You are on page 1of 9

Computer Communications 33 (2010) 21162124

Contents lists available at ScienceDirect

Computer Communications
journal homepage: www.elsevier.com/locate/comcom

Fairness maximization and delay analysis for multi-antenna broadcast channels


using random beamforming
Soroush Akhlaghi *, Alireza Borhani
Faculty of Engineering, Shahed University, Tehran, Iran

a r t i c l e

i n f o

Article history:
Available online 29 July 2010
Keywords:
MIMO broadcast channel
Random beamforming
Power allocation
Delay analysis
Minimum rate constraint

a b s t r a c t
A Multi-Input Multi-Output (MIMO) broadcast channel with large number (n) of users is considered. It is
assumed each user either receives the minimum rate constraint of Rmin or remains silent. Accordingly, for
the case of random beamforming, an user selection strategy together with a proper power allocation
 h1 in the
method is proposed, showing the maximum number of active users scales as M loglogn
Rmin
asymptotic case of n ? 1, where M represents the number of transmit antennas. Noting the asymptotic
sum-rate capacity of such channel is M loglogn, the proposed method is able to approach the asymptotic sum-rate capacity within a constant gap. Moreover, it is shown the expected delay of this fair power



n logn
n
min
x loglogn
, where the expected delay is dened as
allocation strategy behaves like RM
loglognh1
the minimum number of channel uses to make sure each user receives at least one packet. Accordingly,
it is proved that for sufciently large (k) number of channel uses, the average number of services received

q
logk
.
1O
by a randomly selected user scales as kM loglogn
nRmin
k
2010 Elsevier B.V. All rights reserved.

1. Introduction
Recently, Quality of Service (QoS) is recognized as one of the
main concerns in many communication systems. QoS is mainly involved with some features like achievable rate, delay and fairness.
In some applications, it is desirable that users meet a minimum
rate constraint, while having an acceptable delay. On the other
hand, it may happen some users due to their channel conditions
fail to meet the rate constraint, hence, they are required to be silent
during the current transmission. This in turn increases the system
delay as these non-active users impose a queueing delay to the system. Thus, one needs to seek for a proper strategy that aims at
maximizing the number of active users (fairness maximization)
which meet the minimum rate constraint, thereby decreasing the
queueing delay.
This paper addresses the same issue in a Multi-Input MultiOutput (MIMO) Broadcast Channel (BC), where there is a multiantenna transmitter aims at sending individual information to
many single antenna users. Noting the transmitter has a limited
transmit power, thus it is desirable to explore a proper user selection strategy followed by a suitable power allocation method
among selected users to effectively maximize the number of active

* Corresponding author. Tel.: +98 9192696264.


E-mail addresses: akhlaghi@shahed.ac.ir (S. Akhlaghi), aborhani@shahed.ac.ir (A.
Borhani).
0140-3664/$ - see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.comcom.2010.07.014

users. In what follows, we briey discuss recent advances in MIMO


BCs and then justify the idea behind the current work.
The sum-rate capacity of MIMO BC is rst addressed in [13].
This is achieved through using an elegant coding strategy, called
Dirty Paper Coding (DPC) [4]. However, DPC is less likely to nd practical implications as it is too complex to deal with. For example, in
DPC approach, the transmitter needs to have an access to the non
causal information of channel gains. This motivated researchers to
seek an alternative strategy using random beamforming that can
effectively approach the capacity of such channel in a Rayleigh fading environment with lower complexity [58]. Interestingly, it is
demonstrated that random beamforming is able to approach the
sum-rate capacity of such channel, for the case of many users [5].
In random beamforming strategy, considering the transmitter is
equipped with M antennas, M data streams are sent along M random directions, such that each data stream corresponds to the user
which yields the maximum Signal to Interference-plus-Noise Ratio
(SINR) along this direction. Then, an user selection strategy is proposed to effectively select the best user for each direction. In this
method, a partial knowledge of channel gains at the transmitter
is adequate to achieve the sum-rate capacity of such channel.
Accordingly, it is shown the asymptotic sum-rate capacity of such
channel for large number (n) of users scales as M loglogn.
Although, the random beamforming strategy is able to approach
the sum-rate capacity of MIMO BC, a few number of users are selected simultaneously (at most M active users), and a great deal
of information should be sent to each selected user. Thus, this

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

method dramatically increases the waiting list and renders an excess delay to happen. Noting above, this method is practically
infeasible, specically for the case of delay limited applications.
However, in most of practical applications, it is desirable that
each user is supported by a constant rate. Motivated by this, in
[9] for the case of single antenna BC in a Rayleigh fading environment and assuming the transmitter makes use of a multi-layer
superposition code for having concurrent transmissions to many
users, a power allocation strategy across layers is deduced. Accordingly, it is shown this method is able to support loglogn
users at the
Rmin
same time, where n and Rmin are, respectively, the number of users
and the minimum rate constraint. Noting the sum-rate capacity of
such channel is loglogn, this method is able to achieve the maximum number of active users in single antenna broadcast channel.
From now on, we refer to this strategy as Fair Power Allocation
Strategy (FPAS). Although FPAS can be effectively applied to single
antenna Rayleigh fading BC, we will later argue that this method
can not be directly applied to the case of multiple transmit antennas. The main contribution of this work is to address the same issue in multi-antenna BC.
This paper concerns fairness maximization in MIMO BC in a
Rayleigh fading environment by the use of M random beamforming
vectors, assuming the transmitter employs M multi-layer superposition codes, each along one direction. This is achieved through
selecting a set of appropriate users along each direction for which
a proper power allocation strategy similar to what is proposed in
[9] is applied to each superposition code, separately.
It is worth mentioning that the FPAS can not be directly adopted
across the layers of the aforementioned superposition codes. This is
due to the inherent inter-beam interference term arising from other
M  1 directions. However, it is argued that by using a proper user
selection strategy which aims at reducing the inter-beam interference term, one can readily decompose the MIMO BC channel into
M parallel inter-beam interference-free channels. Although, the corresponding channel gains are no longer Rayleigh distributed, it is
demonstrated that the FPAS is again applicable in such channel.
Finally, it is shown the proposed approach is able to achieve the
sum-rate capacity of such channel within a constant gap.
The potential merit of the proposed method is that it alludes to
some implications which may interest both researchers and practitioners. As a practical implication, one can consider a wireless
service provider who aims at giving service to many clients with
a desired QoS to guarantee a minimum rate constraint. The provider needs to employ an effective broadcast strategy to support
as many as users, each having a desired rate, while it is desirable
the provider does not fall short of approaching the sum-rate capacity. Previously, it was widely assumed that in order the transmitter
does not sacrice the sum-rate capacity, it should simultaneously
support at most M users (M is the number of transmit antennas).
For instance, IEEE802.11n and IEEE802.16-2004 standards support
at most four antennas at the transmitter, and similarly their signaling support at most four data streams. However, we will later show
that it is possible to simultaneously support large number of users
at the expense of a modest decrease in the sum-rate capacity. As is
stated earlier, this beats the previous belief of supporting at most
M users (M data streams).
The second part of this work is devoted to delay analysis of the
proposed strategy. As is noted earlier, in some applications, the
users are required to get a service with a constant rate and for a certain time interval. In this regard, [10] proposes an effective way to
support QoS in wireless channels based on maximizing the number
of users with a view to approaching the desired QoS. Some other
works are devoted to the case of delay limited systems [11,12].
The trafc rate of wireless networks under delay constraint with
large number of nodes is presented in [12]. Also, there have been
some attempts to make a balance between the system delay and

