Performance Analysis of Multiuser MIMO Systems With Zero Forcing Receivers
Performance Analysis of Multiuser MIMO Systems With Zero Forcing Receivers
Performance Analysis of Multiuser MIMO Systems With Zero Forcing Receivers
, Jinwoo Choe
n] =
2
I
Nr
and
x stands for the transmitted symbol vector of length N
t
with
E[x
x] =
2
x
I
Nt
. Here I
a
represents an identity matrix of size
a and ()
H)
1
H
n
=
N
t
[(H
H)
1
]
nn
where [A]
ij
denotes the (i, j)-th entry of A and represents
the average received SNR. Assuming N
r
N
t
,
n
is a Chi-
squared random variable with a degree of freedom equal to
2(N
r
N
t
+1) [5]. Here all
n
s are i.i.d. Thus the probability
density function (PDF) of
n
can be written as
f
n
() =
N
t
e
Nt/
(N
r
N
t
)!
_
N
t
_
NrNt
, > 0. (1)
Since the MIMO channel is decomposed into N
t
parallel
channels with the ZF equalizer, the ergodic capacity can be
obtained as the sum capacity of N
t
parallel channels [5]
C
ZF
=
Nt
n=1
E [log
2
(1 +
n
)] (bps/Hz) . (2)
When the ZF receiver model in Fig. 1 is extended to the
multiuser case, we consider the following two architectures:
SDM : Data streams of one user are transmitted through
all transmit antennas at the BS. With the Max-Rate
scheduling criterion [3], the BS can select the user with
the largest
n
log
2
(1 +
n
) at each time slot.
SDMA : The BS can select up to N
t
different user
streams at a given time which can best exploit each
channel. Denoting
k
n
as the output SNR at the nth spatial
subchannel of the kth user, the BS allocates the nth
spatial channel to the user with the highest log
2
(1 +
k
n
)
for n = 1, , N
t
.
III. PERFORMANCE ANALYSIS OF MULTIUSER MIMO
SYSTEMS WITH LINEAR RECEIVERS
In this section, we analyze the performance of multiuser
MIMO systems with ZF receivers. As mentioned earlier,
an antenna conguration which exploits multiuser diversity
most efciently is the system employing an equal number of
antennas at both the transmitter and receiver [3]. Therefore,
from now on, we assume N
t
= N
r
= N and all users
experience statistically independent identical fading process.
Applying the ZF equalizers at the receiver, N spatial channels
are generated. We refer to the data stream transmitted via the
nth spatial channel as the nth stream.
A. Closed Form Expression for Multiuser Capacity
In the case of N
t
= N
r
= N, the PDF of the output SNR
n
in (1) reduces to an i.i.d. exponentially distribution which
can be expressed as
f
n
() =
N
e
N/
. (3)
For the SDMA where the BS can select up to N users
to exploit each antenna, N independent links with K users
whose output SNR for each user can be modeled as (3).
Dening the kth order statistics
n
k:K
as the kth smallest of
1
n
,
2
n
, ,
K
n
[4],
n
K:K
represents the largest output SNR
for the nth stream among K users which is selected through
the max-rate scheduler [3]. The max-rate scheduling scheme
is an opportunistic scheduling scheme which maximizes the
total throughput in the system.
According to the order statistics [4], the PDF of
n
K:K
can
be obtained as
f
n
K:K
() =
dF
n
()
K
d
= KF
n
()
K1
f
n
() . (4)
By substituting (3) to (4), the capacity of the SDMA with K
users is expressed as
C
SDMA
ZF
(K) =
N
n=1
_
0
log
2
(1+)K
N
e
N/
(1e
N/
)
K1
d
=
KN
2
_
0
log
2
(1+)e
N/
(1e
N/
)
K1
d.(5)
Using the binomial power series expansion and the integral
equality [6, pp. 568]
_
0
e
x
ln(1 + x)dx =
1
e
/
E
1
_
_
where the exponential integral function E
1
(x) is dened as
_
x
e
t
t
1
dt, (5) can be rewritten as
C
SDMA
ZF
(K) =
KN
2
log
2
(e)
_
0
ln(1 + )e
K1
k=0
_
K 1
k
_
(1)
k
e
k
d
= KN log
2
(e)
K1
k=0
(1)
k
1
k + 1
_
K 1
k
_
exp
_
N
(k + 1)
_
E
1
_
N
(k + 1)
_
(6)
2136
Note that E
1
(x) can be efciently computed by the approxi-
mation as in [7, pp. 231].
