Adaptive MIMO Channel Estimation using

Sparse Variable Step-Size NLMS Algorithms

Guan Gui1, Li Xu1, Lin Shan2, and Fumiyuki Adachi3
1. Department of Electronics and Information Systems, Akita Prefectural University, Akita, Japan
2. Wireless Network Research Institute, National Institute of Information and Communication Technology, Yokosuka, Japan
3. Department of Communications Engineering, Graduate School of Engineering, Tohoku University, Sendai, Japan
Emails: {guiguan,xuli}@[Link], shanlin@[Link], adachi@[Link]

I.

INTRODUCTION

High-rate data broadband transmission over multiple-input

multiple-output (MIMO) channel is becoming one of
mainstream techniques for the next generation communication
systems [1], [2]. The major motivation is due to the fact that
MIMO technology is a way of using multiple antennas to
simultaneously transmit multiple streams of data in wireless
communications [3] and hence it can bring considerable
improvements such as data rate, reliability and energy
efficiency. However, coherent receivers require accurate
channel state information (CSI) since the received signals are
distorted by multipath fading transmission. The accurate
estimation of channel impulse response (CIR) is a crucial
aspect and challenging issue in coherent modulation and its
accuracy has a significant impact on the overall performance
of the communication system.
During last decades, based on the assumption of dense CIRs,
linear channel estimation methods, e.g., least squares (LS),
were proposed for MIMO systems. By applying these
approaches, the performance of linear methods depend only on

the size of MIMO channel. Note that narrowband MIMO

channel may be modeled as dense channel model because of
its very short time delay spread; however, broadband MIMO
channel is often modeled as sparse channel model [4]. A
typical example of sparse channel is shown in Fig. 1.
0.7

No. of overall taps: 32

No. of nonzero taps: 4

0.6

0.5

Magnitude

AbstractTo estimate multiple-input multiple-output (MIMO)

channels, invariable step-size normalized least mean square
(ISSNLMS) algorithm was applied to adaptive channel
estimation (ACE). Since the MIMO channel is often described by
sparse channel model due to broadband signal transmission, such
sparsity can be exploited by adaptive sparse channel estimation
(ASCE) methods using sparse ISS-NLMS algorithms. It is well
known that step-size is a critical parameter which controls three
aspects: algorithm stability, estimation performance and
computational cost. The previous approaches can exploit channel
sparsity but their step-sizes are keeping invariant which unable
balances well the three aspects and easily cause either estimation
performance loss or instability. In this paper, we propose two
stable sparse variable step-size NLMS (VSS-NLMS) algorithms
to improve the accuracy of MIMO channel estimators. First,
ASCE for estimating MIMO channels is formulated in MIMO
systems. Second, different sparse penalties are introduced to
VSS-NLMS algorithm for ASCE. In addition, difference between
sparse ISSNLMS algorithms and sparse VSS-NLMS ones are
explained. At last, to verify the effectiveness of the proposed
algorithms for ASCE, several selected simulation results are
shown to prove that the proposed sparse VSS-NLMS algorithms
can achieve better estimation performance than the conventional
methods via mean square error (MSE) and bit error rate (BER)
metrics.

0.4

0.3

0.2

0.1

0
0

10

15

Tap index

20

25

30

Fig. 1. A typical example of sparse multipath channel.

Adaptive sparse channel estimation (ASCE) methods using

sparse invariable step-size (ISS) least mean square algorithms
(ISS-LMS) were proposed in [5][7] for single-input single
output (SISO) channels. However, conventional ISS-LMS
methods have two main drawbacks: 1) sensitive to random
scale of training signal and 2) unstable in low signal-to-noise
ratio (SNR) regime. To overcome the two harmful factors on
channel estimation and extend their applications to estimate
MIMO channels, sparse ISS normalized least mean square
(ISS-NLMS) algorithms, e.g., zero-attracting ISS-NLMS (ZAISS-NLMS) and reweight ZA-ISS-NLMS (RZA-ISS-NLMS),
were proposed in [8]. It is well known that step-size is a
critical parameter which controls the estimation performance,
convergence rate and computational cost. Different from
conventional sparse ISS-NLMS algorithms [5][8], zeroattracting variable step size NLMS (ZA-VSS-NLMS)
algorithm was proposed for ASCE to improve estimation
performance in sparse multipath single-input single-output
(SISO) systems [9]. Unlike the previous works, this paper
proposes two sparse VSS-NLMS algorithms for estimating