2117

the achievable throughput in wireless networks [13,14]. The problem of throughput-delay trade off in multicast channels for different
scheduling algorithms is also presented in [15].
This paper aims at investigating the expected delay of a Rayleigh fading broadcast channel when the transmitter makes use
of multiple antennas. In the second part of paper, we rst introduce
and then discuss what is regarded as the expected delay for the
case of single transmit antenna and similarly extend the terminology to multiple antennas. In this work, it is assumed the channel
gain of each user is constant across a coding block (along one transmission interval) and independently varies across different blocks
(quasi-static channel). Also, it is assumed each active user receives
one packet of information during each transmission interval. The
system delay is dened as the minimum number of channel uses
(transmission intervals) which guarantees each user successfully
receives at least one packet.
In [16], it is shown the optimum power allocation strategy for the
case of single antenna BC in terms of maximizing the sum-rate is to
choose the best user, meaning allocating the whole power to the best
user. From now on, we call this strategy as the Best User Selection
Strategy (BUSS). Although, the BUSS is able to approach the sum-rate
capacity of single antenna BC channel, it renders an excess delay to
happen [17], as only one user is being serviced at a given time.
Accordingly, it is proved that in a homogenous single antenna BC,
the expected delay of the BUSS, scales as D n logn [17], where n
represents the total number of users. However, this may not be a
wise strategy, specically, for the case of large users. As is noted earlier, FPAS is able to simultaneously achieve the asymptotic sum-rate
capacity1 and approach the maximum number of active users [9].
However, the expected delay of this method yet to be studied.
Noting our proposed method makes use of FPAS for each direction, this motivated us to rst derive the expected delay of the FPAS
for the case of single antenna channel and then extend it to multiantenna model. It is worth mentioning that for single antenna
broadcast channel, the expected delay lies within two extreme
points, the result of round-robin scheduling with Dmin = n, which
serves as a lower bound and Dmin n logn as the upper bound
which refers to the case that merely one user is being serviced at
a given time [17]. Thus, as is expected, the expected delay of FPAS
resides between the aforementioned bounds.
The remainder of this paper is organized as follows, Section 2
describes the problem formulation. Our proposed algorithm and
main ndings are presented in Section 3. Section 4 computes the
expected delay of the proposed method. The simulation results
are presented in (5) and nally Section 6 summarizes ndings.
Throughout the paper, vectors and scalers are bold face lower
case and simple lower case, respectively. Also the Hermitian operation is denoted by ()H. In this paper, according to the Kunths notation [18], for any function f(n) and g(n), f(n) = O(g(n)), f(n) = o(g(n)),
 
 f n 
f(n) = x(g(n)) and f(n) = h(g(n)) are equivalent to limn!1 gn
 < 1;
 
 
 
 f n 
 f n 
 f n 
limn!1 gn 0; limn!1 gn 1 and limn!1 gn c, respectively,
where c is a constant value.
2. Problem formulation
A MIMO broadcast channel with M transmit antennas and n single antenna users is considered. Assuming the transmitted information vector at time instant t is s(t), the received signal at the
ith receiver can be represented as follows
T

yi t hi st wi t;

i 1; 2; . . . ; n;

where hi is the M  1 channel gain vector of the ith user composed


of M independent and identically distributed (i.i.d.) elements which
1

The optimality of this scheme is limited to the case of large users.

2118

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

are drawn
from a circularly symmetric gaussian distribution, i.e.,
CN 0; M1 , and are assumed to be perfectly available at the afliated
receiver. Moreover, it is assumed the channel gains are constant
across the coding block, and varies for each block. wi(t) represents
a scaler additive white gaussian noise with variance r2, i.e.,
wi t  CN 0; r2 . Furthermore, it is assumed the average transmit
power per antenna is unit, i.e., E{sH(t)s(t)} = M.
It is assumed the transmitter makes use of random beamforming by employing a random unitary matrix, whose columns, U(l)
for l = 1, . . . , M, correspond to M beamforming vectors [19]. As a result, the transmitted vector can be represented as follows

st

M
X

Ul sl t;

l1

where s(l)(t) denotes the transmitted signal to be sent along the lth
direction, U(l), and entails the transmitted signals of users which
are assigned to the lth direction, thus can be expressed as follows

sl t

ml
X
p
ppl k xpl k t for l 1; . . . ; M;
k1

P fpl 1; pl 2; . . . ; pl ml g;

