IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 25, NO.

7, SEPTEMBER 2007 1457

MIMO Broadcast Scheduling with
Limited Feedback
Wei Zhang, Member, IEEE, and Khaled Ben Letaief, Fellow, IEEE
AbstractWe consider multiuser scheduling with limited feed-
back of partial channel state information in MIMO broadcast
channels. By using spatial multiplexing at the base station
(BS) and antenna selection for each user, we propose a mul-
tiuser scheduling method that allocates independent informa-
tion streams from all M transmit antennas to the M most
favorable users with the highest signal-to-interference-plus-noise
ratio (SINR). A close approximation of the achievable sum-rate
throughput for the proposed method is obtained and shown to
match the simulation results very well. Moreover, two reduced
feedback scheduling approaches are proposed. In the rst ap-
proach, which we shall refer to as selected feedback scheduling,
the users are selected based on their SINR compared to a
predesigned threshold. Only those selected users are allowed to
feed back limited information to the BS. The resultant feedback
load and achievable throughput are derived. It will then be
demonstrated that with a proper choice of the threshold, the feed-
back load can be greatly reduced with a negligible performance
loss. The second reduced feedback scheduling approach employs
quantization for each user, in which only few bits of quantized
SINR are fed back to the BS. Performance analysis will show
that even with only 1-bit quantization, the proposed quantized
feedback scheduling approach can exploit the multiuser diversity
at the expense of slight decrease of throughput.
Index TermsAntenna selection, broadcast channel, feedback,
multiple-input multiple-output, multiuser diversity.
I. INTRODUCTION
M
ULTIPLE-INPUT multiple-output (MIMO) systems
with a number of transmit and receive antennas hold
the promise of achieving high capacity over wireless channels
[1], [2]. One key benet of MIMO is the resultant spatial
diversity gain which can be exploited to get a more reliable
reception performance. Space-time block coding (STBC) is
a MIMO technique which can achieve the maximum spatial
diversity gain in an efcient way [3]. Another merit of using
MIMO arises from the high spatial multiplexing gain which
can be obtained by sending independent information streams
over any transmit-receive antenna pair simultaneously [4], [5].
In the context of a broadcast channel (BC) in a MIMO
cellular system, the spatial diversity gain or multiplexing gain
can still be achieved for a single user. From the system
point of view, however, the total throughput may not be
optimal because a potential inherent multiuser diversity is not
exploited. Multiuser diversity is a form of selection diversity
Manuscript received June 1, 2006; revised Jan. 5, 2007. This work was
supported in part by the Hong Kong Research Grant Council under grant No.
HKUST6250/04E.
W. Zhang and K. B. Letaief are with the Department of Electronic and
Computer Engineering, The Hong Kong University of Science and Tech-
nology, Clear Water Bay, Kowloon, Hong Kong (e-mail: eewzhang@ust.hk,
eekhaled@ece.ust.hk).
Digital Object Identier 10.1109/JSAC.2007.070918.
among users which arises from the independent fading chan-
nels between the base station (BS) and the multiple users [6],
[7]. To take advantage of the multiuser diversity, a variety
of multiuser scheduling algorithms have been proposed for
MIMO downlink transmissions in [8][14]. But these are
mostly based on time-division-multiple-access (TDMA) in
which the BS transmits to only one user at a time. It was later
shown that the maximum achievable sum-rate using TDMA
for MIMO broadcast channels is only a small fraction of the
sum-rate capacity of MIMO broadcast channels [15].
Dirty paper coding (DPC) has been shown to be an optimal
approach to achieve MIMO broadcast channel capacity by
serving multiple users simultaneously [16], [17]. Compared
with TDMA, DPC can achieve a linear increase of capacity
in min(M, K) times [15], where M is the number of transmit
antennas and K is the number of users. To alleviate the
computation burden of DPC, simple sum power iterative
water-lling (IWF) algorithms were proposed in [18], [19].
For a large number of users, some practical techniques with
reduced complexity to achieve the sum capacity have also been
investigated in [20][24]. To reduce the amount of channel
feedback and the computation complexity, user selections were
also considered. Through the use of a simple norm-threshold
feedback protocol, the maximal scaling of the sum rate is
achievable [25]. In [26], only those users with the singular
values above a set threshold are allowed to feed back the
channel information to the BS. In [27], a low-complexity
suboptimal user selection scheme, which was referred to
as the semi-orthogonal user selection (SUS) algorithm, was
proposed. When the user selection algorithms in [26] and
[27] are combined with zero-forcing (ZF) beamforming, the
optimum sum-rate throughput can be asymptotically achieved
when K goes to innity.
The aforementioned DPC and the related low-complexity
algorithms require perfect knowledge of the channel state in-
formation at the transmitter (CSIT). However, it is not practical
to assume that the channel coefcients are perfectly known at
the transmitter. Furthermore, with an increasing number of
antennas and users, the feedback amount of CSI becomes
huge. In situations where the feedback is severely limited,
quantization of the information needed at the transmitter seems
to be attractive [28]. With nite rate feedback, it was shown
in [29] that the transmitter can obtain quantized CSI through
a random quantization codebook. Full multiplexing gain can
then be realized if the number of feedback bits scales with
the signal-to-noise ratio (SNR). In [30], and by assuming
perfect knowledge of CSI at the receiver only, the signal-to-
interference-plus-noise ratio (SINR) is made available to the
0733-8716/07/$25.00 c 2007 IEEE
1458 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 25, NO. 7, SEPTEMBER 2007
transmitter and only those users with high SINR are allowed
to receive information. The throughput is then shown to be
asymptotically optimal and scale as M log log(KN), where
N is the number of receive antennas at each user.
In this paper, we will investigate several SINR-based
scheduling algorithms for MIMO broadcast channels in the
absence of CSIT. Firstly, we propose a MIMO broadcast
scheduling with the feedback of SINR from all users, namely
full feedback scheduling. In this scheme, spatial multiplexing
and receive antenna selection are both employed. With antenna
selection for each user and feedback of SINR from all users,
the BS transmits M independent information streams to M
users with high SINR simultaneously. This scheme is similar
to the one proposed by Sharif and Hassibi in [30]. But ours
differs from [30] in that receive antenna selection is used for
each user, whereas in [30] every receive antenna is employed
as an individual user. Because forcing multiple closely spaced
antennas belonging to a single user to receive different signals
would require extra cost and power, the antenna selection
scheme is preferred. Moreover, it will be shown later that our
scheme requires less feedback load without loss of throughput
as opposed to [30].
We further propose two MIMO broadcast scheduling algo-
rithms with limited feedback. In the rst one, which we shall
refer to as selected feedback scheduling, users are selected
based on their SINR compared to a set threshold. Only those
selected users are allowed to send back their SINR to the BS.
The achievable sum-rate throughput and the resultant feedback
load reduction are then studied over Rayleigh fading chan-
nels. Performance analysis will demonstrate that the feedback
load can be substantially reduced with a negligible loss of
throughput. It should be mentioned that a similar technique,
SNR thresholding, was proposed in [31] in the context of
single-antenna multiuser scheduling. In our second scheduling
algorithm with limited feedback, namely, quantized feedback
scheduling, only few bits of quantized SINR of each user are
sent back to the BS. It should be remarked that quantized
feedback was applied to multiuser diversity in single-antenna
systems [32][34]. In contrast, here the quantization of SINR
is employed for each user in a multiuser MIMO system.
Performance analysis will show that even with only 1-bit
quantization, the achievable sum-rate throughput increases
with the increasing number of users.
The rest of this paper is organized as follows. Section
II introduces the system model. In Section III, a multiuser
scheduling algorithm with SINR feedback from all users is
proposed for MIMO broadcast channels and its achievable
throughput is obtained numerically. Section IV proposes a
selected feedback scheduling algorithm. The achievable sum-
rate throughput and the feedback load are derived. Section
V proposes a quantized feedback scheduling algorithm. In
particular, 1-bit feedback scheduling is investigated. Section
VI presents further comparison results of all the proposed
scheduling algorithms. Finally, we draw our conclusions in
Section VII.
II. SYSTEM MODEL
We consider the downlink of a single-cell wireless sys-
tem where the BS with M antennas communicates with K
geographically dispersed users, each having N antennas. In
a typical cellular system, we often have K >> M and
M N. Assume that only J out of K users are allowed to
communicate with the BS simultaneously at any given time
slot t. Let /(t) be the set of active users which are assigned
a non-zero rate at time slot t. Thus, the cardinality of /(t) is
[/(t)[ = J with 1 J K. Let user k (1 k K) be an
active user at time slot t, whose received signal is given by
Y
t
k
=

