On The Performance of The MIMO Zero-Forcing Receiver in The Presence of Channel Estimation Error
On The Performance of The MIMO Zero-Forcing Receiver in The Presence of Channel Estimation Error
On The Performance of The MIMO Zero-Forcing Receiver in The Presence of Channel Estimation Error
3, MARCH 2007
805
I. I NTRODUCTION
ULTIPLE-Input Multiple-Output (MIMO) wireless antenna systems have been recognized as a key technology for future wireless communications. One common
approach to exploit the capacity of MIMO systems is to
employ spatial multiplexing where independent information
streams are transmitted from the antennas. These information streams are then separated at the receiver by means of
appropriate signal processing techniques such as maximum
likelihood (ML) which achieves optimal performance or linear
receivers like Zero-Forcing (ZF) which provide sub-optimal
performance but offer significant computational complexity
reduction with tolerable performance degradation. In these
receiver structures however accurate channel state information
(CSI) is essential for their proper operation.
In this letter we investigate the performance of MIMO
ZF receivers in the presence of channel estimation error. In
practice accurate CSI will not always be available and it is
thus important to be able to characterize the performance
of MIMO receivers in the presence of channel estimation
error. In previous work [1], [2] provide an exponentially tight
upper bound for the performance of MIMO ZF receivers and
Manuscript received May 24, 2005; revised March 05, 2006; accepted April
27, 2006. The editor coordinating the review of this paper and approving it
for publication is T. Guess. This work was supported by the Hong Kong RGC
grant HKUST 6149/03E.
The authors are with the Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Clear Water
Bay, Hong Kong (email: {eeelva, eeedward}@ust.hk; {eermurch, eewhmow,
eecheng, eeknlau}@ee.ust.hk).
Digital Object Identifier 10.1109/TWC.2007.05384.
c 2007 IEEE
1536-1276/07$25.00
806
r = G(Hs + n) = s + Gn,
(2)
1 H
where G = H = HH H
H , denotes the pseudoinverse operation.
However in practice the complex channel gain matrix H has
to be estimated at the receiver for retrieving the transmitted
data symbol vector and imperfect channel estimates arise
in any practical estimation scheme. Following the channel
estimation model for MIMO systems in [5] and [6], we model
the noisy channel estimates as
E n
1 H 1
2
2
H
H H
= N0 + e Nt Es + e N0 tr H H
,
= H + e,
H
(10)
(3)
(11)
1
=
1 + e2
(4)
where hij , h
respectively and denotes the complex conjugate operation.
We define the normalized mean square error (NMSE) of the
channel estimation as
ij 2
E hij h
,
(5)
N M SE =
2
E hij
and it can be easily shown that the NMSE is related to e by
N M SE = e2 .
(6)
where
r = G(Hs
+ n),
(7)
=H
= H + e ,
G
(8)
k = 1, 2, . . . , Nt ,
kk
(12)
Nr
Nt
Nt 1
=
+
. (14)
Nr Nt (Nr Nt )2 1 Nr Nt + 1
(13) and (14), we obtain the variance of
From
H 1
for Nr > Nt + 1 as
tr H H
H 1
NN
t r
.
V ar tr H H
=
(Nr Nt )2 (Nr Nt )2 1
(15)
For practical numbers of transmit and
we
receiveantennas,
1
H
can see from (13) that the mean of tr H H
is quite
small when Nr > Nt , as it is no larger than the number
of transmit antennas. Observing (15) we can see that the
Nt=4, Nr=5
10
10
Nt=4, Nr=6
10
CDF
CDF
10
10
tr((H H) )
10
10
Nt=2, N r=6
10
Nt=4, N r=4
10
tr((HHH)1)
10
CDF
10
CDF
807
10
kk
k = 1, 2, . . . , Nt .