sparse MIMO channels. The main contribution of this paper is

summarized as follows. First, we extend the method in [9]
from SISO to MIMO systems.
Second, a re-weighted ZA-VSS-NLMS (RZA-VSS-NLMS)
is proposed to further improve the estimation performance of
MIMO channels. In addition, we explain the reason why
sparse VSS-NLMS algorithms can achieve better performance
than conventional sparse ISS-NLMS ones. Finally, Monte
Carlo based computer simulations are conducted to confirm
the effectiveness of our proposed algorithms via two metrics:
bit error rate (BER) and mean square error (MSE).
The remainder of this paper is organized as follows. A
baseband MIMO system model is described and problem
formulation is presented in Section II. In section III, sparse
ISS-NLMS algorithms are reviewed and sparse VSS-NLMS
algorithms are proposed. A figure example is also given to
explain the difference between ISS and VSS based algorithms.
Simulation results are presented in section IV in order to
assess the proposed methods. Finally, we conclude the paper
in Section V.
Notation: Throughout the paper, capital bold letters and
small bold letters denote matrices and row/column vectors,
respectively; The discrete Fourier transform (DFT) matrix is
denoted by F with entries [F ]kq = 1 K e j 2 kq K ,
k ,q = 0, 1,..., K 1 ; Matrices and vectors are represented by
boldface upper case letters and boldface lower case letters,
respectively; The superscripts ()T , ()H , Tr () and ()1
denote the transpose, the Hermitian transpose, the trace and
the inverse operators, respectively; E () denotes the
expectation operator; h 0 is the 0 norm operator that counts
the number of nonzero taps in h and h p stands for the p
p
norm operator which is computed as h p = ( i hi )1 p ,
where p {1, 2} is considered in this paper; sgn() is a
component-wise function which is defined by sgn(h ) = 1 for
h > 0 , sgn(h ) = 0 for h = 0 , and sgn(h ) = 1 for h < 0 .

II.

z nt (t ) is an additive white Gaussian noise (AWGN) variable

with distribution (0, n2 ) and -th received multiple-input
single-output (MISO) channel vector hnr : is written as

hnr : = [hnr 1,0 , , hnr 1,L 1 , , hnr nt ,0 , , hnr nt ,L 1

hnTr 1

hnTr nt

(2)

, , hnr N t ,0 , , hnr N t ,L 1 ]T ,

hnTr Nt

and the matrix-vector form of system model (1) is also written

as
y (t ) = Hx (t ) + z (t ),
(3)
where received signal vector y (t ) , noise vector z (t ) and
channel matrix H can be represented, respectively, as follows:

y (t ) = [y1 (t ), y2 (t ), , yNr (t )]T ,

(4)

z (t ) = [z1 (t ), z 2 (t ), , z N r (t )]T ,

(5)

T
T
h1T: h11 h12
T T
T
h22
h h
H = 2: = 21

hT T
T
N r : hN r 1 hN r 2

h1TN t

h2TN t

T
hN

rNt

(6)

where hnr nt , nr = 1, 2, , N t , is assumed equal L -length

sparse channel vector from receiver to nt -th antenna. In
addition, we also assume that each channel vector hnr nt is
only supported by T dominant channel taps.
x

x1
x 11

hnr 1,0

x2
x 1N

x 21

xN t

x 2N

x Nt 1

xNtN

hnr 2,L 1

hnr N t ,0

hnr N t ,L 1

SYSTEM MODEL

A frequency-selective fading MIMO communication system

using OFDM modulation scheme is considered. Initially,
frequency
domain
training
signal
vector
xnt (t ) = [x nt (t , 0),..., x nt (t , L 1)]T , nt = 1, 2,..., N t is fed to
inverse DFT (IDFT) at the -th antenna, where L is the
number of subcarriers and Nt is the number of transmit
antenna.2 Assume that the transmit power is normalized as
E xnt = 1 . The resultant vector xnt (t ) = F H xnt (t ) is
2
padded with cyclic prefix (CP) of length LCP L to avoid
inter-block interference (IBI). After CP removal, the received
signal vector at the -th antenna for time t is written as
ynr (t ) , where nr = 1, 2,..., N r . Then, the received signal
vector y and input signal vector x (t ) are related by

ynr (t ) = n t=1 hnTr nt xnt (t ) + znr (t )