k
H
t
k
X
t
+W
t
k
, k /(t) , (1)
where Y
t
k
is an N 1 vector of the received signal at the kth
user in time slot t, X
t
is an M 1 vector of the transmitted
signal in time slot t, and W
t
k
is an N 1 vector of additive
noise whose entries are i.i.d. complex Gaussian with zero
mean and variance N
0
. H
t
k
is an NM channel matrix, where
the (n, m)th entry h
t
k
(n, m) (^(0, 1) is i.i.d complex
Gaussian with zero mean and unit variance and represents
the complex channel gain from the transmit antenna m to the
receive antenna n of user k at time slot t. The scalar
k
models
the power attenuation due to the path loss and shadow fading.
We assume that the total transmit power is equal to 1, i.e.,
Tr(X
t
[X
t
]
H
) = 1, where Tr() stands for the trace of a
matrix. This ensures that the total transmit power is not de-
pendent on the number of transmit antennas. The block-fading
channel model is considered in which the channel coefcient
h
t
k
(n, m), m, n, k is quasi-static and frequency non-selective
in time slot t but varies independently in different time slots.
Furthermore, the channel h
t
k
(n, m) is uncorrelated with other
channels h
t
k
(n

, m

) for any two distinct pairs (m

, n

, k

) and
(m, n, k). Throughout this paper, we assume that the MIMO
channels are perfectly known at the receiver, but are unknown
at the transmitter.
We may regard model (1) as a general description of several
scenarios in downlink transmissions as follows:
If J = K = 1, (1) represents a single user MIMO
Gaussian channel [1];
If J = K > 1, (1) represents a MIMO Gaussian
broadcast channel [16];
If J = 1, and K > 1, (1) represents a multiuser
scheduling MIMO Gaussian channel, where only one user
is scheduled at a given time [10][13];
If 1 < J < K, (1) represents a multiuser scheduling
MIMO Gaussian channel, where more than one user are
scheduled at a given time [20], [26], [27], [30], [35].
In this paper, we focus on the problem of multiuser scheduling
in the last scenario, where multiple users are scheduled
simultaneously at any time.
The symbol x(m) is transmitted through antenna m with
x(m) assumed to be independent for m = 1, , M. The
average transmit power per antenna is
1
M
, i.e. E[[x(m)[
2
] =
1
M
, thereby ensuring that the total transmit power is 1. The
received signal at the nth antenna of user k can then be derived
from (1) as
y
k
(n) =

k
M

m=1
h
k
(n, m)x(m) +w
k
(n). (2)
Assume that the signal x(m) is the desired signal for user k.
By regarding other signals x(m

), m

,= m as interference, we
ZHANG and LETAIEF: MIMO BROADCAST SCHEDULING WITH LIMITED FEEDBACK 1459
can get the instantaneous SINR of y
k
(n) as
SINR
(k)
m,n

[h
k
(n, m)[
2
M

k
+

=m
[h
k
(n, m

)[
2
, (3)
where
k
=
k
/N
0
is the average SNR of user k.
In [30], a downlink scheduling scheme is proposed with
SINR feedback. For the MIMO case, it treats each receive
antenna at any receiver as an individual user, so that there are
effectively NK single antenna receivers. With the feedback
of the maximum SINR of each user along with the index
m which achieves the maximum SINR, the BS will transmit
M independent signals to M users with the highest SINR
simultaneously. Let o
t
= 1, , M, o
r
= 1, , N and
S
u
= 1, , K. The achievable throughput R can then be
bounded by [30]
R E
_
M

m=1
log
2
_
1 + max
nS
r
,kS
u
SINR
(k)
m,n
_
_
. (4)
The asymptotic analysis in [30] shows that the achievable
throughput is M log log(NK) when K goes to innity. How-
ever, the SINR-based scheduling algorithm in [30] requires the
feedback of KN SINR values and the feedback load increases
with the increase of the number of receive antennas.
III. MIMO BROADCAST SCHEDULING WITH SINR
FEEDBACK
In this section, we propose a MIMO downlink scheduling
algorithm with antenna selection. It only requires a feedback
of total K SINR values. Performance analysis shows that
the reduced feedback scheduling does not induce any loss of
throughput and fairness compared with [30].
A. Scheduling Algorithm
The proposed scheduling algorithm with the full feedback
of SINR can be described as a two-stage algorithm: feedback
by users and scheduling by the BS.
Stage 1: Feedback
1) Initialization: Set B
k
= 0, Tx
k
= 0 and Rx
k
= 0
for k = 1, , K.
2) For k = 1 to K,
B
k
= max
mS
t
,nS
r
SINR
(k)
m,n
,
Tx
k
= arg
m
max
mS
t
,nS
r
SINR
(k)
m,n
,
Rx
k
= arg
n
max
mS
t
,nS
r
SINR
(k)
m,n
,
User k feeds back to the BS Tx
k
and B
k
.
End For
Stage 2: Scheduling
1) Initialization: Set k

m
= 0 and 1
m
= , for m =
1, , M. Get the feedback of Tx
k
and B
k
.
2) For k = 1 to K,
m = Tx
k
,
1
m
= 1
m
k.
End For
3) For m = 1 to M,
If 1
m
,=
k

m
= arg max
kI
m
B
k
.
End If
End For
The key idea behind the above scheduling algorithm is that
each user rstly selects the most favorable transmit antenna
Tx
k
and receive antenna Rx
k
for achieving the maximum
SINR B
k
and then feeds the corresponding B
k
along with
the index Tx
k
to the BS. Since receive antenna selection is
performed at each user, the total feedback load at each time
slot is K real values B
1
, , B
K
plus Klog
2
M| bits of
Tx
k
, k, where x| denotes the smallest integer larger than
x. At the BS, the users with Tx
k
= m are clustered into
the set 1
m
and then the scheduled user k

m
is decided by
achieving max
kI
m
B
k
. Once users k

m
are determined for
all m = 1, , M, the receive antenna Rx
k

m
of user k

m
is
selected to receive the signal x(m).
B. Throughput Analysis
In the following, we will derive the average achiev-
able throughput of the proposed scheduling algorithm over
Rayleigh fading channels.
Under Gaussian codes and the minimum distance decoding
at receiver, the instantaneous sum-rate of the scheduling with
the feedback of SINR is
R =
M

m=1
R
m
=
M

m=1
log
2
_
1 +SINR
(k

m
)
m,Rx
k

m
_
, (5)
where SINR
(k

m
)
m,Rx
k

m
is the SINR of user k

m
scheduled by
the mth transmit antenna, and Rx
k

m
is the receive antenna
selected by user k

m
. Note that
SINR
(k

m
)
m,Rx
k

m
= max
kI
m
B
k
= max
kI
m
max
mS
t
,nS
r
SINR
(k)
m,n
max
l{1, ,NK}
Z
l
, (6)
where
Z
l