where pl(k) for k = 1, . . . , ml denotes the index of kth user in the set
P(l), the set of assigned users to the lth direction, which is assumed
is of size ml. Also, ppl k and xpl k t denote, respectively, the allocated power and the transmitted signal of the kth user in the set
P(l). As a result, by substituting (2) into (1), the received signal of
the ith user in the jth set P(j), can be extracted as
T

ypj i t hpj i

M
X

l l

U s t wpj i t;

l1

or equivalently, after some manipulations and considering (3), it


follows
T

ypj i t hpj i

ml
M X
X

Ul

p
ppl k xpl k t wpj i t

l1 k1

q T
ppj i hpj i Uj xpj i t Ipj i;1 t Ipj i;2 t
wpj i t;

where
T

Ipj i;1 t hpj i

mj
X

Uj

q
ppj k xpj k t;

k1;ki
T

Ipj i;2 t hpj i

M
X

Ul sl t:

l1;lj

Note that, the terms (6) and (7) can be treated as the intra-beam
interference and inter-beam interference terms, respectively.
Assuming the corresponding information for the ith user is sent
through the jth direction, thus the received SINR of the ith user
becomes

SINRpj i


2
 T

ppj i hpj i Uj 
n
o
n
o:

r2 E jIpj i;1 tj2 E jIpj i;2 tj2

Moreover, assuming the transmit power of each beamforming


direction is one (the total transmit power of M directions is M)
n
o
n
2 o Pmj
and noting E jxpj i tj2 1, it follows E sl t i1
ppj i 1,
where mj denotes the total number of users that receive their information from the jth direction.
3. Proposed method
Generally, there are two main issues that should be addressed
when trying to maximize the number of active users in a MIMO

broadcast channel: (i) the beamforming strategy and (ii) the power
allocation method. The former basically concerns to assign one of
the existing beamforming vectors out of M available vectors to
each user which is thought to provide a better channel strength,
thereby reducing the interference. The later, however, concerns
to devise a proper power allocation strategy among users with a
view to increasing the number of active users on each direction.
In the following, the aforementioned issues are discussed and addressed in details.
3.1. Beamforming strategy
In [5], it is proved that for the case of large users, random beamforming is able to achieve the asymptotic capacity of multi-antenna broadcast channels. In this method, M orthogonal random
beams are sent through the channel and each user feeds back the
index of beamforming vector for which the maximum SINR value
is achieved. In other words, the transmitter sends the corresponding information
for

2 the ith user through the jth direction as long as
 T

the term hpl i Ul  is maximized for l = j. This is achieved through
setting up an initialization phase, where in the transmitter broadcasts U(l) for l = 1, . . . , M beamforming vectors. Then, each user individually computes the received SINR for each direction, and feeds
back the corresponding index of the direction which yields the
maximum gain. Finally, the transmitter selects the best user that
its SINR value over this direction outperforms that of the others.
Although the beamforming strategy in the current study is relatively similar to [5], however, each direction is assigned to fairly
large number of users. This is completely different to the method
proposed in [5] which merely one user is being serviced at a given
time.
3.2. Power allocation method
The power allocation strategy, on the other hand, affects the
number of active users on each direction. As noted earlier, referring
to (5), the interference term can be divided into two main parts: (i)
the inter-beam interference and (ii) the intra-beam interference
terms. The rst term is hard to deal with. However, we will later
show that for the case of large users, one can simply discard those
users that their corresponding inter-beam interference term exceeds a certain threshold at the expense of a marginal degradation
in the network throughput. The intra-beam interference term,
however, can be decreased through using the superposition code.
The power allocation method basically attempts to allocate the
available power among users such that the corresponding SINR
value exceeds a certain threshold, thereby achieving a minimum
rate constraint (Rmin). Thus, the interference term plays an important role for power allocation. In what follows, we divide the power
allocation method into two main parts: (i) The user selection algorithm (Section 3.3) which affects the inter-beam interference term,
and (ii) the fair power allocation strategy which basically concerns
optimal power allocation to maximize the number of active users
(Section 3.4).
3.3. User selection algorithm
In this section, we rst formalize and then state the condition
for which the inter-beam interference term can be cancel out for
certain number of users.
~ denotes the number of users selected by the
Theorem 1. Let n
aforementioned beamforming strategy for a given direction, then there

q
 f M
~
logn
^n
~ M1
are at least n
1O
users for which the
~
n
corresponding inter-beam interference power is less than f.

2119

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

Proof. The proof is given in the Appendix A.

3.4. Fair power allocation strategy

~ , the numAlso, the following lemma discuses on the value of n


ber of selected users for transmitting along a specic direction,
according to the beamforming strategy discussed in Section 3.1.
Lemma 1. For sufciently large number (n) of users, the size of each

2
 T

subset, the total number of users for which the value of zl hpl i Ul 
~j
is maximized
along
a specic direction, say l = j, scales as n
q

logn
n
with probability approaching one.
M 1O
n
Proof. It can be readily veried that the probability that a specic
direction out of M directions is assigned to an user is M1 . Thus, the
total number of users for which the jth direction is assigned to


~ j  B n; M1 . Thus, after
them makes a Binomial distribution, i.e., n
some manipulations and following the same approach as is used
in the proof of Theorem 1, we get the result.
^ j , the
Noting Theorem 1 and Lemma 1, one can verify that n
minimum number of users for which the jth direction is assigned
to them and their corresponding inter-beam interference powers
~ j P, where P
fall below a certain threshold, say f, is equal to n
denotes the probability that the inter-beam interference power of
a given user in the jth direction is less than f (see Appendix A).
~ j and referring to the Appendix A, it follows
Thus, noting n

^j
n


M 


n
f
logn
:
1O
M M1
n

Considering f  r2, one can easily discard the inter-beam interfer


2
P
 T
l 
ence term, M
l1;lj hpj i U  , and hence (8) can be simplied to

SINRpj i
where

cpj i

ppj i

;
Pn^j
1=cpj i k1;ki
ppj k


2
 T

hpj i Uj 

r2

10

users, that is2

^
logP logn
,
Rmin

where P is the total transmit power [9].