N

= hnTr :x (t ) + z nr

(1)

(t ),

where x (t ) = [xT1 (t ), xT2 (t ),..., xTNt (t )]T collects all of the input
signal vectors from different antennas at the transmitter;

hnr 1,L 1 hnr 2,0

y nr
enr (n )
Adaptive
algorithm

Fig. 2. MISO channel estimation at

III.

-th antenna of the receiver.

ADAPTIVE CHANNEL ESTIMATION MEHTODS

According to the system model in Eq. (1), the n -th

updating estimation error enr (n ) can be written as

enr (n ) = ynr (t ) ynr (n ) = ynr (t ) hnTr : (n )x (t ),

(7)

where hnr : (n ) denotes an MISO channel estimator of the

hnr : (n ) ; e (n ) = [e1 (n ),e2 (n ), ,eN r (n )]T denotes receive

error vector at the

receive signal at the

-th adaptive update; and ynr (t ) is the

-th receive antenna.

A. Review of ZA-ISS-NLMS and RZA-ISS-NLMS

For estimating hnr : (n ) as shown in Fig. 3, ZA-ISS-NLMS
[7] filtering algorithm was proposed as

hnr : (n + 1) = hnr : (n ) +

en (t )x (t )
r

x (t )x (t )

ZA sgn hnr : (n )

(8)

where ZA = ZA , ZA > 0 is a regularization parameter to

balance the square estimation error enr (t ) and sparse penalty
of hnr : (n ) . Motivated by reweighted -norm minimization
recovery algorithm [10], Chen et al. have proposed a heuristic
approach to reinforce the zero attractor which was termed as
the RZA-ISS-LMS [11]. RZA-ISS-NLMS [7] was proposed as
en (t )x (t ) RZA sgn hnr : (n )
hnr : (n + 1) = hnr : (n ) + T r

, (9)
x (t )x (t )
1 + RZA hnr : (n )

where RZA = RZA RZA , RZA > 0 is the regularization

parameter and reweighted factor RZA > 0 is the positive
threshold. Recall that the ZA-ISS-NLMS algorithm in Eq. (8)
does not make use of the VSS rather than ISS. Inspirited by
the VSS-NLMS algorithm which has been proposed in [12], to
improve estimation performance of MIMO channels, sparse
VSS-NLMS algorithms are proposed. Unlike the sparse ISSNLMS algorithm, sparse VSS-NLMS algorithms are timevariant with respect to the accuracy of updating estimators.

Steps-size for gradient descend

0.8

0.6

n (n ) = max
r

pTnr (n )pnr (n )
pTnr (n )pnr (n ) + C

(11)

where C is a positive threshold parameter which is related to

received signal-to-noise ratio (SNR), C (1 SNR ) .
According to Eq. (11), the range of VSS is given as
nr (n ) (0, max ) , where max is the maximal step-size of
gradient descend. Theoretically, the maximal step-size is less
than 2 to ensure the adaptive algorithm stability [13]. Please
note that pnr (n ) in Eq. (11) is given by
pnr (n ) = pnr (n 1) + (1 )

enr (n )x (t )

xT (t )x (t )

(12)

where [1, 0) is a smoothing factor to trade off VSS and

estimation error.
C. RZA-VSS-NLMS
By introducing the VSS (11) into Eq. (9), improved sparse
channel estimator h (n ) is given by
nr :

hnr : (n + 1) = hnr : (n ) +

(n )en (n )x (t ) RZA sgn hn : (n )

r

x (t )x (t )

1 + RZA hnr : (n )

) . (13)

Please note that the second term in Eq. (13) attracts the
channel coefficients hn n ,l (n ) , l = 0, 1,, L 1
whose
magnitudes are comparable to 1/RZA to zeros. According the
two proposed filtering algorithms in Eqs. (10) and (13),
adaptive channel estimation methods for estimating MIMO
channels are summarized in Algorithm 1.
r

Input: 1) x (t ) and y (t ) ;
2) max = 1 and C ;
3) ZA for ZA-VSS-NLMS
4) RZA and RZA for RZA-VSS-NLMS.
Output: channel estimator H .