[g
l
[
2
M

l
+

M1
j=1
[ g
l,j
[
2
(7)
with g
l
, g
l,j
(^(0, 1) for all l = 1, , NK
and j = 1, , M 1. Considering that 1
m
=
k[Tx
k
= m, k 1, , K and ProbTx
k
= m =
1
M
for any k and m, we can get the average of cardinality
E[1
m
[ K/M. Then, the SINR of the scheduled user is
selected from E[1
m
[ [o
t
[ [o
r
[ NK candidates as shown
in the last step of (6).
Since Z
l
s are i.i.d, we can rewrite (7) as Z =
X
M

+Y
, where
X
2
(2) and Y
2
(2M 2). Then, the PDF and the
CDF of Z
l
can be given by [30]
f
Z
l
(t) =
e
Mt/
l
(1 +t)
M
_
M

l
(1 +t) +M 1
_
, t 0 (8)
F
Z
l
(t) = 1
e
Mt/
l
(1 +t)
M1
, t 0. (9)
Let
U = max
l{1, ,NK}
Z
l
(10)
P = max
kI
m
max
mS
t
,nS
r
SINR
(k)
m,n
. (11)
1460 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 25, NO. 7, SEPTEMBER 2007
5 10 15 20 25
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
P
D
F
f
P
, simulation
f
U
, numerical
K=16
K=32
K=64
Fig. 1. PDF comparison between P and U for K = 16, 32 and 64 at 10
dB SNR for M = 2 and N = 2. Dashed curves denote the simulation results
of the PDF of P. Solid curves denote the numerical results of the PDF of U.
Then, the CDF of U can be easily given by [36]
F
U
(t) =
NK

l=1
F
Z
l
(t).
The PDF of U is given by
f
U
(t) =
d
dt
F
U
(t) =
NK

l=1
_
_
f
Z
l
(t)

j=1, ,NK;j=l
F
Z
j
(t)
_
_
. (12)
Using (5) and (6), the system throughput averaged over
Rayleigh fading channels is given by
E(R) =
M

m=1
E log
2
_
1 +SINR
(k

m
)
m,Rx
k

m
_
= M
_

0
log
2
(1 +t)f
P
(t)dt
M
_

0
log
2
(1 +t)f
U
(t)dt. (13)
In heterogeneous networks, the achievable throughput of the
proposed scheduling algorithm can be obtained by substituting
(12) into (13). Consequently, we get
E(R) M
NK

l=1
_

0
log
2
(1 +t)f
Z
l
(t)

j=1, ,NK;j=l
F
Z
j
(t)dt (14)
where f
Z
l
(t) and F
Z
j
(t) are given by (8) and (9), respectively.
In homogeneous networks, all users have an identical .
Then, (12) reduces to
f
U
(t) = NKf
Z
(t)(F
Z
(t))
NK1
. (15)
Hence, the achievable throughput of the MIMO downlink
scheduling averaged over fading channels is
E(R) KMN
_

0
log
2
(1 +t)f
Z
(t)(F
Z
(t))
NK1
dt. (16)
The numerical approximation of the throughput given in
(16) results from the approximation P U in (6). Since the
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.5
1
1.5
2
2.5
3
P
D
F
f
P
, simulation
f
U
, numerical
K=32
K=64
K=128
Fig. 2. PDF comparison between P and U for K = 32, 64 and 128 at 0
dB SNR for M = 4 and N = 2. Dashed curves denote the simulation results
of the PDF of P. Solid curves denote the numerical results of the PDF of U.
0 20 40 60 80 100 120 140
2
3
4
5
6
7
8
9
Number of users, K
T
o
t
a
l

T
h
r
o
u
g
h
p
u
t

(
b
p
s
)
Numerical
Simulation
10 dB
5 dB
0 dB
Fig. 3. Total throughput of the full feedback multiuser scheduling vs. the
number of users at various SNR, 0, 5 and 10 dB for M = 2 and N = 2.
PDF of U is used to approximate the PDF of P in calculating
the throughput in (13), we should evaluate the accuracy of the
approximation.
In Fig. 1, we compare the PDF f
U
(t) and f
P
(t) at 10
dB SNR with M = 2 and N = 2 for 16, 32 and 64
users, respectively. The PDF f
U
(t) is obtained numerically
from (15), whereas the PDF f
P
(t) is plotted by simulating P
with 100, 000 random runs according to (11). It can be seen
from Fig. 1 that the PDF f
U
(t) matches the PDF f
P
(t) very
well. Thus, the approximation in (13) is very tight. Similar
observations can be obtained from Fig. 2 which shows the
PDF f
U
(t) and f
P
(t) at 0 dB SNR with M = 4 and N = 2
for 32, 64 and 128 users, respectively.
We further perform the simulation of the MIMO downlink
scheduling algorithm with M = 2, N = 2 for 0, 5 and 10 dB
SNR, respectively. The simulation results of the achievable
throughput are compared with the numerical results and are
shown in Fig. 3. Such gure demonstrates that the numerical
results obtained from (16) are accurate enough to characterize
ZHANG and LETAIEF: MIMO BROADCAST SCHEDULING WITH LIMITED FEEDBACK 1461
the simulation results for all examined SNR values and K.
C. Discussion
In [30], three signal reception approaches were proposed
for the case of multiple antennas at each receiver. It was then
shown to achieve the largest throughput, each receive antenna
should be treated as an individual user. By doing so, the K-
user MN system is converted to (NK)-user M1 system
[30]. Note in the case of single antenna for each user, our
approach is the same as the one in [30] because no receive
antenna selection is needed. Thus, (16) can also be used to
evaluate the performance of the approach in [30] for (NK)-
user M 1 system. Then, we get
E(R) NKM
_