It is worth mentioning that in mutli-antenna BC, the corresponding channel strength of each user is no longer exponential
distributed. This is due to the fact that the beamforming strategy
proposed in Section 3.1 aims to assign the beamforming direction
out of M possible choices which yields the maximum channel
strength at the receiver. Thus, cpj i in (11) is no longer exponential
distributed, as it is the maximum of M i.i.d. random variables, each
of which has an exponential distribution with parameter r2. Thus
we have

cpj i maxcp1 i ; cp2 i ; . . . ; cpM i :

12

Thus, the probability distribution function of cpj i becomes

F cp i x F cmax x Prcmax 6 x Prcp1 i 6 x;


j

cp2 i 6 x; . . . ; cpM i 6 x Prcp1 i 6 xM 1  er x M :

13

From now on, for ease of notation, the corresponding probability


density function of cpj i is labeled as W, i.e.:

This makes the original problem can be thought as M parallel single


antenna broadcast channel with non overlapping subset of users,
^ j . For a given direction, say the jth direction, without
each of size n
loss of generality, we assume users are labeled based on their
corresponding channel strength in a descending order, i.e.,
cpj 1 P cpj 2 P    P cpj n^j . It is demonstrated that the single
antenna gaussian broadcast channel is stochastically degraded
[16]. As a result, by using superposition code at the transmitter
and applying successive interference cancelation at the receivers,
one can increase the received SINRpj i of the ith user through canceling out those interfering signals arising from users for which
their corresponding channel strength outperforms that of the ith
user in the ordered list, i.e., the interfering signals arising from users
pj(k) for k = 1, . . . , i  1. As a result, the resulting throughput under
^ j becomes
gaussian code book assumption for i 1; . . . ; n

1
ppj i
@
A:
Rpj i log 1
P
1=cpj i i1
k1 ppj k

Recently, in [9] an elegant power allocation strategy (FPAS) for


the case of single antenna BC in a Rayleigh distributed environment is devised and is shown can approach the theoretical maximum number of active users. Although the channel model in our
work is no longer Rayleigh distributed, in what follows it is shown
that the FPAS leads to the sum-rate capacity within a constant gap.
In the following, the aforementioned power allocation strategy is
briey discussed and then is argued that this strategy can be readily adopted in our model.
^ number of
A single antenna Rayleigh fading BC with large n
single antenna users is considered. The problem is to nd the maximum number of active users, such that each can receive at least
the minimum rate constraint of Rmin P 0, otherwise, if this is not
applicable, this user remains silent for the entire transmission
(its power is set to zero). Accordingly, for the case of many users,
it is shown the FPAS can approach the maximum number of active

11

As a result, the aforementioned MIMO BC can be thought as M independent single antenna BCs. This enables to incorporate single antenna power allocation strategies for MIMO BC when random
beamforming is applied at the transmitter. More specically, any
attempt to maximize the number of active users in single antenna
case maybe applied to this channel. This motivated us to seek for
proper power allocation strategies for the case of single antenna
channel that can be adopted in this work. h

W : fcmax x Mr2 er

2x


M1
2
1  er x
:

14

Accordingly, throughout the paper, the aforementioned BC is called


W-BC. In what follows, it is argued that the FPAS can be readily applied to W-BC, resulting in fairness maximization. Although there is
a discrepancy between the channel model of BC and that of W-BC,
one can verify that the power allocation strategy in both cases
can be considered the same. In what follows, we rst discuss this
power allocation strategy and then exploit it in our work.
Let p0 denotes the probability that the corresponding channel of
a given user is in good condition (the user is selected), meaning the
equivalent noise, i.e., N i c 1 , resides bellow a certain threshold, i.e.,
pj i

N0. Referring to (13), this probability can be readily computed as


M
2
r2
r
p0 1  1  e N0
 Me N0

for N0  1:

15

Assuming xi is a binary random variable representing wether the ith


user is in good condition or not

xi

1 with probability p0 ;
0

with probability 1  p0 :

Clearly, the number of users which are in good condition can be


P^
considered as a random variable, i.e., X ni1 xi , which makes a
2
It is worth mentioning that the asymptotic capacity of such channel is
^ [9].
logP logn

2120

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

^ ; p0 . Thus, referring to (15), the exBinomial distribution as X  Bn


pected number of active users (m) becomes
2

r
N

^ p0 M n
^e
mn

16

The problem is to nd the best value of N0 which guarantees each


selected user can meet the required rate constraint. Note that referring to (16), in order to maximize the number of active users (m),
one need to keep N0 as large as possible, however, noting Ni c 1
pj i
for i = 1, . . . , m should reside below N0, thus the users with lower
channel strength are also selected. On the other hand, the users
with poor channel conditions, need more transmit power to comply
with the rate constraint, which may not be feasible due to the restricted total transmit power. Thus, it is desirable to make a balance
between N0 and total transmit power. In what follows, we address
this issue.
It should be noted that if one could guarantee that the corresponding SINR of each active user exceeds a  1, then Rmin
log1 SINR loga would be achievable. Noting above, in [9]
for the case of Rayleigh channels it is proved that Rmin loga
for best users is achievable if

max16i6m Ni < N0

am

17

where P represents the total transmit power. Thus, substituting (17)


in (16), it follows

^ e
m Mn

r2 am
P

18

which in log scale, it becomes

r2 a m

P
^ log M  log m
) am 2 log n
P
r

P
^ log M  log m

log
n
) m log a log
r2


^ rP2 log M  rP2 log m
log rP2 log n
:
)m
log a

^ 
log m log M n

19

Recursively, it can be veried that



^
^
log log n
log log n
:
h
^
log a
log n

20

^ , it follows
Thus, noting Rmin log a, and for a large value of n

m

^
log log n
:
Rmin

21

Noting above, one can easily verify that the maximum number of
^j
log log n
active users in the jth decomposed W-BC becomes mj
.
R
min

Also using (9), mj is asymptotically equal to

mj


  M 


f
log log Mn M1
1 O logn
n
Rmin

for j 1; . . . ; M:

22

Thus, the total number of active users over all beamforming directions becomes
M
X

mj

Noting the maximum sum-rate of the original MIMO BC scales as


M loglogn [5], it turns out that our method achieves this
throughput within a constant gap.