Signal length: N=16

Smoothing factor: =0.999
Threshold parameter:
C=0.0001

n 1 ; pnr (0) 0 ; H (0) 0 ;

2
While H (n + 1) H (n ) 105 or n 5000 Do
2

nr mod(n 1, N r ) + 1 ;
hnr : (n ) H (nr ;:) ;

0.4

0.2

dnr (n ) y (nr );
enr (n ) dnr (n ) hnTr :x (n );

ISS (=0.5)
ISS (=1.0)
VSS ( =0.5)
max

VSS (

=1.0)

max

0 -5
10

10

Estimation error
Fig. 3. ISS and VSS versus updating estimation error.

B. ZA-VSS-NLMS
At time t , based on the previous research on the ZA-ISSNLMS and VSS-NLMS algorithms, ZA-VSS-NLMS
algorithm is proposed as follows:
hnr : (n + 1) = hnr : (n ) +

n (n )en (n )x (t )
r

xT (t )x (t )

ZA sgn hnr : (n ) , (10)

where nr (n ) is the VSS which is given by

en (n )x (t )
;
pnr (n ) pnr (n 1) + (1 ) Tr
x (t )x (t )
T
pnr (n )pnr (n )
;
nr (n ) max T
pnr (n )pnr (n ) + C

(
n
)
e
(
n
)x (t )
n
n
hnr : (n + 1) hnr : (n ) + r T r
ZA sgn hnr : (n )
x (t )x (t )
for ZA-VSS-NLMS in (10) or
(n )enr (n )x (t ) RZA sgn hnr : (n )
hnr : (n + 1) hnr : (n ) +

xT (t )x (t )
1 + RZA hnr : (n )
For RZA-VSS-NLMS in (13);

End
H (nr ;:) hnr : (n + 1);
Algorithm 1: Sparse VSS-NLMS algorithms for estimating MIMO channels.

Remark: To better understanding the difference between

ISS and VSS, based on Eqs. (8) and (10), it is worth
mentioning that step size for sparse ISS-NLMS algorithm

ISS-NLMS
VSS-NLMS
ZA-ISS-NLMS
ZA-VSS-VNLMS
RZA-ISS-NLMS
RZA-VSS-VNLMS

-1

10

Average MSE

is invariable but the step size ZA (n ) for sparse VSS-NLMS

algorithm is variable as depicted in as Fig. 3, where the
maximal step size and ISS are set as max {0.5, 1.0} and
{0.5, 1.0} , respectively. From the figure, one can easily
find that ISS is kept invariant, while VSS (n ) decreases as
the estimation performance increases and vice versa. In other
words, sparse VSS-NLMS algorithms adopting VSS for
adaptive gradient descend, large step-size is adopted to speed
up convergence rate for reducing computational complexity;
small step-size is adopted to ensure algorithm stable in the
case of high-accuracy estimator for further improving
estimation performance.

-2

N =N =4

10

SNR=10dB
-3

10

-4

10

IV.

COMPUTER SIMULATIONS

To confirm the effectiveness of the proposed methods, two

metrics, i.e., MSE and BER, are adopted for performance
evaluation. Channel estimators are evaluated by average MSE
which is defined by