(16)
1
From [1], [2] we know that 1
HH H
is a chikk
square distributed random variable with 2(Nr Nt +1) degrees
of freedom. Since the SNR distribution of each stream is the
same, we drop the subscript k and denote the SNR per symbol
of each stream as = s x, where
s =
Es /N0
,
1 + e2 Nt Es /N0
(17)
M (12k ) M 1
log2
k1
i2
2
Pb =
(1)
M log2 M k=1
i=0
i 2k1
1
3
s x
k1
2
+
Q (2i + 1)
, (18)
2
M 1
M
10
Fig. 1.
0.5
tr((HHH)1)
1.5
10
10
20
tr((HHH)1)
30
40
estimation error:
log2 M (12 ) M1
2
BERMQAM
=
M log2 M k=1
i=0
k1
i 2k1
i2
1
k1
M
+
(1)
2
2
M
D
D+1
j
1
D+j 1
(1 + i )
(1 i )
j
2
2
j=0
where i =
BERMP SK
=
(20)
3(2i + 1)2 s
, D = Nr Nt .
2(M 1) + 3(2i + 1)2 s
2
max(log2 M, 2)
min(2,M/4)
i=1
D+1
j
1
D + j 1
(1 i )
(1 + i )
,
j
2
2
j=0
(21)
s sin2 (2i 1)/M
, D = Nr Nt .
where i =
1 + s sin2 (2i 1)/M
If only the dominant terms (i = 0, 1) in (20) are considered
min(2,M/4)
[10],
we obtain
(2i 1)
2
Pb
Q
2
s x sin
,
=
max(log2 M, 2)
M
i=1
(19)
BERM QAM
2( M 1)
=
M log2 M
2( M 2)
+
M log2 M
D
D+1
1
D+j
(1 0 )
j
2
j=0
D
D+1
1
D+j
(1 1 )
j
2
j=0
j
1
(1 + 0 )
2
j
1
(1 + 1 )
2
(22)
Note that (20) and (22) are for M-ary square QAM constellations, however it is very easy and straightforward to derive
BER formula for PAM and rectangular QAM constellations
in a similar manner.
808
10
D+j
2D + 1
=
,
(25)
j
D+1
j=0
the expression of the error floor for BPSK and QPSK modulation can be simplified as
2 D+1
e Nt
2D + 1
EFBP SK =
,
(26)
D+1
4
2 D+1
e Nt
2D + 1
EFBP SK =
,
(27)
D+1
2
when e << 1 and Nt is relatively small.
10
BER
10
10
Simu e=0%
Appro e=0%
Simu e=5%
Appro e=5%
Simu e=10%
Appro e=10%
Simu e=20%
Appro e=20%
10
10
10
15
20
25
30
35
40
10
10
BER
D+1
j
1
D + j 1
(1 )
(1 + ) ,
(23)
j
2
2
j=0
10
Simu e=0%
Appro e=0%
Simu e=5%
Appro e=5%
Simu e=10%
Appro e=10%
Simu e=20%
Appro e=20%
10
10
10
15
20
25
30
35
40
10
10
Simu e=0%
Appro e=0%
Simu e=5%
Appro e=5%
Simu e=10%
Appro e=10%
Simu e=20%
Appro e=20%
10
e=20%
2
10
10
BER
BER
809
10
e=10%
e=5%
3
10
10
10
10
15
20
25
30
35
40
H =
. According to [5],
Etrain YStrain Strain Strain
[6], the optimal Strain for the ML channel estimator takes
the form of Strain = Ntrain , where H = INt and
the resulting Nt Nr estimation errors are independent and
Nt
identically distributed as CN 0,
. In the
Ntrain Etrain /N0
simulation we set Ntrain = Nt , so the relationship between
the SNR of the training phase and the quality of the channel
estimation is given by Etrain /N0 = 1/e2 . Note that we use the
un-normalized estimate in the paper because the relationship
between the training SNR and the quality of the channel
estimation is not explicitly known for some channel estimation
algorithms. We test three cases with e = 5%, 10%, 20% and
tune the Etrain /N0 accordingly. From the figure we can see
that our analysis, i.e. (21), shows very good agreement with
the simulation results, where Eb /N0 = Es /(M N0 ). And
we observe that the BER performance of MIMO ZF receiver
is quite sensitive to channel estimation error and degrades
significantly as e increases. According to our approximation
in (26), the error floor for BPSK with Nt = Nr = 4 would
be 2.5 103 and 1 102 for e = 5% and e = 10%
respectively. These results agree very well with our simulation
results shown in Fig. 2.