  M 


f
M log log Mn M1
1 O logn
n

j1

Rmin

4.1. Delay analysis for the proposed method


Generally, in BCs the system delay is dened as the minimum
number of channel uses which guarantees each of n users successfully receives p packets. Accordingly, in [17] it is proved that in a
homogenous single antenna BC and for the case of p = 1, the expected delay of the BUSS strategy scales as n logn. Furthermore,
for multiple antenna BC with M transmit antennas the expected
n logn
delay is shown asymptotically becomes MO
[17].
M2 =n
Similarly, we dene the delay as the minimum number of channel uses for which each user successfully receives at least one packet.
In fact, for the case of single antenna BC with n users, the expected
delay lies within two well known bounds, the result of round-robin
scheduling with Dmin = n, which serves as the lower bound and
n logn as the upper bound which refers to the case that merely
one user is being serviced at a given time [17]. In what follows, we
derive the expected delay of the FPAS for single antenna BC and for
the case of p = 1, and then extend the results to the multi-antenna
BC when random beamforming is applied at the transmitter.
~ numWe consider a single transmit antenna BC with large n
ber of users. The problem is to nd the minimum number of channel uses, dubbed Dmin, for which all users are being serviced
successfully with probability approaching one (w.p.1). In this correspondence, we dene a successful service S as all users receive at
least once. To this end, we propose using a probabilistic viewpoint
to seek for conditions which leads to PrfSg ! 1. Let S 1 ; . . . ; S n~ denote, respectively, the total number of times for which the users in~ are being serviced. As is noted earlier, the successful
dexed 1 to n
event S corresponds to the event that each user is being serviced
at least once, i.e., S 1 0; . . . ; S n~ 0. The following theorem computes Dmin which guarantees the successful event to happen w.p.1.
~ of users, assuming m users are
Theorem 2. For large number n
being serviced simultaneously at a given time, the necessary and
sufcient condition to have PrfSg ! 1 is

Dmin

 
~ logn
~
~
n
n
x
:
m
m

a. Necessary condition:
Assuming D channel uses is being used, it follows

23

As a result, as the minimum rate constraint for each user is set to


Rmin, the achieved sum-rate for the proposed strategy becomes at
least

M loglogn  h1:

There are some applications for which there is a stringent delay


constraint and they can not tolerate an excess delay. One of the
main advantages of our proposed user selection algorithm is
approaching the maximum number of active users at the same
time which dramatically decreases the system delay. In what follows, rst in Section 4.1 the delay analysis for the proposed method is investigated. Then, Section 4.2 is provided to explore the
average number of services given to a randomly selected user.

Proof.

which for the case of large n, asymptotically scales as

M log logn
 h1:
Rmin

4. Delay analysis

24

PrfSg PrfS 1 0; . . . ; S n~ 0g
(~
)
~
n
n
X
[
S i 0 Pa 1 
PrfS i 0g
1  Pr
i1

i1


~ logn
~
mDn
mD c
mD
~ 1
~ e n~ d 1  e n~ ;
 1n
b 1  n
~
n

25

2121

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

where (a) comes from union bound. (b) comes from the fact that for
each channel use the probability that a given user is being serviced
is mn~ , hence, for D channel uses, the probability
of unsuccessful ser D
vice for a given user becomes 1  mn~ . (c) is based on using

n~
~ and nally (d) is based on using
1  mn~ em for large value of n
log x
x e . Thus, referring to (25), the necessary condition to have
PrfSg ! 1 is

once. However, in some applications, it is desirable to investigate


the average number of services received by a randomly selected
user after k trials. The following theorem answers to this inquiry
for a randomly selected user after k transmissions.

~ logn
~ xn
~ ;
mDmin n

~ users in a single antenna BC. The


Theorem 3. Assume there are n
average number of services received by a randomly selected user for

q
logk
large (k) number of channel uses behaves like km
1O
~
k
n

or equivalently

w.p.1.

Dmin

 
~ logn
~
~
n
n
:
x

m
m

b. Sufcient condition:
As is noted earlier, we have

PrfSg Pr

(~
n
\

)
Si 0 :

26

i1

Congruent to what is done in [20], it can be argued that (26) can be


Q~
upper bounded as PrfSg 6 ni1 PrfS i 0g, thus we arrive at the
following

PrfSg 6


n~

m D
mD
~
PrfS i 0g 1  1 
6 a en1 n~
~
n
i1

D log 1m
~
n

27

;
~

where (a) comes from the fact that 1  xn 6 en~x . Clearly in order
m
~ eD log 1 n~ ! 0. Since n
~ goes
to have PrfSg ! 1, we should have n
to innity, we have
mD~
n log~
n
~
n

~ eD log 1 n~  n
~ e n~ 1O n~ e
n
m

mD

Hence by setting Dmin

~ logn
~
n
m

1Omn~ :



~ logn
~
~
n
n
Rmin
x
:
~
~
loglogn
loglogn

~
n

31

~
n

(
r!!
r!!)
km
logk
km
logk
<S<
Pr
1O
1O
~
~
k
k
n
n
 
1
1o
:
k

32

28


x mn~ , the event S happens w.p.1.