In contrast, for SDM systems, the BS selects the user with
the highest sum rate,
n
log
2
(1 +
k
n
). However, it is a
problem of great difculty to describe an exact distribution
of
n
log
2
(1 +
k
n
). To simplify the problem, we modify the
SDM scheduling scheme as follows:
The BS selects the kth user with the largest
n
k
n
instead of
n
log(1 +
k
n
).
Transmit power is distributed equally to each stream in
order to make the received SNR equal for each stream.
Note that the user selection based on
n
k
n
instead of
n
log
2
(1 +
k
n
) does not degrade the performance since
both metrics are monotonically increasing functions and the
polynomial expansion of
n
(1+
k
n
) includes the term
n
k
n
.
Dene a new random variable
k
which stands for a sum of
all
n
s as
k
=
N
n=1
k
n
. Based on the above assumptions,
the capacity formula of the K-user SDM can be rewritten as
C
SDM
ZF
(K) = E
_
max
k
_
N
n=1
log(1 +
k
n
)
__
N
_
0
log
2
_
1 +
x
N
_
f
K:K
(x)dx (7)
where f
K:K
(x) indicates the Kth order statistics of
k
. Since
all
k
n
s are i.i.d, the characteristic function of
k
[8] can be
obtained by multiplying the characteristic functions of
k
n
s.
Therefore, the characteristic function
k
() is written as
k
() =
N
n=1
k
n
() = (1 + j
N
)
N
. (8)
Note that the right-hand side of (8) is the characteristic
function of the gamma distribution [9] with the PDF
f
k
() =
_
N
_
N
N1
e
N/
(N)
(9)
where (N) = (N 1)! for integer N [7].
Since the cumulative density function (CDF) of (9) is given
as [9]
F
k
(x) =
(N, (N/)x)
(N)
,
where (a, x) denotes an incomplete gamma function dened
as (a, x) =
_
x
0
t
a1
e
t
dt [6], the PDF of the Kth order
statistics of
k
can be represented as
f
K:K
(x) = K
_
(N, (N/)x)
(N)
_
K1
(N/)
N
x
N1
e
Nx/
(N)
(10)
To integrate (7) efciently, the PDF in (10) is converted into
[10]
f
K:K
(x) =
K
(N)
K1
k=0
k(N1)
i=0
(1)
k
_
K 1
k
_
a
k
i
_
N
_
N+i
e
(k+1)(N/)x
x
N+i1
(11)
where a
k
i
for 0 i k(N 1) is recursively dened as
a
k
0
= 1, a
k
1
= k,
a
k
i
=
1
i
min(i,N1)
n=1
n(k + 1) i
n!
a
k
in
, 2 i < k(N 1)
a
k
i
=
1
(N)
k
, for i = k(N 1) . (12)
By applying (11) to (7) and using the integral identity [11]
_
0
ln(1 + t) e
t
t
n1
dt = (n 1)!e
j=1
(j n, )
j
where (, x) represents another incomplete gamma function
which is dened as (, x) =
_
x
t
1
e
t
dt [6], the capacity
of the K-user SDM system is nally computed by
C
SDM
ZF
(K) =
KN log
2
(e)
(N 1)!
K1
k=0
k(N1)
i=0
(1)
k
_
K 1
k
_
a
k
i
_
N
2
_
N+i
_
0
ln(1 + t)e
(k+1)
N
2
t
t
N+i1
dt
=
KN log
2
(e)
(N 1)!
K1
k=0
k(N1)
i=0
(1)
k
_
K 1
k
_
a
k
i
_
N
2
_
N+i
(N+i1)!
e
(k+1)
N
2
N+i
j=1
(j N i, (k + 1)
N
2
)
_
(k + 1)
N
2
_
j
. (13)
B. Capacity Bounds
Although the derived formulas for SDM and SDMA systems
in (6) and (13) in the previous section are quite accurate, they
do not provide much information about primary factors which
affect the system performance. In this subsection, in order to
gain an insight on multiuser diversity gains, we develop upper
bounds for channel capacity for both the SDM and SDMA
schemes with K users.
Denoting the mean and variance of the capacity of single
user systems as and
2
, respectively, the system throughput
with K users is bounded as [4]
C
K
+
(K 1)
2K 1
. (14)
For the special case, if the capacity of each user obeys a
Gaussian distribution, a tighter bound can be obtained from
the weak law of large numbers [3] as
C
K
+
2
2
ln K . (15)
As the Shannon capacity can be approximated by log
2
at high SNR ranges (N << ) [2], the average value of
2137
the achievable throughput of the nth stream for the SDMA
is rewritten as
n
SDMA
=
_
0
log
2
(1 + )
_
N
_
exp(
N
)d
log
2
e
_
0
ln
_
N
_
exp(
N
)d .