2
2

},

2000

3000

4000

Iterations

and system performance is evaluated by the BER metric which

adopts different data modulation schemes, such as phase shift
keying (PSK) and quadrature amplitude modulation (QAM).
The results are averaged over 200 independent Monte- Carlo
runs. The length of each channel vector hnr nt ,
nr = 1, 2, , N r nt = 1, 2, , N t is set as equal length with
L = 16 and corresponding number of dominant taps is set to
T {1, 4} . Each dominant channel tap follows random
Gaussian distribution as (0, h2 ) and their positions are
randomly decided within the length of hnr nt . In
addition,
2
MISO channel vector hnr : is subject to E hnr : = 1 . The
2
received SNR is defined as P0 n2 , where P0 is the power of
received signal. Based on the research work in [14], it is worth
mentioning that threshold parameters of sparse VSS-NLMS
algorithm are adopted C = 104 for 5dB and C = 105 for
10dB and 20dB, respectively.
In the first example, average MSE performance of proposed
methods is evaluated in Figs. 46 under two SNR regimes (i.e.,
10dB and 20dB) in the case of T = 1 and 4, respectively. The
effectiveness of the two proposed methods are confirmed
when compared with previous methods, i.e., ISS-NLMS [13],
VSS-NLMS [12] and sparse ISS-NLMS [5][8]. In addition,
one can also find that two proposed methods depend channel
sparsity as well as regularization parameter. Hence, to achieve
a better steady-state estimation performance, regularization
parameters for sparse VSS-NLMS algorithms, i.e., ZA-VSSNLMS and RZA-VSS-NLMS, are adopted from the paper [15],
which depends on the number of nonzero taps of a channel. In
addition, since sparse VSS-NLMS algorithms take advantage
of the channel sparsity as for prior information, hence, they
achieve better estimation performance than standard VSSNLMS algorithm, especially in a very sparse channel case, e.g.,
T =1.

5000

ISS-NLMS
VSS-NLMS
ZA-ISS-NLMS
ZA-VSS-VNLMS
RZA-ISS-NLMS
RZA-VSS-VNLMS

(14)
-1

10

Average MSE

1000

Fig. 4. Average MSE performance versus iterations (T = 1).

N =N =4

-2

10

SNR=10dB
-3

10

-4

10

1000

2000

3000

4000

Iterations

5000

Fig. 5. Average MSE performance versus iterations (T = 4).

ISS-NLMS
VSS-NLMS
ZA-ISS-NLMS
ZA-VSS-VNLMS
RZA-ISS-NLMS
RZA-VSS-VNLMS

-1

10

-2

10

Average MSE

Average MSE H (n ) = E H H (n )

N =N =4
t

-3

SNR=20dB

10

-4

10

-5

10

-6

10

-7

10

1000

2000

3000

Iterations

4000

Fig. 6. Average MSE performance versus iterations (T = 1).

5000

In the second example, system performance using proposed

channel estimators is also evaluated with respect to BER
performance. Multiple QAM schemes are considered.
Received SNR is defined by Es N 0 , where Es is the average
received power of symbol and N 0 is the noise power. In Fig.
7, multiple QAM schemes, i.e., 16QAM, 64QAM and
128QAM, are considered for data modulation. One can easily
find that the proposed method can achieve a better estimation
than previous methods.

performance than previous methods especially in high-order

modulation signal based MIMO communications systems.
Since the empirical parameter C is adopted for the
proposed sparse VSS-NLMS algorithm in Monte Carlo runs, it
may cause the performance loss in different SNR regimes. In
future work, we plan to develop an adaptive C for the
proposed algorithms so that it can learn the estimation error
and SNR while without sacrificing much computation
complexity.
REFERENCES

10

[1]

[2]

-1

10

64QAM

128QAM

Average BER

[3]
-2

10

16QAM

[4]
-3

10

-4

10

-5

10

12

[5]

ISS-NLMS
VSS-NLMS
ZA-ISS-NLMS
ZA-VSS-NLMS
RZA-ISS-NLMS
RZA-VSS-NLMS
14

16

18

[6]
20

22

E /N (dB)
s

24

26

28

30

[7]

Fig. 7. Average BER performance versus received SNR (QAM).

[8]

V.

