, and M. N. S. Swamy, A spatial extrapolation based
blind estimation approach for directions of arrival of closely spaced
sources, IEEE Trans. Aerosp. Electron. Syst., vol. 46, no. 2, pp. 569582,
Apr. 2010.
[40] F. Wan, W.-P. Zhu, and M. N. S. Swamy, A signal perturbation free
whitening-rotation-based semiblind approach for MIMO channel esti-
mation, IEEE Trans. Signal Process., vol. 57, no. 8, pp. 31543166,
Aug. 2009.
[41] F. Wan, W.-P. Zhu, and M. N. S. Swamy, Channel estimation of pulse-
shaped multiple-input multiple-output orthogonal frequency-division
multiplexing systems, IET Commun., vol. 4, no. 17, pp. 21042114,
Nov. 2010.
[42] J. G. Andrews, A. Ghosh, and R. Muhamed, Fundamentals of WiMAX-
Understanding Broadband Wireless Networking. Englewood Cliffs, NJ:
Prentice-Hall, 2007.
[43] F. Gustafsson and F. Gunnarsson, Mobile positioning using wireless
networks: Possibilities and fundamental limitations based on available
wireless network measurements, IEEE Signal Process. Mag., vol. 22,
no. 4, pp. 4153, Jul. 2005.
Feng Wan (S09M09) received the B.E. degree
in electrical engineering from Wuhan University,
Wuhan, China, in 2000, the M.E. degree in electrical
engineering from the Chinese Academy of Sciences,
Beijing, China, in 2003, and the Ph.D. degree in
electrical engineering from Concordia University,
Montreal, QC, Canada, in 2009.
From 2004 to 2009, he was a Research Assistant
with Center for Signal Processing and Communica-
tions, Department of Electrical and Computer Engi-
neering, Concordia University. From 2009 to 2010,
he was a Postdoctoral Fellow with the University of Southern California,
Los Angeles. He is currently with MobilePeak Systems Inc., San Diego,
CA, working on mobile chip design. His research interests include digital
signal processing, statistical estimation theory, and its applications in multiple-
inputmultiple-output, orthogonal frequency-division multiplexing, wideband
code-division multiple access, and smart antenna wireless communications.
Mr. Wan received the Best Paper Award at the 2008 IEEE International
Conference on Neural Networks and Signal Processing, the Doctoral Thesis
Completion Award, the Graduate Fellowship and the International Tuition Fee
Remission Award from Concordia University, the Chinese Government Award
for Outstanding Self-Financed Students Abroad from the China Scholarship
Council, and the Postdoctoral Fellowship from the Natural Sciences and
Engineering Research Council of Canada.
Wei-Ping Zhu (SM97) received the B.E. and M.E.
degrees in electrical engineering from Nanjing Uni-
versity of Posts and Telecommunications, Nanjing,
China, in 1982 and 1985, respectively, and the Ph.D.
degree in electrical engineering from Southeast Uni-
versity, Nanjing, in 1991.
He was a Postdoctoral Fellow from 1991 to 1992
and a Research Associate from 1996 to 1998 with the
Department of Electrical and Computer Engineering,
Concordia University, Montreal, QC, Canada. From
1993 to 1996, he was an Associate Professor with
the Department of Information Engineering, Nanjing University of Posts and
Telecommunications. From 1998 to 2001, he was with hi-tech companies in
Ottawa, ON, Canada, including Nortel Networks and SR Telecom Inc. Since
July 2001, he has been a full-time Faculty Member with the Center for Signal
Processing and Communications, Department of Electrical and Computer
Engineering, Concordia University, where he is currently a Full Professor.
His research interests include digital signal processing fundamentals, speech
and audio processing, and signal processing for wireless communication, with
a particular focus on multiple-inputmultiple-output systems and cooperative
relay networks.