0
log
2
(1 +t)f
Z
(t)(F
Z
(t))
NK1
dt. (17)
It should be mentioned that (16) is the achievable throughput
of our approach using receive antenna selection at each user,
whereas (17) is the achievable throughput of the approach
in [30] which treats each receive antenna as an individual
user. After observing that the right hand side (RHS) of (16)
is identical to the RHS of (17), we can see that our approach
does not induce any performance loss compared with the best
approach proposed in [30].
Next, we will show that our algorithm is in fact equivalent
to the algorithm of [30] in terms of performance though they
are different in terms of receiver design. As shown in (5), for
the signal transmitted from the mth antenna, the best user
k

is selected with the largest SINR

(k

m
)
m,Rx
k

m
. Furthermore,
from (6) we see that SINR
(k

m
)
m,Rx
k

m
can be approximated by
max
l{1, ,NK}
Z
l
, i.e., P U where U and P are given
in (10) and (11), respectively. The close approximation of the
PDFs of U and P along with the resultant throughput has
been validated. Thus, we can conclude that our scheduling
algorithm with antenna selection is equivalent to the algorithm
of [30] that chooses the user with the largest SINR from all
NK users each with single antenna.
In heterogeneous networks, the users usually have various
SNRs due to their different distances from the base station.
A scheduling algorithm that only takes advantage of the best
channel gain will then induce the unfairness of rate allocation
in heterogeneous networks, even though it can maximize the
system throughput. This is because the user located far from
the BS will be seldom scheduled and the user near to the BS
will dominate the resources. Therefore, the fairness issue is
a major concern for the design of practical multiuser MIMO
systems. Next, we would like to evaluate the fairness of our
proposed scheduling algorithm based on the SINR feedback
with antenna selection.
Assume that user k

is the weakest user with the smallest

SNR, i.e.,
k
=
min
= min
k{1,2, ,K}

k
. Let P

min
,m,Rx
k

be the probability of the event that antenna Rx

k
of the
weakest user k

is scheduled for the signal transmitted from

the mth transmit antenna, which is given by
P

min
,m,Rx
k

= ProbSINR
(k

)
m,Rx
k

=
max
kI
m
max
mS
t
,nS
r
SINR
(k)
m,n
(18)
Using (6), we further obtain
P

min
,m,Rx
k

ProbZ
j
> Z
1
, , Z
j1
, Z
j+1
, , Z
NK
,
where Z
j
is the SINR of the virtual user j in the NK-user
M 1 system corresponding to the antenna Rx
k
of the
weakest user k

in the K-user MN system after treating all

NK receive antennas as individual users. Z
l
is given by (7)
and its PDF is given by (8). Next, we recall the main result
of [30, Theorem 6] and get
P

min
,m,Rx
k

1
NK
M 1 +
M

min
M 1 +
M

max

M

min

M

max

e
2 log(NK)
M1
1

+O(1/(NK)
2
),
where
max
= max
k{1,2, ,K}

k
. When
log(NK)
M
becomes
very small, P

min
,m,Rx
k

1
NK
. This implies that the antenna
Rx
k
of the weakest user has a probability of
1
NK
to be
scheduled by the signal transmitted from the mth antenna.
Note that SINR
(k

)
m,n
for m = 1, , M and n = 1, , N
have identical distributions, P

min
,m,n
, which do not depend
on the indices of the transmit antenna m and receive antenna
n. Therefore, each antenna of the weakest user has the same
probability to access the system resources. In other words,
the weakest user has a probability of
1
K
to be scheduled by
the proposed scheduling algorithm with antenna selection. It
achieves the same fairness as that of [30].
Finally, we should stress that our scheduling with antenna
selection does not induce any loss of throughput nor fairness
compared with [30]. However, from the point of view of
practical implementation, receive antenna selection for MIMO
downlink scheduling proposed in this paper is much more
favorable .
IV. MIMO BROADCAST SCHEDULING WITH SELECTED
FEEDBACK
In this section, a scheduling algorithm with selected feed-
back of SINR is proposed for MIMO broadcast channels. The
achievable throughput and the feedback load are also analyzed.
A. Scheduling Algorithm
The proposed scheduling algorithm with the selected feed-
back of SINR can be described as a two-stage algorithm:
selected feedback by users and scheduling by the BS.
Stage 1: Selected Feedback
1) Initialization: Set B
k
= 0, Tx
k
= 0 and Rx
k
= 0
for k = 1, , K.
2) For k = 1 to K,
B
k
= max
mS
t
,nS
r
SINR
(k)
m,n
,
Tx
k
= arg
m
max
mS
t
,nS
r
SINR
(k)
m,n
,
Rx
k
= arg
n
max
mS
t
,nS
r
SINR
(k)
m,n
,
If B
k
,
user k is selected to feed back to the BS Tx
k
and B
k
.
Else,
user k does not feed back anything.
1462 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 25, NO. 7, SEPTEMBER 2007
End If
End For
Stage 2: Scheduling
1) Initialization: Set k

m
= 0 and 1
m
= , for
m = 1, , M. Get the feedback of B
k
and the
corresponding Tx
k
from the selected user k which
has B
k
.
2) For every selected user k,
m = Tx
k
,
1
m
= 1
m
k.
End For
3) For m = 1 to M,
If 1
m
,=
k

m
= arg max
kI
m
B
k
.
End If
End For
The principal idea behind the selected feedback scheduling
is that only those users whose maximum SINR are larger than
a predesigned threshold can be selected to feed back the
SINR to the BS. The advantage of this algorithm is that a low
feedback is obtained due to the user selection.
B. Throughput Analysis
Note that 1
m
is the set of selected users preferring to the
mth transmit antenna. Let o
m
denote the set of all users hav-
ing Tx
k
= m, i.e., o
m
= k[Tx
k
= m, k 1, , K .
Obviously, 1
m
o
m
because 1
m
= k[B
k
, k
o
m
. The system throughput averaged over Rayleigh fading
channels is then given by
E(R) =
M

m=1
E log
2
_
1 +SINR
(k

m
)
m,Rx
k

m
_
, (19)
where
SINR
(k

m
)
m,R
k

m
= max
kI
m
B
k
= max
kS
m
B
k
max
l{1, ,NK}
Z
l
. (20)
By substituting (20) into (19) and using (10), we can obtain
E(R) M
_

log
2
(1 +t)f
U
(t)dt. (21)
After substituting (12) into (21), we can get the achievable
throughput in heterogeneous networks as
E(R) M
NK

l=1
_

log
2
(1 +t)f
Z
l
(t)

j=1, ,NK;j=l
F
Z
j
(t)dt
where f
Z
l
(t) and F
Z
j
(t) are given by (8) and (9), respectively.
In homogeneous networks, we can substitute (15) into (21) and
obtain
E(R) KMN
_

log
2
(1 +t)f
Z
(t)(F
Z
(t))
NK1
dt. (22)
It can be observed that when 0, (22) is equivalent to
(16).
To investigate the accuracy of the approximation in (22),
we perform simulation of the selected feedback scheduling
0 20 40 60 80 100 120 140
1
2
3
4
5
6
7
8
9
Number of users, K
T
o
t
a
l

T
h
r
o
u
g
h
p
u
t

i
n

b
p
s
/
H
z
Numerical
Simulation
= 4 dB
= 8 dB
= 10 dB
Fig. 4. Total throughput of the selected feedback multiuser scheduling vs.
the number of users for various threshold = 4, 8 and 10 dB. M = 2 and
N = 2. SNR is 10 dB.
with M = N = 2 at 10 dB SNR for = 4, 8 and 10
dB, respectively. The simulation results are compared with the
numerical results achieved by (22) and are shown in Fig. 4. We
can see that the numerical results match the simulation results
with negligible difference. Furthermore, with the increase of
the threshold, the throughput is decreased. This is reasonable
because less users are allowed to participate in the feedback
when the threshold is increased.
C. Feedback Load Analysis
Assume that l users are selected for feedback in one time
slot. This indicates that there are l users satisfying B
k