~ users, for m
Theorem 2 states that in a broadcast channel with n
~
~
n
concurrent transmissions, the expected delay behaves like n log
.
m
Referring to the minimum delay Round-Robin scheduling with
~ , the corresponding Dmin of the aforementioned scheduling
D1 n
~ , which is not feasicould theoretically approaches D1 if m logn
~
n
ble as the maximum number of active users is at most loglog
. Also,
Rmin
this is interesting to note that having one active user at a given
time corresponds to the case of m = 1, thus the expected delay be~ logn
~ which is in accordance to the result reported in [17].
comes n
Finally, referring to Theorem 2 and noting the maximum possible
~
n
value of m is at most loglog
, the expected delay of a single antenna
Rmin
SA
BC, Dmin , becomes

DSA
min

0
1


km
d
km
km
B
C
~
n
1  d < S <
1 d  1  2Q @q
Pr

A:
~
~
n
n
km
1m

q
~
n
2m
1 logk
It can be veried that by setting d
and noting the
k
2
x2
1
p

for x 1, it follows
approximation Q x 2px e

~
n
Y

en~e

Proof. Recall that for each channel use, the probability that a spem
cic user is being serviced is
 n~ . Under the assumption of k channel
uses, the average number S of getting services by this user fol

lows a Binomial distribution, that is B k; mn~ .3 Using gaussian
approximation for a binomial distribution when k is large enough,
we have

29

As is proved in (3), using random beamforming strategy together


with FPAS, the MIMO BC can be thought as M parallel channels each
having m loglogn
 h1 concurrent transmissions. As a result,
Rmin
noting mM users are being serviced at a given time along one of
the existing directions (M), the probability that a given user out of
n existing users is selected is simply computed as mM
. Thus, follown
ing the same approach as is used in Theorem 2, it can be concluded


n
that Dmin n logn
x mM
. Finally, after some manipulations, the exmM
pected delay of the proposed strategy, DMA
min , behaves like

It means for sufcient large number of channel uses, the average


number of services to a randomly selected user in a single antenna
BC, S SA , scales as

SA

km
1O

~
n

r!!
logk
k

33

with probability approaching one. h


Accordingly, for the case of MIMO BC with n single antenna
users, the average number of services to a randomly selected user


is a Binomial distribution with parameters B k; mM
. Thus, noting
n
 h1 and following the same approach as is used in
m loglogn
Rmin
Theorem 3, the average number of services achieved by employing
the proposed method, S MA
FPAS , becomes (w.p.1.)

S MA
FPAS

kM loglogn

1O
nRmin

r!!
logk
:
k

34

5. Simulation results

4.2. The average number of services

In this section, we aim at providing some numerical results to


compare the analytical sum-rate capacity and maximum number
of active users to what is experimentally achieved through using
the proposed strategy. To make a better explanation for the experimental results, in the following, we will discuss the procedure
which leads to derive the numerical sum-rate capacity of the proposed method. First, for each user, a vector of size M = 4 associated
to the corresponding channel gain vector from a multi-antenna
transmitter of size M = 4 to each single antenna user is assumed.
The entries of these channel gain vectors are drawn from a
Rayleigh distribution. Then a random unitary matrix of dimension

The preceding subsection aimed at investigating the required


number of services to guarantee each user is being serviced at least

3
B(k, p) denotes a Binomial distribution with parameters k and p, respectively, as
the number of trials and the probability of success.

DMA
min




Rmin
n logn
n
:
x
loglogn
M loglogn  h1

30

2122

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

21

220

The Analytical result for the sumrate capacity, Rmin=0.1,0.2


The numerical result for the proposed method, Rmin=0.1
The numerical result for the proposed method, Rmin=0.2

Number of active users

20

SumRate

19
18
17
16
15

180
160
140
120
100
80

14
2

Total number of users

60
1

10
5
x 10

Fig. 1. The comparison results between the analytical result and the achievable
throughput of the proposed method for two values of Rmin = 0.1, 0.2 for M = 4 and
SNR = 15 dB.

M = 4, composed of M = 4 random beamforming vectors is generated and for each user, the best direction out of existing M = 4 random directions is identied. Then, according to the user selection
algorithm as is addressed in Section 3.3 and assuming the threshold for the interference power (f) is set 10 dB below the noise
power, i.e., f = .1r2 (for more details for the interference power
see the Appendix A), the potential candidates along each direction
are chosen, and the power allocation method provided in Section
3.4 is applied to these candidates. Then, the resulting SINR associated with each selected user is computed and, hence, the achievable rate is derived. Finally, the sum-rate capacity is readily
computed. The above procedure is iteratively repeated until receiving an acceptable condence interval for the obtained results.
Fig. 1 compares the achievable sum-rate throughput of the proposed method as is addressed in (24) and the sum-rate capacity of
the MIMO BC, i.e., M loglogn, which is addressed in [5], versus
the total number of users for two different values of4 Rmin = 0.1,
0.2 bps/Hz, considering the total number of transmit antennas is
M = 4. It is worth mentioning that, the asymptotic sum-rate capacity
of a MIMO BC is independent of Rmin. This conrms the obtained results in Fig. 1, showing the asymptotic throughput of the proposed
method for two distinct values of Rmin = 0.1, 0.2 coincides as the total
number of users tends to innity.
Furthermore, Fig. 1 shows there is a non-vanishing gap between
the achievable throughput through using the proposed method as
compared to that of the analytical result. However, from each
users viewpoint this non-vanishing gap is negligible as it is distributed over fairly large number of users.
Accordingly, Fig. 2 illustrates the total number of active users
for the proposed method (23) versus the total number of users
and compare it to the analytical upper bound, M loglogn
. It is worth
Rmin
mentioning that the method given in [5] has only M = 4 active
users at a time. As a result, the proposed method has a considerable improvement in the total number of active users as compared
to [5]. Moreover, throughout the simulations, the threshold (f) is
set to f = 0.1r2, meaning users for which the inter-beam interference power falls 10 dB below the noise power are merely chosen
throughout the course of user selection.
Fig. 3 is provided to compare the expected delay of the following scheduling methods: (i) the BUSS which is served as a benchmark and is addressed in [5] and (ii) the proposed FPAS, where
again the total number of transmit antennas is set to M = 4 and it

Total number of users

10
5
x 10

Fig. 2. The comparison results between the maximum number of active users (the
upper bound) and that of achieved by using the proposed method for two values of
Rmin = 0.1, 0.2, where M = 4 and SNR = 15 dB.