Then, the mean of C
SDMA
ZF
(K) in (5) with N transmit
antennas and K = 1 can be expressed as
SDMA
ZF
= N
n
SDMA
.
Dene
SDMA
and
2
SDMA
as the mean and variance
of C
SDMA
ZF
(1), respectively. Using the integral identity [6,
pp.567]
_
0
e
x
ln xdx =
1
( + ln )
where = 0.5772 denotes the Euler constant,
SDMA
can
be obtained as
SDMA
= N log
2
e
_
ln
N
_
.
Also
2
SDMA
can be computed as
2
SDMA
= E[(N log
2
(1 + ))
2
]
2
SDMA
= N
2
_
(log
2
e)
2
_
2
6
+
_
+ ln
N
_
2
_
(
i
SDMA
)
2
_
= N
2
(log
2
e)
2
2
6
. (16)
where we have used the integral identity [6, pp.567]
_
0
e
x
(ln x)
2
dx =
1
2
6
+ ( + ln )
2
_
.
In contrast, for SDM systems, the average capacity is
approximated as
SDM
= N
_
0
log
_
1 +
N
__
N
_
N
N1
e
N/
(N 1)!
d
(log
2
e)
_
N
_
N
N
(N 1)!
_
0
ln
N
N1
e
N/
d (17)
for N << . Using the Euler psi function (z) in [7] and the
integral identity [6, pp.569]
_
0
x
1
e
x
ln xdx =
()[() ln ] ,
(17) can be computed as
SDM
log
2
e
_
N
2
_
N
N
(N 1)!
(N)
_
N
2
_
N
_
(N) + ln
N
2
_
= N log
2
e
_
(N) + ln
N
2
_
. (18)
In addition, the variance
2
SDM
is obtained as
2
SDM
N
2
_
E
_
_
ln
N
_
2
_
2
SDM
_
(19)
where E
_
_
ln
N
_
2
_
is computed as
E
_
_
ln
N
_
2
_
_
N
_
N
(log
2
e)
2
(N 1)!
_
0
(ln
N
)
2
N1
e
N/
d
=
(log
2
e)
2
(N 1)!
_
N
2
_
N
_
0
(ln x)
2
x
N1
e
N
2
x
dx. (20)
Applying the integral identity [6, pp.572]
_
0
x
1
e
x
(ln x)
2
dx =
()
_
[() ln ]
2
+ (2, )
_
where (z, q) represents the Riemann zeta function[7], (20)
can be rewritten as
E
_
_
ln
N
_
2
_
(log
2
e)
2
_
[(N) ln
N
2
]
2
+ (2, N)
_
(21)
After some manipulations, substituting (21) to (19) yields
2
SDM
N
2
(log
2
e)
2
_
2
6
N1
n=1
1
n
2
_
(22)
where we have used
n=1
1/n
2
=
2
/6 [7]. As will be shown
in section IV, the channel capacity of the SDM with K users
can be closely expressed by the Gaussian approximation in
(15) with (18) and (22).
The performance of the SDM and the SDMA can be
compared by the following lemma.
Lemma 1 (Capacity Gain of the SDMA Systems): Dene
the capacity gain of the SDMA compared with the SDM as
. Then we have
lim
K
= lim
K
C
K
SDMA
C
K
SDM
> 1 .
Proof: By substituting the results of (16) and (22) to
(14), we have
lim
K
C
K
SDMA
C
K
SDM
=
SDMA
SDM
=
2
6
N1
n=1
(1/n
2
)
> 1. (23)
Thus we can see that the capacity gain of SDMA systems over
SDM systems increases as the number of antenna N grows.
IV. NUMERICAL RESULTS
In this section, we present numerical results to verify
the performance analysis of MIMO systems with the linear
equalizer at the receiver. To evaluate performance, we assume
i.i.d Rayleigh fading channels which is supposed to be at
over transmission bandwith. Fig. 2 exhibits the capacity of
2138
0 5 10 15 20 25 30 35
5
10
15
20
25
30
35
40
45
50
SNR (dB)
C
(
b
p
s
/
H
z
)
SDMA (simulation)
SDMA (analysis)
SDM (simulation)
SDM (analysis)
Fig. 2. Capacity for the MIMO systems with ZF receivers (N = 4, K = 40)
MIMO systems with four transmit and four receive antennas
and 40 users. To verify the accuracy of the derived closed
form expressions, we compare the analysis results with the
Monte Carlo simulations. As shown in the gure, the derived
formulas match well with two MIMO system models. Also
we see that the performance gain of the SDMA system over
the SDM is around 3dB.