. Since B
k
= max
mS
t
,nS
r
SINR
(k)
m,n
k are i.i.d., we
can denote F
B
(t) the CDF of B
k
. Then, the probability of
selecting l users in the current time slot is
T
l
=
_
K
l
_
(1 F
B
())
l
(F
B
())
Kl
(23)
With the notation of Z
l
in (7), we can obtain
B
k
max
l={1, ,MN}
Z
l
(24)
and
F
B
() [F
Z
()]
MN
. (25)
The average feedback load of the selected scheduling is given
by
L =
K

l=1
l T
l
. (26)
To compare with the full feedback scheduling, we dene the
average feedback load ratio (FLR) as

L
K
. (27)
Using (23)-(26), we can rewrite (27) as

K

l=1
l
K
_
K
l
_
_
1 [F
Z
()]
MN
_
l
_
[F
Z
()]
MN
_
Kl
, (28)
ZHANG and LETAIEF: MIMO BROADCAST SCHEDULING WITH LIMITED FEEDBACK 1463
0 20 40 60 80 100 120 140
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Number of users, K
F
e
e
d
b
a
c
k

L
o
a
d

R
a
t
i
o
Numerical
Simulation
= 4 dB
= 8 dB
= 10 dB
Fig. 5. Feedback load ratio of the selected feedback multiuser scheduling
vs. the number of users for various threshold = 4, 8 and 10 dB. M = 2
and N = 2. SNR is 10 dB.
where F
Z
() is given by (9).
Note that using
l
K
_
K
l
_
=
_
K 1
l 1
_
, (29)
we can rewrite (28) as

K

l=1
_
K 1
l 1
_
_
1 [F
Z
()]
MN
_
l
_
[F
Z
()]
MN
_
Kl
=
K1

=0
_
K 1
l

_
_
1 [F
Z
()]
MN
_
l

_
[F
Z
()]
MN
_
(K1)l

_
1 [F
Z
()]
MN
_
= 1 [F
Z
()]
MN
. (30)
From (30), we can see that the FLR is not dependent on the
number of users K.
Fig. 5 shows the comparison in the average FLR for =
4, 8 and 10 dB at 10 dB SNR, respectively. Those results
demonstrate that the numerical approximation given by (30)
is very accurate. Furthermore, it can be observed that the
FLR is not dependent on the number of users. Also, when
the threshold is increased, FLR decreases.
D. Throughput-FLR Tradeoff
From (22) and (30), it can be seen that the throughput
and FLR both depend on the threshold and decrease when
increases. Therefore, a tradeoff exists between increasing
the throughput and decreasing FLR. In practice, two schemes
may be used: throughput-oriented and FLR-oriented. The
former scheme is to minimize FLR while guaranteeing a target
throughput. The latter scheme is to maximize the throughput
while attaining a target FLR. Since the numerical solutions
(22) and (30) closely approximate the throughput and FLR,
respectively, they can be efciently employed to investigate
the throughput-FLR tradeoff performance.
Fig. 6 shows the throughput-threshold curve using (22)
for the throughput-oriented scheme. On the other hand, Fig.
5 0 5 10 15 20
0
1
2
3
4
5
6
7
8
Threshold (dB)
T
o
t
a
l

t
h
r
o
u
g
h
p
u
t

(
b
p
s
)
K=64
K=32
K=8
Fig. 6. Total throughput of the selected feedback multiuser scheduling vs.
threshold (dB) for various number of users K = 8, 32 and 64. M = 2
and N = 2. SNR is 10 dB.
10 5 0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold (dB)
A
v
e
r
a
g
e

F
e
e
d
b
a
c
k

L
o
a
d

R
a
t
i
o
SNR = 10 dB
SNR = 5 dB
SNR = 0 dB
Fig. 7. Feedback load ratio of the selected feedback multiuser scheduling vs.
threshold (dB) at various SNR = 0, 5 and 10 dB. M = 2 and N = 2.
7 shows the FLR-threshold curve using (30) for the FLR-
oriented scheme. Given a target throughput, we could nd the
maximum threshold from Fig. 6 and then use the threshold to
get the corresponding FLR from Fig. 7 which is the minimum
FLR to get the target throughput. For instance, in the case
of M = N = 2 with K = 64 at 10 dB SNR, the target
throughput is 6.3 bps. Firstly, we can read from Fig. 6 that
the threshold corresponding to the target throughput is 10
dB. Then, from Fig. 7, we get that the corresponding FLR
is 0.05. Given a target FLR, say 0.4 for M = N = 2 at 10
dB SNR, we can observe from Fig. 7 that the corresponding
threshold is about 5 dB which is not dependent on K.
From Fig. 6, we can also see that the throughput is 7.7 bps
which is also the maximum achievable throughput. Therefore,
we can conclude that FLR can be greatly reduced without
sacricing the throughput. Finally, in Fig. 8, we show the
system throughput with respect to the required FLR for various
number of users. It shows that the FLR can be reduced to 0.2
without loss of system throughput for more than 32 users when
1464 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 25, NO. 7, SEPTEMBER 2007
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
1
2
3
4
5
6
7
8
Average Feedback Load Ratio
T
o
t
a
l

t
h
r
o
u
g
h
p
u
t

(
b
p
s
)
K=64
K=32
K=8
Fig. 8. Tradeoff between throughput and FLR of the selected feedback
multiuser scheduling for various number of users at SNR = 10 dB. M = 2
and N = 2.
M = N = 2 at 10 dB SNR.
V. MIMO BROADCAST SCHEDULING WITH QUANTIZED
FEEDBACK
In this section, a scheduling algorithm with quantized
feedback of SINR is proposed for MIMO broadcast channels.
The achievable throughput is also derived and in particular
scheduling with 1-bit feedback is investigated.
A. Scheduling Algorithm
The proposed scheduling algorithm with the quantized
feedback of SINR can be described as a two-stage algorithm:
quantized feedback by users and scheduling by the BS.
Stage 1: Quantized Feedback
1) Initialization: Set B
k
= 0, Tx
k
= 0, Rx
k
= 0 and
q
k
= 0 for k = 1, , K.
2) For k = 1 to K,
B
k
= max
mS
t
,nS
r
SINR
(k)
m,n
,
Tx
k
= arg
m
max
mS
t
,nS
r
SINR
(k)
m,n
,
Rx
k
= arg
n
max
mS
t
,nS
r
SINR
(k)
m,n
,
q
k
= Q(B
k
) with q
k
0, 1, , 2
b
1,
User k feeds back to the BS Tx
k
and q
k
.
End For
Stage 2: Scheduling
1) Initialization: Set k

m
= 0 and 1
m
,
m
= , for m =
1, , M. Get the feedback of Tx
k
and q
k
from the
user k, k = 1, 2, , K.
2) For k = 1 to K,
m = Tx
k
,
1
m
= 1
m
k.
End For
3) For m = 1 to M,
If 1
m
,=

m
=

k[

k = arg
k
max
kI
m
q
k
,
k

m
= rand
m
.
End If
End For
Compared with the full feedback scheduling where each
user feeds a real value B
k
to the BS, the quantized feedback
scheduling requires each user to send back a quantized value
Q(B
k
) which can be represented by few bits. In fact, a real
value B
k
is quantized into one of a few levels q
k
= Q(B
k
)
where
Q(B
k
) =
_
_
_
0, 0 B
k
<
1
i,
i
B
k
<
i+1
, i = 1, , L 2
L 1, B
k