CONCLUSIONS AND FUTURE WORK

Traditional adaptive MIMO channel estimation methods

tend to utilize the sparse ISS-NLMS algorithms. One of the
main disadvantages of the traditional methods is the inability
to balance the convergence speed and the estimation accuracy
on the adaptive channel estimation. In this paper, two sparse
VSS-NLMS algorithms, i.e., ZA-VSS-NLMS and RZA-VSSNLMS, were proposed for estimating MIMO channels. Unlike
the traditional sparse ISS-NLMS algorithms, the proposed
algorithms utilized VSS which learns the estimation error and
changes adaptively. Simulation results were provided to
confirm the effectiveness of the proposed methods in three
aspects convergence speed, estimation performance and
system performance. First, convergence speed of sparse VSSNLMS based methods is faster than ISS-NLMS based
methods due to the fact that VSS for adaptive gradient
descend is more efficient than ISS. Second, the proposed
adaptive estimators can achieve better MSE gain than
traditional methods in different SNR regimes especially for
sparser channels. At last, system performance using the
proposed channel estimators can also achieve better BER

[9]

[10]

[11]

[12]

[13]
[14]

[15]

F. Adachi and E. Kudoh, New direction of broadband wireless

technology, Wireless Communications and Mobile Computing, vol. 7,
no. 8, pp. 969983, May 2007.
D. Raychaudhuri and N. Mandayam, Frontiers of wireless and mobile
communications, Proceedings of the IEEE, vol. 100, no. 4, pp. 824
840, Apr. 2012.
A. M. Sayeed, Deconstructing multiantenna fading channels, IEEE
Transactions on Signal Processing, vol. 50, no. 10, pp. 25632579, Oct.
2002.
N. Czink, X. Yin, H. OZcelik, M. Herdin, E. Bonek, and B. Fleury,
Cluster characteristics in a MIMO indoor propagation environment,
IEEE Transactions on Wireless Communications, vol. 6, no. 4, pp.
1465 1475, Apr. 2007.
O. Taheri and S.A. Vorobyov, Sparse channel estimation with lpnorm
and reweighted l1-norm penalized least mean squares, in IEEE
International Conference on Acoustics, Speech, and Signal Processing
(ICASSP), Prague, Czech Republic, 22-27 May 2011, pp. 2864-2867.
O. Taheri and S.A. Vorobyov, Decimated least mean squares for
frequency sparse channel estimation, in IEEE International Conference
on Acoustics, Speech, and Signal Processing (ICASSP), Kyoto, Japan,
2530 Mar. 2012, pp. 31813184.
G. Gui and F. Adachi, Improved least mean square algorithm with
application to adaptive sparse channel estimation, EURASIP Journal on
Wireless Communications and Networking, vol. 2013, no. 1, pp. 118,
Aug. 2013.
G. Gui and F. Adachi, Adaptive sparse channel estimation for timevariant MIMO-OFDM systems, in International Wireless
Communications and Mobile Computing Conference (IWCMC), Cagliari,
Italy, 1-5 July 2013, pp. 878883.
G. Gui, S. Kumagai, A. Mehbodniya, and F. Adachi, Variable is good:
Adaptive sparse channel estimation using VSS-ZA-NLMS algorithm,
in International Conference on Wireless Communications and Signal
Processing (WCSP), Hangzhou, China, 24-26 Oct. 2013, pp. 15.
E. J. Candes, M. B. Wakin, and S. P. Boyd, Enhancing sparsity by
reweighted L1 minimization, Journal of Fourier Anaysis and
Applications, vol. 14, no. 5-6, pp. 877905, Dec. 2008.
Y. Chen, Y. Gu, and A. O. Hero, Sparse LMS for system identification,
in IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), Taipei, Taiwan, 19-24 Apr. 2009, pp. 31253128.
H. Shin, A. H. Sayed, and W. Song, Variable step-size NLMS and
affine projection algorithms, IEEE Signal Processing Letters, vol. 11, no.
2, pp. 132135, Sept. 2004.
B. Widrow and S. D. Stearns, Adaptive signal processing. New Jersey:
Prentice Hall, 1985.
H. Cho and S. W. Kim, Variable step-size normalized LMS algorithm
by approximating correlation matrix of estimation error, Signal
Processcessing, vol. 90, no. 9, pp. 27922799, Sept. 2010.
G. Gui, A. Mehbodniya, and F. Adachi, Least mean square/fourth
algorithm for adaptive sparse channel estimation, in IEEE International
Symposium on Personal, Indoor and Mobile Radio Communications
(PIMRC), London, UK, 8-11 Sep. 2013, pp. 296300.