In Fig. 3, the performance of SDM systems with various
congurations is illustrated. Fig.3 shows that approximations
with (15) describe well the SDM systems. In addition, we
compare the SDM system with the optimal beamforming case
with Gaussian approximation in [3]. Note that the optimal case
can not be applied to the SDMA since orthogonality of unitary
vectors cannot be satised if different users are assigned to
each antenna. As shown in the gure, the performance of the
SDM comes within 3dB compared with the optimal case. Also
the performance of the SDMA in Fig. 2 approaches the optimal
case. To perform the optimal eigen beamforming, channel
information for all propagation paths should be fed back to
the BS. Thus the amount of feedback overhead increases as
N
t
, N
r
and K grow. In comparison, for MIMO systems with
ZF receivers, only scalar SNRs or acceptable data rates of
each stream are sufcient for the feedback information and the
performance gap decreases by combining with the multiuser
diversity. Therefore, the MIMO systems employing the ZF
receiver are better suited for systems where the bandwidth
for feedback is limited as a reasonable performance can be
achieved with reduced feedback overhead.
V. CONCLUSION
In this paper, we analyze the performance of multiuser
MIMO systems with ZF receivers. We rst derive the closed
form expressions of capacity for MIMO systems with the ZF
receiver in multiuser scenarios. As shown in simulation result
section, the derived expressions accurately describe capacity
0 5 10 15 20 25 30 35
0
10
20
30
40
50
60
SNR (dB)
C
(
b
p
s
/
H
z
)
SDM (N=6,K=20,simulation)
SDM (N=6,K=20,approximation)
SDM (N=4,K=40,simulation)
SDM (N=4,K=40,approximation)
SDM (N=2,K=60,simulation)
SDM (N=2,K=60,approximation)
eigen beamforming (N=4,K=40)
Fig. 3. Performance of SDM systems with various congurations
for the multiuser MIMO system. Also, we compare the SDM
and the SDMA by obtaining the capacity bounds and show
that the performance gain decreases as the number of antenna
increases.
ACKNOWLEDGEMENT
This research was supported by the MIC (Ministry of
Information and Communication), Korea, under the ITRC
(Information Technology Research Center) support program
supervised by the IITA (Institute of Information Technology
Assessment).
REFERENCES
[1] I. E. Telatar, Capacity of Muti-antenna Gaussian Channels, AT&T Bell
Laboratories, Lucent Technologies, Murray Hill, NJ, 1994.
[2] G. Foschini and M. Gans, On limits of wireless communication
in a fading environment when using multiple antennas, AT&T Bell
Laboratories Tech. Memo, September 1995.
[3] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, Multiple-Antenna
Channel Hardening and Its Implications for Rate-Feedback and Schedul-
ing, IEEE Transaction on Information Theory, vol. 50, pp. 18931909,
September 2004.
[4] H. David and H. Nagaraja, Order Statistics. Wiley, 2003.
[5] D. A. Gore, J. Robert W. Heath, and A. J. Paulraj, Transmit Selection
in Spatial Multiplexing Systems, IEEE Communications Letters, vol. 6,
pp. 491493, November 2002.
[6] I. S. Gradshteyn and I. M. Razhik, Table of Integrals, Series, and
Products. Academic Press, 6th ed., 2000.
[7] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions
with Formulas, Graphs, and Mathematical Tables. Dover, 1970.
[8] A. Papoulis and S. U. Pillai, Probability, Random Variables and Stochas-
tic Processes. Mc Graw Hill, 4th ed., 2002.
[9] J. G. Proakis, Digital Communications. McGrawHill, 4th ed., 2001.
[10] G. Fedele, N-Branch Diversity Reception of M-ary DPSK Signals in
Slow Fading and Nonselective Nakagami Fading, European Transac-
tions on Telecommunications, vol. 7, pp. 119123, March-April 1996.
[11] M.-S. Alouini and A. J. Goldsmith, A Unied Approach for Calculating
Error Rates of Linearly Modulated Signals over Generalized Fading
Channels, IEEE Transaction on Communications, vol. 47, pp. 1324
1334, September 1999.
2139