L1
The number of levels L is determined by the number of bits
b required to represent a value B
k
and L = 2
b
.
With the feedback of q
k
and Tx
k
from all users, the BS will
rstly cluster the users having Tx
k
= m into the set 1
m
and
then cluster the users achieving the maximum of q
k
for all k
1
m
into the set
m
, i.e.
m
=

k[

k = arg
k
max
kI
m
q
k
.
Obviously,
m
1
m
. Finally, the transmit antenna m is
allocated to user k

m
which is randomly selected from the set

m
.
B. Throughput Analysis
The instantaneous sum-rate achieved by the quantized feed-
back scheduling is given by
R =
M

m=1
log
2
_
1 +SINR
(k

m
)
m,Rx
k

m
_
=
M

m=1
log
2
(1 + randB
k
, k
m
) (31)
where randB
k
, k
m
denotes the random selection of
an element from the set B
k
, k
m
. Note that B
k
is
approximated by (24) and the CDF of B is given by (25).
Let V = randB
k
, k
m
. The average total throughput
is then given by
E(R) = M
_

0
log
2
(1 +v)f
V
(v)dv, (32)
where f
V
(v) denotes the PDF of V and will be derived next.
Let F
V
(v) denote the CDF of V and K

=
K
M
. Note
that SINR
(k)
m,n
in (3) for m = 1, 2, , M have the same
distribution, with each user having an identical chance to select
different transmit antennas, i.e., ProbTx
k
= m =
1
M
. Then,
for a large number of users we can get that the cardinality of
the set 1
m
is equal to K

for all m = 1, 2, , M. Let A

i
denote the range [
i
,
i+1
) for i = 0, 1, , L 1, where

0
= 0 and
L
= . Then, we can derive the CDF F
V
in
the following two cases:
When 0 V <
1
, B
k
of the scheduled user is in range
A
0
. In this case, B
k
for all k 1
m
are also in range A
0
.
Thus, the CDF of V is
F
V
(v) = F
B
(v)[F
B
(
1
)]
K

1
, 0 v <
1
where F
B
(v) denotes the CDF of B and is given by (25).
When
i
V <
i+1
for i = 1, , L 1, B
k
of the
scheduled user is in range A
i
. We can assume that there
are other (r 1) users with B
k
in region A
i
for any
ZHANG and LETAIEF: MIMO BROADCAST SCHEDULING WITH LIMITED FEEDBACK 1465
0 50 100 150 200 250 300
2
3
4
5
6
7
8
Number of users, K
T
o
t
a
l

T
h
r
o
u
g
h
p
u
t

(
b
p
s
)
Numerical
Simulation
SNR = 10 dB
SNR = 5 dB
SNR = 0 dB
Fig. 9. Total throughput of the quantized feedback multiuser scheduling
(1-bit quantization) vs. the number of users at various SNR, 0, 5 and 10 dB.
Quantization threshold is xed to be equal to the SNR. M = 2 and N = 2.
1 r K

. For the remaining (K

r) users, their B
k
must be smaller than
i
. Thus, the CDF of V is
F
V
(v) = [F
B
(
i
)]
K

+
K

r=1
_
K

r
_
[F
B
(v) F
B
(
i
)]
[F
B
(
i
)]
K

r
[F
B
(
i+1
) F
B
(
i
)]
r1
,
for
i
v <
i+1
.
Consequently, the PDF f
V
(v) is given by
f
V
(v) =
_

_
f
B
(v)[F
B
(
1
)]
K

1
, for 0 v <
1

r=1
_
K

r
_
[F
B
(
i
)]
K

r
f
B
(v)
[F
B
(
i+1
) F
B
(
i
)]
r1
,
for
i
v <
i+1
, i = 1, , L 1,
(33)
where F
B
(v) [F
Z
(v)]
MN
and f
B
(v)
MNf
Z
(v)F
Z
(v)
MN1
. The PDF f
Z
(v) and the CDF F
Z
(v)
are, respectively, given by (8) and (9). After substituting (33)
into (32), we can obtain the approximation of the achievable
throughput by the quantized feedback scheduling.
C. 1-bit Feedback
In the case of a quantization with 1-bit, each user feeds 1
or 0 back to the BS according to the threshold
1
, thereby,
obtaining the minimal feedback load. In this case, (33) can be
rewritten as
f
V
(v) =
_
[F
B
(
1
)]
K

1
f
B
(v), 0 v <
1
1[F
B
(
1
)]
K

1F
B
(
1
)
f
B
(v),
1
v.
(34)
Therefore, the average total throughput (32) is given by
E(R) = M[F
B
(
1
)]
K

1
_

1
0
log
2
(1 + v)f
B
(v)dv +
M
1 [F
B
(
1
)]
K

1 F
B
(
1
)
_

1
log
2
(1 +v)f
B
(v)dv,(35)
where the approximation is obtained from F
B
(v)
[F
Z
(v)]
MN
and f
B
(v) MNf
Z
(v)F
Z
(v)
MN1
with K

=
15 10 5 0 5 10 15 20
3.5
4
4.5
5
5.5
6
6.5
7
7.5
8
Quantization Threshold (dB)
T
o
t
a
l

t
h
r
o
u
g
h
p
u
t

(
b
p
s
)
K=128
K=64
K=32
K=16
K=8
Fig. 10. Total throughput of the quantized feedback multiuser scheduling vs.
threshold (1-bit quantization) for various number of users K = 8, 16, 32, 64
and 128. M = 2 and N = 2. SNR is 10 dB.
K/M. The PDF f
Z
(v) and the CDF F
Z
(v) are, respectively,
given by (8) and (9).
When K goes to innity (while K

), it can be
observed that [F
Z
(
1
)]
K

MN
0 for a xed value
1
< .
Then, the throughput of (35) is derived as
lim
K
E(R)
M
2
N
1 [F
Z
(
1
)]
MN

_

1
log
2
(1 +v)f
Z
(v)[F
Z
(v)]
MN1
dv. (36)
If the quantization threshold
1
is xed, it can be seen from
(36) that the total rate will be a constant. This implies that the
multiuser diversity is lost in this case.
To validate the accuracy of (35), the quantized feedback
scheduling is simulated for M = 2, N = 2 at SNR = 0, 5
and 10 dB, respectively. The 1-bit quantization is used and
the threshold
1
is always xed as the value equal to . Fig.
9 shows the simulation results and the numerical results of
the average total throughput achieved by the quantized feed-
back scheduling. The performance comparison between the
numerical and simulation results shows a good approximation
of (35). Moreover, when K goes to innity the throughput
approaches a xed value. It demonstrates that the multiuser
diversity is lost, thereby, corroborating the analysis of (36).
For any given M, N and some in (35), it can be seen that
the throughput is a function of
1
and K, simply denoted by
E(R) = f(K,
1
). Hence, it is not optimal to x the value
of the threshold
1
for various K to enhance the through-
put. Intuitively, an optimal
1
which maximizes throughput
should depend on the value of K. To search for the optimal
quantization threshold, we need to solve
f(K,
1
)

1
= 0. As
seen from (35), a closed form solution to
f(K,
1
)

1
= 0 is not
tractable. However, a numerical solution can be easily given.
From (35), the numerical results of the achievable throughput
with respect to the quantization threshold
1
are shown in
Fig. 10 for various number of users, K = 8, 16, 32, 64 and
128 where M = N = 2 and = 10 dB. It can be observed
that the optimal values of the thresholds are 6, 7, 8.5, 9.5 and
1466 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 25, NO. 7, SEPTEMBER 2007
0 20 40 60 80 100 120 140
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
8
Number of users, K
T
o
t
a
l