2000

Delay (The number of packets)

131

200

Numerical result for the proposed method, Rmin=0.1


Analytical upper bound, M(log(SNRlogn))/Rmin, Rmin=0.1
Numerical result for the proposed method, Rmin=0.2
Analytical upper bound, M(log(SNRlogn))/Rmin, Rmin=0.2

1800

BUSS, Numerical
FPAS, Numerical
FPAS, Analytical
BUSS, Analytical

1600
1400
1200
1000
800
600
400
500

550

600

650

700

750

800

850

900

950 1000

Total number of users


Fig. 3. The expected delay of the BUSS as compared to that of the FPAS, where M = 4
and Rmin is set to one.

is assumed Rmin = 1. Accordingly, for different number of users


ranging from 500 to 1000, the expected delay of these scheduling
methods are simulated and compared to the theoretical results obtained for BUSS n logn and FPAS (Eq. (30) without considering
the ordering parts, i.e., h(), x()), showing there is a close agreement between the computed delays derived by the analytical results to that of the simulations. Referring to Fig. 3, there is a
major difference between the expected delay of these two methods. Moreover, the small gap between the theoretical results to
that of the simulations is due to the fact that only the main term
of the analytical result is considered and the ordering parts are
ignored.
Fig. 4 compares the average number of getting services for a
randomly selected user for both the BUSS and FPAS methods and
for two different values of Rmin, where the total number of channel
uses k is set to DBUSS n logn, and the results are computed for
two values of Rmin = 1, 0.5. Noting k n logn, and referring to Theorem 2, the average number of services to a randomly selected user
for the BUSS with m = M concurrent transmissions and the FPAS with
mM M loglogn
concurrent transmissions respectively scale as
Rmin


 1 
2
S MA
BUSS M logn 1 O n

35

and
4

Consider a system with channel bandwidth of 200Khz. Consequently, the


transmission rate of 0.1 bps/Hz is equivalent to 0.1  200 KHz = 20 Kbps.

S MA
FPAS

 1 
M logn loglogn 
1 O n2 :
Rmin

36

2123

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

Average number of services to a


randomly selected user

160
140
120

6. Conclusion

Numerical, FPAS, Rmin=1


Analytical, FPAS, Rmin=1
Numerical, FPAS, Rmin=0.5
Analytical, FPAS, Rmin=0.5
Numerical, BUSS
Analytical, BUSS

100
80
60
40

20
500

550

600

650

700

750

800

850

900

950 1000

Total number of users


Fig. 4. The average number of services for a randomly selected user for two
different power allocation strategies, and considering Rmin = 0.1.

24

Numerical, SNR=8db
Analytical result for the SumRate capacity, SNR=8db
Numerical, SNR=15db
Analytical result for the SumRate capacity, SNR=15db

22

n logn
strated that the expected delay of FPAS behaves like Rmin loglogn



n
x loglogn
for the case of single antenna BC and generally scales



n logn
n
min
for the case of multi-antenna
x loglogn
as RM
loglognh1

BC (under the assumption of applying random beamforming strategy along the proposed user selection strategy). These are noticeably smaller than the expected delay of the Best User Selection
Strategy (BUSS) for both the single and multi-antenna case which

20
18

SumRate

This paper aims at investigating the fairness maximization in


MIMO broadcast channel, when the transmitter makes use of random beamforming. It is assumed each user either receives the minimum rate constraint of Rmin or remains silent throughout the
transmission. Accordingly, for the case of many users (n), it is
shown the maximum number of active users scales as
M loglogn
 h1, where M represents the total number of transmit
Rmin
antennas. This is achieved through using a novel power allocation
strategy, dubbed Fair Power Allocation Strategy (FPAS), which is recently introduced in [9] for single antenna broadcast channel. Noting the asymptotic sum-rate capacity of such channel is
M loglogn, the proposed method is able to approach the sumrate capacity of such channel within a constant gap, noting this
gap is negligible for the case of large users. Moreover, it is demon-

yield the expected delay of n logn and


16
14

n logn
,
MOM2 =n

respectively.

Appendix A

12
10
8
6

10
5

Total number of users

x 10

Fig. 5. Comparison results between the sum-rate capacity (dashed lines) and the
achievable throughput (solid lines) of the proposed method for two SNR values,
considering Rmin = 0.1.

Moreover, Fig. 4 indicates that there is a major gap between two


power allocation strategies, (35) and (36), when relying on the average number of getting services for a randomly selected user.
Finally, Figs. 5 and 6 are provided to show the impact of different SNR values (SNR = 8, 15 dB) on the sum-rate and maximum
number of active users, where it is assumed Rmin = 0.1, conrming
the double logarithmic and logarithmic behavior of the sum-rate
and maximum number of active users with respect to n and SNR,
respectively.

~ denotes the number of users selected by the


Theorem. Let n
beamforming strategy proposed in Section 3.1 for a given direction,

q
 f M
~
logn
^n
~ M1
1O
users for which
then there are at least n
~
n
the corresponding inter-beam interference power is less than f.

Proof. Assume P denotes the probability that the corresponding


inter-beam interference power of the user indexed by pj(i) in the
jth direction is less than f. Noting


2
M 
 2 h
X
2 i


 T

E Ipj i;2 t
hpj i Ul  E sl t
l1;lj

Number of active users

200

Numerical, SNR=8db
Analytical upper bound, M(log(SNRlogn))/Rmin, SNR=8db
Numerical, SNR=15db
Analytical upper bound, M(log(SNRlogn))/Rmin SNR=15db


2
 T

and setting zl hpl i Ul  , it follows

P Pr

M
X

zl < f :

38

l1;lj
T

Since U(l) for l = 1, . . . , M are orthonormal vectors, thus hpl i Ul is an


i.i.d random variable with the distribution

2 as the entries of hpj i , i.e.,
 T

CN 0; 1. This implies that zl hpl i Ul  are i.i.d over l with expo

2
nential distribution Prx r2 er x ux with parameter r2 = 1,

180
160
140
120

i.e., Pr zl x ex ux. According to the beamforming strategy described in Section 3.1, the jth direction is assigned to the user indexed by pj(i) such that the corresponding channel gain on this
direction has the highest value