Dr. Zhu was an Associate Editor for the IEEE TRANSACTIONS ON
CIRCUITS AND SYSTEMSPART I: FUNDAMENTAL THEORY AND AP-
PLICATIONS from 2001 to 2003. From 2006 to 2009, he was an Associate
Editor for Circuits, Systems and Signal Processing. He is currently a Guest
Editor for the IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS:
BROADBAND WIRELESS COMMUNICATIONS FOR HIGH SPEED VEHICLES
and an Associate Editor for the IEEE TRANSACTIONS ON CIRCUITS AND
SYSTEMS II: EXPRESS BRIEFS.
2582 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 6, JULY 2011
M. N. S. Swamy (S59M62SM74F80) re-
ceived the B.Sc. (Hons.) degree in mathematics
from Mysore University, Mysore, India, in 1954,
the Diploma degree in electrical communication
engineering from the Indian Institute of Science,
Bangalore, in 1957, and the M.Sc. and Ph.D. de-
grees in electrical engineering from the University
of Saskatchewan, Saskatoon, SK, Canada, in 1960
and 1963, respectively. In August 2001, he received
the D.Sc. (honoris causa) degree in engineering from
Ansted University in recognition of his exemplary
contributions to research in electrical and computer engineering and to en-
gineering education, as well as his dedication to the promotion of signal
processing and communications applications.
He is currently a Research Professor and the Director of the Center for Signal
Processing and Communications, Department of Electrical and Computer
Engineering, Concordia University, Montreal, QC, Canada, where he served as
the Founding Chair of the Department of Electrical Engineering from 1970 to
1977 and the Dean of the Faculty of Engineering and Computer Science from
1977 to 1993, during which time he developed the faculty into a research-
oriented faculty from what was primarily an undergraduate department. Since
July 2001, he has been the Concordia Chair (Tier I) in Signal Processing. He
was also with the Department of Electrical Engineering, Technical University
of Nova Scotia, Halifax, NS, Canada; the University of Calgary, Calgary, AB,
Canada; and the Department of Mathematics, University of Saskatchewan.
He has extensively published in number theory, circuits, systems, and signal
processing. He is the holder of ve patents. He is a coauthor of two book
chapters and four books: Graphs, Networks and Algorithms (New York: Wiley,
1981; Russian translation: Moscow, Russia: Mir, 1984; Chinese translation:
Beijing, China: Education, 1987), Graphs: Theory and Algorithms
(New York, Wiley, 1992), Switched Capacitor Filters: Theory, Analysis,
and Design (Prentice Hall International U.K., 1995), and Neural Networks in
a Softcomputing Framework (New York: Springer, 2006). He was a Founding
Member of Micronet from its inception in 1999 as a Canadian Network of
Center of Excellence until its expiration in 2004, as well as the Coordinator
for Concordia University. Recently, Concordia University has instituted the
M.N.S. Swamy Research Chair in Electrical Engineering in recognition of his
research contributions.
Dr. Swamy is a Fellow of a number of professional societies, including the
Institute of Electrical Engineers (U.K.), the Engineering Institute of Canada,
the Institution of Engineers (India), and the Institution of Electronic and
Telecommunication Engineers (India). He has served the IEEE in various
capacities, e.g., as the President-Elect in 2003, the President in 2004, the
Past-President in 2005, a Vice President (Publications) from 2001 to 2002, a
Vice-President in 1976, the Editor-in-Chief of the IEEE TRANSACTIONS ON
CIRCUITS AND SYSTEMS I from June 1999 to December 2001, the Associate
Editor for the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS from June
1985 to May 1987, the Program Chair for the 1973 IEEE Circuits and Systems
(CAS) Symposium, the General Chair for the 1984 IEEE CAS Symposium,
a Vice-Chair for the 1999 CAS Symposium, and a Member of the Board of
Governors of the CAS Society. He has received several IEEE CAS Society
awards, including the Education Award in 2000, the Golden Jubilee Medal in
2000, and the GuilleminCauer Best Paper Award in 1986.

You might also like