T
h
r
o
u
g
h
p
u
t

(
b
p
s
)
Adaptive threshold
Fixed threshold
SNR = 10 dB
SNR = 5 dB
Fig. 11. Throughput comparison of the quantized feedback multiuser
scheduling between using xed quantization and adaptive (optimal) quan-
tization thresholds. For the xed threshold, is chosen the same as the SNR
and irrespective of the number of users K. For the adaptive threshold, is
optimized by maximizing the throughput for differnt K. M = 2 and N = 2.
10 dB, respectively. If the threshold is xed to be 10 dB for all
K, then only the case of K = 128 can achieve the maximum
throughput. As we decrease K at
1
= 10 dB, the gap between
the achievable throughput and the maximum throughput is
increasing. Hence, we conclude that a xed threshold is not
optimal whereas the optimal threshold should be dependent
on K for given M, N and SNR .
In Fig. 11, the throughput of 1-bit quantized feedback
scheduling with the optimal quantization threshold is com-
pared with the one using the xed threshold for M = N = 2
under 5 and 10 dB SNR. The optimal threshold is obtained
by maximizing the throughput whereas the xed threshold is
chosen as the value equal to the SNR. This gure shows that
the xed threshold values achieve the maximum throughput
only when K = 32 for 5 dB SNR (
1
= 5 dB) and K = 128
for 10 dB SNR (
1
= 10 dB), respectively.
VI. FURTHER RESULTS
Fig. 12 shows the performance comparison among the three
scheduling algorithms proposed in this paper for M = N = 2
at 10 dB SNR. The threshold for selected feedback scheduling
is chosen as 5 dB which can get an FLR 0.4 irrespective
of K as shown in Section IV-D. For the 1-bit quantized
feedback scheduling, the optimal quantization threshold is
chosen from (35) adaptively for various K so as to achieve
the maximal throughput. Fig. 12 shows that the selected
feedback scheduling achieves the almost identical throughput
as the full feedback scheduling for more than 32 users while
obtaining a reduction of 60% in feedback load. The 1-bit
quantized feedback scheduling has a lower throughput, but
the multiuser diversity is obtained and suffers a loss around
0.6 bps compared with the full feedback scheduling.
VII. CONCLUSION
We have investigated several multiuser scheduling algo-
rithms with limited feedback in MIMO broadcast channels.
0 20 40 60 80 100 120 140
5
5.5
6
6.5
7
7.5
8
8.5
T
o
t
a
l