100
80
60
1

37

l1;lj

220

M 
2
X
 T

hpj i Ul 

4
5
6
7
Total number of users

10
5
x 10

Fig. 6. The number of active users for two SNR values, considering Rmin = 0.1.

zj max zl :
16l6M

39

As a result, the distribution of zj is the distribution of the maximum


value of M i.i.d random variables with v2(2) distribution

2124

S. Akhlaghi, A. Borhani / Computer Communications 33 (2010) 21162124

F zj x F zmax x Przmax 6 x Prz1 6 x; z2 6 x; . . . ; zM 6 x


M

References

x M

Prz1 6 x 1  e :
On the other hand, the probability that the inter-beam interference
power, which is dened in (38), is lower than a certain threshold,
i.e., f, can be lower bounded as follows

(
P Pr

M
X

zl < f

l1;lj

Pr zj <

f
M1

P Pr fM  1zmax < fg

M
f
1  eM1 :

40

Noting 1  x x2 P ex , thus one can simplies the above lower


bound as follows


M
f

M
f
f
P P 1  eM1
P
1 2
M1
M1

M
f

for f  1:
M1

!M

41

^ for which the inConsequently, the minimum number of users n


ter-beam interference
is lower than f makes a Binomial distri  term

M 5
f
~ ; M1
. Using gaussian approximation for a
bution, that is B n
~ is large enough, we have
binomial distribution when n

( 
)
M

M
f
f
~
^
~
Pr n
1  d < n < n
1 d
M1
M1
1
0
 f M
~
C
B
dn
M1

C
r
 1  2Q B

A:
@  M 


M
f
f
~
1

n
M1
M1

42

r


M
~
2 M1
f 1 logn
It can be veried that by setting d
and noting the
~
n
2
x2
1
e
for
x

1,
it
follows
approximation Q x p
2px

r!!
r!!)
~
~
logn
logn
^
<n<k 1O
Pr k 1  O
~
~
n
n
 
1
1o
;
~
n


43

M
f
. It
M1 

~
^
where k n
means for sufcient large number of users, n
q
 f M
~
logn
~ M1
with probability approaching
1O
scales as n
~
n
one.

5
B(n, p) denotes a Binomial distribution with parameters n and p, respectively, as
the number of trials and the probability of success.

[1] G. Caire, S. Shamai, On the achievable throughput of a multiantenna Gaussian


broadcast channel, IEEE Trans. Inf. Theory 49 (7) (2003) 16911706.
[2] S. Vishwanath, N. Jindal, A. Goldsmith, Duality, achievable rates, and sum-rate
capacity of Gaussian mimo broadcast channels, IEEE Trans. Inf. Theory 49 (10)
(2003) 26582668.
[3] H. Weingarten, Y. Steinberg, S. Shamai, The capacity region of the Gaussian
multiple-input multiple-output broadcast channel, IEEE Trans. Inf. Theory 52
(9) (2006) 39363964.
[4] M. Costa, Writing on dirty paper, IEEE Trans. Inf. Theory 29 (3) (1983) 439441.
[5] M. Sharif, B. Hassibi, On the capacity of mimo broadcast channels with partial
side information, IEEE Trans. Inf. Theory 51 (2) (2005) 506522.
[6] K. Zhang, Z. Niu, Random beamforming with multi-beam selection for MIMO
broadcast channels, IEEE Int. Conf. on Commun. 9 (2006) 41914195.
[7] H.C. Yang, P. Lu, H.K. Sung, Y.C. Ko, Exact sum-rate analysis for mimo broadcast
channels with random unitary beamforming based on quantized sinr feedback,
in: Proc. Int. Conf. Commun., May 2008, pp. 41914195.
[8] H. Wang, A.B. Gershman, T. Kirubarajan, Random unitary beamforming with
partial feedback for multi-antenna downlink transmission using multiuser
diversity, in: IEEE VTC, vol. 1, Spring 2005, pp. 216220.
[9] H. Keshavarz, L.L. Xie, R.R. Mazumdar, On the optimal number of active
receivers in fading broadcast channels, IEEE Trans. Inf. Theory 54 (3) (2008)
13231327.
[10] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whiting, R. Vijayakumar,
Providing quality of service over a shared wireless link, IEEE Commun. Mag. 39
(2) (2001) 150154.
[11] R.A. Berry, R.G. Gallager, Communication over fading channels with delay
constraints, IEEE Trans. Inf. Theory 48 (2002) 11351149.
[12] S. Toumpis, A.J. Goldsmith, Large wireless networks under fading, mobility,
and delay constraints, in: Proc. IEEE INFOCOM, March 2004, pp. 609619.
[13] N. Bansal, Z. Liu, Capacity, delay and mobility in wireless ad-hoc networks, in:
Proc. IEEE INFOCOM, vol. 52(6), April 2003, pp. 15531563.
[14] A. El Gamal, J. Mammen, B. Prabhakar, D. Shah, Optimal throughput-delay
scaling in wireless networks Part I: The uid model, IEEE Trans. Inf. Theory
52 (6) (2006) 25682592.
[15] P.K. Gopala, H. El Gamal, On the throughput-delay tradeoff in cellular
multicast, Int. Conf. Wirel. Netw., Commun. Mobile Comput. 2 (2005) 1401
1406.
[16] T. Cover, J. Thomas, Elements of Information Theory, Wiley, New York, 1991.
[17] M. Sharif, B. Hassibi, A delay analysis for opportunistic transmission in fading
broadcast channels, in: Proc. IEEE INFOCOM, vol. 4, March 2005, pp. 2720
2730.
[18] D.E. Knuth, Big omicron and big omega and big theta, in: ACM SIGACT News,
AprilJune 1967, pp. 1824.
[19] B. Hassibi, T.L. Marzetta, Multiple-antennas and isotropically random unitary
inputs: the received signal density in closed form, IEEE Trans. Inf. Theory 48 (6)
(2002) 14731484.
[20] A. Bayesteh, M.A. Sadrabadi, A.K. Khandani, Throughput and Fairness
Maximization in Wireless Networks, CWIT, June 2007, pp. 168171.

You might also like