In Fig. 3-Fig. 4, BER performance of QPSK and 16QAM
modulations is investigated for system with Nt = 4, Nr = 5
and Nt = 4, Nr = 6 respectively for e = 5%, 10%, 20% using
the channel estimation error model in (3). Again these figures
demonstrate the tightness of our analysis, i.e. (21) and (22).
For e = 5%, 10%, according to our approximation in (27)
the error floor for QPSK with Nt = 4, Nr = 5 would be
7.5 105, 1.2 103 respectively, and these agree with our
simulation results in Fig. 3.
In Fig. 5, the effect of the number of transmit antennas
Nt on the BER performance of QPSK is investigated with
D = 1. The simulation results agree with our analysis that as
Nt increases the BER gets worse with the same value of D
and e.
10
10
e=0
15
20
25
30
35
40
V. C ONCLUSION
In this letter, performance of MIMO ZF receivers over
uncorrelated Rayleigh flat fading channels in the presence of
channel estimation error is investigated. Approximate postprocessing SNR distribution and tight closed-form approximations for the BER expression of M-QAM and M-PSK are
derived. It is found that contrary to the case when there is
no channel estimation error, the BER performance of MIMO
ZF receiver is not only a function of D = Nr Nt , but
also a function of Nt . Besides, for the same value of D and
e the BER performance gets worse as Nt increases. Another
impact of imperfect channel estimation is that when SNR is
high, BER does not approach zero but approaches an error
floor whose value depends on Nt , Nr and e.
A PPENDIX I
In this appendix, we derive the covariance matrix of the
of MIMO ZF receiver in the
effective post-processing noise n
presence of channel estimation error.
= H n eH s eH H n, the covariance
Since n
matrix can be computed as
H
n
E n
=E H n eH s eH H n
H
H n eH s eH H n
H
H
H
=N0 H H eN0 E H H H H
H
H
+e2 Es H E H H eN0 E H H H
H
H
+e2 N0 H E H H H H
H
1
=N0 HH H
+ e2 Es H Nt INr H
1 H H
H
+e2 N0 H E HH H
1 H 1
2
2
H
= N0 + e Nt Es + e N0 tr H H
H H
810
H
H 1
where
,
H we used the factH that H H = H H H
E ss = Es INt , E nn
= N0 INr , E
= Nt INr
1 H
1
= tr HH H
INr .
and E HH H
R EFERENCES
[1] J. H. Winters, J. Salz, and R. D. Gitlin, The capacity increase of wireless
communication systems with antenna diversity, in Proc. 1992 Conference
Information Sciences Syst..
[2] J. H. Winters, J. Salz, and R. D. Gitlin, The impact of antenna
diversity on the capacity of wireless communication systems, IEEE
Trans. Commun., vol. 42, no. 2/3/4, pp. 1740-1751, Feb./Mar./Apr. 1994.
[3] Ho-Yin Fan, MIMO Detection Schemes for Wireless Communications,
MPhil. thesis, the Hong Kong University of Science and Technology,
2002.
[4] X. Zhang and B. Ottersten, Performance analysis of V-BLAST structure
with channel estimation errors, in Proc. SPAWC03, pp. 487-491.
[5] T. L. Marzetta, BLAST training: estimating channel characteristics
for high-capacity space-time wireless, in Proc. 37th Annual Allerton
Conference Communications, Control, and Computing.