t
h
r
o
u
g
h
p
u
t

(
b
p
s
)
Full feedback
Selected feedback, = 5 dB
1bit Quantized feedback, optimal threshold
Fig. 12. Throughput comparison among the three scheduling algorithms:
full feedback, selected feedback and quantized feedback. M = N = 2 and
SNR is 10 dB.
Combined with spatial multiplexing and receive antenna se-
lection, the proposed scheduling algorithms can achieve high
multiuser diversity. It was shown that with user selection
based on SINR, the feedback load can be greatly reduced
with a negligible throughput loss. In particular, the scheduling
with quantized feedback requires a smaller feedback load by
transmitting a few bits of quantized SINR. As a special case,
1-bit quantization was shown to be very attractive as it can
achieve multiuser diversity with minimal feedback load.
The aim of this paper was to consider broadcast scheduling
with limited feedback of partial side information in MIMO
channels. Other interesting topics remain open such as the
investigation of broadcast scheduling algorithms for MIMO-
OFDM systems as well as extensions to packet-based trans-
mission where cross-layer design can be used to achieve the
optimal throughput with quality-of-service (QoS) provision-
ing.
REFERENCES
[1] E. Telatar, Capacity of multi-antenna Gaussian channels, European
Trans. Telecommun., vol. 10, no. 6, pp. 585596, Nov. 1999.
[2] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, Capacity limits
of MIMO channels, IEEE J. Selct. Areas Commun., vol. 21, pp. 684702,
June 2003.
[3] S. M. Alamouti, A simple transmit diversity technique for wireless
communication, IEEE J. Select. Areas Commun., vol. 16, pp. 1451
1458, Oct. 1998.
[4] G. J. Foschini, Layered space-time architechture for wireless communi-
cation in fading environment when using multi-element antennas, Bell
Labs Tech. J., vol. 1, pp. 4159, Nov. 1996.
[5] S. Sfar, L. Dai, and K. B. Letaief, Optimal diversity-multiplexing
tradeoff with group detection for MIMO systems, IEEE Trans. Commun.,
vol. 53, pp. 11781190, July 2005.
[6] R. Knopp and P. A. Humblet, Information capacity and power control
in single cell multiuser communications, in Proc. IEEE Int. Conf.
Commun., June 1995, pp. 331335.
[7] K. B. Letaief and Y. Zhang, Dynamic multiuser resource allocation and
adaptation for wireless systems, IEEE Wireless Commun. Mag., vol. 13,
pp. 3847, Aug. 2006.
[8] W. Ajib and D. Haccoun, An overview of scheduling algorithms in
MIMO-based fourth-generation wireless systems, IEEE Network, pp.
4348, Sep./Oct. 2005.
ZHANG and LETAIEF: MIMO BROADCAST SCHEDULING WITH LIMITED FEEDBACK 1467
[9] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, Multiple-antenna
channel hardening and its implications for rate feedback and scheduling,
IEEE Trans. Inform. Theory, vol. 50, no. 9, pp. 18931909, Sep. 2004.
[10] R. Gozali, R. M. Buehrer, and B. D. Woerner, The impact of multiuser
diversity on space-time block coding, IEEE Commun. Lett., vol. 7, no.
5, pp. 213215, May 2003.
[11] P. Viswanath, D. N. C. Tse, and R. Laroia, Opportunistic beamforming
using dumb antennas, IEEE Trans. Inform. Theory, vol. 48, no. 6, pp.
12771294, June 2002.
[12] E. G. Larsson, On the combination of spatial diversity and multiuser
diversity, IEEE Commun. Lett., vol. 8, no. 8, pp. 517519, Aug. 2004.
[13] J. Chung, C.-S. Hwang, K. Kim and Y. K. Kim, A random beamforming
techniques in MIMO systems exploiting multiuser diversity, IEEE J.
Select. Areas Commun., vol. 21, no. 5, pp. 848855, June 2003.
[14] N. Sharma and L. H. Ozarow, A study of opportunism for multiple-
antenna systems, IEEE Trans. Inform. Theory, vol. 51, no. 5, pp. 1804
1814, May 2005.
[15] N. Jindal and A. Goldsmith, Dirty-paper coding versus TDMA for
MIMO broadcast channels, IEEE Trans. Inform. Theory, vol. 51, pp.
17831794, May 2005.
[16] G. Caire and S. Shamai, On the achievable throughput of a multi-
antenna Gaussian broadcast channel, IEEE Trans. Inform. Theory, vol.
49, no. 7, pp. 16911706, July 2003.
[17] H. Weingarten, Y. Steinberg, and S. Shamai, The capacity region of
the Gaussian MIMO broadcast channel, in Proc. IEEE Int. Symp. Inf.
Theory, Chicago, IL, USA, June 27-July 2, 2004, p. 174.
[18] N. Jindal, W. Rhee, S. Vishwanath, S. A. Jafar, and A. Goldsmith,
Sum power iterative water-lling for multi-antenna Gaussian broadcast
channels, IEEE Trans. Inf. Theory, vol. 51, pp. 15701580, Apr. 2005.
[19] W. Yu, Sum-capacity computation for the Gaussian vector broadcast
channel via dual decomposition, IEEE Trans. Inform. Theory, vol. 52,
pp. 754759, Feb. 2006.
[20] Z. Tu and R. S. Blum, Multiuser diversity for a dirty paper approach,
IEEE Commun. Lett., vol. 7, pp. 370372, Aug. 2003.
[21] G. Dimic and N. D. Sidiropoulos, On downlink beamforming with
greedy user selection: performance analysis and a simple new algorithm,
IEEE Trans. Signal Processing, vol. 53, pp. 38573868, Oct. 2005.
[22] K. K. Wong, R. D. Murch, and K. B. Letaief, A joint channel
diagonalization for multi-user MIMO antenna systems, IEEE Trans.
Wireless Commun., vol. 2, pp. 773786, July 2003.
[23] L.-U. Choi and R. D. Murch, A transmit pre-processing technique for
multi-user MIMO systems using a decomposition approach, IEEE Trans.
Wireless Commun., vol. 3, pp. 2024, Jan. 2004.
[24] C. B. Peel, B. M. Hochwald, and A. L. Swindlehurst, A vector-
perturbation technique for near-capacity multiantenna multiuser commu-
nication Part I: channel inversion and regularization, IEEE Trans.
Commun., vol. 53, pp. 195202, Jan. 2005.
[25] C. Swannack, E. Uysal-Biyikoglu, and G. W. Wornell, MIMO broad-
cast scheduling with limited channel state information, in Proc. Annual
Allerton Conf. on Commun., Control and Computing, Sep. 2005.
[26] A. Bayesteh and A. Khandani, On the user selection for MIMO
broadcast channels, in Proc. IEEE Int. Symp. Inf. Theory, Adelaide,
Australia, Sept. 49, 2005, pp. 23252329.
[27] T. Yoo and A. Goldsmith, On the optimality of multiantenna broadcast
scheduling using zero-forcing beamforming, IEEE J. Select. Areas
Commun., vol. 24, no. 3, pp. 528541, Mar. 2006.
[28] D. J. Love, R. W. Heath, W. Santipach, and M. L. Honig, What is the
value of limited feedback for MIMO channels, IEEE Commun. Mag.,
pp. 5459, Oct. 2004.
[29] N. Jindal, MIMO broadcast channels with nite rate feedback, in Proc.
IEEE Global Commnun. Conf., St. Louis, MO, USA, Nov. 28-Dec. 2,
2005, vol. 3, pp. 1520-1524.
[30] M. Sharif and B. Hassibi, On the capacity of MIMO broadcast channels
with partial side information, IEEE Trans. Inform. Theory, vol. 51, no.
2, pp. 506522, Feb. 2005.
[31] D. Gesbert and M.-S. Alouini, How much feedback is multi-user
diversity really worth, in Proc. IEEE Int. Conf. Commun., June 20-24,
2004, vol. 1, pp. 234238.
[32] F. Floren, O. Edfors, and B.-A. Molin, The effect of feedback quanti-
zation on the throughput of a multiuser diversity scheme, in Proc. IEEE
Global Commun. Conf., 2003, pp. 497501.
[33] M. Johansson, Benets of multiuser diversity with limited feedback, in
Proc. IEEE Workshop on Signal Processing Adv. for Wireless Commun.,
June 2003, pp. 155-159.
[34] S. Sanayei and A. Nosratinia, Exploiting multiuser diversity with only
1-bit feedback, in Proc. IEEE Wireless Commun. and Networking Conf.,
Apr. 2005, pp. 978983.
[35] R. Zhang, J. M. Ciof, and Y.-C. Liang, Throughput comparison of
wireless downlink transmission schemes with multiple antennas, in Proc.
IEEE Int. Conf. Commun., Seoul, Korea, May 16-20, 2005, vol. 4, pp.
27002704.
[36] H. A. David, Order Statistics, 2nd ed. New York: Wiley, 1981.
Wei Zhang (M06) received his B. Eng. degree
in mechanical engineering from Jilin University of
Technology, Changchun, China, his M. Eng. degree
in communications engineering from Jilin Univer-
sity, Changchun, China, and his Ph.D. degree in
electronic engineering from The Chinese University
of Hong Kong, in 1998, 2002, and 2005, respec-
tively.
From 1998 to 1999, he was Research Assistant
with the Institute of Oceanographic Instrumentation,
Shandong Academy of Sciences, Tsingdao, China.
He was Visiting Student at the Department of Electrical and Computer Engi-
neering, University of Delaware, USA, in 2004. He is presently Postdoctoral
Fellow at the Department of Electronic and Computer Engineering, The Hong
Kong University of Science and Technology, Kowloon, Hong Kong. His
current research interests include multiuser MIMO, cooperative diversity, and
cognitive radio.
Khaled Ben Letaief (S85-M86-SM97-F03) re-
ceived the BS degree with distinction in Electri-
cal Engineering from Purdue University at West
Lafayette, Indiana, USA, in December 1984. He
received the MS and Ph.D. Degrees in Electrical En-
gineering from Purdue University, in August 1986,
and May 1990, respectively. From January 1985 and
as a Graduate Instructor in the School of Electrical
Engineering at Purdue University, he has taught
courses in communications and electronics.
From 1990 to 1993, he was a faculty member
at the University of Melbourne, Australia. Since 1993, he has been with
the Hong Kong University of Science & Technology where he is currently
Chair Professor and Head of the Electronic and Computer Engineering
Department. He is also the Director of the Hong Kong Telecom Institute
of Information Technology as well as the Director of the Center for Wireless
Information Technology. His current research interests include wireless and
mobile networks, Broadband wireless access, OFDM, CDMA, Cooperative
networks, Cognitive radio, MIMO, and Beyond 3G systems. In these areas,
he has published over 300 journal and conference papers and given invited
keynote talks as well as courses all over the world.
Dr. Letaief served as consultants for different organizations and is the
founding Editor-in-Chief of the IEEE Transactions on Wireless Communi-
cations. He has served on the editorial board of other prestigious journals
including the IEEE Journal on Selected Areas in Communications - Wireless
Series (as Editor-in-Chief). He has been involved in organizing a number
of major international conferences and events. These include serving as the
Co-Technical Program Chair of the 2004 IEEE International Conference on
Communications, Circuits and Systems, ICCCS04; General Chair of the 2007
IEEE Wireless Communications and Networking Conference, WCNC07; as
well as the Technical Program Co-Chair of the 2008 IEEE International
Conference on Communication, ICC08.
In addition to his active research and professional activities, Professor
Letaief has been a dedicated teacher committed to excellence in teaching
and scholarship. He received the Mangoon Teaching Award from Purdue
University in 1990; the Teaching Excellence Appreciation Award by the
School of Engineering at HKUST (4 times); and the Michael G. Gale Medal
for Distinguished Teaching (Highest university-wide teaching award and only
one recipient/year is honored for his/her contributions).
He is a Fellow of IEEE, an elected member of the IEEE Communications
Society Board of Governors, and an IEEE Distinguished lecturer of the
IEEE Communications Society. He also served as the Chair of the IEEE
Communications Society Technical Committee on Personal Communications.